fix: properly handle tags from Qwen/DeepSeek models

Models like Qwen and DeepSeek wrap reasoning content in  ...  XML-like
tags within the regular content field. This was causing the reasoning
text to appear twice - once as a reasoning block and once as regular text.

Changes:

1. Provider hooks (providers.go):
   - Extract reasoning from  tags and emit proper reasoning events
   - Use openai provider directly with custom ExtraContentFunc and
     StreamExtraFunc hooks to parse thinking content

2. Stream filtering (stream.go):
   - Filter out all text content between  and  tags at the
     streaming level to prevent duplicate rendering
   - Track state with inThinkTag flag across stream chunks

3. Message conversion (content.go):
   - Strip any remaining  tags from text content when converting
     from fantasy messages

The regex patterns use string concatenation to avoid XML tag corruption:
  regexp.MustCompile( +  +  +  +  +  +  + )

Fixes duplicate reasoning text when using custom provider with models
that wrap thinking in  tags.

.gitignore

@@ -1,7 +1,8 @@
 .aider*
 .task/
 .env
-.kit/
+.kit/*
+!.kit/extensions/
 aidocs/
 *.log
 /kit

.kit/extensions/go-edit-lint.go

@@ -0,0 +1,228 @@
+//go:build ignore
+
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"kit/ext"
+)
+
+const (
+	diagnosticsTimeout = 20 * time.Second
+	maxOutputBytes     = 12_000
+)
+
+type toolPathInput struct {
+	Path string `json:"path"`
+}
+
+type lintResult struct {
+	Output string
+	Err    error
+}
+
+func Init(api ext.API) {
+	api.OnSessionStart(func(_ ext.SessionStartEvent, ctx ext.Context) {
+		ctx.Print("go-edit-lint extension loaded - will run gopls and golangci-lint on Go file edits")
+	})
+
+	api.OnToolResult(func(e ext.ToolResultEvent, ctx ext.Context) *ext.ToolResultResult {
+		if e.IsError || !isEditOrWrite(e.ToolName) {
+			return nil
+		}
+
+		absPath, ok := resolveGoFilePath(e.Input, ctx.CWD)
+		if !ok {
+			return nil
+		}
+
+		report := runGoDiagnostics(ctx.CWD, absPath)
+		
+		// Check if there are issues and add explicit prompt for the LLM to react
+		goplsIssues, lintIssues := countIssues(report)
+		hasIssues := goplsIssues > 0 || lintIssues > 0
+		
+		var enhanced string
+		if hasIssues {
+			enhanced = e.Content + "\n\n" + report + "\n\n⚠️ DIAGNOSTICS FOUND: Please review the issues above and fix them before proceeding."
+		} else {
+			enhanced = e.Content + "\n\n" + report
+		}
+
+		// Show TUI message block for diagnostics visibility (only if there are issues)
+		if hasIssues {
+			var msgLines []string
+			msgLines = append(msgLines, fmt.Sprintf("File: %s", filepath.Base(absPath)))
+			if goplsIssues > 0 {
+				msgLines = append(msgLines, fmt.Sprintf("gopls: %d issue(s)", goplsIssues))
+			}
+			if lintIssues > 0 {
+				msgLines = append(msgLines, fmt.Sprintf("golangci-lint: %d issue(s)", lintIssues))
+			}
+			msgLines = append(msgLines, "", "⚠️ Please fix these issues before proceeding.")
+
+			borderColor := "#f9e2af" // yellow
+			if goplsIssues > 0 && lintIssues > 0 {
+				borderColor = "#f38ba8" // red
+			}
+
+			ctx.PrintBlock(ext.PrintBlockOpts{
+				Text:        strings.Join(msgLines, "\n"),
+				BorderColor: borderColor,
+				Subtitle:    "go-edit-lint",
+			})
+		}
+
+		return &ext.ToolResultResult{Content: &enhanced}
+	})
+}
+
+func isEditOrWrite(toolName string) bool {
+	return strings.EqualFold(toolName, "edit") || strings.EqualFold(toolName, "write")
+}
+
+func resolveGoFilePath(inputJSON, cwd string) (string, bool) {
+	var args toolPathInput
+	if err := json.Unmarshal([]byte(inputJSON), &args); err != nil || args.Path == "" {
+		return "", false
+	}
+
+	absPath := args.Path
+	if !filepath.IsAbs(absPath) {
+		absPath = filepath.Join(cwd, absPath)
+	}
+
+	if strings.ToLower(filepath.Ext(absPath)) != ".go" {
+		return "", false
+	}
+
+	return absPath, true
+}
+
+func runGoDiagnostics(cwd, absPath string) string {
+	gopls := runGopls(cwd, absPath)
+	lint := runGolangCILint(cwd, "./...")
+
+	return fmt.Sprintf(
+		"<go_diagnostics file=%q>\n[gopls]\n%s\n\n[golangci-lint]\n%s\n</go_diagnostics>",
+		filepath.Base(absPath),
+		formatToolResult(gopls, "No diagnostics."),
+		formatToolResult(lint, "No lint issues."),
+	)
+}
+
+func runGopls(cwd, absPath string) lintResult {
+	ctx, cancel := context.WithTimeout(context.Background(), diagnosticsTimeout)
+	defer cancel()
+
+	cmd := exec.CommandContext(ctx, "gopls", "check", absPath)
+	cmd.Dir = cwd
+	out, err := cmd.CombinedOutput()
+
+	if ctx.Err() == context.DeadlineExceeded {
+		return lintResult{Err: fmt.Errorf("timed out after %s", diagnosticsTimeout)}
+	}
+
+	if err != nil {
+		return lintResult{Output: truncate(string(out), maxOutputBytes), Err: fmt.Errorf("failed to run gopls check: %w", err)}
+	}
+
+	return lintResult{Output: truncate(string(out), maxOutputBytes)}
+}
+
+func runGolangCILint(cwd, target string) lintResult {
+	ctx, cancel := context.WithTimeout(context.Background(), diagnosticsTimeout)
+	defer cancel()
+
+	args := []string{
+		"run",
+		target,
+		"--show-stats=false",
+		"--output.text.path", "stdout",
+		"--output.text.colors=false",
+		"--output.text.print-issued-lines=false",
+	}
+	cmd := exec.CommandContext(ctx, "golangci-lint", args...)
+	cmd.Dir = cwd
+	out, err := cmd.CombinedOutput()
+
+	if ctx.Err() == context.DeadlineExceeded {
+		return lintResult{Err: fmt.Errorf("timed out after %s", diagnosticsTimeout)}
+	}
+
+	trimmed := truncate(string(out), maxOutputBytes)
+	if err == nil {
+		return lintResult{Output: trimmed}
+	}
+
+	exitErr, ok := err.(*exec.ExitError)
+	if ok && exitErr.ExitCode() == 1 {
+		return lintResult{Output: trimmed}
+	}
+
+	return lintResult{Output: trimmed, Err: fmt.Errorf("failed to run golangci-lint: %w", err)}
+}
+
+func formatToolResult(res lintResult, emptyFallback string) string {
+	var lines []string
+	if res.Err != nil {
+		lines = append(lines, "ERROR: "+res.Err.Error())
+	}
+	out := strings.TrimSpace(res.Output)
+	if out == "" {
+		if res.Err == nil {
+			lines = append(lines, emptyFallback)
+		}
+	} else {
+		lines = append(lines, out)
+	}
+	if len(lines) == 0 {
+		return emptyFallback
+	}
+	return strings.Join(lines, "\n")
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max] + "\n... output truncated ..."
+}
+
+func countIssues(report string) (goplsCount, lintCount int) {
+	// Extract gopls section
+	goplsStart := strings.Index(report, "[gopls]")
+	lintStart := strings.Index(report, "[golangci-lint]")
+	endTag := strings.Index(report, "</go_diagnostics>")
+
+	if goplsStart != -1 && lintStart != -1 {
+		goplsSection := report[goplsStart:lintStart]
+		// Count non-empty lines excluding the header and "No diagnostics." message
+		for _, line := range strings.Split(goplsSection, "\n") {
+			line = strings.TrimSpace(line)
+			if line != "" && line != "[gopls]" && line != "No diagnostics." {
+				goplsCount++
+			}
+		}
+	}
+
+	if lintStart != -1 && endTag != -1 {
+		lintSection := report[lintStart:endTag]
+		// Count non-empty lines excluding the header and "No lint issues." message
+		for _, line := range strings.Split(lintSection, "\n") {
+			line = strings.TrimSpace(line)
+			if line != "" && line != "[golangci-lint]" && line != "No lint issues." {
+				lintCount++
+			}
+		}
+	}
+
+	return goplsCount, lintCount
+}

.kit/extensions/subagent-monitor.go

@@ -0,0 +1,304 @@
+//go:build ignore
+
+// subagent-monitor — live horizontal widget strip for spawned subagents
+//
+// Subscribes to subagents spawned by the main Kit agent and displays a
+// single widget just above the input box. Each subagent occupies one column
+// in a side-by-side horizontal layout. Columns show scrolling real-time
+// output as the subagent works. When a subagent finishes its column is
+// removed automatically.
+//
+// Yaegi-safe design notes:
+// - No sync.Mutex (Yaegi has reflection issues with sync primitives)
+// - No channels in maps (Yaegi panics on range over map[string]chan)
+// - All ctx.* calls guarded with nil checks
+// - Simple data structures only
+package main
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"kit/ext"
+)
+
+// ---------------------------------------------------------------------------
+// Per-subagent state
+// ---------------------------------------------------------------------------
+
+type submonEntry struct {
+	id      int
+	callID  string
+	task    string
+	lines   []string
+	started time.Time
+	elapsed time.Duration
+}
+
+const (
+	submonColWidth = 34 // visible character width per column
+	submonMaxLines = 5  // scrolling output lines per column
+	submonColGap   = 2  // spaces between columns
+)
+
+// ---------------------------------------------------------------------------
+// Package-level state - all simple types
+// ---------------------------------------------------------------------------
+
+var (
+	submonCtx     ext.Context
+	submonHasCtx  bool
+	submonEntries []*submonEntry
+	submonNextID  int
+)
+
+func submonInit() {
+	submonEntries = nil
+	submonNextID = 1
+}
+
+// ---------------------------------------------------------------------------
+// String helpers
+// ---------------------------------------------------------------------------
+
+func submonPad(s string, w int) string {
+	r := []rune(s)
+	if len(r) >= w {
+		return string(r[:w])
+	}
+	return s + strings.Repeat(" ", w-len(r))
+}
+
+func submonTrunc(s string, w int) string {
+	r := []rune(s)
+	if len(r) <= w {
+		return s
+	}
+	if w <= 1 {
+		return "…"
+	}
+	return string(r[:w-1]) + "…"
+}
+
+// ---------------------------------------------------------------------------
+// Widget rendering
+// ---------------------------------------------------------------------------
+
+func submonRenderColumn(e *submonEntry) []string {
+	var rows []string
+
+	// Calculate elapsed time on-demand to avoid race conditions with ticker
+	elapsed := e.elapsed
+	if elapsed == 0 && !e.started.IsZero() {
+		elapsed = time.Since(e.started)
+	}
+	secs := int(elapsed.Seconds())
+	timeStr := fmt.Sprintf("%ds", secs)
+	taskMax := submonColWidth - len(timeStr) - 3
+	taskPart := submonTrunc(e.task, taskMax)
+	header := fmt.Sprintf("#%d %s  %s", e.id, taskPart, timeStr)
+	rows = append(rows, submonPad(header, submonColWidth))
+
+	display := e.lines
+	if len(display) > submonMaxLines {
+		display = display[len(display)-submonMaxLines:]
+	}
+	for _, l := range display {
+		rows = append(rows, submonPad("  "+submonTrunc(l, submonColWidth-2), submonColWidth))
+	}
+	for len(rows) < submonMaxLines+1 {
+		if len(rows) == 1 && len(e.lines) == 0 {
+			rows = append(rows, submonPad("  waiting…", submonColWidth))
+		} else {
+			rows = append(rows, strings.Repeat(" ", submonColWidth))
+		}
+	}
+	return rows
+}
+
+func submonBuildWidget() string {
+	if len(submonEntries) == 0 {
+		return ""
+	}
+
+	numCols := len(submonEntries)
+	numRows := submonMaxLines + 1
+	cols := make([][]string, numCols)
+	for i, e := range submonEntries {
+		rows := submonRenderColumn(e)
+		col := make([]string, numRows)
+		for j := 0; j < numRows; j++ {
+			if j < len(rows) {
+				col[j] = rows[j]
+			} else {
+				col[j] = strings.Repeat(" ", submonColWidth)
+			}
+		}
+		cols[i] = col
+	}
+
+	gap := strings.Repeat(" ", submonColGap)
+	var sb strings.Builder
+	for row := 0; row < numRows; row++ {
+		for ci := range cols {
+			if ci > 0 {
+				sb.WriteString(gap)
+			}
+			sb.WriteString(cols[ci][row])
+		}
+		if row < numRows-1 {
+			sb.WriteString("\n")
+		}
+	}
+	return sb.String()
+}
+
+func submonPushWidget() {
+	if !submonHasCtx {
+		return
+	}
+	if submonCtx.SetWidget == nil {
+		return
+	}
+
+	text := submonBuildWidget()
+	if len(submonEntries) == 0 {
+		if submonCtx.RemoveWidget != nil {
+			submonCtx.RemoveWidget("submon")
+		}
+		return
+	}
+	submonCtx.SetWidget(ext.WidgetConfig{
+		ID:        "submon",
+		Placement: ext.WidgetAbove,
+		Content:   ext.WidgetContent{Text: text},
+		Style:     ext.WidgetStyle{BorderColor: "#89b4fa"},
+		Priority:  0,
+	})
+}
+
+func submonAppendLine(e *submonEntry, line string) {
+	line = strings.TrimRight(line, "\r\n")
+	if strings.TrimSpace(line) == "" {
+		return
+	}
+	e.lines = append(e.lines, line)
+}
+
+// ---------------------------------------------------------------------------
+// Init
+// ---------------------------------------------------------------------------
+
+func Init(api ext.API) {
+	submonInit()
+
+	api.OnSessionStart(func(_ ext.SessionStartEvent, ctx ext.Context) {
+		submonCtx = ctx
+		submonHasCtx = true
+		submonInit()
+		if ctx.RemoveWidget != nil {
+			ctx.RemoveWidget("submon")
+		}
+	})
+
+	api.OnAgentEnd(func(_ ext.AgentEndEvent, ctx ext.Context) {
+		submonCtx = ctx
+		submonHasCtx = true
+	})
+
+	// ── SubagentStart ────────────────────────────────────────────────────────
+	api.OnSubagentStart(func(e ext.SubagentStartEvent, ctx ext.Context) {
+		submonCtx = ctx
+		submonHasCtx = true
+
+		id := submonNextID
+		submonNextID++
+		entry := &submonEntry{
+			id:      id,
+			callID:  e.ToolCallID,
+			task:    e.Task,
+			started: time.Now(),
+		}
+		submonEntries = append(submonEntries, entry)
+
+		submonPushWidget()
+	})
+
+	// ── SubagentChunk ────────────────────────────────────────────────────────
+	api.OnSubagentChunk(func(e ext.SubagentChunkEvent, ctx ext.Context) {
+		submonCtx = ctx
+		submonHasCtx = true
+
+		var entry *submonEntry
+		for _, en := range submonEntries {
+			if en.callID == e.ToolCallID {
+				entry = en
+				break
+			}
+		}
+		if entry == nil {
+			return
+		}
+
+		switch e.ChunkType {
+		case "text":
+			for _, line := range strings.Split(e.Content, "\n") {
+				submonAppendLine(entry, line)
+			}
+		case "tool_call":
+			submonAppendLine(entry, "→ "+e.ToolName)
+		case "tool_execution_start":
+			submonAppendLine(entry, "⚙ "+e.ToolName)
+		case "tool_result":
+			if e.IsError {
+				submonAppendLine(entry, "✗ "+e.ToolName)
+			} else {
+				submonAppendLine(entry, "✓ "+e.ToolName)
+			}
+		}
+
+		submonPushWidget()
+	})
+
+	// ── SubagentEnd ──────────────────────────────────────────────────────────
+	api.OnSubagentEnd(func(e ext.SubagentEndEvent, ctx ext.Context) {
+		submonCtx = ctx
+		submonHasCtx = true
+
+		var entry *submonEntry
+		for _, en := range submonEntries {
+			if en.callID == e.ToolCallID {
+				entry = en
+				break
+			}
+		}
+		if entry != nil {
+			entry.elapsed = time.Since(entry.started)
+			if e.ErrorMsg != "" {
+				submonAppendLine(entry, "✗ "+submonTrunc(e.ErrorMsg, submonColWidth-2))
+			}
+		}
+
+		submonPushWidget()
+
+		// Remove the entry immediately (no goroutine to avoid races)
+		newEntries := submonEntries[:0]
+		for _, en := range submonEntries {
+			if en.callID != e.ToolCallID {
+				newEntries = append(newEntries, en)
+			}
+		}
+		submonEntries = newEntries
+		submonPushWidget()
+	})
+
+	// ── SessionShutdown ──────────────────────────────────────────────────────
+	api.OnSessionShutdown(func(_ ext.SessionShutdownEvent, ctx ext.Context) {
+		submonInit()
+		// Guard ctx access - may be nil during shutdown
+		if ctx.RemoveWidget != nil {
+			ctx.RemoveWidget("submon")
+		}
+	})
+}

AGENTS.md

@@ -1,22 +1,3 @@
-<!-- OPENSPEC:START -->
-# OpenSpec Instructions
-
-These instructions are for AI assistants working in this project.
-
-Always open `@/openspec/AGENTS.md` when the request:
-- Mentions planning or proposals (words like proposal, spec, change, plan)
-- Introduces new capabilities, breaking changes, architecture shifts, or big performance/security work
-- Sounds ambiguous and you need the authoritative spec before coding
-
-Use `@/openspec/AGENTS.md` to learn:
-- How to create and apply change proposals
-- Spec format and conventions
-- Project structure and guidelines
-
-Keep this managed block so 'openspec update' can refresh the instructions.
-
-<!-- OPENSPEC:END -->
-
 # KIT Agent Guidelines
 
 ## Build/Test Commands
@@ -42,6 +23,33 @@ Keep this managed block so 'openspec update' can refresh the instructions.
 - **Extension system** (`internal/extensions/`): Yaegi-interpreted Go, 13 lifecycle events, custom tools/commands/widgets/overlays/editor interceptors
 - **TUI** (`internal/ui/`): Bubble Tea v2 parent-child model (`AppModel` → `InputComponent`, `StreamComponent`, etc.)
 - **Decoupling pattern**: `cmd/root.go` has converter functions (e.g. `widgetProviderForUI()`) that bridge `internal/extensions/` types to `internal/ui/` types — the UI never imports extensions directly
+- **Public SDK** (`pkg/kit/`): The public-facing Go SDK for embedding Kit as a library. See rules below.
+
+## Public SDK (`pkg/kit/`) Rules
+
+`pkg/kit/` is the **public API surface** consumed by external Go developers. All exported symbols, types, function names, and godoc comments in this package are part of the SDK contract.
+
+### No Dependency Name Leakage
+Internal dependency names (e.g. `charm.land/fantasy`, library-specific jargon) **must not** appear in:
+- **Exported function/method names** — use generic terms (`LLM`, `Provider`, `Message`) instead of library names
+- **Exported type names** — type aliases should use domain names (e.g. `LLMMessage`, not `FantasyMessage`)
+- **Godoc comments** on exported symbols — these are visible in `go doc` output and pkg.go.dev
+- **Struct field names and tags** on exported types
+
+Using dependency types directly in **function bodies** (private implementation) is fine — that's invisible to SDK consumers.
+
+### Naming Conventions for SDK Symbols
+- Type aliases re-exporting dependency types: use `LLM*` prefix (e.g. `LLMMessage`, `LLMUsage`, `LLMResponse`)
+- Conversion helpers: use `ConvertToLLM*` / `ConvertFromLLM*` (not the dependency name)
+- Provider queries: use `GetLLMProviders` (not `GetFantasyProviders`)
+- When wrapping internal methods, the `pkg/kit/` name should be dependency-agnostic even if the `internal/` method still uses the old name
+
+### Deprecation Pattern
+When renaming a public SDK symbol, keep the old name as a deprecated wrapper for one release cycle:
+```go
+// Deprecated: Use NewName instead.
+func OldName() { return NewName() }
+```
 
 ## Key Patterns
 

README.md

@@ -18,10 +18,12 @@ A powerful, extensible AI coding agent CLI with multi-provider support, built-in
 ## Features
 
 - **Multi-Provider LLM Support**: Anthropic, OpenAI, Google Gemini, Ollama, Azure OpenAI, AWS Bedrock, OpenRouter, and more
-- **Built-in Core Tools**: bash, read, write, edit, grep, find, ls, spawn_subagent - no MCP overhead
+- **Built-in Core Tools**: bash, read, write, edit, grep, find, ls, subagent - no MCP overhead
 - **MCP Integration**: Connect external MCP servers for expanded capabilities
 - **Extension System**: Write custom tools, commands, widgets, and UI modifications in Go
-- **Theming**: 22 built-in color themes (KITT, Catppuccin, Dracula, Nord, etc.) with runtime switching and custom theme files
+- **Theming**: 22 built-in color themes (KITT, Catppuccin, Dracula, Nord, etc.) with runtime switching, persistence, and custom theme files
+- **Model Persistence**: Model and thinking level selections are automatically saved and restored across sessions
+- **Prompt Templates**: Create reusable prompt templates with shell-style argument substitution
 - **Interactive TUI**: Rich terminal interface powered by Bubble Tea with streaming, syntax highlighting, and custom rendering
 - **Session Management**: Tree-based conversation history with branching support
 - **Non-Interactive Mode**: Script-friendly positional args with JSON output
@@ -71,6 +73,9 @@ kit @main.go @test.go "Review these files"
 # Continue the most recent session
 kit --continue
 
+# Model and thinking level selections are automatically persisted
+# across sessions and restored on next launch
+
 # Use specific model
 kit --model anthropic/claude-sonnet-latest
 ```
@@ -177,6 +182,8 @@ mcpServers:
 # Extensions
 --extension, -e          Load additional extension file(s) (repeatable)
 --no-extensions          Disable all extensions
+--prompt-template        Load a specific prompt template by name
+--no-prompt-templates    Disable prompt template loading
 
 # Generation parameters
 --max-tokens             Maximum tokens in response (default: 4096)
@@ -202,7 +209,7 @@ kit auth status              # Check authentication status
 
 # Model database
 kit models [provider]        # List available models (optionally filter by provider)
-kit models --all             # Show all providers (not just Fantasy-compatible)
+kit models --all             # Show all providers (not just LLM-compatible)
 kit update-models [source]   # Update model database (from models.dev, URL, file, or 'embedded')
 
 # Extension management
@@ -232,6 +239,8 @@ Kit ships with 22 built-in color themes that control all UI elements. Switch at
 /theme tokyonight
 ```
 
+Theme selections are automatically saved and restored on next launch (stored in `~/.config/kit/preferences.yml`). This persistence also applies to **model** and **thinking level** selections — all are saved together and restored on startup.
+
 ### Custom themes
 
 Drop a `.yml` file in `~/.config/kit/themes/` (user) or `.kit/themes/` (project):
@@ -278,7 +287,7 @@ kit -e examples/extensions/minimal.go
 
 ### Extension Capabilities
 
-**Lifecycle Events**: OnSessionStart, OnSessionShutdown, OnBeforeAgentStart, OnAgentStart, OnAgentEnd, OnToolCall, OnToolExecutionStart, OnToolExecutionEnd, OnToolResult, OnInput, OnMessageStart, OnMessageUpdate, OnMessageEnd, OnModelChange, OnContextPrepare, OnBeforeFork, OnBeforeSessionSwitch, OnBeforeCompact
+**Lifecycle Events**: OnSessionStart, OnSessionShutdown, OnBeforeAgentStart, OnAgentStart, OnAgentEnd, OnToolCall, OnToolExecutionStart, OnToolOutput, OnToolExecutionEnd, OnToolResult, OnInput, OnMessageStart, OnMessageUpdate, OnMessageEnd, OnModelChange, OnContextPrepare, OnBeforeFork, OnBeforeSessionSwitch, OnBeforeCompact, OnCustomEvent, OnSubagentStart, OnSubagentChunk, OnSubagentEnd
 
 **Custom Components**:
 - **Tools**: Add new tools the LLM can invoke
@@ -298,6 +307,12 @@ kit -e examples/extensions/minimal.go
 - **Themes**: Register and switch color themes via `RegisterTheme`, `SetTheme`, `ListThemes`
 - **Custom Events**: Inter-extension communication via `EmitCustomEvent`
 
+**Bridged SDK APIs** (NEW): Extensions can now access internal SDK capabilities:
+- **Tree Navigation**: Navigate conversation history (`GetTreeNode`, `GetCurrentBranch`, `NavigateTo`), summarize branches (`SummarizeBranch`), and implement fresh context loops (`CollapseBranch`)
+- **Skill Loading**: Dynamically load and inject skills at runtime (`LoadSkill`, `DiscoverSkills`, `InjectSkillAsContext`)
+- **Template Parsing**: Parse and render templates with `{{variables}}` (`ParseTemplate`, `RenderTemplate`), parse CLI-style arguments (`ParseArguments`, `SimpleParseArguments`), and evaluate model conditionals (`EvaluateModelConditional`, `RenderWithModelConditionals`)
+- **Model Resolution**: Resolve model fallback chains (`ResolveModelChain`), query model capabilities (`GetModelCapabilities`, `CheckModelAvailable`), and extract provider/model ID (`GetCurrentProvider`, `GetCurrentModelID`)
+
 ### Extension Examples
 
 See the `examples/extensions/` directory:
@@ -309,6 +324,7 @@ See the `examples/extensions/` directory:
 - `compact-notify.go` - Notification on compaction
 - `confirm-destructive.go` - Confirm destructive operations
 - `context-inject.go` - Inject context into conversations
+- `conversation-manager.go` - **NEW** Tree navigation, branch summarization, and fresh context loops
 - `custom-editor-demo.go` - Vim-like modal editor
 - `dev-reload.go` - Development live-reload
 - `header-footer-demo.go` - Custom headers and footers
@@ -323,6 +339,7 @@ See the `examples/extensions/` directory:
 - `plan-mode.go` - Read-only planning mode
 - `project-rules.go` - Project-specific rules
 - `prompt-demo.go` - Interactive prompts (select/confirm/input)
+- `prompt-templates.go` - **NEW** Frontmatter-driven templates with model switching and skill injection
 - `protected-paths.go` - Path protection for sensitive files
 - `subagent-widget.go` - Multi-agent orchestration with status widget
 - `subagent-test.go` - Subagent testing utilities
@@ -332,6 +349,8 @@ See the `examples/extensions/` directory:
 - `tool-renderer-demo.go` - Custom tool call rendering
 - `widget-status.go` - Persistent status widgets
 
+Also see `.kit/extensions/go-edit-lint.go` (in this repo) for a project-local extension example that runs gopls and golangci-lint on Go file edits.
+
 ### Loading Extensions
 
 **Auto-discovery** (loads automatically):
@@ -389,6 +408,32 @@ func TestMyExtension(t *testing.T) {
 
 See `examples/extensions/tool-logger_test.go` for a complete example with 14 test cases covering tool calls, input handling, and session lifecycle.
 
+### Prompt Templates
+
+Create reusable prompt templates with shell-style argument substitution. Templates are loaded from `~/.kit/prompts/*.md` and `.kit/prompts/*.md`.
+
+**Example template** (`~/.kit/prompts/review.md`):
+```markdown
+---
+description: Review code for issues
+---
+Review the following code for bugs and security issues.
+Focus on $1 specifically.
+```
+
+**Usage:**
+```
+/review error handling
+```
+
+**Argument placeholders:**
+- `$1`, `$2`, etc. — Individual arguments
+- `$@` or `$ARGUMENTS` — All arguments
+- `${@:2}` — Arguments from position 2 onwards
+- `${@:1:3}` — 3 arguments starting at position 1
+
+Disable templates with `--no-prompt-templates` or load a specific template with `--prompt-template <name>`.
+
 ## Session Management
 
 Kit uses a tree-based session model that supports branching and forking conversations.
@@ -419,6 +464,22 @@ kit -s path/to/session.jsonl
 kit --no-session
 ```
 
+### Interactive Session Commands
+
+During an interactive session, use these slash commands:
+
+| Command | Description |
+|---------|-------------|
+| `/name [name]` | Set or display the session's display name |
+| `/session` | Show session info (path, ID, message count) |
+| `/resume` | Open the session picker to switch sessions |
+| `/export [path]` | Export session as JSONL (auto-generates path if omitted) |
+| `/import <path>` | Import and switch to a session from a JSONL file |
+| `/share` | Upload session to GitHub Gist and get a shareable viewer URL |
+| `/tree` | Navigate the session tree |
+| `/fork` | Branch from an earlier message |
+| `/new` | Start a fresh session |
+
 ## Go SDK
 
 Embed Kit in your Go applications:
@@ -441,7 +502,7 @@ func main() {
     if err != nil {
         log.Fatal(err)
     }
-    defer host.Close()
+    defer func() { _ = host.Close() }()
     
     // Send a prompt
     response, err := host.Prompt(ctx, "What is 2+2?")
@@ -482,23 +543,26 @@ host, err := kit.New(ctx, &kit.Options{
 ### With Callbacks
 
 ```go
-response, err := host.PromptWithCallbacks(
+unsub := host.OnToolCall(func(e kit.ToolCallEvent) {
+    println("Calling tool:", e.ToolName)
+})
+defer unsub()
+
+unsub2 := host.OnToolResult(func(e kit.ToolResultEvent) {
+    if e.IsError {
+        println("Tool failed:", e.ToolName)
+    }
+})
+defer unsub2()
+
+unsub3 := host.OnStreaming(func(e kit.MessageUpdateEvent) {
+    print(e.Chunk)
+})
+defer unsub3()
+
+response, err := host.Prompt(
     ctx,
     "List files in current directory",
-    func(name, args string) {
-        // Tool call started
-        println("Calling tool:", name)
-    },
-    func(name, args, result string, isError bool) {
-        // Tool call completed
-        if isError {
-            println("Tool failed:", name)
-        }
-    },
-    func(chunk string) {
-        // Streaming text chunk
-        print(chunk)
-    },
 )
 ```
 
@@ -651,8 +715,24 @@ npm/                 - NPM package wrapper for distribution
 - **Google Vertex** - Claude on Vertex AI
 - **OpenRouter** - Multi-provider router
 - **Vercel AI** - Vercel AI SDK models
+- **Custom** - Any OpenAI-compatible endpoint via `--provider-url`
 - **Auto-routed** - Any provider from models.dev database
 
+### Custom Provider
+
+Use `custom/custom` when pointing Kit at any OpenAI-compatible endpoint with `--provider-url`:
+
+```bash
+kit --provider-url "http://localhost:8080/v1" "Hello"
+```
+
+This automatically defaults to `custom/custom` without needing to specify a model. The custom provider routes through the `openaicompat` provider and supports:
+
+- Zero cost tracking (input/output = 0)
+- 262K context window, 65K output limit
+- Reasoning and temperature support
+- Optional `CUSTOM_API_KEY` environment variable or `--provider-api-key` flag
+
 ### Model String Format
 
 ```bash

btca.config.jsonc

@@ -76,6 +76,18 @@
       "name": "opencode",
       "url": "https://github.com/anomalyco/opencode",
       "branch": "dev"
+    },
+    {
+      "type": "git",
+      "name": "herald",
+      "url": "https://github.com/indaco/herald",
+      "branch": "main"
+    },
+    {
+      "type": "git",
+      "name": "herald-md",
+      "url": "https://github.com/indaco/herald-md",
+      "branch": "main"
     }
   ],
   "model": "claude-haiku-4-5",

cmd/auth.go

@@ -1,9 +1,13 @@
 package cmd
 
 import (
+	"context"
 	"fmt"
+	"net"
+	"net/http"
 	"os"
 	"strings"
+	"time"
 
 	"charm.land/huh/v2"
 	"github.com/mark3labs/kit/internal/auth"
@@ -14,7 +18,7 @@ import (
 // authCmd represents the auth command for managing AI provider authentication.
 // This command provides subcommands for login, logout, and status checking
 // of authentication credentials for various AI providers, with OAuth support
-// for providers like Anthropic.
+// for providers like Anthropic and OpenAI.
 var authCmd = &cobra.Command{
 	Use:   "auth",
 	Short: "Manage authentication credentials for AI providers",
@@ -25,9 +29,11 @@ using OAuth flows. Stored credentials take precedence over environment variables
 
 Available providers:
   - anthropic: Anthropic Claude API (OAuth)
+  - openai:    OpenAI API (OAuth and API key)
 
 Examples:
   kit auth login anthropic
+  kit auth login openai
   kit auth logout anthropic
   kit auth status`,
 }
@@ -46,9 +52,11 @@ environment variables when making API calls.
 
 Available providers:
   - anthropic: Anthropic Claude API (OAuth)
+  - openai:    OpenAI ChatGPT Plus/Pro (Codex OAuth)
 
 Example:
-  kit auth login anthropic`,
+  kit auth login anthropic
+  kit auth login openai`,
 	Args: cobra.ExactArgs(1),
 	RunE: runAuthLogin,
 }
@@ -61,14 +69,16 @@ var authLogoutCmd = &cobra.Command{
 	Short: "Remove stored authentication credentials for a provider",
 	Long: `Remove stored authentication credentials for an AI provider.
 
-This will delete the stored API key for the specified provider. You will need
-to use environment variables or command-line flags for authentication after logout.
+This will delete the stored API key or OAuth credentials for the specified provider. 
+You will need to use environment variables or command-line flags for authentication after logout.
 
 Available providers:
   - anthropic: Anthropic Claude API
+  - openai:    OpenAI API
 
 Example:
-  kit auth logout anthropic`,
+  kit auth logout anthropic
+  kit auth logout openai`,
 	Args: cobra.ExactArgs(1),
 	RunE: runAuthLogout,
 }
@@ -101,8 +111,10 @@ func runAuthLogin(cmd *cobra.Command, args []string) error {
 	switch provider {
 	case "anthropic":
 		return loginAnthropic()
+	case "openai":
+		return loginOpenAI()
 	default:
-		return fmt.Errorf("unsupported provider: %s. Available providers: anthropic", provider)
+		return fmt.Errorf("unsupported provider: %s. Available providers: anthropic, openai", provider)
 	}
 }
 
@@ -112,8 +124,10 @@ func runAuthLogout(cmd *cobra.Command, args []string) error {
 	switch provider {
 	case "anthropic":
 		return logoutAnthropic()
+	case "openai":
+		return logoutOpenAI()
 	default:
-		return fmt.Errorf("unsupported provider: %s. Available providers: anthropic", provider)
+		return fmt.Errorf("unsupported provider: %s. Available providers: anthropic, openai", provider)
 	}
 }
 
@@ -157,8 +171,44 @@ func runAuthStatus(cmd *cobra.Command, args []string) error {
 		}
 	}
 
+	// Check OpenAI credentials
+	fmt.Print("\nOpenAI: ")
+	if hasOpenAICreds, err := cm.HasOpenAICredentials(); err != nil {
+		fmt.Printf("Error checking credentials: %v\n", err)
+	} else if hasOpenAICreds {
+		if creds, err := cm.GetOpenAICredentials(); err != nil {
+			fmt.Printf("Error reading credentials: %v\n", err)
+		} else {
+			authType := "API Key"
+			status := "✓ Authenticated"
+
+			if creds.Type == "oauth" {
+				authType = "OAuth (ChatGPT/Codex)"
+				if creds.IsExpired() {
+					status = "⚠️  Token expired (will refresh automatically)"
+				} else if creds.NeedsRefresh() {
+					status = "⚠️  Token expires soon (will refresh automatically)"
+				}
+			}
+
+			accountInfo := ""
+			if creds.Type == "oauth" && creds.AccountID != "" {
+				accountInfo = fmt.Sprintf(" [%s]", creds.AccountID)
+			}
+
+			fmt.Printf("%s (%s%s, stored %s)\n", status, authType, accountInfo, creds.CreatedAt.Format("2006-01-02 15:04:05"))
+		}
+	} else {
+		fmt.Println("✗ Not authenticated")
+		// Check if environment variable is set
+		if os.Getenv("OPENAI_API_KEY") != "" {
+			fmt.Println("  (OPENAI_API_KEY environment variable is set)")
+		}
+	}
+
 	fmt.Println("\nTo authenticate with a provider:")
 	fmt.Println("  kit auth login anthropic")
+	fmt.Println("  kit auth login openai")
 
 	return nil
 }
@@ -282,3 +332,246 @@ func logoutAnthropic() error {
 
 	return nil
 }
+
+func loginOpenAI() error {
+	cm, err := kit.NewCredentialManager()
+	if err != nil {
+		return fmt.Errorf("failed to initialize credential manager: %w", err)
+	}
+
+	// Check if already authenticated
+	if hasAuth, err := cm.HasOpenAICredentials(); err == nil && hasAuth {
+		var reauth bool
+		err := huh.NewConfirm().
+			Title("You are already authenticated with OpenAI (ChatGPT/Codex)").
+			Description("Do you want to re-authenticate?").
+			Affirmative("Yes").
+			Negative("No").
+			Value(&reauth).
+			Run()
+		if err != nil || !reauth {
+			fmt.Println("Authentication cancelled.")
+			return nil
+		}
+	}
+
+	// Create OAuth client
+	client := auth.NewOpenAIOAuthClient()
+
+	// Generate authorization URL
+	fmt.Println("🔐 Starting OAuth authentication with OpenAI (ChatGPT/Codex)...")
+	fmt.Println("This will open your browser to authenticate with your ChatGPT account.")
+	fmt.Println()
+
+	authData, err := client.GetAuthorizationURL()
+	if err != nil {
+		return fmt.Errorf("failed to generate authorization URL: %w", err)
+	}
+
+	// Start local callback server
+	callbackServer, err := startOpenAICallbackServer(authData.State)
+	if err != nil {
+		fmt.Printf("⚠️  Could not start local callback server: %v\n", err)
+		fmt.Println("Falling back to manual code entry.")
+	}
+	if callbackServer != nil {
+		defer callbackServer.Close()
+	}
+
+	// Display URL and try to open browser
+	fmt.Println("📱 Opening your browser for authentication...")
+	fmt.Println("If the browser doesn't open automatically, please visit this URL:")
+	fmt.Printf("\n%s\n\n", authData.URL)
+
+	// Try to open browser
+	auth.TryOpenBrowser(authData.URL)
+
+	// Wait for callback or manual input
+	var code string
+	if callbackServer != nil {
+		fmt.Println("Waiting for browser authentication...")
+		select {
+		case callbackCode := <-callbackServer.CodeChan:
+			if callbackCode != "" {
+				code = callbackCode
+				fmt.Println("✓ Received authorization code from browser callback.")
+			}
+		case <-time.After(2 * time.Minute):
+			fmt.Println("\n⏱️  Timeout waiting for browser callback.")
+			callbackServer.Close()
+		}
+	}
+
+	// If no code from callback, prompt for manual entry
+	if code == "" {
+		fmt.Println("\nAfter authorizing, paste the callback URL or authorization code below.")
+		fmt.Println("(The callback URL will look like: http://localhost:1455/auth/callback?code=...&state=...)")
+		fmt.Println()
+
+		var input string
+		err = huh.NewInput().
+			Title("Callback URL or Code").
+			Description("Paste the full callback URL or just the authorization code").
+			Value(&input).
+			Run()
+		if err != nil {
+			return fmt.Errorf("failed to read input: %w", err)
+		}
+		input = strings.TrimSpace(input)
+
+		if input == "" {
+			return fmt.Errorf("authorization code cannot be empty")
+		}
+
+		// Parse the input (could be full URL or just code)
+		parsedCode, parsedState := auth.ParseOpenAIAuthorizationInput(input)
+		if parsedCode == "" {
+			return fmt.Errorf("could not extract authorization code from input")
+		}
+
+		// Validate state if provided
+		if parsedState != "" && parsedState != authData.State {
+			return fmt.Errorf("state mismatch - possible security issue")
+		}
+		code = parsedCode
+	}
+
+	// Exchange code for tokens
+	fmt.Println("\n🔄 Exchanging authorization code for access token...")
+	creds, err := client.ExchangeCode(code, authData.Verifier)
+	if err != nil {
+		return fmt.Errorf("failed to exchange authorization code: %w", err)
+	}
+
+	// Store the credentials
+	if err := cm.SetOpenAIOAuthCredentials(creds); err != nil {
+		return fmt.Errorf("failed to store credentials: %w", err)
+	}
+
+	fmt.Println("✅ Successfully authenticated with OpenAI (ChatGPT/Codex)!")
+	fmt.Printf("📁 Credentials stored in: %s\n", cm.GetCredentialsPath())
+	fmt.Printf("👤 Account ID: %s\n", creds.AccountID)
+	fmt.Println("\n🎉 Your OAuth credentials will now be used for OpenAI API calls.")
+	fmt.Println("💡 You can check your authentication status with: kit auth status")
+
+	return nil
+}
+
+// callbackServer holds the HTTP server and channel for receiving the OAuth callback
+type callbackServer struct {
+	Server   *http.Server
+	CodeChan chan string
+	State    string
+}
+
+// Close shuts down the callback server
+func (cs *callbackServer) Close() {
+	if cs.Server != nil {
+		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+		_ = cs.Server.Shutdown(ctx)
+	}
+}
+
+// startOpenAICallbackServer starts a local HTTP server to receive the OAuth callback
+func startOpenAICallbackServer(expectedState string) (*callbackServer, error) {
+	codeChan := make(chan string, 1)
+
+	mux := http.NewServeMux()
+	server := &http.Server{
+		Addr:    "127.0.0.1:1455",
+		Handler: mux,
+	}
+
+	mux.HandleFunc("/auth/callback", func(w http.ResponseWriter, r *http.Request) {
+		// Check state
+		state := r.URL.Query().Get("state")
+		if state != expectedState {
+			http.Error(w, "State mismatch", http.StatusBadRequest)
+			return
+		}
+
+		code := r.URL.Query().Get("code")
+		if code == "" {
+			http.Error(w, "Missing authorization code", http.StatusBadRequest)
+			return
+		}
+
+		// Send code to channel
+		select {
+		case codeChan <- code:
+		default:
+		}
+
+		// Return success page
+		w.Header().Set("Content-Type", "text/html")
+		w.WriteHeader(http.StatusOK)
+		_, _ = fmt.Fprintf(w, `<!DOCTYPE html>
+<html>
+<head><title>Authentication Successful</title></head>
+<body style="font-family: sans-serif; text-align: center; padding: 50px;">
+<h1>✓ Authentication Successful</h1>
+<p>You can close this window and return to the terminal.</p>
+</body>
+</html>`)
+	})
+
+	// Try to start server
+	listener, err := net.Listen("tcp", "127.0.0.1:1455")
+	if err != nil {
+		return nil, fmt.Errorf("port 1455 not available: %w", err)
+	}
+	_ = listener.Close()
+
+	go func() {
+		_ = server.ListenAndServe()
+	}()
+
+	return &callbackServer{
+		Server:   server,
+		CodeChan: codeChan,
+		State:    expectedState,
+	}, nil
+}
+
+func logoutOpenAI() error {
+	cm, err := kit.NewCredentialManager()
+	if err != nil {
+		return fmt.Errorf("failed to initialize credential manager: %w", err)
+	}
+
+	// Check if authenticated
+	hasAuth, err := cm.HasOpenAICredentials()
+	if err != nil {
+		return fmt.Errorf("failed to check authentication status: %w", err)
+	}
+
+	if !hasAuth {
+		fmt.Println("You are not currently authenticated with OpenAI.")
+		return nil
+	}
+
+	// Confirm logout
+	var confirm bool
+	err = huh.NewConfirm().
+		Title("Remove OpenAI credentials").
+		Description("Are you sure you want to remove your stored credentials?").
+		Affirmative("Yes").
+		Negative("No").
+		Value(&confirm).
+		Run()
+	if err != nil || !confirm {
+		fmt.Println("Logout cancelled.")
+		return nil
+	}
+
+	// Remove credentials
+	if err := cm.RemoveOpenAICredentials(); err != nil {
+		return fmt.Errorf("failed to remove credentials: %w", err)
+	}
+
+	fmt.Println("✓ Successfully logged out from OpenAI!")
+	fmt.Println("You will need to use environment variables or command-line flags for authentication.")
+
+	return nil
+}

cmd/models.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"sort"
 
+	"github.com/mark3labs/kit/internal/models"
 	kit "github.com/mark3labs/kit/pkg/kit"
 	"github.com/spf13/cobra"
 )
@@ -47,11 +48,14 @@ func runModels(_ *cobra.Command, args []string) error {
 }
 
 func printAllProviders(showAll bool) error {
+	// Reload the registry to pick up any custom models from config
+	models.ReloadGlobalRegistry()
+
 	var providerIDs []string
 	if showAll {
 		providerIDs = kit.GetSupportedProviders()
 	} else {
-		providerIDs = kit.GetFantasyProviders()
+		providerIDs = kit.GetLLMProviders()
 	}
 	sort.Strings(providerIDs)
 
@@ -98,6 +102,9 @@ func printAllProviders(showAll bool) error {
 }
 
 func printProvider(provider string) error {
+	// Reload the registry to pick up any custom models from config
+	models.ReloadGlobalRegistry()
+
 	m, err := kit.GetModelsForProvider(provider)
 	if err != nil {
 		return fmt.Errorf("unknown provider %q. Run 'kit models' to see all providers", provider)

cmd/skill.go

@@ -8,19 +8,21 @@ import (
 	"github.com/spf13/cobra"
 )
 
-// skillCmd installs the kit-extensions skill via the skills.sh CLI (npx skills).
-// This teaches AI agents how to create Kit extensions with full knowledge of
-// the extension API, lifecycle events, widgets, tools, commands, and Yaegi constraints.
+// skillCmd installs Kit skills via the skills.sh CLI (npx skills).
 var skillCmd = &cobra.Command{
 	Use:   "skill",
-	Short: "Install the Kit extensions skill via skills.sh",
-	Long: `Install the kit-extensions skill that teaches AI agents how to create
-Kit extensions. Uses the skills.sh CLI (npx skills) to install the skill
-from the Kit repository.
+	Short: "Install Kit skills via skills.sh",
+	Long: `Install Kit skills that teach AI agents how to build with Kit.
+Uses the skills.sh CLI (npx skills) to install all skills from the Kit repository.
 
-The skill provides comprehensive documentation of Kit's extension API including
-lifecycle events, custom tools, slash commands, widgets, editor interceptors,
-tool renderers, and critical Yaegi interpreter constraints.
+Two skills are provided:
+
+  1. Extensions — creating Kit extensions with full knowledge of the extension
+     API, lifecycle events, widgets, tools, commands, editor interceptors,
+     tool renderers, and Yaegi interpreter constraints.
+
+  2. SDK — building AI-powered applications with the Kit Go SDK, including
+     providers, agents, tools, and MCP integration.
 
 Example:
   kit skill`,
@@ -41,8 +43,6 @@ func runSkill(_ *cobra.Command, _ []string) error {
 		"skills",
 		"add",
 		"mark3labs/kit",
-		"--skill",
-		"kit-extensions",
 	}
 
 	cmd := exec.Command(npx, args...)

examples/extensions/all_extensions_load_test.go

@@ -0,0 +1,27 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/mark3labs/kit/pkg/extensions/test"
+)
+
+// TestAllExtensions_Load is a smoke test that verifies every single-file
+// example extension in this directory can be loaded by the Yaegi interpreter
+// without errors. This catches syntax errors, missing symbols, bad imports,
+// and Init signature mismatches.
+func TestAllExtensions_Load(t *testing.T) {
+	files := extensionFiles(t)
+
+	for _, file := range files {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			ext := harness.LoadFile(file)
+			if ext == nil {
+				t.Fatalf("%s: extension should not be nil after loading", file)
+			}
+		})
+	}
+
+	t.Logf("successfully loaded %d extensions", len(files))
+}

examples/extensions/all_extensions_sanity_test.go

@@ -0,0 +1,253 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/mark3labs/kit/internal/extensions"
+	"github.com/mark3labs/kit/pkg/extensions/test"
+)
+
+// extensionFiles returns all single-file extensions in the current directory.
+// It skips test files, the test template, and files without an Init function.
+func extensionFiles(t *testing.T) []string {
+	t.Helper()
+
+	skip := map[string]bool{
+		"extension_test_template.go": true,
+	}
+
+	entries, err := os.ReadDir(".")
+	if err != nil {
+		t.Fatalf("failed to read directory: %v", err)
+	}
+
+	var files []string
+	for _, entry := range entries {
+		name := entry.Name()
+		if entry.IsDir() || filepath.Ext(name) != ".go" {
+			continue
+		}
+		if strings.HasSuffix(name, "_test.go") || skip[name] {
+			continue
+		}
+		src, err := os.ReadFile(name)
+		if err != nil {
+			t.Fatalf("failed to read %s: %v", name, err)
+		}
+		if !strings.Contains(string(src), "func Init(") {
+			continue
+		}
+		files = append(files, name)
+	}
+
+	if len(files) == 0 {
+		t.Fatal("no extensions found — check the directory")
+	}
+	return files
+}
+
+// TestAllExtensions_Lifecycle verifies that every extension survives a full
+// SessionStart → SessionShutdown round-trip without errors.
+func TestAllExtensions_Lifecycle(t *testing.T) {
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			_, err := harness.Emit(extensions.SessionStartEvent{
+				SessionID: "smoke-test-session",
+			})
+			if err != nil {
+				t.Fatalf("SessionStart error: %v", err)
+			}
+
+			_, err = harness.Emit(extensions.SessionShutdownEvent{})
+			if err != nil {
+				t.Fatalf("SessionShutdown error: %v", err)
+			}
+		})
+	}
+}
+
+// TestAllExtensions_CommandSanity checks that every registered command has
+// a non-empty name, a non-empty description, no spaces in the name, no
+// leading slash, a non-nil Execute function, and no duplicate names.
+func TestAllExtensions_CommandSanity(t *testing.T) {
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			cmds := harness.RegisteredCommands()
+			seen := make(map[string]bool)
+			for _, cmd := range cmds {
+				if cmd.Name == "" {
+					t.Error("command has empty name")
+				}
+				if strings.Contains(cmd.Name, " ") {
+					t.Errorf("command %q contains spaces", cmd.Name)
+				}
+				if strings.HasPrefix(cmd.Name, "/") {
+					t.Errorf("command %q has leading slash (framework adds it)", cmd.Name)
+				}
+				if cmd.Description == "" {
+					t.Errorf("command %q has empty description", cmd.Name)
+				}
+				if cmd.Execute == nil {
+					t.Errorf("command %q has nil Execute function", cmd.Name)
+				}
+				if seen[cmd.Name] {
+					t.Errorf("duplicate command name %q", cmd.Name)
+				}
+				seen[cmd.Name] = true
+			}
+		})
+	}
+}
+
+// TestAllExtensions_ToolSanity checks that every registered tool has a
+// non-empty name, a non-empty description, at least one executor, valid
+// JSON in its Parameters field, and no duplicate names.
+func TestAllExtensions_ToolSanity(t *testing.T) {
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			tools := harness.RegisteredTools()
+			seen := make(map[string]bool)
+			for _, tool := range tools {
+				if tool.Name == "" {
+					t.Error("tool has empty name")
+				}
+				if tool.Description == "" {
+					t.Errorf("tool %q has empty description", tool.Name)
+				}
+				if tool.Execute == nil && tool.ExecuteWithContext == nil {
+					t.Errorf("tool %q has no executor (both Execute and ExecuteWithContext are nil)", tool.Name)
+				}
+				if tool.Parameters != "" && !json.Valid([]byte(tool.Parameters)) {
+					t.Errorf("tool %q has invalid JSON in Parameters: %s", tool.Name, tool.Parameters)
+				}
+				if seen[tool.Name] {
+					t.Errorf("duplicate tool name %q", tool.Name)
+				}
+				seen[tool.Name] = true
+			}
+		})
+	}
+}
+
+// TestAllExtensions_ZeroValueEvents fires every event type (as zero-value
+// structs) at each extension and verifies no errors are returned. Extensions
+// should be resilient to events they don't handle and to events with empty
+// fields.
+func TestAllExtensions_ZeroValueEvents(t *testing.T) {
+	// Build the set of zero-value events for every event type.
+	zeroEvents := []extensions.Event{
+		extensions.ToolCallEvent{},
+		extensions.ToolExecutionStartEvent{},
+		extensions.ToolExecutionEndEvent{},
+		extensions.ToolOutputEvent{},
+		extensions.ToolResultEvent{},
+		extensions.InputEvent{},
+		extensions.BeforeAgentStartEvent{},
+		extensions.AgentStartEvent{},
+		extensions.AgentEndEvent{},
+		extensions.MessageStartEvent{},
+		extensions.MessageUpdateEvent{},
+		extensions.MessageEndEvent{},
+		extensions.SessionStartEvent{},
+		extensions.SessionShutdownEvent{},
+		extensions.ModelChangeEvent{},
+		extensions.ContextPrepareEvent{},
+		extensions.BeforeForkEvent{},
+		extensions.BeforeSessionSwitchEvent{},
+		extensions.BeforeCompactEvent{},
+		extensions.SubagentStartEvent{},
+		extensions.SubagentChunkEvent{},
+		extensions.SubagentEndEvent{},
+	}
+
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			for _, ev := range zeroEvents {
+				_, err := harness.Emit(ev)
+				if err != nil {
+					t.Errorf("event %T returned error: %v", ev, err)
+				}
+			}
+		})
+	}
+}
+
+// TestAllExtensions_WidgetSanity emits SessionStart and then checks that
+// any widgets set during initialization have non-empty IDs and valid
+// placements.
+func TestAllExtensions_WidgetSanity(t *testing.T) {
+	validPlacements := map[extensions.WidgetPlacement]bool{
+		"above": true,
+		"below": true,
+	}
+
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			// Trigger SessionStart so extensions that set widgets on init do so.
+			_, _ = harness.Emit(extensions.SessionStartEvent{
+				SessionID: "widget-sanity-test",
+			})
+
+			// Widgets is an exported field on MockContext; reads are safe
+			// here because Emit returned synchronously.
+			for id, w := range harness.Context().Widgets {
+				if w.ID == "" {
+					t.Errorf("widget stored with key %q has empty ID", id)
+				}
+				if w.ID != id {
+					t.Errorf("widget key %q doesn't match widget ID %q", id, w.ID)
+				}
+				if !validPlacements[w.Placement] {
+					t.Errorf("widget %q has invalid placement %q (want \"above\" or \"below\")", id, w.Placement)
+				}
+			}
+		})
+	}
+}
+
+// TestAllExtensions_IdempotentLifecycle verifies that receiving SessionStart
+// twice and SessionShutdown twice doesn't cause errors — extensions should
+// be defensive about repeated lifecycle events.
+func TestAllExtensions_IdempotentLifecycle(t *testing.T) {
+	for _, file := range extensionFiles(t) {
+		t.Run(file, func(t *testing.T) {
+			harness := test.New(t)
+			harness.LoadFile(file)
+
+			for i := range 2 {
+				_, err := harness.Emit(extensions.SessionStartEvent{
+					SessionID: "idempotent-test",
+				})
+				if err != nil {
+					t.Fatalf("SessionStart #%d error: %v", i+1, err)
+				}
+			}
+
+			for i := range 2 {
+				_, err := harness.Emit(extensions.SessionShutdownEvent{})
+				if err != nil {
+					t.Fatalf("SessionShutdown #%d error: %v", i+1, err)
+				}
+			}
+		})
+	}
+}

examples/extensions/bridge_demo.go

@@ -0,0 +1,170 @@
+//go:build ignore
+
+// bridge_demo.go - Demonstrates the new bridged SDK APIs for extensions.
+// This extension showcases tree navigation, skill loading, template parsing,
+// and model resolution capabilities.
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"kit/ext"
+)
+
+var (
+	discoveredSkills []ext.Skill
+	currentBranch    []ext.TreeNode
+)
+
+func Init(api ext.API) {
+	// Register /tree-info command to demonstrate tree navigation
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "tree-info",
+		Description: "Show current conversation tree information",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			branch := ctx.GetCurrentBranch()
+			info := fmt.Sprintf("Current branch has %d nodes:\n", len(branch))
+			for i, node := range branch {
+				info += fmt.Sprintf("  [%d] %s (%s): %s...\n", i, node.Type, node.ID[:8], truncate(node.Content, 40))
+			}
+			ctx.PrintInfo(info)
+			return "", nil
+		},
+	})
+
+	// Register /discover-skills command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "discover-skills",
+		Description: "Discover and list available skills",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			result := ctx.DiscoverSkills()
+			if result.Error != "" {
+				return "", fmt.Errorf("discovery failed: %s", result.Error)
+			}
+			discoveredSkills = result.Skills
+
+			info := fmt.Sprintf("Discovered %d skills:\n", len(result.Skills))
+			for _, s := range result.Skills {
+				info += fmt.Sprintf("  - %s: %s\n", s.Name, s.Description)
+			}
+			ctx.PrintInfo(info)
+			return "", nil
+		},
+	})
+
+	// Register /parse-template command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "parse-template",
+		Description: "Parse a template and show extracted variables",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			if args == "" {
+				args = "Hello {{name}}, welcome to {{place}}!"
+			}
+			tpl := ctx.ParseTemplate("demo", args)
+			info := fmt.Sprintf("Template: %s\nVariables: %v", tpl.Content, tpl.Variables)
+			ctx.PrintInfo(info)
+			return "", nil
+		},
+	})
+
+	// Register /render-template command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "render-template",
+		Description: "Render a template with variables (usage: /render-template name=John place=Kit)",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			tpl := ctx.ParseTemplate("demo", "Hello {{name}}, welcome to {{place}}!")
+			vars := ctx.ParseArguments(args, ext.ArgumentPattern{
+				Flags: map[string]string{"name": "name", "place": "place"},
+			})
+			rendered := ctx.RenderTemplate(tpl, vars.Vars)
+			ctx.PrintInfo("Rendered: " + rendered)
+			return "", nil
+		},
+	})
+
+	// Register /check-model command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "check-model",
+		Description: "Check model capabilities and availability",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			model := args
+			if model == "" {
+				model = ctx.Model
+			}
+
+			available := ctx.CheckModelAvailable(model)
+			caps, err := ctx.GetModelCapabilities(model)
+
+			info := fmt.Sprintf("Model: %s\n", model)
+			info += fmt.Sprintf("Available: %v\n", available)
+			if err == "" {
+				info += fmt.Sprintf("Provider: %s\n", caps.Provider)
+				info += fmt.Sprintf("Context Limit: %d\n", caps.ContextLimit)
+				info += fmt.Sprintf("Reasoning: %v\n", caps.Reasoning)
+			} else {
+				info += fmt.Sprintf("Error: %s\n", err)
+			}
+			ctx.PrintInfo(info)
+			return "", nil
+		},
+	})
+
+	// Register /resolve-chain command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "resolve-chain",
+		Description: "Resolve a model chain (usage: /resolve-chain claude-opus,gpt-4o,claude-sonnet)",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			if args == "" {
+				args = "anthropic/claude-opus-4,anthropic/claude-sonnet-4,openai/gpt-4o"
+			}
+			prefs := ctx.SimpleParseArguments(args, 1)
+			chain := []string{}
+			if len(prefs) > 1 {
+				// Split the first arg by comma
+				for _, p := range strings.Split(prefs[1], ",") {
+					p = strings.TrimSpace(p)
+					if p != "" {
+						chain = append(chain, p)
+					}
+				}
+			}
+
+			result := ctx.ResolveModelChain(chain)
+			info, _ := json.MarshalIndent(result, "", "  ")
+			ctx.PrintInfo("Resolution Result:\n" + string(info))
+			return "", nil
+		},
+	})
+
+	// Register /test-conditional command
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "test-conditional",
+		Description: "Test model conditional rendering",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			content := `<if-model is="claude-*">This is for Claude models<else>This is for other models</if-model>`
+			rendered := ctx.RenderWithModelConditionals(content)
+			ctx.PrintInfo("Input: " + content)
+			ctx.PrintInfo("Output: " + rendered)
+			ctx.PrintInfo(fmt.Sprintf("Current model matches 'claude-*': %v", ctx.EvaluateModelConditional("claude-*")))
+			return "", nil
+		},
+	})
+
+	// OnSessionStart: discover skills automatically
+	api.OnSessionStart(func(e ext.SessionStartEvent, ctx ext.Context) {
+		result := ctx.DiscoverSkills()
+		if result.Error == "" && len(result.Skills) > 0 {
+			discoveredSkills = result.Skills
+			ctx.SetStatus("bridge-demo", fmt.Sprintf("%d skills", len(result.Skills)), 50)
+		}
+	})
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max-3] + "..."
+}

examples/extensions/conversation-manager.go

@@ -0,0 +1,406 @@
+//go:build ignore
+
+// conversation-manager.go - Advanced conversation tree navigation and management.
+// This extension demonstrates:
+// - Tree navigation (GetTreeNode, GetCurrentBranch, NavigateTo)
+// - Branch summarization and collapsing
+// - Interactive tree exploration
+//
+// Commands:
+//   /tree              - Show conversation tree structure
+//   /branch            - Show current branch path
+//   /goto <entry-id>   - Navigate to a specific entry
+//   /summarize <n>     - Summarize last N messages
+//   /fresh-context     - Collapse branch and start fresh
+//   /loop <n> <prompt> - Execute prompt N times with fresh context each iteration
+
+package main
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+
+	"kit/ext"
+)
+
+var (
+	loopActive    bool
+	loopCount     int
+	loopCurrent   int
+	loopPrompt    string
+	loopStartNode string
+)
+
+func Init(api ext.API) {
+	// /tree - Show tree structure
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "tree",
+		Description: "Show conversation tree structure",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			showTree(ctx)
+			return "", nil
+		},
+	})
+
+	// /branch - Show current branch
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "branch",
+		Description: "Show current conversation branch",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			showBranch(ctx)
+			return "", nil
+		},
+	})
+
+	// /goto - Navigate to entry
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "goto",
+		Description: "Navigate to a specific entry ID (usage: /goto <entry-id>)",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			if args == "" {
+				ctx.PrintError("Usage: /goto <entry-id>")
+				return "", nil
+			}
+			result := ctx.NavigateTo(args)
+			if !result.Success {
+				ctx.PrintError(fmt.Sprintf("Navigation failed: %s", result.Error))
+				return "", nil
+			}
+			ctx.PrintInfo(fmt.Sprintf("Navigated to entry: %s", args))
+
+			// Show the node we navigated to
+			node := ctx.GetTreeNode(args)
+			if node != nil {
+				ctx.PrintInfo(fmt.Sprintf("Entry type: %s, Role: %s", node.Type, node.Role))
+			}
+			return "", nil
+		},
+	})
+
+	// /summarize - Summarize recent messages
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "summarize",
+		Description: "Summarize last N messages (usage: /summarize [n=5])",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			n := 5
+			if args != "" {
+				if parsed, err := strconv.Atoi(args); err == nil && parsed > 0 {
+					n = parsed
+				}
+			}
+
+			branch := ctx.GetCurrentBranch()
+			if len(branch) < 2 {
+				ctx.PrintError("Not enough messages to summarize")
+				return "", nil
+			}
+
+			// Find range to summarize
+			startIdx := len(branch) - n - 1
+			if startIdx < 0 {
+				startIdx = 0
+			}
+			endIdx := len(branch) - 1
+
+			fromID := branch[startIdx].ID
+			toID := branch[endIdx].ID
+
+			ctx.PrintInfo(fmt.Sprintf("Summarizing messages %d to %d...", startIdx, endIdx))
+			summary := ctx.SummarizeBranch(fromID, toID)
+
+			if summary == "" {
+				ctx.PrintError("Failed to generate summary")
+				return "", nil
+			}
+
+			ctx.PrintBlock(ext.PrintBlockOpts{
+				Text:        summary,
+				BorderColor: "#89b4fa",
+				Subtitle:    "conversation-manager · Summary",
+			})
+			return "", nil
+		},
+	})
+
+	// /fresh-context - Collapse and restart
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "fresh-context",
+		Description: "Collapse conversation to summary and start fresh",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			branch := ctx.GetCurrentBranch()
+			if len(branch) < 3 {
+				ctx.PrintError("Not enough context to collapse")
+				return "", nil
+			}
+
+			// Keep first message (system), summarize rest
+			fromID := branch[1].ID
+			toID := branch[len(branch)-1].ID
+
+			ctx.PrintInfo("Generating summary for context collapse...")
+			summary := ctx.SummarizeBranch(fromID, toID)
+
+			if summary == "" {
+				ctx.PrintError("Failed to generate summary")
+				return "", nil
+			}
+
+			// Collapse the branch
+			result := ctx.CollapseBranch(fromID, toID, summary)
+			if !result.Success {
+				ctx.PrintError(fmt.Sprintf("Collapse failed: %s", result.Error))
+				return "", nil
+			}
+
+			ctx.PrintInfo("Context collapsed. Starting fresh with summary.")
+			ctx.PrintBlock(ext.PrintBlockOpts{
+				Text:        summary,
+				BorderColor: "#a6e3a1",
+				Subtitle:    "conversation-manager · Collapsed Context",
+			})
+
+			// Set a widget showing we're in fresh mode
+			ctx.SetWidget(ext.WidgetConfig{
+				ID:        "fresh-context",
+				Placement: ext.WidgetAbove,
+				Content:   ext.WidgetContent{Text: "🌱 Fresh Context Mode - Previous conversation collapsed"},
+				Style:     ext.WidgetStyle{BorderColor: "#a6e3a1"},
+			})
+
+			return "", nil
+		},
+	})
+
+	// /loop - Execute with fresh context each iteration
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "loop",
+		Description: "Execute prompt N times with fresh context (usage: /loop 5 analyze this code)",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			if loopActive {
+				ctx.PrintError("Loop already in progress. Wait for completion.")
+				return "", nil
+			}
+
+			// Parse arguments
+			parts := strings.SplitN(args, " ", 2)
+			if len(parts) < 2 {
+				ctx.PrintError("Usage: /loop <count> <prompt>")
+				return "", nil
+			}
+
+			count, err := strconv.Atoi(parts[0])
+			if err != nil || count <= 0 || count > 10 {
+				ctx.PrintError("Invalid count (must be 1-10)")
+				return "", nil
+			}
+
+			loopCount = count
+			loopCurrent = 0
+			loopPrompt = parts[1]
+			loopActive = true
+
+			// Store current branch position
+			branch := ctx.GetCurrentBranch()
+			if len(branch) > 0 {
+				loopStartNode = branch[len(branch)-1].ID
+			}
+
+			ctx.PrintInfo(fmt.Sprintf("Starting loop: %d iterations", loopCount))
+			ctx.SetWidget(ext.WidgetConfig{
+				ID:        "loop-progress",
+				Placement: ext.WidgetAbove,
+				Content:   ext.WidgetContent{Text: fmt.Sprintf("🔄 Loop: 0/%d - %s", loopCount, loopPrompt)},
+				Style:     ext.WidgetStyle{BorderColor: "#fab387"},
+			})
+
+			// Start first iteration
+			executeLoopIteration(ctx)
+			return "", nil
+		},
+	})
+
+	// OnAgentEnd handles loop continuation
+	api.OnAgentEnd(func(e ext.AgentEndEvent, ctx ext.Context) {
+		if !loopActive {
+			return
+		}
+
+		loopCurrent++
+
+		if loopCurrent >= loopCount {
+			// Loop complete
+			loopActive = false
+			ctx.RemoveWidget("loop-progress")
+			ctx.PrintInfo(fmt.Sprintf("✅ Loop complete: %d/%d iterations", loopCurrent, loopCount))
+
+			// Show final summary
+			branch := ctx.GetCurrentBranch()
+			if len(branch) > 0 && loopStartNode != "" {
+				summary := ctx.SummarizeBranch(loopStartNode, branch[len(branch)-1].ID)
+				if summary != "" {
+					ctx.PrintBlock(ext.PrintBlockOpts{
+						Text:        summary,
+						BorderColor: "#a6e3a1",
+						Subtitle:    "conversation-manager · Loop Summary",
+					})
+				}
+			}
+			return
+		}
+
+		// Update progress
+		ctx.SetWidget(ext.WidgetConfig{
+			ID:        "loop-progress",
+			Placement: ext.WidgetAbove,
+			Content:   ext.WidgetContent{Text: fmt.Sprintf("🔄 Loop: %d/%d - %s", loopCurrent, loopCount, loopPrompt)},
+			Style:     ext.WidgetStyle{BorderColor: "#fab387"},
+		})
+
+		// Collapse previous iteration for fresh context
+		branch := ctx.GetCurrentBranch()
+		if len(branch) >= 2 {
+			// Find the user messages (look for the one before the last assistant message)
+			// We want to collapse from the user message that started this iteration
+			// to the last assistant response
+			var collapseStartIdx = -1
+			for i := len(branch) - 1; i >= 0; i-- {
+				if branch[i].Role == "assistant" {
+					// Found the last assistant message, now find the user message before it
+					for j := i - 1; j >= 0; j-- {
+						if branch[j].Role == "user" {
+							collapseStartIdx = j
+							break
+						}
+					}
+					break
+				}
+			}
+
+			if collapseStartIdx >= 0 {
+				fromID := branch[collapseStartIdx].ID
+				toID := branch[len(branch)-1].ID
+
+				ctx.PrintInfo(fmt.Sprintf("Collapsing iteration %d for fresh context...", loopCurrent))
+				summary := ctx.SummarizeBranch(fromID, toID)
+				if summary != "" {
+					result := ctx.CollapseBranch(fromID, toID, summary)
+					if result.Success {
+						ctx.PrintInfo("Context collapsed successfully")
+					} else {
+						ctx.PrintError(fmt.Sprintf("Collapse failed: %s", result.Error))
+					}
+				}
+			}
+		}
+
+		// Small delay to let UI update
+		time.Sleep(500 * time.Millisecond)
+
+		// Trigger next iteration
+		executeLoopIteration(ctx)
+	})
+}
+
+// showTree displays the conversation tree structure
+func showTree(ctx ext.Context) {
+	branch := ctx.GetCurrentBranch()
+	if len(branch) == 0 {
+		ctx.PrintInfo("Tree is empty")
+		return
+	}
+
+	var output strings.Builder
+	output.WriteString(fmt.Sprintf("Conversation Tree (%d nodes):\n\n", len(branch)))
+
+	for i, node := range branch {
+		prefix := "  "
+		if i == len(branch)-1 {
+			prefix = "▶ " // Current node
+		} else {
+			prefix = "  "
+		}
+
+		roleIcon := "💬"
+		switch node.Role {
+		case "user":
+			roleIcon = "👤"
+		case "assistant":
+			roleIcon = "🤖"
+		case "system":
+			roleIcon = "⚙️"
+		}
+
+		content := truncate(node.Content, 50)
+		if node.Type == "branch_summary" {
+			roleIcon = "📋"
+			content = "[Summary] " + truncate(node.Content, 40)
+		}
+
+		output.WriteString(fmt.Sprintf("%s%s %s: %s (%s...)\n", prefix, roleIcon, node.Role, node.ID[:8], content))
+
+		// Show children count if any
+		children := ctx.GetChildren(node.ID)
+		if len(children) > 0 {
+			output.WriteString(fmt.Sprintf("    └─ %d branch(es)\n", len(children)))
+		}
+	}
+
+	ctx.PrintBlock(ext.PrintBlockOpts{
+		Text:        output.String(),
+		BorderColor: "#89b4fa",
+		Subtitle:    "conversation-manager · Tree View",
+	})
+}
+
+// showBranch displays the current branch path
+func showBranch(ctx ext.Context) {
+	branch := ctx.GetCurrentBranch()
+	if len(branch) == 0 {
+		ctx.PrintInfo("No active branch")
+		return
+	}
+
+	var output strings.Builder
+	output.WriteString(fmt.Sprintf("Current Branch (%d nodes from root to leaf):\n\n", len(branch)))
+
+	for i, node := range branch {
+		marker := "  "
+		if i == len(branch)-1 {
+			marker = "▶ " // Current leaf
+		}
+
+		output.WriteString(fmt.Sprintf("%s[%d] %s (%s): %s\n",
+			marker, i, node.Type, node.ID[:8], truncate(node.Content, 40)))
+	}
+
+	// Show current node details
+	leaf := branch[len(branch)-1]
+	output.WriteString(fmt.Sprintf("\nCurrent Leaf:\n"))
+	output.WriteString(fmt.Sprintf("  ID: %s\n", leaf.ID))
+	output.WriteString(fmt.Sprintf("  Type: %s\n", leaf.Type))
+	output.WriteString(fmt.Sprintf("  Role: %s\n", leaf.Role))
+	output.WriteString(fmt.Sprintf("  Model: %s\n", leaf.Model))
+	output.WriteString(fmt.Sprintf("  Children: %d\n", len(leaf.Children)))
+
+	ctx.PrintBlock(ext.PrintBlockOpts{
+		Text:        output.String(),
+		BorderColor: "#cba6f7",
+		Subtitle:    "conversation-manager · Branch View",
+	})
+}
+
+// executeLoopIteration triggers the next loop iteration
+func executeLoopIteration(ctx ext.Context) {
+	iterationPrompt := fmt.Sprintf("[%d/%d] %s", loopCurrent+1, loopCount, loopPrompt)
+	ctx.SendMessage(iterationPrompt)
+}
+
+// truncate helper
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max-3] + "..."
+}

examples/extensions/kit-telegram/main.go

@@ -908,7 +908,7 @@ func summarizeToolAction(toolName string, inputJSON string) string {
 		return "searching " + getStr("pattern", "text")
 	case "ls":
 		return "listing " + getStr("path", "directory")
-	case "spawn_subagent":
+	case "subagent":
 		return "spawning subagent"
 	default:
 		return "using " + toolName

examples/extensions/lsp-diagnostics.go

@@ -2,9 +2,7 @@
 
 // lsp-diagnostics.go — LSP-powered diagnostics for Kit's edit tool.
 //
-// Starts language servers on demand and surfaces diagnostics after file edits,
-// following the same pattern used by Charm's crush editor:
-//
+// Starts language servers on demand and surfaces diagnostics after file edits:
 //  1. After an edit, notify the LSP server of the file change
 //  2. Wait for the server to publish fresh diagnostics
 //  3. Append diagnostic output to the edit tool's result
@@ -412,7 +410,7 @@ func (c *lspClient) changeFile(absPath, content string) {
 }
 
