docs(session): correct DefaultSessionDir convention comment

- Stale comment showed ~/.kit/sessions/--<cwd-path>--/ which does not
  match the actual encoding (no leading/trailing dashes)
- Update to reflect the real format and point to encodeCwdForDir for
  full rules

.kit/prompts/create-pr.md

@@ -0,0 +1,47 @@
+---
+description: Open a GitHub PR for the current branch using the repo's PR template
+---
+
+Open a GitHub pull request for the current branch, filling out the repository's PR template with a description grounded in the actual commits and diff.
+
+## Steps
+
+1. **Verify the branch is pushed**:
+   - `git status -sb` and `git log @{u}..HEAD --oneline 2>/dev/null` — if there is no upstream or unpushed commits, run `git push -u origin "$(git branch --show-current)"` first
+   - If the working tree is dirty, stop and tell the user to commit first (suggest `/commit-push`)
+2. **Gather context**:
+   - `git log origin/main..HEAD --oneline` — list of commits going into the PR
+   - `git diff origin/main...HEAD --stat` then `git diff origin/main...HEAD` — read the actual changes
+   - Identify the linked issue (from commit messages, branch name, or extra user input: $@) — capture as `Fixes #N` if applicable
+3. **Locate the PR template**:
+   - Check `.github/pull_request_template.md`, `.github/PULL_REQUEST_TEMPLATE.md`, or `docs/pull_request_template.md`
+   - If none exists, use a minimal `## Description` / `## Type of Change` / `## Checklist` structure
+4. **Draft the PR body** by filling out the template:
+   - **Description**: 1–3 short paragraphs explaining *what* changed and *why*, grounded in the diff. Include a brief before/after example for new APIs when useful.
+   - **Fixes #N**: only if there is a real linked issue
+   - **Type of Change**: tick the single most accurate box with `[x]` (leave others as `[ ]`)
+   - **Checklist**: tick items that are genuinely true (style, self-review, tests added, docs updated)
+   - **Additional Information**: bullet list of added / modified files and any backward-compatibility notes
+   - Remove template sections explicitly marked "remove if not applicable" (e.g. MCP Spec Compliance) when they don't apply
+5. **Write the body to a temp file**: `/tmp/pr-body-<branch-or-issue>.md` — never inline a long body via `--body`, always use `--body-file`
+6. **Choose the title**: prefer the subject of the primary commit if it already follows Conventional Commits; otherwise craft one in the same style (`<type>(<scope>): <imperative summary>`, ≤72 chars)
+7. **Create the PR**:
+   ```
+   gh pr create \
+     --title "<title>" \
+     --body-file /tmp/pr-body-<...>.md \
+     --base main \
+     --head "$(git branch --show-current)"
+   ```
+   Use the repo's actual default branch if it isn't `main` (`gh repo view --json defaultBranchRef -q .defaultBranchRef.name`)
+8. **Report the PR URL** returned by `gh` and stop
+
+## Guidelines
+
+- Read the diff and commit messages — do **not** invent features that aren't in the code
+- One PR per logical change; if the branch contains unrelated commits, surface that and ask before continuing
+- Keep the description focused on reviewer-relevant information (what / why), not a replay of the diff
+- Only check checklist boxes that are actually satisfied; leave the rest unchecked rather than lying
+- If `gh` is not authenticated (`gh auth status` fails), stop and tell the user
+
+$@

.kit/prompts/feature-request.md

@@ -2,7 +2,7 @@
 description: Create a feature request using the GitHub template
 ---
 
-Create a feature request for the Kit repository. The user wants to request: $+
+Create a feature request for the Kit repository. The user wants to request: $@
 
 ## Feature Request Template
 
@@ -16,7 +16,7 @@ This prompt uses the `feature_request` GitHub template which requires:
 
 ## Steps
 
-1. **Understand the request** from `$+`
+1. **Understand the request** from the user input: $@
    - What capability is missing?
    - What would the ideal behavior look like?
 

.kit/prompts/file-issue.md

@@ -2,7 +2,7 @@
 description: File a GitHub issue using the appropriate template
 ---
 
-File a GitHub issue for the Kit repository. The user wants to create an issue about: $+
+File a GitHub issue for the Kit repository. The user wants to create an issue about: $@
 
 ## Issue Templates Available
 
@@ -16,7 +16,7 @@ This repository has structured issue templates. You MUST use the appropriate tem
 
 ## Steps
 
-1. **Determine the issue type** from `$+`:
+1. **Determine the issue type** from the user input: $@
    - Bug → use `--template bug_report`
    - Feature → use `--template feature_request`  
    - Documentation → use `--template documentation`

.kit/prompts/fix-issue.md

@@ -0,0 +1,61 @@
+---
+description: Implement the fix/feature/docs change requested by a GitHub issue
+---
+
+Resolve GitHub issue #$1 by reading it, classifying it, and producing the appropriate code or doc change. **Stop once the working tree contains the change** — committing, pushing, and opening a PR are handled by `/commit-push` and `/create-pr`.
+
+## Steps
+
+1. **Fetch the issue**:
+   - Run: gh issue view $1 --json number,title,body,labels,state,author,comments
+   - If the issue is closed, stop and ask the user whether to proceed
+   - Read the **entire** thread including comments — the latest comment often refines the ask
+
+2. **Classify the issue** from labels, title prefix, and body content:
+   - `bug` / `fix:` → reproduce, then fix
+   - `enhancement` / `feature` / `feat:` → design, then implement
+   - `documentation` / `docs:` → locate and update docs
+   - `question` / `discussion` → answer in a comment, do **not** write code
+   - Anything else → ask the user how to proceed
+
+3. **Create a working branch** off the default branch:
+   - `git checkout main && git pull --ff-only`
+   - Branch name: <type>/$1-<slug> (e.g. `fix/42-borderColor-ignored`, `feat/57-keyboard-clear`, `docs/63-widget-lifecycle`)
+
+4. **Do the work** based on type:
+
+   ### Bug (`bug` label / `fix:` title)
+   - Reproduce the failure first (write a failing test if feasible) — if you cannot reproduce, comment on the issue asking for clarification and stop
+   - Locate the root cause; do not patch symptoms
+   - Add or extend a regression test that fails before and passes after the fix
+   - Run `go test ./... -race` and `golangci-lint run`
+
+   ### Feature (`enhancement` / `feature` label / `feat:` title)
+   - Re-read the motivation and proposed implementation in the issue body
+   - For large, ambiguous, or breaking changes, sketch the design in a comment on the issue and wait for sign-off before writing code
+   - Implement behind sensible defaults; add godoc on every exported symbol
+   - Add unit tests covering the new behaviour and edge cases
+   - Update `README.md` / `docs/` if the public surface changed
+   - Run `go test ./... -race` and `golangci-lint run`
+
+   ### Documentation (`documentation` label / `docs:` title)
+   - Open the file/URL referenced in the issue's "Documentation Location"
+   - Apply the suggested improvement; verify code samples compile (`go build ./...`)
+   - No tests required, but run `golangci-lint run` if Go files were touched
+
+5. **Report**:
+   - Branch name (`git branch --show-current`)
+   - Summary of files changed (`git status -s`) and the diff highlights
+   - Test/lint results (pass/fail with key output)
+   - Suggest the next step explicitly:
+     - `/commit-push` to commit with a Conventional Commit subject (the message should reference `(#$1)` and include `Fixes #$1` so merge auto-closes)
+     - then `/create-pr $1` to open the pull request
+
+## Guidelines
+
+- This prompt **stops at a clean working tree with the change applied** — do not run `git commit`, `git push`, or `gh pr create`
+- If the issue is unclear, post a clarifying comment on the issue and stop; do not guess
+- Keep the change scoped to the issue; surface unrelated cleanups separately
+- For breaking changes or architecture shifts, propose the design on the issue first and wait for maintainer sign-off
+- If the issue is a duplicate or already fixed on `main`, comment with the reference and stop
+- Do not close the issue manually — the eventual PR's `Fixes #$1` handles that on merge

.kit/prompts/new-prompt.md

@@ -2,7 +2,7 @@
 description: Scaffold a new prompt template in .kit/prompts/
 ---
 
-Create a new kit prompt template. The user wants a prompt that does: $+
+Create a new kit prompt template. The user wants a prompt that does: $@
 
 ## What a prompt template is
 
@@ -16,30 +16,64 @@ It becomes a `/slug` slash command in the kit input box — typed as `/filename`
 description: One-line description shown in autocomplete
 ---
 
-Body text of the prompt. Use $@ for all user-supplied arguments,
-$1 $2 etc. for positional arguments.
+Body text of the prompt. Reference user-supplied arguments
+with positional placeholders (see "Argument placeholders" below).
 ```
 
 - **Filename** → slug: `commit-push.md` becomes `/commit-push`
 - **Frontmatter**: only `description` is recognised; keep it under ~80 chars
 - **Body**: plain markdown; the full text is submitted as the user's message when the template fires
-- **Arguments**: `$+` expands to everything the user typed after the slash command name
-  (requires at least one argument); `$@` is the same but allows zero arguments;
-  `$1`, `$2` for individual positional args; omit entirely if no arguments are needed
+- **Required args**: kit infers required positional args from the highest `$N` it finds *outside* backtick/tilde code fences — a stray `$2` in active prose means kit will refuse to run without 2 arguments
+
+## Argument placeholders
+
+kit performs shell-style substitution before sending the prompt to the model:
+
+- `$1`, `$2`, … — positional arguments (1-indexed)
+- `${1}`, `${2}`, … — same, brace form (use when followed by digits/letters: `${1}_suffix`)
+- `$@` — all arguments joined by spaces (zero or more, optional)
+- `$+` — all arguments, **at least one required**
+- `$ARGUMENTS` / `${ARGUMENTS}` — alias for `$@`
+- `${@:N}` — args from the Nth onwards (1-indexed, bash-style)
+- `${@:N:L}` — `L` args starting from the Nth
+
+### ⚠️ Critical: code fences and inline code preserve placeholders verbatim
+
+Anything inside triple-backtick fences, `~~~` fences, or single-backtick `inline` code spans is **left untouched** so example code samples don't get corrupted. That means:
+
+- An inline-coded `gh issue view $1` stays literal `$1` in the model's input ❌
+- The same command without backticks: gh issue view $1 → expands to `gh issue view 42` ✓
+
+**Rule of thumb:** if you want a placeholder to substitute, keep it outside backticks and fences. If you want a literal `$1` in the output (e.g. teaching the user shell syntax), put it inside backticks.
+
+### Workarounds for "I want it to look like code AND substitute"
+
+1. **Drop the backticks** around just the placeholder portion — the rest can still read as a command line in prose
+2. **Use a 4-space-indented code block** instead of a triple-backtick fence — kit only skips backtick/tilde fences, so indentation-style code blocks still get substitution:
+
+       git push -u origin "$(git branch --show-current)"
+       gh pr create --title "fix: ... (#$1)" --base main
+
+3. **Bind once, reference loosely**: put `Issue: $1` at the top in prose, then leave the backticked examples literal — the model will substitute mentally
 
 ## Steps
 
-1. **Understand the workflow** the user described in `$+` — ask a clarifying question if the intent is ambiguous
+1. **Understand the workflow** the user described in $@ — ask a clarifying question if the intent is ambiguous
 2. **Choose a filename**: short, lowercase, hyphen-separated, descriptive (e.g. `code-review.md`)
 3. **Write the description**: one sentence, imperative, fits in autocomplete
-4. **Draft the body**:
-   - Open with a single sentence stating the goal
+4. **Decide on arguments**:
+   - No args needed → omit placeholders entirely
+   - One required value (issue number, PR url, file path) → use `$1`
+   - Free-form trailing context → end with a single `$@` line
+   - Multiple distinct values → use `$1`, `$2`, … and document each at the top
+5. **Draft the body**:
+   - Open with a single sentence stating the goal, weaving in `$1`/`$@` where the value belongs
    - Use `## Steps` for multi-step workflows; use plain prose for simple prompts
    - Be specific: name commands, flags, and file paths where relevant
-   - End with `$+` on its own line if the user must pass context; use `$@` if arguments
-     are optional; omit if the prompt is self-contained
-5. **Write the file** to `.kit/prompts/<slug>.md`
-6. **Confirm** by showing the final file content and the slash command that activates it
+   - **Audit every backtick and code fence**: any `$N` or `$@` inside them will not expand — was that intentional? If not, apply one of the workarounds above
+6. **Write the file** to `.kit/prompts/<slug>.md`
+7. **Verify substitution** by mentally (or actually) replacing `$1`/`$@` with a sample value and confirming every reference resolves — and that the prompt's *own* example snippets don't accidentally bump the required-arg count (wrap illustrative `$N` examples in triple-backtick fences, not 4-space indentation, so `RequiredArgs()` ignores them)
+8. **Confirm** by showing the final file content and the slash command that activates it (e.g. `/code-review 42`)
 
 ## Guidelines
 
@@ -47,3 +81,4 @@ $1 $2 etc. for positional arguments.
 - Prefer concrete steps over vague instructions
 - A prompt that does one thing well beats one that tries to cover every edge case
 - If the workflow already exists as a prompt, suggest extending it instead of duplicating
+- When in doubt about substitution behaviour, write the file and run `/<slug> testvalue` once to confirm — wrong placement of backticks is the #1 failure mode

.kit/prompts/update-docs.md

@@ -0,0 +1,52 @@
+---
+description: Audit and update project documentation (README and docs site) for a recent change
+---
+
+Review recent code changes, identify all documentation surfaces that should
+mention them, and update each one — grounded in the actual diff, not guesses.
+
+## Steps
+
+1. **Identify the change**:
+   - If the user input ($@) names a commit / PR / branch / topic, use that as the focus
+   - Otherwise inspect `git log origin/main..HEAD --oneline` and `git diff origin/main...HEAD --stat` to discover what shipped on the current branch
+   - Read the actual diff (`git diff origin/main...HEAD`) — never document features that aren't in the code
+
+2. **Inventory the doc surfaces**:
+   - `README.md` at the repo root
+   - Any docs site (commonly `www/`, `docs/`, `site/`) — list its pages and identify the one(s) most thematically related to the change
+   - Inline godoc / API reference comments on the new exported symbols
+   - `CHANGELOG.md` if the project keeps one
+   - Any `examples/` directory entries that demonstrate the affected area
+
+3. **Audit each surface** with `grep`:
+   - Search for the names of related existing APIs (e.g. if you added `IterTools`, grep for `ListTools`) to find every page that already discusses the area
+   - Decide for each hit: does it need a cross-reference, a side-by-side comparison, or to stay untouched?
+
+4. **Decide where new content lives**:
+   - Prefer extending an existing page over creating a new one
+   - For a docs site, place new sections near related content (check the page's `## Heading` outline first)
+   - Skip surfaces that genuinely don't apply (e.g. a server-focused README for a client-only change) and say so explicitly
+
+5. **Draft the updates**:
+   - Lead with a one-sentence statement of what's new and why
+   - Show concrete code examples copied from real signatures — verify against the source files
+   - Include a comparison / "when to use which" table when adding an alternative to an existing API
+   - Note backwards-compatibility behaviour if relevant
+
+6. **Verify the docs build** before committing:
+   - For vocs / docusaurus / mkdocs sites, run the local build command (e.g. `npx vocs build`, `mkdocs build`) and fix any MDX/markdown errors
+   - For godoc, run `go vet ./...` and `go doc <pkg> <Symbol>` to sanity-check rendering
+
+7. **Report**:
+   - List every file changed and every file deliberately left alone (with a one-line reason)
+   - Suggest the next step (typically `/commit-push`) — do not auto-commit unless asked
+
+## Guidelines
+
+- Read the diff before writing anything — invented API names erode trust faster than missing docs
+- One change per doc commit; keep doc updates separate from code changes when possible
+- Match the existing voice and formatting of each surface (headings, code-fence languages, table styles)
+- Prefer linking between pages over duplicating content
+
+$@

.opencode/opencode.jsonc

@@ -1,8 +0,0 @@
-{
-  "$schema": "https://opencode.ai/config.json",
-  "permission": {
-    "external_directory": {
-      "~/go/**": "deny"
-    }
-  }
-}

AUTOSCROLL_FIX_SUMMARY.md

@@ -1,80 +0,0 @@
-# Autoscroll Fix - Final Summary
-
-## Root Cause
-
-The autoscroll was failing for streaming assistant messages due to a bug in how `GotoBottom()` calculated item heights.
-
-### The Problem
-
-1. **Reasoning blocks** (`StreamingMessageItem` with `role="reasoning"`) are **never cached** because they have live duration counters that update every render
-2. The `Height()` method returns `0` when `cachedRender == ""`
-3. `GotoBottom()` was calling:
-   ```go
-   itemHeight := item.Height()  // Returns 0 for reasoning
-   if itemHeight == 0 {
-       item.Render(s.width)  // Renders but doesn't cache (reasoning)
-       itemHeight = item.Height()  // Still returns 0!
-   }
-   ```
-4. This caused incorrect scroll position calculations, especially during reasoning → assistant transitions
-
-## The Solution
-
-Changed `GotoBottom()` and `AtBottom()` to calculate height **directly from the rendered string** instead of relying on the cached height:
-
-```go
-// OLD: item.Height() which checks cached render
-itemHeight := item.Height()
-if itemHeight == 0 {
-    item.Render(s.width)
-    itemHeight = item.Height()  // Still might be 0!
-}
-
-// NEW: Calculate from rendered string directly
-rendered := item.Render(s.width)
-itemHeight := strings.Count(rendered, "\n") + 1
-```
-
-This works for **all** items regardless of whether they cache their render or not.
-
-## Files Changed
-
-### `internal/ui/scrolllist.go`
-- **`GotoBottom()`**: Calculate height from rendered string (2 loops)
-- **`AtBottom()`**: Calculate height from rendered string (1 loop)
-
-### `internal/ui/model.go`
-- **`appendStreamingChunk()`**: For existing messages, call `GotoBottom()` directly (iteratr pattern)
-- **`refreshContent()`**: Simplified to only call `SetItems()` (removed redundant `GotoBottom()`)
-- **Bash streaming handler**: Removed redundant `GotoBottom()` after `refreshContent()`
-
-## Testing Results
-
-✅ **Test prompt**: "explore this repo"
-
-**Before fix**:
-- Autoscroll stopped after reasoning block completed
-- Viewport stuck showing end of reasoning ("Thought for 203ms")
-- Assistant response streamed off-screen below
-
-**After fix**:
-- Autoscroll works throughout reasoning block
-- Autoscroll continues during reasoning → assistant transition  
-- Viewport stays at bottom showing latest assistant content
-- Final position shows end of response (build commands section)
-
-## Behavior Verified
-
-1. ✅ Streaming text auto-scrolls to bottom
-2. ✅ Works across reasoning → assistant transition
-3. ✅ Manual scroll up (PgUp) disables autoscroll
-4. ✅ Scroll to bottom (Alt+End) re-enables autoscroll
-5. ✅ Accurate positioning with no offset errors
-
-## Performance Note
-
-The fix calls `Render()` on all items during `GotoBottom()` calculations. This is acceptable because:
-- `Render()` is already optimized with caching for non-reasoning items
-- `GotoBottom()` is only called during content updates (not every frame)
-- Reasoning blocks need to render anyway for live duration updates
-- This matches iteratr's approach of ensuring items are rendered before height calculations

README.md

@@ -18,7 +18,7 @@ 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, subagent - no MCP overhead
+- **Built-in Core Tools**: bash (with interactive sudo password prompt), read, write, edit, grep, find, ls, subagent - no MCP overhead
 - **Smart @ Attachments**: Binary files auto-detected via MIME type, MCP resources via `@mcp:server:uri`
 - **MCP Integration**: Connect external MCP servers for expanded capabilities
 - **Extension System**: Write custom tools, commands, widgets, and UI modifications in Go
@@ -29,7 +29,7 @@ A powerful, extensible AI coding agent CLI with multi-provider support, built-in
 - **Session Management**: Tree-based conversation history with branching support
 - **Non-Interactive Mode**: Script-friendly positional args with JSON output
 - **ACP Server**: Run Kit as an [Agent Client Protocol](https://agentclientprotocol.com) agent over stdio
-- **Go SDK**: Embed Kit in your own applications
+- **Go SDK**: Embed Kit in your own applications with full agent lifecycle events (30+ event types) and behavior-modifying hooks
 
 ## Installation
 
@@ -126,8 +126,13 @@ model: anthropic/claude-sonnet-latest
 max-tokens: 4096
 temperature: 0.7
 stream: true
+thinking-level: off       # off, none, minimal, low, medium, high
 ```
 
+All of the above keys can also be set programmatically via the SDK
+(`kit.Options.MaxTokens`, `Options.Temperature`, `Options.ThinkingLevel`, etc.)
+without touching config files — see [SDK options](#with-options).
+
 ### Environment Variables
 
 ```bash
@@ -152,6 +157,16 @@ mcpServers:
   search:
     type: remote
     url: "https://mcp.example.com/search"
+
+  pubmed:
+    type: remote
+    url: "https://pubmed.mcp.example.com"
+    noOAuth: true  # skip OAuth for public servers that don't require auth
+
+  builds:
+    type: remote
+    url: "https://builds.mcp.example.com"
+    tasksMode: always  # async task execution — see MCP Tasks below
 ```
 
 ## CLI Reference
@@ -187,14 +202,14 @@ mcpServers:
 --no-prompt-templates    Disable prompt template loading
 
 # Generation parameters
---max-tokens             Maximum tokens in response (default: 4096)
+--max-tokens             Maximum tokens in response (default: 8192, auto-raised up to 32768 for models with larger known output limits)
 --temperature            Randomness 0.0-1.0 (default: 0.7)
 --top-p                  Nucleus sampling 0.0-1.0 (default: 0.95)
 --top-k                  Limit top K tokens (default: 40)
 --stop-sequences         Custom stop sequences (comma-separated)
 --frequency-penalty      Penalize frequent tokens 0.0-2.0 (default: 0.0)
 --presence-penalty       Penalize present tokens 0.0-2.0 (default: 0.0)
---thinking-level         Extended thinking level: off, minimal, low, medium, high (default: off)
+--thinking-level         Extended thinking level: off, none, minimal, low, medium, high (default: off)
 
 # System
 --config                 Config file path (default: ~/.kit.yml)
@@ -206,9 +221,10 @@ mcpServers:
 
 ```bash
 # Authentication (for OAuth-enabled providers)
-kit auth login [provider]    # Start OAuth flow (e.g., anthropic)
-kit auth logout [provider]   # Remove credentials for provider
-kit auth status              # Check authentication status
+kit auth login [provider]          # Start OAuth flow (e.g., anthropic)
+kit auth login [provider] --set-default  # Set provider's default model as system default
+kit auth logout [provider]         # Remove credentials for provider
+kit auth status                    # Check authentication status
 
 # Model database
 kit models [provider]        # List available models (optionally filter by provider)
@@ -290,7 +306,7 @@ kit -e examples/extensions/minimal.go
 
 ### Extension Capabilities
 
-**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
+**Lifecycle Events**: OnSessionStart, OnSessionShutdown, OnBeforeAgentStart, OnAgentStart, OnAgentEnd, OnToolCall, OnToolCallInputStart, OnToolCallInputDelta, OnToolCallInputEnd, 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
@@ -541,6 +557,20 @@ host, err := kit.New(ctx, &kit.Options{
     Streaming:    true,
     Quiet:        true,
 
+    // Generation parameters (override env/config/per-model defaults)
+    MaxTokens:        16384,             // 0 = auto-resolve (env → config → per-model → 8192 floor)
+    ThinkingLevel:    "medium",          // "off", "none", "minimal", "low", "medium", "high"
+    Temperature:      ptr(float32(0.2)), // pointer so 0.0 != unset; nil = provider default
+    TopP:             nil,                // nil = leave provider/per-model default
+    TopK:             nil,
+    FrequencyPenalty: nil,
+    PresencePenalty:  nil,
+
+    // Provider configuration (override env/config without reaching into viper)
+    ProviderAPIKey: "sk-...",                      // "" = use config / provider env var
+    ProviderURL:    "https://proxy.internal/v1",   // "" = provider default
+    TLSSkipVerify:  false,                         // only takes effect when true
+
     // Session options
     SessionPath:  "./session.jsonl",  // Open specific session
     Continue:     true,                // Resume most recent session
@@ -561,6 +591,76 @@ host, err := kit.New(ctx, &kit.Options{
 })
 ```
 
+**Generation & provider fields** (added in v0.55+) let SDK consumers configure
+Kit entirely in-code without `viper.Set()` workarounds or shipping a `.kit.yml`.
+Precedence is `Options` > `KIT_*` env vars > `.kit.yml` > per-model defaults
+(`modelSettings` / `customModels`) > provider-level defaults. Sampling params
+are pointer types so explicit `0.0` is distinguishable from "leave alone"; a
+non-zero `MaxTokens` suppresses automatic right-sizing the same way `--max-tokens`
+does on the CLI.
+
+### MCP OAuth (remote MCP servers)
+
+When a remote MCP server returns 401, Kit runs the full OAuth flow (dynamic
+client registration → PKCE → token exchange → persistence) but delegates the
+user-facing step — showing the authorization URL and receiving the callback —
+to an `MCPAuthHandler` that you pass explicitly via `Options.MCPAuthHandler`.
+If nil, OAuth is disabled and the authorization-required error surfaces to the
+caller; the SDK never auto-opens a browser or binds a localhost port.
+
+```go
+// CLI/TUI apps: opens the system browser + prints status to stderr.
+authHandler, _ := kit.NewCLIMCPAuthHandler()
+defer authHandler.Close()
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPAuthHandler: authHandler,
+})
+
+// Custom UX: reuse the SDK's port + callback server, supply your own
+// presentation via OnAuthURL (TUI modal, QR code, web redirect, etc.).
+//   h, _ := kit.NewDefaultMCPAuthHandler()
+//   h.OnAuthURL = func(server, authURL string) { myUI.Show(server, authURL) }
+//
+// Full control (web apps, daemons): implement kit.MCPAuthHandler yourself —
+// no localhost binding, no side effects.
+```
+
+Tokens are persisted to `$XDG_CONFIG_HOME/.kit/mcp_tokens.json` by default; swap
+in a custom `MCPTokenStoreFactory` for encrypted, DB-backed, or in-memory
+storage. See the [SDK options docs](/sdk/options#mcp-oauth-authorization) for
+the full matrix.
+
+### MCP Tasks (long-running tools)
+
+Kit advertises [MCP task support](https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks)
+during `initialize`, so cooperating MCP servers can respond to `tools/call`
+with a `taskId` instead of blocking the connection. Kit then polls
+`tasks/get` / `tasks/result` until the task reaches a terminal state, and
+best-effort `tasks/cancel`s on context cancellation.
+
+Defaults are safe — a server that doesn't advertise task capability runs
+synchronously, exactly as before. Opt in per server via `tasksMode` in
+`.kit.yml` (`auto` | `never` | `always`) or programmatically through the SDK:
+
+```go
+host, _ := kit.New(ctx, &kit.Options{
+    MCPTaskMode: map[string]kit.MCPTaskMode{
+        "build-server": kit.MCPTaskModeAlways,
+    },
+    MCPTaskTimeout:  15 * time.Minute,
+    MCPTaskProgress: func(p kit.MCPTaskProgress) {
+        log.Printf("%s: %s", p.TaskID, p.Status)
+    },
+})
+
+tasks, _ := host.ListMCPTasks(ctx, "build-server")
+_, _    = host.CancelMCPTask(ctx, "build-server", tasks[0].TaskID)
+```
+
+See the [configuration docs](/configuration#mcp-tasks-long-running-tools) and
+[SDK options → MCP Tasks](/sdk/options#mcp-tasks) for the full surface.
+
 ### Custom Tools
 