 // waitForDiagnostics polls until the server publishes new diagnostics or
-// the timeout elapses. Mirrors crush's WaitForDiagnostics pattern.
+// the timeout elapses.
 func (c *lspClient) waitForDiagnostics(timeout time.Duration) {
 	c.diagMu.Lock()
 	startVersion := c.diagVersion

examples/extensions/prompt-templates.go

@@ -0,0 +1,269 @@
+//go:build ignore
+
+// prompt-templates.go - Frontmatter-driven prompt templates with model switching.
+// This extension demonstrates the new bridged SDK APIs:
+// - Tree navigation for conversation management
+// - Template parsing with {{variable}} substitution
+// - Model resolution with fallback chains
+// - Skill injection
+//
+// Usage:
+//   1. Create ~/.config/kit/prompts/debug.md with frontmatter:
+//      ---
+//      description: Debug Python code
+//      model: claude-sonnet-4-20250514
+//      skill: python
+//      ---
+//      Help me debug this Python code: {{input}}
+//
+//   2. In Kit: /debug my_script.py
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"kit/ext"
+)
+
+// PromptTemplate represents a loaded template with frontmatter
+type PromptTemplate struct {
+	Name        string
+	Description string
+	Model       string
+	Skill       string
+	Content     string
+	Variables   []string
+	Path        string
+}
+
+var (
+	templates   = make(map[string]PromptTemplate)
+	templateDir string
+)
+
+func Init(api ext.API) {
+	// Determine template directory
+	home, _ := os.UserHomeDir()
+	templateDir = filepath.Join(home, ".config", "kit", "prompts")
+
+	// Ensure directory exists
+	os.MkdirAll(templateDir, 0755)
+
+	// Register commands
+	api.RegisterCommand(ext.CommandDef{
+		Name:        "reload-templates",
+		Description: "Reload prompt templates from disk",
+		Execute: func(args string, ctx ext.Context) (string, error) {
+			loadTemplates(ctx)
+			ctx.PrintInfo(fmt.Sprintf("Loaded %d templates from %s", len(templates), templateDir))
+			return "", nil
+		},
+	})
+
+	// Dynamic template commands are registered after loading
+	api.OnSessionStart(func(e ext.SessionStartEvent, ctx ext.Context) {
+		loadTemplates(ctx)
+		registerTemplateCommands(api, ctx)
+	})
+}
+
+// loadTemplates discovers and loads all template files
+func loadTemplates(ctx ext.Context) {
+	templates = make(map[string]PromptTemplate)
+
+	entries, err := os.ReadDir(templateDir)
+	if err != nil {
+		return
+	}
+
+	for _, entry := range entries {
+		if entry.IsDir() || !strings.HasSuffix(entry.Name(), ".md") {
+			continue
+		}
+
+		path := filepath.Join(templateDir, entry.Name())
+		tpl, err := loadTemplateFile(path)
+		if err != nil {
+			continue
+		}
+
+		name := strings.TrimSuffix(entry.Name(), ".md")
+		templates[name] = tpl
+	}
+}
+
+// loadTemplateFile parses a template with YAML frontmatter
+func loadTemplateFile(path string) (PromptTemplate, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return PromptTemplate{}, err
+	}
+
+	content := string(data)
+	tpl := PromptTemplate{Path: path}
+
+	// Parse frontmatter
+	if strings.HasPrefix(content, "---") {
+		parts := strings.SplitN(content[3:], "---", 2)
+		if len(parts) == 2 {
+			frontmatter := strings.TrimSpace(parts[0])
+			body := strings.TrimSpace(parts[1])
+
+			// Simple line-by-line frontmatter parsing
+			for _, line := range strings.Split(frontmatter, "\n") {
+				line = strings.TrimSpace(line)
+				if line == "" || strings.HasPrefix(line, "#") {
+					continue
+				}
+
+				key, value, found := strings.Cut(line, ":")
+				if found {
+					key = strings.TrimSpace(key)
+					value = strings.TrimSpace(value)
+					switch key {
+					case "description":
+						tpl.Description = value
+					case "model":
+						tpl.Model = value
+					case "skill":
+						tpl.Skill = value
+					}
+				}
+			}
+			tpl.Content = body
+		} else {
+			tpl.Content = content
+		}
+	} else {
+		tpl.Content = content
+	}
+
+	// Parse {{variables}} using simple string parsing
+	// (Can't use ctx.ParseTemplate here since we're in Init, not a handler)
+	var vars []string
+	for {
+		start := strings.Index(tpl.Content, "{{")
+		if start == -1 {
+			break
+		}
+		end := strings.Index(tpl.Content[start:], "}}")
+		if end == -1 {
+			break
+		}
+		varName := strings.TrimSpace(tpl.Content[start+2 : start+end])
+		vars = append(vars, varName)
+		tpl.Content = tpl.Content[:start] + "{{" + varName + "}}" + tpl.Content[start+end+2:]
+	}
+	tpl.Variables = vars
+
+	return tpl, nil
+}
+
+// registerTemplateCommands dynamically registers commands for each template
+func registerTemplateCommands(api ext.API, ctx ext.Context) {
+	for name, tpl := range templates {
+		// Skip if already registered (we'd need to track this)
+		tplCopy := tpl // Capture for closure
+		nameCopy := name
+
+		// Build description with metadata
+		desc := tplCopy.Description
+		if desc == "" {
+			desc = fmt.Sprintf("Run %s template", nameCopy)
+		}
+		if tplCopy.Model != "" {
+			desc += fmt.Sprintf(" [%s", tplCopy.Model)
+			if tplCopy.Skill != "" {
+				desc += fmt.Sprintf(" +%s", tplCopy.Skill)
+			}
+			desc += "]"
+		}
+
+		api.RegisterCommand(ext.CommandDef{
+			Name:        nameCopy,
+			Description: desc,
+			Execute: func(args string, ctx ext.Context) (string, error) {
+				return executeTemplate(ctx, tplCopy, args)
+			},
+		})
+	}
+}
+
+// executeTemplate runs a template with the given arguments
+func executeTemplate(ctx ext.Context, tpl PromptTemplate, args string) (string, error) {
+	// Store original model for restoration
+	originalModel := ctx.Model
+
+	// 1. Resolve and switch model if specified
+	if tpl.Model != "" {
+		// Parse model chain (comma-separated)
+		preferences := strings.Split(tpl.Model, ",")
+		for i := range preferences {
+			preferences[i] = strings.TrimSpace(preferences[i])
+		}
+
+		result := ctx.ResolveModelChain(preferences)
+		if result.Error != "" {
+			ctx.PrintError(fmt.Sprintf("Model resolution failed: %s", result.Error))
+			// Continue with current model
+		} else {
+			ctx.PrintInfo(fmt.Sprintf("Switching to model: %s", result.Model))
+			if err := ctx.SetModel(result.Model); err != nil {
+				ctx.PrintError(fmt.Sprintf("Failed to switch model: %s", err.Error()))
+			}
+		}
+	}
+
+	// 2. Inject skill if specified
+	if tpl.Skill != "" {
+		err := ctx.InjectSkillAsContext(tpl.Skill)
+		if err != "" {
+			ctx.PrintError(fmt.Sprintf("Skill injection failed: %s", err))
+		} else {
+			ctx.PrintInfo(fmt.Sprintf("Injected skill: %s", tpl.Skill))
+		}
+	}
+
+	// 3. Parse and render template
+	parsed := ctx.ParseTemplate(tpl.Name, tpl.Content)
+
+	// Build variable map
+	vars := make(map[string]string)
+
+	// Simple argument parsing: first arg is $1 (input), rest is $@
+	if len(parsed.Variables) > 0 {
+		argsList := ctx.SimpleParseArguments(args, len(parsed.Variables))
+		for i, varName := range parsed.Variables {
+			if i < len(parsed.Variables) && i+1 < len(argsList) {
+				vars[varName] = argsList[i+1]
+			}
+		}
+		// If single variable, use full args
+		if len(parsed.Variables) == 1 && vars[parsed.Variables[0]] == "" {
+			vars[parsed.Variables[0]] = args
+		}
+	}
+
+	// Render with model conditionals
+	content := ctx.RenderWithModelConditionals(tpl.Content)
+	rendered := ctx.RenderTemplate(ext.PromptTemplate{Name: tpl.Name, Content: content, Variables: parsed.Variables}, vars)
+
+	// 4. Send the rendered prompt
+	ctx.SendMessage(rendered)
+
+	// 5. Schedule model restoration after turn completes
+	// We use a goroutine to wait and restore
+	if tpl.Model != "" && originalModel != "" {
+		go func() {
+			// Note: In a real implementation, we'd use OnAgentEnd event
+			// For now, the user can manually switch back
+			ctx.SetStatus("template-mode", fmt.Sprintf("Template: %s (model will restore)", tpl.Name), 20)
+		}()
+	}
+
+	return fmt.Sprintf("Executing template: %s", tpl.Name), nil
+}

examples/extensions/subagent-monitor_test.go

@@ -0,0 +1,159 @@
+package main
+
+import (
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/mark3labs/kit/internal/extensions"
+	"github.com/mark3labs/kit/pkg/extensions/test"
+)
+
+// TestSubagentMonitor_SessionStart verifies OnSessionStart initializes state
+// without panicking and properly guards nil ctx calls.
+func TestSubagentMonitor_SessionStart(t *testing.T) {
+	harness := test.New(t)
+	harness.LoadFile("../../.kit/extensions/subagent-monitor.go")
+
+	// Emit SessionStart - should not panic even with nil ctx functions
+	_, err := harness.Emit(extensions.SessionStartEvent{SessionID: "test-session"})
+	if err != nil {
+		t.Fatalf("SessionStart should not error: %v", err)
+	}
+}
+
+// TestSubagentMonitor_SubagentLifecycle verifies the full subagent lifecycle
+// creates entries and emits widget updates.
+func TestSubagentMonitor_SubagentLifecycle(t *testing.T) {
+	harness := test.New(t)
+	harness.LoadFile("../../.kit/extensions/subagent-monitor.go")
+
+	// Start session
+	_, err := harness.Emit(extensions.SessionStartEvent{SessionID: "test-session"})
+	if err != nil {
+		t.Fatalf("SessionStart should not error: %v", err)
+	}
+
+	// Emit SubagentStart
+	_, err = harness.Emit(extensions.SubagentStartEvent{
+		ToolCallID: "call-1",
+		Task:       "test task",
+	})
+	if err != nil {
+		t.Fatalf("SubagentStart should not error: %v", err)
+	}
+
+	// Emit a few chunks
+	for i := range 3 {
+		_, err = harness.Emit(extensions.SubagentChunkEvent{
+			ToolCallID: "call-1",
+			Task:       "test task",
+			ChunkType:  "text",
+			Content:    fmt.Sprintf("line %d", i),
+		})
+		if err != nil {
+			t.Fatalf("SubagentChunk %d should not error: %v", i, err)
+		}
+	}
+
+	// Emit tool call chunk
+	_, err = harness.Emit(extensions.SubagentChunkEvent{
+		ToolCallID: "call-1",
+		Task:       "test task",
+		ChunkType:  "tool_call",
+		ToolName:   "bash",
+	})
+	if err != nil {
+		t.Fatalf("SubagentChunk tool_call should not error: %v", err)
+	}
+
+	// Emit SubagentEnd
+	_, err = harness.Emit(extensions.SubagentEndEvent{
+		ToolCallID: "call-1",
+		Task:       "test task",
+		Response:   "done",
+	})
+	if err != nil {
+		t.Fatalf("SubagentEnd should not error: %v", err)
+	}
+
+	// Give time for cleanup goroutine
+	time.Sleep(100 * time.Millisecond)
+}
+
+// TestSubagentMonitor_MultipleSubagents verifies multiple parallel subagents.
+func TestSubagentMonitor_MultipleSubagents(t *testing.T) {
+	harness := test.New(t)
+	harness.LoadFile("../../.kit/extensions/subagent-monitor.go")
+
+	_, err := harness.Emit(extensions.SessionStartEvent{SessionID: "test-session"})
+	if err != nil {
+		t.Fatalf("SessionStart should not error: %v", err)
+	}
+
+	// Start 3 subagents
+	for i := 1; i <= 3; i++ {
+		_, err := harness.Emit(extensions.SubagentStartEvent{
+			ToolCallID: fmt.Sprintf("call-%d", i),
+			Task:       fmt.Sprintf("task %d", i),
+		})
+		if err != nil {
+			t.Fatalf("SubagentStart %d should not error: %v", i, err)
+		}
+	}
+
+	// Emit chunks for each
+	for i := 1; i <= 3; i++ {
+		_, err := harness.Emit(extensions.SubagentChunkEvent{
+			ToolCallID: fmt.Sprintf("call-%d", i),
+			Task:       fmt.Sprintf("task %d", i),
+			ChunkType:  "text",
+			Content:    fmt.Sprintf("output from agent %d", i),
+		})
+		if err != nil {
+			t.Fatalf("SubagentChunk %d should not error: %v", i, err)
+		}
+	}
+
+	// End all subagents
+	for i := 1; i <= 3; i++ {
+		_, err := harness.Emit(extensions.SubagentEndEvent{
+			ToolCallID: fmt.Sprintf("call-%d", i),
+			Task:       fmt.Sprintf("task %d", i),
+			Response:   "completed",
+		})
+		if err != nil {
+			t.Fatalf("SubagentEnd %d should not error: %v", i, err)
+		}
+	}
+
+	time.Sleep(100 * time.Millisecond)
+}
+
+// TestSubagentMonitor_SessionShutdown verifies shutdown doesn't panic
+// even with nil ctx functions.
+func TestSubagentMonitor_SessionShutdown(t *testing.T) {
+	harness := test.New(t)
+	harness.LoadFile("../../.kit/extensions/subagent-monitor.go")
+
+	// Start then shutdown
+	_, err := harness.Emit(extensions.SessionStartEvent{SessionID: "test-session"})
+	if err != nil {
+		t.Fatalf("SessionStart should not error: %v", err)
+	}
+
+	// Start a subagent
+	_, err = harness.Emit(extensions.SubagentStartEvent{
+		ToolCallID: "call-1",
+		Task:       "test task",
+	})
+	if err != nil {
+		t.Fatalf("SubagentStart should not error: %v", err)
+	}
+
+	// Shutdown - should not panic even with active subagent
+	_, err = harness.Emit(extensions.SessionShutdownEvent{})
+	if err != nil {
+		t.Fatalf("SessionShutdown should not error: %v", err)
+	}
+}

examples/extensions/subagent-test.go

@@ -37,7 +37,7 @@ func Init(api ext.API) {
 			"Subagent Test Extension loaded\n\n" +
 				"/subtest <task>    Spawn blocking subagent\n" +
 				"/subbg <task>      Spawn background subagent\n\n" +
-				"The LLM can also use the spawn_subagent tool.")
+				"The LLM can also use the subagent tool.")
 	})
 
 	api.OnAgentEnd(func(_ ext.AgentEndEvent, ctx ext.Context) {

examples/sdk/README.md

@@ -0,0 +1,45 @@
+# SDK Examples
+
+These examples demonstrate how to use the Kit SDK (`pkg/kit`) to build agents programmatically in Go.
+
+## Examples
+
+### [basic](basic/)
+
+Shows core SDK usage: creating a Kit instance, sending prompts, overriding the model, subscribing to events (tool calls, streaming), and session management.
+
+```bash
+go run ./examples/sdk/basic
+```
+
+### [scripting](scripting/)
+
+A minimal script-friendly wrapper that takes a prompt from the command line and prints the response — useful for piping and automation.
+
+```bash
+go run ./examples/sdk/scripting "Explain what this repo does"
+```
+
+### [crypto-monitor](crypto-monitor/)
+
+A background agent that checks Bitcoin and Ethereum prices every 30 minutes and sends desktop notifications via `notify-send` (dbus). Demonstrates using the SDK for a long-running autonomous task with a single tool.
+
+```bash
+go run ./examples/sdk/crypto-monitor
+
+# Override the check interval:
+CRYPTO_INTERVAL=5m go run ./examples/sdk/crypto-monitor
+```
+
+## Getting Started
+
+```go
+import kit "github.com/mark3labs/kit/pkg/kit"
+
+host, err := kit.New(ctx, nil)        // uses ~/.kit.yml defaults
+defer host.Close()
+
+response, err := host.Prompt(ctx, "Hello!")
+```
+
+See the [SDK README](../../pkg/kit/README.md) for the full API reference.

examples/sdk/crypto-monitor/main.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"os"
+	"os/signal"
+	"time"
+
+	kit "github.com/mark3labs/kit/pkg/kit"
+)
+
+const systemPrompt = `You are a cryptocurrency price monitor. Your job is to:
+
+1. Fetch the current prices of Bitcoin and Ethereum using bash with curl
+2. Send a desktop notification with the results using notify-send
+
+To fetch prices, use this CoinGecko API endpoint (no API key needed):
+  curl -s 'https://api.coingecko.com/api/v3/simple/price?ids=bitcoin,ethereum&vs_currencies=usd&include_24hr_change=true'
+
+To send a desktop notification:
+  notify-send -i dialog-information "Crypto Prices" "BTC: $XX,XXX (+X.X%)\nETH: $X,XXX (+X.X%)"
+
+Include the 24h percentage change in the notification. Use a green arrow (▲) for
+positive changes and a red arrow (▼) for negative. Format prices with commas.
+
+If the API call fails, send a notification about the failure instead.
+
+Always complete both steps: fetch then notify. Be concise — no commentary needed.`
+
+func main() {
+	interval := 30 * time.Minute
+	if os.Getenv("CRYPTO_INTERVAL") != "" {
+		d, err := time.ParseDuration(os.Getenv("CRYPTO_INTERVAL"))
+		if err == nil {
+			interval = d
+		}
+	}
+
+	ctx, cancel := signal.NotifyContext(context.Background(), os.Interrupt)
+	defer cancel()
+
+	host, err := kit.New(ctx, &kit.Options{
+		SystemPrompt: systemPrompt,
+		Tools:        []kit.Tool{kit.NewBashTool()},
+		NoSession:    true,
+		Quiet:        true,
+	})
+	if err != nil {
+		log.Fatalf("Failed to create kit instance: %v", err)
+	}
+	defer func() { _ = host.Close() }()
+
+	fmt.Printf("Crypto price monitor started (every %s)\n", interval)
+	fmt.Println("Press Ctrl+C to stop")
+
+	// Run immediately on startup, then on each tick.
+	check(ctx, host)
+
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ticker.C:
+			check(ctx, host)
+		case <-ctx.Done():
+			fmt.Println("\nStopping price monitor")
+			return
+		}
+	}
+}
+
+func check(ctx context.Context, host *kit.Kit) {
+	fmt.Printf("[%s] Checking prices...\n", time.Now().Format("15:04:05"))
+
+	// Clear session so each check is independent.
+	host.ClearSession()
+
+	_, err := host.Prompt(ctx, "Fetch current Bitcoin and Ethereum prices and send a desktop notification.")
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+	}
+}

go.mod

@@ -1,69 +1,69 @@
 module github.com/mark3labs/kit
 
-go 1.26.0
+go 1.26.1
 
 require (
-	charm.land/bubbles/v2 v2.0.0
+	charm.land/bubbles/v2 v2.1.0
 	charm.land/bubbletea/v2 v2.0.2
-	charm.land/fantasy v0.11.1
-	charm.land/lipgloss/v2 v2.0.1
+	charm.land/fantasy v0.17.1
+	charm.land/huh/v2 v2.0.3
+	charm.land/lipgloss/v2 v2.0.2
 	github.com/alecthomas/chroma/v2 v2.23.1
 	github.com/aymanbagabas/go-udiff v0.4.1
-	github.com/charmbracelet/fang v0.4.4
-	github.com/charmbracelet/log v0.4.2
-	github.com/mark3labs/mcp-go v0.44.1
+	github.com/charmbracelet/fang v1.0.0
+	github.com/charmbracelet/log v1.0.0
+	github.com/coder/acp-go-sdk v0.6.3
+	github.com/indaco/herald v0.10.0
+	github.com/indaco/herald-md v0.1.0
+	github.com/mark3labs/mcp-go v0.46.0
 	github.com/spf13/cobra v1.10.2
 	github.com/spf13/viper v1.21.0
 	github.com/traefik/yaegi v0.16.1
-	golang.org/x/term v0.40.0
+	golang.org/x/term v0.41.0
 	gopkg.in/yaml.v3 v3.0.1
 )
 
 require (
-	charm.land/huh/v2 v2.0.3 // indirect
 	cloud.google.com/go v0.123.0 // indirect
-	cloud.google.com/go/auth v0.18.2 // indirect
+	cloud.google.com/go/auth v0.19.0 // indirect
 	cloud.google.com/go/auth/oauth2adapt v0.2.8 // indirect
 	cloud.google.com/go/compute/metadata v0.9.0 // indirect
 	github.com/Azure/azure-sdk-for-go/sdk/azcore v1.21.0 // indirect
 	github.com/Azure/azure-sdk-for-go/sdk/internal v1.11.2 // indirect
 	github.com/atotto/clipboard v0.1.4 // indirect
-	github.com/aws/aws-sdk-go-v2 v1.41.3 // indirect
-	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.6 // indirect
-	github.com/aws/aws-sdk-go-v2/config v1.32.11 // indirect
-	github.com/aws/aws-sdk-go-v2/credentials v1.19.11 // indirect
-	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.19 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.19 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.19 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/ini v1.8.5 // indirect
-	github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.6 // indirect
-	github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.19 // indirect
-	github.com/aws/aws-sdk-go-v2/service/signin v1.0.7 // indirect
-	github.com/aws/aws-sdk-go-v2/service/sso v1.30.12 // indirect
-	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.16 // indirect
-	github.com/aws/aws-sdk-go-v2/service/sts v1.41.8 // indirect
+	github.com/aws/aws-sdk-go-v2 v1.41.5 // indirect
+	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.8 // indirect
+	github.com/aws/aws-sdk-go-v2/config v1.32.13 // indirect
+	github.com/aws/aws-sdk-go-v2/credentials v1.19.13 // indirect
+	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.21 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.21 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.21 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/ini v1.8.6 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.7 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.21 // indirect
+	github.com/aws/aws-sdk-go-v2/service/signin v1.0.9 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sso v1.30.14 // indirect
+	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.18 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sts v1.41.10 // indirect
 	github.com/aws/smithy-go v1.24.2 // indirect
-	github.com/aymerick/douceur v0.2.0 // indirect
-	github.com/bahlo/generic-list-go v0.2.0 // indirect
-	github.com/buger/jsonparser v1.1.1 // indirect
-	github.com/catppuccin/go v0.2.0 // indirect
+	github.com/catppuccin/go v0.3.0 // indirect
 	github.com/cespare/xxhash/v2 v2.3.0 // indirect
 	github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab // indirect
-	github.com/charmbracelet/colorprofile v0.4.2 // indirect
+	github.com/charmbracelet/colorprofile v0.4.3 // indirect
 	github.com/charmbracelet/harmonica v0.2.0 // indirect
 	github.com/charmbracelet/lipgloss v1.1.1-0.20250404203927-76690c660834 // indirect
-	github.com/charmbracelet/ultraviolet v0.0.0-20260303162955-0b88c25f3fff // indirect
+	github.com/charmbracelet/openai-go v0.0.0-20260319145158-d0740cc34266 // indirect
+	github.com/charmbracelet/ultraviolet v0.0.0-20260330092749-0f94982c930b // indirect
 	github.com/charmbracelet/x/cellbuf v0.0.15 // indirect
-	github.com/charmbracelet/x/exp/charmtone v0.0.0-20260305213658-fe36e8c10185 // indirect
+	github.com/charmbracelet/x/exp/charmtone v0.0.0-20260330094520-2dce04b6f8a4 // indirect
 	github.com/charmbracelet/x/exp/ordered v0.1.0 // indirect
-	github.com/charmbracelet/x/exp/slice v0.0.0-20260305213658-fe36e8c10185 // indirect
-	github.com/charmbracelet/x/exp/strings v0.0.0-20240722160745-212f7b056ed0 // indirect
+	github.com/charmbracelet/x/exp/slice v0.0.0-20260330094520-2dce04b6f8a4 // indirect
+	github.com/charmbracelet/x/exp/strings v0.1.0 // indirect
 	github.com/charmbracelet/x/json v0.2.0 // indirect
 	github.com/charmbracelet/x/termios v0.1.1 // indirect
 	github.com/charmbracelet/x/windows v0.2.2 // indirect
 	github.com/clipperhouse/displaywidth v0.11.0 // indirect
 	github.com/clipperhouse/uax29/v2 v2.7.0 // indirect
-	github.com/coder/acp-go-sdk v0.6.3 // indirect
 	github.com/dlclark/regexp2 v1.11.5 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/felixge/httpsnoop v1.0.4 // indirect
@@ -76,27 +76,22 @@ require (
 	github.com/goccy/go-yaml v1.19.2 // indirect
 	github.com/golang-jwt/jwt/v5 v5.3.0 // indirect
 	github.com/google/go-cmp v0.7.0 // indirect
+	github.com/google/jsonschema-go v0.4.2 // indirect
 	github.com/google/s2a-go v0.1.9 // indirect
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/googleapis/enterprise-certificate-proxy v0.3.14 // indirect
-	github.com/googleapis/gax-go/v2 v2.17.0 // indirect
-	github.com/gorilla/css v1.0.1 // indirect
+	github.com/googleapis/gax-go/v2 v2.20.0 // indirect
 	github.com/gorilla/websocket v1.5.3 // indirect
-	github.com/invopop/jsonschema v0.13.0 // indirect
-	github.com/kaptinlin/go-i18n v0.2.12 // indirect
+	github.com/kaptinlin/go-i18n v0.3.0 // indirect
 	github.com/kaptinlin/jsonpointer v0.4.17 // indirect
-	github.com/kaptinlin/jsonschema v0.7.5 // indirect
-	github.com/kaptinlin/messageformat-go v0.4.18 // indirect
-	github.com/mailru/easyjson v0.9.1 // indirect
-	github.com/microcosm-cc/bluemonday v1.0.27 // indirect
+	github.com/kaptinlin/jsonschema v0.7.7 // indirect
+	github.com/kaptinlin/messageformat-go v0.4.19 // indirect
 	github.com/mitchellh/hashstructure/v2 v2.0.2 // indirect
 	github.com/muesli/mango v0.2.0 // indirect
 	github.com/muesli/mango-cobra v1.3.0 // indirect
 	github.com/muesli/mango-pflag v0.2.0 // indirect
-	github.com/muesli/reflow v0.3.0 // indirect
 	github.com/muesli/roff v0.1.0 // indirect
-	github.com/openai/openai-go/v2 v2.7.1 // indirect
-	github.com/pelletier/go-toml/v2 v2.2.4 // indirect
+	github.com/pelletier/go-toml/v2 v2.3.0 // indirect
 	github.com/sagikazarmark/locafero v0.12.0 // indirect
 	github.com/spf13/afero v1.15.0 // indirect
 	github.com/spf13/cast v1.10.0 // indirect
@@ -105,45 +100,42 @@ require (
 	github.com/tidwall/match v1.2.0 // indirect
 	github.com/tidwall/pretty v1.2.1 // indirect
 	github.com/tidwall/sjson v1.2.5 // indirect
-	github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	github.com/yosida95/uritemplate/v3 v3.0.2 // indirect
-	github.com/yuin/goldmark v1.7.16 // indirect
-	github.com/yuin/goldmark-emoji v1.0.6 // indirect
+	github.com/yuin/goldmark v1.8.2 // indirect
 	go.opentelemetry.io/auto/sdk v1.2.1 // indirect
-	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.66.0 // indirect
-	go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.66.0 // indirect
-	go.opentelemetry.io/otel v1.41.0 // indirect
-	go.opentelemetry.io/otel/metric v1.41.0 // indirect
-	go.opentelemetry.io/otel/trace v1.41.0 // indirect
+	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0 // indirect
+	go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.67.0 // indirect
+	go.opentelemetry.io/otel v1.42.0 // indirect
+	go.opentelemetry.io/otel/metric v1.42.0 // indirect
+	go.opentelemetry.io/otel/trace v1.42.0 // indirect
 	go.yaml.in/yaml/v3 v3.0.4 // indirect
-	golang.org/x/crypto v0.48.0 // indirect
-	golang.org/x/exp v0.0.0-20260218203240-3dfff04db8fa // indirect
-	golang.org/x/net v0.51.0 // indirect
-	golang.org/x/oauth2 v0.35.0 // indirect
-	golang.org/x/time v0.14.0 // indirect
-	google.golang.org/api v0.269.0 // indirect
-	google.golang.org/genai v1.49.0 // indirect
-	google.golang.org/genproto/googleapis/rpc v0.0.0-20260226221140-a57be14db171 // indirect
-	google.golang.org/grpc v1.79.2 // indirect
+	golang.org/x/crypto v0.49.0 // indirect
+	golang.org/x/exp v0.0.0-20260312153236-7ab1446f8b90 // indirect
+	golang.org/x/net v0.52.0 // indirect
+	golang.org/x/oauth2 v0.36.0 // indirect
+	golang.org/x/time v0.15.0 // indirect
+	google.golang.org/api v0.273.0 // indirect
+	google.golang.org/genai v1.52.0 // indirect
+	google.golang.org/genproto/googleapis/rpc v0.0.0-20260319201613-d00831a3d3e7 // indirect
+	google.golang.org/grpc v1.79.3 // indirect
 	google.golang.org/protobuf v1.36.11 // indirect
 	gopkg.in/yaml.v2 v2.4.0 // indirect
 )
 
 require (
 	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
-	github.com/charmbracelet/glamour v0.10.0
-	github.com/charmbracelet/x/ansi v0.11.6 // indirect
+	github.com/charmbracelet/x/ansi v0.11.6
 	github.com/charmbracelet/x/term v0.2.2 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
-	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
+	github.com/lucasb-eyer/go-colorful v1.4.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
-	github.com/mattn/go-runewidth v0.0.20 // indirect
+	github.com/mattn/go-runewidth v0.0.21 // indirect
 	github.com/muesli/cancelreader v0.2.2 // indirect
 	github.com/muesli/termenv v0.16.0 // indirect
 	github.com/rivo/uniseg v0.4.7 // indirect
 	github.com/spf13/pflag v1.0.10 // indirect
-	golang.org/x/sync v0.19.0 // indirect
+	golang.org/x/sync v0.20.0 // indirect
 	golang.org/x/sys v0.42.0 // indirect
-	golang.org/x/text v0.34.0 // indirect
+	golang.org/x/text v0.35.0
 )

go.sum

@@ -1,21 +1,17 @@
-charm.land/bubbles/v2 v2.0.0 h1:tE3eK/pHjmtrDiRdoC9uGNLgpopOd8fjhEe31B/ai5s=
-charm.land/bubbles/v2 v2.0.0/go.mod h1:rCHoleP2XhU8um45NTuOWBPNVHxnkXKTiZqcclL/qOI=
-charm.land/bubbletea/v2 v2.0.1 h1:B8e9zzK7x9JJ+XvHGF4xnYu9Xa0E0y0MyggY6dbaCfQ=
-charm.land/bubbletea/v2 v2.0.1/go.mod h1:3LRff2U4WIYXy7MTxfbAQ+AdfM3D8Xuvz2wbsOD9OHQ=
+charm.land/bubbles/v2 v2.1.0 h1:YSnNh5cPYlYjPxRrzs5VEn3vwhtEn3jVGRBT3M7/I0g=
+charm.land/bubbles/v2 v2.1.0/go.mod h1:l97h4hym2hvWBVfmJDtrEHHCtkIKeTEb3TTJ4ZOB3wY=
 charm.land/bubbletea/v2 v2.0.2 h1:4CRtRnuZOdFDTWSff9r8QFt/9+z6Emubz3aDMnf/dx0=
 charm.land/bubbletea/v2 v2.0.2/go.mod h1:3LRff2U4WIYXy7MTxfbAQ+AdfM3D8Xuvz2wbsOD9OHQ=
-charm.land/fantasy v0.11.1 h1:G1dRqkzEQ0RJN1Ls5mte8HOi0wFKxYd5bfnRAmeYvDk=
-charm.land/fantasy v0.11.1/go.mod h1:C8wNxWlw+b2z54zsTor9r1tG2GE2C4QotvAlgXh9KF8=
+charm.land/fantasy v0.17.1 h1:SQzfnyJPDuQWt6e//KKmQmEEXdqHMC0IZz10XwkLcEM=
+charm.land/fantasy v0.17.1/go.mod h1:FF5ALCCHETacHJPBqU42CtwMInYQ0ul52fdzIHQMbQk=
 charm.land/huh/v2 v2.0.3 h1:2cJsMqEPwSywGHvdlKsJyQKPtSJLVnFKyFbsYZTlLkU=
 charm.land/huh/v2 v2.0.3/go.mod h1:93eEveeeqn47MwiC3tf+2atZ2l7Is88rAtmZNZ8x9Wc=
-charm.land/lipgloss/v2 v2.0.0 h1:sd8N/B3x892oiOjFfBQdXBQp3cAkvjGaU5TvVZC3ivo=
-charm.land/lipgloss/v2 v2.0.0/go.mod h1:w6SnmsBFBmEFBodiEDurGS/sdUY/u1+v72DqUzc6J14=
-charm.land/lipgloss/v2 v2.0.1 h1:6Xzrn49+Py1Um5q/wZG1gWgER2+7dUyZ9XMEufqPSys=
-charm.land/lipgloss/v2 v2.0.1/go.mod h1:KjPle2Qd3YmvP1KL5OMHiHysGcNwq6u83MUjYkFvEkM=
+charm.land/lipgloss/v2 v2.0.2 h1:xFolbF8JdpNkM2cEPTfXEcW1p6NRzOWTSamRfYEw8cs=
+charm.land/lipgloss/v2 v2.0.2/go.mod h1:KjPle2Qd3YmvP1KL5OMHiHysGcNwq6u83MUjYkFvEkM=
 cloud.google.com/go v0.123.0 h1:2NAUJwPR47q+E35uaJeYoNhuNEM9kM8SjgRgdeOJUSE=
 cloud.google.com/go v0.123.0/go.mod h1:xBoMV08QcqUGuPW65Qfm1o9Y4zKZBpGS+7bImXLTAZU=
-cloud.google.com/go/auth v0.18.2 h1:+Nbt5Ev0xEqxlNjd6c+yYUeosQ5TtEUaNcN/3FozlaM=
-cloud.google.com/go/auth v0.18.2/go.mod h1:xD+oY7gcahcu7G2SG2DsBerfFxgPAJz17zz2joOFF3M=
+cloud.google.com/go/auth v0.19.0 h1:DGYwtbcsGsT1ywuxsIoWi1u/vlks0moIblQHgSDgQkQ=
+cloud.google.com/go/auth v0.19.0/go.mod h1:2Aph7BT2KnaSFOM0JDPyiYgNh6PL9vGMiP8CUIXZ+IY=
 cloud.google.com/go/auth/oauth2adapt v0.2.8 h1:keo8NaayQZ6wimpNSmW5OPc283g65QNIiLpZnkHRbnc=
 cloud.google.com/go/auth/oauth2adapt v0.2.8/go.mod h1:XQ9y31RkqZCcwJWNSx2Xvric3RrU88hAYYbjDWYDL+c=
 cloud.google.com/go/compute/metadata v0.9.0 h1:pDUj4QMoPejqq20dK0Pg2N4yG9zIkYGdBtwLoEkH9Zs=
@@ -38,82 +34,78 @@ github.com/alecthomas/repr v0.5.2 h1:SU73FTI9D1P5UNtvseffFSGmdNci/O6RsqzeXJtP0Qs
 github.com/alecthomas/repr v0.5.2/go.mod h1:Fr0507jx4eOXV7AlPV6AVZLYrLIuIeSOWtW57eE/O/4=
 github.com/atotto/clipboard v0.1.4 h1:EH0zSVneZPSuFR11BlR9YppQTVDbh5+16AmcJi4g1z4=
 github.com/atotto/clipboard v0.1.4/go.mod h1:ZY9tmq7sm5xIbd9bOK4onWV4S6X0u6GY7Vn0Yu86PYI=
-github.com/aws/aws-sdk-go-v2 v1.41.3 h1:4kQ/fa22KjDt13QCy1+bYADvdgcxpfH18f0zP542kZA=
-github.com/aws/aws-sdk-go-v2 v1.41.3/go.mod h1:mwsPRE8ceUUpiTgF7QmQIJ7lgsKUPQOUl3o72QBrE1o=
-github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.6 h1:N4lRUXZpZ1KVEUn6hxtco/1d2lgYhNn1fHkkl8WhlyQ=
-github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.6/go.mod h1:lyw7GFp3qENLh7kwzf7iMzAxDn+NzjXEAGjKS2UOKqI=
-github.com/aws/aws-sdk-go-v2/config v1.32.11 h1:ftxI5sgz8jZkckuUHXfC/wMUc8u3fG1vQS0plr2F2Zs=
-github.com/aws/aws-sdk-go-v2/config v1.32.11/go.mod h1:twF11+6ps9aNRKEDimksp923o44w/Thk9+8YIlzWMmo=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.11 h1:NdV8cwCcAXrCWyxArt58BrvZJ9pZ9Fhf9w6Uh5W3Uyc=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.11/go.mod h1:30yY2zqkMPdrvxBqzI9xQCM+WrlrZKSOpSJEsylVU+8=
-github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.19 h1:INUvJxmhdEbVulJYHI061k4TVuS3jzzthNvjqvVvTKM=
-github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.19/go.mod h1:FpZN2QISLdEBWkayloda+sZjVJL+e9Gl0k1SyTgcswU=
-github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.19 h1:/sECfyq2JTifMI2JPyZ4bdRN77zJmr6SrS1eL3augIA=
-github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.19/go.mod h1:dMf8A5oAqr9/oxOfLkC/c2LU/uMcALP0Rgn2BD5LWn0=
-github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.19 h1:AWeJMk33GTBf6J20XJe6qZoRSJo0WfUhsMdUKhoODXE=
-github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.19/go.mod h1:+GWrYoaAsV7/4pNHpwh1kiNLXkKaSoppxQq9lbH8Ejw=
-github.com/aws/aws-sdk-go-v2/internal/ini v1.8.5 h1:clHU5fm//kWS1C2HgtgWxfQbFbx4b6rx+5jzhgX9HrI=
-github.com/aws/aws-sdk-go-v2/internal/ini v1.8.5/go.mod h1:O3h0IK87yXci+kg6flUKzJnWeziQUKciKrLjcatSNcY=
-github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.6 h1:XAq62tBTJP/85lFD5oqOOe7YYgWxY9LvWq8plyDvDVg=
-github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.6/go.mod h1:x0nZssQ3qZSnIcePWLvcoFisRXJzcTVvYpAAdYX8+GI=
-github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.19 h1:X1Tow7suZk9UCJHE1Iw9GMZJJl0dAnKXXP1NaSDHwmw=
-github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.19/go.mod h1:/rARO8psX+4sfjUQXp5LLifjUt8DuATZ31WptNJTyQA=
-github.com/aws/aws-sdk-go-v2/service/signin v1.0.7 h1:Y2cAXlClHsXkkOvWZFXATr34b0hxxloeQu/pAZz2row=
-github.com/aws/aws-sdk-go-v2/service/signin v1.0.7/go.mod h1:idzZ7gmDeqeNrSPkdbtMp9qWMgcBwykA7P7Rzh5DXVU=
-github.com/aws/aws-sdk-go-v2/service/sso v1.30.12 h1:iSsvB9EtQ09YrsmIc44Heqlx5ByGErqhPK1ZQLppias=
-github.com/aws/aws-sdk-go-v2/service/sso v1.30.12/go.mod h1:fEWYKTRGoZNl8tZ77i61/ccwOMJdGxwOhWCkp6TXAr0=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.16 h1:EnUdUqRP1CNzt2DkV67tJx6XDN4xlfBFm+bzeNOQVb0=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.16/go.mod h1:Jic/xv0Rq/pFNCh3WwpH4BEqdbSAl+IyHro8LbibHD8=
-github.com/aws/aws-sdk-go-v2/service/sts v1.41.8 h1:XQTQTF75vnug2TXS8m7CVJfC2nniYPZnO1D4Np761Oo=
-github.com/aws/aws-sdk-go-v2/service/sts v1.41.8/go.mod h1:Xgx+PR1NUOjNmQY+tRMnouRp83JRM8pRMw/vCaVhPkI=
+github.com/aws/aws-sdk-go-v2 v1.41.5 h1:dj5kopbwUsVUVFgO4Fi5BIT3t4WyqIDjGKCangnV/yY=
+github.com/aws/aws-sdk-go-v2 v1.41.5/go.mod h1:mwsPRE8ceUUpiTgF7QmQIJ7lgsKUPQOUl3o72QBrE1o=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.8 h1:eBMB84YGghSocM7PsjmmPffTa+1FBUeNvGvFou6V/4o=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.8/go.mod h1:lyw7GFp3qENLh7kwzf7iMzAxDn+NzjXEAGjKS2UOKqI=
+github.com/aws/aws-sdk-go-v2/config v1.32.13 h1:5KgbxMaS2coSWRrx9TX/QtWbqzgQkOdEa3sZPhBhCSg=
+github.com/aws/aws-sdk-go-v2/config v1.32.13/go.mod h1:8zz7wedqtCbw5e9Mi2doEwDyEgHcEE9YOJp6a8jdSMY=
+github.com/aws/aws-sdk-go-v2/credentials v1.19.13 h1:mA59E3fokBvyEGHKFdnpNNrvaR351cqiHgRg+JzOSRI=
+github.com/aws/aws-sdk-go-v2/credentials v1.19.13/go.mod h1:yoTXOQKea18nrM69wGF9jBdG4WocSZA1h38A+t/MAsk=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.21 h1:NUS3K4BTDArQqNu2ih7yeDLaS3bmHD0YndtA6UP884g=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.21/go.mod h1:YWNWJQNjKigKY1RHVJCuupeWDrrHjRqHm0N9rdrWzYI=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.21 h1:Rgg6wvjjtX8bNHcvi9OnXWwcE0a2vGpbwmtICOsvcf4=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.21/go.mod h1:A/kJFst/nm//cyqonihbdpQZwiUhhzpqTsdbhDdRF9c=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.21 h1:PEgGVtPoB6NTpPrBgqSE5hE/o47Ij9qk/SEZFbUOe9A=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.21/go.mod h1:p+hz+PRAYlY3zcpJhPwXlLC4C+kqn70WIHwnzAfs6ps=
+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.6 h1:qYQ4pzQ2Oz6WpQ8T3HvGHnZydA72MnLuFK9tJwmrbHw=
+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.6/go.mod h1:O3h0IK87yXci+kg6flUKzJnWeziQUKciKrLjcatSNcY=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.7 h1:5EniKhLZe4xzL7a+fU3C2tfUN4nWIqlLesfrjkuPFTY=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.7/go.mod h1:x0nZssQ3qZSnIcePWLvcoFisRXJzcTVvYpAAdYX8+GI=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.21 h1:c31//R3xgIJMSC8S6hEVq+38DcvUlgFY0FM6mSI5oto=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.21/go.mod h1:r6+pf23ouCB718FUxaqzZdbpYFyDtehyZcmP5KL9FkA=
+github.com/aws/aws-sdk-go-v2/service/signin v1.0.9 h1:QKZH0S178gCmFEgst8hN0mCX1KxLgHBKKY/CLqwP8lg=
+github.com/aws/aws-sdk-go-v2/service/signin v1.0.9/go.mod h1:7yuQJoT+OoH8aqIxw9vwF+8KpvLZ8AWmvmUWHsGQZvI=
+github.com/aws/aws-sdk-go-v2/service/sso v1.30.14 h1:GcLE9ba5ehAQma6wlopUesYg/hbcOhFNWTjELkiWkh4=
+github.com/aws/aws-sdk-go-v2/service/sso v1.30.14/go.mod h1:WSvS1NLr7JaPunCXqpJnWk1Bjo7IxzZXrZi1QQCkuqM=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.18 h1:mP49nTpfKtpXLt5SLn8Uv8z6W+03jYVoOSAl/c02nog=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.18/go.mod h1:YO8TrYtFdl5w/4vmjL8zaBSsiNp3w0L1FfKVKenZT7w=
+github.com/aws/aws-sdk-go-v2/service/sts v1.41.10 h1:p8ogvvLugcR/zLBXTXrTkj0RYBUdErbMnAFFp12Lm/U=
+github.com/aws/aws-sdk-go-v2/service/sts v1.41.10/go.mod h1:60dv0eZJfeVXfbT1tFJinbHrDfSJ2GZl4Q//OSSNAVw=
 github.com/aws/smithy-go v1.24.2 h1:FzA3bu/nt/vDvmnkg+R8Xl46gmzEDam6mZ1hzmwXFng=
 github.com/aws/smithy-go v1.24.2/go.mod h1:YE2RhdIuDbA5E5bTdciG9KrW3+TiEONeUWCqxX9i1Fc=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=
-github.com/aymanbagabas/go-udiff v0.4.0 h1:TKnLPh7IbnizJIBKFWa9mKayRUBQ9Kh1BPCk6w2PnYM=
-github.com/aymanbagabas/go-udiff v0.4.0/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
 github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ12Gv5o=
 github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
-github.com/aymerick/douceur v0.2.0 h1:Mv+mAeH1Q+n9Fr+oyamOlAkUNPWPlA8PPGR0QAaYuPk=
-github.com/aymerick/douceur v0.2.0/go.mod h1:wlT5vV2O3h55X9m7iVYN0TBM0NH/MmbLnd30/FjWUq4=
-github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
-github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
-github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=
-github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
-github.com/catppuccin/go v0.2.0 h1:ktBeIrIP42b/8FGiScP9sgrWOss3lw0Z5SktRoithGA=
-github.com/catppuccin/go v0.2.0/go.mod h1:8IHJuMGaUUjQM82qBrGNBv7LFq6JI3NnQCF6MOlZjpc=
+github.com/catppuccin/go v0.3.0 h1:d+0/YicIq+hSTo5oPuRi5kOpqkVA5tAsU6dNhvRu+aY=
+github.com/catppuccin/go v0.3.0/go.mod h1:8IHJuMGaUUjQM82qBrGNBv7LFq6JI3NnQCF6MOlZjpc=
 github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=
 github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
 github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab h1:J7XQLgl9sefgTnTGrmX3xqvp5o6MCiBzEjGv5igAlc4=
 github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab/go.mod h1:hqlYqR7uPKOKfnNeicUbZp0Ps0GeYFlKYtwh5HGDCx8=
-github.com/charmbracelet/colorprofile v0.4.2 h1:BdSNuMjRbotnxHSfxy+PCSa4xAmz7szw70ktAtWRYrY=
-github.com/charmbracelet/colorprofile v0.4.2/go.mod h1:0rTi81QpwDElInthtrQ6Ni7cG0sDtwAd4C4le060fT8=
-github.com/charmbracelet/fang v0.4.4 h1:G4qKxF6or/eTPgmAolwPuRNyuci3hTUGGX1rj1YkHJY=
-github.com/charmbracelet/fang v0.4.4/go.mod h1:P5/DNb9DddQ0Z0dbc0P3ol4/ix5Po7Ofr2KMBfAqoCo=
-github.com/charmbracelet/glamour v0.10.0 h1:MtZvfwsYCx8jEPFJm3rIBFIMZUfUJ765oX8V6kXldcY=
-github.com/charmbracelet/glamour v0.10.0/go.mod h1:f+uf+I/ChNmqo087elLnVdCiVgjSKWuXa/l6NU2ndYk=
+github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=
+github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=
+github.com/charmbracelet/fang v1.0.0 h1:jESBY40agJOlLYnnv9jE0mLqDGTxEk0hkOnx7YGyRlQ=
+github.com/charmbracelet/fang v1.0.0/go.mod h1:P5/DNb9DddQ0Z0dbc0P3ol4/ix5Po7Ofr2KMBfAqoCo=
 github.com/charmbracelet/harmonica v0.2.0 h1:8NxJWRWg/bzKqqEaaeFNipOu77YR5t8aSwG4pgaUBiQ=
 github.com/charmbracelet/harmonica v0.2.0/go.mod h1:KSri/1RMQOZLbw7AHqgcBycp8pgJnQMYYT8QZRqZ1Ao=
 github.com/charmbracelet/lipgloss v1.1.1-0.20250404203927-76690c660834 h1:ZR7e0ro+SZZiIZD7msJyA+NjkCNNavuiPBLgerbOziE=
 github.com/charmbracelet/lipgloss v1.1.1-0.20250404203927-76690c660834/go.mod h1:aKC/t2arECF6rNOnaKaVU6y4t4ZeHQzqfxedE/VkVhA=
-github.com/charmbracelet/log v0.4.2 h1:hYt8Qj6a8yLnvR+h7MwsJv/XvmBJXiueUcI3cIxsyig=
-github.com/charmbracelet/log v0.4.2/go.mod h1:qifHGX/tc7eluv2R6pWIpyHDDrrb/AG71Pf2ysQu5nw=
-github.com/charmbracelet/ultraviolet v0.0.0-20260303162955-0b88c25f3fff h1:uY7A6hTokHPJBHfq7rj9Y/wm+IAjOghZTxKfVW6QLvw=
-github.com/charmbracelet/ultraviolet v0.0.0-20260303162955-0b88c25f3fff/go.mod h1:E6/0abq9uG2SnM8IbLB9Y5SW09uIgfaFETk8aRzgXUQ=
+github.com/charmbracelet/log v1.0.0 h1:HVVVMmfOorfj3BA9i8X8UL69Hoz9lI0PYwXfJvOdRc4=
+github.com/charmbracelet/log v1.0.0/go.mod h1:uYgY3SmLpwJWxmlrPwXvzVYujxis1vAKRV/0VQB7yWA=
+github.com/charmbracelet/openai-go v0.0.0-20260319145158-d0740cc34266 h1:BW/sZtyd1JyYy0h5adMm3tzpNyL857LWjuTRET6OhpY=
+github.com/charmbracelet/openai-go v0.0.0-20260319145158-d0740cc34266/go.mod h1:1DahUaExbUZx/jD+FNT2PKP4L9rLE5+ZBRuI8mZjd/E=
+github.com/charmbracelet/ultraviolet v0.0.0-20260330092749-0f94982c930b h1:ASDO9RT6SNKTQN87jO2bRfxHFJq8cgeYdFzivY2gCeM=
+github.com/charmbracelet/ultraviolet v0.0.0-20260330092749-0f94982c930b/go.mod h1:Vo8TffMf0q7Uho/n8e6XpBZvOWtd3g39yX+9P5rRutA=
 github.com/charmbracelet/x/ansi v0.11.6 h1:GhV21SiDz/45W9AnV2R61xZMRri5NlLnl6CVF7ihZW8=
 github.com/charmbracelet/x/ansi v0.11.6/go.mod h1:2JNYLgQUsyqaiLovhU2Rv/pb8r6ydXKS3NIttu3VGZQ=
 github.com/charmbracelet/x/cellbuf v0.0.15 h1:ur3pZy0o6z/R7EylET877CBxaiE1Sp1GMxoFPAIztPI=
 github.com/charmbracelet/x/cellbuf v0.0.15/go.mod h1:J1YVbR7MUuEGIFPCaaZ96KDl5NoS0DAWkskup+mOY+Q=
-github.com/charmbracelet/x/exp/charmtone v0.0.0-20260305213658-fe36e8c10185 h1:/192monmpmRICpSPrFRzkIO+xfhioV6/nwrQdkDTj10=
-github.com/charmbracelet/x/exp/charmtone v0.0.0-20260305213658-fe36e8c10185/go.mod h1:nsExn0DGyX0lh9LwLHTn2Gg+hafdzfSXnC+QmEJTZFY=
+github.com/charmbracelet/x/conpty v0.1.1 h1:s1bUxjoi7EpqiXysVtC+a8RrvPPNcNvAjfi4jxsAuEs=
+github.com/charmbracelet/x/conpty v0.1.1/go.mod h1:OmtR77VODEFbiTzGE9G1XiRJAga6011PIm4u5fTNZpk=
+github.com/charmbracelet/x/errors v0.0.0-20240508181413-e8d8b6e2de86 h1:JSt3B+U9iqk37QUU2Rvb6DSBYRLtWqFqfxf8l5hOZUA=
+github.com/charmbracelet/x/errors v0.0.0-20240508181413-e8d8b6e2de86/go.mod h1:2P0UgXMEa6TsToMSuFqKFQR+fZTO9CNGUNokkPatT/0=
+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260330094520-2dce04b6f8a4 h1:pIj18ZCZO4WOVj7jwjLoUb1lC7rS/I8oC3fZWXugNaY=
+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260330094520-2dce04b6f8a4/go.mod h1:nsExn0DGyX0lh9LwLHTn2Gg+hafdzfSXnC+QmEJTZFY=
 github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f h1:pk6gmGpCE7F3FcjaOEKYriCvpmIN4+6OS/RD0vm4uIA=
 github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f/go.mod h1:IfZAMTHB6XkZSeXUqriemErjAWCCzT0LwjKFYCZyw0I=
 github.com/charmbracelet/x/exp/ordered v0.1.0 h1:55/qLwjIh0gL0Vni+QAWk7T/qRVP6sBf+2agPBgnOFE=
 github.com/charmbracelet/x/exp/ordered v0.1.0/go.mod h1:5UHwmG+is5THxMyCJHNPCn2/ecI07aKNrW+LcResjJ8=
-github.com/charmbracelet/x/exp/slice v0.0.0-20260305213658-fe36e8c10185 h1:bloHJLweYZeIkBVgi8AF94DrTdx3eoEB57VOpFuFi3U=
-github.com/charmbracelet/x/exp/slice v0.0.0-20260305213658-fe36e8c10185/go.mod h1:vqEfX6xzqW1pKKZUUiFOKg0OQ7bCh54Q2vR/tserrRA=
-github.com/charmbracelet/x/exp/strings v0.0.0-20240722160745-212f7b056ed0 h1:qko3AQ4gK1MTS/de7F5hPGx6/k1u0w4TeYmBFwzYVP4=
-github.com/charmbracelet/x/exp/strings v0.0.0-20240722160745-212f7b056ed0/go.mod h1:pBhA0ybfXv6hDjQUZ7hk1lVxBiUbupdw5R31yPUViVQ=
+github.com/charmbracelet/x/exp/slice v0.0.0-20260330094520-2dce04b6f8a4 h1:VSd4zShIAf/4FgEDFJpapEcAPrc7h3dyyN7V9JlJpQw=
+github.com/charmbracelet/x/exp/slice v0.0.0-20260330094520-2dce04b6f8a4/go.mod h1:vqEfX6xzqW1pKKZUUiFOKg0OQ7bCh54Q2vR/tserrRA=
+github.com/charmbracelet/x/exp/strings v0.1.0 h1:i69S2XI7uG1u4NLGeJPSYU++Nmjvpo9nwd6aoEm7gkA=
+github.com/charmbracelet/x/exp/strings v0.1.0/go.mod h1:/ehtMPNh9K4odGFkqYJKpIYyePhdp1hLBRvyY4bWkH8=
 github.com/charmbracelet/x/json v0.2.0 h1:DqB+ZGx2h+Z+1s98HOuOyli+i97wsFQIxP2ZQANTPrQ=
 github.com/charmbracelet/x/json v0.2.0/go.mod h1:opFIflx2YgXgi49xVUu8gEQ21teFAxyMwvOiZhIvWNM=
 github.com/charmbracelet/x/term v0.2.2 h1:xVRT/S2ZcKdhhOuSP4t5cLi5o+JxklsoEObBSgfgZRk=
@@ -122,6 +114,8 @@ github.com/charmbracelet/x/termios v0.1.1 h1:o3Q2bT8eqzGnGPOYheoYS8eEleT5ZVNYNy8
 github.com/charmbracelet/x/termios v0.1.1/go.mod h1:rB7fnv1TgOPOyyKRJ9o+AsTU/vK5WHJ2ivHeut/Pcwo=
 github.com/charmbracelet/x/windows v0.2.2 h1:IofanmuvaxnKHuV04sC0eBy/smG6kIKrWG2/jYn2GuM=
 github.com/charmbracelet/x/windows v0.2.2/go.mod h1:/8XtdKZzedat74NQFn0NGlGL4soHB0YQZrETF96h75k=
+github.com/charmbracelet/x/xpty v0.1.3 h1:eGSitii4suhzrISYH50ZfufV3v085BXQwIytcOdFSsw=
+github.com/charmbracelet/x/xpty v0.1.3/go.mod h1:poPYpWuLDBFCKmKLDnhBp51ATa0ooD8FhypRwEFtH3Y=
 github.com/clipperhouse/displaywidth v0.11.0 h1:lBc6kY44VFw+TDx4I8opi/EtL9m20WSEFgwIwO+UVM8=
 github.com/clipperhouse/displaywidth v0.11.0/go.mod h1:bkrFNkf81G8HyVqmKGxsPufD3JhNl3dSqnGhOoSD/o0=
 github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJv2v7Vk=
@@ -131,6 +125,8 @@ github.com/cncf/xds/go v0.0.0-20260202195803-dba9d589def2/go.mod h1:qwXFYgsP6T7X
 github.com/coder/acp-go-sdk v0.6.3 h1:LsXQytehdjKIYJnoVWON/nf7mqbiarnyuyE3rrjBsXQ=
 github.com/coder/acp-go-sdk v0.6.3/go.mod h1:yKzM/3R9uELp4+nBAwwtkS0aN1FOFjo11CNPy37yFko=
 github.com/cpuguy83/go-md2man/v2 v2.0.6/go.mod h1:oOW0eioCTA6cOiMLiUPZOpcVxMig6NIQQ7OS05n1F4g=
+github.com/creack/pty v1.1.24 h1:bJrF4RRfyJnbTJqzRLHzcGaZK1NeM5kTC9jGgovnR1s=
+github.com/creack/pty v1.1.24/go.mod h1:08sCNb52WyoAwi2QDyzUCTgcvVFhUzewun7wtTfvcwE=
 github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc h1:U9qPSI2PIWSS1VwoXQT9A3Wy9MM3WgvqSxFWenqJduM=
 github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/dlclark/regexp2 v1.11.5 h1:Q/sSnsKerHeCkc/jSTNq1oCm7KiVgUMZRDUoRu0JQZQ=
@@ -169,51 +165,48 @@ github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek
 github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
 github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
+github.com/google/jsonschema-go v0.4.2 h1:tmrUohrwoLZZS/P3x7ex0WAVknEkBZM46iALbcqoRA8=
+github.com/google/jsonschema-go v0.4.2/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=
 github.com/google/s2a-go v0.1.9 h1:LGD7gtMgezd8a/Xak7mEWL0PjoTQFvpRudN895yqKW0=
 github.com/google/s2a-go v0.1.9/go.mod h1:YA0Ei2ZQL3acow2O62kdp9UlnvMmU7kA6Eutn0dXayM=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/googleapis/enterprise-certificate-proxy v0.3.14 h1:yh8ncqsbUY4shRD5dA6RlzjJaT4hi3kII+zYw8wmLb8=
 github.com/googleapis/enterprise-certificate-proxy v0.3.14/go.mod h1:vqVt9yG9480NtzREnTlmGSBmFrA+bzb0yl0TxoBQXOg=
-github.com/googleapis/gax-go/v2 v2.17.0 h1:RksgfBpxqff0EZkDWYuz9q/uWsTVz+kf43LsZ1J6SMc=
-github.com/googleapis/gax-go/v2 v2.17.0/go.mod h1:mzaqghpQp4JDh3HvADwrat+6M3MOIDp5YKHhb9PAgDY=
-github.com/gorilla/css v1.0.1 h1:ntNaBIghp6JmvWnxbZKANoLyuXTPZ4cAMlo6RyhlbO8=
-github.com/gorilla/css v1.0.1/go.mod h1:BvnYkspnSzMmwRK+b8/xgNPLiIuNZr6vbZBTPQ2A3b0=
+github.com/googleapis/gax-go/v2 v2.20.0 h1:NIKVuLhDlIV74muWlsMM4CcQZqN6JJ20Qcxd9YMuYcs=
+github.com/googleapis/gax-go/v2 v2.20.0/go.mod h1:But/NJU6TnZsrLai/xBAQLLz+Hc7fHZJt/hsCz3Fih4=
 github.com/gorilla/websocket v1.5.3 h1:saDtZ6Pbx/0u+bgYQ3q96pZgCzfhKXGPqt7kZ72aNNg=
 github.com/gorilla/websocket v1.5.3/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/adAjf1fMHhE=
 github.com/hexops/gotextdiff v1.0.3 h1:gitA9+qJrrTCsiCl7+kh75nPqQt1cx4ZkudSTLoUqJM=
 github.com/hexops/gotextdiff v1.0.3/go.mod h1:pSWU5MAI3yDq+fZBTazCSJysOMbxWL1BSow5/V2vxeg=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
-github.com/invopop/jsonschema v0.13.0 h1:KvpoAJWEjR3uD9Kbm2HWJmqsEaHt8lBUpd0qHcIi21E=
-github.com/invopop/jsonschema v0.13.0/go.mod h1:ffZ5Km5SWWRAIN6wbDXItl95euhFz2uON45H2qjYt+0=
-github.com/kaptinlin/go-i18n v0.2.12 h1:ywDsvb4KDFddMC2dpI/rrIzGU2mWUSvHmWUm9BMsdl4=
-github.com/kaptinlin/go-i18n v0.2.12/go.mod h1:pVcu9qsW5pOIOoZFJXesRYmLos1vMQrby70JPAoWmJU=
+github.com/indaco/herald v0.10.0 h1:XzahEKX6cr50qZQrUdA3QrQBHg8uGm5jETD0UDi21BI=
+github.com/indaco/herald v0.10.0/go.mod h1:T5g1+XLYvpjouhzAGHnAHDCKizhESkoV6+QPZ3DhgWA=
+github.com/indaco/herald-md v0.1.0 h1:zmYudYo+uamzKTBcIffJVJYrqk9xDNnVrTh+de2zciw=
+github.com/indaco/herald-md v0.1.0/go.mod h1:Z1HxPCbSn+/+TFzOM/UbsmKeEk/28NNI6JOTileKXto=
+github.com/kaptinlin/go-i18n v0.3.0 h1:wP76dvYg04bvwTb+8NB+CmdZ2kL7lSSCQ9B/kFv7QHo=
+github.com/kaptinlin/go-i18n v0.3.0/go.mod h1:pVcu9qsW5pOIOoZFJXesRYmLos1vMQrby70JPAoWmJU=
 github.com/kaptinlin/jsonpointer v0.4.17 h1:mY9k8ciWncxbsECyaxKnR0MdmxamNdp2tLQkAKVrtSk=
 github.com/kaptinlin/jsonpointer v0.4.17/go.mod h1:SsfsjqnHG5zuKo1DTBzk1VknaHlL4osHw+X9kZKukpU=
-github.com/kaptinlin/jsonschema v0.7.5 h1:jkK4a3NyzNoGlvu12CsL3IcqNMVa5sL51HPVa0nWcPY=
-github.com/kaptinlin/jsonschema v0.7.5/go.mod h1:3gIWnptl+SWMyfMR2r4TXXd0xsQZ1m50AKrwmcUONSg=
-github.com/kaptinlin/messageformat-go v0.4.18 h1:RBlHVWgZyoxTcUgGWBsl2AcyScq/urqbLZvzgryTmSI=
-github.com/kaptinlin/messageformat-go v0.4.18/go.mod h1:ntI3154RnqJgr7GaC+vZBnIExl2V3sv9selvRNNEM24=
+github.com/kaptinlin/jsonschema v0.7.7 h1:41BlQJ9dskH0oE5DSzBUrl/w4JQYIr6N6L0B5GNyDoM=
+github.com/kaptinlin/jsonschema v0.7.7/go.mod h1:rKjWfyySHSxAD7Li2ctYkPlOu960igoKBvZ2ADRtd5Q=
+github.com/kaptinlin/messageformat-go v0.4.19 h1:A5kuuZ1ybXDQ7kD1aoEWGAOemX7hLsMY0yolgSbgpRI=
+github.com/kaptinlin/messageformat-go v0.4.19/go.mod h1:utSDTfiXTxl66OC5RIEuObLH7Ue3YjbA2X86SYMBYWg=
 github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
 github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
 github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
 github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
-github.com/lucasb-eyer/go-colorful v1.3.0 h1:2/yBRLdWBZKrf7gB40FoiKfAWYQ0lqNcbuQwVHXptag=
-github.com/lucasb-eyer/go-colorful v1.3.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=
-github.com/mailru/easyjson v0.9.1 h1:LbtsOm5WAswyWbvTEOqhypdPeZzHavpZx96/n553mR8=
-github.com/mailru/easyjson v0.9.1/go.mod h1:1+xMtQp2MRNVL/V1bOzuP3aP8VNwRW55fQUto+XFtTU=
-github.com/mark3labs/mcp-go v0.44.1 h1:2PKppYlT9X2fXnE8SNYQLAX4hNjfPB0oNLqQVcN6mE8=
-github.com/mark3labs/mcp-go v0.44.1/go.mod h1:YnJfOL382MIWDx1kMY+2zsRHU/q78dBg9aFb8W6Thdw=
+github.com/lucasb-eyer/go-colorful v1.4.0 h1:UtrWVfLdarDgc44HcS7pYloGHJUjHV/4FwW4TvVgFr4=
+github.com/lucasb-eyer/go-colorful v1.4.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=
+github.com/mark3labs/mcp-go v0.46.0 h1:8KRibF4wcKejbLsHxCA/QBVUr5fQ9nwz/n8lGqmaALo=
+github.com/mark3labs/mcp-go v0.46.0/go.mod h1:JKTC7R2LLVagkEWK7Kwu7DbmA6iIvnNAod6yrHiQMag=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
-github.com/mattn/go-runewidth v0.0.12/go.mod h1:RAqKPSqVFrSLVXbA8x7dzmKdmGzieGRCM46jaSJTDAk=
-github.com/mattn/go-runewidth v0.0.20 h1:WcT52H91ZUAwy8+HUkdM3THM6gXqXuLJi9O3rjcQQaQ=
-github.com/mattn/go-runewidth v0.0.20/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=
-github.com/microcosm-cc/bluemonday v1.0.27 h1:MpEUotklkwCSLeH+Qdx1VJgNqLlpY2KXwXFM08ygZfk=
-github.com/microcosm-cc/bluemonday v1.0.27/go.mod h1:jFi9vgW+H7c3V0lb6nR74Ib/DIB5OBs92Dimizgw2cA=
+github.com/mattn/go-runewidth v0.0.21 h1:jJKAZiQH+2mIinzCJIaIG9Be1+0NR+5sz/lYEEjdM8w=
+github.com/mattn/go-runewidth v0.0.21/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=
 github.com/mitchellh/hashstructure/v2 v2.0.2 h1:vGKWl0YJqUNxE8d+h8f6NJLcCJrgbhC4NcD46KavDd4=
 github.com/mitchellh/hashstructure/v2 v2.0.2/go.mod h1:MG3aRVU/N29oo/V/IhBX8GR/zz4kQkprJgF2EVszyDE=
 github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=
@@ -224,24 +217,18 @@ github.com/muesli/mango-cobra v1.3.0 h1:vQy5GvPg3ndOSpduxutqFoINhWk3vD5K2dXo5E8p
 github.com/muesli/mango-cobra v1.3.0/go.mod h1:Cj1ZrBu3806Qw7UjxnAUgE+7tllUBj1NCLQDwwGx19E=
 github.com/muesli/mango-pflag v0.2.0 h1:QViokgKDZQCzKhYe1zH8D+UlPJzBSGoP9yx0hBG0t5k=
 github.com/muesli/mango-pflag v0.2.0/go.mod h1:X9LT1p/pbGA1wjvEbtwnixujKErkP0jVmrxwrw3fL0Y=
-github.com/muesli/reflow v0.3.0 h1:IFsN6K9NfGtjeggFP+68I4chLZV2yIKsXJFNZ+eWh6s=
-github.com/muesli/reflow v0.3.0/go.mod h1:pbwTDkVPibjO2kyvBQRBxTWEEGDGq0FlB1BIKtnHY/8=
 github.com/muesli/roff v0.1.0 h1:YD0lalCotmYuF5HhZliKWlIx7IEhiXeSfq7hNjFqGF8=
 github.com/muesli/roff v0.1.0/go.mod h1:pjAHQM9hdUUwm/krAfrLGgJkXJ+YuhtsfZ42kieB2Ig=
 github.com/muesli/termenv v0.16.0 h1:S5AlUN9dENB57rsbnkPyfdGuWIlkmzJjbFf0Tf5FWUc=
 github.com/muesli/termenv v0.16.0/go.mod h1:ZRfOIKPFDYQoDFF4Olj7/QJbW60Ol/kL1pU3VfY/Cnk=
-github.com/openai/openai-go/v2 v2.7.1 h1:/tfvTJhfv7hTSL8mWwc5VL4WLLSDL5yn9VqVykdu9r8=
-github.com/openai/openai-go/v2 v2.7.1/go.mod h1:jrJs23apqJKKbT+pqtFgNKpRju/KP9zpUTZhz3GElQE=
-github.com/pelletier/go-toml/v2 v2.2.4 h1:mye9XuhQ6gvn5h28+VilKrrPoQVanw5PMw/TB0t5Ec4=
-github.com/pelletier/go-toml/v2 v2.2.4/go.mod h1:2gIqNv+qfxSVS7cM2xJQKtLSTLUE9V8t9Stt+h56mCY=
+github.com/pelletier/go-toml/v2 v2.3.0 h1:k59bC/lIZREW0/iVaQR8nDHxVq8OVlIzYCOJf421CaM=
+github.com/pelletier/go-toml/v2 v2.3.0/go.mod h1:2gIqNv+qfxSVS7cM2xJQKtLSTLUE9V8t9Stt+h56mCY=
 github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c h1:+mdjkGKdHQG3305AYmdv1U2eRNDiU2ErMBj1gwrq8eQ=
 github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c/go.mod h1:7rwL4CYBLnjLxUqIJNnCWiEdr3bn6IUYi15bNlnbCCU=
 github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10 h1:GFCKgmp0tecUJ0sJuv4pzYCqS9+RGSn52M3FUwPs+uo=
 github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10/go.mod h1:t/avpk3KcrXxUnYOhZhMXJlSEyie6gQbtLq5NM3loB8=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 h1:Jamvg5psRIccs7FGNTlIRMkT8wgtp5eCXdBlqhYGL6U=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
-github.com/rivo/uniseg v0.1.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc=
-github.com/rivo/uniseg v0.2.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc=
 github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
 github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
@@ -277,65 +264,63 @@ github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=
 github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28=
 github.com/traefik/yaegi v0.16.1 h1:f1De3DVJqIDKmnasUF6MwmWv1dSEEat0wcpXhD2On3E=
 github.com/traefik/yaegi v0.16.1/go.mod h1:4eVhbPb3LnD2VigQjhYbEJ69vDRFdT2HQNrXx8eEwUY=
-github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=
-github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavMF/ppJZNG9ZpyihvCd0w101no=
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=
 github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=
-github.com/yuin/goldmark v1.7.16 h1:n+CJdUxaFMiDUNnWC3dMWCIQJSkxH4uz3ZwQBkAlVNE=
-github.com/yuin/goldmark v1.7.16/go.mod h1:ip/1k0VRfGynBgxOz0yCqHrbZXhcjxyuS66Brc7iBKg=
-github.com/yuin/goldmark-emoji v1.0.6 h1:QWfF2FYaXwL74tfGOW5izeiZepUDroDJfWubQI9HTHs=
-github.com/yuin/goldmark-emoji v1.0.6/go.mod h1:ukxJDKFpdFb5x0a5HqbdlcKtebh086iJpI31LTKmWuA=
+github.com/yuin/goldmark v1.8.2 h1:kEGpgqJXdgbkhcOgBxkC0X0PmoPG1ZyoZ117rDVp4zE=
+github.com/yuin/goldmark v1.8.2/go.mod h1:ip/1k0VRfGynBgxOz0yCqHrbZXhcjxyuS66Brc7iBKg=
 go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=
 go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=
-go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.66.0 h1:w/o339tDd6Qtu3+ytwt+/jon2yjAs3Ot8Xq8pelfhSo=
-go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.66.0/go.mod h1:pdhNtM9C4H5fRdrnwO7NjxzQWhKSSxCHk/KluVqDVC0=
-go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.66.0 h1:PnV4kVnw0zOmwwFkAzCN5O07fw1YOIQor120zrh0AVo=
-go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.66.0/go.mod h1:ofAwF4uinaf8SXdVzzbL4OsxJ3VfeEg3f/F6CeF49/Y=
-go.opentelemetry.io/otel v1.41.0 h1:YlEwVsGAlCvczDILpUXpIpPSL/VPugt7zHThEMLce1c=
-go.opentelemetry.io/otel v1.41.0/go.mod h1:Yt4UwgEKeT05QbLwbyHXEwhnjxNO6D8L5PQP51/46dE=
-go.opentelemetry.io/otel/metric v1.41.0 h1:rFnDcs4gRzBcsO9tS8LCpgR0dxg4aaxWlJxCno7JlTQ=
-go.opentelemetry.io/otel/metric v1.41.0/go.mod h1:xPvCwd9pU0VN8tPZYzDZV/BMj9CM9vs00GuBjeKhJps=
-go.opentelemetry.io/otel/sdk v1.41.0 h1:YPIEXKmiAwkGl3Gu1huk1aYWwtpRLeskpV+wPisxBp8=
-go.opentelemetry.io/otel/sdk v1.41.0/go.mod h1:ahFdU0G5y8IxglBf0QBJXgSe7agzjE4GiTJ6HT9ud90=
-go.opentelemetry.io/otel/sdk/metric v1.41.0 h1:siZQIYBAUd1rlIWQT2uCxWJxcCO7q3TriaMlf08rXw8=
-go.opentelemetry.io/otel/sdk/metric v1.41.0/go.mod h1:HNBuSvT7ROaGtGI50ArdRLUnvRTRGniSUZbxiWxSO8Y=
-go.opentelemetry.io/otel/trace v1.41.0 h1:Vbk2co6bhj8L59ZJ6/xFTskY+tGAbOnCtQGVVa9TIN0=
-go.opentelemetry.io/otel/trace v1.41.0/go.mod h1:U1NU4ULCoxeDKc09yCWdWe+3QoyweJcISEVa1RBzOis=
+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0 h1:yI1/OhfEPy7J9eoa6Sj051C7n5dvpj0QX8g4sRchg04=
+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0/go.mod h1:NoUCKYWK+3ecatC4HjkRktREheMeEtrXoQxrqYFeHSc=
+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.67.0 h1:OyrsyzuttWTSur2qN/Lm0m2a8yqyIjUVBZcxFPuXq2o=
+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.67.0/go.mod h1:C2NGBr+kAB4bk3xtMXfZ94gqFDtg/GkI7e9zqGh5Beg=
+go.opentelemetry.io/otel v1.42.0 h1:lSQGzTgVR3+sgJDAU/7/ZMjN9Z+vUip7leaqBKy4sho=
+go.opentelemetry.io/otel v1.42.0/go.mod h1:lJNsdRMxCUIWuMlVJWzecSMuNjE7dOYyWlqOXWkdqCc=
+go.opentelemetry.io/otel/metric v1.42.0 h1:2jXG+3oZLNXEPfNmnpxKDeZsFI5o4J+nz6xUlaFdF/4=
+go.opentelemetry.io/otel/metric v1.42.0/go.mod h1:RlUN/7vTU7Ao/diDkEpQpnz3/92J9ko05BIwxYa2SSI=
+go.opentelemetry.io/otel/sdk v1.42.0 h1:LyC8+jqk6UJwdrI/8VydAq/hvkFKNHZVIWuslJXYsDo=
+go.opentelemetry.io/otel/sdk v1.42.0/go.mod h1:rGHCAxd9DAph0joO4W6OPwxjNTYWghRWmkHuGbayMts=
+go.opentelemetry.io/otel/sdk/metric v1.42.0 h1:D/1QR46Clz6ajyZ3G8SgNlTJKBdGp84q9RKCAZ3YGuA=
+go.opentelemetry.io/otel/sdk/metric v1.42.0/go.mod h1:Ua6AAlDKdZ7tdvaQKfSmnFTdHx37+J4ba8MwVCYM5hc=
+go.opentelemetry.io/otel/trace v1.42.0 h1:OUCgIPt+mzOnaUTpOQcBiM/PLQ/Op7oq6g4LenLmOYY=
+go.opentelemetry.io/otel/trace v1.42.0/go.mod h1:f3K9S+IFqnumBkKhRJMeaZeNk9epyhnCmQh/EysQCdc=
 go.yaml.in/yaml/v3 v3.0.4 h1:tfq32ie2Jv2UxXFdLJdh3jXuOzWiL1fo0bu/FbuKpbc=
 go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=
-golang.org/x/crypto v0.48.0 h1:/VRzVqiRSggnhY7gNRxPauEQ5Drw9haKdM0jqfcCFts=
-golang.org/x/crypto v0.48.0/go.mod h1:r0kV5h3qnFPlQnBSrULhlsRfryS2pmewsg+XfMgkVos=
-golang.org/x/exp v0.0.0-20260218203240-3dfff04db8fa h1:Zt3DZoOFFYkKhDT3v7Lm9FDMEV06GpzjG2jrqW+QTE0=
-golang.org/x/exp v0.0.0-20260218203240-3dfff04db8fa/go.mod h1:K79w1Vqn7PoiZn+TkNpx3BUWUQksGO3JcVX6qIjytmA=
-golang.org/x/net v0.51.0 h1:94R/GTO7mt3/4wIKpcR5gkGmRLOuE/2hNGeWq/GBIFo=
-golang.org/x/net v0.51.0/go.mod h1:aamm+2QF5ogm02fjy5Bb7CQ0WMt1/WVM7FtyaTLlA9Y=
-golang.org/x/oauth2 v0.35.0 h1:Mv2mzuHuZuY2+bkyWXIHMfhNdJAdwW3FuWeCPYN5GVQ=
-golang.org/x/oauth2 v0.35.0/go.mod h1:lzm5WQJQwKZ3nwavOZ3IS5Aulzxi68dUSgRHujetwEA=
-golang.org/x/sync v0.19.0 h1:vV+1eWNmZ5geRlYjzm2adRgW2/mcpevXNg50YZtPCE4=
-golang.org/x/sync v0.19.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/crypto v0.49.0 h1:+Ng2ULVvLHnJ/ZFEq4KdcDd/cfjrrjjNSXNzxg0Y4U4=
+golang.org/x/crypto v0.49.0/go.mod h1:ErX4dUh2UM+CFYiXZRTcMpEcN8b/1gxEuv3nODoYtCA=
+golang.org/x/exp v0.0.0-20260312153236-7ab1446f8b90 h1:jiDhWWeC7jfWqR9c/uplMOqJ0sbNlNWv0UkzE0vX1MA=
+golang.org/x/exp v0.0.0-20260312153236-7ab1446f8b90/go.mod h1:xE1HEv6b+1SCZ5/uscMRjUBKtIxworgEcEi+/n9NQDQ=
+golang.org/x/net v0.52.0 h1:He/TN1l0e4mmR3QqHMT2Xab3Aj3L9qjbhRm78/6jrW0=
+golang.org/x/net v0.52.0/go.mod h1:R1MAz7uMZxVMualyPXb+VaqGSa3LIaUqk0eEt3w36Sw=
+golang.org/x/oauth2 v0.36.0 h1:peZ/1z27fi9hUOFCAZaHyrpWG5lwe0RJEEEeH0ThlIs=
+golang.org/x/oauth2 v0.36.0/go.mod h1:YDBUJMTkDnJS+A4BP4eZBjCqtokkg1hODuPjwiGPO7Q=
+golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
+golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.41.0 h1:Ivj+2Cp/ylzLiEU89QhWblYnOE9zerudt9Ftecq2C6k=
-golang.org/x/sys v0.41.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
 golang.org/x/sys v0.42.0 h1:omrd2nAlyT5ESRdCLYdm3+fMfNFE/+Rf4bDIQImRJeo=
 golang.org/x/sys v0.42.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
-golang.org/x/term v0.40.0 h1:36e4zGLqU4yhjlmxEaagx2KuYbJq3EwY8K943ZsHcvg=
-golang.org/x/term v0.40.0/go.mod h1:w2P8uVp06p2iyKKuvXIm7N/y0UCRt3UfJTfZ7oOpglM=
-golang.org/x/text v0.34.0 h1:oL/Qq0Kdaqxa1KbNeMKwQq0reLCCaFtqu2eNuSeNHbk=
-golang.org/x/text v0.34.0/go.mod h1:homfLqTYRFyVYemLBFl5GgL/DWEiH5wcsQ5gSh1yziA=
-golang.org/x/time v0.14.0 h1:MRx4UaLrDotUKUdCIqzPC48t1Y9hANFKIRpNx+Te8PI=
-golang.org/x/time v0.14.0/go.mod h1:eL/Oa2bBBK0TkX57Fyni+NgnyQQN4LitPmob2Hjnqw4=
+golang.org/x/term v0.41.0 h1:QCgPso/Q3RTJx2Th4bDLqML4W6iJiaXFq2/ftQF13YU=
+golang.org/x/term v0.41.0/go.mod h1:3pfBgksrReYfZ5lvYM0kSO0LIkAl4Yl2bXOkKP7Ec2A=
+golang.org/x/text v0.35.0 h1:JOVx6vVDFokkpaq1AEptVzLTpDe9KGpj5tR4/X+ybL8=
+golang.org/x/text v0.35.0/go.mod h1:khi/HExzZJ2pGnjenulevKNX1W67CUy0AsXcNubPGCA=
+golang.org/x/time v0.15.0 h1:bbrp8t3bGUeFOx08pvsMYRTCVSMk89u4tKbNOZbp88U=
+golang.org/x/time v0.15.0/go.mod h1:Y4YMaQmXwGQZoFaVFk4YpCt4FLQMYKZe9oeV/f4MSno=
 gonum.org/v1/gonum v0.16.0 h1:5+ul4Swaf3ESvrOnidPp4GZbzf0mxVQpDCYUQE7OJfk=
 gonum.org/v1/gonum v0.16.0/go.mod h1:fef3am4MQ93R2HHpKnLk4/Tbh/s0+wqD5nfa6Pnwy4E=
-google.golang.org/api v0.269.0 h1:qDrTOxKUQ/P0MveH6a7vZ+DNHxJQjtGm/uvdbdGXCQg=
-google.golang.org/api v0.269.0/go.mod h1:N8Wpcu23Tlccl0zSHEkcAZQKDLdquxK+l9r2LkwAauE=
-google.golang.org/genai v1.49.0 h1:Se+QJaH2GYK1aaR1o5S38mlU2GD5FnVvP76nfkV7LH0=
-google.golang.org/genai v1.49.0/go.mod h1:A3kkl0nyBjyFlNjgxIwKq70julKbIxpSxqKO5gw/gmk=
-google.golang.org/genproto/googleapis/rpc v0.0.0-20260226221140-a57be14db171 h1:ggcbiqK8WWh6l1dnltU4BgWGIGo+EVYxCaAPih/zQXQ=
-google.golang.org/genproto/googleapis/rpc v0.0.0-20260226221140-a57be14db171/go.mod h1:4Hqkh8ycfw05ld/3BWL7rJOSfebL2Q+DVDeRgYgxUU8=
-google.golang.org/grpc v1.79.2 h1:fRMD94s2tITpyJGtBBn7MkMseNpOZU8ZxgC3MMBaXRU=
-google.golang.org/grpc v1.79.2/go.mod h1:KmT0Kjez+0dde/v2j9vzwoAScgEPx/Bw1CYChhHLrHQ=
+google.golang.org/api v0.273.0 h1:r/Bcv36Xa/te1ugaN1kdJ5LoA5Wj/cL+a4gj6FiPBjQ=
+google.golang.org/api v0.273.0/go.mod h1:JbAt7mF+XVmWu6xNP8/+CTiGH30ofmCmk9nM8d8fHew=
+google.golang.org/genai v1.52.0 h1:ekVIxWHtLUNbt+v0WWi4j3JT4yrHDEbysMcHQcaCQoI=
+google.golang.org/genai v1.52.0/go.mod h1:A3kkl0nyBjyFlNjgxIwKq70julKbIxpSxqKO5gw/gmk=
+google.golang.org/genproto v0.0.0-20260319201613-d00831a3d3e7 h1:XzmzkmB14QhVhgnawEVsOn6OFsnpyxNPRY9QV01dNB0=
+google.golang.org/genproto v0.0.0-20260319201613-d00831a3d3e7/go.mod h1:L43LFes82YgSonw6iTXTxXUX1OlULt4AQtkik4ULL/I=
+google.golang.org/genproto/googleapis/api v0.0.0-20260319201613-d00831a3d3e7 h1:41r6JMbpzBMen0R/4TZeeAmGXSJC7DftGINUodzTkPI=
+google.golang.org/genproto/googleapis/api v0.0.0-20260319201613-d00831a3d3e7/go.mod h1:EIQZ5bFCfRQDV4MhRle7+OgjNtZ6P1PiZBgAKuxXu/Y=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260319201613-d00831a3d3e7 h1:ndE4FoJqsIceKP2oYSnUZqhTdYufCYYkqwtFzfrhI7w=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260319201613-d00831a3d3e7/go.mod h1:4Hqkh8ycfw05ld/3BWL7rJOSfebL2Q+DVDeRgYgxUU8=
+google.golang.org/grpc v1.79.3 h1:sybAEdRIEtvcD68Gx7dmnwjZKlyfuc61Dyo9pGXXkKE=
+google.golang.org/grpc v1.79.3/go.mod h1:KmT0Kjez+0dde/v2j9vzwoAScgEPx/Bw1CYChhHLrHQ=
 google.golang.org/protobuf v1.36.11 h1:fV6ZwhNocDyBLK0dj+fg8ektcVegBBuEolpbTQyBNVE=
 google.golang.org/protobuf v1.36.11/go.mod h1:HTf+CrKn2C3g5S8VImy6tdcUvCska2kB7j23XfzDpco=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

internal/acpserver/session.go

@@ -62,8 +62,8 @@ func (r *sessionRegistry) create(ctx context.Context, cwd string) (*acpSession,
 	// work in ACP mode. TUI-dependent features (widgets, prompts, editor)
 	// become no-ops or return cancelled; all data/model/tool APIs work
 	// identically to interactive mode.
-	if kitInstance.HasExtensions() {
-		kitInstance.SetExtensionContext(extensions.Context{
+	if kitInstance.Extensions().HasExtensions() {
+		kitInstance.Extensions().SetContext(extensions.Context{
 			SessionID:   sessionID,
 			CWD:         cwd,
 			Model:       kitInstance.GetModelString(),
@@ -121,31 +121,31 @@ func (r *sessionRegistry) create(ctx context.Context, cwd string) (*acpSession,
 					MessageCount:    s.MessageCount,
 				}
 			},
-			GetMessages:    func() []extensions.SessionMessage { return kitInstance.GetSessionMessages() },
-			GetSessionPath: func() string { return kitInstance.GetSessionFilePath() },
+			GetMessages:    func() []extensions.SessionMessage { return kitInstance.Extensions().GetSessionMessages() },
+			GetSessionPath: func() string { return kitInstance.GetSessionPath() },
 			AppendEntry: func(entryType, data string) (string, error) {
-				return kitInstance.AppendExtensionEntry(entryType, data)
+				return kitInstance.Extensions().AppendEntry(entryType, data)
 			},
 			GetEntries: func(entryType string) []extensions.ExtensionEntry {
-				return kitInstance.GetExtensionEntries(entryType)
+				return kitInstance.Extensions().GetEntries(entryType)
 			},
 
 			// Options, model, and tool management.
-			GetOption: func(name string) string { return kitInstance.GetExtensionOption(name) },
-			SetOption: func(name, value string) { kitInstance.SetExtensionOption(name, value) },
+			GetOption: func(name string) string { return kitInstance.Extensions().GetOption(name) },
+			SetOption: func(name, value string) { kitInstance.Extensions().SetOption(name, value) },
 			SetModel: func(modelString string) error {
-				previousModel := kitInstance.GetExtensionContext().Model
+				previousModel := kitInstance.Extensions().GetContext().Model
 				if err := kitInstance.SetModel(context.Background(), modelString); err != nil {
 					return err
 				}
-				kitInstance.UpdateExtensionContextModel(modelString)
-				kitInstance.EmitModelChange(modelString, previousModel, "extension")
+				kitInstance.Extensions().UpdateContextModel(modelString)
+				kitInstance.Extensions().EmitModelChange(modelString, previousModel, "extension")
 				return nil
 			},
 			GetAvailableModels: func() []extensions.ModelInfoEntry { return kitInstance.GetAvailableModels() },
-			EmitCustomEvent:    func(name, data string) { kitInstance.EmitExtensionCustomEvent(name, data) },
-			GetAllTools:        func() []extensions.ToolInfo { return kitInstance.GetExtensionToolInfos() },
-			SetActiveTools:     func(names []string) { kitInstance.SetExtensionActiveTools(names) },
+			EmitCustomEvent:    func(name, data string) { kitInstance.Extensions().EmitCustomEvent(name, data) },
+			GetAllTools:        func() []extensions.ToolInfo { return kitInstance.Extensions().GetToolInfos() },
+			SetActiveTools:     func(names []string) { kitInstance.Extensions().SetActiveTools(names) },
 
 			// LLM completions and subagents.
 			Complete: func(req extensions.CompleteRequest) (extensions.CompleteResponse, error) {
@@ -173,7 +173,7 @@ func (r *sessionRegistry) create(ctx context.Context, cwd string) (*acpSession,
 				}
 				extResult := &extensions.SubagentResult{
 					Response:  result.Response,
-					Error:     result.Error,
+					Error:     err,
 					SessionID: result.SessionID,
 					Elapsed:   result.Elapsed,
 				}
@@ -188,15 +188,15 @@ func (r *sessionRegistry) create(ctx context.Context, cwd string) (*acpSession,
 
 			// Render — fall back to logging.
 			RenderMessage: func(name, content string) {
-				renderer := kitInstance.GetExtensionMessageRenderer(name)
+				renderer := kitInstance.Extensions().GetMessageRenderer(name)
 				if renderer != nil && renderer.Render != nil {
 					content = renderer.Render(content, 80)
 				}
 				log.Info("extension: message", "renderer", name, "content", content)
 			},
-			ReloadExtensions: func() error { return kitInstance.ReloadExtensions() },
+			ReloadExtensions: func() error { return kitInstance.Extensions().Reload() },
 		})
-		kitInstance.EmitSessionStart()
+		kitInstance.Extensions().EmitSessionStart()
 	}
 
 	sess := &acpSession{

internal/agent/agent.go

@@ -31,7 +31,7 @@ type AgentConfig struct {
 	CoreTools []fantasy.AgentTool
 
 	// ToolWrapper is an optional function that wraps the combined tool list
-	// before it is passed to the Fantasy agent. Used by the extensions system
+	// before it is passed to the LLM agent. Used by the extensions system
 	// to intercept tool calls/results.
 	ToolWrapper func([]fantasy.AgentTool) []fantasy.AgentTool
 
@@ -63,9 +63,21 @@ type ToolCallContentHandler func(content string)
 // ReasoningDeltaHandler is a function type for handling streaming reasoning/thinking deltas.
 type ReasoningDeltaHandler func(delta string)
 
-// Agent represents an AI agent with core tool integration using the fantasy library.
+// ToolOutputHandler is a function type for handling streaming tool output chunks.
+// Used by tools like bash to stream output as it arrives rather than waiting
+// for the command to complete. The isStderr flag indicates if the chunk
+// contains stderr output.
+// Note: This is an alias for core.ToolOutputCallback to avoid import cycles.
+type ToolOutputHandler = core.ToolOutputCallback
+
+// StepUsageHandler is a function type for handling token usage after each
+// complete step in a multi-step agent turn. This enables real-time cost
+// tracking during long-running tool-calling conversations.
+type StepUsageHandler func(inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens int64)
+
+// Agent represents an AI agent with core tool integration using the LLM library.
 // Core tools (bash, read, write, edit, grep, find, ls) are registered as direct
-// fantasy.AgentTool implementations — no MCP layer, no serialization overhead.
+// AgentTool implementations — no MCP layer, no serialization overhead.
 // Additional tools from external MCP servers can be loaded alongside core tools.
 type Agent struct {
 	toolManager      *tools.MCPToolManager
@@ -88,7 +100,7 @@ type GenerateWithLoopResult struct {
 	FinalResponse *fantasy.Response
 	// ConversationMessages contains all messages in the conversation including tool calls and results
 	ConversationMessages []fantasy.Message
-	// Messages contains the conversation as custom content blocks (crush-style)
+	// Messages contains the conversation as custom content blocks
 	Messages []message.Message
 	// TotalUsage contains aggregate token usage across all steps
 	TotalUsage fantasy.Usage
@@ -100,13 +112,13 @@ type GenerateWithLoopResult struct {
 // Core tools (bash, read, write, edit, grep, find, ls) are always registered.
 // External MCP tools are loaded from the config if any MCP servers are configured.
 func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
-	// Create the LLM provider via fantasy
+	// Create the LLM provider
 	providerResult, err := models.CreateProvider(ctx, agentConfig.ModelConfig)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create model provider: %v", err)
 	}
 
-	// Register core tools (direct fantasy implementations, no MCP overhead).
+	// Register core tools (direct AgentTool implementations, no MCP overhead).
 	// Use caller-provided tools if set, otherwise default to all core tools.
 	coreTools := agentConfig.CoreTools
 	if len(coreTools) == 0 {
@@ -146,7 +158,7 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 		allTools = agentConfig.ToolWrapper(allTools)
 	}
 
-	// Build fantasy agent options
+	// Build agent options
 	var agentOpts []fantasy.AgentOption
 
 	if agentConfig.SystemPrompt != "" {
@@ -171,7 +183,8 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 
 	// Pass generation parameters when available.
 	if agentConfig.ModelConfig != nil {
-		if agentConfig.ModelConfig.MaxTokens > 0 {
+		// Skip max_output_tokens for providers that don't support it (e.g., Codex OAuth)
+		if agentConfig.ModelConfig.MaxTokens > 0 && !providerResult.SkipMaxOutputTokens {
 			agentOpts = append(agentOpts, fantasy.WithMaxOutputTokens(int64(agentConfig.ModelConfig.MaxTokens)))
 		}
 		if agentConfig.ModelConfig.Temperature != nil {
@@ -185,7 +198,7 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 		}
 	}
 
-	// Create the fantasy agent
+	// Create the agent
 	fantasyAgent := fantasy.NewAgent(providerResult.Model, agentOpts...)
 
 	// Determine provider type from model string
@@ -218,31 +231,41 @@ func (a *Agent) GenerateWithLoop(ctx context.Context, messages []fantasy.Message
 	onResponse ResponseHandler, onToolCallContent ToolCallContentHandler,
 ) (*GenerateWithLoopResult, error) {
 	return a.GenerateWithLoopAndStreaming(ctx, messages, onToolCall, onToolExecution, onToolResult,
-		onResponse, onToolCallContent, nil, nil)
+		onResponse, onToolCallContent, nil, nil, nil, nil)
 }
 
-// GenerateWithLoopAndStreaming processes messages using the fantasy agent with streaming and callbacks.
-// Fantasy handles the tool call loop internally. We map fantasy's rich callback system
+// GenerateWithLoopAndStreaming processes messages using the agent with streaming and callbacks.
+// The agent handles the tool call loop internally. We map the rich callback system
 // to kit's existing callback interface for UI integration.
 func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fantasy.Message,
 	onToolCall ToolCallHandler, onToolExecution ToolExecutionHandler, onToolResult ToolResultHandler,
 	onResponse ResponseHandler, onToolCallContent ToolCallContentHandler,
 	onStreamingResponse StreamingResponseHandler,
 	onReasoningDelta ReasoningDeltaHandler,
+	onToolOutput ToolOutputHandler,
+	onStepUsage StepUsageHandler,
 ) (*GenerateWithLoopResult, error) {
 
-	// Fantasy requires the current user input as Prompt, with prior messages as history.
+	// Inject tool output handler into context for use by core tools (e.g., bash).
+	if onToolOutput != nil {
+		ctx = core.ContextWithToolOutputCallback(ctx, onToolOutput)
+	}
+
+	// The agent requires the current user input as Prompt, with prior messages as history.
 	// Extract the last user message text and files as the prompt, and pass everything
 	// before it as Messages. Files (e.g. clipboard images) are passed via the Files
-	// field so Fantasy includes them in the API request.
+	// field so the agent includes them in the API request.
 	prompt, files, history := splitPromptAndHistory(messages)
 
-	// Track current tool call info for callbacks
-	var currentToolName string
+	// Apply message-level cache control for Anthropic models.
+	// This avoids type conflicts with provider-level options.
+	history = applyCacheControlToMessages(history)
+
+	// Track current tool call args for callbacks
 	var currentToolArgs string
 
 	// Use the streaming path when streaming is enabled OR when any callbacks are
-	// provided. Fantasy only exposes tool/step callbacks on AgentStreamCall, so
+	// provided. The agent only exposes tool/step callbacks on AgentStreamCall, so
 	// Stream is required to observe tool execution in real time. The non-streaming
 	// Generate path is reserved for the simple case with no callbacks at all.
 	hasCallbacks := onToolCall != nil || onToolExecution != nil || onToolResult != nil ||
@@ -250,13 +273,13 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 
 	if a.streamingEnabled || hasCallbacks {
 		// Track completed step messages so we can return partial results
-		// on cancellation. Fantasy's Stream() discards accumulated steps
+		// on cancellation. The agent's Stream() discards accumulated steps
 		// when it returns an error, but the OnStepFinish callback fires
 		// for every step that completed before the error occurred.
 		var completedStepMessages []fantasy.Message
 
-		// Use fantasy's streaming agent
-		result, err := a.fantasyAgent.Stream(ctx, fantasy.AgentStreamCall{
+		// Use the streaming agent
+		streamCall := fantasy.AgentStreamCall{
 			Prompt:   prompt,
 			Files:    files,
 			Messages: history,
@@ -288,7 +311,6 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				currentToolName = tc.ToolName
 				currentToolArgs = tc.Input
 
 				// Notify about the tool call
@@ -338,9 +360,63 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if text != "" && len(toolCalls) > 0 && onToolCallContent != nil {
 					onToolCallContent(text)
 				}
+				// Emit step usage for real-time cost tracking
+				if onStepUsage != nil {
+					onStepUsage(step.Usage.InputTokens, step.Usage.OutputTokens,
+						step.Usage.CacheReadTokens, step.Usage.CacheCreationTokens)
+				}
 				return nil
 			},
-		})
+		}
+
+		// If a steer channel is attached to the context, wire up a
+		// PrepareStep function that drains the channel between steps
+		// and injects pending steer messages as user messages before
+		// the next LLM call. This enables graceful mid-turn steering
+		// without cancelling in-progress tool execution.
+		if steerCh := steerChFromContext(ctx); steerCh != nil {
+			onConsumed := steerConsumedFromContext(ctx)
+			streamCall.PrepareStep = func(
+				stepCtx context.Context,
+				opts fantasy.PrepareStepFunctionOptions,
+			) (context.Context, fantasy.PrepareStepResult, error) {
+				// Drain all pending steer messages (non-blocking).
+				var steered []string
+				for {
+					select {
+					case msg := <-steerCh:
+						steered = append(steered, msg)
+					default:
+						goto done
+					}
+				}
+			done:
+				result := fantasy.PrepareStepResult{
+					Model:    opts.Model,
+					Messages: opts.Messages,
+				}
+				if len(steered) > 0 {
+					// Inject each steer message as a user message so the
+					// LLM sees the redirection on the next step.
+					for _, text := range steered {
+						result.Messages = append(result.Messages,
+							fantasy.NewUserMessage(text))
+					}
+					// Notify that steer messages were consumed.
+					if onConsumed != nil {
+						onConsumed(len(steered))
+					}
+				}
+
+				// Apply message-level cache control for Anthropic models.
+				// This avoids type conflicts with provider-level options.
+				result.Messages = applyCacheControlToMessages(result.Messages)
+
+				return stepCtx, result, nil
+			}
+		}
+
+		result, err := a.fantasyAgent.Stream(ctx, streamCall)
 		if err != nil {
 			// On cancellation (or any error), return a partial result
 			// containing messages from completed steps so the caller can
@@ -383,13 +459,11 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 		onResponse(result.Response.Content.Text())
 	}
 
-	_ = currentToolName // satisfy compiler for non-streaming path
-
 	return convertAgentResult(result, messages), nil
 }
 
 // splitPromptAndHistory extracts the last user message as the prompt string,
-// and returns everything before it as conversation history. Fantasy's agent
+// and returns everything before it as conversation history. The agent's
 // requires the current turn's input as Prompt (string), with prior messages
 // passed separately as Messages (history).
 func splitPromptAndHistory(messages []fantasy.Message) (string, []fantasy.FilePart, []fantasy.Message) {
@@ -432,8 +506,8 @@ func splitPromptAndHistory(messages []fantasy.Message) (string, []fantasy.FilePa
 	return "", nil, messages
 }
 
-// convertAgentResult converts a fantasy AgentResult to our GenerateWithLoopResult.
-// It builds both the legacy fantasy.Message slice and the new custom content blocks.
+// convertAgentResult converts an AgentResult to our GenerateWithLoopResult.
+// It builds both the message slice and the new custom content blocks.
 func convertAgentResult(result *fantasy.AgentResult, originalMessages []fantasy.Message) *GenerateWithLoopResult {
 	// Collect all conversation messages: original + all step messages
 	var allFantasyMessages []fantasy.Message
@@ -446,7 +520,7 @@ func convertAgentResult(result *fantasy.AgentResult, originalMessages []fantasy.
 	// Convert to custom content blocks
 	var allMessages []message.Message
 	for _, fm := range allFantasyMessages {
-		allMessages = append(allMessages, message.FromFantasyMessage(fm))
+		allMessages = append(allMessages, message.FromLLMMessage(fm))
 	}
 
 	return &GenerateWithLoopResult{
@@ -458,7 +532,7 @@ func convertAgentResult(result *fantasy.AgentResult, originalMessages []fantasy.
 	}
 }
 
-// extractToolResultText extracts the text and error status from a fantasy ToolResultContent.
+// extractToolResultText extracts the text and error status from a ToolResultContent.
 // For core tools, the result is already clean text (no MCP JSON wrapping).
 // For MCP tools, it unwraps the MCP content structure.
 func extractToolResultText(tr fantasy.ToolResultContent) (string, bool) {
@@ -471,7 +545,7 @@ func extractToolResultText(tr fantasy.ToolResultContent) (string, bool) {
 		return errResult.Error.Error(), true
 	}
 
-	// Get text directly from the Fantasy result type.
+	// Get text directly from the result type.
 	if textResult, ok := tr.Result.(fantasy.ToolResultOutputContentText); ok {
 		// Try to unwrap MCP JSON structure (for external MCP tools).
 		// Core tools return plain text, so this is a no-op for them.
@@ -584,7 +658,7 @@ func (a *Agent) SetModel(ctx context.Context, config *models.ProviderConfig) err
 		allTools = a.toolWrapper(allTools)
 	}
 
-	// Rebuild fantasy agent options.
+	// Rebuild agent options.
 	var agentOpts []fantasy.AgentOption
 	if a.systemPrompt != "" {
 		agentOpts = append(agentOpts, fantasy.WithSystemPrompt(a.systemPrompt))
@@ -604,7 +678,8 @@ func (a *Agent) SetModel(ctx context.Context, config *models.ProviderConfig) err
 	}
 
 	// Pass generation parameters when available.
-	if config.MaxTokens > 0 {
+	// Skip max_output_tokens for providers that don't support it (e.g., Codex OAuth)
+	if config.MaxTokens > 0 && !providerResult.SkipMaxOutputTokens {
 		agentOpts = append(agentOpts, fantasy.WithMaxOutputTokens(int64(config.MaxTokens)))
 	}
 	if config.Temperature != nil {
@@ -644,7 +719,7 @@ func (a *Agent) SetModel(ctx context.Context, config *models.ProviderConfig) err
 	return nil
 }
 
-// GetModel returns the underlying fantasy LanguageModel.
+// GetModel returns the underlying LanguageModel.
 func (a *Agent) GetModel() fantasy.LanguageModel {
 	return a.model
 }

internal/agent/cache_control.go

@@ -0,0 +1,84 @@
+package agent
+
+import (
+	"charm.land/fantasy"
+	"charm.land/fantasy/providers/anthropic"
+)
+
+// cacheControlOptions returns provider options for Anthropic cache control.
+// This is used at the message level to avoid type conflicts with provider-level options.
+func cacheControlOptions() fantasy.ProviderOptions {
+	return anthropic.NewProviderCacheControlOptions(&anthropic.ProviderCacheControlOptions{
+		CacheControl: anthropic.CacheControl{
+			Type: "ephemeral",
+		},
+	})
+}
+
+// applyCacheControlToMessages adds cache control to specific messages.
+// Anthropic allows max 4 cache blocks per request.
+// Counts existing cache blocks and only adds new ones up to the limit.
+func applyCacheControlToMessages(messages []fantasy.Message) []fantasy.Message {
+	if len(messages) == 0 {
+		return messages
+	}
+
+	// Make a copy to avoid modifying the original slice
+	result := make([]fantasy.Message, len(messages))
+	copy(result, messages)
+
+	cacheOpts := cacheControlOptions()
+	maxCacheBlocks := 4
+
+	// Helper to check if message already has cache control
+	hasCache := func(msg fantasy.Message) bool {
+		if msg.ProviderOptions == nil {
+			return false
+		}
+		if _, ok := msg.ProviderOptions["anthropic"]; ok {
+			return true
+		}
+		return false
+	}
+
+	// Count existing cache blocks
+	existingCacheCount := 0
+	for _, msg := range result {
+		if hasCache(msg) {
+			existingCacheCount++
+		}
+	}
+
+	// If we're already at or over the limit, don't add more
+	if existingCacheCount >= maxCacheBlocks {
+		return result
+	}
+
+	// How many new cache blocks can we add?
+	remaining := maxCacheBlocks - existingCacheCount
+
+	// First: find and cache the last system message (most important)
+	lastSystemIdx := -1
+	for i, msg := range result {
+		if msg.Role == fantasy.MessageRoleSystem {
+			lastSystemIdx = i
+		}
+	}
+
+	if lastSystemIdx >= 0 && remaining > 0 && !hasCache(result[lastSystemIdx]) {
+		result[lastSystemIdx].ProviderOptions = cacheOpts
+		remaining--
+	}
+
+	// Second: cache the most recent messages (up to remaining limit)
+	// Work backwards from the end to prioritize recent context
+	for i := len(result) - 1; i >= 0 && remaining > 0; i-- {
+		if hasCache(result[i]) {
+			continue
+		}
+		result[i].ProviderOptions = cacheOpts
+		remaining--
+	}
+
+	return result
+}

internal/agent/factory.go

@@ -39,7 +39,7 @@ type AgentCreationOptions struct {
 	// CoreTools overrides the default core tool set. If empty, core.AllTools()
 	// is used.
 	CoreTools []fantasy.AgentTool
-	// ToolWrapper wraps the combined tool list before Fantasy agent creation.
+	// ToolWrapper wraps the combined tool list before agent creation.
 	ToolWrapper func([]fantasy.AgentTool) []fantasy.AgentTool
 	// ExtraTools are additional tools to include (e.g. from extensions).
 	ExtraTools []fantasy.AgentTool

internal/agent/steer.go

@@ -0,0 +1,35 @@
+package agent
+
+import "context"
+
+// steerChKey is the context key for the steer channel.
+type steerChKey struct{}
+
+// steerConsumedKey is the context key for the steer-consumed callback.
+type steerConsumedKey struct{}
+
+// ContextWithSteerCh returns a new context with the steer channel attached.
+// The agent's PrepareStep function checks this channel between steps and
+// injects any pending steer messages as user messages before the next LLM call.
+func ContextWithSteerCh(ctx context.Context, ch <-chan string) context.Context {
+	return context.WithValue(ctx, steerChKey{}, ch)
+}
+
+// ContextWithSteerConsumed returns a new context with a callback that fires
+// when steer messages are consumed by PrepareStep. The count argument is the
+// number of messages injected in this batch.
+func ContextWithSteerConsumed(ctx context.Context, fn func(count int)) context.Context {
+	return context.WithValue(ctx, steerConsumedKey{}, fn)
+}
+
+// steerChFromContext extracts the steer channel from the context, or nil.
+func steerChFromContext(ctx context.Context) <-chan string {
+	ch, _ := ctx.Value(steerChKey{}).(<-chan string)
+	return ch
+}
+
+// steerConsumedFromContext extracts the steer-consumed callback, or nil.
+func steerConsumedFromContext(ctx context.Context) func(int) {
+	fn, _ := ctx.Value(steerConsumedKey{}).(func(int))
+	return fn
+}

internal/app/app.go

@@ -3,7 +3,11 @@ package app
 import (
 	"context"
 	"fmt"
+	"log"
+	"os"
 	"sync"
+	"sync/atomic"
+	"time"
 
 	tea "charm.land/bubbletea/v2"
 	"charm.land/fantasy"
@@ -65,6 +69,15 @@ type App struct {
 	// rootCtx/rootCancel are used to signal shutdown to all goroutines.
 	rootCtx    context.Context
 	rootCancel context.CancelFunc
+
+	// widgetUpdatePending is set to true when a WidgetUpdateEvent has been
+	// sent to the TUI but not yet consumed by its event loop. While the flag
+	// is set, subsequent NotifyWidgetUpdate calls are coalesced (dropped) to
+	// prevent fast extension tickers from flooding the BubbleTea mailbox with
+	// redundant re-render triggers. The flag is cleared after a short debounce
+	// (~1 frame) so new updates are always let through once the TUI has had a
+	// chance to process the pending event.
+	widgetUpdatePending atomic.Bool
 }
 
 // New creates a new App with the provided options and pre-loaded messages.
@@ -159,11 +172,57 @@ func (a *App) QueueLength() int {
 	return len(a.queue)
 }
 
-// Steer cancels the current agent step (if running), clears the queue, and
-// sends a new message that will execute as soon as the current step finishes
-// cancelling. If the agent is idle, the message executes immediately.
-// This is the "steer" delivery mode for SendMessage.
-func (a *App) Steer(prompt string) {
+// Steer injects a steering message into the currently running agent turn.
+// If the agent is in a multi-step tool loop, the message is delivered after
+// the current tool execution finishes but before the next LLM call (graceful
+// mid-turn injection via Fantasy's PrepareStep). If the agent is streaming
+// a text-only response (no pending tool calls), the message waits until the
+// response completes and then executes as the next turn.
+//
+// If the agent is idle, the message starts executing immediately (same as Run).
+//
+// Returns the number of pending steer/queue items (0 = started immediately,
+// >0 = injected/queued). The caller must update UI state based on the return
+// value — Steer does NOT send events to the program to avoid deadlocking
+// when called from within Update().
+//
+// Satisfies ui.AppController.
+func (a *App) Steer(prompt string) int {
+	a.mu.Lock()
+
+	if a.closed {
+		a.mu.Unlock()
+		return 0
+	}
+
+	if !a.busy {
+		// Not busy — start immediately, same as Run().
+		item := queueItem{Prompt: prompt}
+		a.busy = true
+		a.wg.Add(1)
+		a.mu.Unlock()
+		go a.drainQueue(item)
+		return 0
+	}
+
+	a.mu.Unlock()
+
+	// Agent is busy — inject via the SDK's steer channel. The message
+	// will be picked up by PrepareStep between agent steps (after tool
+	// execution, before next LLM call). If PrepareStep doesn't fire
+	// (text-only response), drainQueue will pick it up after the turn.
+	if a.opts.Kit != nil {
+		a.opts.Kit.InjectSteer(prompt)
+	}
+	return 1
+}
+
+// InterruptAndSend cancels the current agent step (if running), clears the
+// queue, and sends a new message that will execute as soon as the current
+// step finishes cancelling. If the agent is idle, the message executes
+// immediately. This is the hard-cancel delivery mode used by extensions'
+// CancelAndSend.
+func (a *App) InterruptAndSend(prompt string) {
 	a.mu.Lock()
 
 	if a.closed {
@@ -212,6 +271,17 @@ func (a *App) ClearMessages() {
 	}
 }
 
+// ReloadMessagesFromTree clears the in-memory message store and reloads it
+// from the tree session's current branch. Unlike ClearMessages, this does NOT
+// reset the tree session's leaf pointer. Used after Branch() to sync the
+// store with the new branch position.
+func (a *App) ReloadMessagesFromTree() {
+	a.store.Clear()
+	if a.opts.TreeSession != nil {
+		a.store.Replace(a.opts.TreeSession.GetLLMMessages())
+	}
+}
+
 // GetTreeSession returns the tree session manager, or nil if not configured.
 func (a *App) GetTreeSession() *session.TreeManager {
 	return a.opts.TreeSession
@@ -226,10 +296,14 @@ func (a *App) SwitchTreeSession(ts *session.TreeManager) {
 		_ = old.Close()
 	}
 	a.opts.TreeSession = ts
+	// Also update the kit SDK's tree session so messages are persisted correctly.
+	if a.opts.Kit != nil {
+		a.opts.Kit.SetTreeSession(ts)
+	}
 	// Reload messages from new session.
 	a.store.Clear()
 	if ts != nil {
-		a.store.Replace(ts.GetFantasyMessages())
+		a.store.Replace(ts.GetLLMMessages())
 	}
 }
 
@@ -245,7 +319,7 @@ func (a *App) AddContextMessage(text string) {
 
 	// Persist to tree session if active.
 	if ts := a.opts.TreeSession; ts != nil {
-		_, _ = ts.AppendFantasyMessage(msg)
+		_, _ = ts.AppendLLMMessage(msg)
 	}
 }
 
@@ -295,7 +369,7 @@ func (a *App) CompactConversation(customInstructions string) error {
 
 		// Sync in-memory store with the compacted session.
 		if a.opts.TreeSession != nil {
-			a.store.Replace(a.opts.TreeSession.GetFantasyMessages())
+			a.store.Replace(a.opts.TreeSession.GetLLMMessages())
 		}
 
 		a.sendEvent(CompactCompleteEvent{
@@ -401,6 +475,13 @@ func (a *App) Close() {
 
 	// Wait for background goroutines.
 	a.wg.Wait()
+
+	// Clean up empty session file on shutdown.
+	if ts := a.opts.TreeSession; ts != nil && ts.IsEmpty() {
+		if path := ts.GetFilePath(); path != "" {
+			_ = os.Remove(path)
+		}
+	}
 }
 
 // --------------------------------------------------------------------------
@@ -434,6 +515,24 @@ func (a *App) drainQueue(first queueItem) {
 		// Process all collected items as a single batch
 		a.runQueueBatch(items)
 
+		// Drain any unconsumed steer messages from the SDK channel.
+		// These arrive when the user steered during a text-only response
+		// (no tool calls, so PrepareStep didn't fire for a second step).
+		// They go to the front of the queue so they run next.
+		if a.opts.Kit != nil {
+			if leftover := a.opts.Kit.DrainSteer(); len(leftover) > 0 {
+				a.mu.Lock()
+				steerItems := make([]queueItem, len(leftover))
+				for i, text := range leftover {
+					steerItems[i] = queueItem{Prompt: text}
+				}
+				a.queue = append(steerItems, a.queue...)
+				a.mu.Unlock()
+				// Notify UI about the consumed steer messages.
+				a.sendEvent(SteerConsumedEvent{})
+			}
+		}
+
 		// Check if more items were queued while we were processing
 		a.mu.Lock()
 		hasMore := len(a.queue) > 0
@@ -486,14 +585,12 @@ func (a *App) runQueueBatch(items []queueItem) {
 	result, err := a.executeBatch(stepCtx, items, eventFn)
 	if err != nil {
 		if stepCtx.Err() != nil {
-			// Step was cancelled by the user (e.g. double-ESC). Sync
-			// the in-memory store from the tree session so that any
-			// tool calls/results that completed before cancellation
-			// are preserved in the conversation history. The SDK's
-			// runTurn already persisted partial progress to the tree
-			// session; we just need to reload it here.
+			// Step was cancelled by the user (double-ESC). The SDK
+			// preserves the user message and any completed tool
+			// call/result pairs; only the in-progress message or tool
+			// call is discarded. Sync the in-memory store to match.
 			if ts := a.opts.TreeSession; ts != nil {
-				a.store.Replace(ts.GetFantasyMessages())
+				a.store.Replace(ts.GetLLMMessages())
 			}
 			a.sendEvent(StepCancelledEvent{})
 			return
@@ -524,9 +621,10 @@ func (a *App) executeStep(ctx context.Context, prompt string, eventFn func(tea.M
 		}
 	}
 
-	// Subscribe to SDK events for TUI rendering. The subscription is
-	// temporary — it lives only for the duration of this step.
-	unsub := a.subscribeSDKEvents(sendFn)
+	// Subscribe to SDK events for TUI rendering and per-step usage updates.
+	// The subscription is temporary — it lives only for the duration of this step.
+	var sawStepUsage atomic.Bool
+	unsub := a.subscribeSDKEvents(sendFn, &sawStepUsage)
 	defer unsub()
 
 	// Show spinner while the agent works.
@@ -546,8 +644,9 @@ func (a *App) executeStep(ctx context.Context, prompt string, eventFn func(tea.M
 	// Sync in-memory store with the SDK's authoritative conversation.
 	a.store.Replace(result.Messages)
 
-	// Update usage tracker.
-	a.updateUsageFromTurnResult(result, prompt)
+	// Update usage tracker. If per-step usage was already recorded from
+	// StepUsageEvent callbacks, avoid double-counting totals.
+	a.updateUsageFromTurnResult(result, prompt, sawStepUsage.Load())
 
 	return result, nil
 }
@@ -571,9 +670,10 @@ func (a *App) executeBatch(ctx context.Context, items []queueItem, eventFn func(
 		}
 	}
 
-	// Subscribe to SDK events for TUI rendering. The subscription is
-	// temporary — it lives only for the duration of this step.
-	unsub := a.subscribeSDKEvents(sendFn)
+	// Subscribe to SDK events for TUI rendering and per-step usage updates.
+	// The subscription is temporary — it lives only for the duration of this step.
+	var sawStepUsage atomic.Bool
+	unsub := a.subscribeSDKEvents(sendFn, &sawStepUsage)
 	defer unsub()
 
 	// Show spinner while the agent works.
@@ -606,8 +706,8 @@ func (a *App) executeBatch(ctx context.Context, items []queueItem, eventFn func(
 			messages = append(messages, item.Prompt)
 		}
 
-		// TODO: Handle file attachments in batch mode
-		// For now, files are ignored in batch mode (rare edge case)
+		// File attachments are not supported in batch mode; fall back to
+		// processing only the first item that carries files.
 		if hasFiles {
 			// If files exist, fall back to processing just the first item with files
 			for _, item := range items {
@@ -628,8 +728,10 @@ func (a *App) executeBatch(ctx context.Context, items []queueItem, eventFn func(
 	// Sync in-memory store with the SDK's authoritative conversation.
 	a.store.Replace(result.Messages)
 
-	// Update usage tracker (using last item's prompt for tracking).
-	a.updateUsageFromTurnResult(result, items[len(items)-1].Prompt)
+	// Update usage tracker (using last item's prompt for fallback estimation).
+	// If per-step usage was already recorded from StepUsageEvent callbacks,
+	// avoid double-counting totals.
+	a.updateUsageFromTurnResult(result, items[len(items)-1].Prompt, sawStepUsage.Load())
 
 	return result, nil
 }
@@ -646,9 +748,10 @@ func (a *App) sendEvent(msg tea.Msg) {
 }
 
 // subscribeSDKEvents registers temporary SDK event subscribers that convert
-// SDK events to tea.Msg events and dispatch them via sendFn. Returns an
-// unsubscribe function that removes all listeners.
-func (a *App) subscribeSDKEvents(sendFn func(tea.Msg)) func() {
+// SDK events to tea.Msg events and dispatch them via sendFn. When stepUsageSeen
+// is provided, it is set to true after any non-zero StepUsageEvent is observed.
+// Returns an unsubscribe function that removes all listeners.
+func (a *App) subscribeSDKEvents(sendFn func(tea.Msg), stepUsageSeen *atomic.Bool) func() {
 	k := a.opts.Kit
 	var unsubs []func()
 
@@ -673,6 +776,17 @@ func (a *App) subscribeSDKEvents(sendFn func(tea.Msg)) func() {
 			sendFn(StreamChunkEvent{Content: ev.Chunk})
 		case kit.ReasoningDeltaEvent:
 			sendFn(ReasoningChunkEvent{Delta: ev.Delta})
+		case kit.ToolOutputEvent:
+			sendFn(ToolOutputEvent{
+				ToolCallID: ev.ToolCallID,
+				ToolName:   ev.ToolName,
+				Chunk:      ev.Chunk,
+				IsStderr:   ev.IsStderr,
+			})
+		case kit.SteerConsumedEvent:
+			sendFn(SteerConsumedEvent{})
+		case kit.StepUsageEvent:
+			a.recordStepUsage(ev, stepUsageSeen)
 		}
 	}))
 
@@ -741,12 +855,32 @@ func (a *App) NotifyModelChanged(provider, model string) {
 // NotifyWidgetUpdate sends a WidgetUpdateEvent to the TUI so it re-renders
 // extension widgets. Called from the extension context's SetWidget/RemoveWidget
 // closures. In non-interactive mode this is a no-op (widgets are TUI-only).
+//
+// Coalescing: if a WidgetUpdateEvent is already queued and not yet consumed
+// by the TUI event loop, additional calls within the same ~16 ms window are
+// dropped. This prevents fast extension tickers from flooding BubbleTea's
+// mailbox with redundant re-render triggers.
 func (a *App) NotifyWidgetUpdate() {
+	// Coalesce: only one pending update at a time.
+	if !a.widgetUpdatePending.CompareAndSwap(false, true) {
+		return
+	}
 	a.mu.Lock()
 	prog := a.program
 	a.mu.Unlock()
 	if prog != nil {
 		prog.Send(WidgetUpdateEvent{})
+		// Reset the pending flag after a short debounce so subsequent calls
+		// within the same render cycle are also coalesced, but new updates
+		// after the cycle are allowed through.
+		go func() {
+			time.Sleep(16 * time.Millisecond) // ~1 frame at 60 fps
+			a.widgetUpdatePending.Store(false)
+		}()
+	} else {
+		// No program registered (non-interactive mode); clear the flag so
+		// future calls are never permanently blocked.
+		a.widgetUpdatePending.Store(false)
 	}
 }
 
@@ -842,29 +976,106 @@ func (a *App) PrintBlockFromExtension(opts extensions.PrintBlockOpts) {
 	}
 }
 
+// recordStepUsage applies token/cost usage reported for a completed step.
+// Step usage events arrive even when a turn is later cancelled, so this keeps
+// the usage widget accurate on all stop paths.
+func (a *App) recordStepUsage(ev kit.StepUsageEvent, stepUsageSeen *atomic.Bool) {
+	hasUsage := ev.InputTokens > 0 || ev.OutputTokens > 0 || ev.CacheReadTokens > 0 || ev.CacheWriteTokens > 0
+	if a.opts.Debug {
+		log.Printf("[DEBUG] recordStepUsage: hasUsage=%v input=%d output=%d cacheRead=%d cacheWrite=%d",
+			hasUsage, ev.InputTokens, ev.OutputTokens, ev.CacheReadTokens, ev.CacheWriteTokens)
+	}
+	if !hasUsage {
+		return
+	}
+	if stepUsageSeen != nil {
+		stepUsageSeen.Store(true)
+	}
+	if a.opts.UsageTracker == nil {
+		return
+	}
+	a.opts.UsageTracker.UpdateUsage(
+		int(ev.InputTokens),
+		int(ev.OutputTokens),
+		int(ev.CacheReadTokens),
+		int(ev.CacheWriteTokens),
+	)
+	// NOTE: We do NOT call SetContextTokens here. Context fill is set once
+	// at turn completion via updateUsageFromTurnResult using FinalUsage.InputTokens,
+	// which reflects the full accumulated context. Per-step context tokens would
+	// cause the display to jump around during multi-step tool calls.
+}
+
 // updateUsageFromTurnResult records token usage from an SDK TurnResult into the
-// configured UsageTracker. This is the SDK-path equivalent of updateUsage.
-func (a *App) updateUsageFromTurnResult(result *kit.TurnResult, userPrompt string) {
+// configured UsageTracker. Called once per turn after the turn completes.
+//
+// When sawStepUsage is true, totals were already accumulated incrementally via
+// StepUsageEvent callbacks; in that case this method only updates context fill.
+// Otherwise it falls back to TotalUsage from the API response.
+//
+// NOTE: We only use ACTUAL token counts from API responses for cost tracking.
+// Estimation is never used for costs - only API-reported tokens are accurate.
+func (a *App) updateUsageFromTurnResult(result *kit.TurnResult, userPrompt string, sawStepUsage bool) {
 	if a.opts.UsageTracker == nil || result == nil {
 		return
 	}
 
-	if result.TotalUsage != nil {
-		inputTokens := int(result.TotalUsage.InputTokens)
-		outputTokens := int(result.TotalUsage.OutputTokens)
-		if inputTokens > 0 && outputTokens > 0 {
-			cacheReadTokens := int(result.TotalUsage.CacheReadTokens)
-			cacheWriteTokens := int(result.TotalUsage.CacheCreationTokens)
-			a.opts.UsageTracker.UpdateUsage(inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens)
+	// Debug logging for token tracking
+	if a.opts.Debug {
+		if result.TotalUsage != nil {
+			log.Printf("[DEBUG] updateUsageFromTurnResult TotalUsage: input=%d output=%d cacheRead=%d cacheCreate=%d",
+				result.TotalUsage.InputTokens, result.TotalUsage.OutputTokens,
+				result.TotalUsage.CacheReadTokens, result.TotalUsage.CacheCreationTokens)
 		} else {
-			a.opts.UsageTracker.EstimateAndUpdateUsage(userPrompt, result.Response)
-			return
+			log.Printf("[DEBUG] updateUsageFromTurnResult: TotalUsage=nil")
 		}
+		if result.FinalUsage != nil {
+			log.Printf("[DEBUG] updateUsageFromTurnResult FinalUsage: input=%d output=%d cacheRead=%d cacheCreate=%d",
+				result.FinalUsage.InputTokens, result.FinalUsage.OutputTokens,
+				result.FinalUsage.CacheReadTokens, result.FinalUsage.CacheCreationTokens)
+		} else {
+			log.Printf("[DEBUG] updateUsageFromTurnResult: FinalUsage=nil")
+		}
+		log.Printf("[DEBUG] updateUsageFromTurnResult: sawStepUsage=%v", sawStepUsage)
 	}
 
-	if result.FinalUsage != nil {
-		if ct := int(result.FinalUsage.InputTokens) + int(result.FinalUsage.OutputTokens); ct > 0 {
-			a.opts.UsageTracker.SetContextTokens(ct)
+	// --- Accumulate cost/token totals for the session ---
+	// Only use actual API-reported tokens for cost tracking.
+	// If sawStepUsage is true, totals were already updated via StepUsageEvent.
+	// Check any token field > 0 (not just InputTokens) because cached prompts
+	// can result in InputTokens=0 while OutputTokens>0 (OpenAI-compatible behavior).
+	hasTotalUsage := result.TotalUsage != nil &&
+		(result.TotalUsage.InputTokens > 0 ||
+			result.TotalUsage.OutputTokens > 0 ||
+			result.TotalUsage.CacheReadTokens > 0 ||
+			result.TotalUsage.CacheCreationTokens > 0)
+	if a.opts.Debug {
+		log.Printf("[DEBUG] updateUsageFromTurnResult: hasTotalUsage=%v", hasTotalUsage)
+	}
+	if !sawStepUsage && hasTotalUsage {
+		if a.opts.Debug {
+			log.Printf("[DEBUG] updateUsageFromTurnResult: calling UpdateUsage input=%d output=%d cacheRead=%d cacheCreate=%d",
+				result.TotalUsage.InputTokens, result.TotalUsage.OutputTokens,
+				result.TotalUsage.CacheReadTokens, result.TotalUsage.CacheCreationTokens)
 		}
+		a.opts.UsageTracker.UpdateUsage(
+			int(result.TotalUsage.InputTokens),
+			int(result.TotalUsage.OutputTokens),
+			int(result.TotalUsage.CacheReadTokens),
+			int(result.TotalUsage.CacheCreationTokens),
+		)
+	}
+
+	// --- Context window fill (drives the % bar) ---
+	// Use FinalUsage.InputTokens as the context window fill. The API's InputTokens
+	// already includes the full conversation history (system prompt + all previous
+	// messages + current user message). Adding OutputTokens would double-count since
+	// the output becomes part of the input for the next turn.
+	if result.FinalUsage != nil && result.FinalUsage.InputTokens > 0 {
+		if a.opts.Debug {
+			log.Printf("[DEBUG] updateUsageFromTurnResult: calling SetContextTokens=%d (FinalUsage.InputTokens)",
+				result.FinalUsage.InputTokens)
+		}
+		a.opts.UsageTracker.SetContextTokens(int(result.FinalUsage.InputTokens))
 	}
 }

internal/app/app_test.go

@@ -7,6 +7,8 @@ import (
 	"testing"
 	"time"
 
+	"charm.land/fantasy"
+
 	kit "github.com/mark3labs/kit/pkg/kit"
 )
 
@@ -14,6 +16,47 @@ import (
 // Helpers
 // --------------------------------------------------------------------------
 
+type usageUpdaterStub struct {
+	mu sync.Mutex
+
+	updateCalls   int
+	estimateCalls int
+	contextCalls  int
+
+	lastUpdateInput      int
+	lastUpdateOutput     int
+	lastUpdateCacheRead  int
+	lastUpdateCacheWrite int
+	lastContextTokens    int
+	lastEstimateInput    string
+	lastEstimateOutput   string
+}
+
+func (s *usageUpdaterStub) UpdateUsage(inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.updateCalls++
+	s.lastUpdateInput = inputTokens
+	s.lastUpdateOutput = outputTokens
+	s.lastUpdateCacheRead = cacheReadTokens
+	s.lastUpdateCacheWrite = cacheWriteTokens
+}
+
+func (s *usageUpdaterStub) EstimateAndUpdateUsage(inputText, outputText string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.estimateCalls++
+	s.lastEstimateInput = inputText
+	s.lastEstimateOutput = outputText
+}
+
+func (s *usageUpdaterStub) SetContextTokens(tokens int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.contextCalls++
+	s.lastContextTokens = tokens
+}
+
 // turnResult builds a minimal TurnResult with response text t.
 func turnResult(t string) *kit.TurnResult {
 	return &kit.TurnResult{Response: t}
@@ -489,3 +532,133 @@ func TestQueueLength_reflects(t *testing.T) {
 		t.Fatalf("expected 3, got %d", got)
 	}
 }
+
+// TestRecordStepUsage_updatesTracker verifies that per-step usage updates are
+// recorded immediately for cost tracking. Context tokens are NOT updated here
+// (only via updateUsageFromTurnResult) to avoid display jumps during multi-step
+// tool calls.
+func TestRecordStepUsage_updatesTracker(t *testing.T) {
+	usage := &usageUpdaterStub{}
+	app := New(Options{UsageTracker: usage}, nil)
+	defer app.Close()
+
+	app.recordStepUsage(kit.StepUsageEvent{
+		InputTokens:      120,
+		OutputTokens:     45,
+		CacheReadTokens:  5,
+		CacheWriteTokens: 2,
+	}, nil)
+
+	usage.mu.Lock()
+	defer usage.mu.Unlock()
+
+	if usage.updateCalls != 1 {
+		t.Fatalf("expected 1 update call, got %d", usage.updateCalls)
+	}
+	if usage.lastUpdateInput != 120 || usage.lastUpdateOutput != 45 || usage.lastUpdateCacheRead != 5 || usage.lastUpdateCacheWrite != 2 {
+		t.Fatalf("unexpected usage update payload: in=%d out=%d cache_read=%d cache_write=%d",
+			usage.lastUpdateInput, usage.lastUpdateOutput, usage.lastUpdateCacheRead, usage.lastUpdateCacheWrite)
+	}
+	// Context tokens should NOT be updated by recordStepUsage (only by updateUsageFromTurnResult)
+	if usage.contextCalls != 0 {
+		t.Fatalf("expected 0 context token updates from recordStepUsage, got %d", usage.contextCalls)
+	}
+}
+
+// TestUpdateUsageFromTurnResult_skipsTotalsWhenStepUsageSeen ensures we avoid
+// double-counting totals once StepUsageEvent-based updates were already applied.
+func TestUpdateUsageFromTurnResult_skipsTotalsWhenStepUsageSeen(t *testing.T) {
+	usage := &usageUpdaterStub{}
+	app := New(Options{UsageTracker: usage}, nil)
+	defer app.Close()
+
+	app.updateUsageFromTurnResult(&kit.TurnResult{
+		Response: "ok",
+		TotalUsage: &fantasy.Usage{
+			InputTokens:         999,
+			OutputTokens:        111,
+			CacheReadTokens:     7,
+			CacheCreationTokens: 3,
+		},
+		FinalUsage: &fantasy.Usage{InputTokens: 456},
+	}, "prompt", true)
+
+	usage.mu.Lock()
+	defer usage.mu.Unlock()
+
+	if usage.updateCalls != 0 {
+		t.Fatalf("expected no total usage update when sawStepUsage=true, got %d", usage.updateCalls)
+	}
+	if usage.estimateCalls != 0 {
+		t.Fatalf("expected no estimate update when sawStepUsage=true, got %d", usage.estimateCalls)
+	}
+	// Context tokens should be InputTokens only (456)
+	if usage.contextCalls != 1 || usage.lastContextTokens != 456 {
+		t.Fatalf("expected final context tokens=456 (InputTokens only), got calls=%d tokens=%d", usage.contextCalls, usage.lastContextTokens)
+	}
+}
+
+// TestUpdateUsageFromTurnResult_recordsWhenInputTokensZero verifies that usage
+// is recorded when InputTokens=0 but OutputTokens>0 (OpenAI-compatible cache behavior).
+func TestUpdateUsageFromTurnResult_recordsWhenInputTokensZero(t *testing.T) {
+	usage := &usageUpdaterStub{}
+	app := New(Options{UsageTracker: usage}, nil)
+	defer app.Close()
+
+	// Simulate OpenAI-compatible behavior: all prompt tokens cached, InputTokens=0
+	app.updateUsageFromTurnResult(&kit.TurnResult{
+		Response: "ok",
+		TotalUsage: &fantasy.Usage{
+			InputTokens:         0,   // All cached - subtracted from prompt
+			OutputTokens:        150, // Actual generated tokens
+			CacheReadTokens:     500, // Cache hit
+			CacheCreationTokens: 0,
+		},
+		FinalUsage: &fantasy.Usage{InputTokens: 0, OutputTokens: 150},
+	}, "prompt", false)
+
+	usage.mu.Lock()
+	defer usage.mu.Unlock()
+
+	if usage.updateCalls != 1 {
+		t.Fatalf("expected 1 update call when InputTokens=0 but OutputTokens>0, got %d", usage.updateCalls)
+	}
+	if usage.lastUpdateInput != 0 || usage.lastUpdateOutput != 150 {
+		t.Fatalf("expected input=0 output=150, got input=%d output=%d",
+			usage.lastUpdateInput, usage.lastUpdateOutput)
+	}
+	if usage.lastUpdateCacheRead != 500 {
+		t.Fatalf("expected cache_read=500, got %d", usage.lastUpdateCacheRead)
+	}
+}
+
+// TestUpdateUsageFromTurnResult_contextTokensUsesInputOnly verifies that context
+// window fill uses InputTokens only (not input+output). The API's InputTokens
+// already includes the full conversation history; adding output would double-count.
+func TestUpdateUsageFromTurnResult_contextTokensUsesInputOnly(t *testing.T) {
+	usage := &usageUpdaterStub{}
+	app := New(Options{UsageTracker: usage}, nil)
+	defer app.Close()
+
+	app.updateUsageFromTurnResult(&kit.TurnResult{
+		Response: "ok",
+		TotalUsage: &fantasy.Usage{
+			InputTokens:  1000,
+			OutputTokens: 200,
+		},
+		FinalUsage: &fantasy.Usage{
+			InputTokens:  1000, // Full context including history
+			OutputTokens: 200,
+		},
+	}, "prompt", false)
+
+	usage.mu.Lock()
+	defer usage.mu.Unlock()
+
+	// Context tokens should be InputTokens only (1000), not input+output (1200)
+	// because InputTokens already includes the full conversation history
+	if usage.contextCalls != 1 || usage.lastContextTokens != 1000 {
+		t.Fatalf("expected context tokens=1000 (InputTokens only), got calls=%d tokens=%d",
+			usage.contextCalls, usage.lastContextTokens)
+	}
+}

internal/app/events.go

@@ -54,6 +54,19 @@ type ToolResultEvent struct {
 	IsError bool
 }
 
+// ToolOutputEvent is sent when a tool produces streaming output chunks (e.g., bash output).
+// This allows the TUI to display tool output as it arrives, before the tool completes.
+type ToolOutputEvent struct {
+	// ToolCallID is the stable identifier for the tool call producing output.
+	ToolCallID string
+	// ToolName is the name of the tool producing output.
+	ToolName string
+	// Chunk is a piece of the tool's output text.
+	Chunk string
+	// IsStderr indicates whether this chunk came from stderr.
+	IsStderr bool
+}
+
 // ToolCallContentEvent is sent when a step includes text content alongside tool calls.
 // This allows the TUI to display assistant commentary that accompanies tool usage.
 type ToolCallContentEvent struct {
@@ -128,6 +141,12 @@ type CompactErrorEvent struct {
 	Err error
 }
 
+// SteerConsumedEvent is sent when one or more steering messages have been
+// consumed — either injected mid-turn via PrepareStep, or drained into the
+// queue after a turn completes. The TUI uses this to clear the steering
+// badge from the display.
+type SteerConsumedEvent struct{}
+
 // ModelChangedEvent is sent when an extension changes the active model via
 // ctx.SetModel. The TUI updates the model name shown in the status bar and
 // message attribution.

internal/auth/credentials.go

@@ -10,9 +10,10 @@ import (
 )
 
 // CredentialStore holds all stored credentials for various providers.
-// Currently supports Anthropic credentials with both OAuth and API key authentication methods.
+// Currently supports Anthropic and OpenAI credentials with both OAuth and API key authentication methods.
 type CredentialStore struct {
 	Anthropic *AnthropicCredentials `json:"anthropic,omitempty"`
+	OpenAI    *OpenAICredentials    `json:"openai,omitempty"`
 }
 
 // AnthropicCredentials holds Anthropic API credentials supporting both OAuth
@@ -28,13 +29,44 @@ type AnthropicCredentials struct {
 	CreatedAt    time.Time `json:"created_at"`
 }
 
+// OpenAICredentials holds OpenAI API credentials supporting both OAuth
+// and API key authentication methods. The Type field indicates which authentication
+// method is being used. For OAuth, tokens are stored with expiration timestamps
+// for automatic refresh. For API keys, only the key itself is stored.
+type OpenAICredentials struct {
+	Type         string    `json:"type"`                    // "oauth" or "api_key"
+	APIKey       string    `json:"api_key,omitempty"`       // For API key auth
+	AccessToken  string    `json:"access_token,omitempty"`  // For OAuth
+	RefreshToken string    `json:"refresh_token,omitempty"` // For OAuth
+	ExpiresAt    int64     `json:"expires_at,omitempty"`    // For OAuth
+	AccountID    string    `json:"account_id,omitempty"`    // For OAuth (ChatGPT account ID)
+	CreatedAt    time.Time `json:"created_at"`
+}
+
+// oauthTokenExpired reports whether an OAuth token with the given type and
+// expiry unix timestamp is past its expiry. Returns false for API key
+// credentials or when no expiry is set.
+func oauthTokenExpired(credType string, expiresAt int64) bool {
+	if credType != "oauth" || expiresAt == 0 {
+		return false
+	}
+	return time.Now().Unix() >= expiresAt
+}
+
+// oauthTokenNeedsRefresh reports whether an OAuth token will expire within the
+// next 5 minutes, allowing proactive refresh before it becomes invalid.
+// Returns false for API key credentials or when no expiry is set.
+func oauthTokenNeedsRefresh(credType string, expiresAt int64) bool {
+	if credType != "oauth" || expiresAt == 0 {
+		return false
+	}
+	return time.Now().Unix() >= (expiresAt - 300) // 5 minutes buffer
+}
+
 // IsExpired checks if the OAuth token is expired based on the ExpiresAt timestamp.
 // Returns false for API key authentication or if no expiration is set.
 func (c *AnthropicCredentials) IsExpired() bool {
-	if c.Type != "oauth" || c.ExpiresAt == 0 {
-		return false
-	}
-	return time.Now().Unix() >= c.ExpiresAt
+	return oauthTokenExpired(c.Type, c.ExpiresAt)
 }
 
 // NeedsRefresh checks if the OAuth token needs refresh, returning true if the token
@@ -42,10 +74,21 @@ func (c *AnthropicCredentials) IsExpired() bool {
 // to avoid authentication failures during operations. Returns false for API key
 // authentication or if no expiration is set.
 func (c *AnthropicCredentials) NeedsRefresh() bool {
-	if c.Type != "oauth" || c.ExpiresAt == 0 {
-		return false
-	}
-	return time.Now().Unix() >= (c.ExpiresAt - 300) // 5 minutes buffer
+	return oauthTokenNeedsRefresh(c.Type, c.ExpiresAt)
+}
+
+// IsExpired checks if the OAuth token is expired based on the ExpiresAt timestamp.
+// Returns false for API key authentication or if no expiration is set.
+func (c *OpenAICredentials) IsExpired() bool {
+	return oauthTokenExpired(c.Type, c.ExpiresAt)
+}
+
+// NeedsRefresh checks if the OAuth token needs refresh, returning true if the token
+// will expire within the next 5 minutes. This allows for proactive token refresh
+// to avoid authentication failures during operations. Returns false for API key
+// authentication or if no expiration is set.
+func (c *OpenAICredentials) NeedsRefresh() bool {
+	return oauthTokenNeedsRefresh(c.Type, c.ExpiresAt)
 }
 
 // CredentialManager handles secure storage and retrieval of authentication credentials.
@@ -212,6 +255,142 @@ func (cm *CredentialManager) HasAnthropicCredentials() (bool, error) {
 	}
 }
 
+// SetOpenAICredentials stores OpenAI API key credentials. It validates the
+// API key format before storing. The API key must start with "sk-" and be
+// at least 20 characters long. Returns an error if the API key is invalid or
+// if storage fails.
+func (cm *CredentialManager) SetOpenAICredentials(apiKey string) error {
+	if err := validateOpenAIAPIKey(apiKey); err != nil {
+		return err
+	}
+
+	store, err := cm.LoadCredentials()
+	if err != nil {
+		return err
+	}
+
+	store.OpenAI = &OpenAICredentials{
+		Type:      "api_key",
+		APIKey:    apiKey,
+		CreatedAt: time.Now(),
+	}
+
+	return cm.SaveCredentials(store)
+}
+
+// GetOpenAICredentials retrieves stored OpenAI credentials. Returns nil if
+// no credentials are stored. The returned credentials may be either OAuth or API
+// key type, check the Type field to determine which.
+func (cm *CredentialManager) GetOpenAICredentials() (*OpenAICredentials, error) {
+	store, err := cm.LoadCredentials()
+	if err != nil {
+		return nil, err
+	}
+
+	return store.OpenAI, nil
+}
+
+// RemoveOpenAICredentials removes stored OpenAI credentials from storage.
+// If this was the only credential stored, the entire credentials file is removed.
+// Returns an error if the removal fails.
+func (cm *CredentialManager) RemoveOpenAICredentials() error {
+	store, err := cm.LoadCredentials()
+	if err != nil {
+		return err
+	}
+
+	store.OpenAI = nil
+
+	// If store is empty, remove the file entirely
+	if store.Anthropic == nil && store.OpenAI == nil {
+		if err := os.Remove(cm.credentialsPath); err != nil && !os.IsNotExist(err) {
+			return fmt.Errorf("failed to remove credentials file: %w", err)
+		}
+		return nil
+	}
+
+	return cm.SaveCredentials(store)
+}
+
+// HasOpenAICredentials checks if valid OpenAI credentials are stored.
+// Returns true if either a non-empty OAuth access token or API key is present,
+// false otherwise. Returns an error if credentials cannot be loaded.
+func (cm *CredentialManager) HasOpenAICredentials() (bool, error) {
+	creds, err := cm.GetOpenAICredentials()
+	if err != nil {
+		return false, err
+	}
+	if creds == nil {
+		return false, nil
+	}
+
+	// Check based on credential type
+	switch creds.Type {
+	case "oauth":
+		return creds.AccessToken != "", nil
+	case "api_key":
+		return creds.APIKey != "", nil
+	default:
+		return false, nil
+	}
+}
+
+// SetOpenAIOAuthCredentials stores OpenAI OAuth credentials in the credential manager's secure storage.
+// The credentials should include access token, refresh token, and expiration information.
+// Returns an error if the credentials cannot be saved.
+func (cm *CredentialManager) SetOpenAIOAuthCredentials(creds *OpenAICredentials) error {
+	store, err := cm.LoadCredentials()
+	if err != nil {
+		return err
+	}
+
+	store.OpenAI = creds
+	return cm.SaveCredentials(store)
+}
+
+// GetValidOpenAIAccessToken returns a valid access token for API requests. For OAuth credentials,
+// it automatically refreshes the token if it's expired or about to expire. For API key
+// credentials, it simply returns the API key. Returns an error if no credentials are found,
+// if token refresh fails, or if the credential type is unknown.
+func (cm *CredentialManager) GetValidOpenAIAccessToken() (string, error) {
+	creds, err := cm.GetOpenAICredentials()
+	if err != nil {
+		return "", err
+	}
+
+	if creds == nil {
+		return "", fmt.Errorf("no credentials found")
+	}
+
+	// For API key auth, return the API key
+	if creds.Type == "api_key" {
+		return creds.APIKey, nil
+	}
+
+	// For OAuth, check if token needs refresh
+	if creds.Type == "oauth" {
+		if creds.NeedsRefresh() {
+			// Refresh the token
+			client := NewOpenAIOAuthClient()
+			newCreds, err := client.RefreshToken(creds.RefreshToken)
+			if err != nil {
+				return "", fmt.Errorf("failed to refresh token: %w", err)
+			}
+
+			// Update stored credentials
+			if err := cm.SetOpenAIOAuthCredentials(newCreds); err != nil {
+				return "", fmt.Errorf("failed to save refreshed token: %w", err)
+			}
+
+			return newCreds.AccessToken, nil
+		}
+
+		return creds.AccessToken, nil
+	}
+
+	return "", fmt.Errorf("unknown credential type: %s", creds.Type)
+}
+
 // GetCredentialsPath returns the absolute path to the credentials JSON file.
 // This is useful for debugging or displaying the storage location to users.
 func (cm *CredentialManager) GetCredentialsPath() string {
@@ -238,6 +417,26 @@ func validateAnthropicAPIKey(apiKey string) error {
 	return nil
 }
 
+// validateOpenAIAPIKey validates the format of an OpenAI API key
+func validateOpenAIAPIKey(apiKey string) error {
+	apiKey = strings.TrimSpace(apiKey)
+
+	if apiKey == "" {
+		return fmt.Errorf("API key cannot be empty")
+	}
+
+	// OpenAI API keys typically start with "sk-" and are quite long
+	if !strings.HasPrefix(apiKey, "sk-") {
+		return fmt.Errorf("invalid OpenAI API key format (should start with 'sk-')")
+	}
+
+	if len(apiKey) < 20 {
+		return fmt.Errorf("API key appears to be too short")
+	}
+
+	return nil
+}
+
 // GetAnthropicAPIKey retrieves an Anthropic API key from multiple sources in priority order:
 // 1. Command-line flag value (highest priority)
 // 2. Stored credentials (OAuth or API key)

internal/auth/oauth.go

@@ -7,6 +7,7 @@ import (
 	"encoding/base64"
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"net/url"
 	"strings"
@@ -30,6 +31,7 @@ type OAuthClient struct {
 type AuthData struct {
 	URL      string
 	Verifier string
+	State    string // Optional state parameter for CSRF protection
 }
 
 // NewOAuthClient creates a new OAuth client configured for Anthropic's OAuth service.
@@ -199,6 +201,270 @@ func (c *OAuthClient) parseCodeAndState(code string) (parsedCode, parsedState st
 	return
 }
 
+// OpenAIOAuthClient handles OAuth 2.0 authentication flow with OpenAI Codex (ChatGPT Plus/Pro).
+// This uses OpenAI's auth0-based OAuth service for ChatGPT account authentication.
+type OpenAIOAuthClient struct {
+	ClientID     string
+	AuthorizeURL string
+	TokenURL     string
+	RedirectURI  string
+	Scopes       string
+}
+
+// NewOpenAIOAuthClient creates a new OAuth client configured for OpenAI Codex OAuth.
+// This uses the public client ID for CLI applications with PKCE for security.
+func NewOpenAIOAuthClient() *OpenAIOAuthClient {
+	return &OpenAIOAuthClient{
+		// Public client ID for OpenAI Codex CLI OAuth
+		ClientID:     "app_EMoamEEZ73f0CkXaXp7hrann",
+		AuthorizeURL: "https://auth.openai.com/oauth/authorize",
+		TokenURL:     "https://auth.openai.com/oauth/token",
+		RedirectURI:  "http://localhost:1455/auth/callback",
+		Scopes:       "openid profile email offline_access",
+	}
+}
+
+// GetAuthorizationURL generates a complete authorization URL for the OAuth flow with
+// PKCE parameters. Returns an AuthData structure containing the URL for user
+// authentication and the PKCE verifier for the subsequent code exchange.
+func (c *OpenAIOAuthClient) GetAuthorizationURL() (*AuthData, error) {
+	verifier, challenge, err := generatePKCE()
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate PKCE: %w", err)
+	}
+
+	// Generate random state
+	stateBytes := make([]byte, 16)
+	if _, err := rand.Read(stateBytes); err != nil {
+		return nil, fmt.Errorf("failed to generate state: %w", err)
+	}
+	state := fmt.Sprintf("%x", stateBytes)
+
+	params := url.Values{
+		"response_type":              {"code"},
+		"client_id":                  {c.ClientID},
+		"redirect_uri":               {c.RedirectURI},
+		"scope":                      {c.Scopes},
+		"code_challenge":             {challenge},
+		"code_challenge_method":      {"S256"},
+		"state":                      {state},
+		"id_token_add_organizations": {"true"},
+		"codex_cli_simplified_flow":  {"true"},
+		"originator":                 {"kit"},
+	}
+
+	authURL := fmt.Sprintf("%s?%s", c.AuthorizeURL, params.Encode())
+
+	return &AuthData{
+		URL:      authURL,
+		Verifier: verifier,
+		State:    state,
+	}, nil
+}
+
+// ExchangeCode exchanges an authorization code for access and refresh tokens.
+// The code parameter should be the authorization code received from the OAuth callback.
+// The verifier parameter must be the same PKCE verifier generated during GetAuthorizationURL.
+// Returns OpenAICredentials containing the tokens, expiration, and account ID.
+func (c *OpenAIOAuthClient) ExchangeCode(code, verifier string) (*OpenAICredentials, error) {
+	return c.exchangeAuthorizationCode(code, verifier, c.RedirectURI)
+}
+
+// exchangeAuthorizationCode performs the token exchange with the OAuth server
+func (c *OpenAIOAuthClient) exchangeAuthorizationCode(code, verifier, redirectUri string) (*OpenAICredentials, error) {
+	data := url.Values{
+		"grant_type":    {"authorization_code"},
+		"client_id":     {c.ClientID},
+		"code":          {code},
+		"code_verifier": {verifier},
+		"redirect_uri":  {redirectUri},
+	}
+
+	req, err := http.NewRequestWithContext(context.Background(), "POST", c.TokenURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+
+	client := &http.Client{Timeout: 30 * time.Second}
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to make token request: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("token exchange failed: %s", string(body))
+	}
+
+	var tokenResp struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int    `json:"expires_in"`
+		IDToken      string `json:"id_token"`
+	}
+
+	if err := json.NewDecoder(resp.Body).Decode(&tokenResp); err != nil {
+		return nil, fmt.Errorf("failed to decode token response: %w", err)
+	}
+
+	if tokenResp.AccessToken == "" || tokenResp.RefreshToken == "" {
+		return nil, fmt.Errorf("token response missing required fields")
+	}
+
+	// Extract account ID from JWT token
+	accountID := extractOpenAIAccountID(tokenResp.AccessToken)
+	if accountID == "" {
+		return nil, fmt.Errorf("failed to extract account ID from token")
+	}
+
+	return &OpenAICredentials{
+		Type:         "oauth",
+		AccessToken:  tokenResp.AccessToken,
+		RefreshToken: tokenResp.RefreshToken,
+		ExpiresAt:    time.Now().Unix() + int64(tokenResp.ExpiresIn),
+		CreatedAt:    time.Now(),
+		AccountID:    accountID,
+	}, nil
+}
+
+// RefreshToken refreshes an expired or expiring access token using a refresh token.
+// Returns new OpenAICredentials with updated access token, refresh token (may be
+// rotated), and new expiration timestamp. Returns an error if the refresh fails or
+// the refresh token is invalid.
+func (c *OpenAIOAuthClient) RefreshToken(refreshToken string) (*OpenAICredentials, error) {
+	data := url.Values{
+		"grant_type":    {"refresh_token"},
+		"refresh_token": {refreshToken},
+		"client_id":     {c.ClientID},
+	}
+
+	req, err := http.NewRequestWithContext(context.Background(), "POST", c.TokenURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+
+	client := &http.Client{Timeout: 30 * time.Second}
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to make refresh request: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("token refresh failed: %s", string(body))
+	}
+
+	var tokenResp struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int    `json:"expires_in"`
+	}
+
+	if err := json.NewDecoder(resp.Body).Decode(&tokenResp); err != nil {
+		return nil, fmt.Errorf("failed to decode refresh response: %w", err)
+	}
+
+	if tokenResp.AccessToken == "" || tokenResp.RefreshToken == "" {
+		return nil, fmt.Errorf("refresh response missing required fields")
+	}
+
+	// Extract account ID from JWT token
+	accountID := extractOpenAIAccountID(tokenResp.AccessToken)
+	if accountID == "" {
+		return nil, fmt.Errorf("failed to extract account ID from refreshed token")
+	}
+
+	return &OpenAICredentials{
+		Type:         "oauth",
+		AccessToken:  tokenResp.AccessToken,
+		RefreshToken: tokenResp.RefreshToken,
+		ExpiresAt:    time.Now().Unix() + int64(tokenResp.ExpiresIn),
+		CreatedAt:    time.Now(),
+		AccountID:    accountID,
+	}, nil
+}
+
+// extractOpenAIAccountID extracts the ChatGPT account ID from a JWT access token.
+// The account ID is stored in the claim path https://api.openai.com/auth.chatgpt_account_id
+func extractOpenAIAccountID(token string) string {
+	// JWT tokens are base64-encoded JSON payloads
+	parts := strings.Split(token, ".")
+	if len(parts) != 3 {
+		return ""
+	}
+
+	// Decode payload (second part)
+	payload := parts[1]
+	// Add padding if needed
+	if len(payload)%4 != 0 {
+		payload += strings.Repeat("=", 4-len(payload)%4)
+	}
+
+	decoded, err := base64.URLEncoding.DecodeString(payload)
+	if err != nil {
+		return ""
+	}
+
+	var claims map[string]any
+	if err := json.Unmarshal(decoded, &claims); err != nil {
+		return ""
+	}
+
+	// Navigate to the claim path: https://api.openai.com/auth.chatgpt_account_id
+	authPath, ok := claims["https://api.openai.com/auth"].(map[string]any)
+	if !ok {
+		return ""
+	}
+
+	accountID, ok := authPath["chatgpt_account_id"].(string)
+	if !ok {
+		return ""
+	}
+
+	return accountID
+}
+
+// ParseOpenAIAuthorizationInput parses various forms of authorization input:
+// - Full callback URL: http://localhost:1455/auth/callback?code=xxx&state=yyy
+// - Code#State format: abc123#state456
+// - Query string: code=abc123&state=state456
+// - Just the code: abc123
+func ParseOpenAIAuthorizationInput(input string) (code, state string) {
+	input = strings.TrimSpace(input)
+	if input == "" {
+		return "", ""
+	}
+
+	// Try parsing as URL
+	if strings.HasPrefix(input, "http") {
+		if u, err := url.Parse(input); err == nil {
+			return u.Query().Get("code"), u.Query().Get("state")
+		}
+	}
+
+	// Try code#state format
+	if strings.Contains(input, "#") {
+		parts := strings.SplitN(input, "#", 2)
+		return parts[0], parts[1]
+	}
+
+	// Try query string format
+	if strings.Contains(input, "code=") {
+		if values, err := url.ParseQuery(input); err == nil {
+			return values.Get("code"), values.Get("state")
+		}
+	}
+
+	// Assume it's just the code
+	return input, ""
+}
+
 // SetOAuthCredentials stores OAuth credentials in the credential manager's secure storage.
 // The credentials should include access token, refresh token, and expiration information.
 // Returns an error if the credentials cannot be saved.