 Create custom tools with automatic schema generation — no external dependencies needed:
@@ -581,7 +681,28 @@ host, _ := kit.New(ctx, &kit.Options{
 })
 ```
 
-Use `kit.NewParallelTool` for tools safe to run concurrently. See the [SDK docs](/sdk/overview) for full details on struct tags, `ToolOutput` fields, and `ToolCallIDFromContext`.
+Use `kit.NewParallelTool` for tools safe to run concurrently. Binary data (images, audio, etc.) in `ToolOutput.Data` is automatically forwarded to the LLM when `MediaType` is set. See the [SDK docs](/sdk/overview) for full details on struct tags, `ToolOutput` fields, and `ToolCallIDFromContext`.
+
+#### Return Helpers
+
+| Helper | Description |
+| --- | --- |
+| `kit.TextResult(content)` | Successful text result |
+| `kit.ErrorResult(content)` | Error result (LLM sees it as a tool error) |
+| `kit.ImageResult(content, data, mediaType)` | Image result with binary data (e.g. `"image/png"`) |
+| `kit.MediaResult(content, data, mediaType)` | Non-image media result (e.g. `"audio/mpeg"`) |
+
+#### ToolOutput Fields
+
+```go
+kit.ToolOutput{
+    Content:   "result text",     // text returned to the LLM
+    IsError:   false,             // true = LLM sees this as an error
+    Data:      pngBytes,          // optional binary data (images, audio)
+    MediaType: "image/png",       // MIME type for binary Data
+    Metadata:  map[string]any{},  // opaque metadata for hooks/UI (not sent to LLM)
+}
+```
 
 ### With Callbacks
 
@@ -598,7 +719,7 @@ unsub2 := host.OnToolResult(func(e kit.ToolResultEvent) {
 })
 defer unsub2()
 
-unsub3 := host.OnStreaming(func(e kit.MessageUpdateEvent) {
+unsub3 := host.OnMessageUpdate(func(e kit.MessageUpdateEvent) {
     print(e.Chunk)
 })
 defer unsub3()

cmd/auth.go

@@ -11,6 +11,7 @@ import (
 
 	"charm.land/huh/v2"
 	"github.com/mark3labs/kit/internal/auth"
+	"github.com/mark3labs/kit/internal/ui"
 	kit "github.com/mark3labs/kit/pkg/kit"
 	"github.com/spf13/cobra"
 )
@@ -54,9 +55,13 @@ Available providers:
   - anthropic: Anthropic Claude API (OAuth)
   - openai:    OpenAI ChatGPT Plus/Pro (Codex OAuth)
 
-Example:
+Flags:
+  --set-default   Set this provider's default model as the system default
+
+Examples:
   kit auth login anthropic
-  kit auth login openai`,
+  kit auth login openai
+  kit auth login openai --set-default`,
 	Args: cobra.ExactArgs(1),
 	RunE: runAuthLogin,
 }
@@ -99,10 +104,43 @@ Example:
 	RunE: runAuthStatus,
 }
 
+var (
+	loginSetDefault bool
+)
+
+// defaultModels maps providers to their recommended default models.
+// These are used when --set-default flag is passed to auth login.
+var defaultModels = map[string]string{
+	"anthropic": "anthropic/claude-sonnet-4-5-20250929",
+	"openai":    "openai/gpt-5.4",
+}
+
+// setDefaultModelIfRequested sets the default model for the given provider
+// if the --set-default flag was provided.
+func setDefaultModelIfRequested(provider string) error {
+	if !loginSetDefault {
+		return nil
+	}
+
+	model, ok := defaultModels[provider]
+	if !ok {
+		return fmt.Errorf("no default model configured for provider: %s", provider)
+	}
+
+	if err := ui.SaveModelPreference(model); err != nil {
+		return fmt.Errorf("failed to save model preference: %w", err)
+	}
+
+	fmt.Printf("\n✓ Set default model to: %s\n", model)
+	return nil
+}
+
 func init() {
 	authCmd.AddCommand(authLoginCmd)
 	authCmd.AddCommand(authLogoutCmd)
 	authCmd.AddCommand(authStatusCmd)
+
+	authLoginCmd.Flags().BoolVar(&loginSetDefault, "set-default", false, "Set this provider's default model as the system default after login")
 }
 
 func runAuthLogin(cmd *cobra.Command, args []string) error {
@@ -288,6 +326,17 @@ func loginAnthropic() error {
 	fmt.Println("\n🎉 Your OAuth credentials will now be used for Anthropic API calls.")
 	fmt.Println("💡 You can check your authentication status with: kit auth status")
 
+	// Set default model if requested
+	if err := setDefaultModelIfRequested("anthropic"); err != nil {
+		return err
+	}
+
+	// Remind users how to set this as default if they didn't use --set-default
+	if !loginSetDefault {
+		fmt.Println("\n💡 To set Anthropic as your default model, run:")
+		fmt.Println("   kit auth login anthropic --set-default")
+	}
+
 	return nil
 }
 
@@ -454,6 +503,17 @@ func loginOpenAI() error {
 	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")
 
+	// Set default model if requested
+	if err := setDefaultModelIfRequested("openai"); err != nil {
+		return err
+	}
+
+	// Remind users how to set this as default if they didn't use --set-default
+	if !loginSetDefault {
+		fmt.Println("\n💡 To set OpenAI as your default model, run:")
+		fmt.Println("   kit auth login openai --set-default")
+	}
+
 	return nil
 }
 
@@ -504,13 +564,13 @@ func startOpenAICallbackServer(expectedState string) (*callbackServer, error) {
 		}
 
 		// Return success page
-		w.Header().Set("Content-Type", "text/html")
+		w.Header().Set("Content-Type", "text/html; charset=utf-8")
 		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>
+<h1>&#10003; Authentication Successful</h1>
 <p>You can close this window and return to the terminal.</p>
 </body>
 </html>`)

cmd/root.go

@@ -297,14 +297,14 @@ func init() {
 	flags.BoolVar(&noPromptTemplates, "no-prompt-templates", false, "disable prompt template discovery")
 
 	// Model generation parameters
-	flags.IntVar(&maxTokens, "max-tokens", 4096, "maximum number of tokens in the response")
+	flags.IntVar(&maxTokens, "max-tokens", 8192, "maximum number of output tokens per response (auto-raised up to 32768 for models with higher known output limits; see internal/models/embedded_models.json)")
 	flags.Float32Var(&temperature, "temperature", 0.7, "controls randomness in responses (0.0-1.0)")
 	flags.Float32Var(&topP, "top-p", 0.95, "controls diversity via nucleus sampling (0.0-1.0)")
 	flags.Int32Var(&topK, "top-k", 40, "controls diversity by limiting top K tokens to sample from")
 	flags.Float32Var(&frequencyPenalty, "frequency-penalty", 0.0, "penalizes tokens based on frequency of appearance (0.0-2.0)")
 	flags.Float32Var(&presencePenalty, "presence-penalty", 0.0, "penalizes tokens based on whether they have appeared (0.0-2.0)")
 	flags.StringSliceVar(&stopSequences, "stop-sequences", nil, "custom stop sequences (comma-separated)")
-	flags.StringVar(&thinkingLevel, "thinking-level", "off", "extended thinking level: off, minimal, low, medium, high")
+	flags.StringVar(&thinkingLevel, "thinking-level", "off", "extended thinking level: off, none, minimal, low, medium, high")
 
 	// Ollama-specific parameters
 	flags.Int32Var(&numGPU, "num-gpu-layers", -1, "number of model layers to offload to GPU for Ollama models (-1 for auto-detect)")

examples/extensions/subagent-monitor_test.go

@@ -130,6 +130,58 @@ func TestSubagentMonitor_MultipleSubagents(t *testing.T) {
 	time.Sleep(100 * time.Millisecond)
 }
 
+// TestSubagentMonitor_ConcurrentSubagents verifies no panics when multiple
+// subagents emit events concurrently from different goroutines.
+func TestSubagentMonitor_ConcurrentSubagents(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 5 subagents concurrently
+	done := make(chan struct{}, 5)
+	for i := range 5 {
+		go func(idx int) {
+			defer func() { done <- struct{}{} }()
+
+			callID := fmt.Sprintf("concurrent-%d", idx)
+			task := fmt.Sprintf("concurrent task %d", idx)
+
+			_, _ = harness.Emit(extensions.SubagentStartEvent{
+				ToolCallID: callID,
+				Task:       task,
+			})
+
+			// Emit many chunks rapidly
+			for j := range 20 {
+				_, _ = harness.Emit(extensions.SubagentChunkEvent{
+					ToolCallID: callID,
+					Task:       task,
+					ChunkType:  "text",
+					Content:    fmt.Sprintf("agent %d chunk %d", idx, j),
+				})
+			}
+
+			_, _ = harness.Emit(extensions.SubagentEndEvent{
+				ToolCallID: callID,
+				Task:       task,
+				Response:   "done",
+			})
+		}(i)
+	}
+
+	// Wait for all goroutines
+	for range 5 {
+		<-done
+	}
+
+	// Allow any final processing
+	time.Sleep(200 * time.Millisecond)
+}
+
 // TestSubagentMonitor_SessionShutdown verifies shutdown doesn't panic
 // even with nil ctx functions.
 func TestSubagentMonitor_SessionShutdown(t *testing.T) {

examples/sdk/basic/main.go

@@ -62,7 +62,7 @@ func main() {
 		}
 	})
 	// Subscribe to streaming chunks.
-	host3.OnStreaming(func(e kit.MessageUpdateEvent) {
+	host3.OnMessageUpdate(func(e kit.MessageUpdateEvent) {
 		fmt.Print(e.Chunk)
 	})
 

go.mod

@@ -4,25 +4,25 @@ go 1.26.2
 
 require (
 	charm.land/bubbles/v2 v2.1.0
-	charm.land/bubbletea/v2 v2.0.5
-	charm.land/fantasy v0.17.2
+	charm.land/bubbletea/v2 v2.0.6
+	charm.land/fantasy v0.23.0
 	charm.land/huh/v2 v2.0.3
 	charm.land/lipgloss/v2 v2.0.3
-	github.com/alecthomas/chroma/v2 v2.23.1
+	github.com/alecthomas/chroma/v2 v2.24.1
 	github.com/atotto/clipboard v0.1.4
 	github.com/aymanbagabas/go-udiff v0.4.1
 	github.com/charmbracelet/fang v1.0.0
 	github.com/charmbracelet/log v1.0.0
 	github.com/charmbracelet/openai-go v0.0.0-20260319145158-d0740cc34266
-	github.com/charmbracelet/ultraviolet v0.0.0-20260414011438-8c69ec811b1e
+	github.com/charmbracelet/ultraviolet v0.0.0-20260428153724-66037269d7be
 	github.com/charmbracelet/x/editor v0.2.0
 	github.com/clipperhouse/displaywidth v0.11.0
 	github.com/clipperhouse/uax29/v2 v2.7.0
-	github.com/coder/acp-go-sdk v0.6.3
-	github.com/fsnotify/fsnotify v1.9.0
+	github.com/coder/acp-go-sdk v0.12.2
+	github.com/fsnotify/fsnotify v1.10.1
 	github.com/indaco/herald v0.13.0
 	github.com/indaco/herald-md v0.3.0
-	github.com/mark3labs/mcp-go v0.48.0
+	github.com/mark3labs/mcp-go v0.51.0
 	github.com/spf13/cobra v1.10.2
 	github.com/spf13/viper v1.21.0
 	github.com/traefik/yaegi v0.16.1
@@ -35,23 +35,23 @@ require (
 	cloud.google.com/go/auth v0.20.0 // 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/azcore v1.21.1 // indirect
 	github.com/Azure/azure-sdk-for-go/sdk/internal v1.12.0 // 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 v1.32.14 // indirect
-	github.com/aws/aws-sdk-go-v2/credentials v1.19.14 // 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 v1.30.15 // indirect
-	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.19 // indirect
-	github.com/aws/aws-sdk-go-v2/service/sts v1.41.10 // indirect
-	github.com/aws/smithy-go v1.24.3 // indirect
+	github.com/aws/aws-sdk-go-v2 v1.41.7 // indirect
+	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.10 // indirect
+	github.com/aws/aws-sdk-go-v2/config v1.32.17 // indirect
+	github.com/aws/aws-sdk-go-v2/credentials v1.19.16 // indirect
+	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.23 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.23 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.23 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.24 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.9 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.23 // indirect
+	github.com/aws/aws-sdk-go-v2/service/signin v1.0.11 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sso v1.30.17 // indirect
+	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.21 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sts v1.42.1 // indirect
+	github.com/aws/smithy-go v1.25.1 // 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
@@ -59,40 +59,41 @@ require (
 	github.com/charmbracelet/harmonica v0.2.0 // indirect
 	github.com/charmbracelet/lipgloss v1.1.1-0.20250404203927-76690c660834 // indirect
 	github.com/charmbracelet/x/cellbuf v0.0.15 // indirect
-	github.com/charmbracelet/x/exp/charmtone v0.0.0-20260413165052-6921c759c913 // indirect
+	github.com/charmbracelet/x/exp/charmtone v0.0.0-20260503005035-c113ba3d2310 // indirect
 	github.com/charmbracelet/x/exp/ordered v0.1.0 // indirect
-	github.com/charmbracelet/x/exp/slice v0.0.0-20260413165052-6921c759c913 // indirect
+	github.com/charmbracelet/x/exp/slice v0.0.0-20260503005035-c113ba3d2310 // 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/dlclark/regexp2 v1.11.5 // indirect
+	github.com/dlclark/regexp2 v1.12.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/felixge/httpsnoop v1.0.4 // indirect
-	github.com/go-json-experiment/json v0.0.0-20260214004413-d219187c3433 // indirect
+	github.com/go-json-experiment/json v0.0.0-20260430182902-b6187a392ed4 // indirect
 	github.com/go-logfmt/logfmt v0.6.1 // indirect
 	github.com/go-logr/logr v1.4.3 // indirect
 	github.com/go-logr/stdr v1.2.2 // indirect
 	github.com/go-viper/mapstructure/v2 v2.5.0 // indirect
 	github.com/goccy/go-yaml v1.19.2 // indirect
 	github.com/google/go-cmp v0.7.0 // indirect
-	github.com/google/jsonschema-go v0.4.2 // indirect
+	github.com/google/jsonschema-go v0.4.3 // 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 v2.21.0 // indirect
+	github.com/googleapis/enterprise-certificate-proxy v0.3.15 // indirect
+	github.com/googleapis/gax-go/v2 v2.22.0 // indirect
 	github.com/gorilla/websocket v1.5.3 // indirect
-	github.com/kaptinlin/go-i18n v0.4.0 // indirect
-	github.com/kaptinlin/jsonpointer v0.4.17 // indirect
-	github.com/kaptinlin/jsonschema v0.7.7 // indirect
-	github.com/kaptinlin/messageformat-go v0.4.20 // indirect
+	github.com/kaptinlin/go-i18n v0.4.7 // indirect
+	github.com/kaptinlin/jsonpointer v0.4.21 // indirect
+	github.com/kaptinlin/jsonschema v0.7.13 // indirect
+	github.com/kaptinlin/messageformat-go v0.6.3 // 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/roff v0.1.0 // indirect
-	github.com/pelletier/go-toml/v2 v2.3.0 // indirect
+	github.com/pelletier/go-toml/v2 v2.3.1 // indirect
 	github.com/sagikazarmark/locafero v0.12.0 // indirect
+	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 // indirect
 	github.com/spf13/afero v1.15.0 // indirect
 	github.com/spf13/cast v1.10.0 // indirect
 	github.com/subosito/gotenv v1.6.0 // indirect
@@ -115,10 +116,10 @@ require (
 	golang.org/x/net v0.53.0 // indirect
 	golang.org/x/oauth2 v0.36.0 // indirect
 	golang.org/x/time v0.15.0 // indirect
-	google.golang.org/api v0.275.0 // indirect
-	google.golang.org/genai v1.54.0 // indirect
-	google.golang.org/genproto/googleapis/rpc v0.0.0-20260414002931-afd174a4e478 // indirect
-	google.golang.org/grpc v1.80.0 // indirect
+	google.golang.org/api v0.277.0 // indirect
+	google.golang.org/genai v1.55.0 // indirect
+	google.golang.org/genproto/googleapis/rpc v0.0.0-20260427160629-7cedc36a6bc4 // indirect
+	google.golang.org/grpc v1.81.0 // indirect
 	google.golang.org/protobuf v1.36.11 // indirect
 	gopkg.in/yaml.v2 v2.4.0 // indirect
 )
@@ -129,12 +130,12 @@ require (
 	github.com/charmbracelet/x/term v0.2.2 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/lucasb-eyer/go-colorful v1.4.0 // indirect
-	github.com/mattn/go-isatty v0.0.21 // indirect
+	github.com/mattn/go-isatty v0.0.22 // indirect
 	github.com/mattn/go-runewidth v0.0.23 // 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
+	github.com/spf13/pflag v1.0.10
 	golang.org/x/sync v0.20.0 // indirect
 	golang.org/x/sys v0.43.0 // indirect
 	golang.org/x/text v0.36.0

go.sum

@@ -1,9 +1,9 @@
 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 v2.0.5 h1:TQlLFqxo39AAHSVuOhJ5D3nH7O9Nk8JGinsfWQ4y1U4=
-charm.land/bubbletea/v2 v2.0.5/go.mod h1:dvbsYZD+MHkdIZl+Z67D212hEvB+GII2tfH8f9SnoDw=
-charm.land/fantasy v0.17.2 h1:ojTMufMxY/PVH7TzYUxht2SVkvD90iCTJfmPR6c8BR8=
-charm.land/fantasy v0.17.2/go.mod h1:V9cCIUMZB9g3Bq40aKEY8xBNzDd48EdfHp2OMS0uzWs=
+charm.land/bubbletea/v2 v2.0.6 h1:UHN/91OyuhaOFGSrBXQ/hMZD8IO1Uc4BvHlgHXL2WJo=
+charm.land/bubbletea/v2 v2.0.6/go.mod h1:MH/D8ZLlN3op37vQvijKuU29g3rqTp+aQapURFonF9g=
+charm.land/fantasy v0.23.0 h1:pocjwC5CxfEg1Bpwb0raML2d5ijo3op33Mmd6hYJyo4=
+charm.land/fantasy v0.23.0/go.mod h1:4yzSsd9XmFEVjRnF1P0LTEbLTmQX6OLnPkrHaf7iruo=
 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 v2.0.3 h1:yM2zJ4Cf5Y51b7RHIwioil4ApI/aypFXXVHSwlM6RzU=
@@ -16,8 +16,8 @@ cloud.google.com/go/auth/oauth2adapt v0.2.8 h1:keo8NaayQZ6wimpNSmW5OPc283g65QNIi
 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=
 cloud.google.com/go/compute/metadata v0.9.0/go.mod h1:E0bWwX5wTnLPedCKqk3pJmVgCBSM6qQI1yTBdEb3C10=
-github.com/Azure/azure-sdk-for-go/sdk/azcore v1.21.0 h1:fou+2+WFTib47nS+nz/ozhEBnvU96bKHy6LjRsY4E28=
-github.com/Azure/azure-sdk-for-go/sdk/azcore v1.21.0/go.mod h1:t76Ruy8AHvUAC8GfMWJMa0ElSbuIcO03NLpynfbgsPA=
+github.com/Azure/azure-sdk-for-go/sdk/azcore v1.21.1 h1:jHb/wfvRikGdxMXYV3QG/SzUOPYN9KEUUuC0Yd0/vC0=
+github.com/Azure/azure-sdk-for-go/sdk/azcore v1.21.1/go.mod h1:pzBXCYn05zvYIrwLgtK8Ap8QcjRg+0i76tMQdWN6wOk=
 github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.13.1 h1:Hk5QBxZQC1jb2Fwj6mpzme37xbCDdNTxU7O9eb5+LB4=
 github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.13.1/go.mod h1:IYus9qsFobWIc2YVwe/WPjcnyCkPKtnHAqUYeebc8z0=
 github.com/Azure/azure-sdk-for-go/sdk/internal v1.12.0 h1:fhqpLE3UEXi9lPaBRpQ6XuRW0nU7hgg4zlmZZa+a9q4=
@@ -28,42 +28,42 @@ github.com/MakeNowJust/heredoc v1.0.0 h1:cXCdzVdstXyiTqTvfqk9SDHpKNjxuom+DOlyEeQ
 github.com/MakeNowJust/heredoc v1.0.0/go.mod h1:mG5amYoWBHf8vpLOuehzbGGw0EHxpZZ6lCpQ4fNJ8LE=
 github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=
 github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=
-github.com/alecthomas/chroma/v2 v2.23.1 h1:nv2AVZdTyClGbVQkIzlDm/rnhk1E9bU9nXwmZ/Vk/iY=
-github.com/alecthomas/chroma/v2 v2.23.1/go.mod h1:NqVhfBR0lte5Ouh3DcthuUCTUpDC9cxBOfyMbMQPs3o=
+github.com/alecthomas/chroma/v2 v2.24.1 h1:m5ffpfZbIb++k8AqFEKy9uVgY12xIQtBsQlc6DfZJQM=
+github.com/alecthomas/chroma/v2 v2.24.1/go.mod h1:l+ohZ9xRXIbGe7cIW+YZgOGbvuVLjMps/FYN/CwuabI=
 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 v1.41.5 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 v1.32.14 h1:opVIRo/ZbbI8OIqSOKmpFaY7IwfFUOCCXBsUpJOwDdI=
-github.com/aws/aws-sdk-go-v2/config v1.32.14/go.mod h1:U4/V0uKxh0Tl5sxmCBZ3AecYny4UNlVmObYjKuuaiOo=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.14 h1:n+UcGWAIZHkXzYt87uMFBv/l8THYELoX6gVcUvgl6fI=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.14/go.mod h1:cJKuyWB59Mqi0jM3nFYQRmnHVQIcgoxjEMAbLkpr62w=
-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 v1.30.15 h1:lFd1+ZSEYJZYvv9d6kXzhkZu07si3f+GQ1AaYwa2LUM=
-github.com/aws/aws-sdk-go-v2/service/sso v1.30.15/go.mod h1:WSvS1NLr7JaPunCXqpJnWk1Bjo7IxzZXrZi1QQCkuqM=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.19 h1:dzztQ1YmfPrxdrOiuZRMF6fuOwWlWpD2StNLTceKpys=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.19/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 v1.24.3 h1:XgOAaUgx+HhVBoP4v8n6HCQoTRDhoMghKqw4LNHsDNg=
-github.com/aws/smithy-go v1.24.3/go.mod h1:YE2RhdIuDbA5E5bTdciG9KrW3+TiEONeUWCqxX9i1Fc=
+github.com/aws/aws-sdk-go-v2 v1.41.7 h1:DWpAJt66FmnnaRIOT/8ASTucrvuDPZASqhhLey6tLY8=
+github.com/aws/aws-sdk-go-v2 v1.41.7/go.mod h1:4LAfZOPHNVNQEckOACQx60Y8pSRjIkNZQz1w92xpMJc=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.10 h1:gx1AwW1Iyk9Z9dD9F4akX5gnN3QZwUB20GGKH/I+Rho=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.10/go.mod h1:qqY157uZoqm5OXq/amuaBJyC9hgBCBQnsaWnPe905GY=
+github.com/aws/aws-sdk-go-v2/config v1.32.17 h1:FpL4/758/diKwqbytU0prpuiu60fgXKUWCpDJtApclU=
+github.com/aws/aws-sdk-go-v2/config v1.32.17/go.mod h1:OXqUMzgXytfoF9JaKkhrOYsyh72t9G+MJH8mMRaexOE=
+github.com/aws/aws-sdk-go-v2/credentials v1.19.16 h1:r3RJBuU7X9ibt8RHbMjWE6y60QbKBiII6wSrXnapxSU=
+github.com/aws/aws-sdk-go-v2/credentials v1.19.16/go.mod h1:6cx7zqDENJDbBIIWX6P8s0h6hqHC8Avbjh9Dseo27ug=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.23 h1:UuSfcORqNSz/ey3VPRS8TcVH2Ikf0/sC+Hdj400QI6U=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.23/go.mod h1:+G/OSGiOFnSOkYloKj/9M35s74LgVAdJBSD5lsFfqKg=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.23 h1:GpT/TrnBYuE5gan2cZbTtvP+JlHsutdmlV2YfEyNde0=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.23/go.mod h1:xYWD6BS9ywC5bS3sz9Xh04whO/hzK2plt2Zkyrp4JuA=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.23 h1:bpd8vxhlQi2r1hiueOw02f/duEPTMK59Q4QMAoTTtTo=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.23/go.mod h1:15DfR2nw+CRHIk0tqNyifu3G1YdAOy68RftkhMDDwYk=
+github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.24 h1:OQqn11BtaYv1WLUowvcA30MpzIu8Ti4pcLPIIyoKZrA=
+github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.24/go.mod h1:X5ZJyfwVrWA96GzPmUCWFQaEARPR7gCrpq2E92PJwAE=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.9 h1:FLudkZLt5ci0ozzgkVo8BJGwvqNaZbTWb3UcucAateA=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.9/go.mod h1:w7wZ/s9qK7c8g4al+UyoF1Sp/Z45UwMGcqIzLWVQHWk=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.23 h1:pbrxO/kuIwgEsOPLkaHu0O+m4fNgLU8B3vxQ+72jTPw=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.23/go.mod h1:/CMNUqoj46HpS3MNRDEDIwcgEnrtZlKRaHNaHxIFpNA=
+github.com/aws/aws-sdk-go-v2/service/signin v1.0.11 h1:TdJ+HdzOBhU8+iVAOGUTU63VXopcumCOF1paFulHWZc=
+github.com/aws/aws-sdk-go-v2/service/signin v1.0.11/go.mod h1:R82ZRExE/nheo0N+T8zHPcLRTcH8MGsnR3BiVGX0TwI=
+github.com/aws/aws-sdk-go-v2/service/sso v1.30.17 h1:7byT8HUWrgoRp6sXjxtZwgOKfhss5fW6SkLBtqzgRoE=
+github.com/aws/aws-sdk-go-v2/service/sso v1.30.17/go.mod h1:xNWknVi4Ezm1vg1QsB/5EWpAJURq22uqd38U8qKvOJc=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.21 h1:+1Kl1zx6bWi4X7cKi3VYh29h8BvsCoHQEQ6ST9X8w7w=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.21/go.mod h1:4vIRDq+CJB2xFAXZ+YgGUTiEft7oAQlhIs71xcSeuVg=
+github.com/aws/aws-sdk-go-v2/service/sts v1.42.1 h1:F/M5Y9I3nwr2IEpshZgh1GeHpOItExNM9L1euNuh/fk=
+github.com/aws/aws-sdk-go-v2/service/sts v1.42.1/go.mod h1:mTNxImtovCOEEuD65mKW7DCsL+2gjEH+RPEAexAzAio=
+github.com/aws/smithy-go v1.25.1 h1:J8ERsGSU7d+aCmdQur5Txg6bVoYelvQJgtZehD12GkI=
+github.com/aws/smithy-go v1.25.1/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 v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ12Gv5o=
@@ -86,8 +86,8 @@ github.com/charmbracelet/log v1.0.0 h1:HVVVMmfOorfj3BA9i8X8UL69Hoz9lI0PYwXfJvOdR
 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 v0.0.0-20260414011438-8c69ec811b1e h1:O5hZFj55wZQWxMiRtQLa3uLKhZGZGS/j8M3OXinQlrw=
-github.com/charmbracelet/ultraviolet v0.0.0-20260414011438-8c69ec811b1e/go.mod h1:bAAz7dh/FTYfC+oiHavL4mX1tOIBZ0ZwYjSi3qE6ivM=
+github.com/charmbracelet/ultraviolet v0.0.0-20260428153724-66037269d7be h1:j7w8VP/D4lu5+/4GamMmFy8nrtadcl82/fjvDgSHwLo=
+github.com/charmbracelet/ultraviolet v0.0.0-20260428153724-66037269d7be/go.mod h1:3YdTxlnV/L0bQ3VN8WOSw8doF7LZV/xawUQ4MuAPDvo=
 github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=
 github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=
 github.com/charmbracelet/x/cellbuf v0.0.15 h1:ur3pZy0o6z/R7EylET877CBxaiE1Sp1GMxoFPAIztPI=
@@ -98,14 +98,14 @@ github.com/charmbracelet/x/editor v0.2.0 h1:7XLUKtaRaB8jN7bWU2p2UChiySyaAuIfYiIR
 github.com/charmbracelet/x/editor v0.2.0/go.mod h1:p3oQ28TSL3YPd+GKJ1fHWcp+7bVGpedHpXmo0D6t1dY=
 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 v0.0.0-20260413165052-6921c759c913 h1:6F/6bu5nBLjodsvaU5xAszTaxtHrDU5UiJarpMPZj48=
-github.com/charmbracelet/x/exp/charmtone v0.0.0-20260413165052-6921c759c913/go.mod h1:nsExn0DGyX0lh9LwLHTn2Gg+hafdzfSXnC+QmEJTZFY=
+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260503005035-c113ba3d2310 h1:rByFKh9JgQScu7oy0+TlUbC2e93woW/QNZmNXbbbw/E=
+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260503005035-c113ba3d2310/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 v0.0.0-20260413165052-6921c759c913 h1:RiZFY92Ug9iz1CenzxSSQla2Z3WflsR7bIuXq40JlpU=
-github.com/charmbracelet/x/exp/slice v0.0.0-20260413165052-6921c759c913/go.mod h1:vqEfX6xzqW1pKKZUUiFOKg0OQ7bCh54Q2vR/tserrRA=
+github.com/charmbracelet/x/exp/slice v0.0.0-20260503005035-c113ba3d2310 h1:PMjHdSo8Vpq9psUw9BoHo9JLPMkm9Hqb+Whk64n3AQQ=
+github.com/charmbracelet/x/exp/slice v0.0.0-20260503005035-c113ba3d2310/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=
@@ -124,15 +124,15 @@ github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJ
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
 github.com/cncf/xds/go v0.0.0-20260202195803-dba9d589def2 h1:aBangftG7EVZoUb69Os8IaYg++6uMOdKK83QtkkvJik=
 github.com/cncf/xds/go v0.0.0-20260202195803-dba9d589def2/go.mod h1:qwXFYgsP6T7XnJtbKlf1HP8AjxZZyzxMmc+Lq5GjlU4=
-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/coder/acp-go-sdk v0.12.2 h1:fpRJ8Z5HMSr5cZ5IywzFlFZcIxZOsto+laNVu7XelFA=
+github.com/coder/acp-go-sdk v0.12.2/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=
-github.com/dlclark/regexp2 v1.11.5/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
+github.com/dlclark/regexp2 v1.12.0 h1:0j4c5qQmnC6XOWNjP3PIXURXN2gWx76rd3KvgdPkCz8=
+github.com/dlclark/regexp2 v1.12.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/dnaeon/go-vcr v1.2.0 h1:zHCHvJYTMh1N7xnV7zf1m1GPBF9Ad0Jk/whtQ1663qI=
 github.com/dnaeon/go-vcr v1.2.0/go.mod h1:R4UdLID7HZT3taECzJs4YgbbH6PIGXB6W/sc5OLb6RQ=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
@@ -146,10 +146,10 @@ github.com/felixge/httpsnoop v1.0.4 h1:NFTV2Zj1bL4mc9sqWACXbQFVBBg2W3GPvqp8/ESS2
 github.com/felixge/httpsnoop v1.0.4/go.mod h1:m8KPJKqk1gH5J9DgRY2ASl2lWCfGKXixSwevea8zH2U=
 github.com/frankban/quicktest v1.14.6 h1:7Xjx+VpznH+oBnejlPUj8oUpdxnVs4f8XU8WnHkI4W8=
 github.com/frankban/quicktest v1.14.6/go.mod h1:4ptaffx2x8+WTWXmUCuVU6aPUX1/Mz7zb5vbUoiM6w0=
-github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=
-github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=
-github.com/go-json-experiment/json v0.0.0-20260214004413-d219187c3433 h1:vymEbVwYFP/L05h5TKQxvkXoKxNvTpjxYKdF1Nlwuao=
-github.com/go-json-experiment/json v0.0.0-20260214004413-d219187c3433/go.mod h1:tphK2c80bpPhMOI4v6bIc2xWywPfbqi1Z06+RcrMkDg=
+github.com/fsnotify/fsnotify v1.10.1 h1:b0/UzAf9yR5rhf3RPm9gf3ehBPpf0oZKIjtpKrx59Ho=
+github.com/fsnotify/fsnotify v1.10.1/go.mod h1:TLheqan6HD6GBK6PrDWyDPBaEV8LspOxvPSjC+bVfgo=
+github.com/go-json-experiment/json v0.0.0-20260430182902-b6187a392ed4 h1:2WmHkJINIjgXXYDGik8d3oJvFA3DAwPy00csDJ3vo+o=
+github.com/go-json-experiment/json v0.0.0-20260430182902-b6187a392ed4/go.mod h1:tphK2c80bpPhMOI4v6bIc2xWywPfbqi1Z06+RcrMkDg=
 github.com/go-logfmt/logfmt v0.6.1 h1:4hvbpePJKnIzH1B+8OR/JPbTx37NktoI9LE2QZBBkvE=
 github.com/go-logfmt/logfmt v0.6.1/go.mod h1:EV2pOAQoZaT1ZXZbqDl5hrymndi4SY9ED9/z6CO0XAk=
 github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=
@@ -167,16 +167,16 @@ 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/jsonschema-go v0.4.3 h1:/DBOLZTfDow7pe2GmaJNhltueGTtDKICi8V8p+DQPd0=
+github.com/google/jsonschema-go v0.4.3/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 v2.21.0 h1:h45NjjzEO3faG9Lg/cFrBh2PgegVVgzqKzuZl/wMbiI=
-github.com/googleapis/gax-go/v2 v2.21.0/go.mod h1:But/NJU6TnZsrLai/xBAQLLz+Hc7fHZJt/hsCz3Fih4=
+github.com/googleapis/enterprise-certificate-proxy v0.3.15 h1:xolVQTEXusUcAA5UgtyRLjelpFFHWlPQ4XfWGc7MBas=
+github.com/googleapis/enterprise-certificate-proxy v0.3.15/go.mod h1:vqVt9yG9480NtzREnTlmGSBmFrA+bzb0yl0TxoBQXOg=
+github.com/googleapis/gax-go/v2 v2.22.0 h1:PjIWBpgGIVKGoCXuiCoP64altEJCj3/Ei+kSU5vlZD4=
+github.com/googleapis/gax-go/v2 v2.22.0/go.mod h1:irWBbALSr0Sk3qlqb9SyJ1h68WjgeFuiOzI4Rqw5+aY=
 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=
@@ -187,14 +187,14 @@ github.com/indaco/herald v0.13.0 h1:+xVG9Fx5NpuWhwku/9IlRL6I009NnX4VUGKvlZHTRxU=
 github.com/indaco/herald v0.13.0/go.mod h1:T5g1+XLYvpjouhzAGHnAHDCKizhESkoV6+QPZ3DhgWA=
 github.com/indaco/herald-md v0.3.0 h1:hN1cKyrexPPM9PeHBsKuaWvIizSi/iYvM9yzRgtdb8M=
 github.com/indaco/herald-md v0.3.0/go.mod h1:RUHVaDSG45ymJjKyxpDwBocLXrZo93FB4OeYMsw9B9s=
-github.com/kaptinlin/go-i18n v0.4.0 h1:i7L3U2yurg+xhokITtJ0k+mjHnXqkoyz8ju5Wb7W8Oc=
-github.com/kaptinlin/go-i18n v0.4.0/go.mod h1:njA6x0+4MWGcLWT0KLrwekhRPmze1Hnstf2+VJFzwpM=
-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 v0.7.7 h1:41BlQJ9dskH0oE5DSzBUrl/w4JQYIr6N6L0B5GNyDoM=
-github.com/kaptinlin/jsonschema v0.7.7/go.mod h1:rKjWfyySHSxAD7Li2ctYkPlOu960igoKBvZ2ADRtd5Q=
-github.com/kaptinlin/messageformat-go v0.4.20 h1:a0ufTd5liiUubIGeGxpSTnNS8ZSrN4DV01/wGFmfzMs=
-github.com/kaptinlin/messageformat-go v0.4.20/go.mod h1:FqdEPfQLkqVBX7OBRMPgYwUPvKYJohFD9Ok1BMzCfIo=
+github.com/kaptinlin/go-i18n v0.4.7 h1:apjIIZHnGRyrkiX3vHj07F1BF6D0JLmV+VGSr1781Jc=
+github.com/kaptinlin/go-i18n v0.4.7/go.mod h1:+i1J0pFq/9i9ESC5qRMVkKwC+mdQTABhhBExpYOlbeM=
+github.com/kaptinlin/jsonpointer v0.4.21 h1:WVkwQbeerbHFcoXG7Yo/mlQhhZjWiTnagECEfwDXXa0=
+github.com/kaptinlin/jsonpointer v0.4.21/go.mod h1:Mo7+DX8RlQTFqS4dnYJl0izSP4ob+Rl5xO/mGDETgaU=
+github.com/kaptinlin/jsonschema v0.7.13 h1:kahVXTy/rURL0XJjyQ9WELm59wEmXi6IY0TWswQEFvU=
+github.com/kaptinlin/jsonschema v0.7.13/go.mod h1:Uh0aUBusnhXDCEXJ2oimL/hx7YTo7F+sKniE+tM0ERc=
+github.com/kaptinlin/messageformat-go v0.6.3 h1:m9ZE/fCjnsk8bdkv7Qs56L/ZoHbmQqhz9mRZSAQLU5g=
+github.com/kaptinlin/messageformat-go v0.6.3/go.mod h1:2KOZ/hgo/SveZ+uyi7vPUpUXieX65Mppzbc3VpGyqKs=
 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=
@@ -203,10 +203,10 @@ github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0
 github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
 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 v0.48.0 h1:o+MXuGW/HCeR2ny5LcAcZQn2bo6I2xaZMEHnpRG+dtw=
-github.com/mark3labs/mcp-go v0.48.0/go.mod h1:JKTC7R2LLVagkEWK7Kwu7DbmA6iIvnNAod6yrHiQMag=
-github.com/mattn/go-isatty v0.0.21 h1:xYae+lCNBP7QuW4PUnNG61ffM4hVIfm+zUzDuSzYLGs=
-github.com/mattn/go-isatty v0.0.21/go.mod h1:ZXfXG4SQHsB/w3ZeOYbR0PrPwLy+n6xiMrJlRFqopa4=
+github.com/mark3labs/mcp-go v0.51.0 h1:e8AhEfxzcYt7XqYzwT7uzWNhnqpu3H1Tn7dEJB9Ygj8=
+github.com/mark3labs/mcp-go v0.51.0/go.mod h1:Zg9cB2HdwdMMVgY0xtTzq3KvYIOJQDsaut+jWjwDaQY=
+github.com/mattn/go-isatty v0.0.22 h1:j8l17JJ9i6VGPUFUYoTUKPSgKe/83EYU2zBC7YNKMw4=
+github.com/mattn/go-isatty v0.0.22/go.mod h1:ZXfXG4SQHsB/w3ZeOYbR0PrPwLy+n6xiMrJlRFqopa4=
 github.com/mattn/go-runewidth v0.0.23 h1:7ykA0T0jkPpzSvMS5i9uoNn2Xy3R383f9HDx3RybWcw=
 github.com/mattn/go-runewidth v0.0.23/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=
 github.com/mitchellh/hashstructure/v2 v2.0.2 h1:vGKWl0YJqUNxE8d+h8f6NJLcCJrgbhC4NcD46KavDd4=
@@ -223,8 +223,8 @@ 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/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/pelletier/go-toml/v2 v2.3.1 h1:MYEvvGnQjeNkRF1qUuGolNtNExTDwct51yp7olPtrEc=
+github.com/pelletier/go-toml/v2 v2.3.1/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=
@@ -238,6 +238,8 @@ github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
 github.com/sagikazarmark/locafero v0.12.0 h1:/NQhBAkUb4+fH1jivKHWusDYFjMOOKU88eegjfxfHb4=
 github.com/sagikazarmark/locafero v0.12.0/go.mod h1:sZh36u/YSZ918v0Io+U9ogLYQJ9tLLBmM4eneO6WwsI=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2/go.mod h1:JXeL+ps8p7/KNMjDQk3TCwPpBy0wYklyWTfbkIzdIFU=
 github.com/spf13/afero v1.15.0 h1:b/YBCLWAJdFWJTN9cLhiXXcD7mzKn9Dm86dNnfyQw1I=
 github.com/spf13/afero v1.15.0/go.mod h1:NC2ByUVxtQs4b3sIUphxK0NioZnmxgyCrfzeuq8lxMg=
 github.com/spf13/cast v1.10.0 h1:h2x0u2shc1QuLHfxi+cTJvs30+ZAHOGRic8uyGTDWxY=
@@ -310,18 +312,18 @@ 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 v0.17.0 h1:VbpOemQlsSMrYmn7T2OUvQ4dqxQXU+ouZFQsZOx50z4=
 gonum.org/v1/gonum v0.17.0/go.mod h1:El3tOrEuMpv2UdMrbNlKEh9vd86bmQ6vqIcDwxEOc1E=
-google.golang.org/api v0.275.0 h1:vfY5d9vFVJeWEZT65QDd9hbndr7FyZ2+6mIzGAh71NI=
-google.golang.org/api v0.275.0/go.mod h1:Fnag/EWUPIcJXuIkP1pjoTgS5vdxlk3eeemL7Do6bvw=
-google.golang.org/genai v1.54.0 h1:ZQCa70WMTJDI11FdqWCzGvZ5PanpcpfoO6jl/lrSnGU=
-google.golang.org/genai v1.54.0/go.mod h1:A3kkl0nyBjyFlNjgxIwKq70julKbIxpSxqKO5gw/gmk=
-google.golang.org/genproto v0.0.0-20260406210006-6f92a3bedf2d h1:N1Ec54vZnIPd7MnxRiYLW+oY4fDR4BOS/LrssdD9+ek=
-google.golang.org/genproto v0.0.0-20260406210006-6f92a3bedf2d/go.mod h1:c2hJ1grtnH0xUiEKGDGkjGNTJ1Hy2LrblyKOHF0sqRM=
-google.golang.org/genproto/googleapis/api v0.0.0-20260406210006-6f92a3bedf2d h1:/aDRtSZJjyLQzm75d+a1wOJaqyKBMvIAfeQmoa3ORiI=
-google.golang.org/genproto/googleapis/api v0.0.0-20260406210006-6f92a3bedf2d/go.mod h1:etfGUgejTiadZAUaEP14NP97xi1RGeawqkjDARA/UOs=
-google.golang.org/genproto/googleapis/rpc v0.0.0-20260414002931-afd174a4e478 h1:RmoJA1ujG+/lRGNfUnOMfhCy5EipVMyvUE+KNbPbTlw=
-google.golang.org/genproto/googleapis/rpc v0.0.0-20260414002931-afd174a4e478/go.mod h1:4Hqkh8ycfw05ld/3BWL7rJOSfebL2Q+DVDeRgYgxUU8=
-google.golang.org/grpc v1.80.0 h1:Xr6m2WmWZLETvUNvIUmeD5OAagMw3FiKmMlTdViWsHM=
-google.golang.org/grpc v1.80.0/go.mod h1:ho/dLnxwi3EDJA4Zghp7k2Ec1+c2jqup0bFkw07bwF4=
+google.golang.org/api v0.277.0 h1:HJfyJUiNeBBUMai7ez8u14wkp/gH/I4wpGbbO9o+cSk=
+google.golang.org/api v0.277.0/go.mod h1:B9TqLBwJqVjp1mtt7WeoQwWRwvu/400y5lETOql+giQ=
+google.golang.org/genai v1.55.0 h1:iLHGk4Bj/IZ/GNNZb7hYqwSJMRBvqLeu2Hb6YQ+rYGw=
+google.golang.org/genai v1.55.0/go.mod h1:A3kkl0nyBjyFlNjgxIwKq70julKbIxpSxqKO5gw/gmk=
+google.golang.org/genproto v0.0.0-20260427160629-7cedc36a6bc4 h1:2iMJZntwvmfgtse+s744JY7v7PgEdSBuFYXucvpOHNM=
+google.golang.org/genproto v0.0.0-20260427160629-7cedc36a6bc4/go.mod h1:v14kaaboYyXQ1Gsu489Q+Hg/oN4B33mWtuOhF1HCeXA=
+google.golang.org/genproto/googleapis/api v0.0.0-20260427160629-7cedc36a6bc4 h1:yOzSCGPx+cp5VO7IxvZ9SBFF7j1tZVcNtlHR2iYKtVo=
+google.golang.org/genproto/googleapis/api v0.0.0-20260427160629-7cedc36a6bc4/go.mod h1:Q9HWtNeE7tM9npdIsEvqXj1QJIvVoeAV3rtXtS715Cw=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260427160629-7cedc36a6bc4 h1:tEkOQcXgF6dH1G+MVKZrfpYvozGrzb91k6ha7jireSM=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260427160629-7cedc36a6bc4/go.mod h1:4Hqkh8ycfw05ld/3BWL7rJOSfebL2Q+DVDeRgYgxUU8=
+google.golang.org/grpc v1.81.0 h1:W3G9N3KQf3BU+YuCtGKJk0CmxQNbAISICD/9AORxLIw=
+google.golang.org/grpc v1.81.0/go.mod h1:xGH9GfzOyMTGIOXBJmXt+BX/V0kcdQbdcuwQ/zNw42I=
 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/agent.go

@@ -177,22 +177,75 @@ func (a *Agent) SetSessionMode(_ context.Context, _ acp.SetSessionModeRequest) (
 	return acp.SetSessionModeResponse{}, nil
 }
 
-// SetSessionModel changes the active model for a session.
-func (a *Agent) SetSessionModel(ctx context.Context, params acp.SetSessionModelRequest) (acp.SetSessionModelResponse, error) {
+// ListSessions returns an empty session list. Kit doesn't persist sessions
+// across restarts in ACP mode, so this is effectively a no-op.
+func (a *Agent) ListSessions(_ context.Context, _ acp.ListSessionsRequest) (acp.ListSessionsResponse, error) {
+	return acp.ListSessionsResponse{
+		Sessions: []acp.SessionInfo{},
+	}, nil
+}
+
+// CloseSession cancels any ongoing work for the session and frees its resources.
+func (a *Agent) CloseSession(_ context.Context, params acp.CloseSessionRequest) (acp.CloseSessionResponse, error) {
 	sessionID := string(params.SessionId)
 	sess, ok := a.registry.get(sessionID)
 	if !ok {
-		return acp.SetSessionModelResponse{}, acp.NewInvalidParams(fmt.Sprintf("session not found: %s", sessionID))
+		return acp.CloseSessionResponse{}, nil
 	}
 
-	modelID := string(params.ModelId)
-	log.Debug("acp: set_session_model", "session", sessionID, "model", modelID)
+	log.Debug("acp: close session", "session", sessionID)
+	sess.cancelPrompt()
+	a.registry.remove(sessionID)
+	return acp.CloseSessionResponse{}, nil
+}
 
-	if err := sess.kit.SetModel(ctx, modelID); err != nil {
-		return acp.SetSessionModelResponse{}, fmt.Errorf("set model: %w", err)
+// ResumeSession is not supported — Kit doesn't persist sessions across
+// restarts in ACP mode. Clients should use NewSession instead.
+func (a *Agent) ResumeSession(_ context.Context, _ acp.ResumeSessionRequest) (acp.ResumeSessionResponse, error) {
+	return acp.ResumeSessionResponse{}, fmt.Errorf("resume session not supported")
+}
+
+// SetSessionConfigOption handles session configuration changes. Currently
+// supports the "model" config option to change the active model for a session.
+func (a *Agent) SetSessionConfigOption(ctx context.Context, params acp.SetSessionConfigOptionRequest) (acp.SetSessionConfigOptionResponse, error) {
+	// Extract session ID and config ID from whichever variant is present.
+	var sessionID string
+	var configID string
+	var value string
+
+	switch {
+	case params.ValueId != nil:
+		sessionID = string(params.ValueId.SessionId)
+		configID = string(params.ValueId.ConfigId)
+		value = string(params.ValueId.Value)
+	case params.Boolean != nil:
+		sessionID = string(params.Boolean.SessionId)
+		configID = string(params.Boolean.ConfigId)
+		// Boolean config options are not used for model selection.
+		log.Debug("acp: set_session_config_option (boolean)", "session", sessionID, "config", configID, "value", params.Boolean.Value)
+		return acp.SetSessionConfigOptionResponse{}, nil
+	default:
+		return acp.SetSessionConfigOptionResponse{}, acp.NewInvalidParams("unsupported config option variant")
 	}
 
-	return acp.SetSessionModelResponse{}, nil
+	sess, ok := a.registry.get(sessionID)
+	if !ok {
+		return acp.SetSessionConfigOptionResponse{}, acp.NewInvalidParams(fmt.Sprintf("session not found: %s", sessionID))
+	}
+
+	log.Debug("acp: set_session_config_option", "session", sessionID, "config", configID, "value", value)
+
+	// Handle known config options.
+	switch configID {
+	case "model":
+		if err := sess.kit.SetModel(ctx, value); err != nil {
+			return acp.SetSessionConfigOptionResponse{}, fmt.Errorf("set model: %w", err)
+		}
+	default:
+		log.Debug("acp: unknown config option", "config", configID)
+	}
+
+	return acp.SetSessionConfigOptionResponse{}, nil
 }
 
 // ---------------------------------------------------------------------------

internal/acpserver/session.go

@@ -232,6 +232,20 @@ func (r *sessionRegistry) closeAll() {
 	}
 }
 
+// remove closes and removes a single session by ID.
+func (r *sessionRegistry) remove(sessionID string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	sess, ok := r.sessions[sessionID]
+	if !ok {
+		return
+	}
+	if sess.kit != nil {
+		_ = sess.kit.Close()
+	}
+	delete(r.sessions, sessionID)
+}
+
 // cancelPrompt cancels the current prompt for a session, if any.
 func (s *acpSession) cancelPrompt() {
 	s.cancelMu.Lock()

internal/agent/agent.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"strings"
+	"time"
 
 	"charm.land/fantasy"
 
@@ -58,6 +59,11 @@ type AgentConfig struct {
 	// loading (successfully or with error). The callback receives the server
 	// name, tool count, and any error. Called from the background goroutine.
 	OnMCPServerLoaded func(serverName string, toolCount int, err error)
+
+	// MCPTaskConfig configures task-augmented tools/call execution. The
+	// zero value preserves historical synchronous-only behaviour for any
+	// server that didn't advertise task support during initialize.
+	MCPTaskConfig tools.MCPTaskConfig
 }
 
 // ToolCallHandler is a function type for handling tool calls as they happen.
@@ -87,6 +93,19 @@ type ReasoningDeltaHandler func(delta string)
 // Called when the last reasoning token has been processed, before text streaming starts.
 type ReasoningCompleteHandler func()
 
+// ToolCallStartHandler is a function type for handling the moment when the LLM
+// begins generating tool call arguments. The tool name is known but the full
+// argument JSON is still streaming.
+type ToolCallStartHandler func(toolCallID, toolName string)
+
+// ToolCallDeltaHandler is a function type for handling streamed fragments of
+// tool call arguments as they arrive from the LLM.
+type ToolCallDeltaHandler func(toolCallID, delta string)
+
+// ToolCallEndHandler is a function type for handling the end of tool argument
+// streaming, before the tool call is parsed and execution begins.
+type ToolCallEndHandler func(toolCallID string)
+
 // 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
@@ -113,6 +132,76 @@ type StepMessagesHandler func(stepMessages []fantasy.Message)
 // tracking during long-running tool-calling conversations.
 type StepUsageHandler func(inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens int64)
 
+// StepStartHandler is called when a new LLM step begins within a turn.
+type StepStartHandler func(stepNumber int)
+
+// StepFinishHandler is called when a step completes with full context.
+type StepFinishHandler func(stepNumber int, hasToolCalls bool, finishReason string, usage fantasy.Usage)
+
+// TextStartHandler is called when the LLM begins generating text content.
+type TextStartHandler func(id string)
+
+// TextEndHandler is called when the LLM finishes generating text content.
+type TextEndHandler func(id string)
+
+// ReasoningStartHandler is called when the LLM begins reasoning/thinking.
+type ReasoningStartHandler func(id string)
+
+// WarningsHandler is called when the LLM provider returns warnings.
+type WarningsHandler func(warnings []string)
+
+// SourceHandler is called when the LLM references a source.
+type SourceHandler func(sourceType, id, url, title string)
+
+// StreamFinishHandler is called when a per-step LLM stream completes.
+type StreamFinishHandler func(usage fantasy.Usage, finishReason string)
+
+// ErrorHandler is called when an agent-level error occurs.
+type ErrorHandler func(err error)
+
+// RetryHandler is called when the LLM request is retried.
+type RetryHandler func(attempt int, err error)
+
+// PrepareStepHandler is called between steps to allow message modification.
+// It receives the step number and current messages, and returns replacement
+// messages (or nil to keep unchanged).
+type PrepareStepHandler func(stepNumber int, messages []fantasy.Message) []fantasy.Message
+
+// GenerateCallbacks consolidates all callback functions for
+// GenerateWithLoopAndStreaming into a single struct. This replaces the previous
+// 16+ positional callback parameters, making it easier to add new callbacks
+// without breaking existing callers (new fields default to nil).
+type GenerateCallbacks struct {
+	OnToolCall          ToolCallHandler
+	OnToolExecution     ToolExecutionHandler
+	OnToolResult        ToolResultHandler
+	OnResponse          ResponseHandler
+	OnToolCallContent   ToolCallContentHandler
+	OnStreamingResponse StreamingResponseHandler
+	OnReasoningDelta    ReasoningDeltaHandler
+	OnReasoningComplete ReasoningCompleteHandler
+	OnToolOutput        ToolOutputHandler
+	OnStepMessages      StepMessagesHandler
+	OnStepUsage         StepUsageHandler
+	OnPasswordPrompt    PasswordPromptHandler
+	OnToolCallStart     ToolCallStartHandler
+	OnToolCallDelta     ToolCallDeltaHandler
+	OnToolCallEnd       ToolCallEndHandler
+
+	// New callbacks for previously unwired Fantasy lifecycle events.
+	OnStepStart      StepStartHandler
+	OnStepFinish     StepFinishHandler
+	OnTextStart      TextStartHandler
+	OnTextEnd        TextEndHandler
+	OnReasoningStart ReasoningStartHandler
+	OnWarnings       WarningsHandler
+	OnSource         SourceHandler
+	OnStreamFinish   StreamFinishHandler
+	OnError          ErrorHandler
+	OnRetry          RetryHandler
+	OnPrepareStep    PrepareStepHandler
+}
+
 // 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
 // AgentTool implementations — no MCP layer, no serialization overhead.
@@ -147,6 +236,10 @@ type Agent struct {
 	authHandler       tools.MCPAuthHandler
 	tokenStoreFactory tools.TokenStoreFactory
 
+	// mcpTaskConfig is stored from AgentConfig so AddMCPServer() can
+	// propagate it to a lazily-created MCPToolManager.
+	mcpTaskConfig tools.MCPTaskConfig
+
 	// mcpReady is closed when background MCP tool loading completes (success
 	// or failure). nil when no MCP servers are configured.
 	mcpReady chan struct{}
@@ -245,6 +338,7 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 		modelConfig:         agentConfig.ModelConfig,
 		authHandler:         agentConfig.AuthHandler,
 		tokenStoreFactory:   agentConfig.TokenStoreFactory,
+		mcpTaskConfig:       agentConfig.MCPTaskConfig,
 	}
 
 	// Start MCP tool loading in the background if servers are configured.
@@ -264,6 +358,8 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 		if agentConfig.OnMCPServerLoaded != nil {
 			toolManager.SetOnServerLoaded(agentConfig.OnMCPServerLoaded)
 		}
+		// Apply task-augmented tool execution config (zero value = no-op).
+		toolManager.SetTaskConfig(agentConfig.MCPTaskConfig)
 		a.toolManager = toolManager
 		a.mcpReady = make(chan struct{})
 
@@ -410,13 +506,20 @@ func (a *Agent) GenerateWithLoop(ctx context.Context, messages []fantasy.Message
 	onToolCall ToolCallHandler, onToolExecution ToolExecutionHandler, onToolResult ToolResultHandler,
 	onResponse ResponseHandler, onToolCallContent ToolCallContentHandler,
 ) (*GenerateWithLoopResult, error) {
-	return a.GenerateWithLoopAndStreaming(ctx, messages, onToolCall, onToolExecution, onToolResult,
-		onResponse, onToolCallContent, nil, nil, nil, nil, nil, nil, nil)
+	return a.GenerateWithCallbacks(ctx, messages, GenerateCallbacks{
+		OnToolCall:        onToolCall,
+		OnToolExecution:   onToolExecution,
+		OnToolResult:      onToolResult,
+		OnResponse:        onResponse,
+		OnToolCallContent: onToolCallContent,
+	})
 }
 
 // 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.
+// The agent handles the tool call loop internally.
+//
+// Deprecated: Use GenerateWithCallbacks instead, which takes a GenerateCallbacks
+// struct and is easier to extend with new callbacks.
 func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fantasy.Message,
 	onToolCall ToolCallHandler, onToolExecution ToolExecutionHandler, onToolResult ToolResultHandler,
 	onResponse ResponseHandler, onToolCallContent ToolCallContentHandler,
@@ -427,6 +530,34 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 	onStepMessages StepMessagesHandler,
 	onStepUsage StepUsageHandler,
 	onPasswordPrompt PasswordPromptHandler,
+	onToolCallStart ToolCallStartHandler,
+	onToolCallDelta ToolCallDeltaHandler,
+	onToolCallEnd ToolCallEndHandler,
+) (*GenerateWithLoopResult, error) {
+	return a.GenerateWithCallbacks(ctx, messages, GenerateCallbacks{
+		OnToolCall:          onToolCall,
+		OnToolExecution:     onToolExecution,
+		OnToolResult:        onToolResult,
+		OnResponse:          onResponse,
+		OnToolCallContent:   onToolCallContent,
+		OnStreamingResponse: onStreamingResponse,
+		OnReasoningDelta:    onReasoningDelta,
+		OnReasoningComplete: onReasoningComplete,
+		OnToolOutput:        onToolOutput,
+		OnStepMessages:      onStepMessages,
+		OnStepUsage:         onStepUsage,
+		OnPasswordPrompt:    onPasswordPrompt,
+		OnToolCallStart:     onToolCallStart,
+		OnToolCallDelta:     onToolCallDelta,
+		OnToolCallEnd:       onToolCallEnd,
+	})
+}
+
+// GenerateWithCallbacks 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) GenerateWithCallbacks(ctx context.Context, messages []fantasy.Message,
+	cb GenerateCallbacks,
 ) (*GenerateWithLoopResult, error) {
 
 	// Wait for background MCP tool loading to complete and rebuild the
@@ -435,13 +566,13 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 	a.ensureMCPTools()
 
 	// Inject tool output handler into context for use by core tools (e.g., bash).
-	if onToolOutput != nil {
-		ctx = core.ContextWithToolOutputCallback(ctx, onToolOutput)
+	if cb.OnToolOutput != nil {
+		ctx = core.ContextWithToolOutputCallback(ctx, cb.OnToolOutput)
 	}
 
 	// Inject password prompt handler into context for use by bash tool.
-	if onPasswordPrompt != nil {
-		ctx = core.ContextWithPasswordPrompt(ctx, onPasswordPrompt)
+	if cb.OnPasswordPrompt != nil {
+		ctx = core.ContextWithPasswordPrompt(ctx, cb.OnPasswordPrompt)
 	}
 
 	// The agent requires the current user input as Prompt, with prior messages as history.
@@ -461,8 +592,13 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 	// 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 ||
-		onToolCallContent != nil || onStreamingResponse != nil || onReasoningDelta != nil
+	hasCallbacks := cb.OnToolCall != nil || cb.OnToolExecution != nil || cb.OnToolResult != nil ||
+		cb.OnToolCallContent != nil || cb.OnStreamingResponse != nil || cb.OnReasoningDelta != nil ||
+		cb.OnToolCallStart != nil || cb.OnToolCallDelta != nil || cb.OnToolCallEnd != nil ||
+		cb.OnStepStart != nil || cb.OnStepFinish != nil || cb.OnTextStart != nil ||
+		cb.OnTextEnd != nil || cb.OnReasoningStart != nil || cb.OnWarnings != nil ||
+		cb.OnSource != nil || cb.OnStreamFinish != nil || cb.OnError != nil ||
+		cb.OnRetry != nil || cb.OnPrepareStep != nil
 
 	if a.streamingEnabled || hasCallbacks {
 		// Track completed step messages so we can return partial results
@@ -471,9 +607,11 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 		// for every step that completed before the error occurred.
 		var completedStepMessages []fantasy.Message
 		// persistedCount tracks how many new messages (beyond the original
-		// input) were persisted incrementally via onStepMessages, so the
+		// input) were persisted incrementally via cb.OnStepMessages, so the
 		// caller can skip them during post-generation persistence.
 		var persistedCount int
+		// stepCounter tracks the current step number for StepStart/StepFinish events.
+		var stepCounter int
 
 		// Use the streaming agent
 		streamCall := fantasy.AgentStreamCall{
@@ -481,13 +619,73 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 			Files:    files,
 			Messages: history,
 
+			// Tool input streaming callbacks — fire during tool argument generation
+			OnToolInputStart: func(id, toolName string) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnToolCallStart != nil {
+					cb.OnToolCallStart(id, toolName)
+				}
+				return nil
+			},
+			OnToolInputDelta: func(id, delta string) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnToolCallDelta != nil {
+					cb.OnToolCallDelta(id, delta)
+				}
+				return nil
+			},
+			OnToolInputEnd: func(id string) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnToolCallEnd != nil {
+					cb.OnToolCallEnd(id)
+				}
+				return nil
+			},
+
+			// Text start/end callbacks
+			OnTextStart: func(id string) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnTextStart != nil {
+					cb.OnTextStart(id)
+				}
+				return nil
+			},
+			OnTextEnd: func(id string) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnTextEnd != nil {
+					cb.OnTextEnd(id)
+				}
+				return nil
+			},
+
+			// Reasoning start callback
+			OnReasoningStart: func(id string, _ fantasy.ReasoningContent) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnReasoningStart != nil {
+					cb.OnReasoningStart(id)
+				}
+				return nil
+			},
+
 			// Reasoning/thinking streaming callback
 			OnReasoningDelta: func(id, delta string) error {
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onReasoningDelta != nil {
-					onReasoningDelta(delta)
+				if cb.OnReasoningDelta != nil {
+					cb.OnReasoningDelta(delta)
 				}
 				return nil
 			},
@@ -497,8 +695,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onReasoningComplete != nil {
-					onReasoningComplete()
+				if cb.OnReasoningComplete != nil {
+					cb.OnReasoningComplete()
 				}
 				return nil
 			},
@@ -508,8 +706,64 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onStreamingResponse != nil {
-					onStreamingResponse(text)
+				if cb.OnStreamingResponse != nil {
+					cb.OnStreamingResponse(text)
+				}
+				return nil
+			},
+
+			// Warnings callback
+			OnWarnings: func(warnings []fantasy.CallWarning) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnWarnings != nil {
+					strs := make([]string, len(warnings))
+					for i, w := range warnings {
+						strs[i] = w.Message
+					}
+					cb.OnWarnings(strs)
+				}
+				return nil
+			},
+
+			// Source callback
+			OnSource: func(source fantasy.SourceContent) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnSource != nil {
+					cb.OnSource(string(source.SourceType), source.ID, source.URL, source.Title)
+				}
+				return nil
+			},
+
+			// Stream finish callback (per-step stream completion)
+			OnStreamFinish: func(usage fantasy.Usage, finishReason fantasy.FinishReason, _ fantasy.ProviderMetadata) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				if cb.OnStreamFinish != nil {
+					cb.OnStreamFinish(usage, string(finishReason))
+				}
+				return nil
+			},
+
+			// Error callback
+			OnError: func(err error) {
+				if cb.OnError != nil {
+					cb.OnError(err)
+				}
+			},
+
+			// Step start callback
+			OnStepStart: func(stepNumber int) error {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				stepCounter = stepNumber
+				if cb.OnStepStart != nil {
+					cb.OnStepStart(stepNumber)
 				}
 				return nil
 			},
@@ -522,13 +776,13 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				currentToolArgs = tc.Input
 
 				// Notify about the tool call
-				if onToolCall != nil {
-					onToolCall(tc.ToolCallID, tc.ToolName, tc.Input)
+				if cb.OnToolCall != nil {
+					cb.OnToolCall(tc.ToolCallID, tc.ToolName, tc.Input)
 				}
 
 				// Notify tool execution starting
-				if onToolExecution != nil {
-					onToolExecution(tc.ToolCallID, tc.ToolName, tc.Input, true)
+				if cb.OnToolExecution != nil {
+					cb.OnToolExecution(tc.ToolCallID, tc.ToolName, tc.Input, true)
 				}
 
 				return nil
@@ -540,14 +794,14 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 					return ctx.Err()
 				}
 				// Notify tool execution finished
-				if onToolExecution != nil {
-					onToolExecution(tr.ToolCallID, tr.ToolName, currentToolArgs, false)
+				if cb.OnToolExecution != nil {
+					cb.OnToolExecution(tr.ToolCallID, tr.ToolName, currentToolArgs, false)
 				}
 
-				if onToolResult != nil {
+				if cb.OnToolResult != nil {
 					// Extract result text and error status
 					resultText, isError := extractToolResultText(tr)
-					onToolResult(tr.ToolCallID, tr.ToolName, currentToolArgs, resultText, tr.ClientMetadata, isError)
+					cb.OnToolResult(tr.ToolCallID, tr.ToolName, currentToolArgs, resultText, tr.ClientMetadata, isError)
 				}
 
 				return nil
@@ -561,8 +815,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 
 				// Persist step messages incrementally so progress is saved
 				// as it happens rather than only at the end of the turn.
-				if onStepMessages != nil && len(step.Messages) > 0 {
-					onStepMessages(step.Messages)
+				if cb.OnStepMessages != nil && len(step.Messages) > 0 {
+					cb.OnStepMessages(step.Messages)
 					persistedCount += len(step.Messages)
 				}
 
@@ -572,65 +826,88 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				// Check if step has text content alongside tool calls
 				text := step.Content.Text()
 				toolCalls := step.Content.ToolCalls()
-				if text != "" && len(toolCalls) > 0 && onToolCallContent != nil {
-					onToolCallContent(text)
+				if text != "" && len(toolCalls) > 0 && cb.OnToolCallContent != nil {
+					cb.OnToolCallContent(text)
 				}
 				// Emit step usage for real-time cost tracking
-				if onStepUsage != nil {
-					onStepUsage(step.Usage.InputTokens, step.Usage.OutputTokens,
+				if cb.OnStepUsage != nil {
+					cb.OnStepUsage(step.Usage.InputTokens, step.Usage.OutputTokens,
 						step.Usage.CacheReadTokens, step.Usage.CacheCreationTokens)
 				}
+				// Emit unified step finish event
+				if cb.OnStepFinish != nil {
+					cb.OnStepFinish(stepCounter, len(toolCalls) > 0, string(step.FinishReason), step.Usage)
+				}
 				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)
+		// Always wire up PrepareStep to handle both steering and the
+		// OnPrepareStep hook. Steering drains its channel first, then
+		// OnPrepareStep hooks run against the (possibly already steered)
+		// messages.
+		steerCh := steerChFromContext(ctx)
+		onConsumed := steerConsumedFromContext(ctx)
+		hasSteering := steerCh != nil
+		hasPrepareStepHook := cb.OnPrepareStep != nil
+
+		if hasSteering || hasPrepareStepHook {
 			streamCall.PrepareStep = func(
 				stepCtx context.Context,
 				opts fantasy.PrepareStepFunctionOptions,
 			) (context.Context, fantasy.PrepareStepResult, error) {
-				// Drain all pending steer messages (non-blocking).
-				var steered []SteerMessage
-				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 _, sm := range steered {
-						result.Messages = append(result.Messages,
-							fantasy.NewUserMessage(sm.Text, sm.Files...))
+
+				// Phase 1: Drain steering channel (if present).
+				if hasSteering {
+					var steered []SteerMessage
+					for {
+						select {
+						case msg := <-steerCh:
+							steered = append(steered, msg)
+						default:
+							goto done
+						}
 					}
-					// Notify that steer messages were consumed.
-					if onConsumed != nil {
-						onConsumed(len(steered))
+				done:
+					if len(steered) > 0 {
+						for _, sm := range steered {
+							result.Messages = append(result.Messages,
+								fantasy.NewUserMessage(sm.Text, sm.Files...))
+						}
+						if onConsumed != nil {
+							onConsumed(len(steered))
+						}
+					}
+				}
+
+				// Phase 2: Run OnPrepareStep hook (if registered).
+				if hasPrepareStepHook {
+					if replacement := cb.OnPrepareStep(opts.StepNumber, result.Messages); replacement != nil {
+						result.Messages = replacement
 					}
 				}
 
 				// 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
 			}
 		}
 
+		// Wire OnRetry callback if provided.
+		if cb.OnRetry != nil {
+			streamCall.OnRetry = func(err *fantasy.ProviderError, _ time.Duration) {
+				// Use the retry number from the error if available; Fantasy
+				// doesn't pass a counter directly, so we approximate with a
+				// counter incremented on each call.
+				cb.OnRetry(0, err)
+			}
+		}
+
 		result, err := a.fantasyAgent.Stream(ctx, streamCall)
 		if err != nil {
 			// On cancellation (or any error), return a partial result
@@ -656,8 +933,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 		// empty (e.g. reasoning-only responses) so the UI properly resets
 		// the stream component and avoids duplicate content on the next
 		// flush.
-		if onResponse != nil {
-			onResponse(result.Response.Content.Text())
+		if cb.OnResponse != nil {
+			cb.OnResponse(result.Response.Content.Text())
 		}
 
 		r := convertAgentResult(result, messages)
@@ -677,8 +954,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 
 	// For non-streaming, fire the response callback so callers can reset
 	// streaming state (see streaming path comment above).
-	if onResponse != nil {
-		onResponse(result.Response.Content.Text())
+	if cb.OnResponse != nil {
+		cb.OnResponse(result.Response.Content.Text())
 	}
 
 	return convertAgentResult(result, messages), nil
@@ -869,6 +1146,7 @@ func (a *Agent) AddMCPServer(ctx context.Context, name string, cfg config.MCPSer
 		if a.tokenStoreFactory != nil {
 			a.toolManager.SetTokenStoreFactory(a.tokenStoreFactory)
 		}
+		a.toolManager.SetTaskConfig(a.mcpTaskConfig)
 		a.toolManager.SetOnToolsChanged(func() {
 			a.rebuildFantasyAgent()
 		})
@@ -1025,6 +1303,22 @@ func (a *Agent) GetModel() fantasy.LanguageModel {
 	return a.model
 }
 
+// GetMaxTokens returns the effective max output tokens the agent currently
+// sends to the LLM provider, after per-model defaults, right-sizing, and any
+// Anthropic thinking-budget adjustments. Returns 0 when no ModelConfig is
+// attached (e.g. early init) or when the provider suppresses the parameter
+// (e.g. Codex OAuth), which allows callers to differentiate "default" from
+// "explicitly capped".
+func (a *Agent) GetMaxTokens() int {
+	if a.skipMaxOutputTokens {
+		return 0
+	}
+	if a.modelConfig == nil {
+		return 0
+	}
+	return a.modelConfig.MaxTokens
+}
+
 // Close closes the agent and cleans up resources.
 // If MCP tools are still loading in the background, Close waits for them
 // to finish before closing connections to avoid resource leaks.

internal/agent/factory.go

@@ -56,6 +56,8 @@ type AgentCreationOptions struct {
 	// OnMCPServerLoaded, if non-nil, is called when each MCP server finishes
 	// loading (successfully or with error). Called from the background goroutine.
 	OnMCPServerLoaded func(serverName string, toolCount int, err error)
+	// MCPTaskConfig configures task-augmented tools/call execution.
+	MCPTaskConfig tools.MCPTaskConfig
 }
 
 // CreateAgent creates an agent with optional spinner for Ollama models.
@@ -76,6 +78,7 @@ func CreateAgent(ctx context.Context, opts *AgentCreationOptions) (*Agent, error
 		ToolWrapper:       opts.ToolWrapper,
 		ExtraTools:        opts.ExtraTools,
 		OnMCPServerLoaded: opts.OnMCPServerLoaded,
+		MCPTaskConfig:     opts.MCPTaskConfig,
 	}
 
 	var agent *Agent

internal/app/app.go

@@ -888,6 +888,12 @@ func (a *App) subscribeSDKEvents(sendFn func(tea.Msg), stepUsageSeen *atomic.Boo
 		switch ev := e.(type) {
 		case kit.ToolCallEvent:
 			sendFn(ToolCallStartedEvent{ToolCallID: ev.ToolCallID, ToolName: ev.ToolName, ToolArgs: ev.ToolArgs})
+		case kit.ToolCallStartEvent:
+			sendFn(ToolCallInputStartEvent{ToolCallID: ev.ToolCallID, ToolName: ev.ToolName, ToolKind: ev.ToolKind})
+		case kit.ToolCallDeltaEvent:
+			sendFn(ToolCallInputDeltaEvent{ToolCallID: ev.ToolCallID, Delta: ev.Delta})
+		case kit.ToolCallEndEvent:
+			sendFn(ToolCallInputEndEvent{ToolCallID: ev.ToolCallID})
 		case kit.ToolExecutionStartEvent:
 			sendFn(ToolExecutionEvent{ToolCallID: ev.ToolCallID, ToolName: ev.ToolName, ToolArgs: ev.ToolArgs, IsStarting: true})
 		case kit.ToolExecutionEndEvent:
@@ -917,7 +923,7 @@ func (a *App) subscribeSDKEvents(sendFn func(tea.Msg), stepUsageSeen *atomic.Boo
 		case kit.SteerConsumedEvent:
 			sendFn(SteerConsumedEvent{})
 		case kit.StepUsageEvent:
-			a.recordStepUsage(ev, stepUsageSeen)
+			a.recordStepUsage(ev, stepUsageSeen, sendFn)
 		case kit.PasswordPromptEvent:
 			// Convert SDK PasswordPromptEvent to app PasswordPromptEvent
 			// The TUI will handle this and send the response back
@@ -932,6 +938,8 @@ func (a *App) subscribeSDKEvents(sendFn func(tea.Msg), stepUsageSeen *atomic.Boo
 				Password:  resp.Password,
 				Cancelled: resp.Cancelled,
 			}
+		case kit.TurnEndEvent:
+			a.handleTurnEnd(ev, sendFn)
 		}
 	}))
 
@@ -942,6 +950,64 @@ func (a *App) subscribeSDKEvents(sendFn func(tea.Msg), stepUsageSeen *atomic.Boo
 	}
 }
 
+// handleTurnEnd inspects a turn's final StopReason and surfaces actionable
+// feedback to the user when the turn ended in a state they can act on.
+//
+// Today the only surfaced case is FinishReasonLength — the model hit its
+// configured max_output_tokens budget and the reply was truncated. Without
+// this banner the TUI used to swallow the truncation silently, leading to
+// "ghost" cut-offs with no indication of why.
+//
+// Separated from subscribeSDKEvents so tests can exercise it directly via a
+// stubbed sendFn without standing up a full Kit.
+func (a *App) handleTurnEnd(ev kit.TurnEndEvent, sendFn func(tea.Msg)) {
+	if sendFn == nil {
+		return
+	}
+	if ev.StopReason != kit.FinishReasonLength {
+		return
+	}
+	sendFn(ExtensionPrintEvent{
+		Level: "info",
+		Text:  a.formatMaxTokensTruncatedMessage(),
+	})
+}
+
+// formatMaxTokensTruncatedMessage builds the user-facing explanation for a
+// truncated turn. It reports the active max_output_tokens budget and, when
+// known, the model's catalog output ceiling so the user can judge how much
+// headroom is available.
+func (a *App) formatMaxTokensTruncatedMessage() string {
+	k := a.opts.Kit
+	if k == nil {
+		// Extremely early / test-stub case: still emit a useful generic hint.
+		return "⚠ Response truncated: the model hit the configured max_output_tokens limit. " +
+			"Raise it with --max-tokens N, KIT_MAX_TOKENS=N, or per-model " +
+			"modelSettings[provider/model].maxTokens in config."
+	}
+	current := k.MaxTokens()
+	ceiling := k.MaxOutputLimit()
+	model := k.GetModelString()
+
+	msg := "⚠ Response truncated: "
+	if model != "" {
+		msg += fmt.Sprintf("%s hit the configured max_output_tokens limit", model)
+	} else {
+		msg += "the model hit the configured max_output_tokens limit"
+	}
+	if current > 0 {
+		msg += fmt.Sprintf(" (%d)", current)
+	}
+	msg += "."
+	if ceiling > 0 && current > 0 && ceiling > current {
+		msg += fmt.Sprintf(" This model supports up to %d output tokens.", ceiling)
+	}
+	msg += "\n\nRaise it with --max-tokens N, KIT_MAX_TOKENS=N, " +
+		"or per-model modelSettings[provider/model].maxTokens in your config. " +
+		"Re-run the last prompt after raising it to get the full response."
+	return msg
+}
+
 // QuitFromExtension triggers a graceful shutdown. In interactive mode it
 // sends a tea.QuitMsg to the program so the TUI exits cleanly. In
 // non-interactive mode it cancels the root context, stopping any in-flight
@@ -1175,7 +1241,16 @@ 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) {
+//
+// Both session totals (cost, token counts) and the context window fill level
+// are updated here so the status bar reflects progress after every LLM call,
+// not just at the end of the full turn. Context fill monotonically increases
+// across steps because each step re-sends the entire conversation plus any
+// new tool results, so the numbers only go up.
+//
+// sendFn is called with a UsageUpdatedEvent to trigger a TUI re-render so
+// the updated values are visible immediately.
+func (a *App) recordStepUsage(ev kit.StepUsageEvent, stepUsageSeen *atomic.Bool, sendFn func(tea.Msg)) {
 	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",
@@ -1196,11 +1271,21 @@ func (a *App) recordStepUsage(ev kit.StepUsageEvent, stepUsageSeen *atomic.Bool)
 		int(ev.CacheReadTokens),
 		int(ev.CacheWriteTokens),
 	)
-	// NOTE: We do NOT call SetContextTokens here. Context fill is set once
-	// at turn completion via updateUsageFromTurnResult, which sums all token
-	// categories (Input + CacheRead + CacheCreate + Output) from FinalUsage.
-	// Per-step context tokens would cause the display to jump around during
-	// multi-step tool calls.
+	// Update context window fill from this step's usage. Each step sends
+	// the full conversation to the LLM, so the reported token counts
+	// represent the actual context utilization at that point.
+	contextFill := int(ev.InputTokens) + int(ev.CacheReadTokens) + int(ev.CacheWriteTokens) + int(ev.OutputTokens)
+	if contextFill > 0 {
+		if a.opts.Debug {
+			log.Printf("[DEBUG] recordStepUsage: SetContextTokens=%d (Input=%d + CacheRead=%d + CacheWrite=%d + Output=%d)",
+				contextFill, ev.InputTokens, ev.CacheReadTokens, ev.CacheWriteTokens, ev.OutputTokens)
+		}
+		a.opts.UsageTracker.SetContextTokens(contextFill)
+	}
+	// Notify the TUI so it re-renders the status bar with updated values.
+	if sendFn != nil {
+		sendFn(UsageUpdatedEvent{})
+	}
 }
 
 // updateUsageFromTurnResult records token usage from an SDK TurnResult into the

internal/app/app_test.go

@@ -3,10 +3,12 @@ package app
 import (
 	"context"
 	"errors"
+	"strings"
 	"sync"
 	"testing"
 	"time"
 
+	tea "charm.land/bubbletea/v2"
 	kit "github.com/mark3labs/kit/pkg/kit"
 )
 
@@ -532,9 +534,9 @@ func TestQueueLength_reflects(t *testing.T) {
 }
 
 // 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.
+// recorded immediately for cost tracking. Context tokens are also updated so
+// the status bar reflects context fill after every LLM call in a multi-step
+// turn, not just at the end.
 func TestRecordStepUsage_updatesTracker(t *testing.T) {
 	usage := &usageUpdaterStub{}
 	app := New(Options{UsageTracker: usage}, nil)
@@ -545,7 +547,7 @@ func TestRecordStepUsage_updatesTracker(t *testing.T) {
 		OutputTokens:     45,
 		CacheReadTokens:  5,
 		CacheWriteTokens: 2,
-	}, nil)
+	}, nil, nil)
 
 	usage.mu.Lock()
 	defer usage.mu.Unlock()
@@ -557,9 +559,13 @@ func TestRecordStepUsage_updatesTracker(t *testing.T) {
 		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)
+	// Context tokens should now be updated per-step (Input + CacheRead + CacheWrite + Output).
+	if usage.contextCalls != 1 {
+		t.Fatalf("expected 1 context token update from recordStepUsage, got %d", usage.contextCalls)
+	}
+	expectedContext := 120 + 45 + 5 + 2
+	if usage.lastContextTokens != expectedContext {
+		t.Fatalf("expected context tokens %d, got %d", expectedContext, usage.lastContextTokens)
 	}
 }
 
@@ -666,3 +672,94 @@ func TestUpdateUsageFromTurnResult_contextTokensUsesAllCategories(t *testing.T)
 			expected, usage.contextCalls, usage.lastContextTokens)
 	}
 }
+
+// TestHandleTurnEnd_LengthEmitsWarning verifies that when the SDK reports a
+// FinishReasonLength (max_output_tokens hit), the app surfaces a user-visible
+// ExtensionPrintEvent with Level="info" so the TUI can render a banner
+// instead of silently showing a truncated reply.
+func TestHandleTurnEnd_LengthEmitsWarning(t *testing.T) {
+	app := New(Options{}, nil)
+	defer app.Close()
+
+	var mu sync.Mutex
+	var received []tea.Msg
+	sendFn := func(m tea.Msg) {
+		mu.Lock()
+		defer mu.Unlock()
+		received = append(received, m)
+	}
+
+	app.handleTurnEnd(kit.TurnEndEvent{StopReason: kit.FinishReasonLength}, sendFn)
+
+	mu.Lock()
+	defer mu.Unlock()
+	if len(received) != 1 {
+		t.Fatalf("expected 1 event on length stop, got %d", len(received))
+	}
+	ev, ok := received[0].(ExtensionPrintEvent)
+	if !ok {
+		t.Fatalf("expected ExtensionPrintEvent, got %T", received[0])
+	}
+	if ev.Level != "info" {
+		t.Errorf("expected Level=info, got %q", ev.Level)
+	}
+	if ev.Text == "" {
+		t.Error("expected non-empty warning text")
+	}
+	if !strings.Contains(ev.Text, "max_output_tokens") {
+		t.Errorf("warning text should mention max_output_tokens, got: %s", ev.Text)
+	}
+}
+
+// TestHandleTurnEnd_NonLengthIgnored verifies that ordinary stop reasons
+// (stop, tool-calls, error, unknown, "") do not produce a warning banner.
+func TestHandleTurnEnd_NonLengthIgnored(t *testing.T) {
+	app := New(Options{}, nil)
+	defer app.Close()
+
+	reasons := []string{
+		kit.FinishReasonStop,
+		kit.FinishReasonToolCalls,
+		kit.FinishReasonError,
+		kit.FinishReasonContentFilter,
+		kit.FinishReasonOther,
+		kit.FinishReasonUnknown,
+		"",
+	}
+	for _, r := range reasons {
+		var called bool
+		app.handleTurnEnd(kit.TurnEndEvent{StopReason: r}, func(m tea.Msg) {
+			called = true
+		})
+		if called {
+			t.Errorf("stop reason %q unexpectedly emitted a warning", r)
+		}
+	}
+}
+
+// TestHandleTurnEnd_NilSendFn guards against panics when no TUI listener is
+// attached (e.g. early init or headless teardown).
+func TestHandleTurnEnd_NilSendFn(t *testing.T) {
+	app := New(Options{}, nil)
+	defer app.Close()
+
+	// Should not panic with a nil sendFn.
+	app.handleTurnEnd(kit.TurnEndEvent{StopReason: kit.FinishReasonLength}, nil)
+}
+
+// TestFormatMaxTokensTruncatedMessage_NoKit verifies the fallback message
+// when Options.Kit is nil (test/stub path).
+func TestFormatMaxTokensTruncatedMessage_NoKit(t *testing.T) {
+	app := New(Options{}, nil)
+	defer app.Close()
+
+	msg := app.formatMaxTokensTruncatedMessage()
+	if msg == "" {
+		t.Fatal("expected non-empty fallback message")
+	}
+	for _, needle := range []string{"max_output_tokens", "--max-tokens", "KIT_MAX_TOKENS", "modelSettings"} {
+		if !strings.Contains(msg, needle) {
+			t.Errorf("fallback message missing %q:\n%s", needle, msg)
+		}
+	}
+}

internal/app/events.go

@@ -32,6 +32,36 @@ type ToolCallStartedEvent struct {
 	ToolArgs string
 }
 
+// ToolCallInputStartEvent is sent when the LLM begins generating tool call
+// arguments. The tool name is known but the full argument JSON is still being
+// streamed. UIs can use this to show a "running" indicator immediately instead
+// of waiting for the full argument JSON to finish streaming.
+type ToolCallInputStartEvent struct {
+	// ToolCallID is the stable identifier for correlating tool lifecycle events.
+	ToolCallID string
+	// ToolName is the name of the tool being called.
+	ToolName string
+	// ToolKind classifies the tool: "execute", "edit", "read", "search", "agent".
+	ToolKind string
+}
+
+// ToolCallInputDeltaEvent is sent for each streamed fragment of tool call
+// arguments as they arrive from the LLM. Useful for live-previewing content
+// or showing a progress indicator with byte count.
+type ToolCallInputDeltaEvent struct {
+	// ToolCallID is the stable identifier for correlating tool lifecycle events.
+	ToolCallID string
+	// Delta is a JSON fragment of tool call arguments.
+	Delta string
+}
+
+// ToolCallInputEndEvent is sent when tool argument streaming is complete,
+// before the tool call is parsed and execution begins.
+type ToolCallInputEndEvent struct {
+	// ToolCallID is the stable identifier for correlating tool lifecycle events.
+	ToolCallID string
+}
+
 // ToolExecutionEvent is sent when a tool starts or finishes executing.
 // The IsStarting flag distinguishes between the start and end of execution.
 type ToolExecutionEvent struct {
@@ -180,6 +210,12 @@ type ModelChangedEvent struct {
 	ModelName string
 }
 
+// UsageUpdatedEvent is sent after each completed LLM step to notify the TUI
+// that token counts and costs have changed. The UsageTracker is updated
+// in-place before this event is sent; the TUI just needs to re-render to
+// reflect the new values in the status bar.
+type UsageUpdatedEvent struct{}
+
 // WidgetUpdateEvent is sent when an extension adds, updates, or removes a
 // widget via ctx.SetWidget or ctx.RemoveWidget. The TUI re-reads widget state
 // from its WidgetProvider on the next render cycle.

internal/auth/credentials.go

@@ -471,5 +471,13 @@ func GetAnthropicAPIKey(flagValue string) (string, string, error) {
 		return envKey, "ANTHROPIC_API_KEY environment variable", nil
 	}
 
+	// Check if OpenAI credentials exist to provide a helpful suggestion
+	if cm != nil {
+		hasOpenAI, _ := cm.HasOpenAICredentials()
+		if hasOpenAI {
+			return "", "", fmt.Errorf("no Anthropic API key found. Use 'kit auth login anthropic', set ANTHROPIC_API_KEY environment variable, or use --provider-api-key flag\n\nNote: OpenAI credentials were detected. To use OpenAI, run with --model openai/gpt-5.4 or set it as default:\n  kit auth login openai --set-default")
+		}
+	}
+
 	return "", "", fmt.Errorf("no Anthropic API key found. Use 'kit auth login anthropic', set ANTHROPIC_API_KEY environment variable, or use --provider-api-key flag")
 }

internal/config/config.go

@@ -30,6 +30,31 @@ type MCPServerConfig struct {
 	OAuthClientSecret string   `json:"oauthClientSecret,omitempty" yaml:"oauthClientSecret,omitempty"`
 	OAuthScopes       []string `json:"oauthScopes,omitempty" yaml:"oauthScopes,omitempty"`
 
+	// NoOAuth disables OAuth transport configuration for this server, even
+	// when the connection pool has an auth handler. Use this for public MCP
+	// servers (e.g. PubMed) that don't require authentication. Without this
+	// flag, the pool would attach OAuth transport to every remote server,
+	// causing proactive dynamic-client-registration attempts that fail on
+	// servers that don't support it.
+	NoOAuth bool `json:"noOAuth,omitempty" yaml:"noOAuth,omitempty"`
+
+	// TasksMode controls when this server's tools/call requests are augmented
+	// with MCP task metadata (turning a synchronous call into an asynchronous,
+	// pollable job — see https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks).
+	//
+	// Valid values:
+	//   - "" or "auto": (default) augment requests with task metadata only
+	//     when the server advertises tasks/toolCalls capability during initialize.
+	//   - "never":      never augment — every tool call is synchronous, regardless
+	//     of server capability.
+	//   - "always":     always augment, even when the server didn't advertise
+	//     task support. The server may still respond synchronously; this just
+	//     opts in unconditionally on the client side.
+	//
+	// In all modes, when the server returns a CreateTaskResult the client polls
+	// tasks/get / tasks/result until the task reaches a terminal state.
+	TasksMode string `json:"tasksMode,omitempty" yaml:"tasksMode,omitempty"`
+
 	// InProcessServer holds a live *server.MCPServer for in-process transport.
 	// When set (and Type is "inprocess"), the connection pool creates an
 	// in-process client instead of spawning a subprocess or making HTTP calls.
@@ -59,6 +84,8 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		OAuthClientID     string            `json:"oauthClientId,omitempty" yaml:"oauthClientId,omitempty"`
 		OAuthClientSecret string            `json:"oauthClientSecret,omitempty" yaml:"oauthClientSecret,omitempty"`
 		OAuthScopes       []string          `json:"oauthScopes,omitempty" yaml:"oauthScopes,omitempty"`
+		NoOAuth           bool              `json:"noOAuth,omitempty" yaml:"noOAuth,omitempty"`
+		TasksMode         string            `json:"tasksMode,omitempty" yaml:"tasksMode,omitempty"`
 	}
 
 	// Also try legacy format
@@ -71,6 +98,7 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		Headers       []string       `json:"headers,omitempty"`
 		AllowedTools  []string       `json:"allowedTools,omitempty" yaml:"allowedTools,omitempty"`
 		ExcludedTools []string       `json:"excludedTools,omitempty" yaml:"excludedTools,omitempty"`
+		TasksMode     string         `json:"tasksMode,omitempty" yaml:"tasksMode,omitempty"`
 	}
 
 	// Try new format first
@@ -86,6 +114,8 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		s.OAuthClientID = newConfig.OAuthClientID
 		s.OAuthClientSecret = newConfig.OAuthClientSecret
 		s.OAuthScopes = newConfig.OAuthScopes
+		s.NoOAuth = newConfig.NoOAuth
+		s.TasksMode = newConfig.TasksMode
 		return nil
 	}
 
@@ -106,6 +136,7 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 	s.Headers = legacyConfig.Headers
 	s.AllowedTools = legacyConfig.AllowedTools
 	s.ExcludedTools = legacyConfig.ExcludedTools
+	s.TasksMode = legacyConfig.TasksMode
 
 	// Infer type from legacy format for better compatibility
 	// Only set Type when it doesn't change existing transport behavior
@@ -314,6 +345,17 @@ func (c *Config) Validate() error {
 			return fmt.Errorf("server %s: allowedTools and excludedTools are mutually exclusive", serverName)
 		}
 
+		// Reject unknown tasksMode values up front so a typo (e.g. "alwasy")
+		// fails loud here instead of being silently downgraded to "auto" by
+		// the runtime parser. Comparison is case-insensitive to match
+		// tools.ParseTaskMode.
+		switch strings.ToLower(strings.TrimSpace(serverConfig.TasksMode)) {
+		case "", "auto", "never", "always":
+			// ok
+		default:
+			return fmt.Errorf("server %s: invalid tasksMode %q (expected one of: auto, never, always)", serverName, serverConfig.TasksMode)
+		}
+
 		transport := serverConfig.GetTransportType()
 		switch transport {
 		case "stdio":

internal/config/config_test.go

@@ -627,3 +627,92 @@ func TestMCPServerConfig_OAuthFields_Omitted(t *testing.T) {
 		t.Errorf("Expected empty OAuthScopes, got %v", cfg.OAuthScopes)
 	}
 }
+
+func TestMCPServerConfig_TasksMode_NewFormat(t *testing.T) {
+	jsonData := `{
+		"type": "remote",
+		"url": "https://my-mcp-server.com",
+		"tasksMode": "always"
+	}`
+	var cfg MCPServerConfig
+	if err := json.Unmarshal([]byte(jsonData), &cfg); err != nil {
+		t.Fatalf("Failed to unmarshal: %v", err)
+	}
+	if cfg.TasksMode != "always" {
+		t.Errorf("expected TasksMode 'always', got %q", cfg.TasksMode)
+	}
+}
+
+func TestMCPServerConfig_TasksMode_LegacyFormat(t *testing.T) {
+	// tasksMode also recognised in the legacy unmarshal path so users on
+	// the older command/args shape can opt in without migrating.
+	jsonData := `{
+		"command": "npx",
+		"args": ["@modelcontextprotocol/server-filesystem", "/path"],
+		"tasksMode": "never"
+	}`
+	var cfg MCPServerConfig
+	if err := json.Unmarshal([]byte(jsonData), &cfg); err != nil {
+		t.Fatalf("Failed to unmarshal: %v", err)
+	}
+	if cfg.TasksMode != "never" {
+		t.Errorf("expected TasksMode 'never', got %q", cfg.TasksMode)
+	}
+}
+
+func TestMCPServerConfig_TasksMode_DefaultEmpty(t *testing.T) {
+	// When tasksMode is not set the field stays empty, which downstream
+	// resolves to "auto" via tools.ParseTaskMode.
+	jsonData := `{"type":"remote","url":"https://x.example"}`
+	var cfg MCPServerConfig
+	if err := json.Unmarshal([]byte(jsonData), &cfg); err != nil {
+		t.Fatalf("Failed to unmarshal: %v", err)
+	}
+	if cfg.TasksMode != "" {
+		t.Errorf("expected default TasksMode to be empty, got %q", cfg.TasksMode)
+	}
+}
+
+func TestConfig_Validate_TasksMode(t *testing.T) {
+	t.Run("empty is valid", func(t *testing.T) {
+		cfg := &Config{
+			MCPServers: map[string]MCPServerConfig{
+				"a": {Type: "remote", URL: "https://x.example"},
+			},
+		}
+		if err := cfg.Validate(); err != nil {
+			t.Errorf("empty TasksMode should validate, got %v", err)
+		}
+	})
+
+	t.Run("known values are valid", func(t *testing.T) {
+		for _, mode := range []string{"auto", "never", "always", "AUTO", " always "} {
+			cfg := &Config{
+				MCPServers: map[string]MCPServerConfig{
+					"a": {Type: "remote", URL: "https://x.example", TasksMode: mode},
+				},
+			}
+			if err := cfg.Validate(); err != nil {
+				t.Errorf("TasksMode=%q should validate, got %v", mode, err)
+			}
+		}
+	})
+
+	t.Run("typo is rejected with a clear error", func(t *testing.T) {
+		cfg := &Config{
+			MCPServers: map[string]MCPServerConfig{
+				"buildbot": {Type: "remote", URL: "https://x.example", TasksMode: "alwasy"},
+			},
+		}
+		err := cfg.Validate()
+		if err == nil {
+			t.Fatal("expected validation error for invalid TasksMode")
+		}
+		// Error must mention the server name AND the bad value so the
+		// user knows where to look.
+		msg := err.Error()
+		if !strings.Contains(msg, "buildbot") || !strings.Contains(msg, `"alwasy"`) {
+			t.Errorf("error %q should mention both server name and bad value", msg)
+		}
+	})
+}

internal/core/edit.go

@@ -21,12 +21,9 @@ type Edit struct {
 }
 
 // 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
+	Path  string `json:"path"`
+	Edits []Edit `json:"edits"`
 }
 
 // replacement represents a normalized edit ready for processing.
@@ -52,20 +49,12 @@ func NewEditTool(opts ...ToolOption) fantasy.AgentTool {
 	return &coreTool{
 		info: fantasy.ToolInfo{
 			Name:        "edit",
-			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.",
+			Description: "Edit a file by replacing exact text. 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",
 					"description": "Path to the file to edit (relative or absolute)",
 				},
-				"old_text": map[string]any{
-					"type":        "string",
-					"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 (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.",
@@ -85,7 +74,7 @@ func NewEditTool(opts ...ToolOption) fantasy.AgentTool {
 					},
 				},
 			},
-			Required: []string{"path"},
+			Required: []string{"path", "edits"},
 		},
 		handler: func(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
 			return executeEdit(ctx, call, cfg.WorkDir)
@@ -163,36 +152,11 @@ func executeEdit(ctx context.Context, call fantasy.ToolCall, workDir string) (fa
 }
 
 // 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 len(args.Edits) == 0 {
+		return nil, fmt.Errorf("edits array is required and must not be empty")
 	}
 
-	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 == "" {

internal/core/edit_test.go

@@ -389,9 +389,11 @@ func TestExecuteEdit_ExactMatch(t *testing.T) {
 	writeFileOrFail(t, path, original)
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "fmt.Println(\"hello\")",
-		NewText: "fmt.Println(\"world\")",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "fmt.Println(\"hello\")",
+			NewText: "fmt.Println(\"world\")",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -426,9 +428,11 @@ func TestExecuteEdit_ExactMatch_DoesNotCorruptRest(t *testing.T) {
 	target := lines[49]
 	replacement := "REPLACED_LINE_50"
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: target,
-		NewText: replacement,
+		Path: path,
+		Edits: []Edit{{
+			OldText: target,
+			NewText: replacement,
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -470,9 +474,11 @@ func TestExecuteEdit_FuzzyMatch_TrailingWhitespace(t *testing.T) {
 
 	// Search without trailing whitespace (common LLM behavior)
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "func foo() {\n\treturn 1\n}",
-		NewText: "func foo() {\n\treturn 2\n}",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "func foo() {\n\treturn 1\n}",
+			NewText: "func foo() {\n\treturn 2\n}",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -519,9 +525,11 @@ func TestExecuteEdit_FuzzyMatch_DoesNotCorruptRest(t *testing.T) {
 	search := strings.Repeat("x", 10) + "\n" + strings.Repeat("x", 10)
 	// But this matches lines 1-2, 2-3, etc. — should fail due to ambiguity.
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: search,
-		NewText: "REPLACED",
+		Path: path,
+		Edits: []Edit{{
+			OldText: search,
+			NewText: "REPLACED",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -546,9 +554,11 @@ func TestExecuteEdit_MultipleMatches_Fails(t *testing.T) {
 	writeFileOrFail(t, path, "hello\nworld\nhello\n")
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "hello",
-		NewText: "goodbye",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "hello",
+			NewText: "goodbye",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -575,9 +585,11 @@ func TestExecuteEdit_NoMatch_Fails(t *testing.T) {
 	writeFileOrFail(t, path, "hello world\n")
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "nonexistent text",
-		NewText: "replacement",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "nonexistent text",
+			NewText: "replacement",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -601,9 +613,11 @@ func TestExecuteEdit_CRLFNormalization(t *testing.T) {
 	writeFileOrFail(t, path, "line1\r\nline2\r\nline3\r\n")
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "line2",
-		NewText: "LINE2",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "line2",
+			NewText: "LINE2",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -622,8 +636,10 @@ func TestExecuteEdit_CRLFNormalization(t *testing.T) {
 
 func TestExecuteEdit_MissingPath(t *testing.T) {
 	input, _ := json.Marshal(editArgs{
-		OldText: "x",
-		NewText: "y",
+		Edits: []Edit{{
+			OldText: "x",
+			NewText: "y",
+		}},
 	})
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, "")
 	if err != nil {
@@ -636,9 +652,11 @@ func TestExecuteEdit_MissingPath(t *testing.T) {
 
 func TestExecuteEdit_NonexistentFile(t *testing.T) {
 	input, _ := json.Marshal(editArgs{
-		Path:    "/tmp/nonexistent_edit_test_file_12345.go",
-		OldText: "x",
-		NewText: "y",
+		Path: "/tmp/nonexistent_edit_test_file_12345.go",
+		Edits: []Edit{{
+			OldText: "x",
+			NewText: "y",
+		}},
 	})
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, "")
 	if err != nil {
@@ -661,9 +679,11 @@ func TestExecuteEdit_DiffContainsHunkHeader(t *testing.T) {
 	writeFileOrFail(t, path, strings.Join(lines, "\n")+"\n")
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "line_10_content",
-		NewText: "REPLACED",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "line_10_content",
+			NewText: "REPLACED",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -684,9 +704,11 @@ func TestExecuteEdit_MetadataContainsFileDiffs(t *testing.T) {
 	writeFileOrFail(t, path, "old content\n")
 
 	input, _ := json.Marshal(editArgs{
-		Path:    path,
-		OldText: "old content",
-		NewText: "new content",
+		Path: path,
+		Edits: []Edit{{
+			OldText: "old content",
+			NewText: "new content",
+		}},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -905,18 +927,14 @@ func TestExecuteEdit_MultiEdit_EmptyArray(t *testing.T) {
 	}
 }
 
-func TestExecuteEdit_MultiEdit_MixedWithSingleMode(t *testing.T) {
+func TestExecuteEdit_EmptyEditsArray_Fails(t *testing.T) {
 	dir := t.TempDir()
-	path := filepath.Join(dir, "mixed.txt")
+	path := filepath.Join(dir, "empty.txt")
 	writeFileOrFail(t, path, "hello\n")
 
-	input, _ := json.Marshal(map[string]any{
-		"path":     path,
-		"old_text": "hello",
-		"new_text": "HELLO",
-		"edits": []Edit{
-			{OldText: "hello", NewText: "HI"},
-		},
+	input, _ := json.Marshal(editArgs{
+		Path:  path,
+		Edits: []Edit{},
 	})
 
 	resp, err := executeEdit(t.Context(), fantasy.ToolCall{Input: string(input)}, dir)
@@ -924,10 +942,10 @@ func TestExecuteEdit_MultiEdit_MixedWithSingleMode(t *testing.T) {
 		t.Fatalf("executeEdit error: %v", err)
 	}
 	if !resp.IsError {
-		t.Error("expected error when mixing single and multi-edit modes")
+		t.Error("expected error for empty edits array")
 	}
-	if !strings.Contains(resp.Content, "cannot use") {
-		t.Errorf("expected 'cannot use' in error, got: %s", resp.Content)
+	if !strings.Contains(resp.Content, "required") {
+		t.Errorf("expected 'required' in error, got: %s", resp.Content)
 	}
 }
 

internal/core/subagent.go

@@ -86,7 +86,7 @@ Example use cases:
 				},
 				"model": map[string]any{
 					"type":        "string",
-					"description": "Optional model override (e.g. 'anthropic/claude-haiku-3-5-20241022' for faster/cheaper tasks)",
+					"description": "Optional model override. Empty string uses the current model.",
 				},
 				"system_prompt": map[string]any{
 					"type":        "string",
@@ -94,7 +94,7 @@ Example use cases:
 				},
 				"timeout_seconds": map[string]any{
 					"type":        "number",
-					"description": "Maximum execution time in seconds (default: 300, max: 1800)",
+					"description": "Maximum execution time in seconds (default: 300, max: 1800, minimum recommended: 240)",
 				},
 			},
 			Required: []string{"task"},

internal/extensions/api.go

@@ -1063,6 +1063,9 @@ type PrintBlockOpts struct {
 type API struct {
 	// Event-specific registration functions (wired by the loader).
 	onToolCall                func(func(ToolCallEvent, Context) *ToolCallResult)
+	onToolCallInputStart      func(func(ToolCallInputStartEvent, Context))
+	onToolCallInputDelta      func(func(ToolCallInputDeltaEvent, Context))
+	onToolCallInputEnd        func(func(ToolCallInputEndEvent, Context))
 	onToolExecStart           func(func(ToolExecutionStartEvent, Context))
 	onToolExecEnd             func(func(ToolExecutionEndEvent, Context))
 	onToolOutput              func(func(ToolOutputEvent, Context))
@@ -1091,6 +1094,14 @@ type API struct {
 	onSubagentStart           func(func(SubagentStartEvent, Context))
 	onSubagentChunk           func(func(SubagentChunkEvent, Context))
 	onSubagentEnd             func(func(SubagentEndEvent, Context))
+	onStepStart               func(func(StepStartEvent, Context))
+	onStepFinish              func(func(StepFinishEvent, Context))
+	onReasoningStart          func(func(ReasoningStartEvent, Context))
+	onWarnings                func(func(WarningsEvent, Context))
+	onSource                  func(func(SourceEvent, Context))
+	onError                   func(func(ErrorEvent, Context))
+	onRetry                   func(func(RetryEvent, Context))
+	onPrepareStep             func(func(PrepareStepEvent, Context) *PrepareStepResult)
 }
 
 // OnToolCall registers a handler that fires before a tool executes.
@@ -1099,6 +1110,26 @@ func (a *API) OnToolCall(handler func(ToolCallEvent, Context) *ToolCallResult) {
 	a.onToolCall(handler)
 }
 
+// OnToolCallInputStart registers a handler that fires when the LLM begins
+// generating tool call arguments. The tool name is known but the full
+// argument JSON is still being streamed. Useful for showing a "running"
+// indicator immediately without waiting for the full arguments.
+func (a *API) OnToolCallInputStart(handler func(ToolCallInputStartEvent, Context)) {
+	a.onToolCallInputStart(handler)
+}
+
+// OnToolCallInputDelta registers a handler that fires for each streamed
+// fragment of tool call arguments as they arrive from the LLM.
+func (a *API) OnToolCallInputDelta(handler func(ToolCallInputDeltaEvent, Context)) {
+	a.onToolCallInputDelta(handler)
+}
+
+// OnToolCallInputEnd registers a handler that fires when tool argument
+// streaming is complete, before the tool call is parsed and execution begins.
+func (a *API) OnToolCallInputEnd(handler func(ToolCallInputEndEvent, Context)) {
+	a.onToolCallInputEnd(handler)
+}
+
 // OnToolExecutionStart registers a handler for tool execution start.
 func (a *API) OnToolExecutionStart(handler func(ToolExecutionStartEvent, Context)) {
 	a.onToolExecStart(handler)
@@ -1278,6 +1309,56 @@ func (a *API) OnBeforeCompact(handler func(BeforeCompactEvent, Context) *BeforeC
 	a.onBeforeCompact(handler)
 }
 
+// OnStepStart registers a handler that fires when a new LLM call begins
+// within a multi-step agent turn.
+func (a *API) OnStepStart(handler func(StepStartEvent, Context)) {
+	a.onStepStart(handler)
+}
+
+// OnStepFinish registers a handler that fires when a step completes,
+// providing step number, finish reason, and decomposed token usage.
+func (a *API) OnStepFinish(handler func(StepFinishEvent, Context)) {
+	a.onStepFinish(handler)
+}
+
+// OnReasoningStart registers a handler that fires when the LLM begins
+// reasoning/thinking.
+func (a *API) OnReasoningStart(handler func(ReasoningStartEvent, Context)) {
+	a.onReasoningStart(handler)
+}
+
+// OnWarnings registers a handler that fires when the LLM provider returns
+// warnings about the request.
+func (a *API) OnWarnings(handler func(WarningsEvent, Context)) {
+	a.onWarnings(handler)
+}
+
+// OnSource registers a handler that fires when the LLM references a source
+// (e.g. from web search tools).
+func (a *API) OnSource(handler func(SourceEvent, Context)) {
+	a.onSource(handler)
+}
+
+// OnError registers a handler that fires when an agent-level error occurs
+// during streaming.
+func (a *API) OnError(handler func(ErrorEvent, Context)) {
+	a.onError(handler)
+}
+
+// OnRetry registers a handler that fires when the LLM provider request is
+// retried after a transient error.
+func (a *API) OnRetry(handler func(RetryEvent, Context)) {
+	a.onRetry(handler)
+}
+
+// OnPrepareStep registers a handler that fires between steps within a
+// multi-step agent turn, after steering messages are injected and before
+// messages are sent to the LLM. Return a non-nil PrepareStepResult with
+// Messages to replace the context window for this step.
+func (a *API) OnPrepareStep(handler func(PrepareStepEvent, Context) *PrepareStepResult) {
+	a.onPrepareStep(handler)
+}
+
 // RegisterToolRenderer registers a custom renderer for a specific tool's
 // display in the TUI. The renderer controls the header (parameter summary)
 // and/or body (result display) of the tool's output block. If multiple
@@ -1890,6 +1971,34 @@ type ToolCallResult struct {
 
 func (ToolCallResult) isResult() {}
 
+// ToolCallInputStartEvent fires when the LLM begins generating tool call
+// arguments. The tool name is known but the full argument JSON is still
+// being streamed.
+type ToolCallInputStartEvent struct {
+	ToolCallID string
+	ToolName   string
+	ToolKind   string // Tool classification: "execute", "edit", "read", "search", "agent"
+}
+
+func (e ToolCallInputStartEvent) Type() EventType { return ToolCallInputStart }
+
+// ToolCallInputDeltaEvent fires for each streamed fragment of tool call
+// arguments as they arrive from the LLM.
+type ToolCallInputDeltaEvent struct {
+	ToolCallID string
+	Delta      string // JSON fragment of tool arguments
+}
+
+func (e ToolCallInputDeltaEvent) Type() EventType { return ToolCallInputDelta }
+
+// ToolCallInputEndEvent fires when tool argument streaming is complete,
+// before the tool call is parsed and execution begins.
+type ToolCallInputEndEvent struct {
+	ToolCallID string
+}
+
+func (e ToolCallInputEndEvent) Type() EventType { return ToolCallInputEnd }
+
 // ToolExecutionStartEvent fires when a tool begins executing.
 type ToolExecutionStartEvent struct {
 	ToolCallID string
@@ -2202,6 +2311,98 @@ type SubagentEndEvent struct {
 
 func (e SubagentEndEvent) Type() EventType { return SubagentEnd }
 
+// ---------------------------------------------------------------------------
+// Step lifecycle events (exposed to Yaegi — concrete structs)
+// ---------------------------------------------------------------------------
+
+// StepStartEvent fires when a new LLM call begins within a multi-step agent turn.
+type StepStartEvent struct {
+	StepNumber int
+}
+
+func (e StepStartEvent) Type() EventType { return StepStart }
+
+// StepFinishEvent fires when a step completes, providing step metadata and
+// token usage. Usage fields are plain int64 (not LLMUsage) because Yaegi
+// cannot handle fantasy types across the interpreter boundary.
+type StepFinishEvent struct {
+	StepNumber       int
+	HasToolCalls     bool
+	FinishReason     string
+	InputTokens      int64
+	OutputTokens     int64
+	CacheReadTokens  int64
+	CacheWriteTokens int64
+}
+
+func (e StepFinishEvent) Type() EventType { return StepFinish }
+
+// ReasoningStartEvent fires when the LLM begins reasoning/thinking.
+type ReasoningStartEvent struct {
+	ID string
+}
+
+func (e ReasoningStartEvent) Type() EventType { return ReasoningStart }
+
+// WarningsEvent fires when the LLM provider returns warnings about the request.
+type WarningsEvent struct {
+	Warnings []string
+}
+
+func (e WarningsEvent) Type() EventType { return Warnings }
+
+// SourceEvent fires when the LLM references a source (e.g. from web search).
+type SourceEvent struct {
+	SourceType string
+	ID         string
+	URL        string
+	Title      string
+}
+
+func (e SourceEvent) Type() EventType { return Source }
+
+// ErrorEvent fires when an agent-level error occurs during streaming.
+// Uses string instead of error because Yaegi cannot handle the error
+// interface reliably across the interpreter boundary.
+type ErrorEvent struct {
+	Error string
+}
+
+func (e ErrorEvent) Type() EventType { return Error }
+
+// RetryEvent fires when the LLM provider request is retried after a
+// transient error.
+type RetryEvent struct {
+	Attempt int
+	Error   string
+}
+
+func (e RetryEvent) Type() EventType { return Retry }
+
+// PrepareStepEvent fires between steps within a multi-step agent turn,
+// after steering messages are injected and before messages are sent to
+// the LLM. Handlers can inspect and replace the context window.
+type PrepareStepEvent struct {
+	// StepNumber is the zero-based step index within the current turn.
+	StepNumber int
+	// Messages is the current context window that will be sent to the LLM.
+	Messages []ContextMessage
+}
+
+func (e PrepareStepEvent) Type() EventType { return PrepareStep }
+
+// PrepareStepResult allows extensions to replace the context window between
+// steps. Return nil Messages to leave the context unchanged.
+type PrepareStepResult struct {
+	// Messages replaces the entire context window for this step. If nil,
+	// the original messages are used unchanged. Messages with a non-negative
+	// Index reuse the original message at that position; messages with
+	// Index < 0 are created fresh from Role + Content.
+	Messages []ContextMessage
+}
+
+func (PrepareStepResult) isResult() {}
+
 // 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

@@ -13,6 +13,19 @@ const (
 	// ToolCall fires before a tool executes. Handlers can block execution.
 	ToolCall EventType = "tool_call"
 
+	// ToolCallInputStart fires when the LLM begins generating tool call
+	// arguments. The tool name is known but the full argument JSON is still
+	// being streamed.
+	ToolCallInputStart EventType = "tool_call_input_start"
+
+	// ToolCallInputDelta fires for each streamed fragment of tool call
+	// arguments as they arrive from the LLM.
+	ToolCallInputDelta EventType = "tool_call_input_delta"
+
+	// ToolCallInputEnd fires when tool argument streaming is complete,
+	// before the tool call is parsed and execution begins.
+	ToolCallInputEnd EventType = "tool_call_input_end"
+
 	// ToolExecutionStart fires when a tool begins executing.
 	ToolExecutionStart EventType = "tool_execution_start"
 
@@ -83,18 +96,50 @@ const (
 	// SubagentEnd fires when a subagent tool call completes (success
 	// or error). Carries the final response and any error message.
 	SubagentEnd EventType = "subagent_end"
+
+	// StepStart fires when a new LLM call begins within a multi-step
+	// agent turn.
+	StepStart EventType = "step_start"
+
+	// StepFinish fires when a step completes, providing step number,
+	// finish reason, and token usage.
+	StepFinish EventType = "step_finish"
+
+	// ReasoningStart fires when the LLM begins reasoning/thinking.
+	ReasoningStart EventType = "reasoning_start"
+
+	// Warnings fires when the LLM provider returns warnings.
+	Warnings EventType = "warnings"
+
+	// Source fires when the LLM references a source (e.g. web search).
+	Source EventType = "source"
+
+	// Error fires when an agent-level error occurs during streaming.
+	Error EventType = "error"
+
+	// Retry fires when the LLM provider request is retried after a
+	// transient error.
+	Retry EventType = "retry"
+
+	// PrepareStep fires between steps within a multi-step agent turn,
+	// after steering messages are injected and before messages are sent
+	// to the LLM. Handlers can replace the context window for this step.
+	PrepareStep EventType = "prepare_step"
 )
 
 // AllEventTypes returns every supported event type.
 func AllEventTypes() []EventType {
 	return []EventType{
-		ToolCall, ToolExecutionStart, ToolExecutionEnd, ToolResult,
+		ToolCall, ToolCallInputStart, ToolCallInputDelta, ToolCallInputEnd,
+		ToolExecutionStart, ToolExecutionEnd, ToolResult,
 		Input, BeforeAgentStart, AgentStart, AgentEnd,
 		MessageStart, MessageUpdate, MessageEnd,
 		SessionStart, SessionShutdown,
 		ModelChange, ContextPrepare,
 		BeforeFork, BeforeSessionSwitch, BeforeCompact,
 		SubagentStart, SubagentChunk, SubagentEnd,
+		StepStart, StepFinish, ReasoningStart, Warnings, Source, Error, Retry,
+		PrepareStep,
 	}
 }
 

internal/extensions/events_test.go

@@ -4,8 +4,8 @@ import "testing"
 
 func TestAllEventTypes_Count(t *testing.T) {
 	all := AllEventTypes()
-	if len(all) != 21 {
-		t.Fatalf("expected 21 event types, got %d", len(all))
+	if len(all) != 32 {
+		t.Fatalf("expected 32 event types, got %d", len(all))
 	}
 }
 
@@ -38,6 +38,9 @@ func TestEventType_TypeMethod(t *testing.T) {
 		want  EventType
 	}{
 		{ToolCallEvent{ToolName: "test"}, ToolCall},
+		{ToolCallInputStartEvent{ToolCallID: "x", ToolName: "test"}, ToolCallInputStart},
+		{ToolCallInputDeltaEvent{ToolCallID: "x", Delta: "{"}, ToolCallInputDelta},
+		{ToolCallInputEndEvent{ToolCallID: "x"}, ToolCallInputEnd},
 		{ToolExecutionStartEvent{ToolName: "test"}, ToolExecutionStart},
 		{ToolExecutionEndEvent{ToolName: "test"}, ToolExecutionEnd},
 		{ToolResultEvent{ToolName: "test"}, ToolResult},

internal/extensions/loader.go

@@ -429,6 +429,24 @@ func loadSingleExtension(path string) (*LoadedExtension, error) {
 				return *r
 			})
 		},
+		onToolCallInputStart: func(h func(ToolCallInputStartEvent, Context)) {
+			reg(ToolCallInputStart, func(e Event, c Context) Result {
+				h(e.(ToolCallInputStartEvent), c)
+				return nil
+			})
+		},
+		onToolCallInputDelta: func(h func(ToolCallInputDeltaEvent, Context)) {
+			reg(ToolCallInputDelta, func(e Event, c Context) Result {
+				h(e.(ToolCallInputDeltaEvent), c)
+				return nil
+			})
+		},
+		onToolCallInputEnd: func(h func(ToolCallInputEndEvent, Context)) {
+			reg(ToolCallInputEnd, func(e Event, c Context) Result {
+				h(e.(ToolCallInputEndEvent), c)
+				return nil
+			})
+		},
 		onToolExecStart: func(h func(ToolExecutionStartEvent, Context)) {
 			reg(ToolExecutionStart, func(e Event, c Context) Result {
 				h(e.(ToolExecutionStartEvent), c)
@@ -600,6 +618,57 @@ func loadSingleExtension(path string) (*LoadedExtension, error) {
 				return nil
 			})
 		},
+		onStepStart: func(h func(StepStartEvent, Context)) {
+			reg(StepStart, func(e Event, c Context) Result {
+				h(e.(StepStartEvent), c)
+				return nil
+			})
+		},
+		onStepFinish: func(h func(StepFinishEvent, Context)) {
+			reg(StepFinish, func(e Event, c Context) Result {
+				h(e.(StepFinishEvent), c)
+				return nil
+			})
+		},
+		onReasoningStart: func(h func(ReasoningStartEvent, Context)) {
+			reg(ReasoningStart, func(e Event, c Context) Result {
+				h(e.(ReasoningStartEvent), c)
+				return nil
+			})
+		},
+		onWarnings: func(h func(WarningsEvent, Context)) {
+			reg(Warnings, func(e Event, c Context) Result {
+				h(e.(WarningsEvent), c)
+				return nil
+			})
+		},
+		onSource: func(h func(SourceEvent, Context)) {
+			reg(Source, func(e Event, c Context) Result {
+				h(e.(SourceEvent), c)
+				return nil
+			})
+		},
+		onError: func(h func(ErrorEvent, Context)) {
+			reg(Error, func(e Event, c Context) Result {
+				h(e.(ErrorEvent), c)
+				return nil
+			})
+		},
+		onRetry: func(h func(RetryEvent, Context)) {
+			reg(Retry, func(e Event, c Context) Result {
+				h(e.(RetryEvent), c)
+				return nil
+			})
+		},
+		onPrepareStep: func(h func(PrepareStepEvent, Context) *PrepareStepResult) {
+			reg(PrepareStep, func(e Event, c Context) Result {
+				r := h(e.(PrepareStepEvent), c)
+				if r == nil {
+					return nil
+				}
+				return *r
+			})
+		},
 	}
 
 	// Call Init — the extension registers its handlers, tools, commands.

internal/extensions/runner.go

@@ -1,21 +1,93 @@
 package extensions
 
 import (
+	"bytes"
 	"fmt"
 	"log"
 	"os"
+	"runtime"
 	"sort"
+	"strconv"
 	"strings"
 	"sync"
 
 	"github.com/spf13/viper"
 )
 
+// ---------------------------------------------------------------------------
+// reentrantMu — a per-extension mutex that allows the same goroutine to
+// re-enter (e.g. handler → ctx.EmitCustomEvent → handler in same extension).
+// Different goroutines are serialized, preventing concurrent state mutation.
+// ---------------------------------------------------------------------------
+
+type reentrantMu struct {
+	mu    sync.Mutex
+	cond  *sync.Cond
+	owner int64 // goroutine ID that holds the lock, or 0
+	depth int   // re-entrancy depth
+}
+
+// initReentrantMu initializes the reentrant mutex in-place. Must be called
+// after the struct is at its final memory location (not before copying).
+func (r *reentrantMu) init() {
+	r.cond = sync.NewCond(&r.mu)
+}
+
+// lock acquires the mutex. If the calling goroutine already holds it, the
+// call succeeds immediately (re-entrant). Every call to lock must be paired
+// with a call to unlock.
+func (r *reentrantMu) lock() {
+	gid := goroutineID()
+	r.mu.Lock()
+	if r.owner == gid {
+		// Re-entrant: same goroutine already holds the lock.
+		r.depth++
+		r.mu.Unlock()
+		return
+	}
+	// Wait for the current owner to release.
+	for r.owner != 0 {
+		r.cond.Wait() // releases mu, blocks, re-acquires mu on wake
+	}
+	r.owner = gid
+	r.depth = 1
+	r.mu.Unlock()
+}
+
+// unlock releases the mutex (or decrements re-entrancy depth).
+func (r *reentrantMu) unlock() {
+	r.mu.Lock()
+	r.depth--
+	if r.depth == 0 {
+		r.owner = 0
+		r.cond.Signal()
+	}
+	r.mu.Unlock()
+}
+
+// goroutineID extracts the current goroutine's ID from runtime.Stack output.
+// This is a well-known technique used by Go testing infrastructure.
+func goroutineID() int64 {
+	var buf [64]byte
+	n := runtime.Stack(buf[:], false)
+	// Stack output starts with "goroutine NNN ["
+	s := buf[:n]
+	s = s[len("goroutine "):]
+	s = s[:bytes.IndexByte(s, ' ')]
+	id, _ := strconv.ParseInt(string(s), 10, 64)
+	return id
+}
+
 // Runner manages loaded extensions and dispatches events to their handlers
 // sequentially. Handlers execute in extension
 // load order; for cancellable events the first blocking result wins.
+//
+// Each extension has a dedicated reentrant mutex so that handlers for the
+// same extension are serialized (preventing data races on shared package-level
+// state), while handlers for different extensions may execute concurrently.
 type Runner struct {
 	extensions      []LoadedExtension
+	extMu           []reentrantMu // per-extension reentrant mutex, indexed by extension position
 	ctx             Context
 	widgets         map[string]WidgetConfig   // keyed by widget ID
 	statusEntries   map[string]StatusBarEntry // keyed by status key
@@ -52,7 +124,11 @@ type LoadedExtension struct {
 
 // NewRunner creates a Runner from a set of loaded extensions.
 func NewRunner(exts []LoadedExtension) *Runner {
-	return &Runner{extensions: exts}
+	mus := make([]reentrantMu, len(exts))
+	for i := range mus {
+		mus[i].init()
+	}
+	return &Runner{extensions: exts, extMu: mus}
 }
 
 // SetContext updates the runtime context (session ID, model, etc.) that is
@@ -367,6 +443,11 @@ func (r *Runner) Emit(event Event) (Result, error) {
 	for i := range r.extensions {
 		ext := &r.extensions[i]
 		handlers := ext.Handlers[event.Type()]
+		if len(handlers) == 0 {
+			continue
+		}
+
+		r.extMu[i].lock()
 		for _, handler := range handlers {
 			result, err := safeCall(handler, event, ctx)
 			if err != nil {
@@ -379,6 +460,7 @@ func (r *Runner) Emit(event Event) (Result, error) {
 
 			// Check for blocking/short-circuit results.
 			if isBlocking(result) {
+				r.extMu[i].unlock()
 				return result, nil
 			}
 
@@ -386,6 +468,7 @@ func (r *Runner) Emit(event Event) (Result, error) {
 			// the caller is responsible for applying the modifications.
 			accumulated = result
 		}
+		r.extMu[i].unlock()
 	}
 	return accumulated, nil
 }
@@ -712,11 +795,17 @@ func (r *Runner) EmitCustomEvent(name, data string) {
 
 	// Extension-registered handlers first (in load order).
 	for i := range r.extensions {
-		for _, h := range r.extensions[i].CustomEventHandlers[name] {
+		extHandlers := r.extensions[i].CustomEventHandlers[name]
+		if len(extHandlers) == 0 {
+			continue
+		}
+		r.extMu[i].lock()
+		for _, h := range extHandlers {
 			safeInvoke(h)
 		}
+		r.extMu[i].unlock()
 	}
-	// Then dynamic subscriptions.
+	// Then dynamic subscriptions (not extension-scoped, no per-ext lock).
 	for _, h := range dynamicHandlers {
 		safeInvoke(h)
 	}

internal/extensions/runner_test.go

@@ -1,6 +1,7 @@
 package extensions
 
 import (
+	"sync"
 	"testing"
 )
 
@@ -571,3 +572,142 @@ func TestRunner_ContextPrintNilSafe(t *testing.T) {
 		t.Fatalf("unexpected error: %v", err)
 	}
 }
+
+func TestRunner_ConcurrentEmitSameExtension(t *testing.T) {
+	// Verify that concurrent Emit calls for the same extension are serialized
+	// and don't cause data races on shared handler state.
+	var counter int
+	ext := makeHandlerExt("shared-state.go", map[EventType][]HandlerFunc{
+		SubagentStart: {
+			func(e Event, c Context) Result {
+				// Read-modify-write: racy without serialization.
+				v := counter
+				counter = v + 1
+				return nil
+			},
+		},
+		SubagentChunk: {
+			func(e Event, c Context) Result {
+				v := counter
+				counter = v + 1
+				return nil
+			},
+		},
+	})
+
+	r := makeRunner(ext)
+	var wg sync.WaitGroup
+	const goroutines = 20
+	const iterations = 50
+	wg.Add(goroutines)
+	for range goroutines {
+		go func() {
+			defer wg.Done()
+			for range iterations {
+				_, _ = r.Emit(SubagentStartEvent{ToolCallID: "x"})
+				_, _ = r.Emit(SubagentChunkEvent{ToolCallID: "x"})
+			}
+		}()
+	}
+	wg.Wait()
+	if counter != goroutines*iterations*2 {
+		t.Errorf("expected counter=%d, got %d (race detected)", goroutines*iterations*2, counter)
+	}
+}
+
+func TestRunner_ConcurrentEmitDifferentExtensions(t *testing.T) {
+	// Two extensions with independent state should not block each other
+	// and should both run correctly under concurrent Emit calls.
+	var counter1, counter2 int
+	ext1 := makeHandlerExt("ext1.go", map[EventType][]HandlerFunc{
+		SubagentStart: {
+			func(e Event, c Context) Result {
+				v := counter1
+				counter1 = v + 1
+				return nil
+			},
+		},
+	})
+	ext2 := makeHandlerExt("ext2.go", map[EventType][]HandlerFunc{
+		SubagentStart: {
+			func(e Event, c Context) Result {
+				v := counter2
+				counter2 = v + 1
+				return nil
+			},
+		},
+	})
+
+	r := makeRunner(ext1, ext2)
+	var wg sync.WaitGroup
+	const goroutines = 20
+	const iterations = 50
+	wg.Add(goroutines)
+	for range goroutines {
+		go func() {
+			defer wg.Done()
+			for range iterations {
+				_, _ = r.Emit(SubagentStartEvent{ToolCallID: "x"})
+			}
+		}()
+	}
+	wg.Wait()
+	expected := goroutines * iterations
+	if counter1 != expected {
+		t.Errorf("ext1 counter: expected %d, got %d", expected, counter1)
+	}
+	if counter2 != expected {
+		t.Errorf("ext2 counter: expected %d, got %d", expected, counter2)
+	}
+}
+
+func TestRunner_ReentrantEmitCustomEvent(t *testing.T) {
+	// Verify that a handler can call EmitCustomEvent (which dispatches to
+	// the same extension's custom event handlers) without deadlocking.
+	var order []string
+	ext := LoadedExtension{
+		Path: "reentrant.go",
+		Handlers: map[EventType][]HandlerFunc{
+			SessionStart: {
+				func(e Event, c Context) Result {
+					order = append(order, "session_start")
+					// This triggers EmitCustomEvent for the same extension
+					// via a direct runner call (simulating ctx.EmitCustomEvent).
+					return nil
+				},
+			},
+		},
+		CustomEventHandlers: map[string][]func(string){
+			"test-event": {
+				func(data string) {
+					order = append(order, "custom:"+data)
+				},
+			},
+		},
+	}
+
+	r := makeRunner(ext)
+
+	// Wire up the handler to call EmitCustomEvent re-entrantly.
+	ext.Handlers[SessionStart] = []HandlerFunc{
+		func(e Event, c Context) Result {
+			order = append(order, "session_start")
+			r.EmitCustomEvent("test-event", "hello")
+			return nil
+		},
+	}
+	r.extensions[0] = ext
+	// Rebuild mutexes after modifying extensions slice.
+	r.extMu = make([]reentrantMu, len(r.extensions))
+	for i := range r.extMu {
+		r.extMu[i].init()
+	}
+
+	_, err := r.Emit(SessionStartEvent{})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(order) != 2 || order[0] != "session_start" || order[1] != "custom:hello" {
+		t.Errorf("expected [session_start, custom:hello], got %v", order)
+	}
+}

internal/extensions/symbols.go

@@ -152,6 +152,9 @@ func Symbols() interp.Exports {
 			// Event structs
 			"ToolCallEvent":           reflect.ValueOf((*ToolCallEvent)(nil)),
 			"ToolCallResult":          reflect.ValueOf((*ToolCallResult)(nil)),
+			"ToolCallInputStartEvent": reflect.ValueOf((*ToolCallInputStartEvent)(nil)),
+			"ToolCallInputDeltaEvent": reflect.ValueOf((*ToolCallInputDeltaEvent)(nil)),
+			"ToolCallInputEndEvent":   reflect.ValueOf((*ToolCallInputEndEvent)(nil)),
 			"ToolExecutionStartEvent": reflect.ValueOf((*ToolExecutionStartEvent)(nil)),
 			"ToolExecutionEndEvent":   reflect.ValueOf((*ToolExecutionEndEvent)(nil)),
 			"ToolOutputEvent":         reflect.ValueOf((*ToolOutputEvent)(nil)),
@@ -169,6 +172,17 @@ func Symbols() interp.Exports {
 			"SessionStartEvent":       reflect.ValueOf((*SessionStartEvent)(nil)),
 			"SessionShutdownEvent":    reflect.ValueOf((*SessionShutdownEvent)(nil)),
 			"ModelChangeEvent":        reflect.ValueOf((*ModelChangeEvent)(nil)),
+
+			// Step lifecycle events
+			"StepStartEvent":      reflect.ValueOf((*StepStartEvent)(nil)),
+			"StepFinishEvent":     reflect.ValueOf((*StepFinishEvent)(nil)),
+			"ReasoningStartEvent": reflect.ValueOf((*ReasoningStartEvent)(nil)),
+			"WarningsEvent":       reflect.ValueOf((*WarningsEvent)(nil)),
+			"SourceEvent":         reflect.ValueOf((*SourceEvent)(nil)),
+			"ErrorEvent":          reflect.ValueOf((*ErrorEvent)(nil)),
+			"RetryEvent":          reflect.ValueOf((*RetryEvent)(nil)),
+			"PrepareStepEvent":    reflect.ValueOf((*PrepareStepEvent)(nil)),
+			"PrepareStepResult":   reflect.ValueOf((*PrepareStepResult)(nil)),
 		},
 	}
 }

internal/extensions/wrapper.go

@@ -90,8 +90,7 @@ func (w *wrappedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.T
 	// 0. Check if tool is disabled via SetActiveTools.
 	if w.runner.IsToolDisabled(toolName) {
 		return fantasy.NewTextErrorResponse(
-				fmt.Sprintf("Error: tool %q is currently disabled", toolName)),
-			fmt.Errorf("tool %q disabled by extension", toolName)
+			fmt.Sprintf("Error: tool %q is currently disabled", toolName)), nil
 	}
 
 	kind := toolKindFor(toolName)
@@ -111,8 +110,7 @@ func (w *wrappedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.T
 			if reason == "" {
 				reason = "blocked by extension"
 			}
-			return fantasy.NewTextErrorResponse(fmt.Sprintf("Error: %s", reason)),
-				fmt.Errorf("tool blocked by extension: %s", reason)
+			return fantasy.NewTextErrorResponse(fmt.Sprintf("Error: %s", reason)), nil
 		}
 	}
 
@@ -238,7 +236,7 @@ func (t *extensionTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy
 	}
 
 	if err != nil {
-		return fantasy.NewTextErrorResponse(err.Error()), err
+		return fantasy.NewTextErrorResponse(err.Error()), nil
 	}
 	return fantasy.NewTextResponse(result), nil
 }

internal/extensions/wrapper_test.go

@@ -142,8 +142,8 @@ func TestWrappedTool_BlockExecution(t *testing.T) {
 	if toolRan {
 		t.Error("tool should not have run after block")
 	}
-	if err == nil {
-		t.Error("expected error from blocked tool")
+	if err != nil {
+		t.Error("expected nil error for blocked tool (error is conveyed via IsError response)")
 	}
 	if resp.IsError != true {
 		t.Error("expected IsError=true from blocked response")
@@ -234,8 +234,8 @@ func TestExtensionTool_Error(t *testing.T) {
 
 	tools := ExtensionToolsAsLLMTools(defs, nil)
 	resp, err := tools[0].Run(context.Background(), fantasy.ToolCall{Input: "x"})
-	if err == nil {
-		t.Error("expected error")
+	if err != nil {
+		t.Error("expected nil error (error is conveyed via IsError response)")
 	}
 	if !resp.IsError {
 		t.Error("expected IsError=true")

internal/kitsetup/setup.go

@@ -72,6 +72,9 @@ type AgentSetupOptions struct {
 	// OnMCPServerLoaded, if non-nil, is called when each MCP server finishes
 	// loading (successfully or with error). Called from the background goroutine.
 	OnMCPServerLoaded func(serverName string, toolCount int, err error)
+	// MCPTaskConfig configures task-augmented tools/call execution. The
+	// zero value preserves historical synchronous-only behaviour.
+	MCPTaskConfig tools.MCPTaskConfig
 }
 
 // AgentSetupResult bundles the created agent and any debug logger so the caller
@@ -229,6 +232,7 @@ func SetupAgent(ctx context.Context, opts AgentSetupOptions) (*AgentSetupResult,
 		ToolWrapper:       toolWrapper,
 		ExtraTools:        extraTools,
 		OnMCPServerLoaded: opts.OnMCPServerLoaded,
+		MCPTaskConfig:     opts.MCPTaskConfig,
 	})
 	if err != nil {
 		return nil, fmt.Errorf("failed to create agent: %w", err)

internal/models/providers.go

@@ -85,6 +85,7 @@ type ThinkingLevel string
 
 const (
 	ThinkingOff     ThinkingLevel = "off"
+	ThinkingNone    ThinkingLevel = "none"
 	ThinkingMinimal ThinkingLevel = "minimal"
 	ThinkingLow     ThinkingLevel = "low"
 	ThinkingMedium  ThinkingLevel = "medium"
@@ -93,12 +94,14 @@ const (
 
 // ThinkingLevels returns the ordered list of available thinking levels for cycling.
 func ThinkingLevels() []ThinkingLevel {
-	return []ThinkingLevel{ThinkingOff, ThinkingMinimal, ThinkingLow, ThinkingMedium, ThinkingHigh}
+	return []ThinkingLevel{ThinkingOff, ThinkingNone, ThinkingMinimal, ThinkingLow, ThinkingMedium, ThinkingHigh}
 }
 
-// thinkingBudgetTokens returns the token budget for a thinking level, or 0 for "off".
+// thinkingBudgetTokens returns the token budget for a thinking level, or 0 for "off" or "none".
 func thinkingBudgetTokens(level ThinkingLevel) int64 {
 	switch level {
+	case ThinkingNone:
+		return 1024
 	case ThinkingMinimal:
 		return 1024
 	case ThinkingLow:
@@ -117,6 +120,8 @@ func ThinkingLevelDescription(level ThinkingLevel) string {
 	switch level {
 	case ThinkingOff:
 		return "No reasoning"
+	case ThinkingNone:
+		return "Minimal reasoning (OpenAI 'none')"
 	case ThinkingMinimal:
 		return "Very brief reasoning (~1k tokens)"
 	case ThinkingLow:
@@ -133,7 +138,7 @@ func ThinkingLevelDescription(level ThinkingLevel) string {
 // ParseThinkingLevel converts a string to a ThinkingLevel, defaulting to ThinkingOff.
 func ParseThinkingLevel(s string) ThinkingLevel {
 	switch ThinkingLevel(s) {
-	case ThinkingMinimal, ThinkingLow, ThinkingMedium, ThinkingHigh:
+	case ThinkingNone, ThinkingMinimal, ThinkingLow, ThinkingMedium, ThinkingHigh:
 		return ThinkingLevel(s)
 	default:
 		return ThinkingOff
@@ -251,6 +256,11 @@ func CreateProvider(ctx context.Context, config *ProviderConfig) (*ProviderResul
 	// via CLI flag or global config.
 	ApplyModelSettings(config, modelInfo)
 
+	// Auto-raise MaxTokens toward the model's known output ceiling when the
+	// user hasn't explicitly set --max-tokens and no per-model override
+	// applied. Runs after ApplyModelSettings so explicit modelSettings win.
+	rightSizeMaxTokens(config, modelInfo)
+
 	// Create the base provider
 	var result *ProviderResult
 	var createErr error
@@ -295,9 +305,18 @@ func CreateProvider(ctx context.Context, config *ProviderConfig) (*ProviderResul
 			// 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
+			//
+			// For OpenAI Responses API models, we skip merging entirely because
+			// ResponsesProviderOptions and ProviderOptions are incompatible types.
+			skipMerge := false
+			if provider == "openai" && openai.IsResponsesModel(modelName) {
+				skipMerge = true
+			}
+			if !skipMerge {
+				for k, v := range cacheOpts {
+					if _, exists := result.ProviderOptions[k]; !exists {
+						result.ProviderOptions[k] = v
+					}
 				}
 			}
 		}
@@ -489,6 +508,37 @@ func validateModelConfig(config *ProviderConfig, modelInfo *ModelInfo) {
 	}
 }
 
+// defaultRightSizeCap bounds auto-raised MaxTokens so that we don't silently
+// allocate enormous output budgets for models with very high ceilings (e.g.
+// Devstral at 262144, Mistral at 128000). Users who genuinely want more can
+// pass --max-tokens explicitly or set modelSettings[...].maxTokens in config.
+const defaultRightSizeCap = 32768
+
+// rightSizeMaxTokens raises config.MaxTokens toward the model's known output
+// ceiling when:
+//   - the user has not explicitly set --max-tokens (or the KIT_MAX_TOKENS env
+//     var, or the top-level max-tokens key in config.yaml), AND
+//   - no per-model override already bumped MaxTokens (ApplyModelSettings runs
+//     before this function), AND
+//   - modelInfo.Limit.Output is known and larger than the current MaxTokens.
+//
+// The raised value is capped at defaultRightSizeCap to keep accidental
+// allocations reasonable on very-large-output models. This prevents the
+// common "ghost" where the agent's reply is silently truncated at the 8192
+// default even though the selected model supports 64k or 262k output tokens.
+func rightSizeMaxTokens(config *ProviderConfig, modelInfo *ModelInfo) {
+	if modelInfo == nil || modelInfo.Limit.Output <= 0 {
+		return
+	}
+	if isExplicitlySet("max-tokens") {
+		return
+	}
+	target := min(modelInfo.Limit.Output, defaultRightSizeCap)
+	if config.MaxTokens < target {
+		config.MaxTokens = target
+	}
+}
+
 // clearConflictingAnthropicSamplingParams ensures that temperature and top_p are
 // not both sent to the Anthropic API, which rejects requests containing both.
 // When both are set (typically from defaults), top_p is cleared so that
@@ -535,6 +585,8 @@ func buildOpenAIProviderOptions(config *ProviderConfig, modelName string) fantas
 // Returns nil for ThinkingOff (use the model's default).
 func thinkingLevelToReasoningEffort(level ThinkingLevel) *openai.ReasoningEffort {
 	switch level {
+	case ThinkingNone:
+		return new(openai.ReasoningEffortNone)
 	case ThinkingMinimal:
 		return new(openai.ReasoningEffortMinimal)
 	case ThinkingLow:
@@ -548,6 +600,56 @@ func thinkingLevelToReasoningEffort(level ThinkingLevel) *openai.ReasoningEffort
 	}
 }
 
+// IsValidThinkingLevelForModel checks if a thinking level is valid for the given
+// model. Some OpenAI models like gpt-5.4 don't support "minimal" and require
+// "none" instead.
+func IsValidThinkingLevelForModel(level ThinkingLevel, modelName string) bool {
+	if level == ThinkingOff {
+		return true
+	}
+
+	// Check if this is an OpenAI model that doesn't support "minimal"
+	// gpt-5.4 and newer gpt-5.x models use "none" instead of "minimal"
+	if level == ThinkingMinimal {
+		if strings.Contains(modelName, "gpt-5.4") ||
+			strings.Contains(modelName, "gpt-5-pro") ||
+			strings.Contains(modelName, "gpt-5-chat") {
+			return false
+		}
+	}
+
+	// Check if this is an OpenAI model that doesn't support "none"
+	// Older gpt-5 models only support "minimal", not "none"
+	if level == ThinkingNone {
+		if strings.Contains(modelName, "gpt-5") &&
+			!strings.Contains(modelName, "gpt-5.4") &&
+			!strings.Contains(modelName, "gpt-5-pro") &&
+			!strings.Contains(modelName, "gpt-5-chat") {
+			// Older gpt-5 models might not support "none"
+			// They only added "none" support in newer versions
+			return false
+		}
+	}
+
+	// All other levels are generally valid for reasoning models
+	return true
+}
+
+// SuggestThinkingLevelFallback returns a recommended fallback level when the
+// requested level is not valid for the model. Returns ThinkingOff if no
+// suitable fallback exists.
+func SuggestThinkingLevelFallback(level ThinkingLevel, modelName string) ThinkingLevel {
+	if level == ThinkingMinimal && !IsValidThinkingLevelForModel(level, modelName) {
+		// For models that don't support "minimal", suggest "none" (~same token budget)
+		return ThinkingNone
+	}
+	if level == ThinkingNone && !IsValidThinkingLevelForModel(level, modelName) {
+		// For models that don't support "none", suggest "minimal" (~same token budget)
+		return ThinkingMinimal
+	}
+	return ThinkingOff
+}
+
 // buildAnthropicProviderOptions returns fantasy.ProviderOptions configured for
 // Anthropic models with extended thinking. When thinking is enabled, it sets
 // SendReasoning to true and configures the thinking budget. For thinking-off

internal/models/registry.go

@@ -4,6 +4,7 @@ import (
 	_ "embed"
 	"encoding/json"
 	"fmt"
+	"maps"
 	"os"
 	"strings"
 
@@ -111,13 +112,30 @@ func NewModelsRegistry() *ModelsRegistry {
 }
 
 // buildFromModelsDB converts models.dev provider data into our internal format.
-// It tries the on-disk cache first and falls back to the embedded database.
+// It starts from the compile-time embedded database and merges on-disk cached
+// data from `kit update-models` on top. Cached provider metadata replaces
+// embedded metadata, and model entries are merged with cached models taking
+// precedence. This means newly synced models are available while embedded
+// models that haven't been synced yet are still reachable.
 func buildFromModelsDB() map[string]ProviderInfo {
-	// Try cached data first (from `kit update-models`)
-	dbProviders, _ := LoadCachedProviders()
-	if len(dbProviders) == 0 {
-		// Fall back to compile-time embedded data
-		dbProviders = loadEmbeddedProviders()
+	// Start with compile-time embedded data as the base.
+	dbProviders := loadEmbeddedProviders()
+	if dbProviders == nil {
+		dbProviders = make(ModelsDBProviders)
+	}
+
+	// Merge on-disk cached data on top (cached takes precedence).
+	if cached, _ := LoadCachedProviders(); len(cached) > 0 {
+		for providerID, cp := range cached {
+			if existing, ok := dbProviders[providerID]; ok {
+				// Merge models: embedded base + cached overrides.
+				mergedModels := make(map[string]modelsDBModel, len(existing.Models)+len(cp.Models))
+				maps.Copy(mergedModels, existing.Models)
+				maps.Copy(mergedModels, cp.Models)
+				cp.Models = mergedModels
+			}
+			dbProviders[providerID] = cp
+		}
 	}
 
 	providers := make(map[string]ProviderInfo, len(dbProviders))

internal/models/rightsize_test.go

@@ -0,0 +1,148 @@
+package models
+
+import (
+	"testing"
+
+	"github.com/spf13/pflag"
+	"github.com/spf13/viper"
+)
+
+// bindMaxTokensFlag wires a fresh pflag-backed "max-tokens" key into viper so
+// isExplicitlySet behaves the same way it does in production. Returns a
+// cleanup function that removes the binding so sibling tests see a clean
+// state.
+func bindMaxTokensFlag(t *testing.T, args []string) func() {
+	t.Helper()
+	fs := pflag.NewFlagSet("test", pflag.ContinueOnError)
+	fs.Int("max-tokens", 8192, "")
+	if err := viper.BindPFlag("max-tokens", fs.Lookup("max-tokens")); err != nil {
+		t.Fatalf("BindPFlag: %v", err)
+	}
+	if err := fs.Parse(args); err != nil {
+		t.Fatalf("fs.Parse: %v", err)
+	}
+	return func() {
+		viper.Reset()
+	}
+}
+
+func TestRightSizeMaxTokens_RaisesWhenBelowCeiling(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil) // no args → flag.Changed = false
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 8192}
+	modelInfo := &ModelInfo{
+		ID:    "claude-sonnet-4-5",
+		Limit: Limit{Context: 200000, Output: 64000},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != 32768 {
+		t.Errorf("expected MaxTokens raised to defaultRightSizeCap (32768), got %d", config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_CapsAtDefaultRightSizeCap(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil)
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 8192}
+	// Mistral Devstral has 262144 output — we should still cap at 32768.
+	modelInfo := &ModelInfo{
+		ID:    "devstral-medium-latest",
+		Limit: Limit{Context: 262144, Output: 262144},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != defaultRightSizeCap {
+		t.Errorf("expected MaxTokens capped at %d, got %d", defaultRightSizeCap, config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_UsesExactOutputWhenBelowCap(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil)
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 4096}
+	// Model with output limit smaller than the cap.
+	modelInfo := &ModelInfo{
+		ID:    "gpt-4",
+		Limit: Limit{Context: 8192, Output: 8192},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != 8192 {
+		t.Errorf("expected MaxTokens raised to model output ceiling (8192), got %d", config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_DoesNotLowerCurrentValue(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil)
+	defer cleanup()
+
+	// User (via per-model settings, applied earlier) already bumped MaxTokens
+	// above the cap — we must not clobber their choice.
+	config := &ProviderConfig{MaxTokens: 100000}
+	modelInfo := &ModelInfo{
+		ID:    "devstral-medium-latest",
+		Limit: Limit{Context: 262144, Output: 262144},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != 100000 {
+		t.Errorf("expected MaxTokens preserved at 100000, got %d", config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_RespectsExplicitFlag(t *testing.T) {
+	// Simulate `--max-tokens 4096` on the command line.
+	cleanup := bindMaxTokensFlag(t, []string{"--max-tokens", "4096"})
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 4096}
+	modelInfo := &ModelInfo{
+		ID:    "claude-sonnet-4-5",
+		Limit: Limit{Context: 200000, Output: 64000},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != 4096 {
+		t.Errorf("expected explicit --max-tokens to be preserved (4096), got %d", config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_NilModelInfo(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil)
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 8192}
+	// Custom model / Ollama / unknown provider → no model info.
+	rightSizeMaxTokens(config, nil)
+
+	if config.MaxTokens != 8192 {
+		t.Errorf("expected MaxTokens unchanged with nil modelInfo, got %d", config.MaxTokens)