internal/compaction/compaction.go

@@ -5,10 +5,18 @@
 // messages (KeepRecentTokens, default 20 000) rather than a fixed message
 // count. Auto-compaction fires when estimated context usage exceeds
 // contextWindow − ReserveTokens.
+//
+// Features modelled after pi's compaction system:
+//   - Tool result truncation (2000 char max) during serialisation
+//   - Split turn handling: when a single turn exceeds the keep budget,
+//     the turn prefix is summarised separately and merged
+//   - Cumulative file tracking: read and modified files extracted from
+//     tool calls and carried forward across compactions
 package compaction
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"strings"
 
@@ -66,10 +74,13 @@ func ShouldCompact(messages []fantasy.Message, contextWindow int, reserveTokens
 
 // CompactionResult contains statistics from a compaction operation.
 type CompactionResult struct {
-	Summary         string // LLM-generated summary of compacted messages
-	OriginalTokens  int    // Estimated token count before compaction
-	CompactedTokens int    // Estimated token count after compaction
-	MessagesRemoved int    // Number of messages replaced by the summary
+	Summary         string   // LLM-generated summary of compacted messages
+	OriginalTokens  int      // Estimated token count before compaction
+	CompactedTokens int      // Estimated token count after compaction
+	MessagesRemoved int      // Number of messages replaced by the summary
+	CutPoint        int      // Index in the original messages where the cut was made
+	ReadFiles       []string // Files read during the compacted conversation
+	ModifiedFiles   []string // Files modified during the compacted conversation
 }
 
 // CompactionOptions configures compaction behaviour. Token-based defaults
@@ -130,8 +141,34 @@ Use this EXACT format:
 - [Any data, examples, or references needed to continue]
 - [Or "(none)" if not applicable]
 
+<read-files>
+[One file path per line for files that were read during the conversation]
+</read-files>
+
+<modified-files>
+[One file path per line for files that were created, edited, or written during the conversation]
+</modified-files>
+
 Keep each section concise. Preserve exact file paths, function names, and error messages.`
 
+// ---------------------------------------------------------------------------
+// Tool result truncation
+// ---------------------------------------------------------------------------
+
+// maxToolResultChars is the maximum length of tool result text preserved
+// during serialisation. Longer results are truncated with a marker.
+const maxToolResultChars = 2000
+
+// truncateToolResult truncates text to maxToolResultChars, appending a
+// marker indicating how many characters were removed.
+func truncateToolResult(text string) string {
+	if len(text) <= maxToolResultChars {
+		return text
+	}
+	truncated := len(text) - maxToolResultChars
+	return text[:maxToolResultChars] + fmt.Sprintf("\n[...%d chars truncated]", truncated)
+}
+
 // ---------------------------------------------------------------------------
 // Cut point (token-based)
 // ---------------------------------------------------------------------------
@@ -143,11 +180,26 @@ func isValidCutPoint(msg fantasy.Message) bool {
 	return msg.Role != fantasy.MessageRoleTool
 }
 
+// findTurnStart returns the index of the user message that starts the turn
+// containing messages[idx]. A "turn" starts with a user message and includes
+// all subsequent assistant/tool messages until the next user message.
+func findTurnStart(messages []fantasy.Message, idx int) int {
+	for i := idx; i >= 0; i-- {
+		if messages[i].Role == fantasy.MessageRoleUser {
+			return i
+		}
+	}
+	return 0
+}
+
 // FindCutPoint walks backward from the end of messages, accumulating tokens
 // until the keepRecentTokens budget is filled. Returns the index that
 // separates "old" messages (0..cutPoint-1, to be summarised) from "recent"
 // messages (cutPoint..end, to be preserved).
 //
+// The cut point prefers turn boundaries (user messages). When a single turn
+// exceeds the budget, the cut lands mid-turn (IsSplitTurn returns true).
+//
 // Returns 0 if there are fewer than 2 messages or all messages fit within
 // the keep budget.
 func FindCutPoint(messages []fantasy.Message, keepRecentTokens int) int {
@@ -193,6 +245,23 @@ func FindCutPoint(messages []fantasy.Message, keepRecentTokens int) int {
 	return 0
 }
 
+// IsSplitTurn returns true if the cut point lands in the middle of a turn
+// (i.e. the message at cutPoint is not a user message, meaning we're
+// splitting a single turn's assistant/tool messages).
+func IsSplitTurn(messages []fantasy.Message, cutPoint int) bool {
+	if cutPoint <= 0 || cutPoint >= len(messages) {
+		return false
+	}
+	// If the cut point is at a user message, it's a clean turn boundary.
+	if messages[cutPoint].Role == fantasy.MessageRoleUser {
+		return false
+	}
+	// Otherwise we're cutting mid-turn — check if the turn started before
+	// the cut point.
+	turnStart := findTurnStart(messages, cutPoint)
+	return turnStart < cutPoint
+}
+
 // forceCutPoint returns a cut point that keeps only the last non-tool
 // message, summarising everything before it. Used when the budget-based
 // FindCutPoint returns 0 but the caller wants to compact anyway (manual
@@ -207,12 +276,104 @@ func forceCutPoint(messages []fantasy.Message) int {
 	return 0
 }
 
+// ---------------------------------------------------------------------------
+// File tracking
+// ---------------------------------------------------------------------------
+
+// fileOps contains cumulative file operation tracking.
+type fileOps struct {
+	ReadFiles     map[string]bool
+	ModifiedFiles map[string]bool
+}
+
+func newFileOps() *fileOps {
+	return &fileOps{
+		ReadFiles:     make(map[string]bool),
+		ModifiedFiles: make(map[string]bool),
+	}
+}
+
+// extractFileOps scans messages for tool calls and extracts file paths.
+// It recognises the built-in Kit tools: read, write, edit, bash, grep, find, ls.
+func extractFileOps(messages []fantasy.Message) *fileOps {
+	ops := newFileOps()
+	for _, msg := range messages {
+		for _, part := range msg.Content {
+			tc, ok := part.(fantasy.ToolCallPart)
+			if !ok {
+				continue
+			}
+
+			// Parse the JSON input to extract path arguments.
+			var args map[string]any
+			if err := json.Unmarshal([]byte(tc.Input), &args); err != nil {
+				continue
+			}
+
+			path, _ := args["path"].(string)
+			if path == "" {
+				continue
+			}
+
+			switch tc.ToolName {
+			case "read", "grep", "find", "ls":
+				ops.ReadFiles[path] = true
+			case "write", "edit":
+				ops.ModifiedFiles[path] = true
+			}
+		}
+	}
+	return ops
+}
+
+// merge combines another fileOps into this one (for cumulative tracking).
+func (f *fileOps) merge(other *fileOps) {
+	if other == nil {
+		return
+	}
+	for k := range other.ReadFiles {
+		f.ReadFiles[k] = true
+	}
+	for k := range other.ModifiedFiles {
+		f.ModifiedFiles[k] = true
+	}
+}
+
+// mergeSlices adds previously tracked file lists (from a prior compaction).
+func (f *fileOps) mergeSlices(readFiles, modifiedFiles []string) {
+	for _, p := range readFiles {
+		f.ReadFiles[p] = true
+	}
+	for _, p := range modifiedFiles {
+		f.ModifiedFiles[p] = true
+	}
+}
+
+// sortedKeys returns the keys of a bool map sorted alphabetically.
+func sortedKeys(m map[string]bool) []string {
+	if len(m) == 0 {
+		return nil
+	}
+	keys := make([]string, 0, len(m))
+	for k := range m {
+		keys = append(keys, k)
+	}
+	// Simple sort — no need for sort package for small lists.
+	for i := 0; i < len(keys); i++ {
+		for j := i + 1; j < len(keys); j++ {
+			if keys[j] < keys[i] {
+				keys[i], keys[j] = keys[j], keys[i]
+			}
+		}
+	}
+	return keys
+}
+
 // ---------------------------------------------------------------------------
 // Message serialisation
 // ---------------------------------------------------------------------------
 
-// roleLabel returns a human-readable label for a fantasy message role,
-
+// roleLabel returns a human-readable label for a fantasy message role.
 func roleLabel(role fantasy.MessageRole) string {
 	switch role {
 	case fantasy.MessageRoleUser:
@@ -229,16 +390,26 @@ func roleLabel(role fantasy.MessageRole) string {
 }
 
 // serializeMessages converts a slice of fantasy messages into a plain-text
-// representation suitable for sending to the summarisation LLM. The format
-
+// representation suitable for sending to the summarisation LLM. Tool result
+// text is truncated to maxToolResultChars to keep the summarisation request
+// within reasonable token budgets.
 func serializeMessages(messages []fantasy.Message) string {
 	var sb strings.Builder
 	for _, msg := range messages {
 		sb.WriteString(roleLabel(msg.Role))
 		sb.WriteString(":\n")
 		for _, part := range msg.Content {
-			if tp, ok := part.(fantasy.TextPart); ok {
-				sb.WriteString(tp.Text)
+			switch p := part.(type) {
+			case fantasy.TextPart:
+				if msg.Role == fantasy.MessageRoleTool {
+					sb.WriteString(truncateToolResult(p.Text))
+				} else {
+					sb.WriteString(p.Text)
+				}
+			case fantasy.ToolCallPart:
+				fmt.Fprintf(&sb, "[Tool call: %s(%s)]", p.ToolName, truncateToolResult(p.Input))
+			case fantasy.ReasoningPart:
+				fmt.Fprintf(&sb, "[Thinking]: %s", truncateToolResult(p.Text))
 			}
 		}
 		sb.WriteString("\n\n")
@@ -250,6 +421,13 @@ func serializeMessages(messages []fantasy.Message) string {
 // Compact
 // ---------------------------------------------------------------------------
 
+// PreviousCompaction carries file tracking state from a prior compaction so
+// that file operations accumulate across multiple compactions.
+type PreviousCompaction struct {
+	ReadFiles     []string
+	ModifiedFiles []string
+}
+
 // Compact summarises older messages using the LLM, returning the compaction
 // result and a new message slice (summary message + preserved recent
 // messages).
@@ -261,12 +439,16 @@ func serializeMessages(messages []fantasy.Message) string {
 // customInstructions is optional text appended to the summary prompt (e.g.
 // "Focus on the API design decisions"). Pass "" to use the default prompt
 // only.
+//
+// prev carries file tracking from a previous compaction for cumulative
+// tracking. Pass nil if there is no prior compaction.
 func Compact(
 	ctx context.Context,
 	model fantasy.LanguageModel,
 	messages []fantasy.Message,
 	opts CompactionOptions,
 	customInstructions string,
+	prev *PreviousCompaction,
 ) (*CompactionResult, []fantasy.Message, error) {
 	opts.defaults()
 
@@ -289,30 +471,30 @@ func Compact(
 	recentMessages := messages[cutPoint:]
 	originalTokens := EstimateMessageTokens(messages)
 
-	// Serialise old messages to text.
-	conversationText := serializeMessages(oldMessages)
-
-	// Build the user-facing prompt: conversation text + summary instructions.
-	userPrompt := opts.SummaryPrompt
-	if userPrompt == "" {
-		userPrompt = defaultSummaryPrompt
-	}
-	if customInstructions != "" {
-		userPrompt += "\n\nAdditional instructions: " + customInstructions
+	// Extract file operations from old messages.
+	ops := extractFileOps(oldMessages)
+	// Accumulate from previous compaction if present.
+	if prev != nil {
+		ops.mergeSlices(prev.ReadFiles, prev.ModifiedFiles)
 	}
+	// Also scan recent messages for file ops (they'll be carried forward).
+	recentOps := extractFileOps(recentMessages)
+	ops.merge(recentOps)
 
-	// Create a lightweight agent (no tools) just for summarisation.
-	summaryAgent := fantasy.NewAgent(model,
-		fantasy.WithSystemPrompt(defaultSystemPrompt),
-	)
-	result, err := summaryAgent.Generate(ctx, fantasy.AgentCall{
-		Prompt: conversationText + "\n\n" + userPrompt,
-	})
+	// Handle split turns: when the cut lands mid-turn, summarise the turn
+	// prefix separately and merge with the history summary.
+	var summaryText string
+	var err error
+
+	if IsSplitTurn(messages, cutPoint) {
+		summaryText, err = compactSplitTurn(ctx, model, oldMessages, messages, cutPoint, opts, customInstructions)
+	} else {
+		summaryText, err = compactNormal(ctx, model, oldMessages, opts, customInstructions)
+	}
 	if err != nil {
-		return nil, nil, fmt.Errorf("compaction summarisation failed: %w", err)
+		return nil, nil, err
 	}
 
-	summaryText := result.Response.Content.Text()
 	if summaryText == "" {
 		return nil, nil, fmt.Errorf("compaction produced an empty summary")
 	}
@@ -338,5 +520,120 @@ func Compact(
 		OriginalTokens:  originalTokens,
 		CompactedTokens: compactedTokens,
 		MessagesRemoved: len(oldMessages),
+		CutPoint:        cutPoint,
+		ReadFiles:       sortedKeys(ops.ReadFiles),
+		ModifiedFiles:   sortedKeys(ops.ModifiedFiles),
 	}, newMessages, nil
 }
+
+// compactNormal generates a summary for a clean turn-boundary cut.
+func compactNormal(
+	ctx context.Context,
+	model fantasy.LanguageModel,
+	oldMessages []fantasy.Message,
+	opts CompactionOptions,
+	customInstructions string,
+) (string, error) {
+	conversationText := serializeMessages(oldMessages)
+	return generateSummary(ctx, model, conversationText, opts, customInstructions)
+}
+
+// compactSplitTurn handles the case where the cut point lands mid-turn.
+// It generates two summaries and merges them:
+//  1. History summary: all complete turns before the split turn
+//  2. Turn prefix summary: the early part of the split turn (from the turn's
+//     user message up to the cut point)
+//
+// The merged result preserves context from both the older history and the
+// beginning of the current long turn.
+func compactSplitTurn(
+	ctx context.Context,
+	model fantasy.LanguageModel,
+	oldMessages []fantasy.Message,
+	allMessages []fantasy.Message,
+	cutPoint int,
+	opts CompactionOptions,
+	customInstructions string,
+) (string, error) {
+	// Find where the split turn starts.
+	turnStart := findTurnStart(allMessages, cutPoint)
+
+	// Messages before the turn are the "history" portion.
+	historyMessages := oldMessages
+	if turnStart > 0 && turnStart < len(oldMessages) {
+		historyMessages = oldMessages[:turnStart]
+	}
+
+	// The turn prefix: from turnStart to cutPoint.
+	turnPrefixMessages := allMessages[turnStart:cutPoint]
+
+	var historySummary string
+	var err error
+
+	// Generate history summary if there are complete turns before the split.
+	if len(historyMessages) >= 2 {
+		historySummary, err = generateSummary(ctx, model,
+			serializeMessages(historyMessages), opts, "")
+		if err != nil {
+			return "", fmt.Errorf("split turn history summary failed: %w", err)
+		}
+	}
+
+	// Generate turn prefix summary.
+	turnPrefixText := serializeMessages(turnPrefixMessages)
+	turnPrefixPrompt := "The messages above are the BEGINNING of a long turn that was split. " +
+		"Summarize the work done so far in this turn, preserving tool call results, " +
+		"file changes, and progress. Another LLM will continue this turn."
+	if customInstructions != "" {
+		turnPrefixPrompt += "\n\nAdditional instructions: " + customInstructions
+	}
+
+	summaryAgent := fantasy.NewAgent(model,
+		fantasy.WithSystemPrompt(defaultSystemPrompt),
+	)
+	result, err := summaryAgent.Generate(ctx, fantasy.AgentCall{
+		Prompt: turnPrefixText + "\n\n" + turnPrefixPrompt,
+	})
+	if err != nil {
+		return "", fmt.Errorf("split turn prefix summary failed: %w", err)
+	}
+	turnPrefixSummary := result.Response.Content.Text()
+
+	// Merge the two summaries.
+	if historySummary != "" && turnPrefixSummary != "" {
+		return historySummary + "\n\n---\n\n## Current Turn (in progress)\n\n" + turnPrefixSummary, nil
+	}
+	if turnPrefixSummary != "" {
+		return turnPrefixSummary, nil
+	}
+	return historySummary, nil
+}
+
+// generateSummary calls the LLM to produce a structured summary.
+func generateSummary(
+	ctx context.Context,
+	model fantasy.LanguageModel,
+	conversationText string,
+	opts CompactionOptions,
+	customInstructions string,
+) (string, error) {
+	userPrompt := opts.SummaryPrompt
+	if userPrompt == "" {
+		userPrompt = defaultSummaryPrompt
+	}
+	if customInstructions != "" {
+		userPrompt += "\n\nAdditional instructions: " + customInstructions
+	}
+
+	summaryAgent := fantasy.NewAgent(model,
+		fantasy.WithSystemPrompt(defaultSystemPrompt),
+	)
+	result, err := summaryAgent.Generate(ctx, fantasy.AgentCall{
+		Prompt: conversationText + "\n\n" + userPrompt,
+	})
+	if err != nil {
+		return "", fmt.Errorf("compaction summarisation failed: %w", err)
+	}
+
+	return result.Response.Content.Text(), nil
+}

internal/compaction/compaction_test.go

@@ -243,7 +243,7 @@ func TestCompact_TooFewMessages(t *testing.T) {
 		makeTextMessageN(fantasy.MessageRoleUser, 400),
 	}
 
-	result, newMsgs, err := Compact(context.TODO(), nil, msgs, CompactionOptions{}, "")
+	result, newMsgs, err := Compact(context.TODO(), nil, msgs, CompactionOptions{}, "", nil)
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -262,7 +262,7 @@ func TestCompact_WithinBudget(t *testing.T) {
 		makeTextMessageN(fantasy.MessageRoleAssistant, 400),
 	}
 
-	result, newMsgs, err := Compact(context.TODO(), nil, msgs, CompactionOptions{}, "")
+	result, newMsgs, err := Compact(context.TODO(), nil, msgs, CompactionOptions{}, "", nil)
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -273,3 +273,169 @@ func TestCompact_WithinBudget(t *testing.T) {
 		t.Errorf("messages changed: got %d, want %d", len(newMsgs), len(msgs))
 	}
 }
+
+// ---------------------------------------------------------------------------
+// Tool result truncation
+// ---------------------------------------------------------------------------
+
+func TestTruncateToolResult(t *testing.T) {
+	// Short text — no truncation.
+	short := strings.Repeat("x", 100)
+	if got := truncateToolResult(short); got != short {
+		t.Errorf("truncated short text unexpectedly")
+	}
+
+	// Exactly at limit.
+	exact := strings.Repeat("x", maxToolResultChars)
+	if got := truncateToolResult(exact); got != exact {
+		t.Errorf("truncated text at exact limit")
+	}
+
+	// Over limit.
+	over := strings.Repeat("x", maxToolResultChars+500)
+	got := truncateToolResult(over)
+	if len(got) > maxToolResultChars+50 { // allow room for marker
+		t.Errorf("truncated text too long: %d chars", len(got))
+	}
+	if !strings.Contains(got, "500 chars truncated") {
+		t.Errorf("truncation marker missing, got: %s", got[maxToolResultChars:])
+	}
+}
+
+func TestSerializeMessages_TruncatesToolResults(t *testing.T) {
+	longResult := strings.Repeat("R", maxToolResultChars+1000)
+	msgs := []fantasy.Message{
+		makeTextMessage(fantasy.MessageRoleUser, "question"),
+		{
+			Role:    fantasy.MessageRoleTool,
+			Content: []fantasy.MessagePart{fantasy.TextPart{Text: longResult}},
+		},
+	}
+
+	serialized := serializeMessages(msgs)
+	if strings.Contains(serialized, longResult) {
+		t.Error("tool result was not truncated during serialisation")
+	}
+	if !strings.Contains(serialized, "chars truncated") {
+		t.Error("truncation marker missing in serialised output")
+	}
+}
+
+func TestSerializeMessages_PreservesNonToolText(t *testing.T) {
+	longText := strings.Repeat("T", maxToolResultChars+1000)
+	msgs := []fantasy.Message{
+		makeTextMessage(fantasy.MessageRoleUser, longText),
+	}
+
+	serialized := serializeMessages(msgs)
+	if !strings.Contains(serialized, longText) {
+		t.Error("non-tool text was unexpectedly truncated")
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Split turn detection
+// ---------------------------------------------------------------------------
+
+func TestIsSplitTurn(t *testing.T) {
+	msgs := []fantasy.Message{
+		makeTextMessageN(fantasy.MessageRoleUser, 400),      // 0: turn 1 user
+		makeTextMessageN(fantasy.MessageRoleAssistant, 400), // 1: turn 1 assistant
+		makeTextMessageN(fantasy.MessageRoleUser, 400),      // 2: turn 2 user
+		makeTextMessageN(fantasy.MessageRoleAssistant, 400), // 3: turn 2 assistant
+		makeTextMessageN(fantasy.MessageRoleTool, 400),      // 4: turn 2 tool result
+		makeTextMessageN(fantasy.MessageRoleAssistant, 400), // 5: turn 2 assistant
+	}
+
+	tests := []struct {
+		name     string
+		cutPoint int
+		want     bool
+	}{
+		{"at user message (turn boundary)", 2, false},
+		{"at assistant mid-turn", 3, true},
+		{"at assistant after tool (mid-turn)", 5, true},
+		{"at 0 (no cut)", 0, false},
+		{"beyond range", 10, false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := IsSplitTurn(msgs, tt.cutPoint)
+			if got != tt.want {
+				t.Errorf("IsSplitTurn(msgs, %d) = %v, want %v", tt.cutPoint, got, tt.want)
+			}
+		})
+	}
+}
+
+// ---------------------------------------------------------------------------
+// File operations extraction
+// ---------------------------------------------------------------------------
+
+func TestExtractFileOps(t *testing.T) {
+	// Create messages with tool calls.
+	msgs := []fantasy.Message{
+		{
+			Role: fantasy.MessageRoleAssistant,
+			Content: []fantasy.MessagePart{
+				fantasy.ToolCallPart{ToolCallID: "1", ToolName: "read", Input: `{"path":"src/main.go"}`},
+				fantasy.ToolCallPart{ToolCallID: "2", ToolName: "write", Input: `{"path":"src/out.go"}`},
+				fantasy.ToolCallPart{ToolCallID: "3", ToolName: "edit", Input: `{"path":"src/edit.go"}`},
+				fantasy.ToolCallPart{ToolCallID: "4", ToolName: "grep", Input: `{"path":"src/search"}`},
+			},
+		},
+	}
+
+	ops := extractFileOps(msgs)
+	if !ops.ReadFiles["src/main.go"] {
+		t.Error("read file not tracked: src/main.go")
+	}
+	if !ops.ReadFiles["src/search"] {
+		t.Error("grep path not tracked as read: src/search")
+	}
+	if !ops.ModifiedFiles["src/out.go"] {
+		t.Error("write file not tracked: src/out.go")
+	}
+	if !ops.ModifiedFiles["src/edit.go"] {
+		t.Error("edit file not tracked: src/edit.go")
+	}
+}
+
+func TestFileOps_MergeSlices(t *testing.T) {
+	ops := newFileOps()
+	ops.ReadFiles["a.go"] = true
+	ops.ModifiedFiles["b.go"] = true
+
+	ops.mergeSlices(
+		[]string{"c.go", "a.go"},
+		[]string{"d.go"},
+	)
+
+	if len(ops.ReadFiles) != 2 { // a.go, c.go
+		t.Errorf("ReadFiles len = %d, want 2", len(ops.ReadFiles))
+	}
+	if len(ops.ModifiedFiles) != 2 { // b.go, d.go
+		t.Errorf("ModifiedFiles len = %d, want 2", len(ops.ModifiedFiles))
+	}
+}
+
+func TestSortedKeys(t *testing.T) {
+	m := map[string]bool{"c": true, "a": true, "b": true}
+	got := sortedKeys(m)
+	want := []string{"a", "b", "c"}
+	if len(got) != len(want) {
+		t.Fatalf("sortedKeys len = %d, want %d", len(got), len(want))
+	}
+	for i, v := range got {
+		if v != want[i] {
+			t.Errorf("sortedKeys[%d] = %q, want %q", i, v, want[i])
+		}
+	}
+}
+
+func TestSortedKeys_Empty(t *testing.T) {
+	got := sortedKeys(nil)
+	if got != nil {
+		t.Errorf("sortedKeys(nil) = %v, want nil", got)
+	}
+}

internal/config/config.go

@@ -157,6 +157,32 @@ type Theme struct {
 	Markdown MarkdownThemeConfig `json:"markdown,omitzero" yaml:"markdown,omitempty"`
 }
 
+// CustomModelConfig defines a custom model that can be used with custom/custom
+// or other custom/ prefixed models. These models are loaded from the config file
+// and merged into the custom provider in the model registry.
+type CustomModelConfig struct {
+	Name        string      `json:"name" yaml:"name"`
+	Family      string      `json:"family,omitempty" yaml:"family,omitempty"`
+	Attachment  bool        `json:"attachment,omitempty" yaml:"attachment,omitempty"`
+	Reasoning   bool        `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
+	Temperature bool        `json:"temperature,omitempty" yaml:"temperature,omitempty"`
+	Knowledge   string      `json:"knowledge,omitempty" yaml:"knowledge,omitempty"`
+	Cost        CostConfig  `json:"cost" yaml:"cost"`
+	Limit       LimitConfig `json:"limit" yaml:"limit"`
+}
+
+// CostConfig defines the pricing for a custom model.
+type CostConfig struct {
+	Input  float64 `json:"input" yaml:"input"`
+	Output float64 `json:"output" yaml:"output"`
+}
+
+// LimitConfig defines context and output limits for a custom model.
+type LimitConfig struct {
+	Context int `json:"context" yaml:"context"`
+	Output  int `json:"output" yaml:"output"`
+}
+
 // Config represents the complete application configuration including MCP servers,
 // model settings, UI preferences, and API credentials. It supports both command-line
 // flags and configuration file settings.
@@ -183,6 +209,13 @@ type Config struct {
 
 	// TLS configuration
 	TLSSkipVerify bool `json:"tls-skip-verify,omitempty" yaml:"tls-skip-verify,omitempty"`
+
+	// Prompt templates configuration
+	Prompts           []string `json:"prompts,omitempty" yaml:"prompts,omitempty"`
+	NoPromptTemplates bool     `json:"no-prompt-templates,omitempty" yaml:"no-prompt-templates,omitempty"`
+
+	// Custom model definitions (under custom/ provider)
+	CustomModels map[string]CustomModelConfig `json:"customModels,omitempty" yaml:"customModels,omitempty"`
 }
 
 // GetTransportType returns the transport type for the server config, mapping
@@ -370,10 +403,9 @@ func FilepathOr[T any](key string, value *T) error {
 				if err != nil {
 					return err
 				}
-				filepath.Join(home, absPath[2:])
+				absPath = filepath.Join(home, absPath[2:])
 			}
 			if !filepath.IsAbs(absPath) {
-				// base := GetConfigPath()
 				base := configPath
 				if base == "" {
 					fmt.Fprintf(os.Stderr, "unable to build relative path to config.")

internal/core/bash.go

@@ -1,34 +1,47 @@
 package core
 
 import (
-	"bytes"
+	"bufio"
 	"context"
 	"fmt"
+	"io"
 	"os"
 	"os/exec"
+	"regexp"
 	"strings"
+	"sync"
 	"time"
 
 	"charm.land/fantasy"
 )
 
+// ToolOutputCallback is the signature for streaming tool output.
+// It receives tool call ID, tool name, output chunk, and whether it's stderr.
+type ToolOutputCallback func(toolCallID, toolName, chunk string, isStderr bool)
+
+// contextKey is a custom type for context keys to avoid collisions.
+type contextKey string
+
+const toolOutputCallbackKey contextKey = "toolOutputCallback"
+
+// ContextWithToolOutputCallback returns a new context with the tool output callback set.
+func ContextWithToolOutputCallback(ctx context.Context, callback ToolOutputCallback) context.Context {
+	return context.WithValue(ctx, toolOutputCallbackKey, callback)
+}
+
+// toolOutputCallbackFromContext retrieves the tool output callback from context.
+func toolOutputCallbackFromContext(ctx context.Context) ToolOutputCallback {
+	if cb, ok := ctx.Value(toolOutputCallbackKey).(ToolOutputCallback); ok {
+		return cb
+	}
+	return nil
+}
+
 const defaultBashTimeout = 120 * time.Second
 const maxBashTimeout = 600 * time.Second
 
-var bannedCommands = []string{
-	"alias ", "bg ", "bind ", "builtin ",
-	"caller ", "command ", "compgen ",
-	"complete ", "compopt ", "coproc ",
-	"dirs ", "disown ", "enable ",
-	"fc ", "fg ", "hash ", "help ",
-	"history ", "jobs ", "kill ",
-	"logout ", "mapfile ", "popd ",
-	"pushd ", "readonly ", "select ",
-	"set ", "shopt ", "source ",
-	"suspend ", "times ", "trap ",
-	"type ", "typeset ", "ulimit ",
-	"umask ", "unalias ", "wait ",
-}
+// bannedCmdRe matches bash builtin commands that are not allowed for security reasons.
+var bannedCmdRe = regexp.MustCompile(`^(alias|bg|bind|builtin|caller|command|compgen|complete|compopt|coproc|dirs|disown|enable|fc|fg|hash|help|history|jobs|kill|logout|mapfile|popd|pushd|readonly|select|set|shopt|source|suspend|times|trap|type|typeset|ulimit|umask|unalias|wait)\s`)
 
 type bashArgs struct {
 	Command string  `json:"command"`
@@ -70,10 +83,8 @@ func executeBash(ctx context.Context, call fantasy.ToolCall, workDir string) (fa
 	}
 
 	// Check for banned commands
-	for _, banned := range bannedCommands {
-		if strings.HasPrefix(args.Command, banned) {
-			return fantasy.NewTextErrorResponse(fmt.Sprintf("command '%s' is not allowed", args.Command)), nil
-		}
+	if bannedCmdRe.MatchString(args.Command) {
+		return fantasy.NewTextErrorResponse(fmt.Sprintf("command '%s' is not allowed", args.Command)), nil
 	}
 
 	// Determine timeout
@@ -99,32 +110,157 @@ func executeBash(ctx context.Context, call fantasy.ToolCall, workDir string) (fa
 	}
 	cmd.Env = append(os.Environ(), "SHELL="+bashPath)
 
-	var stdout, stderr bytes.Buffer
-	cmd.Stdout = &stdout
-	cmd.Stderr = &stderr
+	// Get the output callback if present (for streaming support)
+	outputCallback := toolOutputCallbackFromContext(ctx)
 
-	err = cmd.Run()
+	if outputCallback != nil {
+		// Streaming mode: use pipes to capture output as it arrives
+		return executeBashStreaming(cmdCtx, call, cmd, outputCallback)
+	}
+
+	// Non-streaming mode: collect all output at once (original behavior)
+	return executeBashBuffered(cmdCtx, call, cmd)
+}
+
+// executeBashBuffered collects all output before returning (original behavior).
+// It uses explicit pipes (not cmd.Stdout) so that cmd.WaitDelay can forcibly
+// close them when grandchild processes hold pipe handles open after the
+// direct child exits.
+func executeBashBuffered(cmdCtx context.Context, call fantasy.ToolCall, cmd *exec.Cmd) (fantasy.ToolResponse, error) {
+	stdoutPipe, err := cmd.StdoutPipe()
+	if err != nil {
+		return fantasy.NewTextErrorResponse("failed to create stdout pipe"), nil
+	}
+	stderrPipe, err := cmd.StderrPipe()
+	if err != nil {
+		return fantasy.NewTextErrorResponse("failed to create stderr pipe"), nil
+	}
+
+	if err := cmd.Start(); err != nil {
+		return fantasy.NewTextErrorResponse(fmt.Sprintf("failed to start command: %v", err)), nil
+	}
+
+	// Read pipes concurrently
+	var wg sync.WaitGroup
+	var stdout, stderr strings.Builder
+	var stdoutErr, stderrErr error
+
+	wg.Add(2)
+	go func() {
+		defer wg.Done()
+		_, stdoutErr = io.Copy(&stdout, stdoutPipe)
+	}()
+	go func() {
+		defer wg.Done()
+		_, stderrErr = io.Copy(&stderr, stderrPipe)
+	}()
+
+	// Wait for the process to exit first. cmd.WaitDelay ensures that if
+	// pipes remain open (held by grandchild processes), they'll be forcibly
+	// closed after the grace period, which unblocks the io.Copy goroutines.
+	waitErr := cmd.Wait()
+
+	// Wait for pipe readers to finish draining.
+	wg.Wait()
+
+	// Ignore pipe read errors caused by WaitDelay force-closing —
+	// we still have whatever was read before the close.
+	_ = stdoutErr
+	_ = stderrErr
+
+	exitCode := 0
+	if waitErr != nil {
+		if exitErr, ok := waitErr.(*exec.ExitError); ok {
+			exitCode = exitErr.ExitCode()
+		} else if cmdCtx.Err() == context.DeadlineExceeded {
+			return fantasy.NewTextErrorResponse("command timed out"), nil
+		}
+	}
+
+	return buildBashResponse(stdout.String(), stderr.String(), exitCode)
+}
+
+// executeBashStreaming streams output as it arrives via the callback.
+func executeBashStreaming(cmdCtx context.Context, call fantasy.ToolCall, cmd *exec.Cmd, outputCallback ToolOutputCallback) (fantasy.ToolResponse, error) {
+	stdoutPipe, err := cmd.StdoutPipe()
+	if err != nil {
+		return fantasy.NewTextErrorResponse("failed to create stdout pipe"), nil
+	}
+	stderrPipe, err := cmd.StderrPipe()
+	if err != nil {
+		return fantasy.NewTextErrorResponse("failed to create stderr pipe"), nil
+	}
+
+	// Start command execution
+	if err := cmd.Start(); err != nil {
+		return fantasy.NewTextErrorResponse(fmt.Sprintf("failed to start command: %v", err)), nil
+	}
+
+	// Stream stdout and stderr concurrently
+	var wg sync.WaitGroup
+	var mu sync.Mutex
+	var stdoutChunks, stderrChunks []string
+
+	streamOutput := func(reader io.Reader, isStderr bool) {
+		defer wg.Done()
+		scanner := bufio.NewScanner(reader)
+		// Use larger buffer for long lines
+		buf := make([]byte, 0, 64*1024)
+		scanner.Buffer(buf, 1024*1024)
+
+		for scanner.Scan() {
+			chunk := scanner.Text()
+			// Send chunk to UI
+			outputCallback(call.ID, "bash", chunk, isStderr)
+			// Collect for final result
+			mu.Lock()
+			if isStderr {
+				stderrChunks = append(stderrChunks, chunk)
+			} else {
+				stdoutChunks = append(stdoutChunks, chunk)
+			}
+			mu.Unlock()
+		}
+	}
+
+	wg.Add(2)
+	go streamOutput(stdoutPipe, false)
+	go streamOutput(stderrPipe, true)
+
+	// Wait for the process to exit. cmd.WaitDelay ensures that if pipes
+	// remain open (held by grandchild processes), they'll be forcibly closed
+	// after the grace period, which unblocks the scanners above.
+	err = cmd.Wait()
+
+	// Wait for the pipe readers to finish draining. This will complete
+	// quickly since cmd.Wait() (with WaitDelay) has already ensured
+	// the pipes are closed.
+	wg.Wait()
 
 	exitCode := 0
 	if err != nil {
 		if exitErr, ok := err.(*exec.ExitError); ok {
 			exitCode = exitErr.ExitCode()
 		} else if cmdCtx.Err() == context.DeadlineExceeded {
-			return fantasy.NewTextErrorResponse(fmt.Sprintf("command timed out after %v", timeout)), nil
+			return fantasy.NewTextErrorResponse("command timed out"), nil
 		}
 	}
 
-	// Build result
+	return buildBashResponse(strings.Join(stdoutChunks, "\n"), strings.Join(stderrChunks, "\n"), exitCode)
+}
+
+// buildBashResponse constructs the final tool response from stdout/stderr.
+func buildBashResponse(stdout, stderr string, exitCode int) (fantasy.ToolResponse, error) {
 	var result strings.Builder
-	if stdout.Len() > 0 {
-		result.WriteString(stdout.String())
+	if stdout != "" {
+		result.WriteString(stdout)
 	}
-	if stderr.Len() > 0 {
+	if stderr != "" {
 		if result.Len() > 0 {
 			result.WriteString("\n")
 		}
 		result.WriteString("STDERR:\n")
-		result.WriteString(stderr.String())
+		result.WriteString(stderr)
 	}
 	if exitCode != 0 {
 		if result.Len() > 0 {

internal/core/bash_test.go

@@ -0,0 +1,129 @@
+package core
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+	"time"
+
+	"charm.land/fantasy"
+)
+
+// helper to create a bash tool call with the given command and optional timeout.
+func bashCall(command string, timeout float64) fantasy.ToolCall {
+	args := map[string]any{"command": command}
+	if timeout > 0 {
+		args["timeout"] = timeout
+	}
+	input, _ := json.Marshal(args)
+	return fantasy.ToolCall{
+		ID:    "test-call",
+		Name:  "bash",
+		Input: string(input),
+	}
+}
+
+func TestBash_SimpleCommand(t *testing.T) {
+	resp, err := executeBash(context.Background(), bashCall("echo hello", 0), "")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if resp.IsError {
+		t.Fatalf("expected success, got error: %s", resp.Content)
+	}
+	if resp.Content != "hello\n" {
+		t.Errorf("expected 'hello\\n', got %q", resp.Content)
+	}
+}
+
+func TestBash_TimeoutKillsProcess(t *testing.T) {
+	start := time.Now()
+	resp, err := executeBash(context.Background(), bashCall("sleep 60", 2), "")
+	elapsed := time.Since(start)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !resp.IsError {
+		t.Fatal("expected error response for timed-out command")
+	}
+	if elapsed > 10*time.Second {
+		t.Errorf("command took %v, expected ~2s timeout", elapsed)
+	}
+}
+
+func TestBash_BackgroundProcessDoesNotHang(t *testing.T) {
+	// This command spawns a background sleep that would hold pipes open
+	// forever if we didn't have process group killing + WaitDelay.
+	start := time.Now()
+	resp, err := executeBash(context.Background(), bashCall("echo done; sleep 3600 &", 5), "")
+	elapsed := time.Since(start)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	// The foreground command (echo) should complete quickly
+	if elapsed > 5*time.Second {
+		t.Errorf("command took %v, should complete in <5s (background process should not block)", elapsed)
+	}
+	if resp.IsError {
+		t.Fatalf("expected success, got error: %s", resp.Content)
+	}
+}
+
+func TestBash_BackgroundProcessDoesNotHang_Streaming(t *testing.T) {
+	// Same test but in streaming mode (with output callback).
+	ctx := ContextWithToolOutputCallback(context.Background(), func(_, _, _ string, _ bool) {})
+	start := time.Now()
+	resp, err := executeBash(ctx, bashCall("echo streaming; sleep 3600 &", 5), "")
+	elapsed := time.Since(start)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if elapsed > 5*time.Second {
+		t.Errorf("streaming command took %v, should complete in <5s", elapsed)
+	}
+	if resp.IsError {
+		t.Fatalf("expected success, got error: %s", resp.Content)
+	}
+}
+
+func TestBash_ContextCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		_, _ = executeBash(ctx, bashCall("sleep 60", 0), "")
+	}()
+
+	// Cancel after a short delay
+	time.Sleep(500 * time.Millisecond)
+	cancel()
+
+	// Should return promptly after cancellation
+	select {
+	case <-done:
+		// success
+	case <-time.After(5 * time.Second):
+		t.Fatal("executeBash did not return after context cancellation")
+	}
+}
+
+func TestBash_BannedCommand(t *testing.T) {
+	resp, err := executeBash(context.Background(), bashCall("alias foo=bar", 0), "")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !resp.IsError {
+		t.Fatal("expected error for banned command")
+	}
+}
+
+func TestBash_EmptyCommand(t *testing.T) {
+	resp, err := executeBash(context.Background(), bashCall("", 0), "")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !resp.IsError {
+		t.Fatal("expected error for empty command")
+	}
+}

internal/core/edit.go

@@ -4,25 +4,55 @@ import (
 	"context"
 	"fmt"
 	"os"
+	"sort"
 	"strings"
 	"unicode"
+	"unicode/utf8"
 
 	"charm.land/fantasy"
+
+	udiff "github.com/aymanbagabas/go-udiff"
 )
 
-type editArgs struct {
-	Path    string `json:"path"`
+// Edit represents a single replacement in a multi-edit operation.
+type Edit struct {
 	OldText string `json:"old_text"`
 	NewText string `json:"new_text"`
 }
 
+// editArgs holds the arguments for the edit tool.
+// Supports both single-edit mode (old_text/new_text) and multi-edit mode (edits array).
+type editArgs struct {
+	Path    string `json:"path"`
+	OldText string `json:"old_text"` // Single-edit mode
+	NewText string `json:"new_text"` // Single-edit mode
+	Edits   []Edit `json:"edits"`    // Multi-edit mode
+}
+
+// replacement represents a normalized edit ready for processing.
+type replacement struct {
+	oldText     string // normalized old text for matching
+	newText     string // normalized new text
+	originalOld string // original old text for metadata
+	originalNew string // original new text for metadata
+	index       int    // index in the original edits array (for error messages)
+}
+
+// matchedReplacement represents a replacement with its match location.
+type matchedReplacement struct {
+	replacement
+	start          int  // start index in normalized content
+	end            int  // end index in normalized content
+	usedFuzzyMatch bool // true if fuzzy matching was used
+}
+
 // NewEditTool creates the edit core tool.
 func NewEditTool(opts ...ToolOption) fantasy.AgentTool {
 	cfg := ApplyOptions(opts)
 	return &coreTool{
 		info: fantasy.ToolInfo{
 			Name:        "edit",
-			Description: "Edit a file by replacing exact text. The old_text must match exactly (including whitespace). Use this for precise, surgical edits. Fails if old_text is not found or matches multiple locations.",
+			Description: "Edit a file by replacing exact text. Supports single edit via old_text/new_text, or multiple edits via the edits array. All edits in the array are matched against the original file content (non-incremental) and must be non-overlapping.",
 			Parameters: map[string]any{
 				"path": map[string]any{
 					"type":        "string",
@@ -30,14 +60,32 @@ func NewEditTool(opts ...ToolOption) fantasy.AgentTool {
 				},
 				"old_text": map[string]any{
 					"type":        "string",
-					"description": "Exact text to find and replace (must match exactly)",
+					"description": "Exact text to find and replace (single-edit mode). Must not be used with 'edits' array.",
 				},
 				"new_text": map[string]any{
 					"type":        "string",
-					"description": "New text to replace the old text with",
+					"description": "New text to replace the old text with (single-edit mode). Must not be used with 'edits' array.",
+				},
+				"edits": map[string]any{
+					"type":        "array",
+					"description": "Array of edits for multi-region replacement. Each edit must have unique, non-overlapping old_text. All matches are against the original file content.",
+					"items": map[string]any{
+						"type": "object",
+						"properties": map[string]any{
+							"old_text": map[string]any{
+								"type":        "string",
+								"description": "Exact text to find and replace for this edit",
+							},
+							"new_text": map[string]any{
+								"type":        "string",
+								"description": "New text for this edit",
+							},
+						},
+						"required": []string{"old_text", "new_text"},
+					},
 				},
 			},
-			Required: []string{"path", "old_text", "new_text"},
+			Required: []string{"path"},
 		},
 		handler: func(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
 			return executeEdit(ctx, call, cfg.WorkDir)
@@ -48,7 +96,7 @@ func NewEditTool(opts ...ToolOption) fantasy.AgentTool {
 func executeEdit(ctx context.Context, call fantasy.ToolCall, workDir string) (fantasy.ToolResponse, error) {
 	var args editArgs
 	if err := parseArgs(call.Input, &args); err != nil {
-		return fantasy.NewTextErrorResponse("path, old_text, and new_text parameters are required"), nil
+		return fantasy.NewTextErrorResponse("failed to parse arguments: " + err.Error()), nil
 	}
 	if args.Path == "" {
 		return fantasy.NewTextErrorResponse("path parameter is required"), nil
@@ -66,158 +114,308 @@ func executeEdit(ctx context.Context, call fantasy.ToolCall, workDir string) (fa
 
 	content := string(contentBytes)
 
-	// Normalize line endings for matching
-	normalized := strings.ReplaceAll(content, "\r\n", "\n")
-	normalizedOld := strings.ReplaceAll(args.OldText, "\r\n", "\n")
-
-	// Try exact match first
-	count := strings.Count(normalized, normalizedOld)
-
-	// If no exact match, try fuzzy matching
-	if count == 0 {
-		if idx, matchLen := fuzzyMatch(normalized, normalizedOld); idx >= 0 {
-			// Apply fuzzy match — the matched text is the original content slice
-			matchedText := normalized[idx : idx+matchLen]
-			newContent := normalized[:idx] + args.NewText + normalized[idx+matchLen:]
-			if err := os.WriteFile(absPath, []byte(newContent), 0644); err != nil {
-				return fantasy.NewTextErrorResponse(fmt.Sprintf("failed to write file: %v", err)), nil
-			}
-			diff := generateDiff(absPath, normalized, newContent, idx)
-			resp := fantasy.NewTextResponse(fmt.Sprintf("Applied edit (fuzzy match) to %s\n%s", args.Path, diff))
-			return fantasy.WithResponseMetadata(resp, editDiffMeta(absPath, matchedText, args.NewText)), nil
-		}
-		return fantasy.NewTextErrorResponse(fmt.Sprintf("old_text not found in %s", args.Path)), nil
+	// Normalize and validate input
+	replacements, err := normalizeEditInput(args)
+	if err != nil {
+		return fantasy.NewTextErrorResponse(err.Error()), nil
 	}
 
-	if count > 1 {
-		return fantasy.NewTextErrorResponse(fmt.Sprintf("found %d matches for old_text in %s. Provide more context to identify the correct match.", count, args.Path)), nil
+	// Apply all edits
+	newContent, applied, err := applyEdits(content, replacements)
+	if err != nil {
+		return fantasy.NewTextErrorResponse(err.Error()), nil
 	}
 
-	// Apply the edit
-	newContent := strings.Replace(normalized, normalizedOld, args.NewText, 1)
-
+	// Write the file
 	if err := os.WriteFile(absPath, []byte(newContent), 0644); err != nil {
 		return fantasy.NewTextErrorResponse(fmt.Sprintf("failed to write file: %v", err)), nil
 	}
 
-	idx := strings.Index(normalized, normalizedOld)
-	diff := generateDiff(absPath, normalized, newContent, idx)
-	resp := fantasy.NewTextResponse(fmt.Sprintf("Applied edit to %s\n%s", args.Path, diff))
-	return fantasy.WithResponseMetadata(resp, editDiffMeta(absPath, normalizedOld, args.NewText)), nil
+	// Generate diff
+	normalizedContent := strings.ReplaceAll(content, "\r\n", "\n")
+	diff := generateDiff(absPath, normalizedContent, newContent)
+
+	// Build response with fuzzy match indication
+	fuzzyCount := 0
+	for _, m := range applied {
+		if m.usedFuzzyMatch {
+			fuzzyCount++
+		}
+	}
+
+	var msg string
+	if len(applied) == 1 {
+		if fuzzyCount > 0 {
+			msg = fmt.Sprintf("Applied edit (fuzzy match) to %s\n%s", args.Path, diff)
+		} else {
+			msg = fmt.Sprintf("Applied edit to %s\n%s", args.Path, diff)
+		}
+	} else {
+		if fuzzyCount > 0 {
+			msg = fmt.Sprintf("Applied %d edits (%d fuzzy) to %s\n%s", len(applied), fuzzyCount, args.Path, diff)
+		} else {
+			msg = fmt.Sprintf("Applied %d edits to %s\n%s", len(applied), args.Path, diff)
+		}
+	}
+
+	resp := fantasy.NewTextResponse(msg)
+	return fantasy.WithResponseMetadata(resp, editDiffMeta(absPath, applied)), nil
+}
+
+// normalizeEditInput validates and normalizes the edit input.
+// Returns error if both single-edit and multi-edit modes are used.
+func normalizeEditInput(args editArgs) ([]replacement, error) {
+	singleMode := args.OldText != "" || args.NewText != ""
+	multiMode := len(args.Edits) > 0
+
+	if singleMode && multiMode {
+		return nil, fmt.Errorf("cannot use old_text/new_text together with edits array")
+	}
+
+	if !singleMode && !multiMode {
+		return nil, fmt.Errorf("must provide either old_text/new_text or edits array")
+	}
+
+	if singleMode {
+		if args.OldText == "" {
+			return nil, fmt.Errorf("old_text is required when using single-edit mode")
+		}
+		if args.NewText == "" {
+			return nil, fmt.Errorf("new_text is required when using single-edit mode")
+		}
+		return []replacement{{
+			oldText:     strings.ReplaceAll(args.OldText, "\r\n", "\n"),
+			newText:     strings.ReplaceAll(args.NewText, "\r\n", "\n"),
+			originalOld: args.OldText,
+			originalNew: args.NewText,
+			index:       0,
+		}}, nil
+	}
+
+	// Multi-edit mode
+	var reps []replacement
+	for i, edit := range args.Edits {
+		if edit.OldText == "" {
+			return nil, fmt.Errorf("edits[%d].old_text is required", i)
+		}
+		reps = append(reps, replacement{
+			oldText:     strings.ReplaceAll(edit.OldText, "\r\n", "\n"),
+			newText:     strings.ReplaceAll(edit.NewText, "\r\n", "\n"),
+			originalOld: edit.OldText,
+			originalNew: edit.NewText,
+			index:       i,
+		})
+	}
+	return reps, nil
+}
+
+// applyEdits applies multiple replacements to the content.
+// All matches are against the original content (non-incremental).
+// Returns the new content, the applied matches, and any error.
+func applyEdits(content string, edits []replacement) (string, []matchedReplacement, error) {
+	normalizedContent := strings.ReplaceAll(content, "\r\n", "\n")
+
+	// Find all matches
+	var matched []matchedReplacement
+	for _, edit := range edits {
+		m, err := findMatch(normalizedContent, edit)
+		if err != nil {
+			return "", nil, err
+		}
+		matched = append(matched, *m)
+	}
+
+	// Sort by position
+	sort.Slice(matched, func(i, j int) bool {
+		return matched[i].start < matched[j].start
+	})
+
+	// Check for overlaps
+	for i := 1; i < len(matched); i++ {
+		if matched[i-1].end > matched[i].start {
+			return "", nil, fmt.Errorf("edits[%d] and edits[%d] overlap; merge them into a single edit",
+				matched[i-1].index, matched[i].index)
+		}
+	}
+
+	// Apply edits in reverse order (end to start) to maintain stable offsets
+	result := normalizedContent
+	for i := len(matched) - 1; i >= 0; i-- {
+		m := matched[i]
+		result = result[:m.start] + m.newText + result[m.end:]
+	}
+
+	return result, matched, nil
+}
+
+// findMatch finds a unique match for the edit in the content.
+// Returns error if not found or ambiguous.
+func findMatch(content string, edit replacement) (*matchedReplacement, error) {
+	// Try exact match first
+	count := strings.Count(content, edit.oldText)
+
+	if count == 0 {
+		// Try fuzzy match
+		idx, matchLen := fuzzyMatch(content, edit.oldText)
+		if idx < 0 {
+			return nil, fmt.Errorf("edits[%d]: could not find old_text in file. The text must match exactly (including whitespace)", edit.index)
+		}
+		// Use the matched text from content for the replacement
+		matchedText := content[idx : idx+matchLen]
+		return &matchedReplacement{
+			replacement: replacement{
+				oldText:     matchedText,
+				newText:     edit.newText,
+				originalOld: edit.originalOld,
+				originalNew: edit.originalNew,
+				index:       edit.index,
+			},
+			start:          idx,
+			end:            idx + matchLen,
+			usedFuzzyMatch: true,
+		}, nil
+	}
+
+	if count > 1 {
+		return nil, fmt.Errorf("found %d matches for edits[%d].old_text; each old_text must be unique, provide more context to identify the correct match", count, edit.index)
+	}
+
+	// Single exact match
+	idx := strings.Index(content, edit.oldText)
+	return &matchedReplacement{
+		replacement: edit,
+		start:       idx,
+		end:         idx + len(edit.oldText),
+	}, nil
 }
 
 // editDiffMeta builds the structured metadata attached to edit tool responses.
-func editDiffMeta(path, oldText, newText string) map[string]any {
+func editDiffMeta(path string, applied []matchedReplacement) map[string]any {
+	var diffBlocks []map[string]any
+	totalAdditions, totalDeletions := 0, 0
+
+	for _, m := range applied {
+		diffBlocks = append(diffBlocks, map[string]any{
+			"old_text": m.originalOld,
+			"new_text": m.originalNew,
+		})
+		totalAdditions += strings.Count(m.originalNew, "\n") + 1
+		totalDeletions += strings.Count(m.originalOld, "\n") + 1
+	}
+
 	return map[string]any{
 		"file_diffs": []map[string]any{{
-			"path":      path,
-			"additions": strings.Count(newText, "\n") + 1,
-			"deletions": strings.Count(oldText, "\n") + 1,
-			"diff_blocks": []map[string]any{{
-				"old_text": oldText,
-				"new_text": newText,
-			}},
+			"path":        path,
+			"additions":   totalAdditions,
+			"deletions":   totalDeletions,
+			"diff_blocks": diffBlocks,
 		}},
 	}
 }
 
 // fuzzyMatch tries to find old_text with relaxed matching:
-// - Strips trailing whitespace per line
-// - Normalizes unicode quotes to ASCII
-// - Normalizes unicode dashes/spaces
-// Returns (index, matchLength) or (-1, 0) if not found.
+//   - Strips trailing whitespace per line
+//   - Normalizes unicode quotes to ASCII
+//   - Normalizes unicode dashes/spaces
+//
+// Returns (index, matchLength) in the original content, or (-1, 0) if not
+// found or ambiguous (multiple matches).
 func fuzzyMatch(content, search string) (int, int) {
-	normalizedContent := normalizeForFuzzy(content)
-	normalizedSearch := normalizeForFuzzy(search)
+	normContent, contentMap := normalizeWithMap(content)
+	normSearch := normalizeForFuzzy(search)
 
-	idx := strings.Index(normalizedContent, normalizedSearch)
+	if normSearch == "" {
+		return -1, 0
+	}
+
+	idx := strings.Index(normContent, normSearch)
 	if idx < 0 {
 		return -1, 0
 	}
 
-	// Map back to original content position
-	// Since normalization can change lengths, we need to find the
-	// corresponding region in the original content
-	origIdx := mapFuzzyIndex(content, normalizedContent, idx)
-	origEnd := mapFuzzyIndex(content, normalizedContent, idx+len(normalizedSearch))
+	// Reject ambiguous matches — if there are multiple fuzzy matches
+	// we can't safely pick one.
+	if strings.Count(normContent, normSearch) > 1 {
+		return -1, 0
+	}
 
-	return origIdx, origEnd - origIdx
+	// Map normalized byte positions back to original byte positions.
+	origStart := contentMap[idx]
+	endNorm := idx + len(normSearch)
+	var origEnd int
+	if endNorm >= len(normContent) {
+		origEnd = len(content)
+	} else {
+		origEnd = contentMap[endNorm]
+	}
+
+	return origStart, origEnd - origStart
 }
 
-func normalizeForFuzzy(s string) string {
-	// Strip trailing whitespace per line
+// normalizeWithMap normalizes s for fuzzy matching and returns both the
+// normalized string and a byte-position mapping where mapping[i] is the
+// original byte position corresponding to normalized byte position i.
+//
+// Normalization: trim trailing whitespace per line, replace unicode
+// quotes/dashes/spaces with their ASCII equivalents.
+func normalizeWithMap(s string) (string, []int) {
+	var result []byte
+	var mapping []int // mapping[i] = original byte position for result byte i
+
 	lines := strings.Split(s, "\n")
-	for i, line := range lines {
-		lines[i] = strings.TrimRightFunc(line, unicode.IsSpace)
-	}
-	result := strings.Join(lines, "\n")
-
-	// Normalize smart quotes
-	replacer := strings.NewReplacer(
-		"\u201c", "\"", // left double quote
-		"\u201d", "\"", // right double quote
-		"\u2018", "'", // left single quote
-		"\u2019", "'", // right single quote
-		"\u2013", "-", // en dash
-		"\u2014", "-", // em dash
-		"\u00a0", " ", // non-breaking space
-	)
-	return replacer.Replace(result)
-}
-
-func mapFuzzyIndex(original, normalized string, normIdx int) int {
-	// Simple approach: count runes up to normIdx in normalized,
-	// then advance that many runes in original.
-	// This works because our normalization only replaces runes 1:1.
-	origRunes := []rune(original)
-	normRunes := []rune(normalized)
-
-	if normIdx >= len(normRunes) {
-		return len(original)
-	}
-
-	// Count bytes for the first normIdx runes in original
-	byteCount := 0
-	for i := 0; i < normIdx && i < len(origRunes); i++ {
-		byteCount += len(string(origRunes[i]))
-	}
-	return byteCount
-}
-
-// generateDiff creates a simple unified diff showing the change.
-func generateDiff(path, old, new string, changeIdx int) string {
-	oldLines := strings.Split(old, "\n")
-	newLines := strings.Split(new, "\n")
-
-	// Find the line number where the change starts
-	lineNum := strings.Count(old[:changeIdx], "\n") + 1
-
-	// Show context around the change
-	contextLines := 3
-	start := max(lineNum-contextLines-1, 0)
-
-	var diff strings.Builder
-	fmt.Fprintf(&diff, "--- %s\n+++ %s\n", path, path)
-
-	// Find changed region
-	endOld := min(lineNum+contextLines+countNewlines(old[changeIdx:])+1, len(oldLines))
-	endNew := min(lineNum+contextLines+countNewlines(new[changeIdx:])+1, len(newLines))
-
-	fmt.Fprintf(&diff, "@@ -%d,%d +%d,%d @@\n", start+1, endOld-start, start+1, endNew-start)
-
-	// Very simplified diff: show old lines as removed, new lines as added
-	// around the change region
-	for i := start; i < endOld && i < len(oldLines); i++ {
-		prefix := " "
-		if i >= lineNum-1 && i < lineNum-1+countNewlines(old[changeIdx:])+1 {
-			prefix = "-"
+	origPos := 0
+	for li, line := range lines {
+		if li > 0 {
+			result = append(result, '\n')
+			mapping = append(mapping, origPos)
+			origPos++ // skip \n in original
 		}
-		fmt.Fprintf(&diff, "%s %s\n", prefix, oldLines[i])
+
+		trimmed := strings.TrimRightFunc(line, unicode.IsSpace)
+
+		for j := 0; j < len(trimmed); {
+			r, size := utf8.DecodeRuneInString(trimmed[j:])
+			repl := normalizeRune(r)
+			for k := 0; k < len(repl); k++ {
+				mapping = append(mapping, origPos+j)
+			}
+			result = append(result, repl...)
+			j += size
+		}
+
+		origPos += len(line) // advance past full original line including trailing ws
 	}
 
-	return diff.String()
+	return string(result), mapping
 }
 
-func countNewlines(s string) int {
-	return strings.Count(s, "\n")
+// normalizeRune maps unicode quotes, dashes, and non-breaking spaces to
+// their ASCII equivalents. Returns the original rune as a string for all
+// other characters.
+func normalizeRune(r rune) string {
+	switch r {
+	case '\u201c', '\u201d': // left/right double quote
+		return "\""
+	case '\u2018', '\u2019': // left/right single quote
+		return "'"
+	case '\u2013', '\u2014': // en dash, em dash
+		return "-"
+	case '\u00a0': // non-breaking space
+		return " "
+	default:
+		return string(r)
+	}
+}
+
+// normalizeForFuzzy normalizes s for fuzzy matching (without position mapping).
+// Used for the search string where position mapping is not needed.
+func normalizeForFuzzy(s string) string {
+	norm, _ := normalizeWithMap(s)
+	return norm
+}
+
+// generateDiff creates a unified diff showing the change between old and new
+// file contents. Uses the go-udiff library for correct diff computation.
+func generateDiff(path, old, new string) string {
+	return udiff.Unified(path, path, old, new)
 }

internal/core/subagent.go

@@ -28,12 +28,14 @@ type SubagentSpawnResult struct {
 // SubagentSpawnFunc is a callback that spawns an in-process subagent. The
 // parent Kit instance injects this into the context so the core tool can
 // call back without importing pkg/kit (which would create a cycle).
-type SubagentSpawnFunc func(ctx context.Context, prompt, model, systemPrompt string, timeout time.Duration) (*SubagentSpawnResult, error)
+// The toolCallID parameter is the LLM-assigned ID of the subagent
+// tool call, enabling the parent to correlate subagent events.
+type SubagentSpawnFunc func(ctx context.Context, toolCallID, prompt, model, systemPrompt string, timeout time.Duration) (*SubagentSpawnResult, error)
 
 type subagentCtxKey struct{}
 
 // WithSubagentSpawner stores a spawn function in the context so that the
-// spawn_subagent core tool can create in-process subagents.
+// subagent core tool can create in-process subagents.
 func WithSubagentSpawner(ctx context.Context, fn SubagentSpawnFunc) context.Context {
 	return context.WithValue(ctx, subagentCtxKey{}, fn)
 }
@@ -47,7 +49,7 @@ func getSubagentSpawner(ctx context.Context) SubagentSpawnFunc {
 }
 
 // ---------------------------------------------------------------------------
-// spawn_subagent tool
+// subagent tool
 // ---------------------------------------------------------------------------
 
 type subagentArgs struct {
@@ -57,11 +59,11 @@ type subagentArgs struct {
 	TimeoutSeconds int    `json:"timeout_seconds,omitempty"`
 }
 
-// NewSubagentTool creates the spawn_subagent core tool.
+// NewSubagentTool creates the subagent core tool.
 func NewSubagentTool(opts ...ToolOption) fantasy.AgentTool {
 	return &coreTool{
 		info: fantasy.ToolInfo{
-			Name: "spawn_subagent",
+			Name: "subagent",
 			Description: `Spawn a subagent to perform a task autonomously.
 
 The subagent runs as a separate in-process Kit instance with full tool access
@@ -128,8 +130,16 @@ func executeSubagent(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolRe
 		), fmt.Errorf("no subagent spawner in context")
 	}
 
+	// Detach from the parent's deadline so the subagent gets its own
+	// independent timeout (applied downstream in Kit.Subagent). The parent
+	// context may carry a tight deadline from the LLM generation loop or
+	// other tool timeouts that would prematurely kill the subagent.
+	// We preserve context values (spawner, etc.) and propagate parent
+	// cancellation (e.g. user hits Ctrl-C) without inheriting the deadline.
+	spawnCtx := detachedWithCancel(ctx)
+
 	// Spawn in-process subagent.
-	result, err := spawner(ctx, args.Task, args.Model, args.SystemPrompt, timeout)
+	result, err := spawner(spawnCtx, call.ID, args.Task, args.Model, args.SystemPrompt, timeout)
 	if err != nil || result.Error != nil {
 		spawnErr := err
 		if spawnErr == nil {
@@ -162,6 +172,39 @@ func executeSubagent(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolRe
 	return resp, nil
 }
 
+// ---------------------------------------------------------------------------
+// Context detachment
+// ---------------------------------------------------------------------------
+
+// detachedContext wraps a parent context, preserving its values but removing
+// its deadline and cancellation. This allows the subagent to have its own
+// independent timeout while still accessing context-stored values (e.g. the
+// subagent spawner function).
+type detachedContext struct {
+	parent context.Context
+}
+
+func (d detachedContext) Deadline() (time.Time, bool) { return time.Time{}, false }
+func (d detachedContext) Done() <-chan struct{}       { return nil }
+func (d detachedContext) Err() error                  { return nil }
+func (d detachedContext) Value(key any) any           { return d.parent.Value(key) }
+
+// detachedWithCancel creates a new context that inherits values from the
+// parent but has no deadline. Cancellation of the parent is propagated: when
+// the parent is cancelled the returned context is also cancelled, but the
+// parent's deadline does not apply to the child.
+func detachedWithCancel(parent context.Context) context.Context {
+	child, cancel := context.WithCancel(detachedContext{parent: parent})
+	go func() {
+		select {
+		case <-parent.Done():
+			cancel()
+		case <-child.Done():
+		}
+	}()
+	return child
+}
+
 // truncateResponse limits the response length to avoid overwhelming context windows.
 func truncateResponse(s string, maxLen int) string {
 	if len(s) <= maxLen {

internal/core/tools.go

@@ -86,7 +86,7 @@ func ReadOnlyTools(opts ...ToolOption) []fantasy.AgentTool {
 	}
 }
 
-// SubagentTools returns all core tools except spawn_subagent. This prevents
+// SubagentTools returns all core tools except subagent. This prevents
 // infinite recursion when a subagent is itself a Kit instance.
 func SubagentTools(opts ...ToolOption) []fantasy.AgentTool {
 	return []fantasy.AgentTool{

internal/core/truncate.go

@@ -6,14 +6,17 @@ import (
 )
 
 const (
-	defaultMaxLines = 2000
-	defaultMaxBytes = 50 * 1024 // 50KB
-	grepMaxLineLen  = 500
+	defaultMaxLines   = 2000
+	defaultMaxBytes   = 50 * 1024 // 50KB
+	defaultMaxLineLen = 2000      // max characters per line before truncation
+	grepMaxLineLen    = 500
 
 	// DefaultMaxLines is the exported default line limit for truncation.
 	DefaultMaxLines = defaultMaxLines
 	// DefaultMaxBytes is the exported default byte limit for truncation.
 	DefaultMaxBytes = defaultMaxBytes
+	// DefaultMaxLineLen is the exported default per-line character limit.
+	DefaultMaxLineLen = defaultMaxLineLen
 )
 
 // TruncationResult describes how output was truncated.
@@ -26,6 +29,8 @@ type TruncationResult struct {
 }
 
 // TruncateTail keeps the last maxLines lines and at most maxBytes bytes.
+// Individual lines longer than defaultMaxLineLen are truncated to prevent
+// extremely long single lines from blowing up the TUI when wrapped.
 // Used for bash output where the tail is most relevant.
 func TruncateTail(content string, maxLines, maxBytes int) TruncationResult {
 	if maxLines <= 0 {
@@ -38,11 +43,11 @@ func TruncateTail(content string, maxLines, maxBytes int) TruncationResult {
 	lines := strings.Split(content, "\n")
 	total := len(lines)
 
-	if len(content) <= maxBytes && total <= maxLines {
-		return TruncationResult{Content: content, Total: total, Kept: total}
-	}
+	// Truncate individual long lines first to prevent single lines from
+	// wrapping into hundreds of visual lines in the TUI.
+	lines = truncateLongLines(lines, defaultMaxLineLen)
 
-	// Truncate by lines first (keep tail)
+	// Truncate by lines (keep tail)
 	truncBy := ""
 	if total > maxLines {
 		lines = lines[total-maxLines:]
@@ -78,6 +83,7 @@ func TruncateTail(content string, maxLines, maxBytes int) TruncationResult {
 }
 
 // truncateHead keeps the first maxLines lines and at most maxBytes bytes.
+// Individual lines longer than defaultMaxLineLen are truncated.
 // Used for read, grep, find, ls output where the head is most relevant.
 func truncateHead(content string, maxLines, maxBytes int) TruncationResult {
 	if maxLines <= 0 {
@@ -90,9 +96,8 @@ func truncateHead(content string, maxLines, maxBytes int) TruncationResult {
 	lines := strings.Split(content, "\n")
 	total := len(lines)
 
-	if len(content) <= maxBytes && total <= maxLines {
-		return TruncationResult{Content: content, Total: total, Kept: total}
-	}
+	// Truncate individual long lines first.
+	lines = truncateLongLines(lines, defaultMaxLineLen)
 
 	truncBy := ""
 	if total > maxLines {
@@ -125,6 +130,19 @@ func truncateHead(content string, maxLines, maxBytes int) TruncationResult {
 	}
 }
 
+// truncateLongLines caps each line to maxLen characters, appending a
+// "[...N chars truncated]" marker to any line that exceeds the limit.
+// This prevents a single very long line (e.g. minified JSON/JS) from
+// wrapping into hundreds of visual rows and blowing up the TUI.
+func truncateLongLines(lines []string, maxLen int) []string {
+	for i, line := range lines {
+		if len(line) > maxLen {
+			lines[i] = line[:maxLen] + fmt.Sprintf("... [%d chars truncated]", len(line)-maxLen)
+		}
+	}
+	return lines
+}
+
 // truncateLine truncates a single line to maxChars, appending "..." if cut.
 func truncateLine(line string, maxChars int) string {
 	if maxChars <= 0 {

internal/core/truncate_test.go

@@ -0,0 +1,163 @@
+package core
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestTruncateTail_LongLines(t *testing.T) {
+	// A single line of 5000 chars should be truncated to defaultMaxLineLen.
+	longLine := strings.Repeat("x", 5000)
+	tr := TruncateTail(longLine, 2000, 50*1024)
+
+	if len(tr.Content) > defaultMaxLineLen+100 { // +100 for the "[...N chars truncated]" suffix
+		t.Errorf("single long line not truncated: got %d chars, want <= %d", len(tr.Content), defaultMaxLineLen+100)
+	}
+	if !strings.Contains(tr.Content, "chars truncated]") {
+		t.Error("truncated line should contain truncation marker")
+	}
+}
+
+func TestTruncateTail_NormalLines(t *testing.T) {
+	// Lines within the limit should pass through unchanged.
+	content := "line1\nline2\nline3"
+	tr := TruncateTail(content, 2000, 50*1024)
+	if tr.Content != content {
+		t.Errorf("got %q, want %q", tr.Content, content)
+	}
+	if tr.Truncated {
+		t.Error("should not be marked as truncated")
+	}
+}
+
+func TestTruncateTail_LineCount(t *testing.T) {
+	lines := make([]string, 100)
+	for i := range lines {
+		lines[i] = "line"
+	}
+	content := strings.Join(lines, "\n")
+	tr := TruncateTail(content, 10, 50*1024)
+
+	if !tr.Truncated {
+		t.Error("should be marked as truncated")
+	}
+	if tr.Total != 100 {
+		t.Errorf("total = %d, want 100", tr.Total)
+	}
+	if tr.Kept != 10 {
+		t.Errorf("kept = %d, want 10", tr.Kept)
+	}
+}
+
+func TestTruncateHead_LongLines(t *testing.T) {
+	longLine := strings.Repeat("y", 5000)
+	tr := truncateHead(longLine, 2000, 50*1024)
+
+	if len(tr.Content) > defaultMaxLineLen+100 {
+		t.Errorf("single long line not truncated: got %d chars, want <= %d", len(tr.Content), defaultMaxLineLen+100)
+	}
+	if !strings.Contains(tr.Content, "chars truncated]") {
+		t.Error("truncated line should contain truncation marker")
+	}
+}
+
+func TestTruncateHead_NormalLines(t *testing.T) {
+	content := "line1\nline2\nline3"
+	tr := truncateHead(content, 2000, 50*1024)
+	if tr.Content != content {
+		t.Errorf("got %q, want %q", tr.Content, content)
+	}
+	if tr.Truncated {
+		t.Error("should not be marked as truncated")
+	}
+}
+
+func TestTruncateHead_LineCount(t *testing.T) {
+	lines := make([]string, 100)
+	for i := range lines {
+		lines[i] = "line"
+	}
+	content := strings.Join(lines, "\n")
+	tr := truncateHead(content, 10, 50*1024)
+
+	if !tr.Truncated {
+		t.Error("should be marked as truncated")
+	}
+	if tr.Total != 100 {
+		t.Errorf("total = %d, want 100", tr.Total)
+	}
+	if tr.Kept != 10 {
+		t.Errorf("kept = %d, want 10", tr.Kept)
+	}
+}
+
+func TestTruncateLongLines(t *testing.T) {
+	lines := []string{
+		"short",
+		strings.Repeat("a", 3000),
+		"also short",
+	}
+	result := truncateLongLines(lines, 100)
+
+	if result[0] != "short" {
+		t.Error("short line should be unchanged")
+	}
+	if len(result[1]) > 200 { // 100 chars + marker
+		t.Errorf("long line not truncated: len=%d", len(result[1]))
+	}
+	if !strings.Contains(result[1], "chars truncated]") {
+		t.Error("should contain truncation marker")
+	}
+	if result[2] != "also short" {
+		t.Error("short line should be unchanged")
+	}
+}
+
+func TestTruncateTail_MixedLongAndManyLines(t *testing.T) {
+	// 50 lines, each 3000 chars — tests both per-line and total truncation.
+	lines := make([]string, 50)
+	for i := range lines {
+		lines[i] = strings.Repeat("z", 3000)
+	}
+	content := strings.Join(lines, "\n")
+
+	tr := TruncateTail(content, 10, 50*1024)
+
+	// Should keep 10 lines.
+	if tr.Kept != 10 {
+		t.Errorf("kept = %d, want 10", tr.Kept)
+	}
+	// Each line should be capped at ~defaultMaxLineLen.
+	resultLines := strings.Split(tr.Content, "\n")
+	for i, line := range resultLines {
+		if len(line) > defaultMaxLineLen+100 {
+			t.Errorf("line %d too long: %d chars", i, len(line))
+		}
+	}
+}
+
+func TestTruncateLine(t *testing.T) {
+	short := "hello"
+	if truncateLine(short, 10) != short {
+		t.Error("short line should be unchanged")
+	}
+
+	long := strings.Repeat("x", 100)
+	result := truncateLine(long, 10)
+	if len(result) != 13 { // 10 + "..."
+		t.Errorf("got len %d, want 13", len(result))
+	}
+
+	// Default max for 0 — input shorter than default, so unchanged
+	result2 := truncateLine(long, 0)
+	if result2 != long {
+		t.Errorf("100-char line should be unchanged when maxChars defaults to %d", grepMaxLineLen)
+	}
+
+	// Longer input with default
+	veryLong := strings.Repeat("x", 1000)
+	result3 := truncateLine(veryLong, 0)
+	if len(result3) != grepMaxLineLen+3 {
+		t.Errorf("got len %d, want %d", len(result3), grepMaxLineLen+3)
+	}
+}

internal/extensions/api.go

@@ -572,6 +572,102 @@ type Context struct {
 	//   })
 	//   // handle.Kill() to cancel, handle.Wait() to block
 	SpawnSubagent func(SubagentConfig) (*SubagentHandle, *SubagentResult, error)
+
+	// -------------------------------------------------------------------------
+	// Tree Navigation API (Phase 1 Bridge)
+	// -------------------------------------------------------------------------
+
+	// GetTreeNode returns a node by ID with full metadata and children.
+	// Returns nil if entry not found.
+	GetTreeNode func(entryID string) *TreeNode
+
+	// GetCurrentBranch returns the path from root to current leaf.
+	// Each node contains full metadata (unlike GetMessages which flattens).
+	GetCurrentBranch func() []TreeNode
+
+	// GetChildren returns direct child IDs of an entry.
+	GetChildren func(entryID string) []string
+
+	// NavigateTo branches/forks the session to the specified entry ID.
+	// Equivalent to SDK's Branch() but for extensions.
+	NavigateTo func(entryID string) TreeNavigationResult
+
+	// SummarizeBranch uses LLM to summarize a branch range.
+	// Returns summary text or error string (empty if success).
+	SummarizeBranch func(fromID, toID string) string
+
+	// CollapseBranch replaces a branch range with a summary entry.
+	// This is the "fresh context" primitive for context window management.
+	CollapseBranch func(fromID, toID, summary string) TreeNavigationResult
+
+	// -------------------------------------------------------------------------
+	// Skill Loading API (Phase 2 Bridge)
+	// -------------------------------------------------------------------------
+
+	// LoadSkill loads a single skill file from path.
+	// Parses YAML frontmatter, returns skill with content ready for injection.
+	LoadSkill func(path string) (*Skill, string)
+
+	// LoadSkillsFromDir discovers and loads all skills from a directory.
+	LoadSkillsFromDir func(dir string) SkillLoadResult
+
+	// DiscoverSkills finds skills in standard locations.
+	// Checks ~/.config/kit/skills/, .kit/skills/, .agents/skills/
+	DiscoverSkills func() SkillLoadResult
+
+	// InjectSkillAsContext sends a skill's content as a system message.
+	// Looks up skill by name from discovered skills.
+	InjectSkillAsContext func(skillName string) string
+
+	// InjectRawSkillAsContext loads and immediately injects a skill file.
+	InjectRawSkillAsContext func(path string) string
+
+	// GetAvailableSkills returns all currently loaded/discovered skills.
+	GetAvailableSkills func() []Skill
+
+	// -------------------------------------------------------------------------
+	// Template Parsing API (Phase 3 Bridge)
+	// -------------------------------------------------------------------------
+
+	// ParseTemplate extracts {{variables}} from template content.
+	ParseTemplate func(name, content string) PromptTemplate
+
+	// RenderTemplate substitutes variables into template content.
+	RenderTemplate func(tpl PromptTemplate, vars map[string]string) string
+
+	// ParseArguments parses command-line style arguments.
+	ParseArguments func(input string, pattern ArgumentPattern) ParseResult
+
+	// SimpleParseArguments parses $1, $2, $@ style arguments.
+	// Returns slice where [0]=full input, [1]=$1, [2]=$2, ... [n]=$@
+	SimpleParseArguments func(input string, count int) []string
+
+	// EvaluateModelConditional checks if condition matches current model.
+	// Condition supports wildcards: * matches any, ? matches single char.
+	EvaluateModelConditional func(condition string) bool
+
+	// RenderWithModelConditionals processes <if-model> blocks in content.
+	RenderWithModelConditionals func(content string) string
+
+	// -------------------------------------------------------------------------
+	// Model Resolution API (Phase 4 Bridge)
+	// -------------------------------------------------------------------------
+
+	// ResolveModelChain attempts each model in order until one is available.
+	ResolveModelChain func(preferences []string) ModelResolutionResult
+
+	// GetModelCapabilities returns capabilities for a specific model.
+	// If model is empty, uses current model.
+	GetModelCapabilities func(model string) (ModelCapabilities, string)
+
+	// CheckModelAvailable verifies if a model string is valid.
+	CheckModelAvailable func(model string) bool
+
+	// GetCurrentProvider returns just the provider part of current model.
+	GetCurrentProvider func() string
+
+	// GetCurrentModelID returns just the model ID part of current model.
+	GetCurrentModelID func() string
 }
 
 // ---------------------------------------------------------------------------
@@ -598,6 +694,148 @@ type SessionMessage struct {
 	Timestamp string
 }
 
+// ---------------------------------------------------------------------------
+// Tree navigation types (exposed to Yaegi — concrete structs)
+// ---------------------------------------------------------------------------
+
+// TreeNode represents a node in the session tree for navigation.
+// Extensions use this to traverse conversation history and implement
+// features like "fresh context" loops and branch summarization.
+type TreeNode struct {
+	// ID is the unique entry identifier.
+	ID string
+	// ParentID links this entry to its parent (empty if root).
+	ParentID string
+	// Type is the entry type: "message", "branch_summary", "model_change", "extension_data", "tool_execution".
+	Type string
+	// Role is the message role for message entries: "user", "assistant", "system", "tool".
+	Role string
+	// Content is the text content or summary.
+	Content string
+	// Model is the model that generated this (for assistant messages).
+	Model string
+	// Provider is the provider used.
+	Provider string
+	// Timestamp is the RFC3339-formatted creation time.
+	Timestamp string
+	// Children is the list of child entry IDs for tree traversal.
+	Children []string
+}
+
+// TreeNavigationResult reports success or failure of tree operations.
+type TreeNavigationResult struct {
+	// Success is true if the operation completed.
+	Success bool
+	// Error describes what went wrong (empty if success).
+	Error string
+}
+
+// ---------------------------------------------------------------------------
+// Skill types (exposed to Yaegi — concrete structs)
+// ---------------------------------------------------------------------------
+
+// Skill represents a loaded skill file with parsed YAML frontmatter.
+type Skill struct {
+	// Name is the human-readable identifier.
+	Name string
+	// Description summarizes what this skill provides.
+	Description string
+	// Content is the markdown body (frontmatter stripped).
+	Content string
+	// Path is the absolute filesystem path.
+	Path string
+	// Tags are optional labels for categorization.
+	Tags []string
+	// When controls automatic inclusion: "always", "on-demand", or file-glob.
+	When string
+}
+
+// SkillLoadResult reports skills loaded from a directory.
+type SkillLoadResult struct {
+	// Skills is the list of loaded skills.
+	Skills []Skill
+	// Error describes loading failures (empty if success).
+	Error string
+}
+
+// ---------------------------------------------------------------------------
+// Template parsing types (exposed to Yaegi — concrete structs)
+// ---------------------------------------------------------------------------
+
+// PromptTemplate represents a parsed template with variable placeholders.
+type PromptTemplate struct {
+	// Name is the template identifier.
+	Name string
+	// Content is the original template content.
+	Content string
+	// Variables are the extracted {{variable}} names.
+	Variables []string
+}
+
+// ArgumentPattern defines how to parse command arguments.
+type ArgumentPattern struct {
+	// Positional names for $1, $2, etc.
+	Positional []string
+	// Rest is the variable name for $@ (all remaining).
+	Rest string
+	// Flags maps flag names to variable names (e.g., "--loop" -> "loop").
+	Flags map[string]string
+}
+
+// ParseResult reports argument parsing outcome.
+type ParseResult struct {
+	// Vars maps variable names to values for positional args.
+	Vars map[string]string
+	// Flags maps flag names to values.
+	Flags map[string]string
+	// Rest is remaining unparsed text.
+	Rest string
+	// Error describes parsing failures (empty if success).
+	Error string
+}
+
+// ModelConditional represents an <if-model> block for evaluation.
+type ModelConditional struct {
+	// Condition is the model pattern (e.g., "claude-*", "anthropic/*").
+	Condition string
+	// Content is rendered if condition matches.
+	Content string
+	// Else is rendered if condition doesn't match.
+	Else string
+}
+
+// ---------------------------------------------------------------------------
+// Model resolution types (exposed to Yaegi — concrete structs)
+// ---------------------------------------------------------------------------
+
+// ModelCapabilities describes what a model supports.
+type ModelCapabilities struct {
+	// Provider is the provider ID (e.g., "anthropic").
+	Provider string
+	// ModelID is the model identifier (e.g., "claude-sonnet-4-20250929").
+	ModelID string
+	// ContextLimit is the maximum context window in tokens.
+	ContextLimit int
+	// OutputLimit is the maximum output tokens.
+	OutputLimit int
+	// Reasoning indicates if the model supports reasoning/thinking.
+	Reasoning bool
+	// Streaming indicates if the model supports streaming.
+	Streaming bool
+}
+
+// ModelResolutionResult reports model chain resolution outcome.
+type ModelResolutionResult struct {
+	// Model is the selected model in "provider/model" format.
+	Model string
+	// Capabilities describes the selected model.
+	Capabilities ModelCapabilities
+	// Attempted lists models tried before success.
+	Attempted []string
+	// Error describes resolution failures (empty if success).
+	Error string
+}
+
 // ExtensionEntry represents persisted extension data stored in the session.
 // Extensions use AppendEntry to save custom state and GetEntries to retrieve
 // it on session resume.
@@ -727,6 +965,7 @@ type API struct {
 	onToolCall                func(func(ToolCallEvent, Context) *ToolCallResult)
 	onToolExecStart           func(func(ToolExecutionStartEvent, Context))
 	onToolExecEnd             func(func(ToolExecutionEndEvent, Context))
+	onToolOutput              func(func(ToolOutputEvent, Context))
 	onToolResult              func(func(ToolResultEvent, Context) *ToolResultResult)
 	onInput                   func(func(InputEvent, Context) *InputResult)
 	onBeforeAgentStart        func(func(BeforeAgentStartEvent, Context) *BeforeAgentStartResult)
@@ -749,6 +988,9 @@ type API struct {
 	registerOption            func(OptionDef)
 	registerShortcutFn        func(ShortcutDef, func(Context))
 	registerMessageRendererFn func(MessageRendererConfig)
+	onSubagentStart           func(func(SubagentStartEvent, Context))
+	onSubagentChunk           func(func(SubagentChunkEvent, Context))
+	onSubagentEnd             func(func(SubagentEndEvent, Context))
 }
 
 // OnToolCall registers a handler that fires before a tool executes.
@@ -767,12 +1009,40 @@ func (a *API) OnToolExecutionEnd(handler func(ToolExecutionEndEvent, Context)) {
 	a.onToolExecEnd(handler)
 }
 
+// OnToolOutput registers a handler for streaming tool output chunks.
+// This fires for each output line as it arrives from tools like bash,
+// allowing extensions to observe or process output in real-time.
+func (a *API) OnToolOutput(handler func(ToolOutputEvent, Context)) {
+	a.onToolOutput(handler)
+}
+
 // OnToolResult registers a handler that fires after tool execution.
 // Return a non-nil ToolResultResult to modify the output.
 func (a *API) OnToolResult(handler func(ToolResultEvent, Context) *ToolResultResult) {
 	a.onToolResult(handler)
 }
 
+// OnSubagentStart registers a handler that fires when a subagent tool
+// call begins executing. Use the ToolCallID to correlate with subsequent
+// OnSubagentChunk and OnSubagentEnd events for the same subagent.
+func (a *API) OnSubagentStart(handler func(SubagentStartEvent, Context)) {
+	a.onSubagentStart(handler)
+}
+
+// OnSubagentChunk registers a handler for real-time events from a running
+// subagent. ChunkType identifies the kind of event ("text", "tool_call",
+// "tool_result", "tool_execution_start", "tool_execution_end", etc.).
+// Correlate with OnSubagentStart via the ToolCallID field.
+func (a *API) OnSubagentChunk(handler func(SubagentChunkEvent, Context)) {
+	a.onSubagentChunk(handler)
+}
+
+// OnSubagentEnd registers a handler that fires when a subagent call
+// completes. ErrorMsg is non-empty when the subagent failed.
+func (a *API) OnSubagentEnd(handler func(SubagentEndEvent, Context)) {
+	a.onSubagentEnd(handler)
+}
+
 // OnInput registers a handler that fires when user input is received.
 // Return a non-nil InputResult to transform or handle the input.
 func (a *API) OnInput(handler func(InputEvent, Context) *InputResult) {
@@ -1538,6 +1808,19 @@ type ToolExecutionEndEvent struct {
 
 func (e ToolExecutionEndEvent) Type() EventType { return ToolExecutionEnd }
 
+// ToolOutputEvent fires when a tool produces streaming output chunks.
+// This is primarily used for long-running tools like bash to show output
+// in real-time as it arrives, before the tool completes.
+type ToolOutputEvent struct {
+	ToolCallID string
+	ToolName   string
+	ToolKind   string
+	Chunk      string // Output text chunk
+	IsStderr   bool   // Whether this chunk came from stderr
+}
+
+func (e ToolOutputEvent) Type() EventType { return ToolOutput }
+
 // ToolResultEvent fires after tool execution with the output.
 type ToolResultEvent struct {
 	ToolCallID string
@@ -1743,21 +2026,82 @@ type BeforeCompactEvent struct {
 func (e BeforeCompactEvent) Type() EventType { return BeforeCompact }
 
 // BeforeCompactResult controls whether compaction proceeds. Return
-// Cancel=true with an optional Reason to block compaction.
+// Cancel=true with an optional Reason to block compaction, or provide
+// a custom Summary to replace the default LLM-generated one.
 type BeforeCompactResult struct {
 	// Cancel, when true, prevents compaction from proceeding.
 	Cancel bool
 	// Reason is a human-readable explanation shown to the user when
 	// Cancel is true. Empty string uses a default message.
 	Reason string
+	// Summary, when non-empty, replaces the default LLM-generated summary.
+	// The extension is responsible for generating a useful summary.
+	// Ignored when Cancel is true.
+	Summary string
 }
 
 func (BeforeCompactResult) isResult() {}
 
 // ---------------------------------------------------------------------------
-// Theme types (exposed to Yaegi — concrete structs, string hex colors)
+// Subagent lifecycle events (exposed to Yaegi — concrete structs)
 // ---------------------------------------------------------------------------
 
+// SubagentStartEvent fires when a subagent tool call begins executing.
+type SubagentStartEvent struct {
+	// ToolCallID is the LLM-assigned ID of the subagent tool call.
+	// Use this to correlate SubagentChunkEvent and SubagentEndEvent.
+	ToolCallID string
+	// Task is the task description passed to the subagent.
+	Task string
+}
+
+func (e SubagentStartEvent) Type() EventType { return SubagentStart }
+
+// SubagentChunkEvent fires for each real-time event from a running subagent.
+// Type field indicates the kind of event; read the relevant fields accordingly.
+type SubagentChunkEvent struct {
+	// ToolCallID matches the SubagentStartEvent.ToolCallID for this subagent.
+	ToolCallID string
+	// Task is the task description (repeated for convenience).
+	Task string
+	// ChunkType identifies the event kind:
+	//   "text"                 — LLM text chunk (read Content)
+	//   "reasoning"            — reasoning/thinking delta (read Content)
+	//   "tool_call"            — subagent called a tool (read ToolName, ToolArgs)
+	//   "tool_result"          — tool returned a result (read ToolName, ToolResult, IsError)
+	//   "tool_execution_start" — tool began executing (read ToolName)
+	//   "tool_execution_end"   — tool finished executing (read ToolName)
+	//   "turn_start"           — subagent turn began
+	//   "turn_end"             — subagent turn ended
+	ChunkType string
+	// Content carries text for "text" and "reasoning" chunk types.
+	Content string
+	// ToolName is set on tool-related chunk types.
+	ToolName string
+	// ToolArgs is the JSON-encoded tool arguments for "tool_call" chunks.
+	ToolArgs string
+	// ToolResult is the tool output for "tool_result" chunks.
+	ToolResult string
+	// IsError is true when a "tool_result" chunk represents an error.
+	IsError bool
+}
+
+func (e SubagentChunkEvent) Type() EventType { return SubagentChunk }
+
+// SubagentEndEvent fires when a subagent tool call completes.
+type SubagentEndEvent struct {
+	// ToolCallID matches the SubagentStartEvent.ToolCallID for this subagent.
+	ToolCallID string
+	// Task is the task description.
+	Task string
+	// Response is the subagent's final text response (empty on error).
+	Response string
+	// ErrorMsg is non-empty when the subagent failed.
+	ErrorMsg string
+}
+
+func (e SubagentEndEvent) Type() EventType { return SubagentEnd }
+
 // ThemeColor is an adaptive color pair with light and dark hex values.
 // Either field may be empty to inherit from the default theme.
 type ThemeColor struct {

internal/extensions/events.go

@@ -19,6 +19,9 @@ const (
 	// ToolExecutionEnd fires when a tool finishes executing.
 	ToolExecutionEnd EventType = "tool_execution_end"
 
+	// ToolOutput fires when a tool produces streaming output chunks.
+	ToolOutput EventType = "tool_output"
+
 	// ToolResult fires after a tool executes. Handlers can modify the result.
 	ToolResult EventType = "tool_result"
 
@@ -68,6 +71,18 @@ const (
 	// BeforeCompact fires before context compaction runs. Handlers can
 	// cancel compaction by returning Cancel=true.
 	BeforeCompact EventType = "before_compact"
+
+	// SubagentStart fires when a subagent tool call begins executing.
+	// Carries the tool call ID and the task description.
+	SubagentStart EventType = "subagent_start"
+
+	// SubagentChunk fires for each real-time event emitted by a running
+	// subagent: text chunks, tool calls, tool results, etc.
+	SubagentChunk EventType = "subagent_chunk"
+
+	// SubagentEnd fires when a subagent tool call completes (success
+	// or error). Carries the final response and any error message.
+	SubagentEnd EventType = "subagent_end"
 )
 
 // AllEventTypes returns every supported event type.
@@ -79,6 +94,7 @@ func AllEventTypes() []EventType {
 		SessionStart, SessionShutdown,
 		ModelChange, ContextPrepare,
 		BeforeFork, BeforeSessionSwitch, BeforeCompact,
+		SubagentStart, SubagentChunk, SubagentEnd,
 	}
 }
 

internal/extensions/events_test.go

@@ -4,8 +4,8 @@ import "testing"
 
 func TestAllEventTypes_Count(t *testing.T) {
 	all := AllEventTypes()
-	if len(all) != 18 {
-		t.Fatalf("expected 18 event types, got %d", len(all))
+	if len(all) != 21 {
+		t.Fatalf("expected 21 event types, got %d", len(all))
 	}
 }
 
@@ -55,6 +55,9 @@ func TestEventType_TypeMethod(t *testing.T) {
 		{BeforeForkEvent{TargetID: "abc"}, BeforeFork},
 		{BeforeSessionSwitchEvent{Reason: "new"}, BeforeSessionSwitch},
 		{BeforeCompactEvent{EstimatedTokens: 1000}, BeforeCompact},
+		{SubagentStartEvent{ToolCallID: "x", Task: "t"}, SubagentStart},
+		{SubagentChunkEvent{ToolCallID: "x", ChunkType: "text"}, SubagentChunk},
+		{SubagentEndEvent{ToolCallID: "x"}, SubagentEnd},
 	}
 
 	for _, tt := range tests {

internal/extensions/loader.go

@@ -47,46 +47,56 @@ func LoadExtensions(extraPaths []string) ([]LoadedExtension, error) {
 	return loaded, nil
 }
 
+// pathSet is a thread-safe helper for deduplicating and ordering file paths.
+type pathSet struct {
+	m    map[string]bool
+	list []string
+}
+
+func newPathSet() *pathSet {
+	return &pathSet{m: make(map[string]bool)}
+}
+
+func (ps *pathSet) add(p string) bool {
+	abs, err := filepath.Abs(p)
+	if err != nil {
+		return false
+	}
+	if ps.m[abs] {
+		return false
+	}
+	ps.m[abs] = true
+	ps.list = append(ps.list, abs)
+	return true
+}
+
 // discoverExtensionPaths returns deduplicated paths to extension files in
 // load-order (global first, then project-local, then explicit).
 func discoverExtensionPaths(extraPaths []string) []string {
-	seen := make(map[string]bool)
-	var paths []string
-
-	add := func(p string) {
-		abs, err := filepath.Abs(p)
-		if err != nil {
-			return
-		}
-		if seen[abs] {
-			return
-		}
-		seen[abs] = true
-		paths = append(paths, abs)
-	}
+	ps := newPathSet()
 
 	// Global extensions: $XDG_CONFIG_HOME/kit/extensions/ (default ~/.config/kit/extensions/)
 	globalDir := globalExtensionsDir()
 	for _, p := range findExtensionsInDir(globalDir) {
-		add(p)
+		ps.add(p)
 	}
 
 	// Global installed git packages: $XDG_DATA_HOME/kit/git/
 	globalGitDir := globalGitInstallRoot()
 	for _, p := range findExtensionsInGitPackages(globalGitDir) {
-		add(p)
+		ps.add(p)
 	}
 
 	// Project-local extensions: .kit/extensions/
 	localDir := filepath.Join(".kit", "extensions")
 	for _, p := range findExtensionsInDir(localDir) {
-		add(p)
+		ps.add(p)
 	}
 
 	// Project-local installed git packages: .kit/git/
 	projectGitDir := filepath.Join(".kit", "git")
 	for _, p := range findExtensionsInGitPackages(projectGitDir) {
-		add(p)
+		ps.add(p)
 	}
 
 	// Explicit paths (highest precedence)
@@ -97,14 +107,14 @@ func discoverExtensionPaths(extraPaths []string) []string {
 		}
 		if info.IsDir() {
 			for _, found := range findExtensionsInDir(p) {
-				add(found)
+				ps.add(found)
 			}
 		} else if strings.HasSuffix(p, ".go") {
-			add(p)
+			ps.add(p)
 		}
 	}
 
-	return paths
+	return ps.list
 }
 
 // findExtensionsInDir returns .go files in dir and main.go in immediate subdirs.
@@ -439,6 +449,12 @@ func loadSingleExtension(path string) (*LoadedExtension, error) {
 				return nil
 			})
 		},
+		onToolOutput: func(h func(ToolOutputEvent, Context)) {
+			reg(ToolOutput, func(e Event, c Context) Result {
+				h(e.(ToolOutputEvent), c)
+				return nil
+			})
+		},
 		onToolResult: func(h func(ToolResultEvent, Context) *ToolResultResult) {
 			reg(ToolResult, func(e Event, c Context) Result {
 				r := h(e.(ToolResultEvent), c)
@@ -574,6 +590,24 @@ func loadSingleExtension(path string) (*LoadedExtension, error) {
 		registerShortcutFn: func(def ShortcutDef, handler func(Context)) {
 			ext.Shortcuts = append(ext.Shortcuts, ShortcutEntry{Def: def, Handler: handler})
 		},
+		onSubagentStart: func(h func(SubagentStartEvent, Context)) {
+			reg(SubagentStart, func(e Event, c Context) Result {
+				h(e.(SubagentStartEvent), c)
+				return nil
+			})
+		},
+		onSubagentChunk: func(h func(SubagentChunkEvent, Context)) {
+			reg(SubagentChunk, func(e Event, c Context) Result {
+				h(e.(SubagentChunkEvent), c)
+				return nil
+			})
+		},
+		onSubagentEnd: func(h func(SubagentEndEvent, Context)) {
+			reg(SubagentEnd, func(e Event, c Context) Result {
+				h(e.(SubagentEndEvent), c)
+				return nil
+			})
+		},
 	}
 
 	// Call Init — the extension registers its handlers, tools, commands.

internal/extensions/runner.go

@@ -56,11 +56,261 @@ func NewRunner(exts []LoadedExtension) *Runner {
 }
 
 // SetContext updates the runtime context (session ID, model, etc.) that is
-// passed to every handler invocation. Thread-safe.
+// passed to every handler invocation. Nil function fields are replaced with
+// safe no-ops so extension handlers never panic on a missing callback.
+// Thread-safe.
 func (r *Runner) SetContext(ctx Context) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
-	r.ctx = ctx
+	r.ctx = normalizeContext(ctx)
+}
+
+// normalizeContext replaces nil function fields in ctx with no-op stubs so
+// that extension handlers can call any ctx method without a nil-function panic.
+func normalizeContext(ctx Context) Context {
+	if ctx.Print == nil {
+		ctx.Print = func(string) {}
+	}
+	if ctx.PrintInfo == nil {
+		ctx.PrintInfo = func(string) {}
+	}
+	if ctx.PrintError == nil {
+		ctx.PrintError = func(string) {}
+	}
+	if ctx.PrintBlock == nil {
+		ctx.PrintBlock = func(PrintBlockOpts) {}
+	}
+	if ctx.SendMessage == nil {
+		ctx.SendMessage = func(string) {}
+	}
+	if ctx.CancelAndSend == nil {
+		ctx.CancelAndSend = func(string) {}
+	}
+	if ctx.SetWidget == nil {
+		ctx.SetWidget = func(WidgetConfig) {}
+	}
+	if ctx.RemoveWidget == nil {
+		ctx.RemoveWidget = func(string) {}
+	}
+	if ctx.SetHeader == nil {
+		ctx.SetHeader = func(HeaderFooterConfig) {}
+	}
+	if ctx.RemoveHeader == nil {
+		ctx.RemoveHeader = func() {}
+	}
+	if ctx.SetFooter == nil {
+		ctx.SetFooter = func(HeaderFooterConfig) {}
+	}
+	if ctx.RemoveFooter == nil {
+		ctx.RemoveFooter = func() {}
+	}
+	if ctx.PromptSelect == nil {
+		ctx.PromptSelect = func(PromptSelectConfig) PromptSelectResult {
+			return PromptSelectResult{Cancelled: true}
+		}
+	}
+	if ctx.PromptConfirm == nil {
+		ctx.PromptConfirm = func(PromptConfirmConfig) PromptConfirmResult {
+			return PromptConfirmResult{Cancelled: true}
+		}
+	}
+	if ctx.PromptInput == nil {
+		ctx.PromptInput = func(PromptInputConfig) PromptInputResult {
+			return PromptInputResult{Cancelled: true}
+		}
+	}
+	if ctx.PromptMultiSelect == nil {
+		ctx.PromptMultiSelect = func(PromptMultiSelectConfig) PromptMultiSelectResult {
+			return PromptMultiSelectResult{Cancelled: true}
+		}
+	}
+	if ctx.ShowOverlay == nil {
+		ctx.ShowOverlay = func(OverlayConfig) OverlayResult {
+			return OverlayResult{Cancelled: true, Index: -1}
+		}
+	}
+	if ctx.SetEditor == nil {
+		ctx.SetEditor = func(EditorConfig) {}
+	}
+	if ctx.ResetEditor == nil {
+		ctx.ResetEditor = func() {}
+	}
+	if ctx.SetEditorText == nil {
+		ctx.SetEditorText = func(string) {}
+	}
+	if ctx.SetUIVisibility == nil {
+		ctx.SetUIVisibility = func(UIVisibility) {}
+	}
+	if ctx.SetStatus == nil {
+		ctx.SetStatus = func(string, string, int) {}
+	}
+	if ctx.RemoveStatus == nil {
+		ctx.RemoveStatus = func(string) {}
+	}
+	if ctx.GetContextStats == nil {
+		ctx.GetContextStats = func() ContextStats { return ContextStats{} }
+	}
+	if ctx.GetMessages == nil {
+		ctx.GetMessages = func() []SessionMessage { return nil }
+	}
+	if ctx.GetSessionPath == nil {
+		ctx.GetSessionPath = func() string { return "" }
+	}
+	if ctx.AppendEntry == nil {
+		ctx.AppendEntry = func(string, string) (string, error) { return "", nil }
+	}
+	if ctx.GetEntries == nil {
+		ctx.GetEntries = func(string) []ExtensionEntry { return nil }
+	}
+	if ctx.GetOption == nil {
+		ctx.GetOption = func(string) string { return "" }
+	}
+	if ctx.SetOption == nil {
+		ctx.SetOption = func(string, string) {}
+	}
+	if ctx.SetModel == nil {
+		ctx.SetModel = func(string) error { return nil }
+	}
+	if ctx.GetAvailableModels == nil {
+		ctx.GetAvailableModels = func() []ModelInfoEntry { return nil }
+	}
+	if ctx.EmitCustomEvent == nil {
+		ctx.EmitCustomEvent = func(string, string) {}
+	}
+	if ctx.GetAllTools == nil {
+		ctx.GetAllTools = func() []ToolInfo { return nil }
+	}
+	if ctx.SetActiveTools == nil {
+		ctx.SetActiveTools = func([]string) {}
+	}
+	if ctx.Exit == nil {
+		ctx.Exit = func() {}
+	}
+	if ctx.Complete == nil {
+		ctx.Complete = func(CompleteRequest) (CompleteResponse, error) {
+			return CompleteResponse{}, nil
+		}
+	}
+	if ctx.SuspendTUI == nil {
+		ctx.SuspendTUI = func(callback func()) error { callback(); return nil }
+	}
+	if ctx.RenderMessage == nil {
+		ctx.RenderMessage = func(string, string) {}
+	}
+	if ctx.RegisterTheme == nil {
+		ctx.RegisterTheme = func(string, ThemeColorConfig) {}
+	}
+	if ctx.SetTheme == nil {
+		ctx.SetTheme = func(string) error { return nil }
+	}
+	if ctx.ListThemes == nil {
+		ctx.ListThemes = func() []string { return nil }
+	}
+	if ctx.ReloadExtensions == nil {
+		ctx.ReloadExtensions = func() error { return nil }
+	}
+	if ctx.SpawnSubagent == nil {
+		ctx.SpawnSubagent = func(SubagentConfig) (*SubagentHandle, *SubagentResult, error) {
+			return nil, nil, nil
+		}
+	}
+
+	// -------------------------------------------------------------------------
+	// Tree Navigation API no-ops
+	// -------------------------------------------------------------------------
+	if ctx.GetTreeNode == nil {
+		ctx.GetTreeNode = func(string) *TreeNode { return nil }
+	}
+	if ctx.GetCurrentBranch == nil {
+		ctx.GetCurrentBranch = func() []TreeNode { return nil }
+	}
+	if ctx.GetChildren == nil {
+		ctx.GetChildren = func(string) []string { return nil }
+	}
+	if ctx.NavigateTo == nil {
+		ctx.NavigateTo = func(string) TreeNavigationResult {
+			return TreeNavigationResult{Success: false, Error: "not implemented"}
+		}
+	}
+	if ctx.SummarizeBranch == nil {
+		ctx.SummarizeBranch = func(string, string) string {
+			return ""
+		}
+	}
+	if ctx.CollapseBranch == nil {
+		ctx.CollapseBranch = func(string, string, string) TreeNavigationResult {
+			return TreeNavigationResult{Success: false, Error: "not implemented"}
+		}
+	}
+
+	// -------------------------------------------------------------------------
+	// Skill Loading API no-ops
+	// -------------------------------------------------------------------------
+	if ctx.LoadSkill == nil {
+		ctx.LoadSkill = func(string) (*Skill, string) { return nil, "" }
+	}
+	if ctx.LoadSkillsFromDir == nil {
+		ctx.LoadSkillsFromDir = func(string) SkillLoadResult { return SkillLoadResult{} }
+	}
+	if ctx.DiscoverSkills == nil {
+		ctx.DiscoverSkills = func() SkillLoadResult { return SkillLoadResult{} }
+	}
+	if ctx.InjectSkillAsContext == nil {
+		ctx.InjectSkillAsContext = func(string) string { return "" }
+	}
+	if ctx.InjectRawSkillAsContext == nil {
+		ctx.InjectRawSkillAsContext = func(string) string { return "" }
+	}
+	if ctx.GetAvailableSkills == nil {
+		ctx.GetAvailableSkills = func() []Skill { return nil }
+	}
+
+	// -------------------------------------------------------------------------
+	// Template Parsing API no-ops
+	// -------------------------------------------------------------------------
+	if ctx.ParseTemplate == nil {
+		ctx.ParseTemplate = func(string, string) PromptTemplate { return PromptTemplate{} }
+	}
+	if ctx.RenderTemplate == nil {
+		ctx.RenderTemplate = func(PromptTemplate, map[string]string) string { return "" }
+	}
+	if ctx.ParseArguments == nil {
+		ctx.ParseArguments = func(string, ArgumentPattern) ParseResult { return ParseResult{} }
+	}
+	if ctx.SimpleParseArguments == nil {
+		ctx.SimpleParseArguments = func(string, int) []string { return nil }
+	}
+	if ctx.EvaluateModelConditional == nil {
+		ctx.EvaluateModelConditional = func(string) bool { return false }
+	}
+	if ctx.RenderWithModelConditionals == nil {
+		ctx.RenderWithModelConditionals = func(string) string { return "" }
+	}
+
+	// -------------------------------------------------------------------------
+	// Model Resolution API no-ops
+	// -------------------------------------------------------------------------
+	if ctx.ResolveModelChain == nil {
+		ctx.ResolveModelChain = func([]string) ModelResolutionResult {
+			return ModelResolutionResult{Error: "not implemented"}
+		}
+	}
+	if ctx.GetModelCapabilities == nil {
+		ctx.GetModelCapabilities = func(string) (ModelCapabilities, string) {
+			return ModelCapabilities{}, "not implemented"
+		}
+	}
+	if ctx.CheckModelAvailable == nil {
+		ctx.CheckModelAvailable = func(string) bool { return false }
+	}
+	if ctx.GetCurrentProvider == nil {
+		ctx.GetCurrentProvider = func() string { return "" }
+	}
+	if ctx.GetCurrentModelID == nil {
+		ctx.GetCurrentModelID = func() string { return "" }
+	}
+
+	return ctx
 }
 
 // GetContext returns a snapshot of the current runtime context. Thread-safe.

internal/extensions/subagent.go

@@ -173,10 +173,10 @@ type subagentJSONOutput struct {
 	} `json:"usage,omitempty"`
 }
 
-var subagentCounter uint64
+var subagentCounter atomic.Uint64
 
 func generateSubagentID() string {
-	n := atomic.AddUint64(&subagentCounter, 1)
+	n := subagentCounter.Add(1)
 	return fmt.Sprintf("sub-%d-%d", time.Now().UnixNano(), n)
 }
 

internal/extensions/symbols.go

@@ -119,15 +119,39 @@ func Symbols() interp.Exports {
 			"SubagentHandle": reflect.ValueOf((*SubagentHandle)(nil)),
 			"SubagentEvent":  reflect.ValueOf((*SubagentEvent)(nil)),
 
+			// Subagent lifecycle events
+			"SubagentStartEvent": reflect.ValueOf((*SubagentStartEvent)(nil)),
+			"SubagentChunkEvent": reflect.ValueOf((*SubagentChunkEvent)(nil)),
+			"SubagentEndEvent":   reflect.ValueOf((*SubagentEndEvent)(nil)),
+
 			// Theme types
 			"ThemeColor":       reflect.ValueOf((*ThemeColor)(nil)),
 			"ThemeColorConfig": reflect.ValueOf((*ThemeColorConfig)(nil)),
 
+			// Tree navigation types
+			"TreeNode":             reflect.ValueOf((*TreeNode)(nil)),
+			"TreeNavigationResult": reflect.ValueOf((*TreeNavigationResult)(nil)),
+
+			// Skill types
+			"Skill":           reflect.ValueOf((*Skill)(nil)),
+			"SkillLoadResult": reflect.ValueOf((*SkillLoadResult)(nil)),
+
+			// Template parsing types
+			"PromptTemplate":   reflect.ValueOf((*PromptTemplate)(nil)),
+			"ArgumentPattern":  reflect.ValueOf((*ArgumentPattern)(nil)),
+			"ParseResult":      reflect.ValueOf((*ParseResult)(nil)),
+			"ModelConditional": reflect.ValueOf((*ModelConditional)(nil)),
+
+			// Model resolution types
+			"ModelCapabilities":     reflect.ValueOf((*ModelCapabilities)(nil)),
+			"ModelResolutionResult": reflect.ValueOf((*ModelResolutionResult)(nil)),
+
 			// Event structs
 			"ToolCallEvent":           reflect.ValueOf((*ToolCallEvent)(nil)),
 			"ToolCallResult":          reflect.ValueOf((*ToolCallResult)(nil)),
 			"ToolExecutionStartEvent": reflect.ValueOf((*ToolExecutionStartEvent)(nil)),
 			"ToolExecutionEndEvent":   reflect.ValueOf((*ToolExecutionEndEvent)(nil)),
+			"ToolOutputEvent":         reflect.ValueOf((*ToolOutputEvent)(nil)),
 			"ToolResultEvent":         reflect.ValueOf((*ToolResultEvent)(nil)),
 			"ToolResultResult":        reflect.ValueOf((*ToolResultResult)(nil)),
 			"InputEvent":              reflect.ValueOf((*InputEvent)(nil)),

internal/extensions/test_api.go

@@ -30,6 +30,12 @@ func NewTestAPI(ext *LoadedExtension) API {
 				return nil
 			})
 		},
+		onToolOutput: func(h func(ToolOutputEvent, Context)) {
+			reg(ToolOutput, func(e Event, c Context) Result {
+				h(e.(ToolOutputEvent), c)
+				return nil
+			})
+		},
 		onToolResult: func(h func(ToolResultEvent, Context) *ToolResultResult) {
 			reg(ToolResult, func(e Event, c Context) Result {
 				r := h(e.(ToolResultEvent), c)
@@ -165,5 +171,23 @@ func NewTestAPI(ext *LoadedExtension) API {
 		registerMessageRendererFn: func(config MessageRendererConfig) {
 			ext.MessageRenderers = append(ext.MessageRenderers, config)
 		},
+		onSubagentStart: func(h func(SubagentStartEvent, Context)) {
+			reg(SubagentStart, func(e Event, c Context) Result {
+				h(e.(SubagentStartEvent), c)
+				return nil
+			})
+		},
+		onSubagentChunk: func(h func(SubagentChunkEvent, Context)) {
+			reg(SubagentChunk, func(e Event, c Context) Result {
+				h(e.(SubagentChunkEvent), c)
+				return nil
+			})
+		},
+		onSubagentEnd: func(h func(SubagentEndEvent, Context)) {
+			reg(SubagentEnd, func(e Event, c Context) Result {
+				h(e.(SubagentEndEvent), c)
+				return nil
+			})
+		},
 	}
 }

internal/extensions/wrapper.go

@@ -42,14 +42,14 @@ func ExtensionToolsAsFantasy(defs []ToolDef, runner *Runner) []fantasy.AgentTool
 
 // coreToolKinds maps built-in tool names to their kind classification.
 var coreToolKinds = map[string]string{
-	"bash":           "execute",
-	"edit":           "edit",
-	"write":          "edit",
-	"read":           "read",
-	"ls":             "read",
-	"grep":           "search",
-	"find":           "search",
-	"spawn_subagent": "agent",
+	"bash":     "execute",
+	"edit":     "edit",
+	"write":    "edit",
+	"read":     "read",
+	"ls":       "read",
+	"grep":     "search",
+	"find":     "search",
+	"subagent": "agent",
 }
 
 // toolKindFor returns the ToolKind for a given tool name, defaulting to

internal/message/content.go

@@ -4,11 +4,44 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"regexp"
+	"strings"
 	"time"
 
 	"charm.land/fantasy"
 )
 
+// thinkTagRegex matches  ...  tags that some models (Qwen, DeepSeek) wrap
+// reasoning content in. Used to strip these tags from text content.
+// The (?s) flag makes . match newlines.
+var thinkTagRegex = regexp.MustCompile(`(?s)` + `` + `think` + `` + `(.*?)` + `` + `/think` + ``)
+
+// sanitizeToolCallID ensures the ID matches Anthropic's required pattern:
+// ^[a-zA-Z0-9_-]+$ (alphanumeric, underscores, and hyphens only).
+// Invalid characters are replaced with underscores.
+func sanitizeToolCallID(id string) string {
+	var sb strings.Builder
+	for _, r := range id {
+		switch {
+		case (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z'):
+			sb.WriteRune(r)
+		case r >= '0' && r <= '9':
+			sb.WriteRune(r)
+		case r == '_' || r == '-':
+			sb.WriteRune(r)
+		default:
+			// Replace invalid characters with underscore
+			sb.WriteByte('_')
+		}
+	}
+	result := sb.String()
+	// Ensure non-empty (Anthropic requires at least one character)
+	if result == "" {
+		return "tool_0"
+	}
+	return result
+}
+
 // ContentPart is the marker interface for all message content block types.
 // A message contains a heterogeneous slice of ContentPart values, enabling
 // rich structured messages that carry text, reasoning, tool calls, tool
@@ -88,9 +121,9 @@ const (
 )
 
 // Message is a single conversation message containing a heterogeneous slice
-// of ContentPart blocks. This design (borrowed from crush) enables a single
-// assistant message to carry text, reasoning, and multiple tool calls as
-// discrete, typed blocks rather than flattening everything into strings.
+// of ContentPart blocks. This design enables a single assistant message to
+// carry text, reasoning, and multiple tool calls as discrete, typed blocks
+// rather than flattening everything into strings.
 type Message struct {
 	ID        string        `json:"id"`
 	Role      MessageRole   `json:"role"`
@@ -285,12 +318,18 @@ func UnmarshalParts(data []byte) ([]ContentPart, error) {
 	return parts, nil
 }
 
-// --- Fantasy bridge ---
+// --- LLM bridge ---
 
-// ToFantasyMessages converts a Message to one or more fantasy.Message values.
-// An assistant message with tool calls produces a single fantasy message with
+// ToLLMMessages converts a Message to one or more LLM message values.
+// An assistant message with tool calls produces a single message with
 // mixed TextPart and ToolCallPart content. Tool-role messages produce
 // ToolResultPart entries.
+func (m *Message) ToLLMMessages() []fantasy.Message {
+	return m.ToFantasyMessages()
+}
+
+// Deprecated: Use ToLLMMessages instead.
+// ToFantasyMessages converts a Message to one or more LLM message values.
 func (m *Message) ToFantasyMessages() []fantasy.Message {
 	switch m.Role {
 	case RoleAssistant:
@@ -312,7 +351,7 @@ func (m *Message) ToFantasyMessages() []fantasy.Message {
 		// Add tool calls
 		for _, tc := range m.ToolCalls() {
 			parts = append(parts, fantasy.ToolCallPart{
-				ToolCallID: tc.ID,
+				ToolCallID: sanitizeToolCallID(tc.ID),
 				ToolName:   tc.Name,
 				Input:      tc.Input,
 			})
@@ -340,7 +379,7 @@ func (m *Message) ToFantasyMessages() []fantasy.Message {
 				}
 			}
 			parts = append(parts, fantasy.ToolResultPart{
-				ToolCallID: result.ToolCallID,
+				ToolCallID: sanitizeToolCallID(result.ToolCallID),
 				Output:     output,
 			})
 		}
@@ -389,7 +428,14 @@ func (m *Message) ToFantasyMessages() []fantasy.Message {
 	}
 }
 
-// FromFantasyMessage converts a fantasy.Message into our Message type,
+// FromLLMMessage converts an LLM message into our Message type,
+// extracting all content parts into the appropriate block types.
+func FromLLMMessage(msg fantasy.Message) Message {
+	return FromFantasyMessage(msg)
+}
+
+// Deprecated: Use FromLLMMessage instead.
+// FromFantasyMessage converts an LLM message into our Message type,
 // extracting all content parts into the appropriate block types.
 func FromFantasyMessage(msg fantasy.Message) Message {
 	m := Message{
@@ -403,7 +449,11 @@ func FromFantasyMessage(msg fantasy.Message) Message {
 		switch p := part.(type) {
 		case fantasy.TextPart:
 			if p.Text != "" {
-				m.Parts = append(m.Parts, TextContent{Text: p.Text})
+				// Strip  ...  tags that some models wrap reasoning in
+				cleanedText := thinkTagRegex.ReplaceAllString(p.Text, "")
+				if cleanedText != "" {
+					m.Parts = append(m.Parts, TextContent{Text: cleanedText})
+				}
 			}
 		case fantasy.ToolCallPart:
 			m.Parts = append(m.Parts, ToolCall{

internal/message/content_test.go

@@ -0,0 +1,113 @@
+package message
+
+import (
+	"testing"
+)
+
+func TestSanitizeToolCallID(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		{
+			name:     "valid alphanumeric ID",
+			input:    "call_123abc",
+			expected: "call_123abc",
+		},
+		{
+			name:     "ID with dots (OpenCode/Kimi style)",
+			input:    "call.123.abc",
+			expected: "call_123_abc",
+		},
+		{
+			name:     "ID with colons",
+			input:    "tool:123:abc",
+			expected: "tool_123_abc",
+		},
+		{
+			name:     "ID with special characters",
+			input:    "tool@#$%^&*()",
+			expected: "tool_________",
+		},
+		{
+			name:     "Anthropic style ID (already valid)",
+			input:    "toolu_0123456789ABCDEF",
+			expected: "toolu_0123456789ABCDEF",
+		},
+		{
+			name:     "OpenAI style ID (already valid)",
+			input:    "call_O17Uplv4lJvD6DVdIvFFeRMw",
+			expected: "call_O17Uplv4lJvD6DVdIvFFeRMw",
+		},
+		{
+			name:     "ID with hyphens",
+			input:    "my-tool-call-123",
+			expected: "my-tool-call-123",
+		},
+		{
+			name:     "empty string",
+			input:    "",
+			expected: "tool_0",
+		},
+		{
+			name:     "only special characters",
+			input:    "@#$%",
+			expected: "____",
+		},
+		{
+			name:     "mixed valid and invalid",
+			input:    "call_123.abc-def@ghi",
+			expected: "call_123_abc-def_ghi",
+		},
+		{
+			name:     "Unicode characters",
+			input:    "tool_日本語",
+			expected: "tool____",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := sanitizeToolCallID(tt.input)
+			if result != tt.expected {
+				t.Errorf("sanitizeToolCallID(%q) = %q, want %q", tt.input, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestSanitizeToolCallID_MatchesAnthropicPattern(t *testing.T) {
+	// Test that sanitized IDs match Anthropic's required pattern: ^[a-zA-Z0-9_-]+$
+	// This is a simplified check - in reality the pattern allows alphanumeric, underscore, hyphen
+	testIDs := []string{
+		"call.123.abc",
+		"tool:123:def",
+		"id@#$%^&*()",
+		"mixed.valid-id_test",
+		"",
+	}
+
+	for _, id := range testIDs {
+		sanitized := sanitizeToolCallID(id)
+
+		// Verify each character is valid
+		for i, r := range sanitized {
+			valid := (r >= 'a' && r <= 'z') ||
+				(r >= 'A' && r <= 'Z') ||
+				(r >= '0' && r <= '9') ||
+				r == '_' ||
+				r == '-'
+
+			if !valid {
+				t.Errorf("sanitizeToolCallID(%q) = %q, contains invalid character at position %d: %q",
+					id, sanitized, i, string(r))
+			}
+		}
+
+		// Verify non-empty
+		if sanitized == "" {
+			t.Errorf("sanitizeToolCallID(%q) returned empty string", id)
+		}
+	}
+}

internal/models/cache_options.go

@@ -0,0 +1,87 @@
+package models
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"maps"
+	"os"
+
+	"charm.land/fantasy"
+	"charm.land/fantasy/providers/openai"
+)
+
+// buildCacheProviderOptions returns caching options for supported models.
+// Caching is enabled by default for all supported models to reduce costs.
+// Set KIT_DISABLE_CACHE=1 or ProviderConfig.DisableCaching=true to opt out.
+func buildCacheProviderOptions(modelInfo *ModelInfo, config *ProviderConfig) fantasy.ProviderOptions {
+	// Check explicit opt-out via config
+	if config.DisableCaching {
+		return nil
+	}
+
+	// Check global opt-out via environment
+	if os.Getenv("KIT_DISABLE_CACHE") != "" {
+		return nil
+	}
+
+	// Check if model supports caching
+	if modelInfo == nil || !modelInfo.SupportsCaching() {
+		return nil
+	}
+
+	switch modelInfo.CacheType() {
+	case "anthropic-ephemeral":
+		// Provider-level Anthropic caching disabled - use message-level caching instead.
+		return nil
+	case "openai-prompt-cache":
+		return buildOpenAICacheOptions(config, modelInfo.ID)
+	case "google-cached-content":
+		// Google caching not yet implemented.
+		return nil
+	default:
+		return nil
+	}
+}
+
+// buildOpenAICacheOptions enables prompt caching for OpenAI models.
+// Uses a deterministic cache key based on system prompt and model ID.
+func buildOpenAICacheOptions(config *ProviderConfig, modelID string) fantasy.ProviderOptions {
+	cacheKey := generateCacheKey(config.SystemPrompt, modelID)
+
+	return fantasy.ProviderOptions{
+		openai.Name: &openai.ProviderOptions{
+			PromptCacheKey: &cacheKey,
+		},
+	}
+}
+
+// generateCacheKey creates a deterministic cache key from system prompt and model.
+// This ensures the same system prompt + model combination gets cache hits.
+func generateCacheKey(systemPrompt, modelID string) string {
+	if systemPrompt == "" {
+		systemPrompt = "default"
+	}
+
+	h := sha256.New()
+	h.Write([]byte(systemPrompt))
+	h.Write([]byte(modelID))
+
+	// Prefix with "kit-" to identify KIT-generated cache keys
+	return "kit-" + hex.EncodeToString(h.Sum(nil))[:24]
+}
+
+// mergeProviderOptions merges multiple ProviderOptions maps.
+// Later maps take precedence over earlier ones.
+func mergeProviderOptions(opts ...fantasy.ProviderOptions) fantasy.ProviderOptions {
+	result := make(fantasy.ProviderOptions)
+
+	for _, opt := range opts {
+		maps.Copy(result, opt)
+	}
+
+	if len(result) == 0 {
+		return nil
+	}
+
+	return result
+}

internal/models/cache_options_test.go

@@ -0,0 +1,248 @@
+package models
+
+import (
+	"os"
+	"testing"
+
+	"charm.land/fantasy"
+)
+
+func TestModelInfo_SupportsCaching(t *testing.T) {
+	tests := []struct {
+		name     string
+		family   string
+		expected bool
+	}{
+		{"Claude model", "claude-3-5-sonnet", true},
+		{"Claude 4 model", "claude-4-opus", true},
+		{"GPT model", "gpt-4", true},
+		{"GPT-5 model", "gpt-5", true},
+		{"O1 model", "o1", true},
+		{"O3 model", "o3", true},
+		{"O4 model", "o4-mini", true},
+		{"Codex model", "codex", true},
+		{"Gemini model", "gemini-2.5-pro", true},
+		{"Gemini 1.5 model", "gemini-1.5-flash", true},
+		{"Llama model", "llama-3", false},
+		{"Unknown model", "unknown", false},
+		{"Empty family", "", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := &ModelInfo{Family: tt.family}
+			if got := m.SupportsCaching(); got != tt.expected {
+				t.Errorf("ModelInfo.SupportsCaching() = %v, want %v", got, tt.expected)
+			}
+		})
+	}
+}
+
+func TestModelInfo_CacheType(t *testing.T) {
+	tests := []struct {
+		name     string
+		family   string
+		expected string
+	}{
+		{"Claude model", "claude-3-5-sonnet", "anthropic-ephemeral"},
+		{"GPT model", "gpt-4", "openai-prompt-cache"},
+		{"O1 model", "o1", "openai-prompt-cache"},
+		{"Gemini model", "gemini-2.5-pro", "google-cached-content"},
+		{"Unknown model", "llama-3", ""},
+		{"Empty family", "", ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := &ModelInfo{Family: tt.family}
+			if got := m.CacheType(); got != tt.expected {
+				t.Errorf("ModelInfo.CacheType() = %v, want %v", got, tt.expected)
+			}
+		})
+	}
+}
+
+func TestGenerateCacheKey(t *testing.T) {
+	key1 := generateCacheKey("system prompt", "model-id")
+	key2 := generateCacheKey("system prompt", "model-id")
+	if key1 != key2 {
+		t.Errorf("generateCacheKey should be deterministic: got %q and %q", key1, key2)
+	}
+
+	key3 := generateCacheKey("different prompt", "model-id")
+	if key1 == key3 {
+		t.Errorf("generateCacheKey should produce different keys for different inputs")
+	}
+
+	key4 := generateCacheKey("", "model-id")
+	key5 := generateCacheKey("default", "model-id")
+	if key4 != key5 {
+		t.Errorf("generateCacheKey should treat empty prompt as 'default'")
+	}
+
+	if len(key1) < 4 || key1[:4] != "kit-" {
+		t.Errorf("generateCacheKey should produce keys with 'kit-' prefix, got %q", key1)
+	}
+}
+
+func TestBuildCacheProviderOptions_Disabled(t *testing.T) {
+	config := &ProviderConfig{DisableCaching: true}
+	modelInfo := &ModelInfo{Family: "claude-3", ID: "claude-3-opus"}
+
+	if opts := buildCacheProviderOptions(modelInfo, config); opts != nil {
+		t.Errorf("buildCacheProviderOptions should return nil when DisableCaching=true")
+	}
+}
+
+func TestBuildCacheProviderOptions_EnvironmentVariable(t *testing.T) {
+	_ = os.Setenv("KIT_DISABLE_CACHE", "1")
+	defer func() { _ = os.Unsetenv("KIT_DISABLE_CACHE") }()
+
+	config := &ProviderConfig{DisableCaching: false}
+	modelInfo := &ModelInfo{Family: "claude-3", ID: "claude-3-opus"}
+
+	if opts := buildCacheProviderOptions(modelInfo, config); opts != nil {
+		t.Errorf("buildCacheProviderOptions should return nil when KIT_DISABLE_CACHE is set")
+	}
+}
+
+func TestBuildCacheProviderOptions_UnsupportedModel(t *testing.T) {
+	config := &ProviderConfig{DisableCaching: false}
+	modelInfo := &ModelInfo{Family: "llama-3", ID: "llama-3-70b"}
+
+	if opts := buildCacheProviderOptions(modelInfo, config); opts != nil {
+		t.Errorf("buildCacheProviderOptions should return nil for unsupported model families")
+	}
+}
+
+func TestBuildCacheProviderOptions_NilModelInfo(t *testing.T) {
+	config := &ProviderConfig{DisableCaching: false}
+
+	if opts := buildCacheProviderOptions(nil, config); opts != nil {
+		t.Errorf("buildCacheProviderOptions should return nil when modelInfo is nil")
+	}
+}
+
+func TestBuildCacheProviderOptions_Anthropic(t *testing.T) {
+	_ = os.Unsetenv("KIT_DISABLE_CACHE")
+
+	config := &ProviderConfig{DisableCaching: false}
+	modelInfo := &ModelInfo{Family: "claude-3", ID: "claude-3-opus"}
+
+	opts := buildCacheProviderOptions(modelInfo, config)
+	// Provider-level Anthropic caching is disabled; message-level caching is used instead
+	if opts != nil {
+		t.Logf("Provider-level Anthropic caching disabled; using message-level caching")
+	}
+}
+
+func TestBuildCacheProviderOptions_OpenAI(t *testing.T) {
+	_ = os.Unsetenv("KIT_DISABLE_CACHE")
+
+	config := &ProviderConfig{
+		DisableCaching: false,
+		SystemPrompt:   "test system prompt",
+	}
+	modelInfo := &ModelInfo{Family: "gpt-4", ID: "gpt-4o"}
+
+	opts := buildCacheProviderOptions(modelInfo, config)
+	if opts == nil {
+		t.Fatalf("buildCacheProviderOptions should return options for OpenAI models")
+	}
+
+	if _, ok := opts["openai"]; !ok {
+		t.Errorf("buildCacheProviderOptions should include 'openai' key for GPT models")
+	}
+}
+
+func TestCachingPriorityOverThinking(t *testing.T) {
+	_ = os.Unsetenv("KIT_DISABLE_CACHE")
+
+	// Anthropic uses message-level caching; provider-level returns nil
+	config1 := &ProviderConfig{
+		DisableCaching: false,
+		ThinkingLevel:  ThinkingOff,
+	}
+	modelInfo1 := &ModelInfo{Family: "claude-3", ID: "claude-3-opus"}
+	opts1 := buildCacheProviderOptions(modelInfo1, config1)
+	if opts1 != nil {
+		t.Logf("Provider-level Anthropic caching disabled; using message-level caching")
+	}
+
+	// OpenAI provider-level caching works with thinking enabled
+	config2 := &ProviderConfig{
+		DisableCaching: false,
+		SystemPrompt:   "test prompt",
+		ThinkingLevel:  ThinkingMedium,
+	}
+	modelInfo2 := &ModelInfo{Family: "gpt-4", ID: "gpt-4o"}
+	opts2 := buildCacheProviderOptions(modelInfo2, config2)
+	if opts2 == nil {
+		t.Errorf("OpenAI caching should work with thinking enabled")
+	}
+
+	// OpenAI caching also works with thinking disabled
+	config3 := &ProviderConfig{
+		DisableCaching: false,
+		SystemPrompt:   "test prompt",
+		ThinkingLevel:  ThinkingOff,
+	}
+	opts3 := buildCacheProviderOptions(modelInfo2, config3)
+	if opts3 == nil {
+		t.Errorf("OpenAI caching should work when thinking is OFF")
+	}
+}
+
+func TestMergeProviderOptions(t *testing.T) {
+	opts1 := fantasy.ProviderOptions{
+		"provider1": &testProviderData{value: "value1"},
+	}
+	opts2 := fantasy.ProviderOptions{
+		"provider2": &testProviderData{value: "value2"},
+	}
+
+	merged := mergeProviderOptions(opts1, opts2)
+
+	if len(merged) != 2 {
+		t.Errorf("mergeProviderOptions should combine options from multiple maps, got %d items", len(merged))
+	}
+
+	if _, ok := merged["provider1"]; !ok {
+		t.Errorf("merged options should contain 'provider1' key")
+	}
+
+	if _, ok := merged["provider2"]; !ok {
+		t.Errorf("merged options should contain 'provider2' key")
+	}
+
+	// Later options should override earlier ones
+	opts3 := fantasy.ProviderOptions{
+		"provider1": &testProviderData{value: "overridden"},
+	}
+	merged2 := mergeProviderOptions(opts1, opts3)
+
+	if data, ok := merged2["provider1"].(*testProviderData); ok {
+		if data.value != "overridden" {
+			t.Errorf("later options should override earlier ones, got %q", data.value)
+		}
+	}
+
+	if mergeProviderOptions() != nil {
+		t.Errorf("mergeProviderOptions with no args should return nil")
+	}
+}
+
+// testProviderData is a simple implementation of ProviderOptionsData for testing
+type testProviderData struct {
+	value string
+}
+
+func (t *testProviderData) Options() {}
+
+func (t *testProviderData) MarshalJSON() ([]byte, error) {
+	return []byte(`"` + t.value + `"`), nil
+}
+
+func (t *testProviderData) UnmarshalJSON(data []byte) error {
+	return nil
+}

internal/models/custom.go

@@ -0,0 +1,74 @@
+package models
+
+import (
+	"log"
+
+	"github.com/spf13/viper"
+)
+
+// loadCustomModelsFromConfig loads custom model definitions from the config file
+// and returns them as a map of model ID -> ModelInfo. Returns nil if no custom
+// models are configured.
+func loadCustomModelsFromConfig() map[string]ModelInfo {
+	if !viper.IsSet("customModels") {
+		return nil
+	}
+
+	var customModels map[string]CustomModelConfig
+	if err := viper.UnmarshalKey("customModels", &customModels); err != nil {
+		log.Printf("Warning: Failed to parse customModels: %v", err)
+		return nil
+	}
+
+	result := make(map[string]ModelInfo, len(customModels))
+	for modelID, cfg := range customModels {
+		info := modelConfigToModelInfo(modelID, cfg)
+		result[modelID] = info
+	}
+
+	return result
+}
+
+// modelConfigToModelInfo converts a CustomModelConfig to a ModelInfo.
+func modelConfigToModelInfo(modelID string, cfg CustomModelConfig) ModelInfo {
+	return ModelInfo{
+		ID:          modelID,
+		Name:        cfg.Name,
+		Attachment:  cfg.Attachment,
+		Reasoning:   cfg.Reasoning,
+		Temperature: cfg.Temperature,
+		Cost: Cost{
+			Input:  cfg.Cost.Input,
+			Output: cfg.Cost.Output,
+		},
+		Limit: Limit{
+			Context: cfg.Limit.Context,
+			Output:  cfg.Limit.Output,
+		},
+	}
+}
+
+// CustomModelConfig defines a custom model configuration loaded from the config file.
+// This is a duplicate here to avoid circular dependencies with internal/config.
+type CustomModelConfig struct {
+	Name        string      `json:"name" yaml:"name"`
+	Family      string      `json:"family,omitempty" yaml:"family,omitempty"`
+	Attachment  bool        `json:"attachment,omitempty" yaml:"attachment,omitempty"`
+	Reasoning   bool        `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
+	Temperature bool        `json:"temperature,omitempty" yaml:"temperature,omitempty"`
+	Knowledge   string      `json:"knowledge,omitempty" yaml:"knowledge,omitempty"`
+	Cost        CostConfig  `json:"cost" yaml:"cost"`
+	Limit       LimitConfig `json:"limit" yaml:"limit"`
+}
+
+// CostConfig defines the pricing for a custom model.
+type CostConfig struct {
+	Input  float64 `json:"input" yaml:"input"`
+	Output float64 `json:"output" yaml:"output"`
+}
+
+// LimitConfig defines context and output limits for a custom model.
+type LimitConfig struct {
+	Context int `json:"context" yaml:"context"`
+	Output  int `json:"output" yaml:"output"`
+}

internal/models/modelsdb.go

@@ -17,15 +17,21 @@ type modelsDBProvider struct {
 
 // modelsDBModel represents a model entry from models.dev/api.json.
 type modelsDBModel struct {
-	ID          string        `json:"id"`
-	Name        string        `json:"name"`
-	Family      string        `json:"family,omitempty"`
-	Attachment  bool          `json:"attachment"`
-	Reasoning   bool          `json:"reasoning"`
-	ToolCall    bool          `json:"tool_call"`
-	Temperature bool          `json:"temperature"`
-	Cost        modelsDBCost  `json:"cost"`
-	Limit       modelsDBLimit `json:"limit"`
+	ID          string                 `json:"id"`
+	Name        string                 `json:"name"`
+	Family      string                 `json:"family,omitempty"`
+	Attachment  bool                   `json:"attachment"`
+	Reasoning   bool                   `json:"reasoning"`
+	ToolCall    bool                   `json:"tool_call"`
+	Temperature bool                   `json:"temperature"`
+	Cost        modelsDBCost           `json:"cost"`
+	Limit       modelsDBLimit          `json:"limit"`
+	Provider    *modelsDBModelProvider `json:"provider,omitempty"` // Model-specific provider override
+}
+
+// modelsDBModelProvider represents a provider reference within a model.
+type modelsDBModelProvider struct {
+	NPM string `json:"npm"`
 }
 
 // modelsDBCost represents model pricing from models.dev.
@@ -42,10 +48,10 @@ type modelsDBLimit struct {
 	Output  int `json:"output"`
 }
 
-// npmToFantasyProvider maps npm package names from models.dev to fantasy
+// npmToLLMProvider maps npm package names from models.dev to LLM
 // provider identifiers. Providers not in this map but with an api URL
 // can be auto-routed through openaicompat.
-var npmToFantasyProvider = map[string]string{
+var npmToLLMProvider = map[string]string{
 	"@ai-sdk/anthropic":               "anthropic",
 	"@ai-sdk/openai":                  "openai",
 	"@ai-sdk/google":                  "google",

internal/models/providers.go

@@ -10,6 +10,7 @@ import (
 	"maps"
 	"net/http"
 	"os"
+	"regexp"
 	"strings"
 	"time"
 
@@ -22,6 +23,7 @@ import (
 	"charm.land/fantasy/providers/openaicompat"
 	"charm.land/fantasy/providers/openrouter"
 	"charm.land/fantasy/providers/vercel"
+	openaisdk "github.com/charmbracelet/openai-go"
 
 	"github.com/mark3labs/kit/internal/auth"
 	"github.com/mark3labs/kit/internal/ui/progress"
@@ -155,6 +157,7 @@ type ProviderConfig struct {
 	MainGPU        *int32
 	TLSSkipVerify  bool
 	ThinkingLevel  ThinkingLevel
+	DisableCaching bool // Opt-out: set to true to disable automatic prompt caching
 }
 
 // ProviderResult contains the result of provider creation.
@@ -169,6 +172,9 @@ type ProviderResult struct {
 	// ProviderOptions contains provider-specific options to be passed to the
 	// fantasy agent (e.g. OpenAI Responses API reasoning options).
 	ProviderOptions fantasy.ProviderOptions
+	// SkipMaxOutputTokens indicates that this provider doesn't support the
+	// max_output_tokens parameter (e.g., OpenAI Codex OAuth API).
+	SkipMaxOutputTokens bool
 }
 
 // ParseModelString parses a model string in "provider/model" format (e.g. "anthropic/claude-sonnet-4-5").
@@ -234,47 +240,86 @@ func CreateProvider(ctx context.Context, config *ProviderConfig) (*ProviderResul
 		validateModelConfig(config, modelInfo)
 	}
 
+	// Create the base provider
+	var result *ProviderResult
+	var createErr error
+
 	switch provider {
 	case "anthropic":
-		return createAnthropicProvider(ctx, config, modelName)
+		result, createErr = createAnthropicProvider(ctx, config, modelName)
 	case "openai":
-		return createOpenAIProvider(ctx, config, modelName)
+		result, createErr = createOpenAIProvider(ctx, config, modelName)
 	case "google", "gemini":
-		return createGoogleProvider(ctx, config, modelName)
+		result, createErr = createGoogleProvider(ctx, config, modelName)
 	case "ollama":
-		return createOllamaProvider(ctx, config, modelName)
+		result, createErr = createOllamaProvider(ctx, config, modelName)
 	case "azure":
-		return createAzureProvider(ctx, config, modelName)
+		result, createErr = createAzureProvider(ctx, config, modelName)
 	case "google-vertex-anthropic":
-		return createVertexAnthropicProvider(ctx, config, modelName)
+		result, createErr = createVertexAnthropicProvider(ctx, config, modelName)
 	case "openrouter":
-		return createOpenRouterProvider(ctx, config, modelName)
+		result, createErr = createOpenRouterProvider(ctx, config, modelName)
 	case "bedrock":
-		return createBedrockProvider(ctx, config, modelName)
+		result, createErr = createBedrockProvider(ctx, config, modelName)
 	case "vercel":
-		return createVercelProvider(ctx, config, modelName)
+		result, createErr = createVercelProvider(ctx, config, modelName)
+	case "custom":
+		result, createErr = createCustomProvider(ctx, config, modelName)
 	default:
-		return autoRouteProvider(ctx, config, provider, modelName, registry)
+		result, createErr = autoRouteProvider(ctx, config, provider, modelName, registry)
 	}
+
+	if createErr != nil {
+		return nil, createErr
+	}
+
+	// AUTOMATICALLY ENABLE CACHING for supported models (unless disabled).
+	// This works for BOTH native and auto-routed providers by detecting
+	// the model family from the model metadata.
+	if cacheOpts := buildCacheProviderOptions(modelInfo, config); cacheOpts != nil {
+		if result.ProviderOptions == nil {
+			result.ProviderOptions = cacheOpts
+		} else {
+			// Merge cache options with existing provider options.
+			// Only add cache options for providers that don't already have
+			// options set, to avoid type conflicts (e.g., Anthropic has
+			// different types for regular options vs cache control options).
+			for k, v := range cacheOpts {
+				if _, exists := result.ProviderOptions[k]; !exists {
+					result.ProviderOptions[k] = v
+				}
+			}
+		}
+	}
+
+	return result, nil
 }
 
 // autoRouteProvider attempts to create a provider by looking up its npm package
 // in the models.dev database and routing through the appropriate fantasy provider.
 // For openai-compatible providers, it uses the api URL from models.dev.
+// Models may have a provider override that specifies a different npm package than
+// the provider's default (e.g., opencode's claude-opus-4-6 uses @ai-sdk/anthropic).
 func autoRouteProvider(ctx context.Context, config *ProviderConfig, provider, modelName string, registry *ModelsRegistry) (*ProviderResult, error) {
 	providerInfo := registry.GetProviderInfo(provider)
 	if providerInfo == nil {
 		return nil, fmt.Errorf("unsupported provider: %s (not found in model database)", provider)
 	}
 
-	// Determine the fantasy provider for this npm package
-	fantasyProvider := npmToFantasyProvider[providerInfo.NPM]
-	if fantasyProvider == "" && providerInfo.API != "" {
-		// Unknown npm but has API URL → route through openaicompat
-		fantasyProvider = "openaicompat"
+	// Check for model-specific provider override
+	npmPackage := providerInfo.NPM
+	if modelInfo := registry.LookupModel(provider, modelName); modelInfo != nil && modelInfo.ProviderNPM != "" {
+		npmPackage = modelInfo.ProviderNPM
 	}
 
-	switch fantasyProvider {
+	// Determine the LLM provider for this npm package
+	llmProvider := npmToLLMProvider[npmPackage]
+	if llmProvider == "" && providerInfo.API != "" {
+		// Unknown npm but has API URL → route through openaicompat
+		llmProvider = "openaicompat"
+	}
+
+	switch llmProvider {
 	case "openaicompat":
 		return createAutoRoutedOpenAICompatProvider(ctx, config, modelName, providerInfo)
 	case "anthropic":
@@ -288,7 +333,7 @@ func autoRouteProvider(ctx context.Context, config *ProviderConfig, provider, mo
 		}
 		return createAutoRoutedOpenAIProvider(ctx, config, modelName, providerInfo)
 	default:
-		return nil, fmt.Errorf("unsupported provider: %s (npm: %s has no fantasy mapping)", provider, providerInfo.NPM)
+		return nil, fmt.Errorf("unsupported provider: %s (npm: %s has no LLM provider mapping)", provider, npmPackage)
 	}
 }
 
@@ -346,7 +391,10 @@ func createAutoRoutedAnthropicProvider(ctx context.Context, config *ProviderConf
 	opts = append(opts, anthropic.WithAPIKey(apiKey))
 
 	if config.ProviderURL != "" {
-		opts = append(opts, anthropic.WithBaseURL(config.ProviderURL))
+		// The anthropic client appends "/v1/messages" to the base URL.
+		// If the provider URL ends with "/v1", strip it to avoid double "/v1/v1" paths.
+		baseURL := strings.TrimSuffix(config.ProviderURL, "/v1")
+		opts = append(opts, anthropic.WithBaseURL(baseURL))
 	}
 
 	if config.TLSSkipVerify {
@@ -494,10 +542,15 @@ func thinkingLevelToReasoningEffort(level ThinkingLevel) *openai.ReasoningEffort
 // SendReasoning to true and configures the thinking budget. For thinking-off
 // or non-reasoning models the returned map is nil.
 //
+// NOTE: With message-level caching, thinking and caching can work together.
+// Message-level cache control (ProviderCacheControlOptions) doesn't conflict
+// with provider-level thinking options (ProviderOptions).
+//
 // Anthropic requires max_tokens > thinking.budget_tokens. If the configured
 // MaxTokens is too low, it is bumped to budget + 4096 to leave room for the
 // actual response.
 func buildAnthropicProviderOptions(config *ProviderConfig, modelName string) fantasy.ProviderOptions {
+	// Thinking is OFF by default. If user hasn't explicitly enabled it, return nil.
 	if config.ThinkingLevel == "" || config.ThinkingLevel == ThinkingOff {
 		return nil
 	}
@@ -608,13 +661,52 @@ func createVertexAnthropicProvider(ctx context.Context, config *ProviderConfig,
 
 func createOpenAIProvider(ctx context.Context, config *ProviderConfig, modelName string) (*ProviderResult, error) {
 	apiKey := config.ProviderAPIKey
+	source := "command-line flag"
+	var accountID string
+	var isCodexOAuth bool
+
 	if apiKey == "" {
-		apiKey = os.Getenv("OPENAI_API_KEY")
-	}
-	if apiKey == "" {
-		return nil, fmt.Errorf("OpenAI API key not provided. Use --provider-api-key flag or OPENAI_API_KEY environment variable")
+		// Check stored credentials first
+		cm, err := auth.NewCredentialManager()
+		if err == nil {
+			if creds, err := cm.GetOpenAICredentials(); err == nil && creds != nil {
+				if creds.Type == "oauth" && creds.AccessToken != "" {
+					// For OAuth, get a valid access token (may refresh if needed)
+					token, err := cm.GetValidOpenAIAccessToken()
+					if err == nil && token != "" {
+						apiKey = token
+						accountID = creds.AccountID
+						isCodexOAuth = true
+						source = "stored Codex OAuth credentials"
+					}
+				} else if creds.Type == "api_key" && creds.APIKey != "" {
+					apiKey = creds.APIKey
+					source = "stored API key"
+				}
+			}
+		}
 	}
 
+	// Fall back to environment variable
+	if apiKey == "" {
+		apiKey = os.Getenv("OPENAI_API_KEY")
+		source = "OPENAI_API_KEY environment variable"
+	}
+
+	if apiKey == "" {
+		return nil, fmt.Errorf("OpenAI API key not provided. Use 'kit auth login openai', --provider-api-key flag, or OPENAI_API_KEY environment variable")
+	}
+
+	if os.Getenv("DEBUG") != "" || os.Getenv("KIT_DEBUG") != "" {
+		fmt.Fprintf(os.Stderr, "Using OpenAI API key from: %s\n", source)
+	}
+
+	// For Codex OAuth, use the ChatGPT backend API with custom headers
+	if isCodexOAuth {
+		return createOpenAICodexProvider(ctx, config, modelName, apiKey, accountID)
+	}
+
+	// Regular OpenAI API key flow
 	var opts []openai.Option
 	opts = append(opts, openai.WithAPIKey(apiKey))
 	opts = append(opts, openai.WithUseResponsesAPI())
@@ -643,6 +735,135 @@ func createOpenAIProvider(ctx context.Context, config *ProviderConfig, modelName
 	return &ProviderResult{Model: model, ProviderOptions: providerOpts}, nil
 }
 