+	}
+}
+
+func TestRightSizeMaxTokens_ZeroOutputLimit(t *testing.T) {
+	cleanup := bindMaxTokensFlag(t, nil)
+	defer cleanup()
+
+	config := &ProviderConfig{MaxTokens: 8192}
+	// Model present in catalog but with no known output limit.
+	modelInfo := &ModelInfo{
+		ID:    "unknown-model",
+		Limit: Limit{Context: 0, Output: 0},
+	}
+
+	rightSizeMaxTokens(config, modelInfo)
+
+	if config.MaxTokens != 8192 {
+		t.Errorf("expected MaxTokens unchanged with zero output limit, got %d", config.MaxTokens)
+	}
+}

internal/session/compaction_cycle_test.go

@@ -0,0 +1,66 @@
+package session
+
+import (
+	"testing"
+
+	"github.com/mark3labs/kit/internal/message"
+)
+
+// TestCompactionParentCycleRegression tests that after multiple compactions,
+// newly appended messages always have a valid parent chain and BuildContext
+// returns the correct messages.
+func TestCompactionParentCycleRegression(t *testing.T) {
+	tm := InMemoryTreeSession("/test")
+
+	// Simulate a long conversation with multiple compactions.
+	msg1, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg1"}}})
+	msg2, _ := tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg2"}}})
+
+	// First compaction
+	comp1, _ := tm.AppendCompaction("Summary 1", msg1, 1000, 500, 1, []string{}, []string{})
+
+	msg3, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg3"}}})
+	msg4, _ := tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg4"}}})
+
+	// Second compaction
+	comp2, _ := tm.AppendCompaction("Summary 2", msg3, 1000, 500, 1, []string{}, []string{})
+
+	msg5, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg5"}}})
+	msg6, _ := tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg6"}}})
+
+	// Verify parent chain integrity
+	for _, id := range []string{msg1, msg2, comp1, msg3, msg4, comp2, msg5, msg6} {
+		entry := tm.GetEntry(id)
+		if entry == nil {
+			t.Fatalf("entry %s not found in index", id)
+		}
+	}
+
+	// Walk parent chain from msg6 — must reach root without cycles
+	visited := make(map[string]bool)
+	current := msg6
+	for current != "" {
+		if visited[current] {
+			t.Fatalf("cycle detected at entry %s", current)
+		}
+		visited[current] = true
+		entry := tm.GetEntry(current)
+		if entry == nil {
+			t.Fatalf("entry %s missing from index during parent walk", current)
+		}
+		parent := ""
+		switch e := entry.(type) {
+		case *MessageEntry:
+			parent = e.ParentID
+		case *CompactionEntry:
+			parent = e.ParentID
+		}
+		current = parent
+	}
+
+	// BuildContext should return: Summary2 + msg6 + msg5 + msg3 + msg4 = 5 messages
+	msgs, _, _ := tm.BuildContext()
+	if len(msgs) != 5 {
+		t.Fatalf("expected 5 messages, got %d: %+v", len(msgs), msgs)
+	}
+}

internal/session/session_dir_test.go

@@ -0,0 +1,70 @@
+package session
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestEncodeCwdForDir verifies the working-directory → session-directory
+// name encoding strips characters that are illegal on Windows (notably the
+// drive-letter colon, see issue #18) while preserving the previous output
+// for the typical Unix paths.
+func TestEncodeCwdForDir(t *testing.T) {
+	tests := []struct {
+		name string
+		cwd  string
+		want string
+	}{
+		{
+			name: "unix absolute path",
+			cwd:  "/home/user/proj",
+			want: "home--user--proj",
+		},
+		{
+			name: "unix relative path",
+			cwd:  "proj/sub",
+			want: "proj--sub",
+		},
+		{
+			name: "windows drive root",
+			cwd:  `C:\test`,
+			want: "C--test",
+		},
+		{
+			name: "windows nested path",
+			cwd:  `C:\Users\User\code`,
+			want: "C--Users--User--code",
+		},
+		{
+			name: "windows secondary drive",
+			cwd:  `S:\work\repo`,
+			want: "S--work--repo",
+		},
+		{
+			name: "windows mixed separators",
+			cwd:  `C:\Users/User\code`,
+			want: "C--Users--User--code",
+		},
+		{
+			name: "windows other illegal chars stripped",
+			cwd:  `C:\a<b>c|d?e*f"g`,
+			want: "C--abcdefg",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := encodeCwdForDir(tc.cwd)
+			if got != tc.want {
+				t.Errorf("encodeCwdForDir(%q) = %q, want %q", tc.cwd, got, tc.want)
+			}
+			// Encoded directory must never contain characters that are
+			// illegal in Windows directory names.
+			for _, bad := range []string{":", "<", ">", "\"", "|", "?", "*", "\\", "/"} {
+				if strings.Contains(got, bad) {
+					t.Errorf("encodeCwdForDir(%q) = %q contains illegal char %q", tc.cwd, got, bad)
+				}
+			}
+		})
+	}
+}

internal/session/tree_cycle_test.go

@@ -0,0 +1,109 @@
+package session
+
+import (
+	"testing"
+
+	"github.com/mark3labs/kit/internal/message"
+)
+
+// TestDetectCycleWithCorruptedParentChain tests that cycle detection works
+// when a corrupted session has circular parent references.
+func TestDetectCycleWithCorruptedParentChain(t *testing.T) {
+	tm := InMemoryTreeSession("/test")
+
+	// Create normal chain: msg1 -> msg2 -> msg3
+	id1, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg1"}}})
+	_, _ = tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg2"}}})
+	id3, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg3"}}})
+
+	// Simulate corruption: manually set msg1's parent to msg3, creating cycle
+	// This simulates the condition seen in the user's session
+	for _, entry := range tm.entries {
+		if e, ok := entry.(*MessageEntry); ok && e.ID == id1 {
+			e.ParentID = id3 // Create cycle: msg1 -> msg3 -> ... -> msg1
+			break
+		}
+	}
+
+	// DetectCycle should find the cycle
+	// The cycle is: id1 -> id3 -> id2 -> id1
+	// So detecting from id3 should find id1 as the repeat
+	cycle, entry := tm.DetectCycle(id3)
+	if !cycle {
+		t.Fatal("expected to detect cycle, but none found")
+	}
+	// The cycle entry could be id1 or id3 depending on where we start
+	if entry != id1 && entry != id3 {
+		t.Fatalf("expected cycle at %s or %s, got %s", id1, id3, entry)
+	}
+
+	// BuildContext should still work (it has its own cycle detection)
+	// but will truncate at the cycle point
+	msgs, _, _ := tm.BuildContext()
+	if len(msgs) == 0 {
+		t.Fatal("BuildContext returned no messages")
+	}
+}
+
+// TestAppendMessageRejectsInvalidParent tests that AppendMessage rejects
+// appending when the current leaf has a broken parent chain.
+func TestAppendMessageRejectsInvalidParent(t *testing.T) {
+	tm := InMemoryTreeSession("/test")
+
+	// Create normal message
+	id1, err := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg1"}}})
+	if err != nil {
+		t.Fatalf("failed to append msg1: %v", err)
+	}
+
+	// Simulate corruption: set leafID to a non-existent ID
+	tm.leafID = "non-existent-id"
+
+	// Next append should fail validation
+	_, err = tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg2"}}})
+	if err == nil {
+		t.Fatal("expected error when appending with invalid leafID, got nil")
+	}
+
+	// Restore valid leafID
+	tm.leafID = id1
+
+	// Append should succeed now
+	_, err = tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg3"}}})
+	if err != nil {
+		t.Fatalf("failed to append msg3 after restoring leafID: %v", err)
+	}
+}
+
+// TestBuildContextHandlesCycleGracefully tests that BuildContext handles
+// cycles gracefully by truncating the branch.
+func TestBuildContextHandlesCycleGracefully(t *testing.T) {
+	tm := InMemoryTreeSession("/test")
+
+	// Create messages
+	id1, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg1"}}})
+	_, _ = tm.AppendMessage(message.Message{Role: message.RoleAssistant, Parts: []message.ContentPart{message.TextContent{Text: "msg2"}}})
+	id3, _ := tm.AppendMessage(message.Message{Role: message.RoleUser, Parts: []message.ContentPart{message.TextContent{Text: "msg3"}}})
+
+	// Verify normal case works
+	msgs, _, _ := tm.BuildContext()
+	if len(msgs) != 3 {
+		t.Fatalf("expected 3 messages, got %d", len(msgs))
+	}
+
+	// Simulate cycle: set msg1's parent to msg3
+	for _, entry := range tm.entries {
+		if e, ok := entry.(*MessageEntry); ok && e.ID == id1 {
+			e.ParentID = id3
+			break
+		}
+	}
+
+	// BuildContext should handle cycle gracefully (getBranchLocked has cycle detection)
+	msgs, _, _ = tm.BuildContext()
+	// Should only include messages from the cycle: msg3, msg2, msg1
+	// (msg3 is leaf, walks to msg2 -> msg1 -> msg3 (cycle detected, stops))
+	if len(msgs) != 3 {
+		t.Fatalf("expected 3 messages in cycle case, got %d: %+v", len(msgs), msgs)
+	}
+}

internal/session/tree_manager.go

@@ -63,6 +63,11 @@ type TreeManager struct {
 
 	// file is the open file handle for appending entries. Nil for in-memory.
 	file *os.File
+
+	// writer is a buffered writer wrapping file. Writes go through this
+	// buffer and are flushed to disk at explicit sync points (after each
+	// public Append* call, in Close, etc.) to reduce syscall overhead.
+	writer *bufio.Writer
 }
 
 // --- Constructors ---
@@ -105,11 +110,16 @@ func CreateTreeSession(cwd string) (*TreeManager, error) {
 		return nil, fmt.Errorf("failed to create session file: %w", err)
 	}
 	tm.file = f
+	tm.writer = bufio.NewWriter(f)
 
 	if err := tm.writeEntry(&header); err != nil {
 		_ = f.Close()
 		return nil, fmt.Errorf("failed to write session header: %w", err)
 	}
+	if err := tm.flushLocked(); err != nil {
+		_ = f.Close()
+		return nil, fmt.Errorf("failed to flush session header: %w", err)
+	}
 
 	return tm, nil
 }
@@ -150,6 +160,7 @@ func (tm *TreeManager) ForkToNewSession(cwd string, targetID string) (*TreeManag
 		return nil, fmt.Errorf("failed to recreate session file: %w", err)
 	}
 	newTm.file = f
+	newTm.writer = bufio.NewWriter(f)
 
 	if err := newTm.writeEntry(&newTm.header); err != nil {
 		_ = f.Close()
@@ -289,6 +300,12 @@ func (tm *TreeManager) ForkToNewSession(cwd string, targetID string) (*TreeManag
 		}
 	}
 
+	// Flush all buffered writes from the fork in a single syscall.
+	if err := newTm.flushLocked(); err != nil {
+		_ = f.Close()
+		return nil, fmt.Errorf("failed to flush forked session: %w", err)
+	}
+
 	// Set the leaf to the last entry in the new session.
 	newTm.leafID = prevNewID
 
@@ -365,12 +382,16 @@ func OpenTreeSession(path string) (*TreeManager, error) {
 		tm.leafID = tm.EntryID(tm.entries[len(tm.entries)-1])
 	}
 
+	// Validate tree integrity and log diagnostics
+	tm.LogTreeDiagnostics()
+
 	// Open file for appending.
 	f, err := os.OpenFile(path, os.O_WRONLY|os.O_APPEND, 0644)
 	if err != nil {
 		return nil, fmt.Errorf("failed to open session file for append: %w", err)
 	}
 	tm.file = f
+	tm.writer = bufio.NewWriter(f)
 
 	return tm, nil
 }
@@ -410,6 +431,12 @@ func (tm *TreeManager) AppendMessage(msg message.Message) (string, error) {
 	tm.mu.Lock()
 	defer tm.mu.Unlock()
 
+	// Validate parent chain before appending to detect/prevent cycles
+	// that could be caused by external file corruption or race conditions.
+	if err := tm.validateParentChainLocked(tm.leafID, ""); err != nil {
+		return "", fmt.Errorf("parent chain validation failed: %w", err)
+	}
+
 	entry, err := NewMessageEntry(tm.leafID, msg)
 	if err != nil {
 		return "", err
@@ -418,6 +445,9 @@ func (tm *TreeManager) AppendMessage(msg message.Message) (string, error) {
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush message: %w", err)
+	}
 
 	tm.leafID = entry.ID
 	return entry.ID, nil
@@ -442,6 +472,9 @@ func (tm *TreeManager) AppendModelChange(provider, modelID string) (string, erro
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush model change: %w", err)
+	}
 
 	tm.leafID = entry.ID
 	return entry.ID, nil
@@ -456,6 +489,9 @@ func (tm *TreeManager) AppendBranchSummary(fromID, summary string) (string, erro
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush branch summary: %w", err)
+	}
 
 	tm.leafID = entry.ID
 	return entry.ID, nil
@@ -470,6 +506,9 @@ func (tm *TreeManager) AppendLabel(targetID, label string) (string, error) {
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush label: %w", err)
+	}
 
 	tm.labels[targetID] = label
 	tm.leafID = entry.ID
@@ -485,6 +524,9 @@ func (tm *TreeManager) AppendSessionInfo(name string) (string, error) {
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush session info: %w", err)
+	}
 
 	tm.sessionName = name
 	tm.leafID = entry.ID
@@ -501,6 +543,9 @@ func (tm *TreeManager) AppendExtensionData(extType, data string) (string, error)
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush extension data: %w", err)
+	}
 
 	tm.leafID = entry.ID
 	return entry.ID, nil
@@ -518,6 +563,13 @@ func (tm *TreeManager) AppendCompaction(summary, firstKeptEntryID string, tokens
 	tm.mu.Lock()
 	defer tm.mu.Unlock()
 
+	// Validate that firstKeptEntryID exists if provided
+	if firstKeptEntryID != "" {
+		if _, ok := tm.index[firstKeptEntryID]; !ok {
+			return "", fmt.Errorf("first kept entry %q does not exist", firstKeptEntryID)
+		}
+	}
+
 	// The compaction entry has no parent, making it a new "root" for the
 	// post-compaction branch. This ensures old compacted messages are not
 	// traversed when walking from the current leaf.
@@ -525,6 +577,9 @@ func (tm *TreeManager) AppendCompaction(summary, firstKeptEntryID string, tokens
 	if err := tm.appendAndPersist(entry); err != nil {
 		return "", err
 	}
+	if err := tm.flushLocked(); err != nil {
+		return "", fmt.Errorf("failed to flush compaction: %w", err)
+	}
 
 	tm.leafID = entry.ID
 	return entry.ID, nil
@@ -910,11 +965,31 @@ func (tm *TreeManager) IsEmpty() bool {
 	return tm.MessageCount() == 0
 }
 
-// Close closes the underlying file handle.
+// Flush writes any buffered data to the underlying file.
+func (tm *TreeManager) Flush() error {
+	tm.mu.Lock()
+	defer tm.mu.Unlock()
+	return tm.flushLocked()
+}
+
+// flushLocked writes buffered data to disk. Caller must hold the lock.
+func (tm *TreeManager) flushLocked() error {
+	if tm.writer != nil {
+		return tm.writer.Flush()
+	}
+	return nil
+}
+
+// Close flushes any buffered writes and closes the underlying file handle.
 func (tm *TreeManager) Close() error {
 	tm.mu.Lock()
 	defer tm.mu.Unlock()
 	if tm.file != nil {
+		// Flush buffered data before closing.
+		if tm.writer != nil {
+			_ = tm.writer.Flush()
+			tm.writer = nil
+		}
 		err := tm.file.Close()
 		tm.file = nil
 		return err
@@ -1074,13 +1149,22 @@ func (tm *TreeManager) GetLastCompaction() *CompactionEntry {
 
 // AddLLMMessages appends multiple LLM messages as entries. This is
 // used when syncing from the agent's ConversationMessages after a step.
+// All entries are buffered and flushed to disk in a single batch.
 func (tm *TreeManager) AddLLMMessages(msgs []fantasy.Message) error {
+	tm.mu.Lock()
+	defer tm.mu.Unlock()
+
 	for _, msg := range msgs {
-		if _, err := tm.AppendLLMMessage(msg); err != nil {
+		entry, err := NewMessageEntry(tm.leafID, message.FromLLMMessage(msg))
+		if err != nil {
 			return err
 		}
+		if err := tm.appendAndPersist(entry); err != nil {
+			return err
+		}
+		tm.leafID = entry.ID
 	}
-	return nil
+	return tm.flushLocked()
 }
 
 // Deprecated: Use AddLLMMessages instead.
@@ -1132,12 +1216,20 @@ func (tm *TreeManager) appendAndPersist(entry any) error {
 	return nil
 }
 
-// writeEntry serializes an entry and appends it as a line to the file.
+// writeEntry serializes an entry and appends it to the buffered writer.
+// The data is not flushed to disk until flushLocked is called.
 func (tm *TreeManager) writeEntry(entry any) error {
 	data, err := json.Marshal(entry)
 	if err != nil {
 		return fmt.Errorf("failed to marshal entry: %w", err)
 	}
+	if tm.writer != nil {
+		if _, err := tm.writer.Write(data); err != nil {
+			return err
+		}
+		return tm.writer.WriteByte('\n')
+	}
+	// Fallback for direct file writes (shouldn't happen in normal flow).
 	data = append(data, '\n')
 	_, err = tm.file.Write(data)
 	return err
@@ -1213,12 +1305,32 @@ func (tm *TreeManager) getBranchLocked(fromID string) []any {
 }
 
 // buildTreeNode recursively builds a TreeNode from an entry ID.
+// It includes a depth limit to prevent infinite recursion in case of
+// corrupted parent-child relationships.
 func (tm *TreeManager) buildTreeNode(id string) *TreeNode {
+	return tm.buildTreeNodeDepth(id, 0, make(map[string]bool))
+}
+
+// buildTreeNodeDepth is the internal implementation with depth tracking.
+func (tm *TreeManager) buildTreeNodeDepth(id string, depth int, visited map[string]bool) *TreeNode {
+	const maxDepth = 1000
+	if depth > maxDepth {
+		// Cycle or extremely deep tree detected, stop recursing
+		return nil
+	}
+	if visited[id] {
+		// Cycle detected, stop recursing
+		return nil
+	}
+
 	entry, ok := tm.index[id]
 	if !ok {
 		return nil
 	}
 
+	visited[id] = true
+	defer delete(visited, id)
+
 	node := &TreeNode{
 		Entry:    entry,
 		ID:       id,
@@ -1226,7 +1338,7 @@ func (tm *TreeManager) buildTreeNode(id string) *TreeNode {
 	}
 
 	for _, childID := range tm.childIndex[id] {
-		child := tm.buildTreeNode(childID)
+		child := tm.buildTreeNodeDepth(childID, depth+1, visited)
 		if child != nil {
 			node.Children = append(node.Children, child)
 		}
@@ -1238,15 +1350,44 @@ func (tm *TreeManager) buildTreeNode(id string) *TreeNode {
 // --- Path conventions ---
 
 // DefaultSessionDir returns the default session storage directory for a cwd.
-// Convention: ~/.kit/sessions/--<cwd-path>--/
+// Convention: ~/.kit/sessions/<encoded-cwd>, where path separators are
+// encoded as "--" with no leading or trailing dashes — e.g.
+// /home/user/proj becomes home--user--proj. See encodeCwdForDir for the
+// full encoding rules (including Windows path handling).
 func DefaultSessionDir(cwd string) string {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		home = "."
 	}
-	// Convert path separators to double dashes.
-	safeCwd := strings.ReplaceAll(cwd, string(filepath.Separator), "--")
+	return filepath.Join(home, ".kit", "sessions", encodeCwdForDir(cwd))
+}
+
+// encodeCwdForDir converts a working-directory path into a single, filesystem-
+// safe directory name. Path separators are replaced with double dashes and
+// characters that are illegal in Windows directory names — most importantly
+// the colon that follows the drive letter (e.g. `C:\foo` → `C--foo`) — are
+// stripped. The result is identical to the previous Unix-only encoding for
+// paths that do not contain such characters, so existing session directories
+// are preserved.
+func encodeCwdForDir(cwd string) string {
+	// Convert both `/` and `\` to double dashes so encoding is stable across
+	// platforms and remains correct on Windows where `filepath.Separator`
+	// would otherwise miss forward-slash style paths.
+	safeCwd := strings.ReplaceAll(cwd, "\\", "--")
+	safeCwd = strings.ReplaceAll(safeCwd, "/", "--")
 	// Remove leading separator replacement.
 	safeCwd = strings.TrimPrefix(safeCwd, "--")
-	return filepath.Join(home, ".kit", "sessions", safeCwd)
+	// Strip characters that are illegal in directory names on Windows
+	// (`< > : " | ? *`). On Unix these characters are legal but rare in
+	// practice; stripping them keeps the encoding portable.
+	replacer := strings.NewReplacer(
+		":", "",
+		"<", "",
+		">", "",
+		"\"", "",
+		"|", "",
+		"?", "",
+		"*", "",
+	)
+	return replacer.Replace(safeCwd)
 }

internal/session/tree_validation.go

@@ -0,0 +1,143 @@
+package session
+
+import (
+	"fmt"
+	"log"
+)
+
+// ValidateParentChain checks that the parent ID points to an existing entry
+// and that appending this entry would not create a cycle. This should be called
+// before appending any entry to the tree.
+// Returns an error if the parent is invalid or would create a cycle.
+func (tm *TreeManager) ValidateParentChain(parentID string, newEntryID string) error {
+	if parentID == "" {
+		// Empty parent is valid (root entry)
+		return nil
+	}
+
+	// Check that parent exists
+	if _, ok := tm.index[parentID]; !ok {
+		return fmt.Errorf("parent entry %q does not exist in index", parentID)
+	}
+
+	// Check that we're not creating a cycle by walking up the parent chain
+	// from parentID and ensuring we don't hit newEntryID (or any node that
+	// has newEntryID as an ancestor, but since newEntryID is new, just check
+	// that parentID isn't newEntryID, which it can't be since we check existence)
+	visited := make(map[string]bool)
+	current := parentID
+	for current != "" {
+		if visited[current] {
+			return fmt.Errorf("existing cycle detected at entry %q", current)
+		}
+		visited[current] = true
+
+		// Safety check: if somehow we reach the new entry ID, that's a cycle
+		if current == newEntryID {
+			return fmt.Errorf("would create cycle: entry %q cannot be its own ancestor", newEntryID)
+		}
+
+		entry, ok := tm.index[current]
+		if !ok {
+			return fmt.Errorf("broken parent chain: entry %q not found", current)
+		}
+		current = tm.entryParentID(entry)
+	}
+
+	return nil
+}
+
+// DetectCycle walks the parent chain from the given entry ID and returns true
+// if a cycle is detected. This is used for diagnostics.
+func (tm *TreeManager) DetectCycle(fromID string) (cycleDetected bool, cycleEntry string) {
+	visited := make(map[string]bool)
+	current := fromID
+	for current != "" {
+		if visited[current] {
+			return true, current
+		}
+		visited[current] = true
+		entry, ok := tm.index[current]
+		if !ok {
+			return false, ""
+		}
+		current = tm.entryParentID(entry)
+	}
+	return false, ""
+}
+
+// LogTreeDiagnostics logs information about the tree structure for debugging.
+// Call this after OpenTreeSession or when anomalies are detected.
+func (tm *TreeManager) LogTreeDiagnostics() {
+	tm.mu.RLock()
+	defer tm.mu.RUnlock()
+
+	log.Printf("[TreeManager] Entry count: %d, Leaf ID: %s", len(tm.entries), tm.leafID)
+
+	// Check for cycles from leaf
+	if tm.leafID != "" {
+		if cycle, entry := tm.detectCycleLocked(tm.leafID); cycle {
+			log.Printf("[TreeManager] WARNING: Cycle detected in tree at entry %s", entry)
+		}
+	}
+
+	// Count entries by type
+	counts := make(map[EntryType]int)
+	for _, entry := range tm.entries {
+		var et EntryType
+		switch e := entry.(type) {
+		case *MessageEntry:
+			et = e.Type
+		case *ModelChangeEntry:
+			et = e.Type
+		case *BranchSummaryEntry:
+			et = e.Type
+		case *LabelEntry:
+			et = e.Type
+		case *SessionInfoEntry:
+			et = e.Type
+		case *ExtensionDataEntry:
+			et = e.Type
+		case *CompactionEntry:
+			et = e.Type
+		default:
+			et = "unknown"
+		}
+		counts[et]++
+	}
+	log.Printf("[TreeManager] Entry types: %+v", counts)
+}
+
+// detectCycleLocked is the internal version of DetectCycle (must hold read lock)
+func (tm *TreeManager) detectCycleLocked(fromID string) (bool, string) {
+	visited := make(map[string]bool)
+	current := fromID
+	for current != "" {
+		if visited[current] {
+			return true, current
+		}
+		visited[current] = true
+		entry, ok := tm.index[current]
+		if !ok {
+			return false, ""
+		}
+		current = tm.entryParentID(entry)
+	}
+	return false, ""
+}
+
+// validateParentChainLocked is the internal version used by append methods.
+// Must be called with the write lock held.
+func (tm *TreeManager) validateParentChainLocked(parentID string, newEntryID string) error {
+	if parentID == "" {
+		return nil
+	}
+	if _, ok := tm.index[parentID]; !ok {
+		return fmt.Errorf("parent entry %q does not exist", parentID)
+	}
+	// Check for existing cycles in the parent chain
+	if cycle, entry := tm.detectCycleLocked(parentID); cycle {
+		return fmt.Errorf("existing cycle detected at entry %q in parent chain", entry)
+	}
+	return nil
+}

internal/tools/connection_pool.go

@@ -47,6 +47,7 @@ type MCPConnection struct {
 	client       client.MCPClient
 	serverName   string
 	serverConfig config.MCPServerConfig
+	initResult   *mcp.InitializeResult // captured at handshake; nil before initialize
 	lastUsed     time.Time
 	isHealthy    bool
 	errorCount   int
@@ -243,10 +244,12 @@ func (p *MCPConnectionPool) performHealthCheck(ctx context.Context, conn *MCPCon
 
 // createConnection creates a new connection
 func (p *MCPConnectionPool) createConnection(ctx context.Context, serverName string, serverConfig config.MCPServerConfig) (*MCPConnection, error) {
+	oauthEnabled := p.oauthFlow != nil && !serverConfig.NoOAuth
+
 	mcpClient, err := p.createMCPClient(ctx, serverName, serverConfig)
 	if err != nil {
 		// SSE transport can return OAuth error during Start()
-		if p.oauthFlow != nil && IsOAuthError(err) {
+		if oauthEnabled && IsOAuthError(err) {
 			if flowErr := p.oauthFlow.RunAuthFlow(ctx, serverName, err); flowErr != nil {
 				return nil, fmt.Errorf("OAuth authorization failed: %w", flowErr)
 			}
@@ -260,15 +263,17 @@ func (p *MCPConnectionPool) createConnection(ctx context.Context, serverName str
 		}
 	}
 
-	if err := p.initializeClient(ctx, mcpClient); err != nil {
+	conn := &MCPConnection{}
+
+	if err := p.initializeClient(ctx, mcpClient, conn); err != nil {
 		// Streamable HTTP transport returns OAuth error during Initialize()
-		if p.oauthFlow != nil && IsOAuthError(err) {
+		if oauthEnabled && IsOAuthError(err) {
 			if flowErr := p.oauthFlow.RunAuthFlow(ctx, serverName, err); flowErr != nil {
 				_ = mcpClient.Close()
 				return nil, fmt.Errorf("OAuth authorization failed: %w", flowErr)
 			}
 			// Retry initialization after successful auth
-			if err := p.initializeClient(ctx, mcpClient); err != nil {
+			if err := p.initializeClient(ctx, mcpClient, conn); err != nil {
 				_ = mcpClient.Close()
 				return nil, err
 			}
@@ -278,15 +283,11 @@ func (p *MCPConnectionPool) createConnection(ctx context.Context, serverName str
 		}
 	}
 
-	conn := &MCPConnection{
-		client:       mcpClient,
-		serverName:   serverName,
-		serverConfig: serverConfig,
-		lastUsed:     time.Now(),
-		isHealthy:    true,
-		errorCount:   0,
-		lastError:    nil,
-	}
+	conn.client = mcpClient
+	conn.serverName = serverName
+	conn.serverConfig = serverConfig
+	conn.lastUsed = time.Now()
+	conn.isHealthy = true
 
 	if p.debugLogger != nil && p.debugLogger.IsDebugEnabled() {
 		p.debugLogger.LogDebug(fmt.Sprintf("[POOL] Created connection for %s", serverName))
@@ -363,11 +364,11 @@ func (p *MCPConnectionPool) createSSEClient(ctx context.Context, serverConfig co
 		}
 	}
 
-	// Enable OAuth for remote transports when an auth handler is configured.
-	// The OAuthConfig uses PKCE and the handler's redirect URI. If the server
-	// config provides a pre-registered ClientID (for servers that don't support
-	// dynamic client registration, e.g. GitHub), it is passed through directly.
-	if p.oauthFlow != nil {
+	// Enable OAuth for remote transports when an auth handler is configured
+	// and the server hasn't opted out via NoOAuth. Public MCP servers (e.g.
+	// PubMed) set NoOAuth to skip dynamic client registration and token
+	// exchange, which would otherwise fail with a 404.
+	if p.oauthFlow != nil && !serverConfig.NoOAuth {
 		tokenStore, tsErr := p.createTokenStore(serverConfig.URL)
 		if tsErr != nil {
 			return nil, fmt.Errorf("failed to create token store: %w", tsErr)
@@ -420,11 +421,9 @@ func (p *MCPConnectionPool) createStreamableClient(ctx context.Context, serverCo
 		}
 	}
 
-	// Enable OAuth for remote transports when an auth handler is configured.
-	// The OAuthConfig uses PKCE and the handler's redirect URI. If the server
-	// config provides a pre-registered ClientID (for servers that don't support
-	// dynamic client registration, e.g. GitHub), it is passed through directly.
-	if p.oauthFlow != nil {
+	// Enable OAuth for remote transports when an auth handler is configured
+	// and the server hasn't opted out via NoOAuth.
+	if p.oauthFlow != nil && !serverConfig.NoOAuth {
 		tokenStore, tsErr := p.createTokenStore(serverConfig.URL)
 		if tsErr != nil {
 			return nil, fmt.Errorf("failed to create token store: %w", tsErr)
@@ -484,8 +483,10 @@ func (p *MCPConnectionPool) createTokenStore(serverURL string) (transport.TokenS
 	return NewFileTokenStore(serverURL)
 }
 
-// initializeClient initializes the client
-func (p *MCPConnectionPool) initializeClient(ctx context.Context, client client.MCPClient) error {
+// initializeClient initializes the client and captures the server's
+// initialize result on the supplied connection so callers can later
+// inspect advertised capabilities (e.g. task support).
+func (p *MCPConnectionPool) initializeClient(ctx context.Context, c client.MCPClient, conn *MCPConnection) error {
 	initCtx, cancel := context.WithTimeout(ctx, 5*time.Minute)
 	defer cancel()
 
@@ -495,12 +496,21 @@ func (p *MCPConnectionPool) initializeClient(ctx context.Context, client client.
 		Name:    "kit",
 		Version: "1.0.0",
 	}
-	initRequest.Params.Capabilities = mcp.ClientCapabilities{}
+	// Advertise task support so servers may return CreateTaskResult for
+	// long-running tools/call requests instead of blocking the connection
+	// until completion. The client is responsible for polling tasks/get and
+	// tasks/result until the task reaches a terminal state.
+	initRequest.Params.Capabilities = mcp.ClientCapabilities{
+		Tasks: mcp.NewTasksCapability(),
+	}
 
-	_, err := client.Initialize(initCtx, initRequest)
+	initResult, err := c.Initialize(initCtx, initRequest)
 	if err != nil {
 		return fmt.Errorf("initialization timeout or failed: %w", err)
 	}
+	if conn != nil {
+		conn.initResult = initResult
+	}
 
 	if p.debugLogger != nil && p.debugLogger.IsDebugEnabled() {
 		p.debugLogger.LogDebug("[POOL] Initialized MCP client")
@@ -615,6 +625,54 @@ func (c *MCPConnection) ServerName() string {
 	return c.serverName
 }
 
+// InitializeResult returns the result captured from the server's initialize
+// response, or nil if the connection was created before initialize completed.
+// Callers can inspect ServerCapabilities.Tasks to discover task-related
+// capability advertisements.
+func (c *MCPConnection) InitializeResult() *mcp.InitializeResult {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	return c.initResult
+}
+
+// SupportsToolTasks reports whether the server advertised support for
+// task-augmented tools/call requests. Returns false when the connection has
+// not yet completed initialization or when the server omitted task
+// capabilities.
+func (c *MCPConnection) SupportsToolTasks() bool {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	return supportsToolTasksFromInit(c.initResult)
+}
+
+// supportsToolTasksFromInit reports whether the supplied InitializeResult
+// advertises task-augmented tools/call support. Extracted to a free function
+// for unit testing without standing up a connection.
+func supportsToolTasksFromInit(init *mcp.InitializeResult) bool {
+	if init == nil || init.Capabilities.Tasks == nil {
+		return false
+	}
+	req := init.Capabilities.Tasks.Requests
+	if req == nil || req.Tools == nil {
+		return false
+	}
+	return req.Tools.Call != nil
+}
+
+// ServerSupportsToolTasks reports whether the named server's connection
+// advertises task-augmented tools/call support. Returns false when no
+// connection exists for the server or when the server didn't advertise the
+// capability.
+func (p *MCPConnectionPool) ServerSupportsToolTasks(serverName string) bool {
+	p.mu.RLock()
+	conn, ok := p.connections[serverName]
+	p.mu.RUnlock()
+	if !ok {
+		return false
+	}
+	return conn.SupportsToolTasks()
+}
+
 // GetClients returns a map of all MCP clients currently in the pool.
 // The map keys are server names and values are the corresponding MCP client instances.
 // The returned map is a copy and modifications won't affect the pool.

internal/tools/mcp.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/base64"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"maps"
 	"slices"
@@ -13,6 +14,7 @@ import (
 	log "github.com/charmbracelet/log"
 
 	"github.com/mark3labs/kit/internal/config"
+	"github.com/mark3labs/mcp-go/client"
 	"github.com/mark3labs/mcp-go/mcp"
 )
 
@@ -141,6 +143,11 @@ type MCPToolManager struct {
 	debug             bool
 	debugLogger       DebugLogger
 
+	// taskCfg controls task-augmented tools/call execution. The zero value
+	// means: auto-detect server capability, no progress callback, default
+	// poll/timeout.
+	taskCfg MCPTaskConfig
+
 	// onServerLoaded, if non-nil, is called when each server finishes loading.
 	// Called with server name, tool count, and error (nil on success).
 	onServerLoaded func(serverName string, toolCount int, err error)
@@ -220,6 +227,21 @@ func (m *MCPToolManager) SetOnToolsChanged(cb func()) {
 	m.onToolsChanged = cb
 }
 
+// SetTaskConfig sets the task-augmented tools/call configuration. Call
+// this before LoadTools / AddServer if you want the per-server mode
+// override and progress handler to take effect for the very first call.
+// Subsequent calls replace the previous configuration wholesale.
+func (m *MCPToolManager) SetTaskConfig(cfg MCPTaskConfig) {
+	m.taskCfg = cfg
+}
+
+// TaskConfig returns the manager's current task-augmented tools/call
+// configuration. The zero value means: defer to per-server config and
+// auto-detected capability, with no progress callback and default polling.
+func (m *MCPToolManager) TaskConfig() MCPTaskConfig {
+	return m.taskCfg
+}
+
 // AddServer connects to a new MCP server at runtime and loads its tools.
 // The server's tools are immediately available to the agent after this call.
 // Returns the number of tools loaded from the server.
@@ -551,6 +573,14 @@ func (m *MCPToolManager) loadServerTools(ctx context.Context, serverName string,
 // checks, OAuth re-authorization, and connection error tracking.
 // The inputJSON parameter is the raw JSON arguments from the LLM.
 // Returns the result content, error flag, and any execution error.
+//
+// When the per-server TasksMode resolves to "always", or to "auto" and the
+// server advertised tasks/toolCalls capability during initialize, the call
+// is augmented with TaskParams. If the server elects to respond with a
+// CreateTaskResult the manager polls tasks/get / tasks/result until the
+// task reaches a terminal state, transparently presenting the final
+// CallToolResult-equivalent content to the agent layer. Context
+// cancellation triggers a best-effort tasks/cancel.
 func (m *MCPToolManager) ExecuteTool(ctx context.Context, prefixedName, inputJSON string) (*MCPToolResult, error) {
 	m.mu.Lock()
 	mapping, ok := m.toolMap[prefixedName]
@@ -582,49 +612,221 @@ func (m *MCPToolManager) ExecuteTool(ctx context.Context, prefixedName, inputJSO
 		return nil, fmt.Errorf("failed to get healthy connection from pool: %w", err)
 	}
 
-	callRequest := mcp.CallToolRequest{
-		Request: mcp.Request{
-			Method: "tools/call",
-		},
-		Params: mcp.CallToolParams{
-			Name:      mapping.originalName,
-			Arguments: arguments,
-		},
+	callParams := mcp.CallToolParams{
+		Name:      mapping.originalName,
+		Arguments: arguments,
 	}
 
-	// Call the MCP tool using the original (unprefixed) name
-	result, err := conn.client.CallTool(ctx, callRequest)
-	if err != nil {
-		// Handle OAuth re-authorization: token may have expired mid-session.
-		if m.connectionPool.oauthFlow != nil && IsOAuthError(err) {
-			if flowErr := m.connectionPool.oauthFlow.RunAuthFlow(ctx, mapping.serverName, err); flowErr != nil {
+	// Decide whether to augment the request with TaskParams. Modes:
+	//   never  — never augment (synchronous-only).
+	//   always — always augment, even without server capability.
+	//   auto   — augment only when the server advertised tasks/toolCalls.
+	mode := m.resolveTaskMode(mapping.serverName, mapping.serverConfig)
+	useTask := mode == MCPTaskModeAlways ||
+		(mode == MCPTaskModeAuto && conn.SupportsToolTasks())
+	if useTask {
+		var ttl *int64
+		if m.taskCfg.DefaultTTL > 0 {
+			ms := m.taskCfg.DefaultTTL.Milliseconds()
+			ttl = &ms
+		}
+		callParams.Task = &mcp.TaskParams{TTL: ttl}
+	}
+
+	// Synchronous fast path: no task augmentation. Use the upstream client
+	// helper which keeps content-block typing identical to historical
+	// behaviour.
+	if !useTask {
+		callRequest := mcp.CallToolRequest{
+			Request: mcp.Request{Method: "tools/call"},
+			Params:  callParams,
+		}
+		result, callErr := conn.client.CallTool(ctx, callRequest)
+		if callErr != nil {
+			if m.connectionPool.oauthFlow != nil && IsOAuthError(callErr) {
+				if flowErr := m.connectionPool.oauthFlow.RunAuthFlow(ctx, mapping.serverName, callErr); flowErr != nil {
+					return nil, fmt.Errorf("OAuth re-authorization failed for tool %s: %w", mapping.originalName, flowErr)
+				}
+				result, callErr = conn.client.CallTool(ctx, callRequest)
+				if callErr != nil {
+					m.connectionPool.HandleConnectionError(mapping.serverName, callErr)
+					return nil, fmt.Errorf("failed to call mcp tool after re-auth: %w", callErr)
+				}
+			} else {
+				m.connectionPool.HandleConnectionError(mapping.serverName, callErr)
+				return nil, fmt.Errorf("failed to call mcp tool: %w", callErr)
+			}
+		}
+		marshaledResult, mErr := json.Marshal(result)
+		if mErr != nil {
+			return nil, fmt.Errorf("failed to marshal mcp tool result: %w", mErr)
+		}
+		return &MCPToolResult{
+			Content: string(marshaledResult),
+			IsError: result.IsError,
+		}, nil
+	}
+
+	// Task-augmented path. Bypass the upstream CallTool helper because its
+	// ParseCallToolResult requires a "content" field that is absent from a
+	// CreateTaskResult.
+	rawClient, ok := conn.client.(*client.Client)
+	if !ok {
+		// Older client implementations — fall back to the synchronous shape.
+		callParams.Task = nil
+		callRequest := mcp.CallToolRequest{
+			Request: mcp.Request{Method: "tools/call"},
+			Params:  callParams,
+		}
+		result, callErr := conn.client.CallTool(ctx, callRequest)
+		if callErr != nil {
+			m.connectionPool.HandleConnectionError(mapping.serverName, callErr)
+			return nil, fmt.Errorf("failed to call mcp tool: %w", callErr)
+		}
+		marshaledResult, mErr := json.Marshal(result)
+		if mErr != nil {
+			return nil, fmt.Errorf("failed to marshal mcp tool result: %w", mErr)
+		}
+		return &MCPToolResult{Content: string(marshaledResult), IsError: result.IsError}, nil
+	}
+
+	callResult, taskResult, callErr := callToolWithTask(ctx, rawClient, callParams)
+	if callErr != nil {
+		if m.connectionPool.oauthFlow != nil && IsOAuthError(callErr) {
+			if flowErr := m.connectionPool.oauthFlow.RunAuthFlow(ctx, mapping.serverName, callErr); flowErr != nil {
 				return nil, fmt.Errorf("OAuth re-authorization failed for tool %s: %w", mapping.originalName, flowErr)
 			}
-			// Retry the tool call after successful re-auth.
-			result, err = conn.client.CallTool(ctx, callRequest)
-			if err != nil {
-				m.connectionPool.HandleConnectionError(mapping.serverName, err)
-				return nil, fmt.Errorf("failed to call mcp tool after re-auth: %w", err)
+			callResult, taskResult, callErr = callToolWithTask(ctx, rawClient, callParams)
+			if callErr != nil {
+				m.connectionPool.HandleConnectionError(mapping.serverName, callErr)
+				return nil, fmt.Errorf("failed to call mcp tool after re-auth: %w", callErr)
 			}
 		} else {
-			// Mark connection as unhealthy for automatic recovery
-			m.connectionPool.HandleConnectionError(mapping.serverName, err)
-			return nil, fmt.Errorf("failed to call mcp tool: %w", err)
+			m.connectionPool.HandleConnectionError(mapping.serverName, callErr)
+			return nil, fmt.Errorf("failed to call mcp tool: %w", callErr)
 		}
 	}
 
-	// Marshal the MCP result to JSON string
-	marshaledResult, err := json.Marshal(result)
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal mcp tool result: %w", err)
+	// Server chose to answer synchronously — same shape as the no-task path.
+	if callResult != nil {
+		marshaledResult, mErr := json.Marshal(callResult)
+		if mErr != nil {
+			return nil, fmt.Errorf("failed to marshal mcp tool result: %w", mErr)
+		}
+		return &MCPToolResult{
+			Content: string(marshaledResult),
+			IsError: callResult.IsError,
+		}, nil
 	}
 
+	// Asynchronous task path: poll until terminal, then return the result.
+	if taskResult == nil {
+		return nil, errors.New("mcp tools/call returned neither result nor task")
+	}
+	final, pollErr := pollTaskUntilTerminal(
+		ctx, rawClient, mapping.serverName, taskResult.Task,
+		m.taskCfg, m.taskCfg.Progress,
+	)
+	if pollErr != nil {
+		return nil, fmt.Errorf("task execution failed: %w", pollErr)
+	}
+
+	// Adapt TaskResultResult → CallToolResult for downstream JSON shape parity.
+	adapted := &mcp.CallToolResult{
+		Content:           final.Content,
+		StructuredContent: final.StructuredContent,
+		IsError:           final.IsError,
+	}
+	marshaledResult, mErr := json.Marshal(adapted)
+	if mErr != nil {
+		return nil, fmt.Errorf("failed to marshal mcp tool result: %w", mErr)
+	}
 	return &MCPToolResult{
 		Content: string(marshaledResult),
-		IsError: result.IsError,
+		IsError: final.IsError,
 	}, nil
 }
 
+// resolveTaskMode resolves the effective task mode for a given server.
+// Programmatic overrides via SetTaskConfig take precedence over the
+// per-server TasksMode in MCPServerConfig. Empty / unknown values map to
+// MCPTaskModeAuto.
+func (m *MCPToolManager) resolveTaskMode(name string, cfg config.MCPServerConfig) MCPTaskMode {
+	if m.taskCfg.PerServerMode != nil {
+		if v, ok := m.taskCfg.PerServerMode[name]; ok {
+			return v
+		}
+	}
+	return ParseTaskMode(cfg.TasksMode)
+}
+
+// ListServerTasks queries tasks/list on the named server and returns the
+// active and recent tasks the server is willing to disclose. Errors are
+// returned untouched (callers commonly ignore METHOD_NOT_FOUND when the
+// server didn't advertise tasks/list capability).
+func (m *MCPToolManager) ListServerTasks(ctx context.Context, serverName string) ([]MCPTaskInfo, error) {
+	c, err := m.taskClient(serverName)
+	if err != nil {
+		return nil, err
+	}
+	res, err := c.ListTasks(ctx, mcp.ListTasksRequest{})
+	if err != nil {
+		return nil, fmt.Errorf("tasks/list on %s: %w", serverName, err)
+	}
+	out := make([]MCPTaskInfo, 0, len(res.Tasks))
+	for _, t := range res.Tasks {
+		out = append(out, taskFromMCP(serverName, t))
+	}
+	return out, nil
+}
+
+// GetServerTask queries tasks/get for a single task on the named server.
+func (m *MCPToolManager) GetServerTask(ctx context.Context, serverName, taskID string) (MCPTaskInfo, error) {
+	c, err := m.taskClient(serverName)
+	if err != nil {
+		return MCPTaskInfo{}, err
+	}
+	res, err := c.GetTask(ctx, mcp.GetTaskRequest{Params: mcp.GetTaskParams{TaskId: taskID}})
+	if err != nil {
+		return MCPTaskInfo{}, fmt.Errorf("tasks/get on %s: %w", serverName, err)
+	}
+	return taskFromMCP(serverName, res.Task), nil
+}
+
+// CancelServerTask issues tasks/cancel for a task on the named server.
+// Returns the post-cancel task state when the server responded with one.
+func (m *MCPToolManager) CancelServerTask(ctx context.Context, serverName, taskID string) (MCPTaskInfo, error) {
+	c, err := m.taskClient(serverName)
+	if err != nil {
+		return MCPTaskInfo{}, err
+	}
+	res, err := c.CancelTask(ctx, mcp.CancelTaskRequest{Params: mcp.CancelTaskParams{TaskId: taskID}})
+	if err != nil {
+		return MCPTaskInfo{}, fmt.Errorf("tasks/cancel on %s: %w", serverName, err)
+	}
+	return taskFromMCP(serverName, res.Task), nil
+}
+
+// taskClient returns the *client.Client for a server. Tasks endpoints are
+// not part of the upstream MCPClient interface so callers must work with
+// the concrete client. Returns an error when the connection is missing
+// or backed by a non-standard client type.
+func (m *MCPToolManager) taskClient(serverName string) (*client.Client, error) {
+	if m.connectionPool == nil {
+		return nil, fmt.Errorf("no connection pool available")
+	}
+	clients := m.connectionPool.GetClients()
+	raw, ok := clients[serverName]
+	if !ok {
+		return nil, fmt.Errorf("MCP server %q not loaded", serverName)
+	}
+	c, ok := raw.(*client.Client)
+	if !ok {
+		return nil, fmt.Errorf("MCP server %q does not support task RPCs", serverName)
+	}
+	return c, nil
+}
+
 // GetTools returns all loaded MCP tools from all configured MCP servers.
 // Tools are returned with their prefixed names (serverName__toolName) to ensure uniqueness.
 func (m *MCPToolManager) GetTools() []MCPTool {

internal/tools/mcp_tasks.go

@@ -0,0 +1,404 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"sync/atomic"
+	"time"
+
+	"github.com/mark3labs/mcp-go/client"
+	"github.com/mark3labs/mcp-go/client/transport"
+	"github.com/mark3labs/mcp-go/mcp"
+)
+
+// MCPTaskMode controls when the connection pool augments tools/call requests
+// with MCP task metadata. See https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks.
+type MCPTaskMode string
+
+const (
+	// MCPTaskModeAuto augments tools/call with task metadata only when the
+	// server advertises tasks/toolCalls capability during initialize.
+	MCPTaskModeAuto MCPTaskMode = "auto"
+	// MCPTaskModeNever forces every tools/call to be issued synchronously
+	// (no Task field in the request), regardless of server capability.
+	MCPTaskModeNever MCPTaskMode = "never"
+	// MCPTaskModeAlways always sets a Task field on the tools/call request,
+	// even when the server didn't advertise task support. The server may
+	// still respond synchronously; this just opts in unconditionally on
+	// the client side.
+	MCPTaskModeAlways MCPTaskMode = "always"
+)
+
+// ParseTaskMode normalises a per-server tasks-mode string from
+// configuration. Empty input maps to MCPTaskModeAuto. Unknown values are
+// also treated as MCPTaskModeAuto so a stray config typo never breaks
+// existing flows.
+func ParseTaskMode(s string) MCPTaskMode {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "", "auto":
+		return MCPTaskModeAuto
+	case "never", "off", "disabled":
+		return MCPTaskModeNever
+	case "always", "force":
+		return MCPTaskModeAlways
+	default:
+		return MCPTaskModeAuto
+	}
+}
+
+// MCPTaskInfo is the connection-layer view of an MCP Task. It mirrors the
+// upstream mcp.Task but exposes Go-native types and includes the originating
+// server name. SDK-level wrappers re-export this under public-facing names.
+type MCPTaskInfo struct {
+	// Server is the configured MCP server name this task lives on.
+	Server string
+	// TaskID is the server-assigned identifier for the task.
+	TaskID string
+	// Status is the current task lifecycle state.
+	Status mcp.TaskStatus
+	// StatusMessage is an optional human-readable description.
+	StatusMessage string
+	// CreatedAt is the wall-clock time the task was created (best-effort
+	// parsed from the server's ISO-8601 timestamp; zero on parse failure).
+	CreatedAt time.Time
+	// UpdatedAt is the wall-clock time the task was last updated (best-
+	// effort parsed; zero on parse failure).
+	UpdatedAt time.Time
+	// TTL is the time-to-live the server intends to retain the task after
+	// creation. Zero means the server did not advertise a TTL.
+	TTL time.Duration
+	// PollInterval is the suggested polling interval. Zero means use the
+	// client's default.
+	PollInterval time.Duration
+}
+
+// MCPTaskProgress is emitted while the connection pool is waiting on a
+// task-augmented tool call. It provides minimal feedback for SDK consumers
+// that want to render progress widgets without subscribing to the full
+// notifications/tasks/status channel (Phase 2).
+type MCPTaskProgress struct {
+	Server  string
+	TaskID  string
+	Status  mcp.TaskStatus
+	Message string
+}
+
+// MCPTaskProgressHandler is invoked once after a task is accepted and on
+// every status transition observed by the polling loop. The final
+// invocation always carries a terminal status. Implementations must not
+// block; long work should be queued on a goroutine.
+type MCPTaskProgressHandler func(MCPTaskProgress)
+
+// MCPTaskConfig configures task-aware tool execution on the manager.
+// All fields are optional; the zero value disables progress callbacks and
+// applies sensible defaults.
+type MCPTaskConfig struct {
+	// PerServerMode overrides the per-server TasksMode resolved from
+	// MCPServerConfig. Keys are server names. Missing entries fall back
+	// to the value from config. Used by SDK consumers that want to set
+	// modes programmatically.
+	PerServerMode map[string]MCPTaskMode
+
+	// DefaultTTL is the TTL hint sent in TaskParams when augmenting a
+	// tools/call. Zero means omit the TTL — let the server pick its own.
+	DefaultTTL time.Duration
+
+	// PollInterval is the fallback interval between tasks/get requests
+	// when the server does not suggest one. Zero defaults to 1 second.
+	PollInterval time.Duration
+
+	// MaxPollInterval caps the polling interval. Zero defaults to 5 seconds.
+	MaxPollInterval time.Duration
+
+	// Timeout is the maximum wall-clock duration to wait for a task to
+	// reach a terminal state. Zero defaults to 15 minutes. Independent
+	// of the per-call context deadline; whichever fires first wins.
+	Timeout time.Duration
+
+	// Progress, if non-nil, receives every status transition observed by
+	// the polling loop.
+	Progress MCPTaskProgressHandler
+}
+
+func (c MCPTaskConfig) resolved() MCPTaskConfig {
+	if c.PollInterval <= 0 {
+		c.PollInterval = 1 * time.Second
+	}
+	if c.MaxPollInterval <= 0 {
+		c.MaxPollInterval = 5 * time.Second
+	}
+	if c.Timeout <= 0 {
+		c.Timeout = 15 * time.Minute
+	}
+	return c
+}
+
+// requestIDCounter generates monotonically increasing JSON-RPC request IDs
+// for low-level tools/call invocations that bypass the upstream client's
+// ParseCallToolResult helper (necessary because that helper rejects task
+// responses for lacking a "content" field).
+//
+// The counter is process-wide rather than per-manager so multiple managers
+// or repeated calls within the same connection produce unique IDs.
+var requestIDCounter atomic.Int64
+
+func nextRequestID() mcp.RequestId {
+	return mcp.NewRequestId(requestIDCounter.Add(1))
+}
+
+// callToolWithTask issues tools/call directly on the transport so we can
+// observe both response shapes:
+//
+//   - {"content": [...], ...}  — synchronous CallToolResult.
+//   - {"task": {...}, ...}     — asynchronous CreateTaskResult.
+//
+// On success exactly one of (callResult, taskResult) is non-nil. The
+// upstream client.CallTool helper parses the response with
+// mcp.ParseCallToolResult which requires a "content" field, so it cannot
+// be used for task-augmented calls.
+func callToolWithTask(
+	ctx context.Context,
+	c *client.Client,
+	params mcp.CallToolParams,
+) (callResult *mcp.CallToolResult, taskResult *mcp.CreateTaskResult, err error) {
+	tr := c.GetTransport()
+	if tr == nil {
+		return nil, nil, errors.New("mcp client has no transport")
+	}
+
+	req := transport.JSONRPCRequest{
+		JSONRPC: mcp.JSONRPC_VERSION,
+		ID:      nextRequestID(),
+		Method:  string(mcp.MethodToolsCall),
+		Params:  params,
+	}
+
+	resp, sendErr := tr.SendRequest(ctx, req)
+	if sendErr != nil {
+		return nil, nil, sendErr
+	}
+	if resp.Error != nil {
+		return nil, nil, resp.Error.AsError()
+	}
+
+	// Peek at the raw result to decide which shape we got.
+	var probe struct {
+		Task    json.RawMessage `json:"task"`
+		Content json.RawMessage `json:"content"`
+	}
+	raw := resp.Result
+	if len(raw) == 0 {
+		return nil, nil, errors.New("empty tools/call result")
+	}
+	if uErr := json.Unmarshal(raw, &probe); uErr != nil {
+		return nil, nil, fmt.Errorf("decode tools/call result: %w", uErr)
+	}
+
+	if len(probe.Task) > 0 && string(probe.Task) != "null" {
+		// Task-augmented response.
+		var ct mcp.CreateTaskResult
+		if uErr := json.Unmarshal(raw, &ct); uErr != nil {
+			return nil, nil, fmt.Errorf("decode CreateTaskResult: %w", uErr)
+		}
+		return nil, &ct, nil
+	}
+
+	// Synchronous response — defer to the upstream parser so content blocks
+	// are typed correctly (TextContent, ImageContent, ResourceLink, etc.).
+	cr, pErr := mcp.ParseCallToolResult(&raw)
+	if pErr != nil {
+		return nil, nil, fmt.Errorf("parse CallToolResult: %w", pErr)
+	}
+	return cr, nil, nil
+}
+
+// pollTaskUntilTerminal blocks until the task reaches a terminal status,
+// the context is cancelled, or the configured timeout elapses. On
+// cancellation it best-effort issues tasks/cancel before returning.
+func pollTaskUntilTerminal(
+	ctx context.Context,
+	c *client.Client,
+	serverName string,
+	task mcp.Task,
+	cfg MCPTaskConfig,
+	progress MCPTaskProgressHandler,
+) (*mcp.TaskResultResult, error) {
+	cfg = cfg.resolved()
+	deadline := time.Now().Add(cfg.Timeout)
+
+	emit := func(status mcp.TaskStatus, msg string) {
+		if progress != nil {
+			progress(MCPTaskProgress{Server: serverName, TaskID: task.TaskId, Status: status, Message: msg})
+		}
+	}
+
+	emit(task.Status, task.StatusMessage)
+
+	current := task
+	interval := cfg.PollInterval
+	if current.PollInterval != nil && *current.PollInterval > 0 {
+		interval = time.Duration(*current.PollInterval) * time.Millisecond
+	}
+	if interval > cfg.MaxPollInterval {
+		interval = cfg.MaxPollInterval
+	}
+
+	for !current.Status.IsTerminal() {
+		if time.Now().After(deadline) {
+			cancelTaskBestEffort(c, current.TaskId)
+			return nil, fmt.Errorf("task %s timed out after %s", current.TaskId, cfg.Timeout)
+		}
+
+		// Wait between polls or abort early on context cancellation.
+		select {
+		case <-ctx.Done():
+			cancelTaskBestEffort(c, current.TaskId)
+			return nil, ctx.Err()
+		case <-time.After(interval):
+		}
+
+		got, err := c.GetTask(ctx, mcp.GetTaskRequest{
+			Params: mcp.GetTaskParams{TaskId: current.TaskId},
+		})
+		if err != nil {
+			// Transient transport hiccup — propagate immediately. The
+			// upstream agent layer treats this like any other tool error.
+			return nil, fmt.Errorf("tasks/get failed: %w", err)
+		}
+		current = got.Task
+		if current.Status != task.Status || current.StatusMessage != task.StatusMessage {
+			emit(current.Status, current.StatusMessage)
+			task = current
+		}
+
+		// Honour any updated suggested poll interval, capped at the limit.
+		if current.PollInterval != nil && *current.PollInterval > 0 {
+			interval = min(time.Duration(*current.PollInterval)*time.Millisecond, cfg.MaxPollInterval)
+		}
+	}
+
+	// Terminal state reached. Emit one last progress event and fetch the
+	// definitive tool result.
+	emit(current.Status, current.StatusMessage)
+
+	if current.Status == mcp.TaskStatusCancelled {
+		return nil, fmt.Errorf("task %s was cancelled", current.TaskId)
+	}
+
+	res, err := fetchTaskResult(ctx, c, current.TaskId)
+	if err != nil {
+		return nil, fmt.Errorf("tasks/result failed: %w", err)
+	}
+	if current.Status == mcp.TaskStatusFailed && res != nil && !res.IsError {
+		// The server flagged the task as failed but didn't decorate the
+		// result. Surface the status message so the caller still sees a
+		// useful tool-error.
+		return nil, fmt.Errorf("task %s failed: %s", current.TaskId, current.StatusMessage)
+	}
+	return res, nil
+}
+
+// fetchTaskResult issues tasks/result on the transport and parses the raw
+// response. The upstream client.TaskResult helper delegates to
+// mcp.ParseTaskResultResult which (as of mcp-go v0.51.0) looks for the
+// content array under a nested "result" key that never exists in the
+// wire format — leading to systematically empty Content. Doing the
+// parse here keeps the polling path working until that is fixed upstream.
+func fetchTaskResult(ctx context.Context, c *client.Client, taskID string) (*mcp.TaskResultResult, error) {
+	tr := c.GetTransport()
+	if tr == nil {
+		return nil, errors.New("mcp client has no transport")
+	}
+	req := transport.JSONRPCRequest{
+		JSONRPC: mcp.JSONRPC_VERSION,
+		ID:      nextRequestID(),
+		Method:  string(mcp.MethodTasksResult),
+		Params:  mcp.TaskResultParams{TaskId: taskID},
+	}
+	resp, err := tr.SendRequest(ctx, req)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Error != nil {
+		return nil, resp.Error.AsError()
+	}
+
+	// Manually decode the wire shape: {"_meta": {...}, "content": [...],
+	// "structuredContent": ..., "isError": bool}.
+	var shape struct {
+		Meta              json.RawMessage   `json:"_meta"`
+		Content           []json.RawMessage `json:"content"`
+		StructuredContent any               `json:"structuredContent"`
+		IsError           bool              `json:"isError"`
+	}
+	if err := json.Unmarshal(resp.Result, &shape); err != nil {
+		return nil, fmt.Errorf("decode tasks/result: %w", err)
+	}
+
+	out := &mcp.TaskResultResult{
+		StructuredContent: shape.StructuredContent,
+		IsError:           shape.IsError,
+	}
+	if len(shape.Meta) > 0 && string(shape.Meta) != "null" {
+		var metaMap map[string]any
+		if err := json.Unmarshal(shape.Meta, &metaMap); err == nil {
+			out.Meta = mcp.NewMetaFromMap(metaMap)
+		}
+	}
+	for _, raw := range shape.Content {
+		var contentMap map[string]any
+		if err := json.Unmarshal(raw, &contentMap); err != nil {
+			return nil, fmt.Errorf("decode content block: %w", err)
+		}
+		parsed, err := mcp.ParseContent(contentMap)
+		if err != nil {
+			return nil, fmt.Errorf("parse content block: %w", err)
+		}
+		out.Content = append(out.Content, parsed)
+	}
+	return out, nil
+}
+
+// cancelTaskBestEffort issues tasks/cancel and ignores any error. Used on
+// context cancellation paths where the connection is already going away.
+func cancelTaskBestEffort(c *client.Client, taskID string) {
+	if c == nil || taskID == "" {
+		return
+	}
+	cancelCtx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	_, _ = c.CancelTask(cancelCtx, mcp.CancelTaskRequest{
+		Params: mcp.CancelTaskParams{TaskId: taskID},
+	})
+}
+
+// taskFromMCP converts a wire-format mcp.Task to our richer connection-
+// layer view. Unparseable timestamps surface as the zero time.
+func taskFromMCP(serverName string, t mcp.Task) MCPTaskInfo {
+	out := MCPTaskInfo{
+		Server:        serverName,
+		TaskID:        t.TaskId,
+		Status:        t.Status,
+		StatusMessage: t.StatusMessage,
+	}
+	if t.CreatedAt != "" {
+		if v, err := time.Parse(time.RFC3339, t.CreatedAt); err == nil {
+			out.CreatedAt = v
+		}
+	}
+	if t.LastUpdatedAt != "" {
+		if v, err := time.Parse(time.RFC3339, t.LastUpdatedAt); err == nil {
+			out.UpdatedAt = v
+		}
+	}
+	if t.TTL != nil {
+		out.TTL = time.Duration(*t.TTL) * time.Millisecond
+	}
+	if t.PollInterval != nil {
+		out.PollInterval = time.Duration(*t.PollInterval) * time.Millisecond
+	}
+	return out
+}

internal/tools/mcp_tasks_test.go

@@ -0,0 +1,294 @@
+package tools
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/mark3labs/kit/internal/config"
+	"github.com/mark3labs/mcp-go/mcp"
+	"github.com/mark3labs/mcp-go/server"
+)
+
+// newTaskTestInProcessServer builds an in-process MCP server with a
+// task-augmented tool. The handler simulates work by sleeping briefly
+// before completing.
+//
+// Important: the upstream mcp-go server cancels the request context as
+// soon as the synchronous part of the tools/call returns (see
+// request_handler.go:85, `defer cancel()`). Task goroutines spawned by
+// AddTaskTool inherit that context and therefore see context.Canceled
+// the instant they start. Real-world transports (stdio, SSE, streamable
+// HTTP) don't trip this because they keep the connection — and a
+// background context — alive across the async work, but the in-process
+// transport runs entirely on the request goroutine. To test the polling
+// path realistically we detach from the request context here.
+func newTaskTestInProcessServer(t *testing.T, workDuration time.Duration) *server.MCPServer {
+	t.Helper()
+	srv := server.NewMCPServer("task-test", "1.0.0",
+		server.WithToolCapabilities(true),
+		// list=true, cancel=true, toolCallTasks=true so capability detection,
+		// cancellation, and tool augmentation all flow through.
+		server.WithTaskCapabilities(true, true, true),
+	)
+	srv.AddTaskTool(
+		mcp.Tool{
+			Name:        "long_running",
+			Description: "Sleep, then echo the input string.",
+			InputSchema: mcp.ToolInputSchema{
+				Type: "object",
+				Properties: map[string]any{
+					"msg": map[string]any{"type": "string"},
+				},
+			},
+			Execution: &mcp.ToolExecution{
+				TaskSupport: mcp.TaskSupportRequired,
+			},
+		},
+		func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CreateTaskResult, error) {
+			msg, _ := req.GetArguments()["msg"].(string)
+			// Detach from the request context so the task handler can
+			// outlive the synchronous request — see comment above.
+			time.Sleep(workDuration)
+			_ = ctx
+			return &mcp.CreateTaskResult{
+				Content: []mcp.Content{
+					mcp.TextContent{Type: "text", Text: "echo:" + msg},
+				},
+			}, nil
+		},
+	)
+	return srv
+}
+
+// newSyncOnlyServer is a server that does NOT advertise task capability.
+// Used to verify the auto-detect path keeps the sync semantics.
+func newSyncOnlyServer() *server.MCPServer {
+	srv := server.NewMCPServer("sync-only", "1.0.0",
+		server.WithToolCapabilities(true),
+	)
+	srv.AddTool(
+		mcp.NewTool("greet",
+			mcp.WithDescription("Say hello"),
+			mcp.WithString("name", mcp.Required()),
+		),
+		func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+			name, _ := req.GetArguments()["name"].(string)
+			return mcp.NewToolResultText("hi " + name), nil
+		},
+	)
+	return srv
+}
+
+func TestConnectionPoolAdvertisesTaskCapability(t *testing.T) {
+	pool := NewMCPConnectionPool(DefaultConnectionPoolConfig(), false, nil, nil)
+	defer func() { _ = pool.Close() }()
+
+	srv := newTaskTestInProcessServer(t, 0)
+	cfg := config.MCPServerConfig{Type: "inprocess", InProcessServer: srv}
+
+	conn, err := pool.GetConnection(context.Background(), "tasks", cfg)
+	if err != nil {
+		t.Fatalf("GetConnection: %v", err)
+	}
+
+	init := conn.InitializeResult()
+	if init == nil {
+		t.Fatal("InitializeResult is nil after GetConnection")
+	}
+	if init.Capabilities.Tasks == nil {
+		t.Fatal("server did not advertise Tasks capability — initialize handshake regressed")
+	}
+	if !conn.SupportsToolTasks() {
+		t.Error("SupportsToolTasks should be true for a server with toolCallTasks=true")
+	}
+	if !pool.ServerSupportsToolTasks("tasks") {
+		t.Error("ServerSupportsToolTasks should mirror the connection's value")
+	}
+}
+
+func TestConnectionPoolDetectsAbsentTaskCapability(t *testing.T) {
+	pool := NewMCPConnectionPool(DefaultConnectionPoolConfig(), false, nil, nil)
+	defer func() { _ = pool.Close() }()
+
+	cfg := config.MCPServerConfig{Type: "inprocess", InProcessServer: newSyncOnlyServer()}
+	conn, err := pool.GetConnection(context.Background(), "sync", cfg)
+	if err != nil {
+		t.Fatalf("GetConnection: %v", err)
+	}
+	if conn.SupportsToolTasks() {
+		t.Error("SupportsToolTasks should be false for a server that didn't advertise the capability")
+	}
+}
+
+func TestSupportsToolTasksFromInit(t *testing.T) {
+	cases := []struct {
+		name string
+		in   *mcp.InitializeResult
+		want bool
+	}{
+		{"nil", nil, false},
+		{"no tasks", &mcp.InitializeResult{}, false},
+		{"tasks no requests", &mcp.InitializeResult{
+			Capabilities: mcp.ServerCapabilities{Tasks: &mcp.TasksCapability{}},
+		}, false},
+		{"tasks with toolCalls", &mcp.InitializeResult{
+			Capabilities: mcp.ServerCapabilities{Tasks: mcp.NewTasksCapability()},
+		}, true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := supportsToolTasksFromInit(tc.in); got != tc.want {
+				t.Errorf("supportsToolTasksFromInit() = %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestParseTaskMode(t *testing.T) {
+	cases := []struct {
+		in   string
+		want MCPTaskMode
+	}{
+		{"", MCPTaskModeAuto},
+		{"auto", MCPTaskModeAuto},
+		{"AUTO", MCPTaskModeAuto},
+		{"never", MCPTaskModeNever},
+		{"off", MCPTaskModeNever},
+		{"always", MCPTaskModeAlways},
+		{"force", MCPTaskModeAlways},
+		{"bogus", MCPTaskModeAuto},
+	}
+	for _, tc := range cases {
+		if got := ParseTaskMode(tc.in); got != tc.want {
+			t.Errorf("ParseTaskMode(%q) = %q, want %q", tc.in, got, tc.want)
+		}
+	}
+}
+
+func TestExecuteToolPollsTaskToCompletion(t *testing.T) {
+	mgr := NewMCPToolManager()
+	mgr.SetTaskConfig(MCPTaskConfig{
+		PollInterval:    20 * time.Millisecond,
+		MaxPollInterval: 50 * time.Millisecond,
+		Timeout:         10 * time.Second,
+	})
+
+	cfg := config.MCPServerConfig{
+		Type:            "inprocess",
+		InProcessServer: newTaskTestInProcessServer(t, 50*time.Millisecond),
+	}
+
+	if _, err := mgr.AddServer(context.Background(), "tasks", cfg); err != nil {
+		t.Fatalf("AddServer: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	res, err := mgr.ExecuteTool(ctx, "tasks__long_running", `{"msg":"hello"}`)
+	if err != nil {
+		t.Fatalf("ExecuteTool: %v", err)
+	}
+	if res.IsError {
+		t.Fatalf("expected non-error result, got %s", res.Content)
+	}
+	if !strings.Contains(res.Content, "echo:hello") {
+		t.Errorf("expected result to contain 'echo:hello', got %s", res.Content)
+	}
+}
+
+func TestExecuteToolHonorsNeverMode(t *testing.T) {
+	// Even though the server advertises tasks/toolCalls, "never" should
+	// keep the call synchronous. Since the tool is TaskSupportRequired,
+	// the server returns an error rather than running it sync — we just
+	// verify the error surfaces (not a poll-loop hang).
+	mgr := NewMCPToolManager()
+	mgr.SetTaskConfig(MCPTaskConfig{
+		PerServerMode: map[string]MCPTaskMode{"tasks": MCPTaskModeNever},
+		Timeout:       2 * time.Second,
+	})
+
+	cfg := config.MCPServerConfig{
+		Type:            "inprocess",
+		InProcessServer: newTaskTestInProcessServer(t, 0),
+	}
+
+	if _, err := mgr.AddServer(context.Background(), "tasks", cfg); err != nil {
+		t.Fatalf("AddServer: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+
+	// We don't care which way the server fails the sync call; we just want
+	// to confirm we didn't hang in the polling loop and didn't panic.