+// createOpenAICodexProvider creates a provider for ChatGPT/Codex OAuth tokens.
+// Uses the chatgpt.com/backend-api/codex endpoint with special headers.
+func createOpenAICodexProvider(ctx context.Context, config *ProviderConfig, modelName, token, accountID string) (*ProviderResult, error) {
+	// Check for spark models which are not accessible via OAuth
+	if detectCodexModelFamily(modelName) == "gpt-codex-spark" {
+		return nil, fmt.Errorf("gpt-codex-spark models are not accessible via ChatGPT OAuth. " +
+			"These models require special access or a different authentication method. " +
+			"Please use regular Codex models like 'openai/gpt-5.3-codex' instead")
+	}
+
+	// Use the ChatGPT backend API with /codex path
+	baseURL := "https://chatgpt.com/backend-api/codex"
+	if config.ProviderURL != "" {
+		baseURL = config.ProviderURL
+	}
+
+	// Build custom HTTP client with required headers
+	httpClient := createCodexHTTPClient(token, accountID, config.TLSSkipVerify)
+
+	var opts []openai.Option
+	opts = append(opts, openai.WithAPIKey(token))
+	opts = append(opts, openai.WithBaseURL(baseURL))
+	opts = append(opts, openai.WithUseResponsesAPI())
+	opts = append(opts, openai.WithHTTPClient(httpClient))
+
+	provider, err := openai.New(opts...)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create OpenAI Codex provider: %w", err)
+	}
+
+	model, err := provider.LanguageModel(ctx, modelName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create OpenAI Codex model: %w", err)
+	}
+
+	providerOpts := buildCodexProviderOptions(config, modelName)
+
+	return &ProviderResult{
+		Model:               model,
+		ProviderOptions:     providerOpts,
+		SkipMaxOutputTokens: true,
+	}, nil
+}
+
+// buildCodexProviderOptions returns fantasy.ProviderOptions configured for
+// OpenAI Codex API. The Codex API requires the system prompt to be passed
+// as 'instructions' rather than as a system message.
+func buildCodexProviderOptions(config *ProviderConfig, modelName string) fantasy.ProviderOptions {
+	store := false
+	opts := &openai.ResponsesProviderOptions{
+		Store: &store,
+	}
+
+	if config.SystemPrompt != "" {
+		opts.Instructions = &config.SystemPrompt
+	}
+
+	if openai.IsResponsesReasoningModel(modelName) {
+		opts.ReasoningEffort = thinkingLevelToReasoningEffort(config.ThinkingLevel)
+	}
+
+	return fantasy.ProviderOptions{openai.Name: opts}
+}
+
+// detectCodexModelFamily determines the model family from the model name
+func detectCodexModelFamily(modelName string) string {
+	modelName = strings.ToLower(modelName)
+	if strings.Contains(modelName, "spark") {
+		return "gpt-codex-spark"
+	}
+	if strings.Contains(modelName, "codex-mini") || strings.Contains(modelName, "mini-latest") {
+		return "gpt-codex-mini"
+	}
+	if strings.Contains(modelName, "codex") {
+		return "gpt-codex"
+	}
+	return ""
+}
+
+// createCodexHTTPClient creates an HTTP client with headers required for ChatGPT/Codex API
+func createCodexHTTPClient(token, accountID string, skipVerify bool) *http.Client {
+	var base http.RoundTripper
+	if skipVerify {
+		base = &http.Transport{
+			TLSClientConfig: &tls.Config{
+				InsecureSkipVerify: true,
+			},
+		}
+	} else {
+		base = http.DefaultTransport
+	}
+
+	return &http.Client{
+		Transport: &codexTransport{
+			base:      base,
+			token:     token,
+			accountID: accountID,
+		},
+		Timeout: 120 * time.Second,
+	}
+}
+
+// codexTransport is a custom RoundTripper that adds ChatGPT/Codex specific headers
+type codexTransport struct {
+	base      http.RoundTripper
+	token     string
+	accountID string
+}
+
+func (t *codexTransport) RoundTrip(req *http.Request) (*http.Response, error) {
+	newReq := req.Clone(req.Context())
+
+	// Add required headers for ChatGPT/Codex API
+	// These headers mimic the official pi client to avoid Cloudflare blocking
+	newReq.Header.Set("Authorization", "Bearer "+t.token)
+	if t.accountID != "" {
+		newReq.Header.Set("chatgpt-account-id", t.accountID)
+	}
+	newReq.Header.Set("originator", "kit")
+	newReq.Header.Set("User-Agent", "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36")
+	newReq.Header.Set("OpenAI-Beta", "responses=experimental")
+	newReq.Header.Set("Accept", "text/event-stream")
+	newReq.Header.Set("Accept-Language", "en-US,en;q=0.9")
+	newReq.Header.Set("Cache-Control", "no-cache")
+	newReq.Header.Set("Pragma", "no-cache")
+
+	return t.base.RoundTrip(newReq)
+}
+
 func createGoogleProvider(ctx context.Context, config *ProviderConfig, modelName string) (*ProviderResult, error) {
 	apiKey := firstNonEmpty(
 		config.ProviderAPIKey,
@@ -779,6 +1000,176 @@ func createVercelProvider(ctx context.Context, config *ProviderConfig, modelName
 	return &ProviderResult{Model: model}, nil
 }
 
+// thinkTagRegex matches <think>...</think> tags for extracting reasoning content
+// from models that wrap thinking in XML-like tags (e.g., Qwen, DeepSeek).
+var thinkTagRegex = regexp.MustCompile(`(?s)<think>(.*?)</think>`)
+
+// customExtraContentFunc extracts reasoning from <think> tags in the content field.
+// This handles models like Qwen and DeepSeek that return reasoning wrapped in XML tags
+// rather than using a separate reasoning_content field.
+func customExtraContentFunc(choice openaisdk.ChatCompletionChoice) []fantasy.Content {
+	var content []fantasy.Content
+	if choice.Message.Content == "" {
+		return content
+	}
+
+	// Check for <think> tags in the content
+	matches := thinkTagRegex.FindStringSubmatch(choice.Message.Content)
+	if len(matches) > 1 {
+		// Found reasoning content in <think> tags
+		reasoning := strings.TrimSpace(matches[1])
+		if reasoning != "" {
+			content = append(content, fantasy.ReasoningContent{
+				Text: reasoning,
+			})
+		}
+	}
+
+	return content
+}
+
+// customStreamExtraFunc handles streaming responses with <think> tags.
+// It extracts reasoning content and emits proper reasoning events.
+func customStreamExtraFunc(
+	chunk openaisdk.ChatCompletionChunk,
+	yield func(fantasy.StreamPart) bool,
+	ctx map[string]any,
+) (map[string]any, bool) {
+	if len(chunk.Choices) == 0 {
+		return ctx, true
+	}
+
+	const reasoningStartedKey = "reasoning_started"
+	const reasoningBufferKey = "reasoning_buffer"
+	const inThinkTagKey = "in_think_tag"
+
+	reasoningStarted, _ := ctx[reasoningStartedKey].(bool)
+	inThinkTag, _ := ctx[inThinkTagKey].(bool)
+	reasoningBuffer, _ := ctx[reasoningBufferKey].(string)
+
+	for i, choice := range chunk.Choices {
+		content := choice.Delta.Content
+		if content == "" {
+			continue
+		}
+
+		// Check for <think> tag start
+		if strings.Contains(content, "<think>") {
+			inThinkTag = true
+			ctx[inThinkTagKey] = true
+
+			// Emit reasoning start event
+			if !reasoningStarted {
+				reasoningStarted = true
+				ctx[reasoningStartedKey] = true
+				if !yield(fantasy.StreamPart{
+					Type: fantasy.StreamPartTypeReasoningStart,
+					ID:   fmt.Sprintf("%d", i),
+				}) {
+					return ctx, false
+				}
+			}
+
+			// Extract content after <think>
+			parts := strings.SplitN(content, "<think>", 2)
+			if len(parts) > 1 && parts[1] != "" {
+				reasoningBuffer += parts[1]
+				ctx[reasoningBufferKey] = reasoningBuffer
+			}
+			continue
+		}
+
+		// Check for </think> tag end
+		if strings.Contains(content, "</think>") {
+			inThinkTag = false
+			ctx[inThinkTagKey] = false
+
+			// Extract content before </think>
+			parts := strings.SplitN(content, "</think>", 2)
+			if len(parts) > 0 {
+				reasoningBuffer += parts[0]
+			}
+
+			// Emit the accumulated reasoning
+			if reasoningBuffer != "" {
+				if !yield(fantasy.StreamPart{
+					Type:  fantasy.StreamPartTypeReasoningDelta,
+					ID:    fmt.Sprintf("%d", i),
+					Delta: reasoningBuffer,
+				}) {
+					return ctx, false
+				}
+				ctx[reasoningBufferKey] = ""
+			}
+
+			// Emit reasoning end
+			if !yield(fantasy.StreamPart{
+				Type: fantasy.StreamPartTypeReasoningEnd,
+				ID:   fmt.Sprintf("%d", i),
+			}) {
+				return ctx, false
+			}
+			continue
+		}
+
+		// Accumulate reasoning content while in think tag
+		if inThinkTag {
+			reasoningBuffer += content
+			ctx[reasoningBufferKey] = reasoningBuffer
+		}
+	}
+
+	return ctx, true
+}
+
+// customToPromptFunc converts prompts to OpenAI format using the default conversion.
+func customToPromptFunc(prompt fantasy.Prompt, systemPrompt, user string) ([]openaisdk.ChatCompletionMessageParamUnion, []fantasy.CallWarning) {
+	return openai.DefaultToPrompt(prompt, systemPrompt, user)
+}
+
+func createCustomProvider(ctx context.Context, config *ProviderConfig, modelName string) (*ProviderResult, error) {
+	if config.ProviderURL == "" {
+		return nil, fmt.Errorf("custom provider requires --provider-url")
+	}
+
+	apiKey := config.ProviderAPIKey
+	if apiKey == "" {
+		apiKey = os.Getenv("CUSTOM_API_KEY")
+	}
+	if apiKey == "" {
+		// Many local/custom endpoints don't require a key; use a placeholder.
+		apiKey = "custom"
+	}
+
+	// Use the openai provider directly with custom hooks to handle <think> tags
+	// from models like Qwen and DeepSeek that wrap reasoning in XML tags.
+	var opts []openai.Option
+	opts = append(opts, openai.WithBaseURL(config.ProviderURL))
+	opts = append(opts, openai.WithAPIKey(apiKey))
+	opts = append(opts, openai.WithName("custom"))
+	opts = append(opts, openai.WithLanguageModelOptions(
+		openai.WithLanguageModelExtraContentFunc(customExtraContentFunc),
+		openai.WithLanguageModelStreamExtraFunc(customStreamExtraFunc),
+		openai.WithLanguageModelToPromptFunc(customToPromptFunc),
+	))
+
+	if config.TLSSkipVerify {
+		opts = append(opts, openai.WithHTTPClient(createHTTPClientWithTLSConfig(true)))
+	}
+
+	p, err := openai.New(opts...)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create custom provider: %w", err)
+	}
+
+	model, err := p.LanguageModel(ctx, modelName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create custom model: %w", err)
+	}
+
+	return &ProviderResult{Model: model}, nil
+}
+
 func createOllamaProvider(ctx context.Context, config *ProviderConfig, modelName string) (*ProviderResult, error) {
 	baseURL := "http://localhost:11434"
 	if host := os.Getenv("OLLAMA_HOST"); host != "" {

internal/models/registry.go

@@ -17,11 +17,51 @@ var embeddedModelsJSON []byte
 type ModelInfo struct {
 	ID          string
 	Name        string
+	Family      string // Model family (e.g., "claude", "gpt", "gemini")
 	Attachment  bool
 	Reasoning   bool
 	Temperature bool
 	Cost        Cost
 	Limit       Limit
+	ProviderNPM string // Model-specific provider npm override (e.g. "@ai-sdk/anthropic")
+}
+
+// SupportsCaching returns true if this model family supports prompt caching.
+// This enables automatic cost savings for supported models regardless of provider.
+func (m *ModelInfo) SupportsCaching() bool {
+	switch {
+	case strings.HasPrefix(m.Family, "claude"):
+		return true
+	case strings.HasPrefix(m.Family, "gpt"),
+		strings.HasPrefix(m.Family, "o1"),
+		strings.HasPrefix(m.Family, "o3"),
+		strings.HasPrefix(m.Family, "o4"),
+		strings.HasPrefix(m.Family, "codex"):
+		return true
+	case strings.HasPrefix(m.Family, "gemini"):
+		return true
+	default:
+		return false
+	}
+}
+
+// CacheType returns the appropriate cache mechanism for this model family.
+// Returns empty string if caching is not supported.
+func (m *ModelInfo) CacheType() string {
+	switch {
+	case strings.HasPrefix(m.Family, "claude"):
+		return "anthropic-ephemeral"
+	case strings.HasPrefix(m.Family, "gpt"),
+		strings.HasPrefix(m.Family, "o1"),
+		strings.HasPrefix(m.Family, "o3"),
+		strings.HasPrefix(m.Family, "o4"),
+		strings.HasPrefix(m.Family, "codex"):
+		return "openai-prompt-cache"
+	case strings.HasPrefix(m.Family, "gemini"):
+		return "google-cached-content"
+	default:
+		return ""
+	}
 }
 
 // Cost represents the pricing information for a model.
@@ -78,9 +118,14 @@ func buildFromModelsDB() map[string]ProviderInfo {
 	for providerID, dp := range dbProviders {
 		modelsMap := make(map[string]ModelInfo, len(dp.Models))
 		for modelID, dm := range dp.Models {
+			providerNPM := ""
+			if dm.Provider != nil {
+				providerNPM = dm.Provider.NPM
+			}
 			modelsMap[modelID] = ModelInfo{
 				ID:          dm.ID,
 				Name:        dm.Name,
+				Family:      dm.Family,
 				Attachment:  dm.Attachment,
 				Reasoning:   dm.Reasoning,
 				Temperature: dm.Temperature,
@@ -94,6 +139,7 @@ func buildFromModelsDB() map[string]ProviderInfo {
 					Context: dm.Limit.Context,
 					Output:  dm.Limit.Output,
 				},
+				ProviderNPM: providerNPM,
 			}
 		}
 
@@ -116,6 +162,47 @@ func buildFromModelsDB() map[string]ProviderInfo {
 		}
 	}
 
+	// Register the "custom" provider stub for --provider-url without --model.
+	// This allows users to point kit at any OpenAI-compatible endpoint without
+	// needing to specify a model from the database.
+	providers["custom"] = ProviderInfo{
+		ID:   "custom",
+		Name: "Custom",
+		Models: map[string]ModelInfo{
+			"custom": {
+				ID:          "custom",
+				Name:        "Custom",
+				Attachment:  false,
+				Reasoning:   true,
+				Temperature: true,
+				Cost: Cost{
+					Input:  0,
+					Output: 0,
+				},
+				Limit: Limit{
+					Context: 262_144,
+					Output:  65_536,
+				},
+			},
+		},
+	}
+
+	// Load custom models from config file and merge into custom provider.
+	// Config file models take precedence - if a model ID exists in both
+	// models.dev and config, the config version wins.
+	if customModels := loadCustomModelsFromConfig(); customModels != nil {
+		for modelID, info := range customModels {
+			// Validate custom model config
+			if info.Limit.Context <= 0 {
+				fmt.Fprintf(os.Stderr, "Warning: custom model %q has invalid context limit: %d\n", modelID, info.Limit.Context)
+			}
+			if info.Limit.Output <= 0 {
+				fmt.Fprintf(os.Stderr, "Warning: custom model %q has invalid output limit: %d\n", modelID, info.Limit.Output)
+			}
+			providers["custom"].Models[modelID] = info
+		}
+	}
+
 	return providers
 }
 
@@ -178,6 +265,15 @@ func (r *ModelsRegistry) ValidateEnvironment(provider string, apiKey string) err
 		}
 	}
 
+	// For openai, check stored credentials (OAuth / API key)
+	if provider == "openai" {
+		if cm, err := auth.NewCredentialManager(); err == nil {
+			if has, _ := cm.HasOpenAICredentials(); has {
+				return nil
+			}
+		}
+	}
+
 	envVars, err := r.getRequiredEnvVars(provider)
 	if err != nil {
 		// Unknown provider — nothing to validate
@@ -252,27 +348,32 @@ func (r *ModelsRegistry) GetSupportedProviders() []string {
 	return providers
 }
 
-// GetFantasyProviders returns provider IDs that can be used with fantasy,
+// GetLLMProviders returns provider IDs that have LLM support,
 // either through a native provider or via openaicompat auto-routing.
-func (r *ModelsRegistry) GetFantasyProviders() []string {
+func (r *ModelsRegistry) GetLLMProviders() []string {
 	var providers []string
 	for providerID, info := range r.providers {
-		if isProviderFantasySupported(providerID, &info) {
+		if isProviderLLMSupported(providerID, &info) {
 			providers = append(providers, providerID)
 		}
 	}
 	return providers
 }
 
-// isProviderFantasySupported checks if a provider can be used with fantasy.
-func isProviderFantasySupported(providerID string, info *ProviderInfo) bool {
+// Deprecated: Use GetLLMProviders instead.
+func (r *ModelsRegistry) GetFantasyProviders() []string {
+	return r.GetLLMProviders()
+}
+
+// isProviderLLMSupported checks if a provider can be used with the LLM layer.
+func isProviderLLMSupported(providerID string, info *ProviderInfo) bool {
 	// Ollama is always supported (via openaicompat pointed at localhost)
 	if providerID == "ollama" {
 		return true
 	}
 
-	// Check if npm maps to a fantasy provider
-	if _, ok := npmToFantasyProvider[info.NPM]; ok {
+	// Check if npm maps to an LLM provider
+	if _, ok := npmToLLMProvider[info.NPM]; ok {
 		return true
 	}
 

internal/prompts/frontmatter.go

@@ -0,0 +1,25 @@
+package prompts
+
+import (
+	"fmt"
+
+	"gopkg.in/yaml.v3"
+)
+
+// frontmatterSep is the YAML frontmatter delimiter.
+const frontmatterSep = "---"
+
+// Frontmatter represents the YAML frontmatter in a prompt template file.
+type Frontmatter struct {
+	// Description summarises what this template provides.
+	Description string `yaml:"description"`
+}
+
+// ParseFrontmatter parses YAML frontmatter content into a Frontmatter struct.
+func ParseFrontmatter(content string) (*Frontmatter, error) {
+	var fm Frontmatter
+	if err := yaml.Unmarshal([]byte(content), &fm); err != nil {
+		return nil, fmt.Errorf("parsing frontmatter: %w", err)
+	}
+	return &fm, nil
+}

internal/prompts/loader.go

@@ -0,0 +1,217 @@
+package prompts
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/charmbracelet/log"
+)
+
+// LoadOptions configures how templates are discovered and loaded.
+type LoadOptions struct {
+	// Cwd is the current working directory for project-local discovery.
+	// If empty, the current working directory is used.
+	Cwd string
+	// HomeDir is the user's home directory. If empty, os.UserHomeDir() is used.
+	HomeDir string
+	// ExtraPaths are additional explicit paths to search for templates.
+	ExtraPaths []string
+	// ConfigPaths are paths from configuration files to search.
+	ConfigPaths []string
+	// IncludeDefaults determines whether to include built-in default templates.
+	IncludeDefaults bool
+}
+
+// Diagnostic reports a template collision or loading issue.
+type Diagnostic struct {
+	// Name is the template name that had a collision.
+	Name string
+	// KeptPath is the path of the template that was kept (higher precedence).
+	KeptPath string
+	// DroppedPath is the path of the template that was dropped.
+	DroppedPath string
+	// Reason explains why the collision occurred.
+	Reason string
+}
+
+// LoadAll discovers and loads all prompt templates from standard locations
+// and any extra paths. Templates are loaded in order of precedence (lowest
+// to highest), with later templates overriding earlier ones of the same name.
+//
+// Discovery paths searched in order:
+//  1. Default templates (if IncludeDefaults)
+//  2. ~/.kit/prompts/ (global user templates)
+//  3. .kit/prompts/ (project-local templates)
+//  4. ConfigPaths (from configuration)
+//  5. ExtraPaths (explicit paths, highest precedence)
+func LoadAll(opts LoadOptions) ([]*PromptTemplate, []Diagnostic, error) {
+	if opts.Cwd == "" {
+		opts.Cwd, _ = os.Getwd()
+	}
+
+	if opts.HomeDir == "" {
+		home, err := os.UserHomeDir()
+		if err != nil {
+			return nil, nil, fmt.Errorf("getting home directory: %w", err)
+		}
+		opts.HomeDir = home
+	}
+
+	var all []*PromptTemplate
+	var diagnostics []Diagnostic
+	seen := make(map[string]*PromptTemplate) // name -> template
+
+	// Helper to add templates with deduplication tracking
+	addTemplates := func(templates []*PromptTemplate, source string) {
+		for _, tpl := range templates {
+			if existing, ok := seen[tpl.Name]; ok {
+				// Collision: report diagnostic, keep existing (lower precedence wins)
+				diagnostics = append(diagnostics, Diagnostic{
+					Name:        tpl.Name,
+					KeptPath:    existing.FilePath,
+					DroppedPath: tpl.FilePath,
+					Reason:      fmt.Sprintf("template from %s overridden by %s", source, existing.Source),
+				})
+				log.Debug("template collision",
+					"name", tpl.Name,
+					"dropped", tpl.FilePath,
+					"kept", existing.FilePath)
+			} else {
+				tpl.Source = source
+				seen[tpl.Name] = tpl
+				all = append(all, tpl)
+			}
+		}
+	}
+
+	// 1. Default templates (lowest precedence)
+	if opts.IncludeDefaults {
+		defaults := loadDefaultTemplates()
+		addTemplates(defaults, "default")
+	}
+
+	// 2. Global user templates: ~/.kit/prompts/
+	globalDir := filepath.Join(opts.HomeDir, ".kit", "prompts")
+	if templates, err := LoadFromDir(globalDir); err == nil {
+		addTemplates(templates, "global")
+	}
+
+	// 3. Project-local templates: .kit/prompts/
+	localDir := filepath.Join(opts.Cwd, ".kit", "prompts")
+	if templates, err := LoadFromDir(localDir); err == nil {
+		addTemplates(templates, "local")
+	}
+
+	// 4. Config paths
+	for _, path := range opts.ConfigPaths {
+		info, err := os.Stat(path)
+		if err != nil {
+			continue
+		}
+		if info.IsDir() {
+			if templates, err := LoadFromDir(path); err == nil {
+				addTemplates(templates, "config")
+			}
+		} else if strings.HasSuffix(path, ".md") {
+			if tpl, err := ParseTemplate(path); err == nil {
+				addTemplates([]*PromptTemplate{tpl}, "config")
+			}
+		}
+	}
+
+	// 5. Extra paths (highest precedence)
+	for _, path := range opts.ExtraPaths {
+		info, err := os.Stat(path)
+		if err != nil {
+			continue
+		}
+		if info.IsDir() {
+			if templates, err := LoadFromDir(path); err == nil {
+				addTemplates(templates, "explicit")
+			}
+		} else if strings.HasSuffix(path, ".md") {
+			if tpl, err := ParseTemplate(path); err == nil {
+				addTemplates([]*PromptTemplate{tpl}, "explicit")
+			}
+		}
+	}
+
+	return all, diagnostics, nil
+}
+
+// LoadFromDir scans a directory for .md files and loads them as templates.
+// It looks for *.md files directly in the directory.
+// Files that fail to parse are logged and skipped.
+func LoadFromDir(dir string) ([]*PromptTemplate, error) {
+	info, err := os.Stat(dir)
+	if err != nil || !info.IsDir() {
+		return nil, nil // directory doesn't exist — not an error
+	}
+
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return nil, fmt.Errorf("reading prompts directory %s: %w", dir, err)
+	}
+
+	var templates []*PromptTemplate
+	var errs []string
+
+	for _, entry := range entries {
+		if entry.IsDir() {
+			continue
+		}
+
+		name := entry.Name()
+		if !strings.HasSuffix(name, ".md") {
+			continue
+		}
+
+		full := filepath.Join(dir, name)
+		tpl, err := ParseTemplate(full)
+		if err != nil {
+			errs = append(errs, err.Error())
+			continue
+		}
+		templates = append(templates, tpl)
+	}
+
+	if len(errs) > 0 {
+		return templates, fmt.Errorf("some templates failed to load: %s", strings.Join(errs, "; "))
+	}
+	return templates, nil
+}
+
+// Deduplicate removes duplicate templates by name, keeping the first occurrence.
+// It returns the deduplicated list and diagnostics for any collisions.
+// This is a standalone function for when you need to deduplicate an existing list.
+func Deduplicate(templates []*PromptTemplate) ([]*PromptTemplate, []Diagnostic) {
+	seen := make(map[string]*PromptTemplate)
+	var result []*PromptTemplate
+	var diagnostics []Diagnostic
+
+	for _, tpl := range templates {
+		if existing, ok := seen[tpl.Name]; ok {
+			diagnostics = append(diagnostics, Diagnostic{
+				Name:        tpl.Name,
+				KeptPath:    existing.FilePath,
+				DroppedPath: tpl.FilePath,
+				Reason:      "duplicate template name (first-match-wins)",
+			})
+		} else {
+			seen[tpl.Name] = tpl
+			result = append(result, tpl)
+		}
+	}
+
+	return result, diagnostics
+}
+
+// loadDefaultTemplates returns the built-in default templates.
+// These are embedded templates that ship with Kit.
+func loadDefaultTemplates() []*PromptTemplate {
+	// Default templates can be added here as needed
+	// For now, return an empty slice - users can define their own templates
+	return nil
+}

internal/prompts/loader_test.go

@@ -0,0 +1,126 @@
+package prompts
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestLoadAll_Integration(t *testing.T) {
+	// Create a temp directory for testing
+	tempDir := t.TempDir()
+
+	// Create the .kit/prompts subdirectory structure
+	promptsDir := filepath.Join(tempDir, ".kit", "prompts")
+	if err := os.MkdirAll(promptsDir, 0755); err != nil {
+		t.Fatalf("Failed to create prompts dir: %v", err)
+	}
+
+	// Create a test template file
+	templateContent := `---
+description: Test template for integration
+---
+Review $1 with focus on $2`
+
+	testFile := filepath.Join(promptsDir, "test.md")
+	if err := os.WriteFile(testFile, []byte(templateContent), 0644); err != nil {
+		t.Fatalf("Failed to create test file: %v", err)
+	}
+
+	// Test loading from the temp directory
+	tpls, diags, err := LoadAll(LoadOptions{
+		HomeDir:         tempDir,
+		IncludeDefaults: false, // Skip default locations for this test
+	})
+
+	if err != nil {
+		t.Fatalf("LoadAll failed: %v", err)
+	}
+
+	if len(diags) > 0 {
+		t.Logf("Got %d diagnostics", len(diags))
+	}
+
+	if len(tpls) != 1 {
+		t.Fatalf("Expected 1 template, got %d", len(tpls))
+	}
+
+	tpl := tpls[0]
+	if tpl.Name != "test" {
+		t.Errorf("Expected name 'test', got '%s'", tpl.Name)
+	}
+
+	if tpl.Description != "Test template for integration" {
+		t.Errorf("Expected description 'Test template for integration', got '%s'", tpl.Description)
+	}
+
+	// Test expansion
+	expanded := tpl.Expand("code security")
+	expected := "Review code with focus on security"
+	if expanded != expected {
+		t.Errorf("Expected '%s', got '%s'", expected, expanded)
+	}
+}
+
+func TestParseTemplate_WithFrontmatter(t *testing.T) {
+	// Create a temp file with frontmatter
+	tempDir := t.TempDir()
+	templateContent := `---
+description: A test template
+---
+Create a $1 component with $2 features`
+
+	testFile := filepath.Join(tempDir, "component.md")
+	if err := os.WriteFile(testFile, []byte(templateContent), 0644); err != nil {
+		t.Fatalf("Failed to create test file: %v", err)
+	}
+
+	tpl, err := ParseTemplate(testFile)
+	if err != nil {
+		t.Fatalf("ParseTemplate failed: %v", err)
+	}
+
+	if tpl.Name != "component" {
+		t.Errorf("Expected name 'component', got '%s'", tpl.Name)
+	}
+
+	if tpl.Description != "A test template" {
+		t.Errorf("Expected description 'A test template', got '%s'", tpl.Description)
+	}
+
+	expectedContent := "Create a $1 component with $2 features"
+	if tpl.Content != expectedContent {
+		t.Errorf("Expected content '%s', got '%s'", expectedContent, tpl.Content)
+	}
+}
+
+func TestParseTemplate_WithoutFrontmatter(t *testing.T) {
+	// Create a temp file without frontmatter
+	tempDir := t.TempDir()
+	templateContent := `Simple template without frontmatter
+Supports $1 and $2 placeholders`
+
+	testFile := filepath.Join(tempDir, "simple.md")
+	if err := os.WriteFile(testFile, []byte(templateContent), 0644); err != nil {
+		t.Fatalf("Failed to create test file: %v", err)
+	}
+
+	tpl, err := ParseTemplate(testFile)
+	if err != nil {
+		t.Fatalf("ParseTemplate failed: %v", err)
+	}
+
+	if tpl.Name != "simple" {
+		t.Errorf("Expected name 'simple', got '%s'", tpl.Name)
+	}
+
+	// Description should be empty since there's no frontmatter
+	if tpl.Description != "" {
+		t.Errorf("Expected empty description, got '%s'", tpl.Description)
+	}
+
+	// Content should include everything
+	if tpl.Content != templateContent {
+		t.Errorf("Content mismatch\nExpected:\n%s\nGot:\n%s", templateContent, tpl.Content)
+	}
+}

internal/prompts/template.go

@@ -0,0 +1,279 @@
+package prompts
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+// PromptTemplate is a named prompt template with shell-style argument placeholders.
+// It supports Pi-style $1, $2, $@, $ARGUMENTS, ${@:N}, ${@:N:L} syntax.
+type PromptTemplate struct {
+	// Name is the human-readable identifier for this template.
+	Name string
+	// Description summarises what this template provides.
+	Description string
+	// Content is the raw template text with placeholders.
+	Content string
+	// Source indicates where the template was loaded from (e.g., "default", "user").
+	Source string
+	// FilePath is the absolute filesystem path the template was loaded from.
+	FilePath string
+}
+
+// ParseTemplate reads a template from a file. The template name is derived
+// from the filename (without extension). If the file contains YAML frontmatter,
+// the description is extracted from it.
+func ParseTemplate(path string) (*PromptTemplate, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("reading template %s: %w", path, err)
+	}
+
+	abs, err := filepath.Abs(path)
+	if err != nil {
+		abs = path
+	}
+
+	content := string(data)
+	tpl := &PromptTemplate{
+		FilePath: abs,
+		Content:  content,
+	}
+
+	// Parse frontmatter if present
+	if strings.HasPrefix(strings.TrimSpace(content), frontmatterSep) {
+		trimmed := strings.TrimSpace(content)
+		rest := trimmed[len(frontmatterSep):]
+		frontmatter, body, found := strings.Cut(rest, "\n"+frontmatterSep)
+		if found {
+			body = strings.TrimPrefix(body, "\n")
+			fm, err := ParseFrontmatter(frontmatter)
+			if err == nil {
+				tpl.Description = fm.Description
+			}
+			tpl.Content = strings.TrimSpace(body)
+		}
+	}
+
+	// Derive name from filename
+	base := filepath.Base(path)
+	ext := filepath.Ext(base)
+	tpl.Name = strings.TrimSuffix(base, ext)
+
+	return tpl, nil
+}
+
+// ParseCommandArgs splits a command line into arguments respecting quotes.
+// It handles single quotes, double quotes, and backslash escaping.
+func ParseCommandArgs(input string) []string {
+	var args []string
+	var current strings.Builder
+	inSingleQuote := false
+	inDoubleQuote := false
+	escaped := false
+
+	for i, r := range input {
+		if escaped {
+			current.WriteRune(r)
+			escaped = false
+			continue
+		}
+
+		if r == '\\' && !inSingleQuote {
+			// Backslash escapes next char, but not in single quotes
+			escaped = true
+			continue
+		}
+
+		if r == '\'' && !inDoubleQuote {
+			inSingleQuote = !inSingleQuote
+			continue
+		}
+
+		if r == '"' && !inSingleQuote {
+			inDoubleQuote = !inDoubleQuote
+			continue
+		}
+
+		if r == ' ' && !inSingleQuote && !inDoubleQuote {
+			if current.Len() > 0 {
+				args = append(args, current.String())
+				current.Reset()
+			}
+			continue
+		}
+
+		current.WriteRune(r)
+		_ = i // silence unused warning when we need position later
+	}
+
+	if current.Len() > 0 {
+		args = append(args, current.String())
+	}
+
+	return args
+}
+
+// argPlaceholder matches shell-style argument placeholders:
+//   - $1, $2, etc. - positional arguments
+//   - $@ - all arguments
+//   - $ARGUMENTS - all arguments (alias for $@)
+//   - ${@:N} - arguments from N onwards
+//   - ${@:N:L} - L arguments starting from N
+var argPlaceholder = regexp.MustCompile(`\$\{(\d+)\}|\$\{(\d+):(\d+)\}|\$\{ARGUMENTS\}|\$\{@(:\d+)?(:\d+)?\}|\$(\d+)|\$@|\$ARGUMENTS`)
+
+// SubstituteArgs replaces argument placeholders in content with values from args.
+// Supported placeholders:
+//   - $N, ${N} - the Nth argument (1-indexed)
+//   - $@, $ARGUMENTS, ${ARGUMENTS} - all arguments joined with spaces
+//   - ${@:N} - arguments from index N onwards (0-indexed)
+//   - ${@:N:L} - L arguments starting from index N (0-indexed)
+func SubstituteArgs(content string, args []string) string {
+	return argPlaceholder.ReplaceAllStringFunc(content, func(match string) string {
+		// Check for ${N} or ${N:M} format
+		if strings.HasPrefix(match, "${") && strings.Contains(match, "}") {
+			inner := match[2 : len(match)-1] // Remove ${ and }
+
+			// Check for ${ARGUMENTS}
+			if inner == "ARGUMENTS" {
+				return strings.Join(args, " ")
+			}
+
+			// Check for ${@...} format
+			if strings.HasPrefix(inner, "@") {
+				return expandAtArgs(inner, args)
+			}
+
+			// Check for ${N:M} format (positional with length)
+			if colonIdx := strings.Index(inner, ":"); colonIdx > 0 {
+				startStr := inner[:colonIdx]
+				rest := inner[colonIdx+1:]
+
+				start, err := strconv.Atoi(startStr)
+				if err != nil || start < 1 {
+					return match
+				}
+
+				// Check if there's a second colon for length ${N:M:L}
+				lengthStr, _, ok := strings.Cut(rest, ":")
+				if ok {
+					length, err := strconv.Atoi(lengthStr)
+					if err != nil || length < 0 {
+						return match
+					}
+					return joinArgsRange(args, start-1, length)
+				}
+
+				// Single colon ${N:M} - M is length
+				length, err := strconv.Atoi(rest)
+				if err != nil || length < 0 {
+					return match
+				}
+				return joinArgsRange(args, start-1, length)
+			}
+
+			// Simple ${N} format
+			n, err := strconv.Atoi(inner)
+			if err != nil || n < 1 {
+				return match
+			}
+			if n <= len(args) {
+				return args[n-1]
+			}
+			return ""
+		}
+
+		// Check for $N format (without braces)
+		if strings.HasPrefix(match, "$") && !strings.HasPrefix(match, "${") {
+			suffix := match[1:]
+
+			// $@ or $ARGUMENTS
+			if suffix == "@" || suffix == "ARGUMENTS" {
+				return strings.Join(args, " ")
+			}
+
+			// $N
+			n, err := strconv.Atoi(suffix)
+			if err != nil || n < 1 {
+				return match
+			}
+			if n <= len(args) {
+				return args[n-1]
+			}
+			return ""
+		}
+
+		return match
+	})
+}
+
+// expandAtArgs handles ${@...} patterns (1-indexed like bash)
+func expandAtArgs(inner string, args []string) string {
+	// Remove the @ prefix
+	rest := inner[1:]
+
+	if rest == "" {
+		// ${@} - all arguments
+		return strings.Join(args, " ")
+	}
+
+	// Must start with :
+	if !strings.HasPrefix(rest, ":") {
+		return "${" + inner + "}"
+	}
+	rest = rest[1:]
+
+	// Parse start index
+	startStr, lengthStr, hasLength := strings.Cut(rest, ":")
+
+	start, err := strconv.Atoi(startStr)
+	if err != nil || start < 0 {
+		return "${" + inner + "}"
+	}
+
+	// Convert from 1-indexed to 0-indexed (bash convention)
+	// Treat 0 as 1 (bash convention: args start at 1)
+	if start > 0 {
+		start--
+	}
+
+	if hasLength {
+		length, err := strconv.Atoi(lengthStr)
+		if err != nil || length < 0 {
+			return "${" + inner + "}"
+		}
+		return joinArgsRange(args, start, length)
+	}
+
+	// ${@:N} - from N to end
+	if start >= len(args) {
+		return ""
+	}
+	return strings.Join(args[start:], " ")
+}
+
+// joinArgsRange joins args from start index, taking up to length elements
+func joinArgsRange(args []string, start, length int) string {
+	if start >= len(args) || length <= 0 {
+		return ""
+	}
+	end := start + length
+	end = min(end, len(args))
+	return strings.Join(args[start:end], " ")
+}
+
+// Expand substitutes arguments into the template content and returns the result.
+// It first parses args from the input string, then substitutes them into the template.
+func (t *PromptTemplate) Expand(argsInput string) string {
+	args := ParseCommandArgs(argsInput)
+	return SubstituteArgs(t.Content, args)
+}
+
+// ExpandWithArgs substitutes the provided arguments into the template content.
+func (t *PromptTemplate) ExpandWithArgs(args []string) string {
+	return SubstituteArgs(t.Content, args)
+}

internal/prompts/template_test.go

@@ -0,0 +1,215 @@
+package prompts
+
+import (
+	"testing"
+)
+
+func TestParseCommandArgs(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected []string
+	}{
+		{"", []string{}},
+		{"hello", []string{"hello"}},
+		{"hello world", []string{"hello", "world"}},
+		{`"hello world"`, []string{"hello world"}},
+		{`'hello world'`, []string{"hello world"}},
+		{`hello "world foo" bar`, []string{"hello", "world foo", "bar"}},
+		{`hello 'world foo' bar`, []string{"hello", "world foo", "bar"}},
+		{`hello \"world\"`, []string{"hello", `"world"`}},
+		{`hello \\world`, []string{"hello", `\world`}},
+		{`  hello   world  `, []string{"hello", "world"}},
+		{`Button "onClick handler" "disabled support"`, []string{"Button", "onClick handler", "disabled support"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			got := ParseCommandArgs(tt.input)
+			if len(got) != len(tt.expected) {
+				t.Errorf("ParseCommandArgs(%q) = %v, want %v", tt.input, got, tt.expected)
+				return
+			}
+			for i := range got {
+				if got[i] != tt.expected[i] {
+					t.Errorf("ParseCommandArgs(%q)[%d] = %q, want %q", tt.input, i, got[i], tt.expected[i])
+				}
+			}
+		})
+	}
+}
+
+func TestSubstituteArgs(t *testing.T) {
+	tests := []struct {
+		name     string
+		content  string
+		args     []string
+		expected string
+	}{
+		{
+			name:     "no placeholders",
+			content:  "Hello world",
+			args:     []string{},
+			expected: "Hello world",
+		},
+		{
+			name:     "positional $1",
+			content:  "Hello $1",
+			args:     []string{"world"},
+			expected: "Hello world",
+		},
+		{
+			name:     "positional $1 $2",
+			content:  "$1 and $2",
+			args:     []string{"first", "second"},
+			expected: "first and second",
+		},
+		{
+			name:     "missing arg",
+			content:  "Hello $1 and $2",
+			args:     []string{"world"},
+			expected: "Hello world and ",
+		},
+		{
+			name:     "$@ wildcard",
+			content:  "Args: $@",
+			args:     []string{"a", "b", "c"},
+			expected: "Args: a b c",
+		},
+		{
+			name:     "$ARGUMENTS wildcard",
+			content:  "Args: $ARGUMENTS",
+			args:     []string{"a", "b", "c"},
+			expected: "Args: a b c",
+		},
+		{
+			name:     "${@} all args",
+			content:  "Args: ${@}",
+			args:     []string{"a", "b", "c"},
+			expected: "Args: a b c",
+		},
+		{
+			name:     "${@:2} slice from index 2",
+			content:  "Rest: ${@:2}",
+			args:     []string{"a", "b", "c", "d"},
+			expected: "Rest: b c d",
+		},
+		{
+			name:     "${@:1:2} slice with length",
+			content:  "First two: ${@:1:2}",
+			args:     []string{"a", "b", "c", "d"},
+			expected: "First two: a b",
+		},
+		{
+			name:     "${@:0} from start",
+			content:  "All: ${@:0}",
+			args:     []string{"a", "b", "c"},
+			expected: "All: a b c",
+		},
+		{
+			name:     "${@:3:1} single arg",
+			content:  "Third: ${@:3:1}",
+			args:     []string{"a", "b", "c", "d"},
+			expected: "Third: c",
+		},
+		{
+			name:     "combined placeholders",
+			content:  "Create $1 with features: $ARGUMENTS",
+			args:     []string{"Button", "onClick", "disabled"},
+			expected: "Create Button with features: Button onClick disabled",
+		},
+		{
+			name:     "slice beyond bounds",
+			content:  "${@:10}",
+			args:     []string{"a", "b"},
+			expected: "",
+		},
+		{
+			name:     "empty args with wildcard",
+			content:  "Args: $@",
+			args:     []string{},
+			expected: "Args: ",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := SubstituteArgs(tt.content, tt.args)
+			if got != tt.expected {
+				t.Errorf("SubstituteArgs(%q, %v) = %q, want %q", tt.content, tt.args, got, tt.expected)
+			}
+		})
+	}
+}
+
+func TestParseFrontmatter(t *testing.T) {
+	tests := []struct {
+		name     string
+		content  string
+		wantDesc string
+		wantErr  bool
+	}{
+		{
+			name:     "simple description",
+			content:  "description: Review code\n",
+			wantDesc: "Review code",
+		},
+		{
+			name:     "empty",
+			content:  "",
+			wantDesc: "",
+		},
+		{
+			name:     "invalid yaml",
+			content:  "description: [unclosed",
+			wantDesc: "",
+			wantErr:  true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			fm, err := ParseFrontmatter(tt.content)
+			if (err != nil) != tt.wantErr {
+				t.Errorf("ParseFrontmatter() error = %v, wantErr %v", err, tt.wantErr)
+				return
+			}
+			if err != nil {
+				return
+			}
+			if fm.Description != tt.wantDesc {
+				t.Errorf("ParseFrontmatter() Description = %q, want %q", fm.Description, tt.wantDesc)
+			}
+		})
+	}
+}
+
+func TestPromptTemplateExpand(t *testing.T) {
+	tpl := &PromptTemplate{
+		Name:        "component",
+		Description: "Create a component",
+		Content:     "Create a React component named $1 with features: $ARGUMENTS",
+	}
+
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{
+			input:    "Button",
+			expected: "Create a React component named Button with features: Button",
+		},
+		{
+			input:    `Button "onClick handler"`,
+			expected: "Create a React component named Button with features: Button onClick handler",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			got := tpl.Expand(tt.input)
+			if got != tt.expected {
+				t.Errorf("Expand(%q) = %q, want %q", tt.input, got, tt.expected)
+			}
+		})
+	}
+}

internal/session/entry.go

@@ -23,6 +23,7 @@ const (
 	EntryTypeLabel         EntryType = "label"
 	EntryTypeSessionInfo   EntryType = "session_info"
 	EntryTypeExtensionData EntryType = "extension_data"
+	EntryTypeCompaction    EntryType = "compaction"
 )
 
 // CurrentVersion is the session format version for JSONL tree sessions.
@@ -102,6 +103,20 @@ type ExtensionDataEntry struct {
 	Data    string `json:"data"`     // Extension-defined data (JSON or plain text)
 }
 
+// CompactionEntry records an LLM-generated summary of older messages.
+// Instead of deleting old messages, the tree manager skips entries before
+// FirstKeptEntryID when building the LLM context, preserving full history.
+type CompactionEntry struct {
+	Entry
+	Summary          string   `json:"summary"`
+	FirstKeptEntryID string   `json:"first_kept_entry_id"`
+	TokensBefore     int      `json:"tokens_before"`
+	TokensAfter      int      `json:"tokens_after"`
+	MessagesRemoved  int      `json:"messages_removed"`
+	ReadFiles        []string `json:"read_files,omitempty"`
+	ModifiedFiles    []string `json:"modified_files,omitempty"`
+}
+
 // GenerateEntryID creates a unique entry identifier (16 hex chars).
 func GenerateEntryID() string {
 	bytes := make([]byte, 8)
@@ -188,6 +203,20 @@ func NewExtensionDataEntry(parentID, extType, data string) *ExtensionDataEntry {
 	}
 }
 
+// NewCompactionEntry creates a CompactionEntry.
+func NewCompactionEntry(parentID, summary, firstKeptEntryID string, tokensBefore, tokensAfter, messagesRemoved int, readFiles, modifiedFiles []string) *CompactionEntry {
+	return &CompactionEntry{
+		Entry:            NewEntry(EntryTypeCompaction, parentID),
+		Summary:          summary,
+		FirstKeptEntryID: firstKeptEntryID,
+		TokensBefore:     tokensBefore,
+		TokensAfter:      tokensAfter,
+		MessagesRemoved:  messagesRemoved,
+		ReadFiles:        readFiles,
+		ModifiedFiles:    modifiedFiles,
+	}
+}
+
 // --- JSONL marshaling helpers ---
 
 // MarshalEntry serializes any entry to a JSON line (no trailing newline).
@@ -259,6 +288,13 @@ func UnmarshalEntry(data []byte) (any, error) {
 		}
 		return &e, nil
 
+	case EntryTypeCompaction:
+		var e CompactionEntry
+		if err := json.Unmarshal(data, &e); err != nil {
+			return nil, fmt.Errorf("failed to unmarshal compaction entry: %w", err)
+		}
+		return &e, nil
+
 	default:
 		return nil, fmt.Errorf("unknown entry type: %q", env.Type)
 	}

internal/session/store.go

@@ -96,6 +96,7 @@ func ListAllSessions() ([]SessionInfo, error) {
 }
 
 // listSessionsInDir reads all .jsonl files in a directory and extracts session info.
+// Empty sessions (no messages) are automatically cleaned up and not returned.
 func listSessionsInDir(dir string) ([]SessionInfo, error) {
 	if _, err := os.Stat(dir); os.IsNotExist(err) {
 		return nil, nil
@@ -117,6 +118,11 @@ func listSessionsInDir(dir string) ([]SessionInfo, error) {
 		if err != nil {
 			continue // skip malformed session files
 		}
+		// Clean up and skip empty sessions (no messages)
+		if info.MessageCount == 0 {
+			_ = os.Remove(path)
+			continue
+		}
 		sessions = append(sessions, *info)
 	}
 

internal/session/tree_manager.go

@@ -4,6 +4,7 @@ import (
 	"bufio"
 	"encoding/json"
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
 	"strings"
@@ -128,10 +129,34 @@ func OpenTreeSession(path string) (*TreeManager, error) {
 		filePath:   path,
 	}
 
-	scanner := bufio.NewScanner(strings.NewReader(string(data)))
+	reader := bufio.NewReader(strings.NewReader(string(data)))
 	lineNum := 0
-	for scanner.Scan() {
-		line := scanner.Text()
+	for {
+		line, err := reader.ReadString('\n')
+		if err != nil {
+			if err == io.EOF {
+				// Process the last line if it's not empty
+				if strings.TrimSpace(line) != "" {
+					lineNum++
+					entry, err := UnmarshalEntry([]byte(line))
+					if err != nil {
+						return nil, fmt.Errorf("line %d: %w", lineNum, err)
+					}
+					if lineNum == 1 {
+						h, ok := entry.(*SessionHeader)
+						if !ok {
+							return nil, fmt.Errorf("first line must be a session header, got %T", entry)
+						}
+						tm.header = *h
+					} else {
+						tm.addEntryToIndex(entry)
+					}
+				}
+				break
+			}
+			return nil, fmt.Errorf("failed to read session file: %w", err)
+		}
+
 		if strings.TrimSpace(line) == "" {
 			continue
 		}
@@ -153,13 +178,10 @@ func OpenTreeSession(path string) (*TreeManager, error) {
 
 		tm.addEntryToIndex(entry)
 	}
-	if err := scanner.Err(); err != nil {
-		return nil, fmt.Errorf("failed to scan session file: %w", err)
-	}
 
 	// Set leaf to the last entry.
 	if len(tm.entries) > 0 {
-		tm.leafID = tm.entryID(tm.entries[len(tm.entries)-1])
+		tm.leafID = tm.EntryID(tm.entries[len(tm.entries)-1])
 	}
 
 	// Open file for appending.
@@ -220,9 +242,14 @@ func (tm *TreeManager) AppendMessage(msg message.Message) (string, error) {
 	return entry.ID, nil
 }
 
-// AppendFantasyMessage converts a fantasy.Message and appends it.
+// AppendLLMMessage converts an LLM message and appends it.
+func (tm *TreeManager) AppendLLMMessage(msg fantasy.Message) (string, error) {
+	return tm.AppendMessage(message.FromLLMMessage(msg))
+}
+
+// Deprecated: Use AppendLLMMessage instead.
 func (tm *TreeManager) AppendFantasyMessage(msg fantasy.Message) (string, error) {
-	return tm.AppendMessage(message.FromFantasyMessage(msg))
+	return tm.AppendLLMMessage(msg)
 }
 
 // AppendModelChange records a model/provider change.
@@ -298,6 +325,22 @@ func (tm *TreeManager) AppendExtensionData(extType, data string) (string, error)
 	return entry.ID, nil
 }
 
+// AppendCompaction adds a compaction entry to the tree. The entry records
+// the summary and the ID of the first entry that should be preserved in the
+// LLM context. Messages before that entry are replaced by the summary.
+func (tm *TreeManager) AppendCompaction(summary, firstKeptEntryID string, tokensBefore, tokensAfter, messagesRemoved int, readFiles, modifiedFiles []string) (string, error) {
+	tm.mu.Lock()
+	defer tm.mu.Unlock()
+
+	entry := NewCompactionEntry(tm.leafID, summary, firstKeptEntryID, tokensBefore, tokensAfter, messagesRemoved, readFiles, modifiedFiles)
+	if err := tm.appendAndPersist(entry); err != nil {
+		return "", err
+	}
+
+	tm.leafID = entry.ID
+	return entry.ID, nil
+}
+
 // GetExtensionData returns all extension data entries matching the given type,
 // walking the current branch from root to leaf. If extType is empty, all
 // extension data entries on the branch are returned.
@@ -441,8 +484,9 @@ func (tm *TreeManager) GetTree() []*TreeNode {
 // --- Context building ---
 
 // BuildContext walks from the current leaf to the root and returns the
-// conversation messages suitable for sending to the LLM. Branch summaries
-// are converted to user messages to provide context from abandoned branches.
+// conversation messages suitable for sending to the LLM. Compaction entries
+// cause older messages to be replaced by the summary. Branch summaries are
+// converted to user messages to provide context from abandoned branches.
 // Also returns the latest model/provider settings encountered on the path.
 func (tm *TreeManager) BuildContext() (messages []fantasy.Message, provider string, modelID string) {
 	tm.mu.RLock()
@@ -455,14 +499,48 @@ func (tm *TreeManager) BuildContext() (messages []fantasy.Message, provider stri
 	// Walk from leaf to root collecting entries.
 	branch := tm.getBranchLocked(tm.leafID)
 
+	// Find the last compaction entry on this branch — it determines
+	// which older messages are replaced by the summary.
+	var lastCompaction *CompactionEntry
+	for i := len(branch) - 1; i >= 0; i-- {
+		if c, ok := branch[i].(*CompactionEntry); ok {
+			lastCompaction = c
+			break
+		}
+	}
+
+	// If there is a compaction, inject the summary first.
+	if lastCompaction != nil {
+		messages = append(messages, fantasy.Message{
+			Role: fantasy.MessageRoleSystem,
+			Content: []fantasy.MessagePart{
+				fantasy.TextPart{
+					Text: fmt.Sprintf("[Conversation summary — earlier messages were compacted]\n\n%s", lastCompaction.Summary),
+				},
+			},
+		})
+	}
+
+	// Determine whether to skip entries (everything before firstKeptEntryID).
+	skipping := lastCompaction != nil
 	for _, entry := range branch {
+		// Once we reach the first kept entry, stop skipping.
+		if skipping {
+			entryID := tm.EntryID(entry)
+			if entryID == lastCompaction.FirstKeptEntryID {
+				skipping = false
+			} else {
+				continue
+			}
+		}
+
 		switch e := entry.(type) {
 		case *MessageEntry:
 			msg, err := e.ToMessage()
 			if err != nil {
 				continue // skip malformed entries
 			}
-			msgs := msg.ToFantasyMessages()
+			msgs := msg.ToLLMMessages()
 			messages = append(messages, msgs...)
 
 		case *BranchSummaryEntry:
@@ -481,6 +559,10 @@ func (tm *TreeManager) BuildContext() (messages []fantasy.Message, provider stri
 		case *ModelChangeEntry:
 			provider = e.Provider
 			modelID = e.ModelID
+
+		case *CompactionEntry:
+			// Already handled above (the last one on the branch).
+			continue
 		}
 	}
 
@@ -551,6 +633,11 @@ func (tm *TreeManager) MessageCount() int {
 	return count
 }
 
+// IsEmpty returns true if the session has no messages (only header).
+func (tm *TreeManager) IsEmpty() bool {
+	return tm.MessageCount() == 0
+}
+
 // Close closes the underlying file handle.
 func (tm *TreeManager) Close() error {
 	tm.mu.Lock()
@@ -563,33 +650,133 @@ func (tm *TreeManager) Close() error {
 	return nil
 }
 
+// GetContextEntryIDs returns the entry IDs corresponding to the fantasy
+// messages returned by BuildContext, in the same order. Each entry ID maps
+// to the session entry that produced the fantasy message at the same index.
+// This is used by compaction to map a cut point index back to an entry ID.
+//
+// Note: A single MessageEntry produces at most one fantasy message. Branch
+// summary entries also produce one message each. The returned slice has the
+// same length as the messages slice from BuildContext (excluding the
+// compaction summary system message, which has no entry ID — it gets the
+// empty string "").
+func (tm *TreeManager) GetContextEntryIDs() []string {
+	tm.mu.RLock()
+	defer tm.mu.RUnlock()
+
+	if tm.leafID == "" {
+		return nil
+	}
+
+	branch := tm.getBranchLocked(tm.leafID)
+
+	// Find the last compaction entry for skip logic.
+	var lastCompaction *CompactionEntry
+	for i := len(branch) - 1; i >= 0; i-- {
+		if c, ok := branch[i].(*CompactionEntry); ok {
+			lastCompaction = c
+			break
+		}
+	}
+
+	var ids []string
+
+	// If there's a compaction summary injected, it has no entry ID.
+	if lastCompaction != nil {
+		ids = append(ids, "") // placeholder for the summary system message
+	}
+
+	skipping := lastCompaction != nil
+	for _, entry := range branch {
+		if skipping {
+			entryID := tm.EntryID(entry)
+			if entryID == lastCompaction.FirstKeptEntryID {
+				skipping = false
+			} else {
+				continue
+			}
+		}
+
+		switch e := entry.(type) {
+		case *MessageEntry:
+			msg, err := e.ToMessage()
+			if err != nil {
+				continue
+			}
+			msgs := msg.ToLLMMessages()
+			for range msgs {
+				ids = append(ids, e.ID)
+			}
+
+		case *BranchSummaryEntry:
+			if e.Summary != "" {
+				ids = append(ids, e.ID)
+			}
+
+		case *CompactionEntry:
+			continue
+		}
+	}
+
+	return ids
+}
+
+// GetLastCompaction returns the most recent CompactionEntry on the current
+// branch, or nil if none exists. Used to carry forward file tracking.
+func (tm *TreeManager) GetLastCompaction() *CompactionEntry {
+	tm.mu.RLock()
+	defer tm.mu.RUnlock()
+
+	if tm.leafID == "" {
+		return nil
+	}
+
+	branch := tm.getBranchLocked(tm.leafID)
+	for i := len(branch) - 1; i >= 0; i-- {
+		if c, ok := branch[i].(*CompactionEntry); ok {
+			return c
+		}
+	}
+	return nil
+}
+
 // --- Legacy bridge ---
 
-// AddFantasyMessages appends multiple fantasy messages as entries. This is
+// AddLLMMessages appends multiple LLM messages as entries. This is
 // used when syncing from the agent's ConversationMessages after a step.
-func (tm *TreeManager) AddFantasyMessages(msgs []fantasy.Message) error {
+func (tm *TreeManager) AddLLMMessages(msgs []fantasy.Message) error {
 	for _, msg := range msgs {
-		if _, err := tm.AppendFantasyMessage(msg); err != nil {
+		if _, err := tm.AppendLLMMessage(msg); err != nil {
 			return err
 		}
 	}
 	return nil
 }
 
-// GetFantasyMessages builds the context and returns just the messages.
+// Deprecated: Use AddLLMMessages instead.
+func (tm *TreeManager) AddFantasyMessages(msgs []fantasy.Message) error {
+	return tm.AddLLMMessages(msgs)
+}
+
+// GetLLMMessages builds the context and returns just the messages.
 // This satisfies the same conceptual role as the old Manager.GetMessages().
-func (tm *TreeManager) GetFantasyMessages() []fantasy.Message {
+func (tm *TreeManager) GetLLMMessages() []fantasy.Message {
 	msgs, _, _ := tm.BuildContext()
 	return msgs
 }
 
+// Deprecated: Use GetLLMMessages instead.
+func (tm *TreeManager) GetFantasyMessages() []fantasy.Message {
+	return tm.GetLLMMessages()
+}
+
 // --- Internal helpers ---
 
 // addEntryToIndex adds an entry to the in-memory indices.
 func (tm *TreeManager) addEntryToIndex(entry any) {
 	tm.entries = append(tm.entries, entry)
 
-	id := tm.entryID(entry)
+	id := tm.EntryID(entry)
 	parentID := tm.entryParentID(entry)
 
 	if id != "" {
@@ -626,8 +813,8 @@ func (tm *TreeManager) writeEntry(entry any) error {
 	return err
 }
 
-// entryID extracts the ID from any entry type.
-func (tm *TreeManager) entryID(entry any) string {
+// EntryID extracts the ID from any entry type.
+func (tm *TreeManager) EntryID(entry any) string {
 	switch e := entry.(type) {
 	case *MessageEntry:
 		return e.ID
@@ -641,6 +828,8 @@ func (tm *TreeManager) entryID(entry any) string {
 		return e.ID
 	case *ExtensionDataEntry:
 		return e.ID
+	case *CompactionEntry:
+		return e.ID
 	default:
 		return ""
 	}
@@ -661,6 +850,8 @@ func (tm *TreeManager) entryParentID(entry any) string {
 		return e.ParentID
 	case *ExtensionDataEntry:
 		return e.ParentID
+	case *CompactionEntry:
+		return e.ParentID
 	default:
 		return ""
 	}

internal/tools/connection_pool.go

@@ -127,9 +127,7 @@ func (p *MCPConnectionPool) GetConnection(ctx context.Context, serverName string
 			return conn, nil
 		} else {
 			if p.debugLogger != nil && p.debugLogger.IsDebugEnabled() {
-				if p.debugLogger != nil && p.debugLogger.IsDebugEnabled() {
-					p.debugLogger.LogDebug(fmt.Sprintf("[POOL] Connection %s unhealthy, removing", serverName))
-				}
+				p.debugLogger.LogDebug(fmt.Sprintf("[POOL] Connection %s unhealthy, removing", serverName))
 			}
 			_ = conn.client.Close()
 			delete(p.connections, serverName)

internal/tools/mcp_test.go

@@ -3,6 +3,7 @@ package tools
 import (
 	"context"
 	"encoding/json"
+	"strings"
 	"testing"
 	"time"
 
@@ -70,7 +71,7 @@ func TestMCPToolManager_LoadTools_GracefulFailure(t *testing.T) {
 	}
 
 	// The error should mention that all servers failed
-	if err != nil && !contains(err.Error(), "all MCP servers failed") {
+	if err != nil && !strings.Contains(err.Error(), "all MCP servers failed") {
 		t.Errorf("Expected error message to mention all servers failed, got: %v", err)
 	}
 
@@ -459,13 +460,3 @@ func sliceEqual(a, b []any) bool {
 	}
 	return true
 }
-
-// Helper function to check if a string contains a substring
-func contains(s, substr string) bool {
-	for i := 0; i <= len(s)-len(substr); i++ {
-		if s[i:i+len(substr)] == substr {
-			return true
-		}
-	}
-	return false
-}

internal/ui/children_test.go

@@ -149,11 +149,13 @@ func TestInputComponent_QuitReturnsTeaQuit(t *testing.T) {
 }
 
 // --------------------------------------------------------------------------
-// TestInputComponent_ClearCallsClearMessages verifies that /clear (and its
-// aliases) calls appCtrl.ClearMessages() and returns no submitMsg.
+// TestInputComponent_ClearForwardsAsSubmitMsg verifies that /clear (and its
+// aliases) are forwarded as submitMsg to the parent model so that the parent
+// can call ClearMessages(), update scrollback, and print the confirmation
+// message in one place. InputComponent must NOT call ClearMessages() directly.
 // --------------------------------------------------------------------------
 
-func TestInputComponent_ClearCallsClearMessages(t *testing.T) {
+func TestInputComponent_ClearForwardsAsSubmitMsg(t *testing.T) {
 	aliases := []string{"/clear", "/c", "/cls"}
 	for _, alias := range aliases {
 		t.Run(alias, func(t *testing.T) {
@@ -164,22 +166,29 @@ func TestInputComponent_ClearCallsClearMessages(t *testing.T) {
 
 			_, cmd := sendInputMsg(c, tea.KeyPressMsg{Code: tea.KeyEnter})
 
-			if ctrl.clearMsgCalled != 1 {
-				t.Fatalf("%s: expected ClearMessages() called once, got %d", alias, ctrl.clearMsgCalled)
+			// InputComponent must NOT call ClearMessages() directly.
+			if ctrl.clearMsgCalled != 0 {
+				t.Fatalf("%s: InputComponent must not call ClearMessages(), got %d", alias, ctrl.clearMsgCalled)
 			}
-			// No cmd should be returned (no submitMsg forwarded to parent).
-			if cmd != nil {
-				msg := runCmd(cmd)
-				if _, ok := msg.(submitMsg); ok {
-					t.Fatalf("%s: /clear should not emit submitMsg, got submitMsg", alias)
-				}
+			// A submitMsg must be emitted so the parent model handles /clear.
+			if cmd == nil {
+				t.Fatalf("%s: expected submitMsg cmd, got nil", alias)
+			}
+			msg := runCmd(cmd)
+			sm, ok := msg.(submitMsg)
+			if !ok {
+				t.Fatalf("%s: expected submitMsg, got %T", alias, msg)
+			}
+			if sm.Text != alias {
+				t.Fatalf("%s: expected submitMsg text %q, got %q", alias, alias, sm.Text)
 			}
 		})
 	}
 }
 
 // TestInputComponent_ClearNilCtrl_NoPanic verifies that /clear with a nil
-// appCtrl does not panic.
+// appCtrl does not panic. Since /clear is now forwarded to the parent via
+// submitMsg, no appCtrl interaction happens in InputComponent at all.
 func TestInputComponent_ClearNilCtrl_NoPanic(t *testing.T) {
 	c := newTestInput(nil)
 	c.textarea.SetValue("/clear")
@@ -349,7 +358,7 @@ func TestStreamComponent_SpinnerKeepsRunningDuringStreaming(t *testing.T) {
 	c = sendStreamMsg(c, app.StreamChunkEvent{Content: "hello"})
 
 	// Flush pending chunks (simulates the 16ms tick firing).
-	c = sendStreamMsg(c, streamFlushTickMsg{})
+	c = sendStreamMsg(c, streamFlushTickMsg{generation: c.flushGeneration})
 
 	if !c.spinning {
 		t.Fatal("expected spinning=true after first chunk")
@@ -376,7 +385,7 @@ func TestStreamComponent_ChunkAccumulation(t *testing.T) {
 	}
 
 	// Flush pending chunks (simulates the 16ms tick firing).
-	c = sendStreamMsg(c, streamFlushTickMsg{})
+	c = sendStreamMsg(c, streamFlushTickMsg{generation: c.flushGeneration})
 
 	got := c.streamContent.String()
 	want := "Hello, world!"
@@ -396,6 +405,7 @@ func TestStreamComponent_ToolExecution_IsStarting_ShowsSpinner(t *testing.T) {
 	c := newTestStream()
 
 	_, cmd := c.Update(app.ToolExecutionEvent{
+		ToolCallID: "call-exec-1",
 		ToolName:   "exec_tool",
 		IsStarting: true,
 	})
@@ -403,8 +413,9 @@ func TestStreamComponent_ToolExecution_IsStarting_ShowsSpinner(t *testing.T) {
 	if !c.spinning {
 		t.Fatal("expected spinning=true during tool execution")
 	}
-	if len(c.activeTools) != 1 || !strings.Contains(c.activeTools[0], "exec_tool") {
-		t.Fatalf("expected activeTools to contain tool name, got %v", c.activeTools)
+	tools := c.activeToolDisplays()
+	if len(tools) != 1 || !strings.Contains(tools[0], "exec_tool") {
+		t.Fatalf("expected activeTools to contain tool name, got %v", tools)
 	}
 	if cmd == nil {
 		t.Fatal("expected tick cmd from ToolExecutionEvent{IsStarting:true}")
@@ -418,11 +429,13 @@ func TestStreamComponent_ToolExecution_NotStarting_KeepsSpinning(t *testing.T) {
 	c = sendStreamMsg(c, app.SpinnerEvent{Show: true})
 	// Simulate a tool starting
 	c = sendStreamMsg(c, app.ToolExecutionEvent{
+		ToolCallID: "call-some-1",
 		ToolName:   "some_tool",
 		IsStarting: true,
 	})
 
 	c = sendStreamMsg(c, app.ToolExecutionEvent{
+		ToolCallID: "call-some-1",
 		ToolName:   "some_tool",
 		IsStarting: false,
 	})
@@ -440,9 +453,9 @@ func TestStreamComponent_ParallelToolExecution(t *testing.T) {
 	c := newTestStream()
 
 	// Start three tools in parallel
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "read", IsStarting: true})
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "grep", IsStarting: true})
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "find", IsStarting: true})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read", ToolName: "read", IsStarting: true})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-grep", ToolName: "grep", IsStarting: true})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-find", ToolName: "find", IsStarting: true})
 
 	if len(c.activeTools) != 3 {
 		t.Fatalf("expected 3 active tools, got %d: %v", len(c.activeTools), c.activeTools)
@@ -455,19 +468,44 @@ func TestStreamComponent_ParallelToolExecution(t *testing.T) {
 	}
 
 	// Finish one tool
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "grep", IsStarting: false})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-grep", ToolName: "grep", IsStarting: false})
 	if len(c.activeTools) != 2 {
 		t.Fatalf("expected 2 active tools after one finished, got %d: %v", len(c.activeTools), c.activeTools)
 	}
 
 	// Finish remaining tools
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "read", IsStarting: false})
-	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolName: "find", IsStarting: false})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read", ToolName: "read", IsStarting: false})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-find", ToolName: "find", IsStarting: false})
 	if len(c.activeTools) != 0 {
 		t.Fatalf("expected 0 active tools after all finished, got %d: %v", len(c.activeTools), c.activeTools)
 	}
 }
 
+// TestStreamComponent_ParallelSameToolName_UsesToolCallID verifies finishing one
+// tool call does not remove another concurrent call with the same tool name.
+func TestStreamComponent_ParallelSameToolName_UsesToolCallID(t *testing.T) {
+	c := newTestStream()
+
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read-1", ToolName: "read", IsStarting: true})
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read-2", ToolName: "read", IsStarting: true})
+
+	tools := c.activeToolDisplays()
+	if len(tools) != 2 {
+		t.Fatalf("expected 2 active read calls, got %d (%v)", len(tools), tools)
+	}
+
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read-1", ToolName: "read", IsStarting: false})
+	tools = c.activeToolDisplays()
+	if len(tools) != 1 {
+		t.Fatalf("expected 1 active read call after finishing one ID, got %d (%v)", len(tools), tools)
+	}
+
+	c = sendStreamMsg(c, app.ToolExecutionEvent{ToolCallID: "call-read-2", ToolName: "read", IsStarting: false})
+	if len(c.activeToolDisplays()) != 0 {
+		t.Fatalf("expected no active tools after finishing both IDs, got %v", c.activeToolDisplays())
+	}
+}
+
 // --------------------------------------------------------------------------
 // TestStreamComponent_GetRenderedContent verifies the method returns rendered
 // text when content is accumulated, and empty string when not.
@@ -621,3 +659,43 @@ func TestStreamComponent_StaleTick_Discarded(t *testing.T) {
 		t.Fatal("current-gen tick should reschedule")
 	}
 }
+
+// TestStreamComponent_StaleFlushTick_Discarded verifies that flush ticks from a
+// previous generation (e.g. pre-Reset) are ignored.
+func TestStreamComponent_StaleFlushTick_Discarded(t *testing.T) {
+	c := newTestStream()
+
+	// Start a pending flush and capture its generation.
+	c = sendStreamMsg(c, app.StreamChunkEvent{Content: "old"})
+	staleGen := c.flushGeneration
+	if !c.flushPending {
+		t.Fatal("precondition: expected flushPending=true after first chunk")
+	}
+
+	// Reset should invalidate in-flight flush ticks.
+	c.Reset()
+	if c.flushGeneration == staleGen {
+		t.Fatal("expected flushGeneration to change after Reset")
+	}
+
+	// New content in a new generation.
+	c = sendStreamMsg(c, app.StreamChunkEvent{Content: "new"})
+	if got := c.pendingStream.String(); got != "new" {
+		t.Fatalf("expected pendingStream='new', got %q", got)
+	}
+
+	// Stale flush tick should be ignored.
+	c = sendStreamMsg(c, streamFlushTickMsg{generation: staleGen})
+	if got := c.pendingStream.String(); got != "new" {
+		t.Fatalf("stale flush tick should not commit pending stream, got %q", got)
+	}
+
+	// Current generation flush should commit.
+	c = sendStreamMsg(c, streamFlushTickMsg{generation: c.flushGeneration})
+	if got := c.pendingStream.String(); got != "" {
+		t.Fatalf("expected pendingStream empty after current flush, got %q", got)
+	}
+	if got := c.streamContent.String(); got != "new" {
+		t.Fatalf("expected streamContent='new' after current flush, got %q", got)
+	}
+}

internal/ui/cli.go

@@ -179,9 +179,8 @@ func (c *CLI) DisplayDebugConfig(config map[string]any) {
 }
 
 // UpdateUsageFromResponse records token usage using metadata from the fantasy
-// response when available. Falls back to text-based estimation if the metadata is
-// missing or appears unreliable. This provides more accurate usage tracking when
-// providers supply token count information.
+// response. Only actual API-reported tokens are used for cost tracking.
+// If the provider doesn't report token counts, no usage is recorded.
 func (c *CLI) UpdateUsageFromResponse(response *fantasy.Response, inputText string) {
 	if c.usageTracker == nil {
 		return
@@ -191,19 +190,19 @@ func (c *CLI) UpdateUsageFromResponse(response *fantasy.Response, inputText stri
 	inputTokens := int(usage.InputTokens)
 	outputTokens := int(usage.OutputTokens)
 
-	// Validate that the metadata seems reasonable
-	if inputTokens > 0 && outputTokens > 0 {
+	// Only use actual API-reported tokens for cost tracking.
+	// We intentionally do NOT estimate tokens - estimation is inaccurate
+	// and should never be used for cost calculations.
+	if inputTokens > 0 {
 		cacheReadTokens := int(usage.CacheReadTokens)
 		cacheWriteTokens := int(usage.CacheCreationTokens)
 		c.usageTracker.UpdateUsage(inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens)
 		// Per-response usage is a single API call, so it represents the
 		// actual context window fill level.
 		c.usageTracker.SetContextTokens(inputTokens + outputTokens)
-	} else {
-		// Fallback to estimation if no metadata is available.
-		// EstimateAndUpdateUsage sets context tokens internally.
-		c.usageTracker.EstimateAndUpdateUsage(inputText, response.Content.Text())
 	}
+	// If inputTokens is 0, the provider didn't report usage - we skip recording
+	// rather than estimating, to ensure cost accuracy.
 }
 
 // DisplayUsageAfterResponse renders and displays token usage information immediately

internal/ui/commands.go

@@ -152,6 +152,11 @@ var SlashCommands = []SlashCommand{
 		Description: "Export session (JSONL by default, or /export path.jsonl)",
 		Category:    "System",
 	},
+	{
+		Name:        "/share",
+		Description: "Share session via GitHub Gist (requires gh CLI)",
+		Category:    "System",
+	},
 	{
 		Name:        "/import",
 		Description: "Import a session from a JSONL file (/import path.jsonl)",

internal/ui/compact_renderer.go

@@ -43,7 +43,7 @@ func (r *CompactRenderer) SetWidth(width int) {
 // while minimizing vertical space usage. Returns a UIMessage with formatted content
 // and metadata.
 func (r *CompactRenderer) RenderUserMessage(content string, timestamp time.Time) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Info).Render(">")
 	label := lipgloss.NewStyle().Foreground(theme.Info).Bold(true).Render("User")
 
@@ -82,10 +82,21 @@ func (r *CompactRenderer) RenderUserMessage(content string, timestamp time.Time)
 }
 
 // RenderAssistantMessage renders an AI assistant's response in compact format with
-// a distinctive symbol (<) and the model name as label. Empty content is displayed
-// as "(no output)". Returns a UIMessage with formatted content and metadata.
+// a distinctive symbol (<) and the model name as label. Empty content is ignored
+// and returns an empty message. Returns a UIMessage with formatted content and metadata.
 func (r *CompactRenderer) RenderAssistantMessage(content string, timestamp time.Time, modelName string) UIMessage {
-	theme := getTheme()
+	// Ignore empty responses - don't render anything
+	compactContent := r.formatUserAssistantContent(content)
+	if compactContent == "" {
+		return UIMessage{
+			Type:      AssistantMessage,
+			Content:   "",
+			Height:    0,
+			Timestamp: timestamp,
+		}
+	}
+
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Primary).Render("<")
 
 	// Use the full model name, fallback to "Assistant" if empty
@@ -94,12 +105,6 @@ func (r *CompactRenderer) RenderAssistantMessage(content string, timestamp time.
 	}
 	label := lipgloss.NewStyle().Foreground(theme.Primary).Bold(true).Render(modelName)
 
-	// Format content for assistant messages (preserve formatting, no truncation)
-	compactContent := r.formatUserAssistantContent(content)
-	if compactContent == "" {
-		compactContent = lipgloss.NewStyle().Foreground(theme.Muted).Italic(true).Render("(no output)")
-	}
-
 	// Handle multi-line content
 	lines := strings.Split(compactContent, "\n")
 	var formattedLines []string
@@ -122,35 +127,11 @@ func (r *CompactRenderer) RenderAssistantMessage(content string, timestamp time.
 	}
 }
 
-// RenderToolCallMessage renders a tool call notification in compact format, showing
-// the tool being executed with its arguments in a single line. The tool name is
-// highlighted and arguments are displayed in a muted color for visual distinction.
-func (r *CompactRenderer) RenderToolCallMessage(toolName, toolArgs string, timestamp time.Time) UIMessage {
-	theme := getTheme()
-	symbol := lipgloss.NewStyle().Foreground(theme.Tool).Render("[")
-	label := lipgloss.NewStyle().Foreground(theme.Tool).Bold(true).Render(toolName)
-
-	// Format args for compact display
-	argsDisplay := r.formatToolArgs(toolArgs)
-	if argsDisplay != "" {
-		argsDisplay = lipgloss.NewStyle().Foreground(theme.Muted).Render(argsDisplay)
-	}
-
-	line := fmt.Sprintf("%s  %s %s", symbol, label, argsDisplay)
-
-	return UIMessage{
-		Type:      ToolCallMessage,
-		Content:   line,
-		Height:    1,
-		Timestamp: timestamp,
-	}
-}
-
 // RenderToolMessage renders a unified tool block in compact format, combining
 // the tool invocation header (icon + display name + params) with the execution
 // result body. Status is indicated by icon: checkmark for success, cross for error.
 func (r *CompactRenderer) RenderToolMessage(toolName, toolArgs, toolResult string, isError bool) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 
 	// Resolve extension renderer once for all overrides.
 	var extRd *ToolRendererData
@@ -239,7 +220,7 @@ func (r *CompactRenderer) RenderToolMessage(toolName, toolArgs, toolResult strin
 // compact format with a distinctive symbol (*) and "System" label. Content is
 // formatted to fit on a single line for minimal space usage.
 func (r *CompactRenderer) RenderSystemMessage(content string, timestamp time.Time) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Muted).Render("◇")
 	label := lipgloss.NewStyle().Foreground(theme.Muted).Bold(true).Render("System")
 