+	_, err := mgr.ExecuteTool(ctx, "tasks__long_running", `{"msg":"x"}`)
+	if err == nil {
+		t.Fatal("expected an error when forcing sync execution of a task-required tool")
+	}
+}
+
+func TestExecuteToolEmitsProgress(t *testing.T) {
+	var statuses []mcp.TaskStatus
+	mgr := NewMCPToolManager()
+	mgr.SetTaskConfig(MCPTaskConfig{
+		PollInterval:    10 * time.Millisecond,
+		MaxPollInterval: 25 * time.Millisecond,
+		Timeout:         5 * time.Second,
+		Progress: func(p MCPTaskProgress) {
+			statuses = append(statuses, p.Status)
+		},
+	})
+
+	cfg := config.MCPServerConfig{
+		Type:            "inprocess",
+		InProcessServer: newTaskTestInProcessServer(t, 30*time.Millisecond),
+	}
+	if _, err := mgr.AddServer(context.Background(), "tasks", cfg); err != nil {
+		t.Fatalf("AddServer: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	if _, err := mgr.ExecuteTool(ctx, "tasks__long_running", `{"msg":"hi"}`); err != nil {
+		t.Fatalf("ExecuteTool: %v", err)
+	}
+	if len(statuses) == 0 {
+		t.Fatal("expected at least one progress event")
+	}
+	last := statuses[len(statuses)-1]
+	if !last.IsTerminal() {
+		t.Errorf("last progress event should be terminal, got %q", last)
+	}
+}
+
+func TestListGetCancelMCPTasksOnLoadedServer(t *testing.T) {
+	mgr := NewMCPToolManager()
+	cfg := config.MCPServerConfig{
+		Type:            "inprocess",
+		InProcessServer: newTaskTestInProcessServer(t, 0),
+	}
+	if _, err := mgr.AddServer(context.Background(), "tasks", cfg); err != nil {
+		t.Fatalf("AddServer: %v", err)
+	}
+
+	ctx := context.Background()
+
+	// tasks/list — no in-flight tasks yet, so we just verify the call
+	// succeeds and returns an empty slice (or any slice; the exact length
+	// depends on server retention policy).
+	if _, err := mgr.ListServerTasks(ctx, "tasks"); err != nil {
+		t.Errorf("ListServerTasks: %v", err)
+	}
+
+	// Unknown server should error cleanly without panicking.
+	if _, err := mgr.GetServerTask(ctx, "unknown", "abc"); err == nil {
+		t.Error("GetServerTask on unknown server should error")
+	}
+	if _, err := mgr.CancelServerTask(ctx, "unknown", "abc"); err == nil {
+		t.Error("CancelServerTask on unknown server should error")
+	}
+}

internal/ui/children_test.go

@@ -19,7 +19,7 @@ import (
 
 // newTestInput creates an InputComponent with the given AppController (may be nil).
 func newTestInput(ctrl AppController) *InputComponent {
-	return NewInputComponent(80, "test input", ctrl)
+	return NewInputComponent(80, ctrl)
 }
 
 // sendInputMsg calls component.Update with the given message, returns the
@@ -69,30 +69,6 @@ func TestInputComponent_SubmitEmitsSubmitMsg(t *testing.T) {
 	}
 }
 
-// TestInputComponent_CtrlD_SubmitEmitsSubmitMsg verifies that ctrl+d also
-// submits the text.
-func TestInputComponent_CtrlD_SubmitEmitsSubmitMsg(t *testing.T) {
-	ctrl := &stubAppController{}
-	c := newTestInput(ctrl)
-
-	c.textarea.SetValue("ctrl+d submit")
-	c.lastValue = "ctrl+d submit"
-
-	_, cmd := sendInputMsg(c, tea.KeyPressMsg{Code: 'd', Mod: tea.ModCtrl})
-
-	msg := runCmd(cmd)
-	if msg == nil {
-		t.Fatal("expected a cmd from ctrl+d on non-empty input")
-	}
-	sm, ok := msg.(core.SubmitMsg)
-	if !ok {
-		t.Fatalf("expected submitMsg from ctrl+d, got %T", msg)
-	}
-	if sm.Text != "ctrl+d submit" {
-		t.Fatalf("expected Text='ctrl+d submit', got %q", sm.Text)
-	}
-}
-
 // TestInputComponent_EmptySubmit_NoCmd verifies that submitting an empty or
 // whitespace-only string produces no cmd.
 func TestInputComponent_EmptySubmit_NoCmd(t *testing.T) {

internal/ui/commands/commands.go

@@ -84,7 +84,7 @@ var SlashCommands = []SlashCommand{
 	},
 	{
 		Name:        "/thinking",
-		Description: "Set thinking/reasoning level (off, minimal, low, medium, high)",
+		Description: "Set thinking/reasoning level (off, none, minimal, low, medium, high)",
 		Category:    "System",
 		Aliases:     []string{"/think"},
 		Complete: func(prefix string) []string {

internal/ui/core/events.go

@@ -25,6 +25,11 @@ type SubmitMsg struct {
 // presses ESC a second time, the canceling state is reset to false.
 type CancelTimerExpiredMsg struct{}
 
+// CtrlCResetMsg is sent after a short delay when the user presses Ctrl+C to
+// clear input. If the user doesn't press Ctrl+C again within the timeout,
+// the ctrlCPressedOnce flag is reset so the next Ctrl+C will clear again.
+type CtrlCResetMsg struct{}
+
 // --- Tree session events ---
 
 // TreeNodeSelectedMsg is sent when the user selects a node in the tree selector.

internal/ui/file_suggestions.go

@@ -6,6 +6,8 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
+	"sync"
+	"time"
 )
 
 // FileSuggestion represents a single file, directory, or MCP resource
@@ -31,6 +33,51 @@ type FileSuggestion struct {
 // maxFileSuggestions is the maximum number of file suggestions returned.
 const maxFileSuggestions = 20
 
+// fileListCache caches the result of listFiles() keyed by directory to avoid
+// re-running git subprocesses on every keystroke during @file completion.
+var fileListCache struct {
+	mu       sync.Mutex
+	dir      string           // searchDir that produced the cached entries
+	cwd      string           // cwd used for the git query
+	entries  []FileSuggestion // cached file list
+	expireAt time.Time        // when the cache entry expires
+}
+
+// fileListCacheTTL controls how long a cached file list stays valid.
+// During rapid typing the list is reused; after the TTL a fresh git
+// ls-files is executed so newly created files become visible.
+const fileListCacheTTL = 3 * time.Second
+
+// getCachedFileList returns the file list for searchDir, using a short-lived
+// cache to avoid repeated subprocess calls during @file autocompletion.
+func getCachedFileList(searchDir, cwd string) []FileSuggestion {
+	fileListCache.mu.Lock()
+	defer fileListCache.mu.Unlock()
+
+	now := time.Now()
+	if fileListCache.dir == searchDir &&
+		fileListCache.cwd == cwd &&
+		now.Before(fileListCache.expireAt) {
+		// Return a copy so callers can mutate (e.g. prepend baseDir).
+		cp := make([]FileSuggestion, len(fileListCache.entries))
+		copy(cp, fileListCache.entries)
+		return cp
+	}
+
+	// Cache miss or expired — run the real (potentially expensive) lookup.
+	files := listFiles(searchDir, cwd)
+
+	fileListCache.dir = searchDir
+	fileListCache.cwd = cwd
+	fileListCache.entries = files
+	fileListCache.expireAt = now.Add(fileListCacheTTL)
+
+	// Return a copy.
+	cp := make([]FileSuggestion, len(files))
+	copy(cp, files)
+	return cp
+}
+
 // ExtractAtPrefix checks the current line for an @-file trigger at cursorCol.
 // It returns:
 //   - hasAt: true if a valid @ trigger was found
@@ -99,7 +146,7 @@ func GetFileSuggestions(prefix string, cwd string) []FileSuggestion {
 		}
 	}
 
-	files := listFiles(searchDir, cwd)
+	files := getCachedFileList(searchDir, cwd)
 	if len(files) == 0 {
 		return nil
 	}

internal/ui/fileutil/processor_test.go

@@ -145,7 +145,13 @@ func TestDetectMediaType(t *testing.T) {
 		content  []byte
 		expected string
 	}{
-		{".go", nil, "text/plain"}, // .go falls back to content sniffing → text/plain
+		// An intentionally-synthetic extension that is not registered
+		// in any system MIME database. Exercises the "unknown ext +
+		// no content" branch, which must return the text/plain default.
+		// Do not use real extensions (e.g. .go) here: CI images often
+		// ship /etc/mime.types with entries like ".go → text/x-go",
+		// which would make the assertion environment-dependent.
+		{".kitsyntheticext", nil, "text/plain"},
 		{".png", []byte{0x89, 0x50, 0x4E, 0x47}, "image/png"},
 		{".jpg", []byte{0xFF, 0xD8, 0xFF}, "image/jpeg"},
 		{".pdf", []byte{0x25, 0x50, 0x44, 0x46}, "application/pdf"},

internal/ui/format.go

@@ -17,6 +17,7 @@ type Renderer interface {
 	RenderReasoningBlock(content string, timestamp time.Time) UIMessage
 	RenderToolMessage(toolName, toolArgs, toolResult string, isError bool) UIMessage
 	RenderSystemMessage(content string, timestamp time.Time) UIMessage
+	RenderCustomMessage(content, label string, timestamp time.Time) UIMessage
 	RenderErrorMessage(errorMsg string, timestamp time.Time) UIMessage
 	RenderDebugMessage(message string, timestamp time.Time) UIMessage
 	RenderDebugConfigMessage(config map[string]any, timestamp time.Time) UIMessage

internal/ui/input.go

@@ -40,7 +40,6 @@ type InputComponent struct {
 	width       int
 	lastValue   string
 	popupHeight int
-	title       string
 	submitNext  bool // defer submit one tick so popup dismisses cleanly
 
 	// Argument completion state. When the user types "/cmd " followed by
@@ -106,17 +105,17 @@ type clipboardImageMsg struct {
 	err   error
 }
 
-// NewInputComponent creates a new InputComponent with the given width, title,
-// and optional AppController. If appCtrl is nil the component still works but
+// NewInputComponent creates a new InputComponent with the given width and
+// optional AppController. If appCtrl is nil the component still works but
 // /clear and /clear-queue are no-ops.
-func NewInputComponent(width int, title string, appCtrl AppController) *InputComponent {
+func NewInputComponent(width int, appCtrl AppController) *InputComponent {
 	ta := textarea.New()
 	ta.Placeholder = "Type your message..."
 	ta.ShowLineNumbers = false
 	ta.Prompt = ""
 	ta.CharLimit = 0
 	ta.SetWidth(width - 8) // Account for container padding, border and internal padding
-	ta.SetHeight(3)        // Default to 3 lines like huh
+	ta.SetHeight(4)        // 4 lines for comfortable multi-line input
 	ta.Focus()
 
 	// Override InsertNewline so only ctrl+j and shift+enter insert newlines.
@@ -141,8 +140,8 @@ func NewInputComponent(width int, title string, appCtrl AppController) *InputCom
 		commands:    commands.SlashCommands,
 		width:       width,
 		popupHeight: 7,
-		title:       title,
 		appCtrl:     appCtrl,
+		hideHint:    true,
 	}
 }
 
@@ -201,7 +200,7 @@ func (s *InputComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 	case tea.KeyPressMsg:
 		if !s.showPopup {
 			switch msg.String() {
-			case "ctrl+d", "enter":
+			case "enter":
 				value := s.textarea.Value()
 				s.pushHistory(value)
 				s.textarea.SetValue("")
@@ -520,19 +519,13 @@ func (s *InputComponent) resetHistoryBrowsing() {
 	s.savedInput = ""
 }
 
-// View implements tea.Model. Renders the title, textarea, autocomplete popup
+// View implements tea.Model. Renders the textarea, autocomplete popup
 // (if visible), and help text.
 func (s *InputComponent) View() tea.View {
 	containerStyle := lipgloss.NewStyle()
 
 	theme := style.GetTheme()
 
-	// PaddingLeft(3) aligns with message content: border(1) + paddingLeft(2).
-	titleStyle := lipgloss.NewStyle().
-		Foreground(theme.Text).
-		MarginBottom(1).
-		PaddingLeft(3)
-
 	inputBoxStyle := lipgloss.NewStyle().
 		Border(lipgloss.ThickBorder()).
 		BorderLeft(true).
@@ -540,12 +533,12 @@ func (s *InputComponent) View() tea.View {
 		BorderTop(false).
 		BorderBottom(false).
 		BorderForeground(theme.Primary).
+		MarginTop(1).
+		MarginBottom(1).
 		PaddingLeft(2).    // match message block paddingLeft
 		Width(s.width - 1) // full width minus left border
 
 	var view strings.Builder
-	view.WriteString(titleStyle.Render(s.title))
-	view.WriteString("\n")
 	view.WriteString(inputBoxStyle.Render(s.textarea.View()))
 
 	// Popup is now rendered as a centered overlay in AppModel.View()
@@ -708,9 +701,25 @@ func (s *InputComponent) renderPopupWithOptions(centered bool) string {
 				}
 				content = indicator + displayName
 			} else {
-				nameWidth := 15
-				if innerWidth < 25 {
-					nameWidth = max(innerWidth*2/5+1, 8)
+				// Compute nameWidth from the longest command name in the
+				// visible slice so we never truncate unnecessarily.
+				nameWidth := 0
+				for _, fm := range s.filtered {
+					if n := len([]rune(fm.Command.Name)); n > nameWidth {
+						nameWidth = n
+					}
+				}
+				nameWidth += 3 // account for indicator prefix (2) + gap before description (1)
+				// Ensure descriptions still get at least 20 chars when possible.
+				maxForName := innerWidth - 20
+				if maxForName < 8 {
+					maxForName = innerWidth * 2 / 3
+				}
+				if nameWidth > maxForName {
+					nameWidth = maxForName
+				}
+				if nameWidth < 8 {
+					nameWidth = 8
 				}
 				maxNameChars := nameWidth - 2
 				displayName := sc.Name
@@ -843,6 +852,21 @@ func (s *InputComponent) PendingImageCount() int {
 	return len(s.pendingImages)
 }
 
+// Clear clears the textarea content and resets related state. Returns true if
+// there was content to clear, false if the input was already empty.
+func (s *InputComponent) Clear() bool {
+	hadContent := s.textarea.Value() != ""
+	s.textarea.SetValue("")
+	s.textarea.CursorEnd()
+	s.lastValue = ""
+	s.showPopup = false
+	s.argMode = false
+	s.fileMode = false
+	s.browsingHistory = false
+	s.savedInput = ""
+	return hadContent
+}
+
 // applyFileCompletion replaces the @prefix in the textarea with the selected
 // file or MCP resource suggestion. For directories, it keeps the popup open
 // for further drilling. For files and resources, it closes the popup and adds

internal/ui/message_items.go

@@ -109,8 +109,8 @@ func (m *TextMessageItem) renderContent(width int) string {
 // It accumulates content chunks and re-renders on each update for live display.
 type StreamingMessageItem struct {
 	id            string
-	role          string // "assistant" or "reasoning"
-	content       string // Accumulated streaming content
+	role          string          // "assistant" or "reasoning"
+	content       strings.Builder // Accumulated streaming content
 	timestamp     time.Time
 	startTime     time.Time // When streaming started (for live duration counter)
 	modelName     string
@@ -156,10 +156,10 @@ func (s *StreamingMessageItem) Render(width int) string {
 			durationMs = time.Since(s.startTime).Milliseconds()
 		}
 		ty := createTypography(style.GetTheme())
-		rendered = render.ReasoningBlock(s.content, durationMs, ty, style.GetTheme())
+		rendered = render.ReasoningBlock(s.content.String(), durationMs, width, ty, style.GetTheme())
 	} else {
 		// Render as assistant message
-		rendered = render.AssistantBlock(s.content, width, style.GetTheme())
+		rendered = render.AssistantBlock(s.content.String(), width, style.GetTheme())
 	}
 
 	// Cache and return (but reasoning is never cached due to live duration)
@@ -187,7 +187,7 @@ func (s *StreamingMessageItem) Height() int {
 
 // AppendChunk adds a content chunk and invalidates the render cache.
 func (s *StreamingMessageItem) AppendChunk(chunk string) {
-	s.content += chunk
+	s.content.WriteString(chunk)
 	s.cachedWidth = 0 // Invalidate cache
 }
 
@@ -243,9 +243,7 @@ func (m *StreamingBashOutputItem) Render(width int) string {
 
 	// Header with command
 	if m.command != "" {
-		headerStyle := lipgloss.NewStyle().
-			Foreground(theme.Muted).
-			Italic(true)
+		headerStyle := style.GetCachedStyles().BashHeader
 		parts = append(parts, headerStyle.Render(fmt.Sprintf("▸ %s", m.command)))
 	}
 

internal/ui/messages.go

@@ -88,13 +88,9 @@ func formatToolParams(toolArgs string, maxWidth int) string {
 	}
 
 	bodyKeys := map[string]bool{
-		"content":  true,
-		"old_text": true,
-		"new_text": true,
-		"oldText":  true,
-		"newText":  true,
-		"edits":    true,
-		"todos":    true,
+		"content": true,
+		"edits":   true,
+		"todos":   true,
 	}
 	var remaining []string
 	for key, val := range params {
@@ -150,9 +146,26 @@ func (r *MessageRenderer) SetWidth(width int) {
 	r.width = width
 }
 
-// RenderUserMessage renders a user's input message using herald Tip alert
+// RenderUserMessage renders a user's input message with a colored left border.
 func (r *MessageRenderer) RenderUserMessage(content string, timestamp time.Time) UIMessage {
-	rendered := render.UserBlock(content, r.width, r.ty, style.GetTheme())
+	if strings.TrimSpace(content) == "" {
+		content = "(empty message)"
+	}
+
+	theme := style.GetTheme()
+
+	// Highlight @file tokens with accent color.
+	content = render.HighlightFileTokens(content, theme)
+
+	rendered := renderContentBlock(
+		content,
+		r.width,
+		WithAlign(lipgloss.Left),
+		WithBorderColor(theme.Success),
+		WithPaddingTop(0),
+		WithPaddingBottom(0),
+		WithMarginBottom(1),
+	)
 
 	return UIMessage{
 		Type:      UserMessage,
@@ -178,7 +191,7 @@ func (r *MessageRenderer) RenderAssistantMessage(content string, timestamp time.
 // as live streaming: muted italic text with margin. This is used when resuming
 // sessions to display saved reasoning content.
 func (r *MessageRenderer) RenderReasoningBlock(content string, timestamp time.Time) UIMessage {
-	rendered := render.ReasoningBlock(content, 0, r.ty, style.GetTheme())
+	rendered := render.ReasoningBlock(content, 0, r.width, r.ty, style.GetTheme())
 
 	return UIMessage{
 		Type:      AssistantMessage,
@@ -200,6 +213,19 @@ func (r *MessageRenderer) RenderSystemMessage(content string, timestamp time.Tim
 	}
 }
 
+// RenderCustomMessage renders a message with a custom alert label (e.g. "Help").
+// Content is rendered as markdown.
+func (r *MessageRenderer) RenderCustomMessage(content, label string, timestamp time.Time) UIMessage {
+	rendered := render.CustomBlock(content, label, r.width, style.GetTheme())
+
+	return UIMessage{
+		Type:      SystemMessage,
+		Content:   rendered,
+		Height:    lipgloss.Height(rendered),
+		Timestamp: timestamp,
+	}
+}
+
 // RenderDebugMessage renders diagnostic and debugging information
 func (r *MessageRenderer) RenderDebugMessage(message string, timestamp time.Time) UIMessage {
 	header := r.ty.H6("🔍 Debug Output")
@@ -308,7 +334,7 @@ func (r *MessageRenderer) RenderToolMessage(toolName, toolArgs, toolResult strin
 	// Build the content: icon + name + params on first line, then body
 	headerLine := styledIcon + " " + styledName
 	if params != "" {
-		headerLine += " " + lipgloss.NewStyle().Foreground(theme.Muted).Render(params)
+		headerLine += " " + style.GetCachedStyles().ToolMuted.Render(params)
 	}
 
 	// Get body content
@@ -399,7 +425,8 @@ func createTypography(theme style.Theme) *herald.Typography {
 		herald.WithCodeLineNumbers(true),
 		// Customize alert labels
 		herald.WithAlertLabel(herald.AlertNote, "Info"),
-		herald.WithAlertLabel(herald.AlertTip, "You"),
+		herald.WithAlertLabel(herald.AlertTip, ""),
+		herald.WithAlertIcon(herald.AlertTip, ""),
 		herald.WithAlertLabel(herald.AlertWarning, "Working"),
 		herald.WithAlertLabel(herald.AlertCaution, "Error"),
 	)

internal/ui/model.go

@@ -720,6 +720,10 @@ type AppModel struct {
 	// disables alt screen to restore the terminal properly.
 	quitting bool
 
+	// ctrlCPressedOnce tracks if Ctrl+C was pressed once to clear input.
+	// A second Ctrl+C (or Ctrl+C when input is empty) will quit the app.
+	ctrlCPressedOnce bool
+
 	// streamingBashOutput holds the current streaming bash output lines.
 	// Lines are accumulated as they arrive and displayed in the stream region.
 	streamingBashOutput []string
@@ -869,7 +873,7 @@ func NewAppModel(appCtrl AppController, opts AppModelOptions) *AppModel {
 	m.messages = []MessageItem{}
 
 	// Wire up child components now that we have the concrete implementations.
-	m.input = NewInputComponent(width, "Enter your prompt (Type /help for commands, Ctrl+C to quit)", appCtrl)
+	m.input = NewInputComponent(width, appCtrl)
 
 	// Wire up cwd for @file autocomplete.
 	if ic, ok := m.input.(*InputComponent); ok && opts.Cwd != "" {
@@ -1040,7 +1044,7 @@ func (m *AppModel) AddStartupMessageToScrollList() {
 	// Add a visual separator after startup info: blank line + HR + blank line.
 	// Uses a single pre-rendered item so there are no left borders on the spacing.
 	theme := style.GetTheme()
-	separator := strings.Repeat("─", 80)
+	separator := strings.Repeat("─", m.width)
 	separatorStyled := lipgloss.NewStyle().
 		Foreground(theme.Border).
 		Render(separator)
@@ -1138,6 +1142,31 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		m.state = stateInput
 		if m.setModel != nil {
 			previousModel := m.providerName + "/" + m.modelName
+
+			// Check if thinking level needs adjustment for the new model.
+			// Some models (e.g., OpenAI gpt-5.4) don't support "minimal" and require "none".
+			if m.thinkingLevel != "" && m.thinkingLevel != "off" {
+				parts := strings.SplitN(msg.ModelString, "/", 2)
+				if len(parts) == 2 {
+					modelName := parts[1]
+					currentLevel := models.ParseThinkingLevel(m.thinkingLevel)
+					if !models.IsValidThinkingLevelForModel(currentLevel, modelName) {
+						fallback := models.SuggestThinkingLevelFallback(currentLevel, modelName)
+						if fallback != models.ThinkingOff {
+							m.printSystemMessage(fmt.Sprintf(
+								"Note: Model %s doesn't support '%s' thinking level. Adjusted to '%s'.",
+								modelName, currentLevel, fallback,
+							))
+							m.thinkingLevel = string(fallback)
+							if m.setThinkingLevel != nil {
+								_ = m.setThinkingLevel(string(fallback))
+							}
+							go func() { _ = prefs.SaveThinkingLevelPreference(string(fallback)) }()
+						}
+					}
+				}
+			}
+
 			if err := m.setModel(msg.ModelString); err != nil {
 				m.printSystemMessage(fmt.Sprintf("Failed to switch model: %v", err))
 			} else {
@@ -1283,10 +1312,22 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 				m.overlayResponseCh = nil
 				m.overlay = nil
 			}
-			// Set quitting flag so View() disables alt screen for clean exit.
-			m.quitting = true
-			// Graceful quit: app.Close() is deferred in cmd/root.go.
-			return m, tea.Quit
+
+			// Second Ctrl+C within the timeout window — quit.
+			if m.ctrlCPressedOnce {
+				m.quitting = true
+				return m, tea.Quit
+			}
+
+			// First Ctrl+C — clear input if it has content, then arm the quit flag.
+			if m.state == stateInput {
+				if ic, ok := m.input.(*InputComponent); ok {
+					ic.Clear()
+				}
+			}
+			m.ctrlCPressedOnce = true
+			// Start reset timer so the flag clears after 3 seconds.
+			return m, ctrlCResetCmd()
 		}
 
 		// Check extension-registered global keyboard shortcuts. These fire
@@ -1318,11 +1359,11 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 					m.scrollList.autoScroll = true
 				}
 				return m, tea.Batch(cmds...)
-			case "alt+home":
+			case "ctrl+home":
 				m.scrollList.GotoTop()
 				m.scrollList.autoScroll = false
 				return m, tea.Batch(cmds...)
-			case "alt+end":
+			case "ctrl+end":
 				m.scrollList.GotoBottom()
 				m.scrollList.autoScroll = true
 				return m, tea.Batch(cmds...)
@@ -1330,15 +1371,10 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		}
 
 		// Thinking keybindings — only when the model supports reasoning.
+		// Note: thinking visibility toggle is under leader chord (Ctrl+X t)
+		// to avoid conflicts with terminal multiplexers.
 		if m.isReasoningModel {
 			switch msg.String() {
-			case "ctrl+t":
-				// Toggle thinking block visibility.
-				m.thinkingVisible = !m.thinkingVisible
-				if m.stream != nil {
-					m.stream.SetThinkingVisible(m.thinkingVisible)
-				}
-				return m, tea.Batch(cmds...)
 			case "shift+tab":
 				// Cycle thinking level.
 				m.cycleThinkingLevel()
@@ -1439,6 +1475,14 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 						}
 					}
 				}
+			case "t":
+				// Ctrl+X t → Toggle thinking block visibility.
+				if m.isReasoningModel {
+					m.thinkingVisible = !m.thinkingVisible
+					if m.stream != nil {
+						m.stream.SetThinkingVisible(m.thinkingVisible)
+					}
+				}
 			case "e":
 				// Ctrl+X e → open $EDITOR to compose/edit the prompt.
 				editorApp := os.Getenv("VISUAL")
@@ -1561,10 +1605,16 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 	case uicore.CancelTimerExpiredMsg:
 		m.canceling = false
 
+	// ── Ctrl+C reset timer expired ────────────────────────────────────────────
+	case uicore.CtrlCResetMsg:
+		m.ctrlCPressedOnce = false
+
 	// ── Input submitted ──────────────────────────────────────────────────────
 	case uicore.SubmitMsg:
 		// Re-enable auto-scroll when user submits a new message.
 		m.scrollList.autoScroll = true
+		// Reset Ctrl+C flag so next Ctrl+C clears input instead of quitting.
+		m.ctrlCPressedOnce = false
 
 		// Handle slash commands locally — they should never reach app.Run().
 		// Parse once: split on the first space so argument-bearing commands
@@ -1831,6 +1881,10 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		} else {
 			bashItem.AppendStdout(msg.Chunk)
 		}
+		// Invalidate cached height after mutation.
+		if m.scrollList != nil {
+			m.scrollList.InvalidateItemHeight(bashItem.ID())
+		}
 
 		// Check height and cap if needed - we don't want streaming output to grow forever
 		const maxStreamingBashHeight = 20 // Max lines to show during streaming
@@ -2022,6 +2076,12 @@ func (m *AppModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		m.providerName = msg.ProviderName
 		m.modelName = msg.ModelName
 
+	case app.UsageUpdatedEvent:
+		// Token usage was updated after a completed LLM step. No state
+		// changes needed — the UsageTracker was already mutated in-place.
+		// Returning from Update() triggers View() which re-renders the
+		// status bar with the latest token counts, cost, and context %.
+
 	case app.WidgetUpdateEvent:
 		// Extension widget changed — recalculate height distribution so the
 		// stream region accounts for widget space. View() will read the
@@ -2390,8 +2450,10 @@ func (m *AppModel) View() tea.View {
 	scrollbackView := m.renderScrollback()
 
 	// Propagate hint visibility to the input component before rendering.
+	// Hints are hidden by default for a cleaner UI; extensions cannot
+	// override this.
 	if ic, ok := m.input.(*InputComponent); ok {
-		ic.hideHint = vis.HideInputHint
+		ic.hideHint = true
 		ic.agentBusy = m.state == stateWorking
 	}
 
@@ -2433,6 +2495,14 @@ func (m *AppModel) View() tea.View {
 		parts = append(parts, warning)
 	}
 
+	if m.ctrlCPressedOnce {
+		warning := lipgloss.NewStyle().
+			Foreground(theme.Warning).
+			Bold(true).
+			Render("  ⚠ Press Ctrl+C again to quit")
+		parts = append(parts, warning)
+	}
+
 	if !vis.HideSeparator {
 		parts = append(parts, m.renderSeparator())
 	}
@@ -2577,9 +2647,14 @@ func (m *AppModel) renderStatusBar() string {
 		middleSide = "  " + middleSide
 	}
 
-	// Right side: provider · model + usage stats.
+	// Right side: help hint + provider · model + usage stats.
+	// Order matters for progressive truncation — least important first.
 	var rightParts []string
 
+	rightParts = append(rightParts, lipgloss.NewStyle().
+		Foreground(theme.VeryMuted).
+		Render("/help for help"))
+
 	var modelLabel string
 	if m.providerName != "" && m.modelName != "" {
 		modelLabel = m.providerName + " · " + m.modelName
@@ -2598,11 +2673,11 @@ func (m *AppModel) renderStatusBar() string {
 		}
 	}
 
-	rightSide := strings.Join(rightParts, "  ")
+	rightSide := strings.Join(rightParts, "  |  ")
 
 	// Progressive truncation to keep the status bar on one line.
 	// When content exceeds terminal width, drop sections in order:
-	// middle (extensions/thinking) → usage stats → model label → right side.
+	// middle (extensions/thinking) → help hint → usage → model → all.
 	leftW := lipgloss.Width(leftSide)
 	middleW := lipgloss.Width(middleSide)
 	rightW := lipgloss.Width(rightSide)
@@ -2613,13 +2688,19 @@ func (m *AppModel) renderStatusBar() string {
 		middleSide = ""
 		middleW = 0
 	}
+	if leftW+rightW+1 > m.width && len(rightParts) > 2 {
+		// Drop help hint first.
+		rightParts = rightParts[1:]
+		rightSide = strings.Join(rightParts, "  |  ")
+		rightW = lipgloss.Width(rightSide)
+	}
 	if leftW+rightW+1 > m.width && len(rightParts) > 1 {
-		// Drop usage stats, keep model label.
-		rightSide = rightParts[0]
+		// Drop usage (last) next, keep model label.
+		rightParts = rightParts[:len(rightParts)-1]
+		rightSide = strings.Join(rightParts, "  |  ")
 		rightW = lipgloss.Width(rightSide)
 	}
 	if leftW+rightW+1 > m.width {
-		// Drop right side entirely.
 		rightSide = ""
 		rightW = 0
 	}
@@ -2630,7 +2711,7 @@ func (m *AppModel) renderStatusBar() string {
 
 // cycleThinkingLevel advances to the next thinking level and applies it.
 func (m *AppModel) cycleThinkingLevel() {
-	levels := []string{"off", "minimal", "low", "medium", "high"}
+	levels := []string{"off", "none", "minimal", "low", "medium", "high"}
 	current := m.thinkingLevel
 	if current == "" {
 		current = "off"
@@ -2663,7 +2744,7 @@ func (m *AppModel) cycleThinkingLevel() {
 // renderSeparator renders the separator line with an optional queue/steer count badge.
 func (m *AppModel) renderSeparator() string {
 	theme := style.GetTheme()
-	lineStyle := lipgloss.NewStyle().Foreground(theme.Muted)
+	lineStyle := lipgloss.NewStyle().Foreground(theme.Border)
 	queueLen := len(m.queuedMessages)
 	steerLen := len(m.steeringMessages)
 
@@ -3040,6 +3121,16 @@ func (m *AppModel) printSystemMessage(text string) {
 	m.refreshContent()
 }
 
+// printCustomMessage renders a message with a custom alert label into the ScrollList.
+func (m *AppModel) printCustomMessage(text, label string) {
+	styledMsg := m.renderer.RenderCustomMessage(text, label, time.Now())
+
+	msg := NewStyledMessageItem(generateMessageID(), "system", styledMsg.Content, styledMsg.Content)
+	m.messages = append(m.messages, msg)
+
+	m.refreshContent()
+}
+
 // printExtensionBlock renders a custom styled block from an extension with
 // caller-chosen border color and optional subtitle into the ScrollList.
 func (m *AppModel) printExtensionBlock(evt app.ExtensionPrintEvent) {
@@ -3419,13 +3510,14 @@ func (m *AppModel) printHelpMessage() {
 		"- `!command`: Run shell command, output included in LLM context\n" +
 		"- `!!command`: Run shell command, output excluded from LLM context\n\n" +
 		"**Keys:**\n" +
-		"- `Ctrl+C`: Exit at any time\n" +
+		"- `Ctrl+C`: Clear input and arm quit (press again to exit)\n" +
 		"- `ESC` (x2): Cancel ongoing LLM generation\n" +
 		"- `Ctrl+X s`: Steer — redirect the agent mid-turn (injected between tool calls)\n" +
 		"- `Ctrl+X e`: Open `$EDITOR` to compose/edit your prompt\n" +
+		"- `Ctrl+V`: Paste image from clipboard\n" +
 		"- `Enter` (while working): Queue message for after the agent finishes\n\n" +
 		"You can also just type your message to chat with the AI assistant."
-	m.printSystemMessage(help)
+	m.printCustomMessage(help, "Help")
 }
 
 // printToolsMessage renders the list of available tools.
@@ -3614,6 +3706,10 @@ func (m *AppModel) appendStreamingChunk(role, content string) {
 	// If last message is a StreamingMessageItem with matching role, append to it
 	if streamMsg, ok := lastMsg.(*StreamingMessageItem); ok && streamMsg.role == role {
 		streamMsg.AppendChunk(content)
+		// Invalidate cached height so GotoBottom sees the new size.
+		if m.scrollList != nil {
+			m.scrollList.InvalidateItemHeight(streamMsg.ID())
+		}
 		// Auto-scroll to bottom if enabled (iteratr pattern)
 		// Don't call SetItems() - the slice reference hasn't changed
 		if m.scrollList != nil {
@@ -3678,15 +3774,16 @@ func (m *AppModel) distributeHeight() {
 	}
 
 	// Propagate hint visibility before measuring input height.
+	// Hints are always hidden for a cleaner UI.
 	if ic, ok := m.input.(*InputComponent); ok {
-		ic.hideHint = vis.HideInputHint
+		ic.hideHint = true
 	}
 
 	// Measure the actual rendered input (or prompt overlay) height so we
 	// don't rely on a fragile constant that drifts when styling changes.
 	// Use renderInput() which includes the editor interceptor's Render
 	// wrapper so the measured height matches what View() actually renders.
-	inputLines := 9 // fallback: title(1)+margin(1)+nl(1)+textarea(3)+nl(1)+margin(1)+help(1)
+	inputLines := 8 // fallback: marginTop(1)+textarea(4)+border-chrome(2)+marginBottom(1)
 	if m.state == statePrompt && m.prompt != nil {
 		if rendered := m.prompt.Render(); rendered != "" {
 			inputLines = lipgloss.Height(rendered)
@@ -3815,6 +3912,30 @@ func (m *AppModel) handleModelCommand(args string) tea.Cmd {
 		return nil
 	}
 
+	// Check if thinking level needs adjustment for the new model.
+	// Some models (e.g., OpenAI gpt-5.4) don't support "minimal" and require "none".
+	if m.thinkingLevel != "" && m.thinkingLevel != "off" {
+		parts := strings.SplitN(args, "/", 2)
+		if len(parts) == 2 {
+			modelName := parts[1]
+			currentLevel := models.ParseThinkingLevel(m.thinkingLevel)
+			if !models.IsValidThinkingLevelForModel(currentLevel, modelName) {
+				fallback := models.SuggestThinkingLevelFallback(currentLevel, modelName)
+				if fallback != models.ThinkingOff {
+					m.printSystemMessage(fmt.Sprintf(
+						"Note: Model %s doesn't support '%s' thinking level. Adjusted to '%s'.",
+						modelName, currentLevel, fallback,
+					))
+					m.thinkingLevel = string(fallback)
+					if m.setThinkingLevel != nil {
+						_ = m.setThinkingLevel(string(fallback))
+					}
+					go func() { _ = prefs.SaveThinkingLevelPreference(string(fallback)) }()
+				}
+			}
+		}
+	}
+
 	// Direct model switch with the provided model string.
 	previousModel := m.providerName + "/" + m.modelName
 	if err := m.setModel(args); err != nil {
@@ -3919,7 +4040,7 @@ func (m *AppModel) handleThinkingCommand(args string) tea.Cmd {
 	// Parse and validate the level.
 	level := models.ParseThinkingLevel(args)
 	if string(level) != strings.ToLower(args) {
-		m.printSystemMessage(fmt.Sprintf("Unknown thinking level: %q. Use: off, minimal, low, medium, high", args))
+		m.printSystemMessage(fmt.Sprintf("Unknown thinking level: %q. Use: off, none, minimal, low, medium, high", args))
 		return nil
 	}
 
@@ -4506,6 +4627,14 @@ func cancelTimerCmd() tea.Cmd {
 	})
 }
 
+// ctrlCResetCmd returns a tea.Cmd that fires CtrlCResetMsg after 3s.
+// This resets the ctrlCPressedOnce flag so the next Ctrl+C will clear input again.
+func ctrlCResetCmd() tea.Cmd {
+	return tea.Tick(3*time.Second, func(_ time.Time) tea.Msg {
+		return uicore.CtrlCResetMsg{}
+	})
+}
+
 // --------------------------------------------------------------------------
 // Interactive prompt support
 // --------------------------------------------------------------------------
@@ -4577,9 +4706,12 @@ func (m *AppModel) updatePromptState(msg tea.Msg) (tea.Model, tea.Cmd) {
 	switch msg := msg.(type) {
 	case tea.KeyPressMsg:
 		if msg.String() == "ctrl+c" {
-			// Cancel prompt and quit the application.
+			// Cancel the prompt but don't quit — let the main handler's
+			// double-Ctrl+C logic handle quitting.
 			m.resolvePrompt(app.PromptResponse{Cancelled: true})
-			return m, tea.Quit
+			// Don't consume the keypress — re-dispatch so the main
+			// ctrl+c handler can track the double-press state.
+			return m.Update(msg)
 		}
 		result, cmd := m.prompt.Update(msg)
 		if cmd != nil {
@@ -4646,9 +4778,12 @@ func (m *AppModel) updateOverlayState(msg tea.Msg) (tea.Model, tea.Cmd) {
 	switch msg := msg.(type) {
 	case tea.KeyPressMsg:
 		if msg.String() == "ctrl+c" {
-			// Cancel overlay and quit the application.
+			// Cancel the overlay but don't quit — let the main handler's
+			// double-Ctrl+C logic handle quitting.
 			m.resolveOverlay(app.OverlayResponse{Cancelled: true})
-			return m, tea.Quit
+			// Don't consume the keypress — re-dispatch so the main
+			// ctrl+c handler can track the double-press state.
+			return m.Update(msg)
 		}
 		result, cmd := m.overlay.Update(msg)
 		if cmd != nil {

internal/ui/model_test.go

@@ -515,12 +515,12 @@ func TestWindowResize_distributeHeight(t *testing.T) {
 	ctrl := &stubAppController{}
 	m, _, _ := newTestAppModel(ctrl)
 
-	// With height=30, scroll height = 30 - 1 (separator) - 9 (input) - 1 (statusBar) = 19
+	// With height=30, scroll height = 30 - 1 (separator) - 8 (input) - 1 (statusBar) = 20
 	m = sendMsg(m, tea.WindowSizeMsg{Width: 80, Height: 30})
 	_ = m
 
-	if m.scrollList.height != 19 {
-		t.Fatalf("expected scroll list height=19, got %d", m.scrollList.height)
+	if m.scrollList.height != 20 {
+		t.Fatalf("expected scroll list height=20, got %d", m.scrollList.height)
 	}
 }
 
@@ -853,23 +853,165 @@ func TestSpinnerEvent_hideDoesNotTransitionState(t *testing.T) {
 }
 
 // --------------------------------------------------------------------------
-// ctrl+c produces tea.Quit
+// ctrl+c double-press to quit
 // --------------------------------------------------------------------------
 
-// TestCtrlC_producesQuit verifies that ctrl+c always returns a tea.Quit cmd.
+// TestCtrlC_producesQuit verifies that double ctrl+c returns a tea.Quit cmd.
 func TestCtrlC_producesQuit(t *testing.T) {
 	ctrl := &stubAppController{}
 	m, _, _ := newTestAppModel(ctrl)
 
+	// First Ctrl+C arms the quit flag.
+	updated, cmd := m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	m = updated.(*AppModel)
+	if cmd == nil {
+		t.Fatal("expected a command after first ctrl+c, got nil")
+	}
+	// Should be a reset timer, not quit.
+	msg := cmd()
+	if _, ok := msg.(core.CtrlCResetMsg); !ok {
+		t.Fatalf("expected CtrlCResetMsg after first ctrl+c, got %T", msg)
+	}
+
+	// Second Ctrl+C should quit.
+	_, cmd = m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	if cmd == nil {
+		t.Fatal("expected tea.Quit cmd on second ctrl+c, got nil")
+	}
+	msg = cmd()
+	if _, ok := msg.(tea.QuitMsg); !ok {
+		t.Fatalf("expected QuitMsg from second ctrl+c, got %T", msg)
+	}
+}
+
+// TestCtrlC_clearsInput_firstPress tests that Ctrl+C clears input on first
+// press when there's content, and requires a second press to quit.
+func TestCtrlC_clearsInput_firstPress(t *testing.T) {
+	// Create a real InputComponent to test the clear behavior
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+
+	// Replace with real InputComponent that has content
+	input := NewInputComponent(80, ctrl)
+	input.textarea.SetValue("some text content")
+	m.input = input
+
+	// First Ctrl+C should clear input, not quit
 	_, cmd := m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
 
-	if cmd == nil {
-		t.Fatal("expected tea.Quit cmd on ctrl+c, got nil")
+	// Should have cleared the input
+	if input.textarea.Value() != "" {
+		t.Fatalf("expected input to be cleared, got %q", input.textarea.Value())
+	}
+
+	// Should have set ctrlCPressedOnce flag
+	if !m.ctrlCPressedOnce {
+		t.Fatal("expected ctrlCPressedOnce to be true after first Ctrl+C")
+	}
+
+	// The command should be a ctrlCResetCmd (not tea.Quit)
+	if cmd == nil {
+		t.Fatal("expected a command after first Ctrl+C, got nil")
 	}
-	// We verify it's a quit command by running it and checking the message type.
 	msg := cmd()
+	if _, ok := msg.(core.CtrlCResetMsg); !ok {
+		t.Fatalf("expected CtrlCResetMsg, got %T", msg)
+	}
+
+	// Second Ctrl+C should now quit
+	_, cmd = m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	if cmd == nil {
+		t.Fatal("expected tea.Quit cmd on second Ctrl+C, got nil")
+	}
+	msg = cmd()
 	if _, ok := msg.(tea.QuitMsg); !ok {
-		t.Fatalf("expected QuitMsg from ctrl+c cmd, got %T", msg)
+		t.Fatalf("expected QuitMsg on second Ctrl+C, got %T", msg)
+	}
+}
+
+// TestCtrlC_resetAfterSubmit tests that the Ctrl+C flag is reset after
+// submitting a message, so the next Ctrl+C clears input again.
+func TestCtrlC_resetAfterSubmit(t *testing.T) {
+	// Use newTestAppModel but replace the input with a real InputComponent
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+
+	// Replace with real InputComponent
+	input := NewInputComponent(80, ctrl)
+	input.textarea.SetValue("content")
+	m.input = input
+
+	// First Ctrl+C clears input
+	updated, _ := m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	m = updated.(*AppModel)
+	if input.textarea.Value() != "" {
+		t.Fatal("expected input to be cleared")
+	}
+
+	// Flag should be set
+	if !m.ctrlCPressedOnce {
+		t.Fatal("expected ctrlCPressedOnce to be true after first Ctrl+C")
+	}
+
+	// Simulate CtrlCResetMsg being processed (timer expired)
+	updated, _ = m.Update(core.CtrlCResetMsg{})
+	m = updated.(*AppModel)
+
+	// Flag should be reset
+	if m.ctrlCPressedOnce {
+		t.Fatal("expected ctrlCPressedOnce to be false after CtrlCResetMsg")
+	}
+
+	// Add new content to input
+	input.textarea.SetValue("new content")
+
+	// Next Ctrl+C should clear again (not quit) because flag was reset
+	_, cmd := m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	if input.textarea.Value() != "" {
+		t.Fatalf("expected input to be cleared again, got %q", input.textarea.Value())
+	}
+	if cmd == nil {
+		t.Fatal("expected a command after Ctrl+C, got nil")
+	}
+	msg := cmd()
+	if _, ok := msg.(core.CtrlCResetMsg); !ok {
+		t.Fatalf("expected CtrlCResetMsg, got %T", msg)
+	}
+}
+
+// TestCtrlC_emptyInput_armsQuit tests that Ctrl+C on empty input still
+// requires a second press to quit (consistent double-press behavior).
+func TestCtrlC_emptyInput_armsQuit(t *testing.T) {
+	ctrl := &stubAppController{}
+	m, _, _ := newTestAppModel(ctrl)
+
+	// Replace with real InputComponent (empty by default)
+	input := NewInputComponent(80, ctrl)
+	m.input = input
+
+	// First Ctrl+C on empty input should arm the flag, not quit.
+	updated, cmd := m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	m = updated.(*AppModel)
+
+	if !m.ctrlCPressedOnce {
+		t.Fatal("expected ctrlCPressedOnce to be true after first Ctrl+C")
+	}
+	if cmd == nil {
+		t.Fatal("expected a command (reset timer), got nil")
+	}
+	msg := cmd()
+	if _, ok := msg.(core.CtrlCResetMsg); !ok {
+		t.Fatalf("expected CtrlCResetMsg, got %T", msg)
+	}
+
+	// Second Ctrl+C should quit.
+	_, cmd = m.Update(tea.KeyPressMsg{Code: 'c', Mod: tea.ModCtrl})
+	if cmd == nil {
+		t.Fatal("expected tea.Quit cmd on second Ctrl+C, got nil")
+	}
+	msg = cmd()
+	if _, ok := msg.(tea.QuitMsg); !ok {
+		t.Fatalf("expected QuitMsg on second Ctrl+C, got %T", msg)
 	}
 }
 

internal/ui/render/blocks.go

@@ -36,16 +36,16 @@ func UserBlock(content string, width int, ty *herald.Typography, theme style.The
 
 	// Highlight @file tokens with accent color so file references are
 	// visually distinct from surrounding prompt text.
-	content = highlightFileTokens(content, theme)
+	content = HighlightFileTokens(content, theme)
 
 	rendered := ty.Tip(content)
 	return styleMarginBottom(theme, rendered)
 }
 
-// highlightFileTokens wraps @file tokens in the given text with the theme
+// HighlightFileTokens wraps @file tokens in the given text with the theme
 // accent color so they stand out visually in rendered user messages.
-func highlightFileTokens(text string, theme style.Theme) string {
-	accentStyle := lipgloss.NewStyle().Foreground(theme.Accent).Bold(true)
+func HighlightFileTokens(text string, theme style.Theme) string {
+	accentStyle := style.GetCachedStyles().FileTokenAccent
 	return fileTokenPattern.ReplaceAllStringFunc(text, func(token string) string {
 		return accentStyle.Render(token)
 	})
@@ -63,16 +63,20 @@ func AssistantBlock(content string, width int, theme style.Theme) string {
 
 // ReasoningBlock renders a reasoning/thinking block with muted italic text.
 // If duration > 0, shows "Thought for Xs" label. Otherwise shows just "Thought".
-func ReasoningBlock(content string, duration int64, ty *herald.Typography, theme style.Theme) string {
+// The width parameter controls soft-wrapping so long reasoning lines don't get cut off.
+func ReasoningBlock(content string, duration int64, width int, ty *herald.Typography, theme style.Theme) string {
 	if strings.TrimSpace(content) == "" {
 		return ""
 	}
 
-	// Match live streaming styling: muted italic text
+	// Match live streaming styling: muted italic text.
 	lines := strings.Split(strings.TrimRight(content, "\n"), "\n")
 	contentStr := strings.TrimLeft(strings.Join(lines, "\n"), " \t\n")
-	mutedStyle := lipgloss.NewStyle().Foreground(theme.Muted)
-	contentRendered := mutedStyle.Render(ty.Italic(contentStr))
+	if width > 4 {
+		contentStr = wrapText(contentStr, width-4)
+	}
+	cs := style.GetCachedStyles()
+	contentRendered := cs.Muted.Render(ty.Italic(contentStr))
 
 	// Build label based on duration
 	if duration > 0 {
@@ -82,14 +86,14 @@ func ReasoningBlock(content string, duration int64, ty *herald.Typography, theme
 		} else {
 			durationStr = fmt.Sprintf("%.1fs", float64(duration)/1000)
 		}
-		labelPart := lipgloss.NewStyle().Foreground(theme.VeryMuted).Render("Thought for ")
-		durationPart := lipgloss.NewStyle().Foreground(theme.Accent).Render(durationStr)
+		labelPart := cs.VeryMuted.Render("Thought for ")
+		durationPart := cs.Accent.Render(durationStr)
 		label := labelPart + durationPart
 		rendered := contentRendered + "\n" + label
 		return styleMarginBottom(theme, rendered)
 	}
 
-	label := lipgloss.NewStyle().Foreground(theme.VeryMuted).Render("Thought")
+	label := cs.VeryMuted.Render("Thought")
 	rendered := contentRendered + "\n" + label
 
 	return styleMarginBottom(theme, rendered)
@@ -105,6 +109,45 @@ func SystemBlock(content string, ty *herald.Typography, theme style.Theme) strin
 	return styleMarginBottom(theme, rendered)
 }
 
+// CustomBlock renders a message with herald Note styling and a custom label.
+// Content is rendered as markdown before being wrapped in the alert. This
+// creates a one-off Typography instance with the given label so callers
+// can use any title (e.g. "Help", "Warning") without changing the shared
+// typography's default "Info" label.
+func CustomBlock(content, label string, width int, theme style.Theme) string {
+	if strings.TrimSpace(content) == "" {
+		content = "No content available"
+	}
+
+	// Render markdown first — subtract 4 for the alert bar prefix ("│ ").
+	mdWidth := max(width-4, 10)
+	rendered := style.ToMarkdown(content, mdWidth)
+
+	ty := 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.WithAlertLabel(herald.AlertNote, label),
+	)
+	alertRendered := ty.Note(rendered)
+	return styleMarginBottom(theme, alertRendered)
+}
+
 // ErrorBlock renders an error message with herald Caution styling.
 func ErrorBlock(errorMsg string, ty *herald.Typography, theme style.Theme) string {
 	rendered := ty.Caution(errorMsg)
@@ -151,5 +194,11 @@ func ToolBlock(displayName, params, body string, isError bool, width int, ty *he
 
 // styleMarginBottom applies a 1-line margin bottom using the theme.
 func styleMarginBottom(theme style.Theme, content string) string {
-	return lipgloss.NewStyle().MarginBottom(1).Render(content)
+	return style.GetCachedStyles().MarginBottom1.Render(content)
+}
+
+// wrapText soft-wraps a string to the given width using lipgloss, which is
+// ANSI-aware and preserves escape sequences across line breaks.
+func wrapText(s string, width int) string {
+	return lipgloss.NewStyle().Width(width).Render(s)
 }

internal/ui/render/blocks_test.go

@@ -23,7 +23,8 @@ func testTypography(theme style.Theme) *herald.Typography {
 			Surface:   theme.Background,
 			Base:      theme.CodeBg,
 		}),
-		herald.WithAlertLabel(herald.AlertTip, "You"),
+		herald.WithAlertLabel(herald.AlertTip, ""),
+		herald.WithAlertIcon(herald.AlertTip, ""),
 	)
 }
 
@@ -70,18 +71,18 @@ func TestHighlightFileTokens(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result := highlightFileTokens(tt.input, theme)
+			result := HighlightFileTokens(tt.input, theme)
 
 			for _, want := range tt.wantHas {
 				if !strings.Contains(result, want) {
-					t.Errorf("highlightFileTokens(%q) = %q, want substring %q", tt.input, result, want)
+					t.Errorf("HighlightFileTokens(%q) = %q, want substring %q", tt.input, result, want)
 				}
 			}
 
 			// If there were @tokens, the result should contain ANSI escape
 			// sequences (from lipgloss styling).
 			if fileTokenPattern.MatchString(tt.input) && !strings.Contains(result, "\x1b[") {
-				t.Errorf("highlightFileTokens(%q) should contain ANSI escapes for @tokens but got %q", tt.input, result)
+				t.Errorf("HighlightFileTokens(%q) should contain ANSI escapes for @tokens but got %q", tt.input, result)
 			}
 		})
 	}

internal/ui/scrolllist.go

@@ -35,6 +35,12 @@ type ScrollList struct {
 	autoScroll bool // Whether to auto-scroll to bottom on new content
 	itemGap    int  // Number of blank lines between items (0 = no gap)
 
+	// heightCache maps item ID → rendered line count at current width.
+	// Avoids redundant Render() calls in GotoBottom/clampOffset/AtBottom.
+	// Invalidated on width change; individual entries are refreshed in
+	// View() when an item is actually rendered.
+	heightCache map[string]int
+
 	// Character-level text selection (crush-style).
 	sel selection.State
 }
@@ -42,13 +48,14 @@ type ScrollList struct {
 // NewScrollList creates a new ScrollList with the given dimensions.
 func NewScrollList(width, height int) *ScrollList {
 	return &ScrollList{
-		items:      []MessageItem{},
-		offsetIdx:  0,
-		offsetLine: 0,
-		width:      width,
-		height:     height,
-		autoScroll: true,
-		sel:        selection.NewState(),
+		items:       []MessageItem{},
+		offsetIdx:   0,
+		offsetLine:  0,
+		width:       width,
+		height:      height,
+		autoScroll:  true,
+		heightCache: make(map[string]int, 64),
+		sel:         selection.NewState(),
 	}
 }
 
@@ -61,6 +68,13 @@ func (s *ScrollList) SetItems(items []MessageItem) {
 	}
 }
 
+// InvalidateItemHeight removes the cached height for the given item ID,
+// forcing a re-render on the next height query. Call this after mutating
+// an item's content (e.g. AppendChunk on a streaming message).
+func (s *ScrollList) InvalidateItemHeight(id string) {
+	delete(s.heightCache, id)
+}
+
 // SetHeight updates the viewport height. Called when the terminal is resized.
 func (s *ScrollList) SetHeight(height int) {
 	s.height = height
@@ -68,9 +82,11 @@ func (s *ScrollList) SetHeight(height int) {
 }
 
 // SetWidth updates the viewport width. Called when the terminal is resized.
-// This may invalidate cached renders in MessageItems.
+// This invalidates the height cache since rendered heights are width-dependent.
 func (s *ScrollList) SetWidth(width int) {
 	s.width = width
+	// Width change invalidates all cached heights.
+	clear(s.heightCache)
 	s.clampOffset()
 }
 
@@ -338,9 +354,8 @@ func (s *ScrollList) ScrollBy(lines int) {
 			if s.offsetIdx >= len(s.items) {
 				break
 			}
-			currentItem := s.items[s.offsetIdx]
-			itemHeight := currentItem.Height()
-			remainingLines := itemHeight - s.offsetLine
+			ih := s.itemHeight(s.items[s.offsetIdx])
+			remainingLines := ih - s.offsetLine
 
 			if lines >= remainingLines {
 				// Move to next item
@@ -387,14 +402,13 @@ func (s *ScrollList) ScrollBy(lines int) {
 				// Move to previous item
 				s.offsetIdx--
 				if s.offsetIdx < len(s.items) {
-					currentItem := s.items[s.offsetIdx]
-					itemHeight := currentItem.Height()
+					ih := s.itemHeight(s.items[s.offsetIdx])
 
-					if lines >= itemHeight {
-						lines -= itemHeight
+					if lines >= ih {
+						lines -= ih
 						s.offsetLine = 0
 					} else {
-						s.offsetLine = itemHeight - lines
+						s.offsetLine = ih - lines
 						lines = 0
 					}
 				}
@@ -405,6 +419,8 @@ func (s *ScrollList) ScrollBy(lines int) {
 }
 
 // GotoBottom scrolls to the end of the list.
+// Uses cached heights and walks backwards from the end to avoid rendering
+// every item in the list.
 func (s *ScrollList) GotoBottom() {
 	if len(s.items) == 0 {
 		s.offsetIdx = 0
@@ -412,42 +428,31 @@ func (s *ScrollList) GotoBottom() {
 		return
 	}
 
-	// Calculate total height including gaps
-	totalHeight := 0
-	for i, item := range s.items {
-		rendered := item.Render(s.width)
-		itemHeight := strings.Count(rendered, "\n") + 1
-		totalHeight += itemHeight
-		if s.itemGap > 0 && i < len(s.items)-1 {
-			totalHeight += s.itemGap
+	// Walk backwards from the last item, accumulating height until we
+	// exceed the viewport. This is O(visible) instead of O(all items).
+	budget := s.height
+	for idx := len(s.items) - 1; idx >= 0; idx-- {
+		ih := s.itemHeight(s.items[idx])
+
+		// Account for gap *above* this item (gap between idx-1 and idx).
+		gap := 0
+		if s.itemGap > 0 && idx < len(s.items)-1 {
+			gap = s.itemGap
 		}
-	}
 
-	// If content fits in viewport, start at top
-	if totalHeight <= s.height {
-		s.offsetIdx = 0
-		s.offsetLine = 0
-		return
-	}
-
-	// Otherwise, position viewport at bottom
-	remaining := totalHeight - s.height
-	for idx := 0; idx < len(s.items); idx++ {
-		rendered := s.items[idx].Render(s.width)
-		itemHeight := strings.Count(rendered, "\n") + 1
-		if remaining < itemHeight {
+		if ih+gap >= budget {
+			// This item (partially) fills the remaining budget.
+			// When the gap consumed part of the budget, offsetLine would go
+			// negative — clamp to 0 so the item is shown fully.
 			s.offsetIdx = idx
-			s.offsetLine = remaining
+			s.offsetLine = max(0, ih-budget)
 			return
 		}
-		remaining -= itemHeight
-		if s.itemGap > 0 && idx < len(s.items)-1 {
-			remaining -= s.itemGap
-		}
+		budget -= ih + gap
 	}
 
-	// Fallback: show last item
-	s.offsetIdx = max(0, len(s.items)-1)
+	// All content fits in viewport — start at top.
+	s.offsetIdx = 0
 	s.offsetLine = 0
 }
 
@@ -465,14 +470,12 @@ func (s *ScrollList) AtBottom() bool {
 
 	visibleHeight := 0
 	for idx := s.offsetIdx; idx < len(s.items); idx++ {
-		item := s.items[idx]
-		rendered := item.Render(s.width)
-		itemHeight := strings.Count(rendered, "\n") + 1
+		ih := s.itemHeight(s.items[idx])
 
 		if idx == s.offsetIdx {
-			visibleHeight += itemHeight - s.offsetLine
+			visibleHeight += ih - s.offsetLine
 		} else {
-			visibleHeight += itemHeight
+			visibleHeight += ih
 		}
 
 		if s.itemGap > 0 && idx < len(s.items)-1 {
@@ -520,6 +523,9 @@ func (s *ScrollList) View() string {
 			content := item.Render(s.width)
 			contentLines := strings.Split(content, "\n")
 
+			// Refresh height cache from the actual render (authoritative).
+			s.heightCache[item.ID()] = len(contentLines)
+
 			startLine := 0
 			if idx == s.offsetIdx {
 				startLine = s.offsetLine
@@ -568,7 +574,7 @@ func (s *ScrollList) ScrollPercent() float64 {
 
 	totalHeight := 0
 	for _, item := range s.items {
-		totalHeight += item.Height()
+		totalHeight += s.itemHeight(item)
 	}
 
 	if totalHeight <= s.height {
@@ -577,7 +583,7 @@ func (s *ScrollList) ScrollPercent() float64 {
 
 	linesAbove := 0
 	for i := 0; i < s.offsetIdx && i < len(s.items); i++ {
-		linesAbove += s.items[i].Height()
+		linesAbove += s.itemHeight(s.items[i])
 	}
 	linesAbove += s.offsetLine
 
@@ -597,7 +603,8 @@ func (s *ScrollList) ScrollPercent() float64 {
 }
 
 // clampOffset ensures the offset values are within valid bounds after
-// resizing or scrolling operations.
+// resizing or scrolling operations. Uses cached heights to avoid
+// redundant Render() calls.
 func (s *ScrollList) clampOffset() {
 	if len(s.items) == 0 {
 		s.offsetIdx = 0
@@ -605,6 +612,7 @@ func (s *ScrollList) clampOffset() {
 		return
 	}
 
+	// Clamp offsetIdx to valid item range.
 	if s.offsetIdx >= len(s.items) {
 		s.offsetIdx = len(s.items) - 1
 	}
@@ -612,37 +620,38 @@ func (s *ScrollList) clampOffset() {
 		s.offsetIdx = 0
 	}
 
+	// Clamp offsetLine within current item.
 	if s.offsetIdx < len(s.items) {
-		rendered := s.items[s.offsetIdx].Render(s.width)
-		itemHeight := strings.Count(rendered, "\n") + 1
-		if s.offsetLine >= itemHeight {
-			s.offsetLine = max(0, itemHeight-1)
+		ih := s.itemHeight(s.items[s.offsetIdx])
+		if s.offsetLine >= ih {
+			s.offsetLine = max(0, ih-1)
 		}
 	}
 	if s.offsetLine < 0 {
 		s.offsetLine = 0
 	}
 
-	// Prevent scrolling past the bottom
+	// Prevent scrolling past the bottom — compute total height and check
+	// whether remaining content from the current offset fills the viewport.
 	totalHeight := 0
 	for i, item := range s.items {
-		rendered := item.Render(s.width)
-		totalHeight += strings.Count(rendered, "\n") + 1
+		totalHeight += s.itemHeight(item)
 		if s.itemGap > 0 && i < len(s.items)-1 {
 			totalHeight += s.itemGap
 		}
 	}
 
+	// If content fits in viewport, force start at top.
 	if totalHeight <= s.height {
 		s.offsetIdx = 0
 		s.offsetLine = 0
 		return
 	}
 
+	// Compute lines above the viewport.
 	linesAbove := 0
 	for i := 0; i < s.offsetIdx; i++ {
-		rendered := s.items[i].Render(s.width)
-		linesAbove += strings.Count(rendered, "\n") + 1
+		linesAbove += s.itemHeight(s.items[i])
 		if s.itemGap > 0 && i < len(s.items)-1 {
 			linesAbove += s.itemGap
 		}
@@ -651,20 +660,21 @@ func (s *ScrollList) clampOffset() {
 
 	linesFromCurrentToEnd := totalHeight - linesAbove
 	if linesFromCurrentToEnd < s.height {
+		// We've scrolled past the bottom — reposition so the last line
+		// of content sits at the bottom of the viewport.
 		targetLine := totalHeight - s.height
 		currentLine := 0
 
 		for idx := 0; idx < len(s.items); idx++ {
-			rendered := s.items[idx].Render(s.width)
-			itemHeight := strings.Count(rendered, "\n") + 1
+			ih := s.itemHeight(s.items[idx])
 
-			if currentLine+itemHeight > targetLine {
+			if currentLine+ih > targetLine {
 				s.offsetIdx = idx
 				s.offsetLine = targetLine - currentLine
 				return
 			}
 
-			currentLine += itemHeight
+			currentLine += ih
 			if s.itemGap > 0 && idx < len(s.items)-1 {
 				currentLine += s.itemGap
 			}
@@ -672,6 +682,26 @@ func (s *ScrollList) clampOffset() {
 	}
 }
 
+// itemHeight returns the cached rendered height for an item, computing and
+// caching it on first access. This avoids calling Render() purely to
+// count lines — the most common source of redundant work in the scroll
+// list (GotoBottom, clampOffset, AtBottom, ScrollBy all need heights but
+// never use the rendered content).
+//
+// The cache is invalidated wholesale on width changes (SetWidth) and
+// individual entries are refreshed in View() after an item is actually
+// rendered, so stale entries are self-correcting within one frame.
+func (s *ScrollList) itemHeight(item MessageItem) int {
+	id := item.ID()
+	if h, ok := s.heightCache[id]; ok {
+		return h
+	}
+	// Cache miss — render to measure.
+	h := s.renderedHeight(item)
+	s.heightCache[id] = h
+	return h
+}
+
 // renderedHeight returns the height of a message item in lines by actually
 // rendering it. This is the single source of truth for item height — it
 // matches exactly what View() produces, unlike item.Height() which may

internal/ui/stream.go

@@ -21,12 +21,11 @@ func knightRiderFrames() []string {
 	const numDots = 8
 	const dot = "▪"
 
-	theme := style.GetTheme()
-
-	bright := lipgloss.NewStyle().Foreground(theme.Primary)
-	med := lipgloss.NewStyle().Foreground(theme.Muted)
-	dim := lipgloss.NewStyle().Foreground(theme.VeryMuted)
-	off := lipgloss.NewStyle().Foreground(theme.MutedBorder)
+	cs := style.GetCachedStyles()
+	bright := cs.SpinnerBright
+	med := cs.SpinnerMed
+	dim := cs.SpinnerDim
+	off := cs.SpinnerOff
 
 	// Scanner bounces: 0→7→0
 	positions := make([]int, 0, 2*numDots-2)
@@ -472,9 +471,12 @@ func (s *StreamComponent) renderReasoningBlock(reasoning string) string {
 
 	// 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)))
+	// Soft-wrap to the available width so long lines don't get cut off.
+	if s.width > 4 {
+		content = lipgloss.NewStyle().Width(s.width - 4).Render(content)
+	}
+	cs := style.GetCachedStyles()
+	parts = append(parts, cs.Muted.Render(s.ty.Italic(content)))
 
 	// Duration footer with VeryMuted label and Accent duration.
 	var duration time.Duration
@@ -490,8 +492,8 @@ func (s *StreamComponent) renderReasoningBlock(reasoning string) string {
 		} else {
 			durationStr = fmt.Sprintf("%.1fs", duration.Seconds())
 		}
-		label := lipgloss.NewStyle().Foreground(theme.VeryMuted).Render("Thought for ")
-		durationStyled := lipgloss.NewStyle().Foreground(theme.Accent).Render(durationStr)
+		label := cs.VeryMuted.Render("Thought for ")
+		durationStyled := cs.Accent.Render(durationStr)
 		parts = append(parts, label+durationStyled)
 	}
 
@@ -588,8 +590,10 @@ func formatToolExecutionMessage(toolName string) string {
 	return toolName
 }
 
-// UpdateTheme refreshes the component's typography instance with colors from
-// the current theme. This is called when the user changes themes via /theme.
+// UpdateTheme refreshes the component's typography instance and spinner
+// animation frames with colors from the current theme. This is called when
+// the user changes themes via /theme.
 func (s *StreamComponent) UpdateTheme() {
 	s.ty = createTypography(GetTheme())
+	s.spinnerFrames = knightRiderFrames()
 }

internal/ui/style/enhanced.go

@@ -40,6 +40,70 @@ func GetTheme() Theme {
 func SetTheme(theme Theme) {
 	currentTheme = theme
 	markdownTypographyCache = nil // invalidate cached renderer; colors may have changed
+	styleCache = nil              // invalidate cached styles; colors may have changed
+}
+
+// CachedStyles holds pre-built lipgloss styles that are reused across
+// render frames. Invalidated by SetTheme, lazily rebuilt on next access.
+// Only accessed from BubbleTea's single-threaded Update/View cycle.
+type CachedStyles struct {
+	// render/blocks.go
+	FileTokenAccent lipgloss.Style // Foreground(Accent).Bold(true)
+	Muted           lipgloss.Style // Foreground(Muted)
+	VeryMuted       lipgloss.Style // Foreground(VeryMuted)
+	Accent          lipgloss.Style // Foreground(Accent)
+	MarginBottom1   lipgloss.Style // MarginBottom(1)
+
+	// stream.go - spinner phases
+	SpinnerBright lipgloss.Style // Foreground(Primary)
+	SpinnerMed    lipgloss.Style // Foreground(Muted)
+	SpinnerDim    lipgloss.Style // Foreground(VeryMuted)
+	SpinnerOff    lipgloss.Style // Foreground(MutedBorder)
+
+	// message_items.go - bash output
+	BashHeader lipgloss.Style // Foreground(Muted).Italic(true)
+	BashStderr lipgloss.Style // Foreground(Error)
+
+	// render/blocks.go - tool block
+	ToolSuccess lipgloss.Style // Foreground(Success)
+	ToolError   lipgloss.Style // Foreground(Error)
+	ToolInfo    lipgloss.Style // Foreground(Info).Bold(true)
+	ToolMuted   lipgloss.Style // Foreground(Muted)
+
+	// common
+	ErrorFg  lipgloss.Style // Foreground(Error)
+	TextBold lipgloss.Style // Foreground(Text).Bold(true)
+}
+
+var styleCache *CachedStyles
+
+// GetCachedStyles returns the pre-built style cache, creating it lazily
+// from the current theme. Invalidated by SetTheme.
+func GetCachedStyles() *CachedStyles {
+	if styleCache != nil {
+		return styleCache
+	}
+	theme := GetTheme()
+	styleCache = &CachedStyles{
+		FileTokenAccent: lipgloss.NewStyle().Foreground(theme.Accent).Bold(true),
+		Muted:           lipgloss.NewStyle().Foreground(theme.Muted),
+		VeryMuted:       lipgloss.NewStyle().Foreground(theme.VeryMuted),
+		Accent:          lipgloss.NewStyle().Foreground(theme.Accent),
+		MarginBottom1:   lipgloss.NewStyle().MarginBottom(1),
+		SpinnerBright:   lipgloss.NewStyle().Foreground(theme.Primary),
+		SpinnerMed:      lipgloss.NewStyle().Foreground(theme.Muted),
+		SpinnerDim:      lipgloss.NewStyle().Foreground(theme.VeryMuted),
+		SpinnerOff:      lipgloss.NewStyle().Foreground(theme.MutedBorder),
+		BashHeader:      lipgloss.NewStyle().Foreground(theme.Muted).Italic(true),
+		BashStderr:      lipgloss.NewStyle().Foreground(theme.Error),
+		ToolSuccess:     lipgloss.NewStyle().Foreground(theme.Success),
+		ToolError:       lipgloss.NewStyle().Foreground(theme.Error),
+		ToolInfo:        lipgloss.NewStyle().Foreground(theme.Info).Bold(true),
+		ToolMuted:       lipgloss.NewStyle().Foreground(theme.Muted),
+		ErrorFg:         lipgloss.NewStyle().Foreground(theme.Error),
+		TextBold:        lipgloss.NewStyle().Foreground(theme.Text).Bold(true),
+	}
+	return styleCache
 }
 
 // MarkdownThemeColors defines colors for markdown rendering and syntax highlighting.

internal/ui/tool_renderers.go

@@ -79,8 +79,7 @@ func renderToolBody(toolName, toolArgs, toolResult string, width int) string {
 // Edit tool — side-by-side diff
 // ---------------------------------------------------------------------------
 
-// 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).
+// renderEditBody renders a side-by-side diff from the edits array in toolArgs.
 func renderEditBody(toolArgs, toolResult string, width int) string {
 	var args map[string]any
 	if err := json.Unmarshal([]byte(toolArgs), &args); err != nil {
@@ -90,35 +89,28 @@ func renderEditBody(toolArgs, toolResult string, width int) string {
 	// 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)
-					}
+	editsArr, ok := args["edits"].([]any)
+	if !ok || len(editsArr) == 0 {
+		return ""
+	}
+
+	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 ""
+	if len(results) > 0 {
+		return strings.Join(results, "\n")
 	}
-
-	return renderDiffBlock(oldText, newText, startLine, width)
+	return ""
 }
 
 // extractDiffStartLine parses the first @@ hunk header from a unified diff

internal/ui/tree_selector.go

@@ -200,10 +200,6 @@ func (ts *TreeSelectorComponent) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		case key.Matches(msg, key.NewBinding(key.WithKeys("ctrl+l"))):
 			ts.filter = TreeFilterLabelOnly
 			ts.rebuildFlatList()
-		case key.Matches(msg, key.NewBinding(key.WithKeys("ctrl+a"))):
-			ts.filter = TreeFilterAll
-			ts.rebuildFlatList()
-
 		default:
 			// Typing search.
 			if msg.Text != "" && len(msg.Text) == 1 {

pkg/kit/README.md

@@ -106,7 +106,7 @@ unsub2 := host.OnToolResult(func(e kit.ToolResultEvent) {
 })
 defer unsub2()
 
-unsub3 := host.OnStreaming(func(e kit.MessageUpdateEvent) {
+unsub3 := host.OnMessageUpdate(func(e kit.MessageUpdateEvent) {
     fmt.Print(e.Chunk)
 })
 defer unsub3()
@@ -190,6 +190,41 @@ msg, err := host.GetMCPPrompt(ctx, "server-name", "prompt-name", map[string]stri
 })
 ```
 
+### MCP Tasks (long-running tools)
+
+Kit advertises [MCP task support](https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks)
+during `initialize`. Cooperating servers can respond to `tools/call` with a
+`taskId` immediately; Kit then polls `tasks/get` / `tasks/result` until the
+task reaches a terminal state, and best-effort `tasks/cancel`s on context
+cancellation. Servers that don't advertise the capability keep their previous
+synchronous behaviour.
+
+```go
+host, _ := kit.New(ctx, &kit.Options{
+    // Per-server mode: auto (default), never, or always.
+    MCPTaskMode: map[string]kit.MCPTaskMode{
+        "build-server": kit.MCPTaskModeAlways,
+    },
+    MCPTaskTimeout:  15 * time.Minute, // total wall-clock cap
+    MCPTaskProgress: func(p kit.MCPTaskProgress) {
+        log.Printf("%s/%s: %s", p.Server, p.TaskID, p.Status)
+    },
+})
+
+// Inspect / cancel in-flight tasks
+tasks, _ := host.ListMCPTasks(ctx, "build-server")
+t, _    := host.GetMCPTask(ctx, "build-server", tasks[0].TaskID)
+if !t.Status.IsTerminal() {
+    _, _ = host.CancelMCPTask(ctx, "build-server", t.TaskID)
+}
+```
+
+The progress handler fires once when a task is accepted and again on every
+observed status transition; the final invocation always carries a terminal
+status (`MCPTaskStatusCompleted`, `MCPTaskStatusFailed`, or
+`MCPTaskStatusCancelled`). Don't block in the handler — dispatch long work on
+a goroutine.
+
 ### Session Management
 