@@ -259,7 +240,7 @@ func (r *CompactRenderer) RenderSystemMessage(content string, timestamp time.Tim
 // distinctive error symbol (!) and styling to ensure visibility. The error
 // content is displayed in a single line with appropriate color highlighting.
 func (r *CompactRenderer) RenderErrorMessage(errorMsg string, timestamp time.Time) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Error).Render("!")
 	label := lipgloss.NewStyle().Foreground(theme.Error).Bold(true).Render("Error")
 
@@ -279,7 +260,7 @@ func (r *CompactRenderer) RenderErrorMessage(errorMsg string, timestamp time.Tim
 // mode is enabled. Messages are truncated if they exceed the available width to
 // maintain single-line display.
 func (r *CompactRenderer) RenderDebugMessage(message string, timestamp time.Time) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Tool).Render("*")
 	label := lipgloss.NewStyle().Foreground(theme.Tool).Bold(true).Render("Debug")
 
@@ -303,7 +284,7 @@ func (r *CompactRenderer) RenderDebugMessage(message string, timestamp time.Time
 // debugging purposes. Config entries are displayed as key=value pairs separated
 // by commas, truncated if necessary to fit on a single line.
 func (r *CompactRenderer) RenderDebugConfigMessage(config map[string]any, timestamp time.Time) UIMessage {
-	theme := getTheme()
+	theme := GetTheme()
 	symbol := lipgloss.NewStyle().Foreground(theme.Tool).Render("*")
 	label := lipgloss.NewStyle().Foreground(theme.Tool).Bold(true).Render("Debug")
 
@@ -421,32 +402,6 @@ func (r *CompactRenderer) wrapText(text string, width int) string {
 	return strings.Join(wrappedLines, "\n")
 }
 
-// formatToolArgs formats tool arguments for compact display
-func (r *CompactRenderer) formatToolArgs(args string) string {
-	if args == "" || args == "{}" {
-		return ""
-	}
-
-	// Remove JSON braces and format compactly
-	args = strings.TrimSpace(args)
-	if strings.HasPrefix(args, "{") && strings.HasSuffix(args, "}") {
-		args = strings.TrimPrefix(args, "{")
-		args = strings.TrimSuffix(args, "}")
-		args = strings.TrimSpace(args)
-	}
-
-	// Remove quotes around simple values
-	args = strings.ReplaceAll(args, `"`, "")
-
-	// Remove parameter names (e.g., "command: ls" -> "ls", "path: /home" -> "/home")
-	// Look for pattern "key: value" and extract just the value
-	if colonIndex := strings.Index(args, ":"); colonIndex != -1 {
-		args = strings.TrimSpace(args[colonIndex+1:])
-	}
-
-	return r.formatCompactContent(args)
-}
-
 // formatToolResult formats tool results preserving formatting but limiting to 5 lines
 func (r *CompactRenderer) formatToolResult(result string) string {
 	if result == "" {
@@ -485,5 +440,5 @@ func (r *CompactRenderer) formatToolResult(result string) string {
 // and styling appropriately. Delegates tag parsing to the shared parseBashOutput
 // helper.
 func (r *CompactRenderer) formatBashOutput(result string) string {
-	return parseBashOutput(result, getTheme())
+	return parseBashOutput(result, GetTheme())
 }

internal/ui/enhanced_styles.go

@@ -35,8 +35,11 @@ func GetTheme() Theme {
 
 // SetTheme updates the global UI theme, affecting all subsequent rendering
 // operations. This allows runtime theme switching for different visual preferences.
+// It also invalidates the markdownTypographyCache so the next call to
+// GetMarkdownTypography picks up the new theme.
 func SetTheme(theme Theme) {
 	currentTheme = theme
+	markdownTypographyCache = nil // invalidate cached renderer; colors may have changed
 }
 
 // MarkdownThemeColors defines colors for markdown rendering and syntax highlighting.
@@ -291,45 +294,3 @@ func ApplyGradient(text string, colorA, colorB color.Color) string {
 
 	return result.String()
 }
-
-// CreateGradientText creates styled text with a gradient effect between two colors.
-func CreateGradientText(text string, startColor, endColor color.Color) string {
-	return ApplyGradient(text, startColor, endColor)
-}
-
-// Compact styling utilities
-
-// StyleCompactSymbol creates a lipgloss style for message type indicators in
-// compact mode, using bold colored text to distinguish different message categories.
-func StyleCompactSymbol(symbol string, c color.Color) lipgloss.Style {
-	return lipgloss.NewStyle().
-		Foreground(c).
-		Bold(true)
-}
-
-// StyleCompactLabel creates a lipgloss style for message labels in compact mode
-// with fixed width for alignment and bold colored text for readability.
-func StyleCompactLabel(c color.Color) lipgloss.Style {
-	return lipgloss.NewStyle().
-		Foreground(c).
-		Bold(true).
-		Width(8)
-}
-
-// StyleCompactContent creates a simple lipgloss style for message content in
-// compact mode, applying only color without additional formatting.
-func StyleCompactContent(c color.Color) lipgloss.Style {
-	return lipgloss.NewStyle().
-		Foreground(c)
-}
-
-// FormatCompactLine assembles a complete compact mode message line with consistent
-// spacing and styling. Combines a symbol, fixed-width label, and content with their
-// respective colors to create a uniform appearance across all message types.
-func FormatCompactLine(symbol, label, content string, symbolColor, labelColor, contentColor color.Color) string {
-	styledSymbol := StyleCompactSymbol(symbol, symbolColor).Render(symbol)
-	styledLabel := StyleCompactLabel(labelColor).Render(label)
-	styledContent := StyleCompactContent(contentColor).Render(content)
-
-	return fmt.Sprintf("%s  %-8s %s", styledSymbol, styledLabel, styledContent)
-}

internal/ui/file_suggestions.go

@@ -6,7 +6,6 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
-	"unicode/utf8"
 )
 
 // FileSuggestion represents a single file or directory suggestion for the @
@@ -345,44 +344,14 @@ func scoreFilePath(query, path string) int {
 	}
 
 	// Fuzzy character match on basename.
-	if score := fuzzyCharMatch(query, baseNameLower); score > 0 {
+	if score := fuzzyCharacterMatch(query, baseNameLower); score > 0 {
 		return score
 	}
 
 	// Fuzzy character match on full path.
-	if score := fuzzyCharMatch(query, pathLower); score > 0 {
+	if score := fuzzyCharacterMatch(query, pathLower); score > 0 {
 		return score - 50
 	}
 
 	return 0
 }
-
-// fuzzyCharMatch performs character-by-character fuzzy matching. Returns a
-// positive score if all query characters appear in order in the target.
-func fuzzyCharMatch(query, target string) int {
-	if utf8.RuneCountInString(query) > utf8.RuneCountInString(target) {
-		return 0
-	}
-
-	qRunes := []rune(query)
-	tRunes := []rune(target)
-	qi := 0
-	score := 100
-	consecutive := 0
-
-	for ti := 0; ti < len(tRunes) && qi < len(qRunes); ti++ {
-		if tRunes[ti] == qRunes[qi] {
-			qi++
-			consecutive++
-			score += consecutive * 5
-		} else {
-			consecutive = 0
-			score -= 2
-		}
-	}
-
-	if qi < len(qRunes) {
-		return 0
-	}
-	return score
-}

internal/ui/fuzzy.go

@@ -113,19 +113,23 @@ func fuzzyScore(query string, cmd *SlashCommand) int {
 	return 0
 }
 
-// fuzzyCharacterMatch performs character-by-character fuzzy matching
+// fuzzyCharacterMatch performs character-by-character fuzzy matching using
+// rune-safe iteration so multi-byte Unicode characters are handled correctly.
+// Returns a positive score if all query runes appear in order within target.
 func fuzzyCharacterMatch(query, target string) int {
-	if len(query) > len(target) {
+	qRunes := []rune(query)
+	tRunes := []rune(target)
+	if len(qRunes) > len(tRunes) {
 		return 0
 	}
 
-	queryIdx := 0
+	qi := 0
 	score := 100
 	consecutiveMatches := 0
 
-	for i := 0; i < len(target) && queryIdx < len(query); i++ {
-		if target[i] == query[queryIdx] {
-			queryIdx++
+	for ti := 0; ti < len(tRunes) && qi < len(qRunes); ti++ {
+		if tRunes[ti] == qRunes[qi] {
+			qi++
 			consecutiveMatches++
 			score += consecutiveMatches * 10
 		} else {
@@ -135,7 +139,7 @@ func fuzzyCharacterMatch(query, target string) int {
 	}
 
 	// Must match all characters in query
-	if queryIdx < len(query) {
+	if qi < len(qRunes) {
 		return 0
 	}
 

internal/ui/input.go

@@ -65,6 +65,10 @@ type InputComponent struct {
 	// hideHint suppresses the "enter submit · ctrl+j..." hint text.
 	hideHint bool
 
+	// agentBusy indicates the agent is currently working. When true, the
+	// hint text shows steering shortcut (Ctrl+S) instead of submit.
+	agentBusy bool
+
 	// pendingImages holds clipboard images attached to the next submission.
 	// Images are added via Ctrl+V and cleared on submit or Ctrl+U.
 	pendingImages []ImageAttachment
@@ -405,21 +409,14 @@ func (s *InputComponent) handleSubmit(value string) tea.Cmd {
 	}
 
 	// Resolve via canonical command lookup so aliases are handled uniformly.
-	// Only /quit and /clear are handled locally — /clear-queue must go
-	// through the parent model so it can update queueCount directly
-	// (calling ClearQueue here would skip the UI state update since we
-	// can't send events from within Update without deadlocking).
+	// Only /quit is handled locally — all other slash commands (including
+	// /clear and /clear-queue) are forwarded to the parent model via
+	// submitMsg so the parent can update its own state (scrollback, queue
+	// counts, etc.) in one place.
 	if sc := GetCommandByName(trimmed); sc != nil {
 		switch sc.Name {
 		case "/quit":
 			return tea.Quit
-
-		case "/clear":
-			if s.appCtrl != nil {
-				s.appCtrl.ClearMessages()
-			}
-			// Don't forward to app.Run(); just clear silently.
-			return nil
 		}
 	}
 
@@ -514,7 +511,16 @@ func (s *InputComponent) View() tea.View {
 		// Adapt hint text to available width (accounting for left padding of 3).
 		var hint string
 		availableHintWidth := s.width - 3
-		if availableHintWidth >= 67 {
+		if s.agentBusy {
+			// When the agent is working, show steering shortcut.
+			if availableHintWidth >= 55 {
+				hint = "enter queue • ctrl+s steer • esc esc cancel"
+			} else if availableHintWidth >= 35 {
+				hint = "↵ queue • ^S steer • esc×2 cancel"
+			} else {
+				hint = "^S steer"
+			}
+		} else if availableHintWidth >= 67 {
 			hint = "enter submit • ctrl+j / shift+enter new line • ctrl+v paste image"
 		} else if availableHintWidth >= 40 {
 			hint = "↵ submit • ctrl+j newline • ctrl+v image"

internal/ui/messages.go

@@ -9,6 +9,7 @@ import (
 	"time"
 
 	"charm.land/lipgloss/v2"
+	"github.com/indaco/herald"
 )
 
 // ansiEscapeRe matches ANSI escape sequences used for terminal styling.
@@ -22,9 +23,9 @@ const (
 	UserMessage MessageType = iota
 	AssistantMessage
 	ToolMessage
-	ToolCallMessage // New type for showing tool calls in progress
-	SystemMessage   // New type for KIT system messages (help, tools, etc.)
-	ErrorMessage    // New type for error messages
+	ToolCallMessage
+	SystemMessage
+	ErrorMessage
 )
 
 // UIMessage encapsulates a fully rendered message ready for display in the UI,
@@ -40,29 +41,9 @@ type UIMessage struct {
 	Streaming bool
 }
 
-// Helper functions to get theme colors
-func getTheme() Theme {
-	return GetTheme()
-}
-
-// toolDisplayNames maps raw tool names to human-friendly display names.
-var toolDisplayNames = map[string]string{
-	"bash":          "Bash",
-	"read":          "Read",
-	"write":         "Write",
-	"edit":          "Edit",
-	"grep":          "Grep",
-	"find":          "Find",
-	"ls":            "Ls",
-	"run_shell_cmd": "Bash",
-}
-
-// toolDisplayName returns a human-friendly display name for a tool.
-// Falls back to capitalizing the first letter of the raw name.
+// toolDisplayName returns a human-friendly display name for a tool,
+// title-casing the first letter of the raw name.
 func toolDisplayName(rawName string) string {
-	if display, ok := toolDisplayNames[rawName]; ok {
-		return display
-	}
 	if rawName != "" {
 		return strings.ToUpper(rawName[:1]) + rawName[1:]
 	}
@@ -70,8 +51,6 @@ func toolDisplayName(rawName string) string {
 }
 
 // formatToolParams formats tool input parameters for inline header display.
-// Extracts the primary parameter (command/filePath) first, then shows
-// remaining params as (key=val, ...). Truncates to maxWidth.
 func formatToolParams(toolArgs string, maxWidth int) string {
 	args := strings.TrimSpace(toolArgs)
 	if args == "" || args == "{}" {
@@ -80,7 +59,6 @@ func formatToolParams(toolArgs string, maxWidth int) string {
 
 	var params map[string]any
 	if err := json.Unmarshal([]byte(args), &params); err != nil {
-		// Fallback: strip braces and return raw content
 		args = strings.TrimPrefix(args, "{")
 		args = strings.TrimSuffix(args, "}")
 		args = strings.TrimSpace(args)
@@ -94,7 +72,6 @@ func formatToolParams(toolArgs string, maxWidth int) string {
 		return ""
 	}
 
-	// Identify primary parameter by checking known keys in priority order
 	primaryKeys := []string{"command", "filePath", "path", "pattern", "query", "url"}
 	var primaryKey string
 	var primaryVal string
@@ -111,14 +88,24 @@ func formatToolParams(toolArgs string, maxWidth int) string {
 		result.WriteString(primaryVal)
 	}
 
-	// Collect remaining parameters (skip large values like file content)
+	bodyKeys := map[string]bool{
+		"content":  true,
+		"old_text": true,
+		"new_text": true,
+		"oldText":  true,
+		"newText":  true,
+		"edits":    true,
+		"todos":    true,
+	}
 	var remaining []string
 	for key, val := range params {
 		if key == primaryKey {
 			continue
 		}
+		if bodyKeys[key] {
+			continue
+		}
 		valStr := fmt.Sprintf("%v", val)
-		// Skip very large values (e.g., oldString, newString, content, todos)
 		if len(valStr) > 100 {
 			continue
 		}
@@ -143,65 +130,35 @@ func formatToolParams(toolArgs string, maxWidth int) string {
 }
 
 // MessageRenderer handles the formatting and rendering of different message types
-// with consistent styling, markdown support, and appropriate visual hierarchies
-// for the standard (non-compact) display mode.
 type MessageRenderer struct {
-	width int
-	debug bool
-
-	// getToolRenderer returns extension-provided rendering overrides for a
-	// specific tool. May be nil if no extensions are loaded. Used in
-	// RenderToolMessage to check for custom header/body formatting before
-	// falling back to builtin renderers.
+	width           int
+	debug           bool
+	ty              *herald.Typography
 	getToolRenderer func(toolName string) *ToolRendererData
 }
 
-// newMessageRenderer creates and initializes a new MessageRenderer with the specified
-// terminal width and debug mode setting. The width parameter determines line wrapping
-// and layout calculations.
+// newMessageRenderer creates and initializes a new MessageRenderer
 func newMessageRenderer(width int, debug bool) *MessageRenderer {
 	return &MessageRenderer{
 		width: width,
 		debug: debug,
+		ty:    createTypography(GetTheme()),
 	}
 }
 
-// SetWidth updates the terminal width for the renderer, affecting how content
-// is wrapped and formatted in subsequent render operations.
+// SetWidth updates the terminal width for the renderer
 func (r *MessageRenderer) SetWidth(width int) {
 	r.width = width
 }
 
-// RenderUserMessage renders a user's input message with distinctive right-aligned
-// formatting, including the system username, timestamp, and markdown-rendered content.
-// The message is displayed with a colored right border for visual distinction.
+// RenderUserMessage renders a user's input message using herald Tip alert
 func (r *MessageRenderer) RenderUserMessage(content string, timestamp time.Time) UIMessage {
-	theme := getTheme()
-
-	// Only run markdown rendering when the message contains code spans or
-	// fenced code blocks. Plain text is rendered directly so that newlines
-	// are preserved without the extra paragraph spacing glamour adds.
-	var messageContent string
-	if strings.Contains(content, "`") {
-		// Glamour treats single \n as a soft break, so convert to paragraph
-		// breaks and collapse the resulting blank lines after rendering.
-		mdContent := strings.ReplaceAll(content, "\n", "\n\n")
-		messageContent = r.renderMarkdown(mdContent, r.width-8)
-		messageContent = removeBlankLines(messageContent)
-	} else {
-		messageContent = content
+	if strings.TrimSpace(content) == "" {
+		content = "(empty message)"
 	}
 
-	fullContent := strings.TrimSuffix(messageContent, "\n")
-
-	// Left border with Blue color for user messages.
-	rendered := renderContentBlock(
-		fullContent,
-		r.width,
-		WithAlign(lipgloss.Left),
-		WithBorderColor(theme.Info),
-		WithMarginBottom(1),
-	)
+	rendered := r.ty.Tip(content)
+	rendered = styleMarginBottom1.Render(rendered)
 
 	return UIMessage{
 		Type:      UserMessage,
@@ -211,33 +168,20 @@ func (r *MessageRenderer) RenderUserMessage(content string, timestamp time.Time)
 	}
 }
 
-// RenderAssistantMessage renders an AI assistant's response with left-aligned formatting,
-// including the model name, timestamp, and markdown-rendered content. Empty responses
-// are displayed with a special "Finished without output" message. The message features
-// a colored left border for visual distinction.
+// RenderAssistantMessage renders an AI assistant's response
 func (r *MessageRenderer) RenderAssistantMessage(content string, timestamp time.Time, modelName string) UIMessage {
-	theme := getTheme()
-
-	var messageContent string
 	if strings.TrimSpace(content) == "" {
-		messageContent = lipgloss.NewStyle().
-			Italic(true).
-			Foreground(theme.Muted).
-			Align(lipgloss.Center).
-			Render("Finished without output")
-	} else {
-		messageContent = r.renderMarkdown(content, r.width-8)
+		return UIMessage{
+			Type:      AssistantMessage,
+			Content:   "",
+			Height:    0,
+			Timestamp: timestamp,
+		}
 	}
 
-	fullContent := strings.TrimSuffix(messageContent, "\n")
-
-	// Left border with Primary (Mauve) color for assistant messages.
-	rendered := renderContentBlock(
-		fullContent,
-		r.width,
-		WithBorderColor(theme.Primary),
-		WithMarginBottom(1),
-	)
+	// Use markdown rendering with Chroma syntax highlighting
+	rendered := toMarkdown(content, r.width-4)
+	rendered = styleMarginBottom1.Render(rendered)
 
 	return UIMessage{
 		Type:      AssistantMessage,
@@ -247,30 +191,14 @@ func (r *MessageRenderer) RenderAssistantMessage(content string, timestamp time.
 	}
 }
 
-// RenderSystemMessage renders KIT system messages such as help text, command outputs,
-// and informational notifications. These messages are displayed with a distinctive system
-// color border and "KIT System" label to differentiate them from user and AI content.
+// RenderSystemMessage renders KIT system messages using herald Note alert
 func (r *MessageRenderer) RenderSystemMessage(content string, timestamp time.Time) UIMessage {
-	theme := getTheme()
-
-	var messageContent string
 	if strings.TrimSpace(content) == "" {
-		messageContent = "No content available"
-	} else if strings.Contains(content, "`") {
-		messageContent = r.renderMarkdown(content, r.width-8)
-	} else {
-		messageContent = content
+		content = "No content available"
 	}
 
-	fullContent := "◇ " + strings.TrimSuffix(messageContent, "\n")
-
-	rendered := renderContentBlock(
-		fullContent,
-		r.width,
-		WithNoBorder(),
-		WithForeground(theme.Muted),
-		WithMarginBottom(1),
-	)
+	rendered := r.ty.Note(content)
+	rendered = styleMarginBottom1.Render(rendered)
 
 	return UIMessage{
 		Type:      SystemMessage,
@@ -280,27 +208,9 @@ func (r *MessageRenderer) RenderSystemMessage(content string, timestamp time.Tim
 	}
 }
 
-// RenderDebugMessage renders diagnostic and debugging information with special formatting
-// including a debug icon, colored border, and structured layout. Debug messages are only
-// displayed when debug mode is enabled and help developers troubleshoot issues.
+// RenderDebugMessage renders diagnostic and debugging information
 func (r *MessageRenderer) RenderDebugMessage(message string, timestamp time.Time) UIMessage {
-	baseStyle := lipgloss.NewStyle()
-
-	theme := getTheme()
-	style := baseStyle.
-		Width(r.width - 3).
-		BorderLeft(true).
-		Foreground(theme.Muted).
-		BorderForeground(theme.Tool).
-		BorderStyle(lipgloss.ThickBorder()).
-		PaddingLeft(1).
-		MarginLeft(2).
-		MarginBottom(1)
-
-	header := baseStyle.
-		Foreground(theme.Tool).
-		Bold(true).
-		Render("🔍 Debug Output")
+	header := r.ty.H6("🔍 Debug Output")
 
 	lines := strings.Split(message, "\n")
 	var formattedLines []string
@@ -310,87 +220,52 @@ func (r *MessageRenderer) RenderDebugMessage(message string, timestamp time.Time
 		}
 	}
 
-	content := baseStyle.
-		Foreground(theme.Muted).
-		Render(strings.Join(formattedLines, "\n"))
-
-	fullContent := lipgloss.JoinVertical(lipgloss.Left,
+	content := r.ty.Compose(
 		header,
-		content,
+		r.ty.P(strings.Join(formattedLines, "\n")),
 	)
+	content = styleMarginBottom1.Render(content)
 
 	return UIMessage{
-		Content: style.Render(fullContent),
-		Height:  lipgloss.Height(style.Render(fullContent)),
+		Content: content,
+		Height:  lipgloss.Height(content),
 	}
 }
 
-// RenderDebugConfigMessage renders configuration settings in a formatted debug display
-// with key-value pairs shown in a structured layout. Used to display runtime configuration
-// for debugging purposes with a distinctive icon and border styling.
+// RenderDebugConfigMessage renders configuration settings
 func (r *MessageRenderer) RenderDebugConfigMessage(config map[string]any, timestamp time.Time) UIMessage {
-	baseStyle := lipgloss.NewStyle()
-
-	theme := getTheme()
-	style := baseStyle.
-		Width(r.width - 1).
-		BorderLeft(true).
-		Foreground(theme.Muted).
-		BorderForeground(theme.Tool).
-		BorderStyle(lipgloss.ThickBorder()).
-		PaddingLeft(1)
-
-	header := baseStyle.
-		Foreground(theme.Tool).
-		Bold(true).
-		Render("🔧 Debug Configuration")
+	header := r.ty.H6("🔧 Debug Configuration")
 
 	var configLines []string
 	for key, value := range config {
 		if value != nil {
-			configLines = append(configLines, fmt.Sprintf("  %s: %v", key, value))
+			configLines = append(configLines, fmt.Sprintf("%s: %v", key, value))
 		}
 	}
 
-	configContent := baseStyle.
-		Foreground(theme.Muted).
-		Render(strings.Join(configLines, "\n"))
-
-	parts := []string{header}
+	var content string
 	if len(configLines) > 0 {
-		parts = append(parts, configContent)
+		content = r.ty.Compose(
+			header,
+			r.ty.P(strings.Join(configLines, "\n")),
+		)
+	} else {
+		content = header
 	}
-
-	rendered := style.Render(
-		lipgloss.JoinVertical(lipgloss.Left, parts...),
-	)
+	content = styleMarginBottom1.Render(content)
 
 	return UIMessage{
 		Type:      SystemMessage,
-		Content:   rendered,
-		Height:    lipgloss.Height(rendered),
+		Content:   content,
+		Height:    lipgloss.Height(content),
 		Timestamp: timestamp,
 	}
 }
 
-// RenderErrorMessage renders error notifications with distinctive red coloring and
-// bold text to ensure visibility. Error messages include timestamp information and
-// are displayed with an error-colored border for immediate recognition.
+// RenderErrorMessage renders error notifications
 func (r *MessageRenderer) RenderErrorMessage(errorMsg string, timestamp time.Time) UIMessage {
-	theme := getTheme()
-
-	errorContent := lipgloss.NewStyle().
-		Foreground(theme.Error).
-		Bold(true).
-		Render(errorMsg)
-
-	rendered := renderContentBlock(
-		errorContent,
-		r.width,
-		WithAlign(lipgloss.Left),
-		WithBorderColor(theme.Error),
-		WithMarginBottom(1),
-	)
+	rendered := r.ty.Caution(errorMsg)
+	rendered = styleMarginBottom1.Render(rendered)
 
 	return UIMessage{
 		Type:      ErrorMessage,
@@ -400,93 +275,18 @@ func (r *MessageRenderer) RenderErrorMessage(errorMsg string, timestamp time.Tim
 	}
 }
 
-// RenderToolCallMessage renders a notification that a tool is being executed, showing
-// the tool name, formatted arguments (if any), and execution timestamp. The message
-// uses tool-specific coloring to distinguish it from regular conversation messages.
-func (r *MessageRenderer) RenderToolCallMessage(toolName, toolArgs string, timestamp time.Time) UIMessage {
-	// Format timestamp
-	timeStr := timestamp.Local().Format("15:04")
-
-	// Format arguments with better presentation
-	theme := getTheme()
-	var argsContent string
-	if toolArgs != "" && toolArgs != "{}" {
-		argsContent = lipgloss.NewStyle().
-			Foreground(theme.Muted).
-			Italic(true).
-			Render(fmt.Sprintf("Arguments: %s", r.formatToolArgs(toolArgs)))
-	}
-
-	// Create info line
-	info := fmt.Sprintf(" Executing %s (%s)", toolName, timeStr)
-
-	// Combine parts
-	var fullContent string
-	if argsContent != "" {
-		fullContent = argsContent + "\n" +
-			lipgloss.NewStyle().Foreground(theme.VeryMuted).Render(info)
-	} else {
-		fullContent = lipgloss.NewStyle().Foreground(theme.VeryMuted).Render(info)
-	}
-
-	// Use the new block renderer
-	rendered := renderContentBlock(
-		fullContent,
-		r.width,
-		WithAlign(lipgloss.Left),
-		WithBorderColor(theme.Tool),
-		WithMarginBottom(1),
-	)
-
-	return UIMessage{
-		Type:      ToolCallMessage,
-		Content:   rendered,
-		Height:    lipgloss.Height(rendered),
-		Timestamp: timestamp,
-	}
-}
-
-// RenderToolMessage renders a unified tool block combining the tool invocation
-// header (icon + display name + params) with the execution result body. The
-// border color indicates status: green for success, red for error. This replaces
-// the previous two-block approach (separate call + result blocks).
+// RenderToolMessage renders a unified tool block
 func (r *MessageRenderer) RenderToolMessage(toolName, toolArgs, toolResult string, isError bool) UIMessage {
-	theme := getTheme()
-
-	// Resolve extension renderer once for all overrides.
 	var extRd *ToolRendererData
 	if r.getToolRenderer != nil {
 		extRd = r.getToolRenderer(toolName)
 	}
 
-	// --- Header: [icon] [name] [params] ---
-	var icon string
-	borderColor := theme.Success
-	iconColor := theme.Success
-	if isError {
-		icon = "×"
-		borderColor = theme.Error
-		iconColor = theme.Error
-	} else {
-		icon = "✓"
-	}
-
-	// Extension can override border color (applies to both success and error).
-	if extRd != nil && extRd.BorderColor != "" {
-		borderColor = lipgloss.Color(extRd.BorderColor)
-	}
-
-	iconStr := lipgloss.NewStyle().Foreground(iconColor).Bold(true).Render(icon)
-
-	// Extension can override display name.
 	displayName := toolDisplayName(toolName)
 	if extRd != nil && extRd.DisplayName != "" {
 		displayName = extRd.DisplayName
 	}
-	nameStr := lipgloss.NewStyle().Foreground(theme.Info).Bold(true).Render(displayName)
 
-	// Format params with width budget for the header line.
-	// Check extension renderer for custom header params first.
 	paramBudget := max(r.width-10-len(displayName), 20)
 	var params string
 	if extRd != nil && extRd.RenderHeader != nil {
@@ -496,97 +296,70 @@ func (r *MessageRenderer) RenderToolMessage(toolName, toolArgs, toolResult strin
 		params = formatToolParams(toolArgs, paramBudget)
 	}
 
-	header := iconStr + " " + nameStr
-	if params != "" {
-		header += " " + lipgloss.NewStyle().Foreground(theme.Muted).Render(params)
+	var icon string
+	iconColor := GetTheme().Success
+	if isError {
+		icon = "×"
+		iconColor = GetTheme().Error
+	} else {
+		icon = "✓"
 	}
 
-	// --- Body: check extension renderer first, then builtin, then default ---
+	// Style the tool name with color
+	theme := GetTheme()
+	nameColor := theme.Info
+	if isError {
+		nameColor = theme.Error
+	}
+	styledName := lipgloss.NewStyle().Foreground(nameColor).Bold(true).Render(displayName)
+	styledIcon := lipgloss.NewStyle().Foreground(iconColor).Render(icon)
+
+	// Build the content: icon + name + params on first line, then body
+	headerLine := styledIcon + " " + styledName
+	if params != "" {
+		headerLine += " " + lipgloss.NewStyle().Foreground(theme.Muted).Render(params)
+	}
+
+	// Get body content
 	var body string
 	if extRd != nil && extRd.RenderBody != nil {
 		body = extRd.RenderBody(toolResult, isError, r.width-8)
-		// Apply markdown rendering if requested and body is non-empty.
 		if body != "" && extRd.BodyMarkdown {
 			body = strings.TrimSuffix(toMarkdown(body, r.width-8), "\n")
 		}
 	}
 	if body == "" {
 		if isError {
-			body = lipgloss.NewStyle().
-				Foreground(theme.Error).
-				Render(toolResult)
+			body = r.formatToolResult(toolName, toolResult)
 		} else {
 			body = renderToolBody(toolName, toolArgs, toolResult, r.width-8)
 			if body == "" {
-				body = r.formatToolResult(toolName, toolResult, r.width-8)
+				body = r.formatToolResult(toolName, toolResult)
 			}
 		}
 	}
 
 	if strings.TrimSpace(body) == "" {
-		body = lipgloss.NewStyle().
-			Italic(true).
-			Foreground(theme.Muted).
-			Render("(no output)")
+		body = r.ty.Italic("(no output)")
 	}
 
-	// Combine header + body into a single block.
-	fullContent := header + "\n\n" + strings.TrimSuffix(body, "\n")
-
-	// Build rendering options; extension can override background.
-	blockOpts := []renderingOption{
-		WithAlign(lipgloss.Left),
-		WithBorderColor(borderColor),
-		WithMarginBottom(1),
-	}
-	if extRd != nil && extRd.Background != "" {
-		blockOpts = append(blockOpts, WithBackground(lipgloss.Color(extRd.Background)))
-	}
-
-	rendered := renderContentBlock(
-		fullContent,
-		r.width,
-		blockOpts...,
+	// Compose: icon + name + params, then body
+	fullContent := r.ty.Compose(
+		headerLine,
+		"",
+		body,
 	)
+	fullContent = styleMarginBottom1.Render(fullContent)
 
 	return UIMessage{
 		Type:    ToolMessage,
-		Content: rendered,
-		Height:  lipgloss.Height(rendered),
+		Content: fullContent,
+		Height:  lipgloss.Height(fullContent),
 	}
 }
 
-// formatToolArgs formats tool arguments for display
-func (r *MessageRenderer) formatToolArgs(args string) string {
-	// Remove outer braces and clean up JSON formatting
-	args = strings.TrimSpace(args)
-	if strings.HasPrefix(args, "{") && strings.HasSuffix(args, "}") {
-		args = strings.TrimPrefix(args, "{")
-		args = strings.TrimSuffix(args, "}")
-		args = strings.TrimSpace(args)
-	}
-
-	// If it's empty after cleanup, return a placeholder
-	if args == "" {
-		return "(no arguments)"
-	}
-
-	// Truncate if too long, but skip truncation in debug mode
-	if !r.debug {
-		maxLen := 100
-		if len(args) > maxLen {
-			return args[:maxLen] + "..."
-		}
-	}
-
-	return args
-}
-
 // formatToolResult formats tool results based on tool type
-func (r *MessageRenderer) formatToolResult(toolName, result string, width int) string {
-	baseStyle := lipgloss.NewStyle()
-
-	// Truncate very long results only if not in debug mode
+func (r *MessageRenderer) formatToolResult(toolName, result string) string {
 	if !r.debug {
 		maxLines := 10
 		lines := strings.Split(result, "\n")
@@ -595,51 +368,47 @@ func (r *MessageRenderer) formatToolResult(toolName, result string, width int) s
 		}
 	}
 
-	// Format bash/command output with better formatting
-	if strings.Contains(toolName, "bash") || strings.Contains(toolName, "command") || strings.Contains(toolName, "shell") || toolName == "run_shell_cmd" {
-		theme := getTheme()
-
-		// Split result into sections if it contains both stdout and stderr
+	if strings.Contains(toolName, "bash") || strings.Contains(toolName, "command") ||
+		strings.Contains(toolName, "shell") {
 		if strings.Contains(result, "<stdout>") || strings.Contains(result, "<stderr>") {
-			return r.formatBashOutput(result, width, theme)
+			return parseBashOutput(result, GetTheme())
 		}
-
-		// For simple output, just render as monospace text with proper line breaks
-		return baseStyle.
-			Width(width).
-			Foreground(theme.Muted).
-			Render(result)
 	}
 
-	// For other tools, render as muted text
-	theme := getTheme()
-	return baseStyle.
-		Width(width).
-		Foreground(theme.Muted).
-		Render(result)
+	return result
 }
 
-// formatBashOutput formats bash command output with proper section handling.
-// Delegates tag parsing to the shared parseBashOutput helper.
-func (r *MessageRenderer) formatBashOutput(result string, width int, theme Theme) string {
-	parsed := parseBashOutput(result, theme)
-	return lipgloss.NewStyle().
-		Width(width).
-		Foreground(theme.Muted).
-		Render(parsed)
-}
-
-// renderMarkdown renders markdown content using glamour
-func (r *MessageRenderer) renderMarkdown(content string, width int) string {
-	rendered := toMarkdown(content, width)
-	return strings.TrimSuffix(rendered, "\n")
+// createTypography creates a typography instance from theme
+func createTypography(theme Theme) *herald.Typography {
+	return herald.New(
+		herald.WithPalette(herald.ColorPalette{
+			Primary:   theme.Primary,
+			Secondary: theme.Secondary,
+			Tertiary:  theme.Info,
+			Accent:    theme.Accent,
+			Highlight: theme.Highlight,
+			Muted:     theme.Muted,
+			Text:      theme.Text,
+			Surface:   theme.Background,
+			Base:      theme.CodeBg,
+		}),
+		herald.WithAlertPalette(herald.AlertPalette{
+			Note:      theme.Info,
+			Tip:       theme.Success,
+			Important: theme.Accent,
+			Warning:   theme.Warning,
+			Caution:   theme.Error,
+		}),
+		herald.WithCodeLineNumbers(true),
+		// Customize alert labels
+		herald.WithAlertLabel(herald.AlertNote, "Info"),
+		herald.WithAlertLabel(herald.AlertTip, "You"),
+		herald.WithAlertLabel(herald.AlertWarning, "Working"),
+		herald.WithAlertLabel(herald.AlertCaution, "Error"),
+	)
 }
 
 // removeBlankLines removes lines that are visually blank from rendered output.
-// Glamour wraps every character (including padding spaces) with ANSI color
-// codes, so we must strip escape sequences before checking whether a line is
-// empty. This collapses paragraph spacing so user messages render without
-// extra vertical gaps.
 func removeBlankLines(s string) string {
 	lines := strings.Split(s, "\n")
 	filtered := lines[:0]

internal/ui/model_selector.go

@@ -50,7 +50,7 @@ func NewModelSelector(currentModel string, width, height int) *ModelSelectorComp
 	registry := models.GetGlobalRegistry()
 	var allModels []ModelEntry
 
-	for _, providerID := range registry.GetFantasyProviders() {
+	for _, providerID := range registry.GetLLMProviders() {
 		// Only include providers with valid API keys configured.
 		if err := registry.ValidateEnvironment(providerID, ""); err != nil {
 			continue

internal/ui/model_test.go

@@ -46,6 +46,10 @@ func (s *stubAppController) ClearMessages() {
 	s.clearMsgCalled++
 }
 
+func (s *stubAppController) ReloadMessagesFromTree() {
+	// no-op in tests
+}
+
 func (s *stubAppController) CompactConversation(_ string) error {
 	return nil
 }
@@ -54,6 +58,10 @@ func (s *stubAppController) GetTreeSession() *session.TreeManager {
 	return nil
 }
 
+func (s *stubAppController) SwitchTreeSession(_ *session.TreeManager) {
+	// no-op in tests
+}
+
 func (s *stubAppController) SendEvent(_ tea.Msg) {
 	// no-op in tests
 }
@@ -67,6 +75,11 @@ func (s *stubAppController) RunWithFiles(prompt string, _ []fantasy.FilePart) in
 	return s.queueLen
 }
 
+func (s *stubAppController) Steer(prompt string) int {
+	s.runCalls = append(s.runCalls, prompt)
+	return s.queueLen
+}
+
 // --------------------------------------------------------------------------
 // Stub child components
 // --------------------------------------------------------------------------
@@ -112,15 +125,16 @@ func newTestAppModel(ctrl AppController) (*AppModel, *stubStreamComponent, *stub
 	stream := &stubStreamComponent{}
 	input := &stubInputComponent{}
 	m := &AppModel{
-		state:       stateInput,
-		appCtrl:     ctrl,
-		stream:      stream,
-		input:       input,
-		renderer:    newMessageRenderer(80, false),
-		compactMode: false,
-		modelName:   "test-model",
-		width:       80,
-		height:      24,
+		state:                 stateInput,
+		appCtrl:               ctrl,
+		stream:                stream,
+		input:                 input,
+		renderer:              newMessageRenderer(80, false),
+		compactMode:           false,
+		modelName:             "test-model",
+		width:                 80,
+		height:                24,
+		streamingBashMaxLines: 50, // Initialize buffer cap like NewAppModel does
 	}
 	return m, stream, input
 }
@@ -132,7 +146,11 @@ func newTestAppModel(ctrl AppController) (*AppModel, *stubStreamComponent, *stub
 // sendMsg calls m.Update once with the given message and returns the updated model.
 func sendMsg(m *AppModel, msg tea.Msg) *AppModel {
 	updated, _ := m.Update(msg)
-	return updated.(*AppModel)
+	result := updated.(*AppModel)
+	// Simulate BubbleTea's frame cycle: View() is called after every Update().
+	// This flushes any pending layoutDirty work (e.g. distributeHeight).
+	_ = result.View()
+	return result
 }
 
 // --------------------------------------------------------------------------
@@ -602,6 +620,133 @@ func TestToolResult_printsAndStartsSpinner(t *testing.T) {
 	}
 }
 
+// TestToolOutputEvent_accumulatesBashOutput verifies that ToolOutputEvent
+// accumulates stdout and stderr lines into the streaming bash output buffers.
+func TestToolOutputEvent_accumulatesBashOutput(t *testing.T) {
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+	m.state = stateWorking
+
+	// Send stdout chunk.
+	m = sendMsg(m, app.ToolOutputEvent{
+		ToolCallID: "call-1",
+		ToolName:   "bash",
+		Chunk:      "line one\n",
+		IsStderr:   false,
+	})
+
+	if len(m.streamingBashOutput) != 1 || m.streamingBashOutput[0] != "line one\n" {
+		t.Fatalf("expected streamingBashOutput=['line one\\n'], got %v", m.streamingBashOutput)
+	}
+	if len(m.streamingBashStderr) != 0 {
+		t.Fatalf("expected empty streamingBashStderr, got %v", m.streamingBashStderr)
+	}
+
+	// Send another stdout chunk.
+	m = sendMsg(m, app.ToolOutputEvent{
+		ToolCallID: "call-1",
+		ToolName:   "bash",
+		Chunk:      "line two\n",
+		IsStderr:   false,
+	})
+
+	if len(m.streamingBashOutput) != 2 {
+		t.Fatalf("expected 2 stdout lines, got %d", len(m.streamingBashOutput))
+	}
+
+	// Send stderr chunk.
+	m = sendMsg(m, app.ToolOutputEvent{
+		ToolCallID: "call-1",
+		ToolName:   "bash",
+		Chunk:      "error: something failed\n",
+		IsStderr:   true,
+	})
+
+	if len(m.streamingBashStderr) != 1 {
+		t.Fatalf("expected 1 stderr line, got %d", len(m.streamingBashStderr))
+	}
+	if m.streamingBashStderr[0] != "error: something failed\n" {
+		t.Fatalf("expected stderr 'error: something failed\\n', got %q", m.streamingBashStderr[0])
+	}
+}
+
+// TestToolResult_clearsStreamingBashOutput verifies that ToolResultEvent clears
+// the streaming bash output buffers since the final result will be printed.
+func TestToolResult_clearsStreamingBashOutput(t *testing.T) {
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+	m.state = stateWorking
+
+	// Accumulate some bash output.
+	m.streamingBashOutput = []string{"output line"}
+	m.streamingBashStderr = []string{"error line"}
+
+	_, _ = m.Update(app.ToolResultEvent{
+		ToolName: "bash",
+		ToolArgs: `{"cmd":"ls"}`,
+		Result:   "output line\nerror line\n",
+		IsError:  false,
+	})
+
+	if len(m.streamingBashOutput) != 0 {
+		t.Fatalf("expected streamingBashOutput cleared, got %v", m.streamingBashOutput)
+	}
+	if len(m.streamingBashStderr) != 0 {
+		t.Fatalf("expected streamingBashStderr cleared, got %v", m.streamingBashStderr)
+	}
+}
+
+// TestToolCallStarted_extractsBashCommand verifies that ToolCallStartedEvent
+// extracts the bash command from ToolArgs and stores it for the streaming output header.
+func TestToolCallStarted_extractsBashCommand(t *testing.T) {
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+	m.state = stateWorking
+
+	// Send ToolCallStartedEvent with bash command.
+	m = sendMsg(m, app.ToolCallStartedEvent{
+		ToolCallID: "call-1",
+		ToolName:   "bash",
+		ToolArgs:   `{"command":"ls -la /home"}`,
+	})
+
+	if m.streamingBashCommand != "ls -la /home" {
+		t.Fatalf("expected streamingBashCommand='ls -la /home', got %q", m.streamingBashCommand)
+	}
+
+	// ToolResultEvent should clear the command.
+	m = sendMsg(m, app.ToolResultEvent{
+		ToolCallID: "call-1",
+		ToolName:   "bash",
+		ToolArgs:   `{"command":"ls -la /home"}`,
+		Result:     "output",
+		IsError:    false,
+	})
+
+	if m.streamingBashCommand != "" {
+		t.Fatalf("expected streamingBashCommand cleared, got %q", m.streamingBashCommand)
+	}
+}
+
+// TestToolCallStarted_nonBashTool_doesNotSetCommand verifies that non-bash tools
+// do not set the streamingBashCommand field.
+func TestToolCallStarted_nonBashTool_doesNotSetCommand(t *testing.T) {
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+	m.state = stateWorking
+
+	// Send ToolCallStartedEvent with a non-bash tool.
+	m = sendMsg(m, app.ToolCallStartedEvent{
+		ToolCallID: "call-1",
+		ToolName:   "read",
+		ToolArgs:   `{"file":"/etc/passwd"}`,
+	})
+
+	if m.streamingBashCommand != "" {
+		t.Fatalf("expected streamingBashCommand to remain empty for non-bash tools, got %q", m.streamingBashCommand)
+	}
+}
+
 // TestStepError_printCmd verifies that StepErrorEvent with a non-nil error
 // produces a non-nil cmd (the tea.Println call for the error message).
 func TestStepError_printCmd(t *testing.T) {

internal/ui/preferences.go

@@ -0,0 +1,129 @@
+package ui
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+
+	"gopkg.in/yaml.v3"
+)
+
+// preferences holds user-mutable runtime state that persists across sessions.
+// Stored at ~/.config/kit/preferences.yml, separate from the declarative
+// .kit.yml config so we never clobber user comments or formatting.
+type preferences struct {
+	Theme         string `yaml:"theme,omitempty"`
+	Model         string `yaml:"model,omitempty"`
+	ThinkingLevel string `yaml:"thinking_level,omitempty"`
+}
+
+// preferencesPath returns ~/.config/kit/preferences.yml.
+// Returns "" if the config directory cannot be determined.
+func preferencesPath() string {
+	cfgDir, err := os.UserConfigDir()
+	if err != nil {
+		return ""
+	}
+	return filepath.Join(cfgDir, "kit", "preferences.yml")
+}
+
+// loadPreferences reads and parses the preferences file.
+// Returns zero-value preferences if the file is missing or invalid.
+func loadPreferences() preferences {
+	path := preferencesPath()
+	if path == "" {
+		return preferences{}
+	}
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return preferences{}
+	}
+	var prefs preferences
+	if err := yaml.Unmarshal(data, &prefs); err != nil {
+		return preferences{}
+	}
+	return prefs
+}
+
+// savePreferences atomically writes the preferences file, merging into any
+// existing content. The mutate function receives the current preferences and
+// should modify them in place.
+func savePreferences(mutate func(*preferences)) error {
+	path := preferencesPath()
+	if path == "" {
+		return nil // silently skip if config dir unavailable
+	}
+
+	// Load existing preferences to preserve other fields.
+	var prefs preferences
+	if data, err := os.ReadFile(path); err == nil {
+		_ = yaml.Unmarshal(data, &prefs)
+	}
+
+	mutate(&prefs)
+
+	data, err := yaml.Marshal(&prefs)
+	if err != nil {
+		return err
+	}
+
+	// Ensure parent directory exists.
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return err
+	}
+
+	// Atomic write: write to temp file, then rename.
+	tmp := path + ".tmp"
+	if err := os.WriteFile(tmp, data, 0o644); err != nil {
+		return err
+	}
+	return os.Rename(tmp, path)
+}
+
+// ── Theme preference ────────────────────────────────────────────────────────
+
+// LoadThemePreference reads the persisted theme name from preferences.yml.
+// Returns "" if no preference is saved or the file doesn't exist.
+func LoadThemePreference() string {
+	return strings.TrimSpace(loadPreferences().Theme)
+}
+
+// SaveThemePreference persists the theme name to ~/.config/kit/preferences.yml.
+// Preserves other preference fields. Uses atomic write (temp + rename) to
+// avoid corruption from concurrent Kit instances.
+func SaveThemePreference(name string) error {
+	return savePreferences(func(p *preferences) {
+		p.Theme = name
+	})
+}
+
+// ── Model preference ────────────────────────────────────────────────────────
+
+// LoadModelPreference reads the persisted model string (e.g.
+// "anthropic/claude-sonnet-4-5-20250929") from preferences.yml.
+// Returns "" if no preference is saved.
+func LoadModelPreference() string {
+	return strings.TrimSpace(loadPreferences().Model)
+}
+
+// SaveModelPreference persists the model string to preferences.yml.
+func SaveModelPreference(model string) error {
+	return savePreferences(func(p *preferences) {
+		p.Model = model
+	})
+}
+
+// ── Thinking level preference ───────────────────────────────────────────────
+
+// LoadThinkingLevelPreference reads the persisted thinking level from
+// preferences.yml. Returns "" if no preference is saved.
+func LoadThinkingLevelPreference() string {
+	return strings.TrimSpace(loadPreferences().ThinkingLevel)
+}
+
+// SaveThinkingLevelPreference persists the thinking level to preferences.yml.
+func SaveThinkingLevelPreference(level string) error {
+	return savePreferences(func(p *preferences) {
+		p.ThinkingLevel = level
+	})
+}

internal/ui/preferences_test.go

@@ -0,0 +1,180 @@
+package ui
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestSaveAndLoadThemePreference(t *testing.T) {
+	// Use a temp dir as XDG_CONFIG_HOME so we don't touch the real config.
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Initially no preference is saved.
+	if got := LoadThemePreference(); got != "" {
+		t.Fatalf("expected empty preference, got %q", got)
+	}
+
+	// Save a preference.
+	if err := SaveThemePreference("dracula"); err != nil {
+		t.Fatalf("SaveThemePreference: %v", err)
+	}
+
+	// Load it back.
+	if got := LoadThemePreference(); got != "dracula" {
+		t.Fatalf("expected %q, got %q", "dracula", got)
+	}
+
+	// Overwrite with a different theme.
+	if err := SaveThemePreference("nord"); err != nil {
+		t.Fatalf("SaveThemePreference: %v", err)
+	}
+	if got := LoadThemePreference(); got != "nord" {
+		t.Fatalf("expected %q, got %q", "nord", got)
+	}
+
+	// Verify the file exists and is valid YAML.
+	path := filepath.Join(tmp, "kit", "preferences.yml")
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("reading preferences file: %v", err)
+	}
+	if len(data) == 0 {
+		t.Fatal("preferences file is empty")
+	}
+}
+
+func TestLoadThemePreference_MissingFile(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// No file exists — should return empty string, not error.
+	if got := LoadThemePreference(); got != "" {
+		t.Fatalf("expected empty string for missing file, got %q", got)
+	}
+}
+
+func TestLoadThemePreference_InvalidYAML(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Write invalid YAML.
+	dir := filepath.Join(tmp, "kit")
+	_ = os.MkdirAll(dir, 0o755)
+	_ = os.WriteFile(filepath.Join(dir, "preferences.yml"), []byte(":::bad yaml"), 0o644)
+
+	// Should return empty string, not panic.
+	if got := LoadThemePreference(); got != "" {
+		t.Fatalf("expected empty string for invalid YAML, got %q", got)
+	}
+}
+
+func TestSaveThemePreference_PreservesOtherFields(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Pre-populate with extra content (simulating future fields).
+	dir := filepath.Join(tmp, "kit")
+	_ = os.MkdirAll(dir, 0o755)
+	_ = os.WriteFile(filepath.Join(dir, "preferences.yml"), []byte("theme: old\n"), 0o644)
+
+	// Overwrite theme.
+	if err := SaveThemePreference("catppuccin"); err != nil {
+		t.Fatalf("SaveThemePreference: %v", err)
+	}
+
+	if got := LoadThemePreference(); got != "catppuccin" {
+		t.Fatalf("expected %q, got %q", "catppuccin", got)
+	}
+}
+
+func TestSaveAndLoadModelPreference(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Initially empty.
+	if got := LoadModelPreference(); got != "" {
+		t.Fatalf("expected empty, got %q", got)
+	}
+
+	// Save a model.
+	if err := SaveModelPreference("anthropic/claude-sonnet-4-5-20250929"); err != nil {
+		t.Fatalf("SaveModelPreference: %v", err)
+	}
+	if got := LoadModelPreference(); got != "anthropic/claude-sonnet-4-5-20250929" {
+		t.Fatalf("expected %q, got %q", "anthropic/claude-sonnet-4-5-20250929", got)
+	}
+
+	// Overwrite.
+	if err := SaveModelPreference("openai/gpt-4o"); err != nil {
+		t.Fatalf("SaveModelPreference: %v", err)
+	}
+	if got := LoadModelPreference(); got != "openai/gpt-4o" {
+		t.Fatalf("expected %q, got %q", "openai/gpt-4o", got)
+	}
+}
+
+func TestSaveAndLoadThinkingLevelPreference(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Initially empty.
+	if got := LoadThinkingLevelPreference(); got != "" {
+		t.Fatalf("expected empty, got %q", got)
+	}
+
+	// Save a level.
+	if err := SaveThinkingLevelPreference("medium"); err != nil {
+		t.Fatalf("SaveThinkingLevelPreference: %v", err)
+	}
+	if got := LoadThinkingLevelPreference(); got != "medium" {
+		t.Fatalf("expected %q, got %q", "medium", got)
+	}
+
+	// Overwrite.
+	if err := SaveThinkingLevelPreference("high"); err != nil {
+		t.Fatalf("SaveThinkingLevelPreference: %v", err)
+	}
+	if got := LoadThinkingLevelPreference(); got != "high" {
+		t.Fatalf("expected %q, got %q", "high", got)
+	}
+}
+
+func TestPreferencesPreserveEachOther(t *testing.T) {
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", tmp)
+
+	// Save all three preferences.
+	if err := SaveThemePreference("dracula"); err != nil {
+		t.Fatal(err)
+	}
+	if err := SaveModelPreference("anthropic/claude-haiku-3-5-20241022"); err != nil {
+		t.Fatal(err)
+	}
+	if err := SaveThinkingLevelPreference("high"); err != nil {
+		t.Fatal(err)
+	}
+
+	// All three should be preserved.
+	if got := LoadThemePreference(); got != "dracula" {
+		t.Fatalf("theme: expected %q, got %q", "dracula", got)
+	}
+	if got := LoadModelPreference(); got != "anthropic/claude-haiku-3-5-20241022" {
+		t.Fatalf("model: expected %q, got %q", "anthropic/claude-haiku-3-5-20241022", got)
+	}
+	if got := LoadThinkingLevelPreference(); got != "high" {
+		t.Fatalf("thinking_level: expected %q, got %q", "high", got)
+	}
+
+	// Updating one should not affect the others.
+	if err := SaveModelPreference("openai/gpt-4o"); err != nil {
+		t.Fatal(err)
+	}
+	if got := LoadThemePreference(); got != "dracula" {
+		t.Fatalf("theme after model update: expected %q, got %q", "dracula", got)
+	}
+	if got := LoadThinkingLevelPreference(); got != "high" {
+		t.Fatalf("thinking_level after model update: expected %q, got %q", "high", got)
+	}
+}

internal/ui/slash_command_input.go

@@ -1,352 +0,0 @@
-package ui
-
-import (
-	"strings"
-
-	"charm.land/bubbles/v2/key"
-	"charm.land/bubbles/v2/textarea"
-	tea "charm.land/bubbletea/v2"
-	"charm.land/lipgloss/v2"
-)
-
-// SlashCommandInput provides an interactive text input field with intelligent
-// slash command autocomplete functionality. It displays a popup menu of matching
-// commands as the user types, supporting fuzzy matching and keyboard navigation.
-type SlashCommandInput struct {
-	textarea      textarea.Model
-	commands      []SlashCommand
-	showPopup     bool
-	filtered      []FuzzyMatch
-	selected      int
-	width         int
-	lastValue     string
-	popupHeight   int
-	title         string
-	quitting      bool
-	value         string
-	submitNext    bool // Flag to submit on next update
-	renderedLines int  // Track how many lines were rendered
-	hideHint      bool // Suppress the "enter submit · ctrl+j..." hint
-}
-
-// NewSlashCommandInput creates and initializes a new slash command input field with
-// the specified width and title. The input supports multi-line text entry, command
-// autocomplete, and is styled to match the application's theme.
-func NewSlashCommandInput(width int, title string) *SlashCommandInput {
-	ta := textarea.New()
-	ta.Placeholder = "Type your message..."
-	ta.ShowLineNumbers = false
-	ta.Prompt = ""
-	ta.CharLimit = 5000
-	ta.SetWidth(width - 8) // Account for container padding, border and internal padding
-	ta.SetHeight(3)        // Default to 3 lines like huh
-	ta.Focus()
-
-	// Override InsertNewline so only ctrl+j and shift+enter insert newlines.
-	// Enter always submits the input.
-	ta.KeyMap.InsertNewline = key.NewBinding(
-		key.WithKeys("ctrl+j", "shift+enter"),
-		key.WithHelp("ctrl+j", "insert newline"),
-	)
-
-	// Style the textarea using theme colors.
-	theme := GetTheme()
-	styles := ta.Styles()
-	styles.Focused.Base = lipgloss.NewStyle()
-	styles.Focused.Placeholder = lipgloss.NewStyle().Foreground(theme.VeryMuted)
-	styles.Focused.Text = lipgloss.NewStyle().Foreground(theme.Text)
-	styles.Focused.Prompt = lipgloss.NewStyle()
-	styles.Focused.CursorLine = lipgloss.NewStyle()
-	ta.SetStyles(styles)
-
-	return &SlashCommandInput{
-		textarea:    ta,
-		commands:    SlashCommands,
-		width:       width,
-		popupHeight: 7,
-		title:       title,
-	}
-}
-
-// Init implements the tea.Model interface, returning the initial command to start
-// the cursor blinking animation for the text input field.
-func (s *SlashCommandInput) Init() tea.Cmd {
-	return textarea.Blink
-}
-
-// Update implements the tea.Model interface, handling keyboard input for text entry,
-// command selection, and navigation. Manages the autocomplete popup display and
-// processes submission or cancellation actions.
-func (s *SlashCommandInput) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
-	var cmd tea.Cmd
-
-	// Check if we need to submit after updating the view
-	if s.submitNext {
-		s.value = s.textarea.Value()
-		s.quitting = true
-		return s, tea.Quit
-	}
-
-	switch msg := msg.(type) {
-	case tea.KeyPressMsg: // Check for quit keys first (when popup is not shown)
-		if !s.showPopup {
-			switch msg.String() {
-			case "ctrl+c", "esc":
-				s.quitting = true
-				return s, tea.Quit
-			case "ctrl+d", "enter": // Enter always submits
-				s.value = s.textarea.Value()
-				s.quitting = true
-				return s, tea.Quit
-			}
-		}
-
-		// Handle popup navigation
-		if s.showPopup {
-			switch {
-			case key.Matches(msg, key.NewBinding(key.WithKeys("up"), key.WithHelp("↑", "up"))):
-				if s.selected > 0 {
-					s.selected--
-				}
-				return s, nil
-			case key.Matches(msg, key.NewBinding(key.WithKeys("down"), key.WithHelp("↓", "down"))):
-				if s.selected < len(s.filtered)-1 {
-					s.selected++
-				}
-				return s, nil
-			case key.Matches(msg, key.NewBinding(key.WithKeys("tab"))):
-				if s.selected < len(s.filtered) {
-					// Complete with selected command
-					s.textarea.SetValue(s.filtered[s.selected].Command.Name)
-					s.showPopup = false
-					s.selected = 0
-					// Move cursor to end
-					s.textarea.CursorEnd()
-				}
-				return s, nil
-			case key.Matches(msg, key.NewBinding(key.WithKeys("enter"))):
-				if s.selected < len(s.filtered) {
-					// Populate the field with the selected command
-					s.textarea.SetValue(s.filtered[s.selected].Command.Name)
-					s.textarea.CursorEnd()
-					// Hide the popup
-					s.showPopup = false
-					s.selected = 0
-					// Set flag to submit on next update (after view refresh)
-					s.submitNext = true
-					// Force a refresh
-					return s, nil
-				}
-				return s, nil
-			case key.Matches(msg, key.NewBinding(key.WithKeys("esc"))):
-				s.showPopup = false
-				s.selected = 0
-				return s, nil
-			}
-		}
-
-		// Update textarea
-		s.textarea, cmd = s.textarea.Update(msg)
-
-		// Check if we should show/update popup
-		value := s.textarea.Value()
-		if value != s.lastValue {
-			s.lastValue = value
-			// Only show popup if we're on the first line and it starts with /
-			lines := strings.Split(value, "\n")
-			if len(lines) > 0 && strings.HasPrefix(lines[0], "/") && !strings.Contains(lines[0], " ") && len(lines) == 1 {
-				// Show and update popup
-				s.showPopup = true
-				s.filtered = FuzzyMatchCommands(lines[0], s.commands)
-				s.selected = 0
-			} else {
-				// Hide popup
-				s.showPopup = false
-			}
-		}
-		return s, cmd
-
-	default:
-		// Pass through other messages
-		s.textarea, cmd = s.textarea.Update(msg)
-		return s, cmd
-	}
-}
-
-// View implements the tea.Model interface, rendering the complete input field
-// including the title, text area, autocomplete popup (when active), and help text.
-// The view adapts based on whether single or multi-line input is detected.
-func (s *SlashCommandInput) View() tea.View {
-	containerStyle := lipgloss.NewStyle()
-
-	theme := GetTheme()
-
-	// PaddingLeft(3) aligns with message content: border(1) + paddingLeft(2).
-	titleStyle := lipgloss.NewStyle().
-		Foreground(theme.Text).
-		MarginBottom(1).
-		PaddingLeft(3)
-
-	// Input box with huh-like styling
-	inputBoxStyle := lipgloss.NewStyle().
-		Border(lipgloss.ThickBorder()).
-		BorderLeft(true).
-		BorderRight(false).
-		BorderTop(false).
-		BorderBottom(false).
-		BorderForeground(theme.Primary).
-		PaddingLeft(2).    // match message block paddingLeft
-		Width(s.width - 1) // full width minus left border
-
-	// Build the view
-	var view strings.Builder
-	view.WriteString(titleStyle.Render(s.title))
-	view.WriteString("\n")
-	view.WriteString(inputBoxStyle.Render(s.textarea.View()))
-	// Count rendered lines
-	s.renderedLines = 2 + s.textarea.Height() // title + newline + textarea height
-
-	// Add popup if visible
-	if s.showPopup && len(s.filtered) > 0 {
-		view.WriteString("\n")
-		view.WriteString(s.renderPopup())
-		// Add popup lines
-		visibleItems := min(len(s.filtered), s.popupHeight)
-		scrollIndicators := 0
-		if s.selected >= s.popupHeight {
-			scrollIndicators++ // top indicator
-		}
-		if len(s.filtered) > s.popupHeight {
-			scrollIndicators++ // bottom indicator
-		}
-		popupLines := visibleItems + scrollIndicators + 5 // items + scroll + border + padding + footer
-		s.renderedLines += 1 + popupLines                 // newline + popup
-	}
-
-	// Add help text at bottom (unless hidden by extension).
-	if !s.hideHint {
-		helpStyle := lipgloss.NewStyle().
-			Foreground(theme.VeryMuted).
-			MarginTop(1).
-			PaddingLeft(3)
-
-		helpText := "enter submit • ctrl+j / shift+enter new line"
-
-		view.WriteString("\n")
-		view.WriteString(helpStyle.Render(helpText))
-		s.renderedLines += 2 // newline + help text
-	}
-
-	// Apply container padding to entire view
-	return tea.NewView(containerStyle.Render(view.String()))
-}
-
-// renderPopup renders the autocomplete popup
-func (s *SlashCommandInput) renderPopup() string {
-	theme := GetTheme()
-
-	// Popup styling
-	popupStyle := lipgloss.NewStyle().
-		Border(lipgloss.RoundedBorder()).
-		BorderForeground(theme.MutedBorder).
-		Padding(1, 2).
-		Width(s.width - 4). // Account for container padding
-		MarginLeft(0)       // No extra margin needed due to container padding
-
-	var items []string
-
-	// Calculate visible window
-	visibleItems := min(len(s.filtered), s.popupHeight)
-	startIdx := 0
-
-	// Adjust window to keep selected item visible
-	if s.selected >= s.popupHeight {
-		startIdx = s.selected - s.popupHeight + 1
-	}
-
-	endIdx := min(startIdx+visibleItems, len(s.filtered))
-
-	for i := startIdx; i < endIdx; i++ {
-		match := s.filtered[i]
-		cmd := match.Command
-		// Create the selection indicator
-		var indicator string
-		if i == s.selected {
-			indicator = lipgloss.NewStyle().
-				Foreground(theme.Primary).
-				Render("> ")
-		} else {
-			indicator = "  "
-		}
-
-		// Format item
-		nameStyle := lipgloss.NewStyle().
-			Foreground(theme.Secondary).
-			Bold(true)
-
-		descStyle := lipgloss.NewStyle().
-			Foreground(theme.Muted)
-
-		// Highlight selected item
-		if i == s.selected {
-			nameStyle = nameStyle.Foreground(theme.Primary)
-			descStyle = descStyle.Foreground(theme.Text)
-		}
-
-		// Format with proper spacing
-		nameWidth := 15
-		name := nameStyle.Width(nameWidth - 2).Render(cmd.Name)
-
-		// Truncate description if needed
-		desc := cmd.Description
-		maxDescLen := s.width - nameWidth - 14 // Account for padding and indicator
-		if len(desc) > maxDescLen && maxDescLen > 3 {
-			desc = desc[:maxDescLen-3] + "..."
-		}
-
-		line := indicator + name + descStyle.Render(desc)
-		items = append(items, line)
-	}
-
-	// Add scroll indicators if needed
-	if startIdx > 0 {
-		scrollUpStyle := lipgloss.NewStyle().Foreground(theme.VeryMuted)
-		items = append([]string{scrollUpStyle.Render("  ↑ more above")}, items...)
-	}
-	if endIdx < len(s.filtered) {
-		scrollDownStyle := lipgloss.NewStyle().Foreground(theme.VeryMuted)
-		items = append(items, scrollDownStyle.Render("  ↓ more below"))
-	}
-	// Join items
-	content := strings.Join(items, "\n")
-
-	// Add footer hint
-	footerStyle := lipgloss.NewStyle().
-		Foreground(theme.VeryMuted).
-		Italic(true)
-	footer := footerStyle.Render("↑↓ navigate • tab complete • ↵ select • esc dismiss")
-
-	// Combine content and footer
-	popupContent := content + "\n\n" + footer
-
-	return popupStyle.Render(popupContent)
-}
-
-// Value returns the final text value entered by the user after submission.
-// This will be empty if the input was cancelled.
-func (s *SlashCommandInput) Value() string {
-	return s.value
-}
-
-// Cancelled returns true if the user cancelled the input operation (e.g., by
-// pressing ESC or Ctrl+C) without submitting any text.
-func (s *SlashCommandInput) Cancelled() bool {
-	return s.quitting && s.value == ""
-}
-
-// RenderedLines returns the total number of terminal lines used by the last
-// rendered view, including the title, input area, popup, and help text. This
-// is used for proper screen clearing when the input is dismissed.
-func (s *SlashCommandInput) RenderedLines() int {
-	return s.renderedLines
-}

internal/ui/stream.go

@@ -2,14 +2,27 @@ package ui
 
 import (
 	"fmt"
+	"regexp"
 	"strings"
 	"time"
 
 	tea "charm.land/bubbletea/v2"
 	"charm.land/lipgloss/v2"
+	"github.com/indaco/herald"
 	"github.com/mark3labs/kit/internal/app"
 )
 
+// thinkTagRegex matches  ...  tags that some models (Qwen, DeepSeek) wrap
+// reasoning content in. Used to strip these tags from streaming text content.
+// The (?s) flag makes . match newlines.
+var thinkTagRegex = regexp.MustCompile(`(?s)` + `` + `think` + `` + `(.*?)` + `` + `/think` + ``)
+
+// thinkTagOpen and thinkTagClose are the opening and closing think tag strings.
+const (
+	thinkTagOpen  = "<think>"
+	thinkTagClose = "</think>"
+)
+
 // knightRiderFrames generates a KITT-style scanning animation where a bright
 // light bounces back and forth across a row of dots with a trailing glow.
 // Colors are derived from the active theme. Used by StreamComponent (TUI
@@ -79,7 +92,12 @@ func streamSpinnerTickCmd(generation uint64) tea.Cmd {
 // streamFlushTickMsg fires when it's time to commit pending chunks to the
 // main content builders and trigger a re-render. This coalesces rapid
 // streaming chunks into fewer expensive markdown re-renders.
-type streamFlushTickMsg struct{}
+//
+// generation ties the tick to the pending flush session that created it so
+// stale ticks from a prior Reset() are discarded.
+type streamFlushTickMsg struct {
+	generation uint64
+}
 
 // streamFlushInterval is the coalescing window for stream chunks. Chunks
 // arriving within this window are batched into a single render pass.
@@ -89,9 +107,9 @@ const streamFlushInterval = 16 * time.Millisecond
 
 // streamFlushTickCmd returns a tea.Cmd that fires streamFlushTickMsg after
 // the coalescing interval.
-func streamFlushTickCmd() tea.Cmd {
+func streamFlushTickCmd(generation uint64) tea.Cmd {
 	return tea.Tick(streamFlushInterval, func(_ time.Time) tea.Msg {
-		return streamFlushTickMsg{}
+		return streamFlushTickMsg{generation: generation}
 	})
 }
 
@@ -149,9 +167,11 @@ type StreamComponent struct {
 	// spinnerFrame is the current frame index.
 	spinnerFrame int
 
-	// activeTools tracks the names of tools currently executing in parallel.
-	// When multiple tools run concurrently, all are displayed in the spinner.
-	activeTools []string
+	// activeTools maps ToolCallID -> display label for currently running tools.
+	activeTools map[string]string
+
+	// activeToolOrder preserves deterministic display order for active tools.
+	activeToolOrder []string
 
 	// streamContent holds committed streaming text (flushed from pending).
 	streamContent strings.Builder
@@ -172,6 +192,10 @@ type StreamComponent struct {
 	// the same coalescing window.
 	flushPending bool
 
+	// flushGeneration is incremented when stream state resets so stale flush
+	// ticks from a previous step can be discarded.
+	flushGeneration uint64
+
 	// renderCache holds the last rendered output string. Reused by View()
 	// between flush ticks to avoid redundant markdown re-parsing.
 	renderCache string
@@ -190,14 +214,12 @@ type StreamComponent struct {
 	// reasoningDuration holds the total reasoning time, frozen when streaming text begins.
 	reasoningDuration time.Duration
 
-	// messageRenderer renders assistant messages in standard mode.
-	messageRenderer *MessageRenderer
+	// inThinkTag tracks whether we're currently inside a  section
+	// from models that wrap reasoning in XML-like tags (Qwen, DeepSeek).
+	inThinkTag bool
 
-	// compactRenderer renders assistant messages in compact mode.
-	compactRenderer *CompactRenderer
-
-	// compactMode selects which renderer to use.
-	compactMode bool
+	// renderer renders streaming assistant text in either compact or standard mode.
+	renderer Renderer
 
 	// modelName is displayed in the streaming text header.
 	modelName string
@@ -211,6 +233,9 @@ type StreamComponent struct {
 	// height constrains the render output to at most this many lines.
 	// 0 means unconstrained.
 	height int
+
+	// ty provides typography functions for rendering text.
+	ty *herald.Typography
 }
 
 // NewStreamComponent creates a new StreamComponent ready to be embedded in AppModel.
@@ -218,13 +243,20 @@ func NewStreamComponent(compactMode bool, width int, modelName string) *StreamCo
 	if width == 0 {
 		width = 80
 	}
+
+	var renderer Renderer
+	if compactMode {
+		renderer = NewCompactRenderer(width, false)
+	} else {
+		renderer = newMessageRenderer(width, false)
+	}
+
 	return &StreamComponent{
-		spinnerFrames:   knightRiderFrames(),
-		compactMode:     compactMode,
-		modelName:       modelName,
-		messageRenderer: newMessageRenderer(width, false),
-		compactRenderer: NewCompactRenderer(width, false),
-		width:           width,
+		spinnerFrames: knightRiderFrames(),
+		modelName:     modelName,
+		renderer:      renderer,
+		width:         width,
+		ty:            createTypography(GetTheme()),
 	}
 }
 
@@ -251,11 +283,13 @@ func (s *StreamComponent) Reset() {
 	s.spinnerGeneration++ // invalidate any in-flight tick commands
 	s.spinnerFrame = 0
 	s.activeTools = nil
+	s.activeToolOrder = nil
 	s.streamContent.Reset()
 	s.reasoningContent.Reset()
 	s.pendingStream.Reset()
 	s.pendingReasoning.Reset()
 	s.flushPending = false
+	s.flushGeneration++
 	s.renderCache = ""
 	s.renderDirty = false
 	s.timestamp = time.Time{}
@@ -282,7 +316,8 @@ func (s *StreamComponent) GetRenderedContent() string {
 
 	text := s.streamContent.String()
 	if text != "" {
-		sections = append(sections, s.renderStreamingText(text))
+		rendered := s.renderStreamingText(text)
+		sections = append(sections, rendered)
 	}
 
 	if len(sections) == 0 {
@@ -295,7 +330,9 @@ func (s *StreamComponent) GetRenderedContent() string {
 // Called before reading content for scrollback output or on flush tick.
 func (s *StreamComponent) commitPending() {
 	if s.pendingStream.Len() > 0 {
-		s.streamContent.WriteString(s.pendingStream.String())
+		// Strip  ...  tags that some models wrap reasoning in
+		cleanedText := thinkTagRegex.ReplaceAllString(s.pendingStream.String(), "")
+		s.streamContent.WriteString(cleanedText)
 		s.pendingStream.Reset()
 		s.renderDirty = true
 	}
@@ -322,8 +359,9 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 
 	case tea.WindowSizeMsg:
 		s.width = msg.Width
-		s.messageRenderer.SetWidth(s.width)
-		s.compactRenderer.SetWidth(s.width)
+		if s.renderer != nil {
+			s.renderer.SetWidth(s.width)
+		}
 		// Invalidate render cache — width change affects wrapping/styling.
 		s.renderCache = ""
 		s.renderDirty = true
@@ -359,6 +397,9 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		}
 
 	case streamFlushTickMsg:
+		if msg.generation != s.flushGeneration {
+			break
+		}
 		s.flushPending = false
 		s.commitPending()
 
@@ -373,7 +414,7 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		s.pendingReasoning.WriteString(msg.Delta)
 		if !s.flushPending {
 			s.flushPending = true
-			return s, streamFlushTickCmd()
+			return s, streamFlushTickCmd(s.flushGeneration)
 		}
 
 	case app.StreamChunkEvent:
@@ -385,17 +426,66 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		if s.reasoningDuration == 0 && !s.reasoningStartTime.IsZero() {
 			s.reasoningDuration = time.Since(s.reasoningStartTime)
 		}
-		s.pendingStream.WriteString(msg.Content)
-		if !s.flushPending {
+
+		// Handle models that wrap reasoning in  tags (Qwen, DeepSeek)
+		// Filter out all content between  and  tags
+		content := msg.Content
+
+		// Check for opening tag
+		if strings.Contains(content, thinkTagOpen) {
+			parts := strings.SplitN(content, thinkTagOpen, 2)
+			// Content before the tag can be written
+			if !s.inThinkTag && parts[0] != "" {
+				s.pendingStream.WriteString(parts[0])
+			}
+			s.inThinkTag = true
+			// Content after the opening tag is reasoning - don't write it
+			if len(parts) > 1 && parts[1] != "" {
+				// Check if the same chunk contains the closing tag
+				if strings.Contains(parts[1], thinkTagClose) {
+					innerParts := strings.SplitN(parts[1], thinkTagClose, 2)
+					s.inThinkTag = false
+					// Content after closing tag can be written
+					if len(innerParts) > 1 && innerParts[1] != "" {
+						s.pendingStream.WriteString(innerParts[1])
+					}
+				}
+			}
+		} else if strings.Contains(content, thinkTagClose) {
+			// Closing tag found
+			parts := strings.SplitN(content, thinkTagClose, 2)
+			s.inThinkTag = false
+			// Content after closing tag can be written
+			if len(parts) > 1 && parts[1] != "" {
+				s.pendingStream.WriteString(parts[1])
+			}
+		} else if !s.inThinkTag {
+			// Normal content, not inside think tags
+			s.pendingStream.WriteString(content)
+		}
+		// else: inside think tag, don't write this content
+
+		if !s.flushPending && s.pendingStream.Len() > 0 {
 			s.flushPending = true
-			return s, streamFlushTickCmd()
+			return s, streamFlushTickCmd(s.flushGeneration)
 		}
 
 	case app.ToolExecutionEvent:
+		toolID := msg.ToolCallID
+		if toolID == "" {
+			// Defensive fallback for older/third-party emitters that may omit
+			// ToolCallID. Best-effort only: same-name+args concurrent calls can
+			// still collide without a stable ID.
+			toolID = fmt.Sprintf("%s|%s", msg.ToolName, msg.ToolArgs)
+		}
 		if msg.IsStarting {
-			// Add tool to active list for parallel execution display.
-			toolDisplay := formatToolExecutionMessage(msg.ToolName, msg.ToolArgs)
-			s.activeTools = append(s.activeTools, toolDisplay)
+			if s.activeTools == nil {
+				s.activeTools = make(map[string]string)
+			}
+			if _, exists := s.activeTools[toolID]; !exists {
+				s.activeToolOrder = append(s.activeToolOrder, toolID)
+			}
+			s.activeTools[toolID] = formatToolExecutionMessage(msg.ToolName)
 			s.spinnerFrame = 0
 			if !s.spinning {
 				s.phase = streamPhaseActive
@@ -404,9 +494,10 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 				return s, streamSpinnerTickCmd(s.spinnerGeneration)
 			}
 		} else {
-			// Tool finished — remove from active list but keep spinning if others remain.
-			toolDisplay := formatToolExecutionMessage(msg.ToolName, msg.ToolArgs)
-			s.activeTools = removeFromSlice(s.activeTools, toolDisplay)
+			if s.activeTools != nil {
+				delete(s.activeTools, toolID)
+			}
+			s.activeToolOrder = removeToolID(s.activeToolOrder, toolID)
 		}
 	}
 