 Maintain conversation context:
@@ -224,10 +259,13 @@ kit.LLMResponse     // {Content, FinishReason, Usage}
 kit.LLMFilePart     // {Filename, Data []byte, MediaType}
 
 // MCP OAuth types
-kit.MCPServer        // *server.MCPServer for in-process MCP transport
-kit.MCPServerConfig  // Configuration for an MCP server (stdio, SSE, or in-process)
-kit.MCPTokenStore    // Persists OAuth tokens for a single MCP server
-kit.MCPToken         // OAuth token (access token, refresh token, expiry)
+kit.MCPServer            // *server.MCPServer for in-process MCP transport
+kit.MCPServerConfig      // Configuration for an MCP server (stdio, SSE, or in-process)
+kit.MCPAuthHandler       // Interface: handles user-facing OAuth authorization
+kit.DefaultMCPAuthHandler // Port + callback-server mechanics; set OnAuthURL for presentation
+kit.CLIMCPAuthHandler    // CLI wrapper: opens browser, prints status
+kit.MCPTokenStore        // Persists OAuth tokens for a single MCP server
+kit.MCPToken             // OAuth token (access token, refresh token, expiry)
 kit.MCPTokenStoreFactory // Creates an MCPTokenStore for a given server URL
 
 // Conversion helpers

pkg/kit/config.go

@@ -38,20 +38,37 @@ Guidelines:
 - Be concise in your responses
 - Show file paths clearly when working with files`
 
-// setSDKDefaults registers the same viper defaults that the CLI sets via
-// cobra flag bindings. This ensures the SDK behaves identically to the CLI
-// even when cobra is not used.
+// sdkDefaultMaxTokens is the last-resort ceiling applied when the SDK caller
+// has not configured max-tokens via Options, env, config, or a per-model
+// default. It matches the CLI's --max-tokens cobra default so SDK and CLI
+// callers see the same base value before per-model right-sizing runs.
+// It is intentionally applied on the *models.ProviderConfig struct
+// (not via viper) so that viper.IsSet("max-tokens") remains false and the
+// right-sizing + per-model-default paths continue to work.
+const sdkDefaultMaxTokens = 8192
+
+// setSDKDefaults registers viper defaults that match the CLI's cobra flag
+// defaults for keys where SetDefault does not interfere with downstream
+// viper.IsSet() checks.
+//
+// Keys that participate in "explicit vs unset" precedence downstream —
+// max-tokens, temperature, top-p, top-k, frequency-penalty, presence-penalty,
+// thinking-level — are deliberately NOT registered here. viper.SetDefault
+// causes viper.IsSet() to return true, which would suppress per-model
+// defaults (ApplyModelSettings) and automatic right-sizing (rightSizeMaxTokens)
+// for every SDK-created Kit. Those defaults are instead applied:
+//
+//   - max-tokens: as a last-resort struct-level floor (sdkDefaultMaxTokens)
+//     in kit.New() after BuildProviderConfig returns, when the resolved
+//     value is still zero.
+//   - thinking-level: handled implicitly by models.ParseThinkingLevel("")
+//     which returns models.ThinkingOff.
+//   - sampling params (temperature, top-p, top-k, frequency/presence-penalty):
+//     left as nil pointers so provider libraries apply their own defaults.
 func setSDKDefaults() {
 	viper.SetDefault("model", "anthropic/claude-sonnet-4-5-20250929")
 	viper.SetDefault("system-prompt", defaultSystemPrompt)
-	viper.SetDefault("max-tokens", 4096)
-	viper.SetDefault("temperature", 0.7)
-	viper.SetDefault("top-p", 0.95)
-	viper.SetDefault("top-k", 40)
-	viper.SetDefault("frequency-penalty", 0.0)
-	viper.SetDefault("presence-penalty", 0.0)
 	viper.SetDefault("stream", true)
-	viper.SetDefault("thinking-level", "off")
 	viper.SetDefault("num-gpu-layers", -1)
 	viper.SetDefault("main-gpu", 0)
 }
@@ -102,6 +119,10 @@ func InitConfig(configFile string, debug bool) error {
 	}
 
 	viper.SetEnvPrefix("KIT")
+	// Map hyphenated config keys (e.g. "max-tokens") to underscored env
+	// var names (e.g. KIT_MAX_TOKENS). Without this, AutomaticEnv looks
+	// for KIT_MAX-TOKENS and silently misses valid env overrides.
+	viper.SetEnvKeyReplacer(strings.NewReplacer("-", "_"))
 	viper.AutomaticEnv()
 	return nil
 }

pkg/kit/events.go

@@ -23,6 +23,14 @@ const (
 	EventMessageUpdate EventType = "message_update"
 	// EventMessageEnd fires when the assistant message is complete.
 	EventMessageEnd EventType = "message_end"
+	// EventToolCallStart fires when the LLM begins generating tool call arguments.
+	// The tool name is known but arguments are still streaming.
+	EventToolCallStart EventType = "tool_call_start"
+	// EventToolCallDelta fires for each streamed fragment of tool call arguments.
+	EventToolCallDelta EventType = "tool_call_delta"
+	// EventToolCallEnd fires when tool argument streaming is complete, before
+	// the tool call is parsed and execution begins.
+	EventToolCallEnd EventType = "tool_call_end"
 	// EventToolCall fires when a tool call has been parsed and is about to execute.
 	EventToolCall EventType = "tool_call"
 	// EventToolExecutionStart fires when a tool begins executing.
@@ -50,6 +58,31 @@ const (
 	// EventSteerConsumed fires when one or more steering messages have been
 	// injected into the agent turn via PrepareStep.
 	EventSteerConsumed EventType = "steer_consumed"
+	// EventStepStart fires when a new LLM call begins within a turn.
+	EventStepStart EventType = "step_start"
+	// EventStepFinish fires when a step completes, providing full step context
+	// including whether tool calls were made, the finish reason, and usage stats.
+	EventStepFinish EventType = "step_finish"
+	// EventTextStart fires when the LLM begins generating text content.
+	EventTextStart EventType = "text_start"
+	// EventTextEnd fires when the LLM finishes generating text content.
+	EventTextEnd EventType = "text_end"
+	// EventReasoningStart fires when the LLM begins reasoning/thinking.
+	EventReasoningStart EventType = "reasoning_start"
+	// EventWarnings fires when the LLM provider returns warnings.
+	EventWarnings EventType = "warnings"
+	// EventSource fires when the LLM references a source (e.g. from web search).
+	EventSource EventType = "source"
+	// EventStreamFinish fires when a per-step LLM stream completes with
+	// usage stats and a finish reason.
+	EventStreamFinish EventType = "stream_finish"
+	// EventError fires when an agent-level error occurs during streaming.
+	// This is distinct from TurnEndEvent.Error — it fires at the point of
+	// failure, before the turn ends.
+	EventError EventType = "error"
+	// EventRetry fires when the LLM provider request is retried after a
+	// transient error.
+	EventRetry EventType = "retry"
 )
 
 // ---------------------------------------------------------------------------
@@ -110,6 +143,38 @@ func parseToolArgs(toolArgs string) map[string]any {
 	return nil
 }
 
+// ---------------------------------------------------------------------------
+// Finish reason constants
+// ---------------------------------------------------------------------------
+
+// Finish reasons reported by the LLM provider on a completed turn. These
+// mirror fantasy.FinishReason string values so comparisons against
+// TurnEndEvent.StopReason / TurnResult.StopReason are stable across
+// providers.
+const (
+	// FinishReasonStop: the model produced a natural stop (e.g. stop sequence
+	// or end-of-turn signal).
+	FinishReasonStop = "stop"
+	// FinishReasonLength: the model hit the configured max_output_tokens
+	// budget. The response is truncated. Surface this to the user and
+	// consider raising --max-tokens / KIT_MAX_TOKENS / modelSettings[...]
+	// .maxTokens.
+	FinishReasonLength = "length"
+	// FinishReasonToolCalls: the model stopped to emit tool calls (normal
+	// mid-turn state during agentic loops).
+	FinishReasonToolCalls = "tool-calls"
+	// FinishReasonContentFilter: the provider's safety filter stopped
+	// generation.
+	FinishReasonContentFilter = "content-filter"
+	// FinishReasonError: the model stopped because of an error.
+	FinishReasonError = "error"
+	// FinishReasonOther: provider-specific reason that doesn't map to any of
+	// the above.
+	FinishReasonOther = "other"
+	// FinishReasonUnknown: the provider didn't report a finish reason.
+	FinishReasonUnknown = "unknown"
+)
+
 // ---------------------------------------------------------------------------
 // Concrete event structs
 // ---------------------------------------------------------------------------
@@ -124,9 +189,13 @@ func (e TurnStartEvent) EventType() EventType { return EventTurnStart }
 