@@ -415,7 +506,9 @@ func (s *StreamComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 
 // View implements tea.Model. Renders the current stream region content.
 func (s *StreamComponent) View() tea.View {
-	return tea.NewView(s.render())
+	fullContent := s.render()
+	visibleContent := s.viewContent(fullContent)
+	return tea.NewView(visibleContent)
 }
 
 // --------------------------------------------------------------------------
@@ -458,52 +551,51 @@ func (s *StreamComponent) render() string {
 
 	content := strings.Join(sections, "\n")
 
-	// Clamp to height if constrained: keep the last h lines so the most
-	// recent output is always visible.
-	if s.height > 0 && content != "" {
-		lines := strings.Split(content, "\n")
-		if len(lines) > s.height {
-			lines = lines[len(lines)-s.height:]
-			content = strings.Join(lines, "\n")
-		}
-	}
-
+	// Cache FULL content without height clamping.
+	// Height clamping is applied in View() for display only.
 	s.renderCache = content
 	s.renderDirty = false
 	return content
 }
 
-// renderReasoningBlock renders the reasoning/thinking content in a surface-tinted
-// box. When collapsed, shows the last 10 lines with a truncation hint. When
+// viewContent returns the visible portion of content based on height constraint.
+// This is called by View() to get the slice that fits in the terminal.
+func (s *StreamComponent) viewContent(fullContent string) string {
+	if s.height > 0 && fullContent != "" {
+		lines := strings.Split(fullContent, "\n")
+		if len(lines) > s.height {
+			// Keep only the last h lines so the most recent output is visible.
+			lines = lines[len(lines)-s.height:]
+			return strings.Join(lines, "\n")
+		}
+	}
+	return fullContent
+}
+
+// renderReasoningBlock renders the reasoning/thinking content using blockquote.
+// When collapsed, shows the last 10 lines with a truncation hint. When
 // expanded, shows all lines. Includes a "Thought for Xs" duration footer.
 func (s *StreamComponent) renderReasoningBlock(reasoning string) string {
-	theme := GetTheme()
-	maxWidth := max(s.width-4, 20)
-
 	lines := strings.Split(strings.TrimRight(reasoning, "\n"), "\n")
 
-	contentStyle := lipgloss.NewStyle().
-		Foreground(theme.Muted).
-		Italic(true)
-
 	var parts []string
 
 	// When collapsed and content exceeds 10 lines, show only the last 10
-	// with a truncation hint (matching iteratr's thinking block pattern).
+	// with a truncation hint.
 	const maxCollapsedLines = 10
 	if !s.thinkingVisible && len(lines) > maxCollapsedLines {
 		hidden := len(lines) - maxCollapsedLines
-		hintStyle := lipgloss.NewStyle().
-			Foreground(theme.VeryMuted).
-			Italic(true)
-		parts = append(parts, hintStyle.Render(fmt.Sprintf("... (%d lines hidden)", hidden)))
+		parts = append(parts, s.ty.Italic(fmt.Sprintf("... (%d lines hidden)", hidden)))
 		lines = lines[len(lines)-maxCollapsedLines:]
 	}
 
-	// Render reasoning text.
-	parts = append(parts, contentStyle.Width(maxWidth).Render(strings.Join(lines, "\n")))
+	// Main content using Italic with Muted color for visual distinction.
+	content := strings.TrimLeft(strings.Join(lines, "\n"), " \t\n")
+	theme := GetTheme()
+	mutedStyle := lipgloss.NewStyle().Foreground(theme.Muted)
+	parts = append(parts, mutedStyle.Render(s.ty.Italic(content)))
 
-	// Duration footer.
+	// Duration footer with VeryMuted label and Accent duration.
 	var duration time.Duration
 	if s.reasoningDuration > 0 {
 		duration = s.reasoningDuration
@@ -517,21 +609,21 @@ func (s *StreamComponent) renderReasoningBlock(reasoning string) string {
 		} else {
 			durationStr = fmt.Sprintf("%.1fs", duration.Seconds())
 		}
-		footer := lipgloss.NewStyle().Foreground(theme.VeryMuted).Render("Thought for ") +
-			lipgloss.NewStyle().Foreground(theme.Info).Render(durationStr)
-		parts = append(parts, footer)
+		label := lipgloss.NewStyle().Foreground(theme.VeryMuted).Render("Thought for ")
+		durationStyled := lipgloss.NewStyle().Foreground(theme.Accent).Render(durationStr)
+		parts = append(parts, label+durationStyled)
 	}
 
-	innerContent := strings.Join(parts, "\n")
-
-	// Wrap in box with surface background for visual distinction.
-	boxStyle := lipgloss.NewStyle().
-		Background(theme.MutedBorder). // Surface0 (#313244)
-		PaddingLeft(1).
-		Width(maxWidth + 2).
-		MarginBottom(1)
-
-	return boxStyle.Render(innerContent)
+	// Concatenate parts with newline between blockquote and footer
+	var result string
+	if len(parts) == 1 {
+		result = parts[0]
+	} else if len(parts) == 2 {
+		result = parts[0] + "\n" + parts[1]
+	} else {
+		result = strings.Join(parts, "\n")
+	}
+	return styleMarginBottom1.Render(result)
 }
 
 // SetThinkingVisible sets whether reasoning blocks are shown or collapsed.
@@ -557,7 +649,8 @@ func (s *StreamComponent) SpinnerView() string {
 		return ""
 	}
 	frame := s.spinnerFrames[s.spinnerFrame%len(s.spinnerFrames)]
-	if len(s.activeTools) == 0 {
+	tools := s.activeToolDisplays()
+	if len(tools) == 0 {
 		return "  " + frame
 	}
 	theme := GetTheme()
@@ -567,10 +660,10 @@ func (s *StreamComponent) SpinnerView() string {
 
 	// Format active tools list
 	var toolsMsg string
-	if len(s.activeTools) == 1 {
-		toolsMsg = s.activeTools[0]
+	if len(tools) == 1 {
+		toolsMsg = tools[0]
 	} else {
-		toolsMsg = "Running: " + strings.Join(s.activeTools, ", ")
+		toolsMsg = "Running: " + strings.Join(tools, ", ")
 	}
 	return "  " + frame + " " + msgStyle.Render(toolsMsg)
 }
@@ -582,30 +675,37 @@ func (s *StreamComponent) renderStreamingText(text string) string {
 	if ts.IsZero() {
 		ts = time.Now()
 	}
-
-	if s.compactMode {
-		msg := s.compactRenderer.RenderAssistantMessage(text, ts, s.modelName)
-		return msg.Content
+	if s.renderer == nil {
+		return text
 	}
-	msg := s.messageRenderer.RenderAssistantMessage(text, ts, s.modelName)
+	msg := s.renderer.RenderAssistantMessage(text, ts, s.modelName)
 	return msg.Content
 }
 
-// removeFromSlice removes the first occurrence of a string from a slice.
-func removeFromSlice(slice []string, s string) []string {
-	for i, v := range slice {
-		if v == s {
-			return append(slice[:i], slice[i+1:]...)
+func (s *StreamComponent) activeToolDisplays() []string {
+	if len(s.activeTools) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(s.activeToolOrder))
+	for _, id := range s.activeToolOrder {
+		if display, ok := s.activeTools[id]; ok {
+			out = append(out, display)
 		}
 	}
-	return slice
+	return out
+}
+
+// removeToolID removes the first occurrence of a tool ID from a slice.
+func removeToolID(ids []string, id string) []string {
+	for i, v := range ids {
+		if v == id {
+			return append(ids[:i], ids[i+1:]...)
+		}
+	}
+	return ids
 }
 
 // formatToolExecutionMessage creates a descriptive spinner message for tool execution.
-// For spawn_subagent, it shows simply as "Subagent" with optional task preview.
-func formatToolExecutionMessage(toolName, toolArgs string) string {
-	if toolName == "spawn_subagent" {
-		return "Subagent"
-	}
+func formatToolExecutionMessage(toolName string) string {
 	return toolName
 }

internal/ui/styles.go

@@ -1,19 +1,11 @@
 package ui
 
 import (
-	"fmt"
-	"image/color"
-
 	"charm.land/lipgloss/v2"
-	"github.com/charmbracelet/glamour"
-	"github.com/charmbracelet/glamour/ansi"
+	"github.com/indaco/herald"
+	heraldmd "github.com/indaco/herald-md"
 )
 
-// uintPtr returns a pointer to u. Used by ansi.StyleConfig fields.
-//
-//go:fix inline
-func uintPtr(u uint) *uint { return new(u) }
-
 // BaseStyle returns a new, empty lipgloss style that can be customized with
 // additional styling methods. This serves as the foundation for building more
 // complex styled components.
@@ -21,248 +13,83 @@ func BaseStyle() lipgloss.Style {
 	return lipgloss.NewStyle()
 }
 
-// colorHex converts a color.Color to a hex string suitable for ansi.StyleConfig.
-func colorHex(c color.Color) string {
-	r, g, b, _ := c.RGBA()
-	return fmt.Sprintf("#%02x%02x%02x", r>>8, g>>8, b>>8)
-}
+// markdownTypographyCache holds the last-created Typography instance for
+// herald-md rendering. It is cached to avoid re-initialization on every
+// streaming flush tick. The cache is invalidated by SetTheme when the
+// active theme changes.
+// This is only accessed from BubbleTea's single-threaded Update/View cycle,
+// so no mutex is required.
+var markdownTypographyCache *herald.Typography
 
-// colorHexPtr returns a pointer to the hex string of a color.Color.
-func colorHexPtr(c color.Color) *string {
-	s := colorHex(c)
-	return &s
-}
-
-// GetMarkdownRenderer creates and returns a configured glamour.TermRenderer for
-// rendering markdown content with syntax highlighting and proper formatting. The
-// renderer is customized with our theme colors and adapted to the specified width.
-func GetMarkdownRenderer(width int) *glamour.TermRenderer {
-	r, _ := glamour.NewTermRenderer(
-		glamour.WithStyles(generateMarkdownStyleConfig()),
-		glamour.WithWordWrap(width),
-	)
-	return r
-}
-
-// generateMarkdownStyleConfig creates an ansi.StyleConfig from the active theme.
-func generateMarkdownStyleConfig() ansi.StyleConfig {
-	md := GetTheme().Markdown
-	text := colorHexPtr(md.Text)
-	muted := colorHexPtr(md.Muted)
-	heading := colorHexPtr(md.Heading)
-	emph := colorHexPtr(md.Emph)
-	strong := colorHexPtr(md.Strong)
-	link := colorHexPtr(md.Link)
-	code := colorHexPtr(md.Code)
-	errClr := colorHexPtr(md.Error)
-	keyword := colorHexPtr(md.Keyword)
-	str := colorHexPtr(md.String)
-	number := colorHexPtr(md.Number)
-	comment := colorHexPtr(md.Comment)
-
-	return ansi.StyleConfig{
-		Document: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				BlockPrefix: "",
-				BlockSuffix: "",
-				Color:       text,
-			},
-			Margin: uintPtr(0),
-		},
-		BlockQuote: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Color:  muted,
-				Italic: new(true),
-				Prefix: "┃ ",
-			},
-			Indent: uintPtr(1),
-		},
-		List: ansi.StyleList{
-			LevelIndent: 0,
-			StyleBlock: ansi.StyleBlock{
-				StylePrimitive: ansi.StylePrimitive{
-					Color: text,
-				},
-			},
-		},
-		Heading: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				BlockSuffix: "\n",
-				Color:       heading,
-				Bold:        new(true),
-			},
-		},
-		H1: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "# ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		H2: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "## ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		H3: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "### ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		H4: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "#### ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		H5: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "##### ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		H6: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Prefix: "###### ",
-				Color:  heading,
-				Bold:   new(true),
-			},
-		},
-		Strikethrough: ansi.StylePrimitive{
-			CrossedOut: new(true),
-			Color:      muted,
-		},
-		Emph: ansi.StylePrimitive{
-			Color:  emph,
-			Italic: new(true),
-		},
-		Strong: ansi.StylePrimitive{
-			Bold:  new(true),
-			Color: strong,
-		},
-		HorizontalRule: ansi.StylePrimitive{
-			Color:  muted,
-			Format: "\n─────────────────────────────────────────\n",
-		},
-		Item: ansi.StylePrimitive{
-			BlockPrefix: "• ",
-			Color:       text,
-		},
-		Enumeration: ansi.StylePrimitive{
-			BlockPrefix: ". ",
-			Color:       text,
-		},
-		Task: ansi.StyleTask{
-			StylePrimitive: ansi.StylePrimitive{},
-			Ticked:         "[✓] ",
-			Unticked:       "[ ] ",
-		},
-		Link: ansi.StylePrimitive{
-			Color:     link,
-			Underline: new(true),
-		},
-		LinkText: ansi.StylePrimitive{
-			Color: link,
-			Bold:  new(true),
-		},
-		Image: ansi.StylePrimitive{
-			Color:     link,
-			Underline: new(true),
-			Format:    "🖼 {{.text}}",
-		},
-		ImageText: ansi.StylePrimitive{
-			Color:  link,
-			Format: "{{.text}}",
-		},
-		Code: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Color:  code,
-				Prefix: "",
-				Suffix: "",
-			},
-		},
-		CodeBlock: ansi.StyleCodeBlock{
-			StyleBlock: ansi.StyleBlock{
-				StylePrimitive: ansi.StylePrimitive{
-					Prefix: "",
-					Color:  code,
-				},
-				Margin: uintPtr(0),
-			},
-			Chroma: &ansi.Chroma{
-				Text:             ansi.StylePrimitive{Color: text},
-				Error:            ansi.StylePrimitive{Color: errClr},
-				Comment:          ansi.StylePrimitive{Color: comment},
-				CommentPreproc:   ansi.StylePrimitive{Color: keyword},
-				Keyword:          ansi.StylePrimitive{Color: keyword},
-				KeywordReserved:  ansi.StylePrimitive{Color: keyword},
-				KeywordNamespace: ansi.StylePrimitive{Color: keyword},
-				KeywordType:      ansi.StylePrimitive{Color: keyword},
-				Operator:         ansi.StylePrimitive{Color: text},
-				Punctuation:      ansi.StylePrimitive{Color: text},
-				Name:             ansi.StylePrimitive{Color: text},
-				NameBuiltin:      ansi.StylePrimitive{Color: text},
-				NameTag:          ansi.StylePrimitive{Color: keyword},
-				NameAttribute:    ansi.StylePrimitive{Color: text},
-				NameClass:        ansi.StylePrimitive{Color: keyword},
-				NameConstant:     ansi.StylePrimitive{Color: text},
-				NameDecorator:    ansi.StylePrimitive{Color: text},
-				NameFunction:     ansi.StylePrimitive{Color: text},
-				LiteralNumber:    ansi.StylePrimitive{Color: number},
-				LiteralString:    ansi.StylePrimitive{Color: str},
-				LiteralStringEscape: ansi.StylePrimitive{
-					Color: keyword,
-				},
-				GenericDeleted: ansi.StylePrimitive{Color: errClr},
-				GenericEmph: ansi.StylePrimitive{
-					Color:  emph,
-					Italic: new(true),
-				},
-				GenericInserted: ansi.StylePrimitive{Color: str},
-				GenericStrong: ansi.StylePrimitive{
-					Color: strong,
-					Bold:  new(true),
-				},
-				GenericSubheading: ansi.StylePrimitive{
-					Color: heading,
-				},
-			},
-		},
-		Table: ansi.StyleTable{
-			StyleBlock: ansi.StyleBlock{
-				StylePrimitive: ansi.StylePrimitive{
-					BlockPrefix: "\n",
-					BlockSuffix: "\n",
-				},
-			},
-			CenterSeparator: new("┼"),
-			ColumnSeparator: new("│"),
-			RowSeparator:    new("─"),
-		},
-		DefinitionDescription: ansi.StylePrimitive{
-			BlockPrefix: "\n ❯ ",
-			Color:       link,
-		},
-		Text: ansi.StylePrimitive{
-			Color: text,
-		},
-		Paragraph: ansi.StyleBlock{
-			StylePrimitive: ansi.StylePrimitive{
-				Color: text,
-			},
-		},
+// GetMarkdownTypography returns a herald.Typography configured with our
+// active theme colors. The typography is cached and only rebuilt when
+// the theme changes via SetTheme.
+func GetMarkdownTypography() *herald.Typography {
+	if markdownTypographyCache != nil {
+		return markdownTypographyCache
 	}
+
+	theme := GetTheme()
+	md := theme.Markdown
+
+	// Build herald theme from our theme colors
+	hty := herald.Theme{
+		// Headings - use heading color
+		H1: lipgloss.NewStyle().Foreground(md.Heading).Bold(true),
+		H2: lipgloss.NewStyle().Foreground(md.Heading).Bold(true),
+		H3: lipgloss.NewStyle().Foreground(md.Heading).Bold(true),
+		H4: lipgloss.NewStyle().Foreground(md.Heading).Bold(true),
+		H5: lipgloss.NewStyle().Foreground(md.Heading).Bold(true),
+		H6: lipgloss.NewStyle().Foreground(md.Muted).Bold(true),
+
+		// Text blocks
+		Paragraph:  lipgloss.NewStyle().Foreground(md.Text),
+		Blockquote: lipgloss.NewStyle().Foreground(md.Muted).Italic(true),
+		CodeInline: lipgloss.NewStyle().Foreground(md.Code),
+		CodeBlock:  lipgloss.NewStyle().Foreground(md.Code),
+		HR:         lipgloss.NewStyle().Foreground(md.Muted),
+
+		// Lists
+		ListBullet: lipgloss.NewStyle().Foreground(md.Text),
+		ListItem:   lipgloss.NewStyle().Foreground(md.Text),
+
+		// Inline styles
+		Bold:          lipgloss.NewStyle().Foreground(md.Strong).Bold(true),
+		Italic:        lipgloss.NewStyle().Foreground(md.Emph).Italic(true),
+		Strikethrough: lipgloss.NewStyle().Foreground(md.Muted).Strikethrough(true),
+		Link:          lipgloss.NewStyle().Foreground(md.Link).Underline(true),
+
+		// Definition lists
+		DT: lipgloss.NewStyle().Foreground(md.Text).Bold(true),
+		DD: lipgloss.NewStyle().Foreground(md.Muted),
+
+		// Key-value
+		KVKey:   lipgloss.NewStyle().Foreground(md.Text).Bold(true),
+		KVValue: lipgloss.NewStyle().Foreground(md.Text),
+
+		// Badges/Tags - use semantic colors
+		Badge:        lipgloss.NewStyle().Foreground(md.Text).Bold(true),
+		SuccessBadge: lipgloss.NewStyle().Foreground(theme.Success).Bold(true),
+		WarningBadge: lipgloss.NewStyle().Foreground(theme.Warning).Bold(true),
+		ErrorBadge:   lipgloss.NewStyle().Foreground(theme.Error).Bold(true),
+		InfoBadge:    lipgloss.NewStyle().Foreground(theme.Info).Bold(true),
+
+		// Heading decorations
+		H1UnderlineChar: "═",
+		H2UnderlineChar: "─",
+		H3UnderlineChar: "·",
+	}
+
+	ty := herald.New(herald.WithTheme(hty))
+	markdownTypographyCache = ty
+	return ty
 }
 
-// toMarkdown renders markdown content using glamour.
+// toMarkdown renders markdown content using herald-md.
+// The width parameter is currently unused as herald handles wrapping
+// based on terminal width internally.
 func toMarkdown(content string, width int) string {
-	r := GetMarkdownRenderer(width)
-	rendered, _ := r.Render(content)
+	ty := GetMarkdownTypography()
+	rendered := heraldmd.Render(ty, []byte(content))
 	return rendered
 }

internal/ui/themes.go

@@ -7,11 +7,86 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"strconv"
 	"strings"
 
 	"gopkg.in/yaml.v3"
 )
 
+// ---------------------------------------------------------------------------
+// Color derivation helpers
+// ---------------------------------------------------------------------------
+
+// parseHexColor parses a "#RRGGBB" hex string into r, g, b components (0-255).
+func parseHexColor(hex string) (r, g, b int) {
+	hex = strings.TrimPrefix(hex, "#")
+	if len(hex) == 6 {
+		if v, err := strconv.ParseUint(hex[0:2], 16, 8); err == nil {
+			r = int(v)
+		}
+		if v, err := strconv.ParseUint(hex[2:4], 16, 8); err == nil {
+			g = int(v)
+		}
+		if v, err := strconv.ParseUint(hex[4:6], 16, 8); err == nil {
+			b = int(v)
+		}
+	}
+	return
+}
+
+// blendHex linearly interpolates between two hex colors by amount (0.0–1.0).
+func blendHex(base, tint string, amount float64) string {
+	br, bg, bb := parseHexColor(base)
+	tr, tg, tb := parseHexColor(tint)
+	clamp := func(v int) int {
+		if v < 0 {
+			return 0
+		}
+		if v > 255 {
+			return 255
+		}
+		return v
+	}
+	r := clamp(int(float64(br)*(1-amount) + float64(tr)*amount))
+	g := clamp(int(float64(bg)*(1-amount) + float64(tg)*amount))
+	b := clamp(int(float64(bb)*(1-amount) + float64(tb)*amount))
+	return fmt.Sprintf("#%02x%02x%02x", r, g, b)
+}
+
+// deriveDiffBg computes diff / code background colors from the theme's
+// background, success, and error hex pairs. Returns an adaptive color for each
+// diff element. The tint amounts are tuned for subtle differentiation.
+func deriveDiffBg(bgPair, successPair, errorPair [2]string) (diffInsert, diffDelete, diffEqual, diffMissing, codeBg, gutterBg, writeBg color.Color) {
+	derive := func(idx int) (color.Color, color.Color, color.Color, color.Color) {
+		bg := bgPair[idx]
+		// Contrast target: darken for light mode (idx 0), lighten for dark (idx 1).
+		contrast := "#000000"
+		if idx == 1 {
+			contrast = "#ffffff"
+		}
+		ins := blendHex(bg, successPair[idx], 0.13)
+		del := blendHex(bg, errorPair[idx], 0.13)
+		eq := blendHex(bg, contrast, 0.05)
+		miss := blendHex(bg, contrast, 0.03)
+		return AdaptiveColor(ins, ins), AdaptiveColor(del, del), AdaptiveColor(eq, eq), AdaptiveColor(miss, miss)
+	}
+
+	// Pick the correct index based on detected background.
+	idx := 0
+	if isDarkBg {
+		idx = 1
+	}
+	insL, delL, eqL, missL := derive(idx)
+	diffInsert = insL
+	diffDelete = delL
+	diffEqual = eqL
+	diffMissing = missL
+	codeBg = eqL
+	gutterBg = missL
+	writeBg = insL
+	return
+}
+
 // ThemeEntry is a named, loadable theme — either built-in or discovered from disk.
 type ThemeEntry struct {
 	Name   string // Display name (filename stem or preset name)
@@ -54,13 +129,20 @@ type presetColors struct {
 }
 
 func makeTheme(p presetColors) Theme {
-	ac := func(pair [2]string) color.Color { return AdaptiveColor(pair[0], pair[1]) }
 	def := DefaultTheme()
-	acOr := func(pair [2]string, fb color.Color) color.Color {
+	ac := func(pair [2]string) color.Color {
+		c := AdaptiveColor(pair[0], pair[1])
 		if pair[0] == "" && pair[1] == "" {
+			return nil
+		}
+		return c
+	}
+	acOr := func(pair [2]string, fb color.Color) color.Color {
+		c := ac(pair)
+		if c == nil {
 			return fb
 		}
-		return ac(pair)
+		return c
 	}
 	t := Theme{
 		Primary:     ac(p.primary),
@@ -80,14 +162,9 @@ func makeTheme(p presetColors) Theme {
 		Accent:      acOr(p.accent, ac(p.primary)),
 		Highlight:   acOr(p.highlight, def.Highlight),
 	}
-	// Derive diff/code backgrounds from the base background.
-	t.DiffInsertBg = def.DiffInsertBg
-	t.DiffDeleteBg = def.DiffDeleteBg
-	t.DiffEqualBg = def.DiffEqualBg
-	t.DiffMissingBg = def.DiffMissingBg
-	t.CodeBg = def.CodeBg
-	t.GutterBg = def.GutterBg
-	t.WriteBg = def.WriteBg
+	// Derive diff/code backgrounds from the theme's own palette.
+	t.DiffInsertBg, t.DiffDeleteBg, t.DiffEqualBg, t.DiffMissingBg,
+		t.CodeBg, t.GutterBg, t.WriteBg = deriveDiffBg(p.background, p.success, p.error_)
 	// Markdown colors.
 	t.Markdown = MarkdownThemeColors{
 		Text:    t.Text,
@@ -439,7 +516,23 @@ func LoadThemeByName(name string) (Theme, error) {
 }
 
 // ApplyTheme loads a theme by name and sets it as the active global theme.
+// The selection is persisted to ~/.config/kit/preferences.yml so it survives
+// across sessions. Persistence errors are silently ignored — the theme is
+// still applied in-memory even if the write fails.
 func ApplyTheme(name string) error {
+	t, err := LoadThemeByName(name)
+	if err != nil {
+		return err
+	}
+	SetTheme(t)
+	_ = SaveThemePreference(name)
+	return nil
+}
+
+// ApplyThemeWithoutSave loads a theme by name and sets it as the active global
+// theme without persisting the choice. Used at startup to restore a previously
+// saved preference without redundantly re-writing it.
+func ApplyThemeWithoutSave(name string) error {
 	t, err := LoadThemeByName(name)
 	if err != nil {
 		return err
@@ -593,6 +686,17 @@ func loadThemeFile(path string) (Theme, error) {
 
 func fileConfigToTheme(cfg themeFileConfig) Theme {
 	def := DefaultTheme()
+
+	// Resolve the base background/success/error hex pairs for diff derivation.
+	// We need the raw hex strings to feed deriveDiffBg.
+	bgPair := resolveHexPair(cfg.Background, [2]string{"#F0F0F0", "#0D0D0D"})
+	successPair := resolveHexPair(cfg.Success, [2]string{"#998800", "#CCAA00"})
+	errorPair := resolveHexPair(cfg.Error, [2]string{"#CC0000", "#FF3333"})
+
+	// Derive diff backgrounds from the theme's own palette.
+	derivedInsert, derivedDelete, derivedEqual, derivedMissing,
+		derivedCodeBg, derivedGutterBg, derivedWriteBg := deriveDiffBg(bgPair, successPair, errorPair)
+
 	return Theme{
 		Primary:     cfg.Primary.resolve(def.Primary),
 		Secondary:   cfg.Secondary.resolve(def.Secondary),
@@ -611,13 +715,13 @@ func fileConfigToTheme(cfg themeFileConfig) Theme {
 		Accent:      cfg.Accent.resolve(def.Accent),
 		Highlight:   cfg.Highlight.resolve(def.Highlight),
 
-		DiffInsertBg:  cfg.DiffInsertBg.resolve(def.DiffInsertBg),
-		DiffDeleteBg:  cfg.DiffDeleteBg.resolve(def.DiffDeleteBg),
-		DiffEqualBg:   cfg.DiffEqualBg.resolve(def.DiffEqualBg),
-		DiffMissingBg: cfg.DiffMissingBg.resolve(def.DiffMissingBg),
-		CodeBg:        cfg.CodeBg.resolve(def.CodeBg),
-		GutterBg:      cfg.GutterBg.resolve(def.GutterBg),
-		WriteBg:       cfg.WriteBg.resolve(def.WriteBg),
+		DiffInsertBg:  cfg.DiffInsertBg.resolve(derivedInsert),
+		DiffDeleteBg:  cfg.DiffDeleteBg.resolve(derivedDelete),
+		DiffEqualBg:   cfg.DiffEqualBg.resolve(derivedEqual),
+		DiffMissingBg: cfg.DiffMissingBg.resolve(derivedMissing),
+		CodeBg:        cfg.CodeBg.resolve(derivedCodeBg),
+		GutterBg:      cfg.GutterBg.resolve(derivedGutterBg),
+		WriteBg:       cfg.WriteBg.resolve(derivedWriteBg),
 
 		Markdown: MarkdownThemeColors{
 			Text:    cfg.Markdown.Text.resolve(def.Markdown.Text),
@@ -635,3 +739,17 @@ func fileConfigToTheme(cfg themeFileConfig) Theme {
 		},
 	}
 }
+
+// resolveHexPair returns the hex pair from an adaptiveColorPair, falling back
+// to defaults when the pair is empty.
+func resolveHexPair(a adaptiveColorPair, fallback [2]string) [2]string {
+	light := a.Light
+	if light == "" {
+		light = fallback[0]
+	}
+	dark := a.Dark
+	if dark == "" {
+		dark = fallback[1]
+	}
+	return [2]string{light, dark}
+}

internal/ui/themes_test.go

@@ -0,0 +1,85 @@
+package ui
+
+import (
+	"testing"
+)
+
+func TestParseHexColor(t *testing.T) {
+	tests := []struct {
+		hex     string
+		r, g, b int
+	}{
+		{"#000000", 0, 0, 0},
+		{"#ffffff", 255, 255, 255},
+		{"#1e1e2e", 0x1e, 0x1e, 0x2e},
+		{"#a6e3a1", 0xa6, 0xe3, 0xa1},
+		{"#f38ba8", 0xf3, 0x8b, 0xa8},
+	}
+	for _, tt := range tests {
+		r, g, b := parseHexColor(tt.hex)
+		if r != tt.r || g != tt.g || b != tt.b {
+			t.Errorf("parseHexColor(%q) = (%d,%d,%d), want (%d,%d,%d)",
+				tt.hex, r, g, b, tt.r, tt.g, tt.b)
+		}
+	}
+}
+
+func TestBlendHex(t *testing.T) {
+	// Blending with 0 amount should return the base color.
+	got := blendHex("#1e1e2e", "#a6e3a1", 0.0)
+	if got != "#1e1e2e" {
+		t.Errorf("blendHex with 0.0 = %q, want #1e1e2e", got)
+	}
+
+	// Blending with 1.0 amount should return the tint color.
+	got = blendHex("#1e1e2e", "#a6e3a1", 1.0)
+	if got != "#a6e3a1" {
+		t.Errorf("blendHex with 1.0 = %q, want #a6e3a1", got)
+	}
+
+	// Blending black and white at 0.5 should give mid gray.
+	got = blendHex("#000000", "#ffffff", 0.5)
+	// 127 = int(0 + 255*0.5) — truncated, so #7f7f7f
+	if got != "#7f7f7f" {
+		t.Errorf("blendHex black/white at 0.5 = %q, want #7f7f7f", got)
+	}
+}
+
+func TestDeriveDiffBgProducesDifferentColorsPerTheme(t *testing.T) {
+	// Catppuccin palette
+	catBg := [2]string{"#eff1f5", "#1e1e2e"}
+	catSuccess := [2]string{"#40a02b", "#a6e3a1"}
+	catError := [2]string{"#d20f39", "#f38ba8"}
+
+	// KITT palette
+	kittBg := [2]string{"#F0F0F0", "#0D0D0D"}
+	kittSuccess := [2]string{"#998800", "#CCAA00"}
+	kittError := [2]string{"#CC0000", "#FF3333"}
+
+	catInsert, catDelete, _, _, _, _, _ := deriveDiffBg(catBg, catSuccess, catError)
+	kittInsert, kittDelete, _, _, _, _, _ := deriveDiffBg(kittBg, kittSuccess, kittError)
+
+	if catInsert == kittInsert {
+		t.Error("catppuccin DiffInsertBg should differ from kitt DiffInsertBg")
+	}
+	if catDelete == kittDelete {
+		t.Error("catppuccin DiffDeleteBg should differ from kitt DiffDeleteBg")
+	}
+}
+
+func TestMakeThemeDerivesUniqueDiffColors(t *testing.T) {
+	themes := builtinThemes()
+	kitt := themes["kitt"]
+	cat := themes["catppuccin"]
+
+	// The catppuccin diff backgrounds should NOT equal the kitt defaults.
+	if cat.DiffInsertBg == kitt.DiffInsertBg {
+		t.Error("catppuccin DiffInsertBg should differ from kitt default")
+	}
+	if cat.DiffDeleteBg == kitt.DiffDeleteBg {
+		t.Error("catppuccin DiffDeleteBg should differ from kitt default")
+	}
+	if cat.DiffEqualBg == kitt.DiffEqualBg {
+		t.Error("catppuccin DiffEqualBg should differ from kitt default")
+	}
+}

internal/ui/tool_renderers.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"path/filepath"
 	"regexp"
 	"strconv"
 	"strings"
@@ -14,6 +15,8 @@ import (
 	"github.com/alecthomas/chroma/v2/lexers"
 	"github.com/alecthomas/chroma/v2/styles"
 	udiff "github.com/aymanbagabas/go-udiff"
+	xansi "github.com/charmbracelet/x/ansi"
+	"github.com/indaco/herald"
 )
 
 // Maximum visible lines per tool type before truncation.
@@ -22,8 +25,17 @@ const (
 	maxCodeLines  = 20 // lines for Read / code blocks
 	maxWriteLines = 10 // lines for Write blocks
 	maxBashLines  = 20 // lines for Bash output (matches Read)
+	maxLsLines    = 20 // lines for Ls directory listings
 )
 
+// isShellTool reports if the tool name matches a shell-like tool (bash, grep, find, or
+// tools with "shell"/"command" in the name). Used by both renderToolBody and
+// renderToolBodyCompact to avoid code duplication.
+func isShellTool(toolName string) bool {
+	return toolName == "bash" || toolName == "grep" || toolName == "find" ||
+		strings.Contains(toolName, "shell") || strings.Contains(toolName, "command")
+}
+
 // renderToolBody dispatches to tool-specific body renderers based on tool name.
 // Returns the styled body string, or empty string to fall back to default rendering.
 func renderToolBody(toolName, toolArgs, toolResult string, width int) string {
@@ -44,12 +56,11 @@ func renderToolBody(toolName, toolArgs, toolResult string, width int) string {
 		if body := renderWriteBody(toolArgs, toolResult, width); body != "" {
 			return body
 		}
-	case toolName == "bash" || toolName == "run_shell_cmd" ||
-		strings.Contains(toolName, "shell") || strings.Contains(toolName, "command"):
+	case isShellTool(toolName):
 		if body := renderBashBody(toolResult, width); body != "" {
 			return body
 		}
-	case toolName == "spawn_subagent":
+	case toolName == "subagent":
 		if body := renderSubagentBody(toolResult, width); body != "" {
 			return body
 		}
@@ -62,21 +73,44 @@ func renderToolBody(toolName, toolArgs, toolResult string, width int) string {
 // ---------------------------------------------------------------------------
 
 // renderEditBody renders a side-by-side diff from old_text/new_text in toolArgs.
+// Supports both single-edit mode and multi-edit mode (edits array).
 func renderEditBody(toolArgs, toolResult string, width int) string {
 	var args map[string]any
 	if err := json.Unmarshal([]byte(toolArgs), &args); err != nil {
 		return ""
 	}
 
+	// Try to extract the starting line number from the unified diff in the result
+	startLine := extractDiffStartLine(toolResult)
+
+	// Check for multi-edit mode (edits array)
+	if editsArr, ok := args["edits"].([]any); ok && len(editsArr) > 0 {
+		var results []string
+		for _, edit := range editsArr {
+			if e, ok := edit.(map[string]any); ok {
+				oldText, _ := e["old_text"].(string)
+				newText, _ := e["new_text"].(string)
+				if oldText != "" || newText != "" {
+					diff := renderDiffBlock(oldText, newText, startLine, width)
+					if diff != "" {
+						results = append(results, diff)
+					}
+				}
+			}
+		}
+		if len(results) > 0 {
+			return strings.Join(results, "\n")
+		}
+		return ""
+	}
+
+	// Single-edit mode (legacy)
 	oldText, _ := args["old_text"].(string)
 	newText, _ := args["new_text"].(string)
 	if oldText == "" && newText == "" {
 		return ""
 	}
 
-	// Try to extract the starting line number from the unified diff in the result
-	startLine := extractDiffStartLine(toolResult)
-
 	return renderDiffBlock(oldText, newText, startLine, width)
 }
 
@@ -219,7 +253,7 @@ func renderDiffBlock(before, after string, startLine int, width int) string {
 	gutterWidth := max(len(fmt.Sprintf("%d", maxLineNum)), 3)
 	contentWidth := max(panelWidth-gutterWidth-4, 10) // gutter + " - " or " + "
 
-	theme := getTheme()
+	theme := GetTheme()
 
 	// Styles for each cell type
 	gutterInsert := lipgloss.NewStyle().Foreground(theme.Muted).Background(theme.DiffInsertBg)
@@ -228,7 +262,7 @@ func renderDiffBlock(before, after string, startLine int, width int) string {
 	gutterMissing := lipgloss.NewStyle().Background(theme.DiffMissingBg)
 
 	contentInsert := lipgloss.NewStyle().Background(theme.DiffInsertBg)
-	contentDelete := lipgloss.NewStyle().Background(theme.DiffDeleteBg).Strikethrough(true)
+	contentDelete := lipgloss.NewStyle().Background(theme.DiffDeleteBg)
 	contentEqual := lipgloss.NewStyle().Foreground(theme.Muted).Background(theme.DiffEqualBg)
 	contentMissing := lipgloss.NewStyle().Background(theme.DiffMissingBg)
 
@@ -314,18 +348,34 @@ func renderLsBody(toolResult string, width int) string {
 
 	lines := strings.Split(content, "\n")
 
+	// Truncate to maxLsLines for display
+	var hiddenCount int
+	if len(lines) > maxLsLines {
+		hiddenCount = len(lines) - maxLsLines
+		lines = lines[:maxLsLines]
+	}
+
 	const indent = "  "
 	codeWidth := max(width-len(indent), 20)
 
-	theme := getTheme()
+	theme := GetTheme()
 	codeStyle := lipgloss.NewStyle().Background(theme.CodeBg).PaddingLeft(1)
 
 	var result []string
 	for _, line := range lines {
+		// Truncate before styling to prevent wrapping.
+		line = truncateLine(line, codeWidth-1) // account for PaddingLeft(1)
 		styled := codeStyle.Width(codeWidth).Render(line)
 		result = append(result, indent+styled)
 	}
 
+	if hiddenCount > 0 {
+		hint := fmt.Sprintf("...(%d more entries)", hiddenCount)
+		hintContent := codeStyle.Width(codeWidth).
+			Foreground(theme.Muted).Italic(true).Render(hint)
+		result = append(result, indent+hintContent)
+	}
+
 	return strings.Join(result, "\n")
 }
 
@@ -333,133 +383,106 @@ func renderLsBody(toolResult string, width int) string {
 // Read tool — code block with line numbers + syntax highlighting
 // ---------------------------------------------------------------------------
 
-// renderReadBody renders Read tool output with styled line numbers and optional
-// syntax highlighting based on file extension.
+// renderReadBody renders Read tool output using herald.CodeBlock with line numbers
+// and syntax highlighting. Uses WithCodeLineNumberOffset to show correct offsets
+// based on the Read tool's offset parameter.
 func renderReadBody(toolArgs, toolResult string, width int) string {
 	if strings.TrimSpace(toolResult) == "" {
 		return ""
 	}
 
-	// Extract file path for syntax highlighting
+	// Extract file path and offset from tool args
 	var fileName string
+	var offset = 1
 	var args map[string]any
 	if err := json.Unmarshal([]byte(toolArgs), &args); err == nil {
 		if p, ok := args["path"].(string); ok {
 			fileName = p
 		}
+		if o, ok := args["offset"].(float64); ok {
+			offset = int(o)
+		}
 	}
 
-	return renderCodeBlock(toolResult, fileName, width)
-}
-
-// codeLine holds a parsed line with optional line number.
-type codeLine struct {
-	lineNum string
-	code    string
-}
-
-// renderCodeBlock renders content with a styled gutter (line numbers) and
-// optional syntax highlighting.
-func renderCodeBlock(content, fileName string, width int) string {
-	rawLines := strings.Split(content, "\n")
-
-	// Parse lines: detect "N: content" format from Read tool
-	var parsed []codeLine
-	maxNumWidth := 0
-	var codeOnly []string
+	// Parse lines to extract pure code content (removing "N: " prefixes)
+	rawLines := strings.Split(toolResult, "\n")
+	var codeLines []string
+	var footerLines []string
+	var codeHiddenCount int
 
 	for _, line := range rawLines {
+		// Detect "N: content" format from Read tool
 		if idx := strings.Index(line, ": "); idx > 0 && idx <= 7 {
 			numPart := line[:idx]
 			if _, err := strconv.Atoi(strings.TrimSpace(numPart)); err == nil {
-				parsed = append(parsed, codeLine{lineNum: numPart, code: line[idx+2:]})
-				if len(numPart) > maxNumWidth {
-					maxNumWidth = len(numPart)
-				}
-				codeOnly = append(codeOnly, line[idx+2:])
+				codeLines = append(codeLines, line[idx+2:])
 				continue
 			}
 		}
-		// No line number — treat as metadata/footer
-		parsed = append(parsed, codeLine{code: line})
-		codeOnly = append(codeOnly, line)
+		// No line number — treat as footer/metadata (e.g., truncation notice)
+		footerLines = append(footerLines, line)
 	}
 
-	if len(parsed) == 0 {
-		return ""
+	// Apply maxCodeLines truncation
+	totalCodeLines := len(codeLines)
+	if totalCodeLines > maxCodeLines {
+		codeHiddenCount = totalCodeLines - maxCodeLines
+		codeLines = codeLines[:maxCodeLines]
 	}
 
-	// Truncate to maxCodeLines visible lines (preserve footer/metadata lines)
-	var codeHiddenCount int
-	totalParsed := len(parsed)
-	if totalParsed > maxCodeLines {
-		// Check if last line is a footer (no line number) — keep it
-		var footerLines []codeLine
-		for totalParsed > 0 && parsed[totalParsed-1].lineNum == "" {
-			footerLines = append([]codeLine{parsed[totalParsed-1]}, footerLines...)
-			totalParsed--
-		}
-		if totalParsed > maxCodeLines {
-			codeHiddenCount = totalParsed - maxCodeLines
-			parsed = append(parsed[:maxCodeLines], footerLines...)
-			codeOnly = codeOnly[:maxCodeLines]
-			for _, fl := range footerLines {
-				codeOnly = append(codeOnly, fl.code)
-			}
-		} else {
-			// Restore — footer trimming was enough
-			parsed = parsed[:totalParsed]
-			parsed = append(parsed, footerLines...)
+	// Build language hint from file extension
+	lang := ""
+	if fileName != "" {
+		// Extract extension without the dot
+		if ext := strings.TrimPrefix(filepath.Ext(fileName), "."); ext != "" {
+			lang = ext
 		}
 	}
 
-	// Syntax highlight the code portion
-	highlighted := syntaxHighlight(strings.Join(codeOnly, "\n"), fileName)
-	highlightedLines := strings.Split(highlighted, "\n")
-
-	// Layout
-	const codeIndent = "  "
-	gutterWidth := max(maxNumWidth+2, 5)
-	codeWidth := max(width-gutterWidth-len(codeIndent), 20)
-
-	theme := getTheme()
-	gutterStyle := lipgloss.NewStyle().Foreground(theme.Muted).Background(theme.GutterBg).PaddingRight(1)
-	codeStyle := lipgloss.NewStyle().Background(theme.CodeBg).PaddingLeft(1)
-
-	var result []string
-	for i, p := range parsed {
-		// If this line has no line number, it's a metadata/footer line (e.g. truncation notice).
-		if p.lineNum == "" {
-			// Render footer lines with code background but no gutter
-			footer := codeStyle.Width(codeWidth).Render(p.code)
-			emptyGutter := gutterStyle.Width(gutterWidth).Render("")
-			result = append(result, codeIndent+lipgloss.JoinHorizontal(lipgloss.Top, emptyGutter, footer))
-			continue
-		}
-
-		gutter := gutterStyle.Width(gutterWidth).Render(p.lineNum)
-
-		var codePart string
-		if i < len(highlightedLines) {
-			codePart = highlightedLines[i]
-		} else {
-			codePart = p.code
-		}
-		styledCode := codeStyle.Width(codeWidth).Render(codePart)
-
-		result = append(result, codeIndent+lipgloss.JoinHorizontal(lipgloss.Top, gutter, styledCode))
+	// Create typography with line number offset and custom formatter
+	// Match Write tool: GutterBg for line numbers, CodeBg for content
+	codeContent := strings.Join(codeLines, "\n")
+	theme := GetTheme()
+	hty := herald.Theme{
+		CodeBlock: lipgloss.NewStyle().
+			Background(theme.CodeBg).
+			PaddingLeft(1),
+		CodeLineNumber: lipgloss.NewStyle().
+			Foreground(theme.Muted).
+			Background(theme.GutterBg),
 	}
+	ty := herald.New(
+		herald.WithTheme(hty),
+		herald.WithCodeLineNumbers(true),
+		herald.WithCodeLineNumberOffset(offset),
+		herald.WithCodeFormatter(func(code, _ string) string {
+			// Use our syntax highlighter with the filename for lexer detection
+			return syntaxHighlight(code, fileName)
+		}),
+	)
 
-	// Truncation hint
+	// Render the code block
+	result := ty.CodeBlock(codeContent, lang)
+
+	// Add truncation hint if needed
 	if codeHiddenCount > 0 {
 		hint := fmt.Sprintf("...(%d more lines)", codeHiddenCount)
-		emptyGutter := gutterStyle.Width(gutterWidth).Render("")
-		hintContent := codeStyle.Width(codeWidth).
-			Foreground(theme.Muted).Italic(true).Render(hint)
-		result = append(result, codeIndent+lipgloss.JoinHorizontal(lipgloss.Top, emptyGutter, hintContent))
+		result += "\n" + lipgloss.NewStyle().Foreground(GetTheme().Muted).Italic(true).Render(hint)
 	}
 
-	return strings.Join(result, "\n")
+	// Add any footer lines
+	if len(footerLines) > 0 {
+		footer := strings.Join(footerLines, "\n")
+		result += "\n" + lipgloss.NewStyle().Foreground(GetTheme().Muted).Render(footer)
+	}
+
+	// Indent entire block to match Write/Edit tools (2 spaces)
+	const blockIndent = "  "
+	lines := strings.Split(result, "\n")
+	for i, line := range lines {
+		lines[i] = blockIndent + line
+	}
+	return strings.Join(lines, "\n")
 }
 
 // ---------------------------------------------------------------------------
@@ -513,7 +536,7 @@ func renderWriteBlock(content, fileName string, width int) string {
 	gutterWidth := numDigits + 2
 	codeWidth := max(width-gutterWidth-len(codeIndent), 20)
 
-	theme := getTheme()
+	theme := GetTheme()
 	gutterStyle := lipgloss.NewStyle().Foreground(theme.Muted).Background(theme.GutterBg).PaddingRight(1)
 	writeStyle := lipgloss.NewStyle().Background(theme.WriteBg).PaddingLeft(1)
 
@@ -528,6 +551,9 @@ func renderWriteBlock(content, fileName string, width int) string {
 		} else {
 			codePart = line
 		}
+		// Truncate the (possibly ANSI-highlighted) line to fit within
+		// the code column, preventing lipgloss from wrapping it.
+		codePart = truncateLine(codePart, codeWidth-1) // account for PaddingLeft(1)
 		styledCode := writeStyle.Width(codeWidth).Render(codePart)
 
 		result = append(result, codeIndent+lipgloss.JoinHorizontal(lipgloss.Top, gutter, styledCode))
@@ -562,7 +588,7 @@ func renderBashBody(toolResult string, width int) string {
 		return ""
 	}
 
-	theme := getTheme()
+	theme := GetTheme()
 	outputStyle := lipgloss.NewStyle().Background(theme.CodeBg).PaddingLeft(1)
 	stderrStyle := lipgloss.NewStyle().Foreground(theme.Error).Background(theme.CodeBg).PaddingLeft(1)
 
@@ -578,9 +604,15 @@ func renderBashBody(toolResult string, width int) string {
 	}
 
 	const lineIndent = "  "
+	// Truncate individual lines to the available width so they never wrap.
+	lineWidth := max(width-len(lineIndent), 20)
+	// Account for PaddingLeft(1) on the output/stderr styles
+	maxLineChars := lineWidth - 1
+
 	var rendered []string
 	inStderr := false
 	for _, line := range lines {
+		line = truncateLine(line, maxLineChars)
 		// Detect the STDERR: label that Kit's bash tool emits
 		if strings.TrimSpace(line) == "STDERR:" {
 			inStderr = true
@@ -682,23 +714,28 @@ func syntaxHighlight(source, fileName string) string {
 // Helpers
 // ---------------------------------------------------------------------------
 
-// padRight pads s with spaces to exactly width characters.
+// padRight pads s with spaces to exactly width visual characters.
+// This is ANSI-aware: it measures the visual width of s (ignoring escape
+// codes and accounting for wide characters) before padding or truncating.
 func padRight(s string, width int) string {
-	if len(s) >= width {
-		return s[:width]
+	w := xansi.StringWidth(s)
+	if w >= width {
+		return xansi.Truncate(s, width, "")
 	}
-	return s + strings.Repeat(" ", width-len(s))
+	return s + strings.Repeat(" ", width-w)
 }
 
-// truncateLine truncates a line to maxWidth, adding "…" if truncated.
+// truncateLine truncates a line to maxWidth visual characters, adding "…"
+// if truncated. This is ANSI-aware: escape codes are preserved and wide
+// characters are measured correctly.
 func truncateLine(s string, maxWidth int) string {
-	if len(s) <= maxWidth {
+	if xansi.StringWidth(s) <= maxWidth {
 		return s
 	}
 	if maxWidth < 2 {
-		return s[:maxWidth]
+		return xansi.Truncate(s, maxWidth, "")
 	}
-	return s[:maxWidth-1] + "…"
+	return xansi.Truncate(s, maxWidth, "…")
 }
 
 // ---------------------------------------------------------------------------
@@ -717,10 +754,9 @@ func renderToolBodyCompact(toolName, toolArgs, toolResult string, width int) str
 		return renderReadCompact(toolResult)
 	case toolName == "write":
 		return renderWriteCompact(toolArgs)
-	case toolName == "bash" || toolName == "run_shell_cmd" ||
-		strings.Contains(toolName, "shell") || strings.Contains(toolName, "command"):
+	case isShellTool(toolName):
 		return renderBashCompact(toolResult, width)
-	case toolName == "spawn_subagent":
+	case toolName == "subagent":
 		return renderSubagentCompact(toolResult)
 	}
 	return ""
@@ -749,7 +785,7 @@ func renderReadCompact(toolResult string) string {
 		return ""
 	}
 
-	theme := getTheme()
+	theme := GetTheme()
 	summary := fmt.Sprintf("%d lines", codeLines)
 	return lipgloss.NewStyle().Foreground(theme.Muted).Italic(true).Render(summary)
 }
@@ -770,7 +806,7 @@ func renderEditCompact(toolArgs, toolResult string) string {
 	oldCount := len(strings.Split(oldText, "\n"))
 	newCount := len(strings.Split(newText, "\n"))
 
-	theme := getTheme()
+	theme := GetTheme()
 	var summary string
 	if oldCount == newCount {
 		summary = fmt.Sprintf("%d lines modified", oldCount)
@@ -793,7 +829,7 @@ func renderWriteCompact(toolArgs string) string {
 	}
 
 	count := len(strings.Split(content, "\n"))
-	theme := getTheme()
+	theme := GetTheme()
 	summary := fmt.Sprintf("%d lines written", count)
 	return lipgloss.NewStyle().Foreground(theme.Muted).Italic(true).Render(summary)
 }
@@ -806,7 +842,7 @@ func renderLsCompact(toolResult string) string {
 	}
 
 	entries := strings.Split(content, "\n")
-	theme := getTheme()
+	theme := GetTheme()
 	summary := fmt.Sprintf("%d entries", len(entries))
 	return lipgloss.NewStyle().Foreground(theme.Muted).Italic(true).Render(summary)
 }
@@ -844,26 +880,24 @@ func renderBashCompact(toolResult string, width int) string {
 
 	if len(outputLines) == 0 {
 		if exitCode != "" {
-			theme := getTheme()
+			theme := GetTheme()
 			return lipgloss.NewStyle().Foreground(theme.Error).Render(exitCode)
 		}
 		return ""
 	}
 
 	const maxLines = 3
-	theme := getTheme()
+	theme := GetTheme()
 
 	display := outputLines
 	if len(display) > maxLines {
 		display = display[:maxLines]
 	}
 
-	// Truncate each line to available width
+	// Truncate each line to available width (ANSI-aware)
 	lineMax := max(width-4, 20)
 	for i, line := range display {
-		if len(line) > lineMax {
-			display[i] = line[:lineMax-3] + "..."
-		}
+		display[i] = truncateLine(line, lineMax)
 	}
 
 	summary := strings.Join(display, "\n")
@@ -881,10 +915,10 @@ func renderBashCompact(toolResult string, width int) string {
 // Subagent tool renderers — show only summary, not full output
 // ---------------------------------------------------------------------------
 
-// renderSubagentBody renders a clean summary of subagent results.
-// Extracts timing/token info and shows only a brief summary instead of raw output.
+// renderSubagentBody renders a clean summary of subagent results with bash-style
+// background styling for consistency with other tools.
 func renderSubagentBody(toolResult string, width int) string {
-	theme := getTheme()
+	theme := GetTheme()
 	result := strings.TrimSpace(toolResult)
 	if result == "" {
 		return ""
@@ -902,9 +936,19 @@ func renderSubagentBody(toolResult string, width int) string {
 	// First line is always the status summary
 	statusLine := lines[0]
 
-	// Build a clean summary
-	var summary strings.Builder
-	summary.WriteString(lipgloss.NewStyle().Foreground(theme.Muted).Render(statusLine))
+	// Build content lines for display with bash-style background
+	outputStyle := lipgloss.NewStyle().Background(theme.CodeBg).PaddingLeft(1)
+	errorStyle := lipgloss.NewStyle().Foreground(theme.Error).Background(theme.CodeBg).PaddingLeft(1)
+
+	const lineIndent = "  "
+	lineWidth := max(width-len(lineIndent), 20)
+	maxLineChars := lineWidth - 1 // account for PaddingLeft(1)
+
+	var contentLines []string
+
+	// Add status line
+	styledStatus := outputStyle.Width(lineWidth).Render(truncateLine(statusLine, maxLineChars))
+	contentLines = append(contentLines, lineIndent+styledStatus)
 
 	// For successful results, extract a brief preview of the actual result
 	if strings.Contains(statusLine, "successfully") {
@@ -912,25 +956,45 @@ func renderSubagentBody(toolResult string, width int) string {
 		if _, resultContent, found := strings.Cut(result, "Result:\n"); found {
 			resultContent = strings.TrimSpace(resultContent)
 			if resultContent != "" {
-				// Show first 3 meaningful lines as preview
-				preview := extractSubagentPreview(resultContent, 3, width-4)
-				if preview != "" {
-					summary.WriteString("\n\n")
-					summary.WriteString(lipgloss.NewStyle().
-						Foreground(theme.Muted).
-						Italic(true).
-						Render(preview))
+				// Show first few meaningful lines as preview
+				previewLines := extractSubagentPreviewLines(resultContent, 5, maxLineChars)
+				if len(previewLines) > 0 {
+					// Add blank separator line
+					blankLine := outputStyle.Width(lineWidth).Render("")
+					contentLines = append(contentLines, lineIndent+blankLine)
+
+					for _, line := range previewLines {
+						styled := outputStyle.Width(lineWidth).Render(line)
+						contentLines = append(contentLines, lineIndent+styled)
+					}
+				}
+			}
+		}
+	} else {
+		// For failed results, show error info
+		if _, errorContent, found := strings.Cut(result, "Error:\n"); found {
+			errorContent = strings.TrimSpace(errorContent)
+			if errorContent != "" {
+				previewLines := extractSubagentPreviewLines(errorContent, 3, maxLineChars)
+				if len(previewLines) > 0 {
+					blankLine := outputStyle.Width(lineWidth).Render("")
+					contentLines = append(contentLines, lineIndent+blankLine)
+
+					for _, line := range previewLines {
+						styled := errorStyle.Width(lineWidth).Render(line)
+						contentLines = append(contentLines, lineIndent+styled)
+					}
 				}
 			}
 		}
 	}
 
-	return summary.String()
+	return strings.Join(contentLines, "\n")
 }
 
-// extractSubagentPreview extracts the first N non-empty lines from content,
-// truncating each line to maxWidth.
-func extractSubagentPreview(content string, maxLines, maxWidth int) string {
+// extractSubagentPreviewLines extracts the first N non-empty lines from content,
+// truncating each line to maxWidth. Returns as a slice of strings.
+func extractSubagentPreviewLines(content string, maxLines, maxWidth int) []string {
 	lines := strings.Split(content, "\n")
 	var preview []string
 
@@ -940,10 +1004,8 @@ func extractSubagentPreview(content string, maxLines, maxWidth int) string {
 			continue
 		}
 
-		// Truncate long lines
-		if len(trimmed) > maxWidth {
-			trimmed = trimmed[:maxWidth-3] + "..."
-		}
+		// Truncate long lines (ANSI-aware)
+		trimmed = truncateLine(trimmed, maxWidth)
 		preview = append(preview, trimmed)
 
 		if len(preview) >= maxLines {
@@ -951,12 +1013,6 @@ func extractSubagentPreview(content string, maxLines, maxWidth int) string {
 		}
 	}
 
-	if len(preview) == 0 {
-		return ""
-	}
-
-	result := strings.Join(preview, "\n")
-
 	// Count remaining lines for "more" indicator
 	totalLines := 0
 	for _, line := range lines {
@@ -965,10 +1021,10 @@ func extractSubagentPreview(content string, maxLines, maxWidth int) string {
 		}
 	}
 	if totalLines > maxLines {
-		result += fmt.Sprintf("\n...(%d more lines)", totalLines-maxLines)
+		preview = append(preview, fmt.Sprintf("...(%d more lines)", totalLines-maxLines))
 	}
 
-	return result
+	return preview
 }
 
 // renderSubagentCompact returns a brief one-line summary for subagent results.
@@ -978,7 +1034,7 @@ func renderSubagentCompact(toolResult string) string {
 		return ""
 	}
 
-	theme := getTheme()
+	theme := GetTheme()
 
 	// Extract just the first line which contains the status
 	lines := strings.Split(result, "\n")

internal/ui/tree_selector.go

@@ -430,6 +430,8 @@ func (ts *TreeSelectorComponent) renderNode(node FlatNode, isCursor, isLeaf bool
 		}
 	case *session.BranchSummaryEntry:
 		style = lipgloss.NewStyle().Foreground(theme.Warning).Italic(true)
+	case *session.CompactionEntry:
+		style = lipgloss.NewStyle().Foreground(theme.Info).Italic(true)
 	default:
 		style = lipgloss.NewStyle().Foreground(theme.Muted)
 	}
@@ -483,6 +485,13 @@ func (ts *TreeSelectorComponent) entryDisplayText(entry any) string {
 		}
 		return fmt.Sprintf("branch summary: %s", summary)
 
+	case *session.CompactionEntry:
+		summary := e.Summary
+		if len(summary) > 60 {
+			summary = summary[:60] + "..."
+		}
+		return fmt.Sprintf("compaction: %s", summary)
+
 	case *session.LabelEntry:
 		return fmt.Sprintf("label: %s", e.Label)
 

internal/ui/usage_tracker.go

@@ -134,13 +134,23 @@ func (ut *UsageTracker) EstimateAndUpdateUsage(inputText, outputText string) {
 }
 
 // SetContextTokens records the approximate current context window utilization.
-// This should be set from the final API call's input + output tokens (i.e.
-// FinalResponse.Usage) rather than the aggregate TotalUsage, because TotalUsage
+// This should be set from FinalUsage.InputTokens, which already includes the
+// full conversation history (system prompt + all previous messages). Do NOT
+// add OutputTokens as that would double-count (output becomes input next turn).
+// Use FinalResponse.Usage rather than aggregate TotalUsage, because TotalUsage
 // sums across all tool-calling steps and overstates the actual window fill level.
 func (ut *UsageTracker) SetContextTokens(tokens int) {
 	ut.mu.Lock()
 	defer ut.mu.Unlock()
-	ut.contextTokens = tokens
+	// Track the maximum context seen so far. In multi-step tool calls,
+	// FinalUsage.InputTokens may reflect only the last step's input, which
+	// can be smaller than previous steps. We want to show the largest context
+	// the model has processed in this session.
+	if tokens > ut.contextTokens {
+		ut.contextTokens = tokens
+	}
+	// If tokens < current, we keep the larger value (no-op)
+	// This prevents the display from dropping during multi-step tool calls.
 }
 
 // RenderUsageInfo generates a formatted string displaying current usage statistics
@@ -151,10 +161,6 @@ func (ut *UsageTracker) RenderUsageInfo() string {
 	ut.mu.RLock()
 	defer ut.mu.RUnlock()
 
-	if ut.sessionStats.RequestCount == 0 {
-		return ""
-	}
-
 	baseStyle := lipgloss.NewStyle()
 
 	// Display the current context window token count (from the last API call),
@@ -266,3 +272,14 @@ func (ut *UsageTracker) SetWidth(width int) {
 	defer ut.mu.Unlock()
 	ut.width = width
 }
+
+// UpdateModelInfo updates the model information and OAuth status when the model
+// is switched mid-session. This ensures token costs and context limits are
+// calculated correctly for the new model.
+func (ut *UsageTracker) UpdateModelInfo(modelInfo *models.ModelInfo, provider string, isOAuth bool) {
+	ut.mu.Lock()
+	defer ut.mu.Unlock()
+	ut.modelInfo = modelInfo
+	ut.provider = provider
+	ut.isOAuth = isOAuth
+}

internal/ui/usage_tracker_render_test.go

@@ -67,3 +67,62 @@ func TestUsageTracker_RenderUsageInfo_OAuth(t *testing.T) {
 		t.Errorf("Expected regular rendered output to show actual cost, got: %s", regularRendered)
 	}
 }
+
+func TestUsageTracker_RenderUsageInfo_StartupState(t *testing.T) {
+	// Create a mock model info with costs and context limit
+	modelInfo := &models.ModelInfo{
+		ID:   "claude-3-5-sonnet-20241022",
+		Name: "Claude 3.5 Sonnet v2",
+		Cost: models.Cost{
+			Input:  3.0,
+			Output: 15.0,
+		},
+		Limit: models.Limit{
+			Context: 200000,
+			Output:  8192,
+		},
+	}
+
+	// Test startup state (no requests made yet) - Regular API key
+	regularTracker := NewUsageTracker(modelInfo, "anthropic", 80, false)
+	rendered := stripAnsi(regularTracker.RenderUsageInfo())
+
+	// Should NOT return empty string on startup
+	if rendered == "" {
+		t.Errorf("Expected non-empty output on startup, got empty string")
+	}
+
+	// Should show 0 tokens
+	if !strings.Contains(rendered, "Tokens: 0") {
+		t.Errorf("Expected 'Tokens: 0' on startup, got: %s", rendered)
+	}
+
+	// Should NOT show percentage when tokens are 0
+	if strings.Contains(rendered, "(%") {
+		t.Errorf("Expected no percentage on startup with 0 tokens, got: %s", rendered)
+	}
+
+	// Should show $0.0000 cost for regular API key
+	if !strings.Contains(rendered, "Cost: $0.0000") {
+		t.Errorf("Expected 'Cost: $0.0000' on startup, got: %s", rendered)
+	}
+
+	// Test startup state (no requests made yet) - OAuth
+	oauthTracker := NewUsageTracker(modelInfo, "anthropic", 80, true)
+	oauthRendered := stripAnsi(oauthTracker.RenderUsageInfo())
+
+	// Should NOT return empty string on startup
+	if oauthRendered == "" {
+		t.Errorf("Expected non-empty output on startup for OAuth, got empty string")
+	}
+
+	// Should show 0 tokens for OAuth
+	if !strings.Contains(oauthRendered, "Tokens: 0") {
+		t.Errorf("Expected 'Tokens: 0' on startup for OAuth, got: %s", oauthRendered)
+	}
+
+	// Should show $0.00 cost for OAuth
+	if !strings.Contains(oauthRendered, "Cost: $0.00") {
+		t.Errorf("Expected 'Cost: $0.00' on startup for OAuth, got: %s", oauthRendered)
+	}
+}

pkg/extensions/test/harness.go

@@ -52,7 +52,6 @@ type Harness struct {
 	t       *testing.T
 	runner  *extensions.Runner
 	context *MockContext
-	extPath string
 }
 
 // New creates a new test harness for the given test.
@@ -72,15 +71,9 @@ func New(t *testing.T) *Harness {
 func (h *Harness) LoadFile(path string) *extensions.LoadedExtension {
 	h.t.Helper()
 
-	// Verify file exists
-	if _, err := os.Stat(path); err != nil {
-		h.t.Fatalf("extension file not found: %s: %v", path, err)
-	}
-
-	// Read extension source
 	src, err := os.ReadFile(path)
 	if err != nil {
-		h.t.Fatalf("failed to read extension file: %v", err)
+		h.t.Fatalf("failed to read extension file %s: %v", path, err)
 	}
 
 	return h.loadSource(string(src), path)
@@ -144,7 +137,6 @@ func (h *Harness) loadSource(src string, path string) *extensions.LoadedExtensio
 
 	// Create runner with the loaded extension
 	h.runner = extensions.NewRunner([]extensions.LoadedExtension{*ext})
-	h.extPath = path
 
 	// Wire the mock context
 	h.runner.SetContext(h.context.ToContext())
@@ -222,11 +214,3 @@ func (h *Harness) RegisteredCommands() []extensions.CommandDef {
 	}
 	return h.runner.RegisteredCommands()
 }
-
-// MustLoad is like LoadFile but fails the test immediately on error.
-// It returns the harness for chaining.
-func (h *Harness) MustLoad(path string) *Harness {
-	h.t.Helper()
-	h.LoadFile(path)
-	return h
-}

pkg/extensions/test/mock.go

@@ -59,29 +59,12 @@ type MockContext struct {
 	Overlays []extensions.OverlayConfig
 }
 
-// StatusBarEntry represents a recorded status bar entry
-type StatusBarEntry struct {
-	Key      string
-	Text     string
-	Priority int
-}
-
 // NewMockContext creates a new mock context with default values.
 func NewMockContext() *MockContext {
 	return &MockContext{
-		Prints:        make([]string, 0),
-		PrintInfos:    make([]string, 0),
-		PrintErrors:   make([]string, 0),
-		PrintBlocks:   make([]extensions.PrintBlockOpts, 0),
-		Messages:      make([]string, 0),
-		CancelSends:   make([]string, 0),
 		Widgets:       make(map[string]extensions.WidgetConfig),
-		RemovedIDs:    make([]string, 0),
 		StatusEntries: make(map[string]extensions.StatusBarEntry),
-		RemovedStatus: make([]string, 0),
-		EditorTexts:   make([]string, 0),
 		Options:       make(map[string]string),
-		Overlays:      make([]extensions.OverlayConfig, 0),
 		Interactive:   true,
 		SessionID:     "test-session",
 		CWD:           "/test",

pkg/kit/README.md

@@ -1,6 +1,6 @@
 # KIT SDK
 
-The KIT SDK allows you to use KIT programmatically from Go applications without spawning OS processes.
+The KIT SDK (`pkg/kit`) lets you embed Kit's full agent capabilities — LLM interactions, tool execution, session management, streaming, hooks — into any Go application.
 
 ## Installation
 
@@ -17,26 +17,26 @@ import (
     "context"
     "fmt"
     "log"
-    
+
     kit "github.com/mark3labs/kit/pkg/kit"
 )
 
 func main() {
     ctx := context.Background()
-    
+
     // Create Kit instance with default configuration
     host, err := kit.New(ctx, nil)
     if err != nil {
         log.Fatal(err)
     }
-    defer host.Close()
-    
+    defer func() { _ = host.Close() }()
+
     // Send a prompt
     response, err := host.Prompt(ctx, "What is 2+2?")
     if err != nil {
         log.Fatal(err)
     }
-    
+
     fmt.Println(response)
 }
 ```
@@ -56,11 +56,23 @@ You can override specific settings:
 ```go
 host, err := kit.New(ctx, &kit.Options{
     Model:        "ollama/llama3",            // Override model
-    SystemPrompt: "You are a helpful bot",   // Override system prompt
-    ConfigFile:   "/path/to/config.yml",     // Use specific config file
-    MaxSteps:     10,                        // Override max steps
-    Streaming:    true,                      // Enable streaming
-    Quiet:        true,                      // Suppress debug output
+    SystemPrompt: "You are a helpful bot",    // Override system prompt
+    ConfigFile:   "/path/to/config.yml",      // Use specific config file
+    MaxSteps:     10,                         // Override max steps
+    Streaming:    true,                       // Enable streaming
+    Quiet:        true,                       // Suppress debug output
+
+    // Session options
+    SessionPath:  "./session.jsonl",          // Open specific session
+    Continue:     true,                       // Resume most recent session
+    NoSession:    true,                       // Ephemeral mode
+
+    // Tool options
+    Tools:        []kit.Tool{kit.NewBashTool()}, // Replace default tool set
+    ExtraTools:   []kit.Tool{myTool},            // Add alongside defaults
+
+    // Compaction
+    AutoCompact:  true,                       // Auto-compact near context limit
 })
 ```
 
@@ -71,22 +83,28 @@ host, err := kit.New(ctx, &kit.Options{
 Monitor tool execution in real-time:
 
 ```go
-response, err := host.PromptWithCallbacks(
+unsub := host.OnToolCall(func(e kit.ToolCallEvent) {
+    fmt.Printf("Calling tool: %s\n", e.ToolName)
+})
+defer unsub()
+
+unsub2 := host.OnToolResult(func(e kit.ToolResultEvent) {
+    if e.IsError {
+        fmt.Printf("Tool %s failed: %s\n", e.ToolName, e.Result)
+    } else {
+        fmt.Printf("Tool %s succeeded\n", e.ToolName)
+    }
+})
+defer unsub2()
+
+unsub3 := host.OnStreaming(func(e kit.MessageUpdateEvent) {
+    fmt.Print(e.Chunk)
+})
+defer unsub3()
+
+response, err := host.Prompt(
     ctx,
     "List files in the current directory",
-    func(name, args string) {
-        fmt.Printf("Calling tool: %s\n", name)
-    },
-    func(name, args, result string, isError bool) {
-        if isError {
-            fmt.Printf("Tool %s failed: %s\n", name, result)
-        } else {
-            fmt.Printf("Tool %s succeeded\n", name)
-        }
-    },
-    func(chunk string) {
-        fmt.Print(chunk) // Stream output
-    },
 )
 ```
 
@@ -102,35 +120,52 @@ host.Prompt(ctx, "My name is Alice")
 response, _ := host.Prompt(ctx, "What's my name?")
 // Response: "Your name is Alice"
 
-// Save session
-host.SaveSession("./session.json")
-
-// Load session later
-host.LoadSession("./session.json")
-
-// Clear session
+// Clear conversation history
 host.ClearSession()
 ```
 
+## Re-exported Types
+
+The SDK re-exports types so you don't need direct internal imports:
+
+```go
+// Message types
+kit.Message, kit.MessageRole, kit.ContentPart
+kit.TextContent, kit.ReasoningContent, kit.ToolCall, kit.ToolResult, kit.Finish
+kit.RoleUser, kit.RoleAssistant, kit.RoleTool, kit.RoleSystem
+
+// LLM types (re-exported from the underlying LLM library)
+kit.LLMMessage, kit.LLMUsage, kit.LLMResponse, kit.LLMFilePart
+
+// Conversion helpers
+msgs := kit.ConvertToLLMMessages(&msg)   // SDK message → LLM messages
+msg := kit.ConvertFromLLMMessage(fMsg)    // LLM message → SDK message
+```
+
 ## API Reference
 
 ### Types
 
 - `Kit` - Main SDK type
 - `Options` - Configuration options
-- `Message` - Conversation message
-- `ToolCall` - Tool invocation details
+- `Message` - Conversation message with typed content parts
+- `Tool` - Agent tool interface
+- `TurnResult` - Full result from a prompt including usage stats
 
-### Methods
+### Key Methods
 
 - `New(ctx, opts)` - Create new Kit instance
-- `Prompt(ctx, message)` - Send message and get response
-- `PromptWithCallbacks(ctx, message, ...)` - Send message with progress callbacks
-- `LoadSession(path)` - Load session from file
-- `SaveSession(path)` - Save session to file
-- `ClearSession()` - Clear conversation history
-- `GetSessionManager()` - Get session manager for advanced usage
+- `Prompt(ctx, message)` - Send message and get response string
+- `PromptResult(ctx, message)` - Send message and get full TurnResult
+- `PromptWithOptions(ctx, message, opts)` - Prompt with per-call options
+- `Steer(ctx, instruction)` - System-level steering
+- `FollowUp(ctx, text)` - Continue without new user input
+- `SetModel(ctx, model)` - Switch model at runtime
 - `GetModelString()` - Get current model string
+- `GetModelInfo()` - Get model capabilities and limits
+- `ClearSession()` - Clear conversation history
+- `GetSessionPath()` - Get session file path
+- `GetSessionID()` - Get session UUID
 - `Close()` - Clean up resources
 
 ## Environment Variables

pkg/kit/auth.go

@@ -1,6 +1,10 @@
 package kit
 
-import "github.com/mark3labs/kit/internal/auth"
+import (
+	"os"
+
+	"github.com/mark3labs/kit/internal/auth"
+)
 
 // CredentialManager manages API keys and OAuth credentials.
 type CredentialManager = auth.CredentialManager
@@ -9,6 +13,10 @@ type CredentialManager = auth.CredentialManager
 // and API key authentication methods.
 type AnthropicCredentials = auth.AnthropicCredentials
 
+// OpenAICredentials holds OpenAI API credentials supporting both OAuth
+// and API key authentication methods.
+type OpenAICredentials = auth.OpenAICredentials
+
 // CredentialStore holds all stored credentials for various providers.
 type CredentialStore = auth.CredentialStore
 
@@ -42,3 +50,32 @@ func GetAnthropicAPIKey() string {
 	}
 	return key
 }
+
+// HasOpenAICredentials checks if valid OpenAI credentials are stored
+// (either OAuth token or API key).
+func HasOpenAICredentials() bool {
+	cm, err := auth.NewCredentialManager()
+	if err != nil {
+		return false
+	}
+	has, err := cm.HasOpenAICredentials()
+	if err != nil {
+		return false
+	}
+	return has
+}
+
+// GetOpenAIAPIKey resolves the OpenAI API key using the standard
+// resolution order: stored credentials -> OPENAI_API_KEY env var.
+// Returns an empty string if no key is found.
+func GetOpenAIAPIKey() string {
+	cm, err := auth.NewCredentialManager()
+	if err == nil {
+		// Try to get valid access token (handles OAuth refresh)
+		if token, err := cm.GetValidOpenAIAccessToken(); err == nil && token != "" {
+			return token
+		}
+	}
+	// Fall back to environment variable
+	return os.Getenv("OPENAI_API_KEY")
+}