 // TurnEndEvent fires after the agent finishes processing.
 type TurnEndEvent struct {
-	Response   string
-	Error      error
-	StopReason string // "end_turn", "max_tokens", "tool_use", "error", etc.
+	Response string
+	Error    error
+	// StopReason is the LLM provider's finish reason for the final step of
+	// the turn. Compare against the FinishReason* constants — in particular,
+	// FinishReasonLength indicates the response was truncated because the
+	// agent hit its max_output_tokens budget.
+	StopReason string
 }
 
 // EventType implements Event.
@@ -180,6 +249,40 @@ type MessageEndEvent struct {
 // EventType implements Event.
 func (e MessageEndEvent) EventType() EventType { return EventMessageEnd }
 
+// ToolCallStartEvent fires when the LLM begins generating tool call arguments.
+// The tool name is known at this point but the full arguments are still being
+// streamed. UIs can use this to show a "running" indicator immediately instead
+// of waiting for the full argument JSON to finish streaming.
+type ToolCallStartEvent struct {
+	ToolCallID string // Stable ID for correlating tool lifecycle events
+	ToolName   string
+	ToolKind   string // Tool classification: "execute", "edit", "read", "search", "agent"
+}
+
+// EventType implements Event.
+func (e ToolCallStartEvent) EventType() EventType { return EventToolCallStart }
+
+// ToolCallDeltaEvent fires for each streamed fragment of tool call arguments.
+// Useful for live-previewing artifact content as it's generated, or showing a
+// progress indicator with byte count.
+type ToolCallDeltaEvent struct {
+	ToolCallID string // Stable ID for correlating tool lifecycle events
+	Delta      string // JSON fragment of tool arguments
+}
+
+// EventType implements Event.
+func (e ToolCallDeltaEvent) EventType() EventType { return EventToolCallDelta }
+
+// ToolCallEndEvent fires when tool argument streaming is complete, before
+// the tool call is parsed and execution begins. UIs can use this to
+// transition from an "generating args" state to an "executing" state.
+type ToolCallEndEvent struct {
+	ToolCallID string // Stable ID for correlating tool lifecycle events
+}
+
+// EventType implements Event.
+func (e ToolCallEndEvent) EventType() EventType { return EventToolCallEnd }
+
 // ToolCallEvent fires when a tool call has been parsed.
 type ToolCallEvent struct {
 	ToolCallID string // Stable ID for correlating tool lifecycle events
@@ -301,6 +404,100 @@ type SteerConsumedEvent struct {
 // EventType implements Event.
 func (e SteerConsumedEvent) EventType() EventType { return EventSteerConsumed }
 
+// StepStartEvent fires when a new LLM call begins within a multi-step agent turn.
+type StepStartEvent struct {
+	StepNumber int
+}
+
+// EventType implements Event.
+func (e StepStartEvent) EventType() EventType { return EventStepStart }
+
+// StepFinishEvent fires when a step completes, providing full step context.
+// This is a unified event that carries the same data as the existing
+// ToolCallContentEvent and StepUsageEvent, plus additional step metadata.
+type StepFinishEvent struct {
+	StepNumber   int
+	HasToolCalls bool
+	FinishReason string
+	Usage        LLMUsage
+}
+
+// EventType implements Event.
+func (e StepFinishEvent) EventType() EventType { return EventStepFinish }
+
+// TextStartEvent fires when the LLM begins generating text content.
+// Paired with MessageUpdateEvent (deltas) and TextEndEvent.
+type TextStartEvent struct {
+	ID string
+}
+
+// EventType implements Event.
+func (e TextStartEvent) EventType() EventType { return EventTextStart }
+
+// TextEndEvent fires when the LLM finishes generating text content.
+type TextEndEvent struct {
+	ID string
+}
+
+// EventType implements Event.
+func (e TextEndEvent) EventType() EventType { return EventTextEnd }
+
+// ReasoningStartEvent fires when the LLM begins reasoning/thinking.
+// Paired with ReasoningDeltaEvent (deltas) and ReasoningCompleteEvent.
+type ReasoningStartEvent struct {
+	ID string
+}
+
+// EventType implements Event.
+func (e ReasoningStartEvent) EventType() EventType { return EventReasoningStart }
+
+// WarningsEvent fires when the LLM provider returns warnings about the request.
+type WarningsEvent struct {
+	Warnings []string
+}
+
+// EventType implements Event.
+func (e WarningsEvent) EventType() EventType { return EventWarnings }
+
+// SourceEvent fires when the LLM references a source (e.g. from web search tools).
+type SourceEvent struct {
+	SourceType string
+	ID         string
+	URL        string
+	Title      string
+}
+
+// EventType implements Event.
+func (e SourceEvent) EventType() EventType { return EventSource }
+
+// StreamFinishEvent fires when a per-step LLM stream completes.
+// Provides per-stream usage stats and finish reason.
+type StreamFinishEvent struct {
+	Usage        LLMUsage
+	FinishReason string
+}
+
+// EventType implements Event.
+func (e StreamFinishEvent) EventType() EventType { return EventStreamFinish }
+
+// ErrorEvent fires when an agent-level error occurs during streaming.
+// This is distinct from TurnEndEvent.Error — it fires at the point of failure.
+type ErrorEvent struct {
+	Error error
+}
+
+// EventType implements Event.
+func (e ErrorEvent) EventType() EventType { return EventError }
+
+// RetryEvent fires when the LLM provider request is retried after a transient error.
+type RetryEvent struct {
+	Attempt int
+	Error   error
+}
+
+// EventType implements Event.
+func (e RetryEvent) EventType() EventType { return EventRetry }
+
 // PasswordPromptEvent fires when a sudo command needs a password.
 // The TUI should display a password prompt and send the result back via ResponseCh.
 type PasswordPromptEvent struct {
@@ -384,6 +581,39 @@ func (m *Kit) OnToolCall(handler func(ToolCallEvent)) func() {
 	})
 }
 
+// OnToolCallStart registers a handler that fires only for ToolCallStartEvent.
+// This fires when the LLM begins generating tool call arguments — before the
+// full argument JSON is available. Returns an unsubscribe function.
+func (m *Kit) OnToolCallStart(handler func(ToolCallStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tcs, ok := e.(ToolCallStartEvent); ok {
+			handler(tcs)
+		}
+	})
+}
+
+// OnToolCallDelta registers a handler that fires only for ToolCallDeltaEvent.
+// Each delta contains a JSON fragment of tool call arguments as they stream in.
+// Returns an unsubscribe function.
+func (m *Kit) OnToolCallDelta(handler func(ToolCallDeltaEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tcd, ok := e.(ToolCallDeltaEvent); ok {
+			handler(tcd)
+		}
+	})
+}
+
+// OnToolCallEnd registers a handler that fires only for ToolCallEndEvent.
+// This fires when tool argument streaming is complete, before the tool call
+// is parsed and execution begins. Returns an unsubscribe function.
+func (m *Kit) OnToolCallEnd(handler func(ToolCallEndEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tce, ok := e.(ToolCallEndEvent); ok {
+			handler(tce)
+		}
+	})
+}
+
 // OnToolResult registers a handler that fires only for ToolResultEvent.
 // Returns an unsubscribe function.
 func (m *Kit) OnToolResult(handler func(ToolResultEvent)) func() {
@@ -406,7 +636,16 @@ func (m *Kit) OnToolOutput(handler func(ToolOutputEvent)) func() {
 
 // OnStreaming registers a handler that fires only for MessageUpdateEvent
 // (streaming text chunks). Returns an unsubscribe function.
+//
+// Deprecated: Use OnMessageUpdate instead. OnStreaming will be removed in a
+// future release.
 func (m *Kit) OnStreaming(handler func(MessageUpdateEvent)) func() {
+	return m.OnMessageUpdate(handler)
+}
+
+// OnMessageUpdate registers a handler that fires only for MessageUpdateEvent
+// (streaming text chunks). Returns an unsubscribe function.
+func (m *Kit) OnMessageUpdate(handler func(MessageUpdateEvent)) func() {
 	return m.Subscribe(func(e Event) {
 		if mu, ok := e.(MessageUpdateEvent); ok {
 			handler(mu)
@@ -444,6 +683,214 @@ func (m *Kit) OnTurnEnd(handler func(TurnEndEvent)) func() {
 	})
 }
 
+// ---------------------------------------------------------------------------
+// Typed subscribers for previously unsubscribed event types
+// ---------------------------------------------------------------------------
+
+// OnMessageStart registers a handler that fires only for MessageStartEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnMessageStart(handler func(MessageStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if ms, ok := e.(MessageStartEvent); ok {
+			handler(ms)
+		}
+	})
+}
+
+// OnMessageEnd registers a handler that fires only for MessageEndEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnMessageEnd(handler func(MessageEndEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if me, ok := e.(MessageEndEvent); ok {
+			handler(me)
+		}
+	})
+}
+
+// OnReasoningDelta registers a handler that fires only for ReasoningDeltaEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnReasoningDelta(handler func(ReasoningDeltaEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if rd, ok := e.(ReasoningDeltaEvent); ok {
+			handler(rd)
+		}
+	})
+}
+
+// OnReasoningComplete registers a handler that fires only for ReasoningCompleteEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnReasoningComplete(handler func(ReasoningCompleteEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if rc, ok := e.(ReasoningCompleteEvent); ok {
+			handler(rc)
+		}
+	})
+}
+
+// OnToolExecutionStart registers a handler that fires only for ToolExecutionStartEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnToolExecutionStart(handler func(ToolExecutionStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tes, ok := e.(ToolExecutionStartEvent); ok {
+			handler(tes)
+		}
+	})
+}
+
+// OnToolExecutionEnd registers a handler that fires only for ToolExecutionEndEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnToolExecutionEnd(handler func(ToolExecutionEndEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tee, ok := e.(ToolExecutionEndEvent); ok {
+			handler(tee)
+		}
+	})
+}
+
+// OnToolCallContent registers a handler that fires only for ToolCallContentEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnToolCallContent(handler func(ToolCallContentEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if tcc, ok := e.(ToolCallContentEvent); ok {
+			handler(tcc)
+		}
+	})
+}
+
+// OnStepUsage registers a handler that fires only for StepUsageEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnStepUsage(handler func(StepUsageEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if su, ok := e.(StepUsageEvent); ok {
+			handler(su)
+		}
+	})
+}
+
+// OnCompaction registers a handler that fires only for CompactionEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnCompaction(handler func(CompactionEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if ce, ok := e.(CompactionEvent); ok {
+			handler(ce)
+		}
+	})
+}
+
+// OnSteerConsumed registers a handler that fires only for SteerConsumedEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnSteerConsumed(handler func(SteerConsumedEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if sc, ok := e.(SteerConsumedEvent); ok {
+			handler(sc)
+		}
+	})
+}
+
+// ---------------------------------------------------------------------------
+// Typed subscribers for new event types
+// ---------------------------------------------------------------------------
+
+// OnStepStart registers a handler that fires only for StepStartEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnStepStart(handler func(StepStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if ss, ok := e.(StepStartEvent); ok {
+			handler(ss)
+		}
+	})
+}
+
+// OnStepFinish registers a handler that fires only for StepFinishEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnStepFinish(handler func(StepFinishEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if sf, ok := e.(StepFinishEvent); ok {
+			handler(sf)
+		}
+	})
+}
+
+// OnTextStart registers a handler that fires only for TextStartEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnTextStart(handler func(TextStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if ts, ok := e.(TextStartEvent); ok {
+			handler(ts)
+		}
+	})
+}
+
+// OnTextEnd registers a handler that fires only for TextEndEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnTextEnd(handler func(TextEndEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if te, ok := e.(TextEndEvent); ok {
+			handler(te)
+		}
+	})
+}
+
+// OnReasoningStart registers a handler that fires only for ReasoningStartEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnReasoningStart(handler func(ReasoningStartEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if rs, ok := e.(ReasoningStartEvent); ok {
+			handler(rs)
+		}
+	})
+}
+
+// OnWarnings registers a handler that fires only for WarningsEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnWarnings(handler func(WarningsEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if w, ok := e.(WarningsEvent); ok {
+			handler(w)
+		}
+	})
+}
+
+// OnSource registers a handler that fires only for SourceEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnSource(handler func(SourceEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if s, ok := e.(SourceEvent); ok {
+			handler(s)
+		}
+	})
+}
+
+// OnStreamFinish registers a handler that fires only for StreamFinishEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnStreamFinish(handler func(StreamFinishEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if sf, ok := e.(StreamFinishEvent); ok {
+			handler(sf)
+		}
+	})
+}
+
+// OnError registers a handler that fires only for ErrorEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnError(handler func(ErrorEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if ee, ok := e.(ErrorEvent); ok {
+			handler(ee)
+		}
+	})
+}
+
+// OnRetry registers a handler that fires only for RetryEvent.
+// Returns an unsubscribe function.
+func (m *Kit) OnRetry(handler func(RetryEvent)) func() {
+	return m.Subscribe(func(e Event) {
+		if r, ok := e.(RetryEvent); ok {
+			handler(r)
+		}
+	})
+}
+
 // ---------------------------------------------------------------------------
 // Subagent event subscriptions
 // ---------------------------------------------------------------------------

pkg/kit/events_test.go

@@ -1,6 +1,7 @@
 package kit
 
 import (
+	"fmt"
 	"sync"
 	"sync/atomic"
 	"testing"
@@ -190,6 +191,74 @@ func TestEventTypes(t *testing.T) {
 	}
 }
 
+// TestNewEventTypes verifies that each new event struct returns the correct EventType.
+func TestNewEventTypes(t *testing.T) {
+	tests := []struct {
+		event    Event
+		expected EventType
+	}{
+		{StepStartEvent{StepNumber: 0}, EventStepStart},
+		{StepFinishEvent{StepNumber: 1, HasToolCalls: true}, EventStepFinish},
+		{TextStartEvent{ID: "text-1"}, EventTextStart},
+		{TextEndEvent{ID: "text-1"}, EventTextEnd},
+		{ReasoningStartEvent{ID: "reason-1"}, EventReasoningStart},
+		{WarningsEvent{Warnings: []string{"test"}}, EventWarnings},
+		{SourceEvent{URL: "https://example.com", Title: "Example"}, EventSource},
+		{StreamFinishEvent{FinishReason: "stop"}, EventStreamFinish},
+		{ErrorEvent{Error: fmt.Errorf("test error")}, EventError},
+		{RetryEvent{Attempt: 1, Error: fmt.Errorf("retry error")}, EventRetry},
+		{ToolCallStartEvent{}, EventToolCallStart},
+		{ToolCallDeltaEvent{}, EventToolCallDelta},
+		{ToolCallEndEvent{}, EventToolCallEnd},
+		{PasswordPromptEvent{}, EventPasswordPrompt},
+	}
+
+	for _, tt := range tests {
+		if got := tt.event.EventType(); got != tt.expected {
+			t.Errorf("%T.EventType() = %q, want %q", tt.event, got, tt.expected)
+		}
+	}
+}
+
+// TestNewEventEmission verifies that new event types are properly emitted and received.
+func TestNewEventEmission(t *testing.T) {
+	bus := newEventBus()
+	var received []Event
+
+	bus.subscribe(func(e Event) {
+		received = append(received, e)
+	})
+
+	bus.emit(StepStartEvent{StepNumber: 0})
+	bus.emit(TextStartEvent{ID: "text-1"})
+	bus.emit(TextEndEvent{ID: "text-1"})
+	bus.emit(ReasoningStartEvent{ID: "reason-1"})
+	bus.emit(WarningsEvent{Warnings: []string{"low confidence"}})
+	bus.emit(SourceEvent{URL: "https://example.com", Title: "Example"})
+	bus.emit(StreamFinishEvent{FinishReason: "stop"})
+	bus.emit(StepFinishEvent{StepNumber: 0, HasToolCalls: false, FinishReason: "stop"})
+	bus.emit(ErrorEvent{Error: fmt.Errorf("test error")})
+	bus.emit(RetryEvent{Attempt: 1, Error: fmt.Errorf("retry")})
+
+	if len(received) != 10 {
+		t.Fatalf("expected 10 events, got %d", len(received))
+	}
+
+	// Verify specific event fields
+	if ss, ok := received[0].(StepStartEvent); !ok || ss.StepNumber != 0 {
+		t.Errorf("event 0: expected StepStartEvent{StepNumber:0}, got %T %+v", received[0], received[0])
+	}
+	if ts, ok := received[1].(TextStartEvent); !ok || ts.ID != "text-1" {
+		t.Errorf("event 1: expected TextStartEvent{ID:text-1}, got %T %+v", received[1], received[1])
+	}
+	if w, ok := received[4].(WarningsEvent); !ok || len(w.Warnings) != 1 || w.Warnings[0] != "low confidence" {
+		t.Errorf("event 4: expected WarningsEvent with 1 warning, got %T %+v", received[4], received[4])
+	}
+	if sf, ok := received[7].(StepFinishEvent); !ok || sf.StepNumber != 0 || sf.HasToolCalls {
+		t.Errorf("event 7: expected StepFinishEvent{StepNumber:0, HasToolCalls:false}, got %T %+v", received[7], received[7])
+	}
+}
+
 // TestEventBusListenerCanUnsubscribeInCallback verifies that a listener can
 // safely call its own unsubscribe function from within the callback.
 func TestEventBusListenerCanUnsubscribeInCallback(t *testing.T) {

pkg/kit/extensions_bridge.go

@@ -100,6 +100,38 @@ func (m *Kit) bridgeExtensions(runner *extensions.Runner) {
 		})
 	}
 
+	// Tool call input streaming events — fire as the LLM generates tool arguments.
+	if runner.HasHandlers(extensions.ToolCallInputStart) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(ToolCallStartEvent); ok {
+				_, _ = runner.Emit(extensions.ToolCallInputStartEvent{
+					ToolCallID: ev.ToolCallID,
+					ToolName:   ev.ToolName,
+					ToolKind:   ev.ToolKind,
+				})
+			}
+		})
+	}
+	if runner.HasHandlers(extensions.ToolCallInputDelta) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(ToolCallDeltaEvent); ok {
+				_, _ = runner.Emit(extensions.ToolCallInputDeltaEvent{
+					ToolCallID: ev.ToolCallID,
+					Delta:      ev.Delta,
+				})
+			}
+		})
+	}
+	if runner.HasHandlers(extensions.ToolCallInputEnd) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(ToolCallEndEvent); ok {
+				_, _ = runner.Emit(extensions.ToolCallInputEndEvent{
+					ToolCallID: ev.ToolCallID,
+				})
+			}
+		})
+	}
+
 	if runner.HasHandlers(extensions.AgentEnd) {
 		m.Subscribe(func(e Event) {
 			if ev, ok := e.(TurnEndEvent); ok {
@@ -324,4 +356,134 @@ func (m *Kit) bridgeExtensions(runner *extensions.Runner) {
 			return nil
 		})
 	}
+
+	// --- Step lifecycle observation events ---
+
+	if runner.HasHandlers(extensions.StepStart) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(StepStartEvent); ok {
+				_, _ = runner.Emit(extensions.StepStartEvent{StepNumber: ev.StepNumber})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.StepFinish) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(StepFinishEvent); ok {
+				_, _ = runner.Emit(extensions.StepFinishEvent{
+					StepNumber:       ev.StepNumber,
+					HasToolCalls:     ev.HasToolCalls,
+					FinishReason:     ev.FinishReason,
+					InputTokens:      ev.Usage.InputTokens,
+					OutputTokens:     ev.Usage.OutputTokens,
+					CacheReadTokens:  ev.Usage.CacheReadTokens,
+					CacheWriteTokens: ev.Usage.CacheCreationTokens,
+				})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.ReasoningStart) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(ReasoningStartEvent); ok {
+				_, _ = runner.Emit(extensions.ReasoningStartEvent{ID: ev.ID})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.Warnings) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(WarningsEvent); ok {
+				_, _ = runner.Emit(extensions.WarningsEvent{Warnings: ev.Warnings})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.Source) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(SourceEvent); ok {
+				_, _ = runner.Emit(extensions.SourceEvent{
+					SourceType: ev.SourceType,
+					ID:         ev.ID,
+					URL:        ev.URL,
+					Title:      ev.Title,
+				})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.Error) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(ErrorEvent); ok {
+				_, _ = runner.Emit(extensions.ErrorEvent{Error: ev.Error.Error()})
+			}
+		})
+	}
+
+	if runner.HasHandlers(extensions.Retry) {
+		m.Subscribe(func(e Event) {
+			if ev, ok := e.(RetryEvent); ok {
+				_, _ = runner.Emit(extensions.RetryEvent{
+					Attempt: ev.Attempt,
+					Error:   ev.Error.Error(),
+				})
+			}
+		})
+	}
+
+	// --- PrepareStep hook ---
+	// Extension PrepareStep → SDK PrepareStep hook.
+	// Same pattern as ContextPrepare: convert LLMMessage ↔ ContextMessage.
+	if runner.HasHandlers(extensions.PrepareStep) {
+		m.OnPrepareStep(HookPriorityNormal, func(h PrepareStepHook) *PrepareStepResult {
+			// Convert LLM message slice to extension ContextMessage slice.
+			extMsgs := make([]extensions.ContextMessage, len(h.Messages))
+			for i, msg := range h.Messages {
+				var sb strings.Builder
+				for _, part := range msg.Content {
+					if tp, ok := part.(LLMTextPart); ok {
+						sb.WriteString(tp.Text)
+					}
+				}
+				extMsgs[i] = extensions.ContextMessage{
+					Index:   i,
+					Role:    string(msg.Role),
+					Content: sb.String(),
+				}
+			}
+
+			result, _ := runner.Emit(extensions.PrepareStepEvent{
+				StepNumber: h.StepNumber,
+				Messages:   extMsgs,
+			})
+			r, ok := result.(extensions.PrepareStepResult)
+			if !ok || r.Messages == nil {
+				return nil
+			}
+
+			// Rebuild LLM message slice from extension result.
+			rebuilt := make([]LLMMessage, 0, len(r.Messages))
+			for _, cm := range r.Messages {
+				if cm.Index >= 0 && cm.Index < len(h.Messages) {
+					rebuilt = append(rebuilt, h.Messages[cm.Index])
+				} else {
+					role := LLMRoleUser
+					switch cm.Role {
+					case "assistant":
+						role = LLMRoleAssistant
+					case "system":
+						role = LLMRoleSystem
+					case "tool":
+						role = LLMRoleTool
+					}
+					rebuilt = append(rebuilt, LLMMessage{
+						Role:    role,
+						Content: []LLMMessagePart{LLMTextPart{Text: cm.Content}},
+					})
+				}
+			}
+
+			return &PrepareStepResult{Messages: rebuilt}
+		})
+	}
 }

pkg/kit/hooks.go

@@ -5,8 +5,6 @@ import (
 	"fmt"
 	"sort"
 	"sync"
-
-	"charm.land/fantasy"
 )
 
 // ---------------------------------------------------------------------------
@@ -121,6 +119,32 @@ type BeforeCompactResult struct {
 	Summary string
 }
 
+// PrepareStepHook is the input for hooks that fire between steps within a
+// multi-step agent turn, with full message replacement capability. This is
+// the most powerful interception point — it fires after the existing steering
+// logic (if any) and before the messages are sent to the LLM.
+//
+// Use cases:
+//   - Transforming tool results (e.g. converting image tool results to FilePart
+//     user messages for vision models that don't support media in tool results)
+//   - Dynamic tool filtering per step
+//   - Mid-turn context injection beyond simple steering
+//   - Custom stop conditions that inspect message history
+type PrepareStepHook struct {
+	// StepNumber is the zero-based step index within the current turn.
+	StepNumber int
+	// Messages is the current context window that will be sent to the LLM.
+	// This includes any steering messages already injected in this step.
+	Messages []LLMMessage
+}
+
+// PrepareStepResult can replace the context window between steps.
+type PrepareStepResult struct {
+	// Messages replaces the entire context window for this step. If nil,
+	// the original messages (including any steering) are used unchanged.
+	Messages []LLMMessage
+}
+
 // ---------------------------------------------------------------------------
 // Generic hook registry with priority ordering
 // ---------------------------------------------------------------------------
@@ -248,6 +272,19 @@ func (m *Kit) OnBeforeCompact(p HookPriority, h func(BeforeCompactHook) *BeforeC
 	return m.beforeCompact.register(p, h)
 }
 
+// OnPrepareStep registers a hook that fires between steps within a multi-step
+// agent turn, after steering messages are injected and before the messages are
+// sent to the LLM. Return a non-nil PrepareStepResult with Messages to replace
+// the entire context window for this step. Hooks execute in priority order;
+// the first non-nil result wins. Returns an unregister function.
+//
+// This is the most powerful interception point in the agent lifecycle. It
+// enables patterns like transforming tool results, dynamic tool filtering,
+// and mid-turn context injection.
+func (m *Kit) OnPrepareStep(p HookPriority, h func(PrepareStepHook) *PrepareStepResult) func() {
+	return m.prepareStep.register(p, h)
+}
+
 // ---------------------------------------------------------------------------
 // Tool wrapping via hooks
 // ---------------------------------------------------------------------------
@@ -256,16 +293,16 @@ func (m *Kit) OnBeforeCompact(p HookPriority, h func(BeforeCompactHook) *BeforeC
 // AfterToolResult hooks around each execution. The registries are referenced
 // by pointer so hooks added after agent creation are still invoked.
 type hookedTool struct {
-	inner           fantasy.AgentTool
+	inner           Tool
 	beforeToolCall  *hookRegistry[BeforeToolCallHook, BeforeToolCallResult]
 	afterToolResult *hookRegistry[AfterToolResultHook, AfterToolResultResult]
 }
 
-func (h *hookedTool) Info() fantasy.ToolInfo                       { return h.inner.Info() }
-func (h *hookedTool) ProviderOptions() fantasy.ProviderOptions     { return h.inner.ProviderOptions() }
-func (h *hookedTool) SetProviderOptions(o fantasy.ProviderOptions) { h.inner.SetProviderOptions(o) }
+func (h *hookedTool) Info() LLMToolInfo                       { return h.inner.Info() }
+func (h *hookedTool) ProviderOptions() LLMProviderOptions     { return h.inner.ProviderOptions() }
+func (h *hookedTool) SetProviderOptions(o LLMProviderOptions) { h.inner.SetProviderOptions(o) }
 
-func (h *hookedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
+func (h *hookedTool) Run(ctx context.Context, call LLMToolCall) (LLMToolResponse, error) {
 	toolName := h.inner.Info().Name
 
 	// 1. BeforeToolCall — can block execution.
@@ -279,7 +316,7 @@ func (h *hookedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.To
 			if reason == "" {
 				reason = "blocked by hook"
 			}
-			return fantasy.NewTextErrorResponse(fmt.Sprintf("Error: %s", reason)),
+			return newLLMTextErrorResponse(fmt.Sprintf("Error: %s", reason)),
 				fmt.Errorf("tool blocked by hook: %s", reason)
 		}
 	}
@@ -314,9 +351,9 @@ func (h *hookedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.To
 func hookToolWrapper(
 	beforeToolCall *hookRegistry[BeforeToolCallHook, BeforeToolCallResult],
 	afterToolResult *hookRegistry[AfterToolResultHook, AfterToolResultResult],
-) func([]fantasy.AgentTool) []fantasy.AgentTool {
-	return func(tools []fantasy.AgentTool) []fantasy.AgentTool {
-		wrapped := make([]fantasy.AgentTool, len(tools))
+) func([]Tool) []Tool {
+	return func(tools []Tool) []Tool {
+		wrapped := make([]Tool, len(tools))
 		for i, tool := range tools {
 			wrapped[i] = &hookedTool{
 				inner:           tool,

pkg/kit/hooks_test.go

@@ -5,8 +5,6 @@ import (
 	"fmt"
 	"sync"
 	"testing"
-
-	"charm.land/fantasy"
 )
 
 // ---------------------------------------------------------------------------
@@ -177,20 +175,20 @@ func TestHookRegistry_ConcurrentAccess(t *testing.T) {
 // mockAgentTool implements the AgentTool interface for testing.
 type mockAgentTool struct {
 	name  string
-	runFn func(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error)
-	popts fantasy.ProviderOptions
+	runFn func(ctx context.Context, call LLMToolCall) (LLMToolResponse, error)
+	popts LLMProviderOptions
 }
 
-func (m *mockAgentTool) Info() fantasy.ToolInfo {
-	return fantasy.ToolInfo{Name: m.name, Description: "mock tool"}
+func (m *mockAgentTool) Info() LLMToolInfo {
+	return LLMToolInfo{Name: m.name, Description: "mock tool"}
 }
-func (m *mockAgentTool) ProviderOptions() fantasy.ProviderOptions     { return m.popts }
-func (m *mockAgentTool) SetProviderOptions(o fantasy.ProviderOptions) { m.popts = o }
-func (m *mockAgentTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
+func (m *mockAgentTool) ProviderOptions() LLMProviderOptions     { return m.popts }
+func (m *mockAgentTool) SetProviderOptions(o LLMProviderOptions) { m.popts = o }
+func (m *mockAgentTool) Run(ctx context.Context, call LLMToolCall) (LLMToolResponse, error) {
 	if m.runFn != nil {
 		return m.runFn(ctx, call)
 	}
-	return fantasy.NewTextResponse("default output"), nil
+	return newLLMTextResponse("default output"), nil
 }
 
 // newEmptyHookedTool creates a hookedTool with empty hook registries and the given mock tool.
@@ -203,14 +201,14 @@ func newEmptyHookedTool(mock *mockAgentTool) *hookedTool {
 func TestHookedTool_Passthrough(t *testing.T) {
 	mock := &mockAgentTool{
 		name: "test_tool",
-		runFn: func(_ context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {
-			return fantasy.NewTextResponse("hello world"), nil
+		runFn: func(_ context.Context, _ LLMToolCall) (LLMToolResponse, error) {
+			return newLLMTextResponse("hello world"), nil
 		},
 	}
 
 	ht := newEmptyHookedTool(mock)
 
-	resp, err := ht.Run(context.Background(), fantasy.ToolCall{Input: "{}"})
+	resp, err := ht.Run(context.Background(), LLMToolCall{Input: "{}"})
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -226,9 +224,9 @@ func TestHookedTool_BeforeToolCallBlock(t *testing.T) {
 	toolRan := false
 	mock := &mockAgentTool{
 		name: "dangerous_tool",
-		runFn: func(_ context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {
+		runFn: func(_ context.Context, _ LLMToolCall) (LLMToolResponse, error) {
 			toolRan = true
-			return fantasy.NewTextResponse("should not run"), nil
+			return newLLMTextResponse("should not run"), nil
 		},
 	}
 
@@ -241,7 +239,7 @@ func TestHookedTool_BeforeToolCallBlock(t *testing.T) {
 
 	ht := &hookedTool{inner: mock, beforeToolCall: before, afterToolResult: after}
 
-	resp, err := ht.Run(context.Background(), fantasy.ToolCall{Input: "{}"})
+	resp, err := ht.Run(context.Background(), LLMToolCall{Input: "{}"})
 	if err == nil {
 		t.Fatal("expected error from blocked tool")
 	}
@@ -263,7 +261,7 @@ func TestHookedTool_BeforeToolCallBlockDefaultReason(t *testing.T) {
 	})
 
 	ht := &hookedTool{inner: mock, beforeToolCall: before, afterToolResult: after}
-	resp, _ := ht.Run(context.Background(), fantasy.ToolCall{})
+	resp, _ := ht.Run(context.Background(), LLMToolCall{})
 	if resp.Content != "Error: blocked by hook" {
 		t.Errorf("expected default block reason, got %q", resp.Content)
 	}
@@ -275,8 +273,8 @@ func TestHookedTool_AfterToolResultModify(t *testing.T) {
 
 	mock := &mockAgentTool{
 		name: "tool",
-		runFn: func(_ context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {
-			return fantasy.NewTextResponse("secret data"), nil
+		runFn: func(_ context.Context, _ LLMToolCall) (LLMToolResponse, error) {
+			return newLLMTextResponse("secret data"), nil
 		},
 	}
 
@@ -286,7 +284,7 @@ func TestHookedTool_AfterToolResultModify(t *testing.T) {
 	})
 
 	ht := &hookedTool{inner: mock, beforeToolCall: before, afterToolResult: after}
-	resp, err := ht.Run(context.Background(), fantasy.ToolCall{Input: "{}"})
+	resp, err := ht.Run(context.Background(), LLMToolCall{Input: "{}"})
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -301,8 +299,8 @@ func TestHookedTool_AfterToolResultModifyIsError(t *testing.T) {
 
 	mock := &mockAgentTool{
 		name: "tool",
-		runFn: func(_ context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {
-			return fantasy.NewTextResponse("ok"), nil
+		runFn: func(_ context.Context, _ LLMToolCall) (LLMToolResponse, error) {
+			return newLLMTextResponse("ok"), nil
 		},
 	}
 
@@ -312,7 +310,7 @@ func TestHookedTool_AfterToolResultModifyIsError(t *testing.T) {
 	})
 
 	ht := &hookedTool{inner: mock, beforeToolCall: before, afterToolResult: after}
-	resp, err := ht.Run(context.Background(), fantasy.ToolCall{})
+	resp, err := ht.Run(context.Background(), LLMToolCall{})
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -327,8 +325,8 @@ func TestHookedTool_HookReceivesToolInfo(t *testing.T) {
 
 	mock := &mockAgentTool{
 		name: "my_tool",
-		runFn: func(_ context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {
-			return fantasy.NewTextResponse("result"), nil
+		runFn: func(_ context.Context, _ LLMToolCall) (LLMToolResponse, error) {
+			return newLLMTextResponse("result"), nil
 		},
 	}
 
@@ -345,7 +343,7 @@ func TestHookedTool_HookReceivesToolInfo(t *testing.T) {
 	})
 
 	ht := &hookedTool{inner: mock, beforeToolCall: before, afterToolResult: after}
-	_, _ = ht.Run(context.Background(), fantasy.ToolCall{Input: `{"key":"value"}`})
+	_, _ = ht.Run(context.Background(), LLMToolCall{Input: `{"key":"value"}`})
 
 	if capturedBefore.ToolName != "my_tool" {
 		t.Errorf("BeforeToolCall: expected tool name 'my_tool', got %q", capturedBefore.ToolName)
@@ -380,7 +378,7 @@ func TestHookToolWrapper(t *testing.T) {
 
 	wrapper := hookToolWrapper(before, after)
 
-	tools := []fantasy.AgentTool{
+	tools := []Tool{
 		&mockAgentTool{name: "tool_a"},
 		&mockAgentTool{name: "tool_b"},
 	}
@@ -407,7 +405,7 @@ func TestHookToolWrapper(t *testing.T) {
 		return &BeforeToolCallResult{Block: true, Reason: "late hook"}
 	})
 
-	_, err := wrapped[0].Run(context.Background(), fantasy.ToolCall{})
+	_, err := wrapped[0].Run(context.Background(), LLMToolCall{})
 	if err == nil {
 		t.Error("expected error from late-registered blocking hook")
 	}
@@ -538,3 +536,75 @@ func TestKit_HookMethodsExist(t *testing.T) {
 	u3()
 	u4()
 }
+
+// TestPrepareStepHookRegistry verifies registration and execution of PrepareStep hooks.
+func TestPrepareStepHookRegistry(t *testing.T) {
+	hr := newHookRegistry[PrepareStepHook, PrepareStepResult]()
+
+	// Register a hook that appends a message.
+	hr.register(HookPriorityNormal, func(h PrepareStepHook) *PrepareStepResult {
+		if h.StepNumber == 0 {
+			// On step 0, prepend a system message.
+			newMsgs := make([]LLMMessage, 0, len(h.Messages)+1)
+			newMsgs = append(newMsgs, NewLLMSystemMessage("injected"))
+			newMsgs = append(newMsgs, h.Messages...)
+			return &PrepareStepResult{Messages: newMsgs}
+		}
+		return nil // No modification for other steps.
+	})
+
+	// Test step 0 — should modify messages.
+	input := PrepareStepHook{
+		StepNumber: 0,
+		Messages:   []LLMMessage{NewLLMUserMessage("hello")},
+	}
+	result := hr.run(input)
+	if result == nil {
+		t.Fatal("expected non-nil result for step 0")
+	}
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+	if result.Messages[0].Role != LLMRoleSystem {
+		t.Errorf("expected system message first, got role %q", result.Messages[0].Role)
+	}
+
+	// Test step 1 — should return nil (no modification).
+	input.StepNumber = 1
+	result = hr.run(input)
+	if result != nil {
+		t.Errorf("expected nil result for step 1, got %+v", result)
+	}
+}
+
+// TestPrepareStepHookPriority verifies that PrepareStep hooks respect priority ordering.
+func TestPrepareStepHookPriority(t *testing.T) {
+	hr := newHookRegistry[PrepareStepHook, PrepareStepResult]()
+
+	var order []string
+
+	// Low priority — should run second.
+	hr.register(HookPriorityLow, func(_ PrepareStepHook) *PrepareStepResult {
+		order = append(order, "low")
+		return nil
+	})
+
+	// High priority — should run first and win.
+	hr.register(HookPriorityHigh, func(h PrepareStepHook) *PrepareStepResult {
+		order = append(order, "high")
+		return &PrepareStepResult{Messages: h.Messages}
+	})
+
+	input := PrepareStepHook{
+		StepNumber: 0,
+		Messages:   []LLMMessage{NewLLMUserMessage("test")},
+	}
+	result := hr.run(input)
+
+	if result == nil {
+		t.Fatal("expected non-nil result")
+	}
+	if len(order) != 1 || order[0] != "high" {
+		t.Errorf("expected [high] (first non-nil wins), got %v", order)
+	}
+}

pkg/kit/kit.go

@@ -51,6 +51,7 @@ type Kit struct {
 	bufferedLogger *tools.BufferedDebugLogger
 	authHandler    MCPAuthHandler // OAuth handler for remote MCP servers (may need Close)
 	opts           *Options       // stored for reload operations (skills, etc.)
+	mcpConfig      *config.Config // loaded MCP/server config, shared with subagents
 
 	// hasCustomSystemPrompt is true when the user explicitly configured a
 	// system prompt (via --system-prompt flag, config file, or SDK option).
@@ -65,6 +66,7 @@ type Kit struct {
 	afterTurn       *hookRegistry[AfterTurnHook, AfterTurnResult]
 	contextPrepare  *hookRegistry[ContextPrepareHook, ContextPrepareResult]
 	beforeCompact   *hookRegistry[BeforeCompactHook, BeforeCompactResult]
+	prepareStep     *hookRegistry[PrepareStepHook, PrepareStepResult]
 
 	// lastInputTokens stores the API-reported input token count from the
 	// most recent turn. Used by GetContextStats() to return accurate usage
@@ -542,6 +544,23 @@ func (m *Kit) SetModel(ctx context.Context, modelString string) error {
 	systemPrompt, _ := config.LoadSystemPrompt(viper.GetString("system-prompt"))
 	thinkingLevel := models.ParseThinkingLevel(viper.GetString("thinking-level"))
 
+	// Validate and adjust thinking level for the target model.
+	// Some models (e.g., OpenAI gpt-5.4) don't support "minimal" and require "none".
+	if thinkingLevel != models.ThinkingOff {
+		parts := strings.SplitN(modelString, "/", 2)
+		if len(parts) == 2 {
+			modelName := parts[1]
+			if !models.IsValidThinkingLevelForModel(thinkingLevel, modelName) {
+				fallback := models.SuggestThinkingLevelFallback(thinkingLevel, modelName)
+				if fallback != models.ThinkingOff {
+					// Adjust the thinking level in viper so the change persists.
+					viper.Set("thinking-level", string(fallback))
+					thinkingLevel = fallback
+				}
+			}
+		}
+	}
+
 	// With message-level caching, thinking and caching can work together.
 	// No need to disable caching when thinking is enabled.
 	cfg := &models.ProviderConfig{
@@ -713,7 +732,7 @@ func (m *Kit) ExecuteCompletion(ctx context.Context, req extensions.CompleteRequ
 		llmModel    fantasy.LanguageModel
 		closer      func()
 		usedModel   string
-		providerOps fantasy.ProviderOptions
+		providerOps LLMProviderOptions
 	)
 
 	if req.Model == "" {
@@ -810,6 +829,29 @@ func (m *Kit) ExecuteCompletion(ctx context.Context, req extensions.CompleteRequ
 // Options configures Kit creation with optional overrides for model,
 // prompts, configuration, and behavior settings. All fields are optional
 // and will use CLI defaults if not specified.
+//
+// Global viper state warning:
+// Options are applied by [New] via [viper.Set] calls against viper's
+// process-global store. This store is shared with every downstream reader
+// (e.g. [Kit.SetModel], [Kit.GetThinkingLevel], BuildProviderConfig, and
+// any other code path that calls viper.Get*). Two consequences:
+//
+//  1. Kit instances are NOT isolated from each other within a single
+//     process. Values set by the second New() call overwrite the first,
+//     and any code that later reads viper will see the most recent Set.
+//  2. Fields left at the zero value do NOT clear prior viper state; they
+//     simply skip the viper.Set. Callers that need a clean slate between
+//     constructions should invoke viper.Reset() (the test suite uses a
+//     private resetViper() helper that wraps it) before the next New().
+//
+// Recommended usage: create one Kit per process, or reset viper between
+// constructions. Concurrent calls to New are serialized internally by
+// [viperInitMu], but that mutex does not prevent later viper reads (from
+// a different Kit) from observing mutated keys.
+//
+// TODO: refactor New to use a per-instance *viper.Viper (constructed via
+// viper.New()) so each Kit owns its own isolated config store and Options
+// no longer leak through the global singleton.
 type Options struct {
 	Model        string // Override model (e.g., "anthropic/claude-sonnet-4-5-20250929")
 	SystemPrompt string // Override system prompt
@@ -820,6 +862,76 @@ type Options struct {
 	Tools        []Tool // Custom tool set. If empty, AllTools() is used.
 	ExtraTools   []Tool // Additional tools added alongside core/MCP/extension tools.
 
+	// Generation parameters. These override the corresponding values from
+	// .kit.yml / KIT_* environment variables. Leaving a field at its
+	// zero/nil value means "use the configured default", which in turn
+	// falls back to per-model defaults (modelSettings / customModels) and
+	// finally to a last-resort SDK floor of 8192 for MaxTokens (matching
+	// the CLI --max-tokens default; sampling params fall through to
+	// provider-level defaults).
+	//
+	// Pointer types are used for sampling parameters so the SDK can
+	// distinguish "explicitly set to 0" from "leave alone".
+
+	// MaxTokens overrides the maximum output tokens per LLM response.
+	// 0 = let the precedence chain resolve a value (env → config →
+	// per-model → 8192 SDK floor, matching the CLI default). Setting a
+	// non-zero value here suppresses automatic right-sizing, matching
+	// the CLI's --max-tokens flag semantics. Bump this when generating
+	// long outputs (HTML artifacts, large refactors, etc.) to avoid
+	// silent truncation mid-tool-call. The cap also applies after
+	// model switches via [Kit.SetModel].
+	MaxTokens int
+
+	// ThinkingLevel sets the reasoning effort for models that support
+	// extended thinking. Valid values: "off", "none", "minimal", "low",
+	// "medium", "high". "" = let the precedence chain resolve a level
+	// (env → config → per-model → "off"). Use [Kit.SetThinkingLevel]
+	// to change at runtime.
+	ThinkingLevel string
+
+	// Temperature controls sampling randomness (typically 0.0–2.0).
+	// nil = leave provider/per-model default in place. Pointer type
+	// so explicit 0.0 (deterministic) is distinguishable from "unset".
+	Temperature *float32
+
+	// TopP is the nucleus-sampling cutoff (0.0–1.0).
+	// nil = leave provider/per-model default in place.
+	TopP *float32
+
+	// TopK limits sampling to the top K tokens.
+	// nil = leave provider/per-model default in place.
+	TopK *int32
+
+	// FrequencyPenalty discourages repeated tokens (OpenAI-family models).
+	// nil = leave provider/per-model default in place.
+	FrequencyPenalty *float32
+
+	// PresencePenalty discourages repeating topics (OpenAI-family models).
+	// nil = leave provider/per-model default in place.
+	PresencePenalty *float32
+
+	// Provider configuration. These override values normally read from
+	// .kit.yml or provider-specific environment variables. Useful when
+	// loading credentials from a secrets manager, pointing at custom
+	// OpenAI-compatible endpoints (LiteLLM, vLLM, Azure OpenAI, internal
+	// proxies), or running against self-hosted infrastructure.
+
+	// ProviderAPIKey overrides the API key used to authenticate with the
+	// model provider. "" = use the value from config or the
+	// provider-specific environment variable.
+	ProviderAPIKey string
+
+	// ProviderURL overrides the provider endpoint. "" = use the provider's
+	// default URL.
+	ProviderURL string
+
+	// TLSSkipVerify disables TLS certificate verification on provider
+	// HTTP clients. Only set this for self-signed certificates in
+	// development. Once enabled here it cannot be disabled via Options
+	// (use the config file or env var to opt back out).
+	TLSSkipVerify bool
+
 	// SkipConfig, when true, skips loading .kit.yml configuration files.
 	// Viper defaults (setSDKDefaults) and environment variables (KIT_*)
 	// are still applied. Use this for fully programmatic configuration.
@@ -849,6 +961,13 @@ type Options struct {
 	// (e.g. AGENTS.md) from the working directory.
 	NoContextFiles bool
 
+	// MCPConfig provides a pre-loaded MCP configuration. When set,
+	// LoadAndValidateConfig is skipped during Kit creation — avoiding
+	// viper access entirely. This is set automatically for in-process
+	// subagents (inheriting the parent's loaded config) and can be used
+	// by SDK consumers who build config programmatically.
+	MCPConfig *config.Config
+
 	// InProcessMCPServers registers mcp-go servers that run in the same
 	// process. Each key is the server name (used to prefix tool names, e.g.
 	// "docs__search"). The value must be a *[server.MCPServer].
@@ -879,15 +998,23 @@ type Options struct {
 	Debug bool
 
 	// MCPAuthHandler handles OAuth authorization for remote MCP servers.
-	// When set, remote transports (streamable HTTP, SSE) are configured with
-	// OAuth support. If the server returns a 401, the handler is invoked to
-	// let the user authorize via browser.
+	// When set, remote transports (streamable HTTP, SSE) are configured
+	// with OAuth support. If the server returns a 401, the handler is
+	// invoked to let the user authorize.
 	//
-	// If nil, a [DefaultMCPAuthHandler] is created automatically — opening the
-	// system browser and listening on a local callback server.
+	// If nil, OAuth is disabled: remote MCP servers requiring authorization
+	// will fail to connect and the underlying authorization-required error
+	// is surfaced to the caller. The SDK deliberately does not construct a
+	// default handler — doing so would bind a local TCP port and trigger
+	// presentation I/O (browser open, stderr writes) without the consumer
+	// opting in, which is wrong for library, daemon, or web-app embedders.
 	//
-	// Set to a custom implementation to control the authorization UX (e.g.
-	// display a URL in a custom UI, redirect to a web app, etc.).
+	// CLI consumers: pass [NewCLIMCPAuthHandler] to get the standard
+	// "open browser + print status" behavior.
+	//
+	// Custom UX: implement [MCPAuthHandler] directly, or use
+	// [DefaultMCPAuthHandler] and set its OnAuthURL hook to plug in your
+	// own presentation (TUI modal, QR code, web redirect, etc.).
 	MCPAuthHandler MCPAuthHandler
 
 	// MCPTokenStoreFactory, if non-nil, is called to create a token store for
@@ -908,6 +1035,41 @@ type Options struct {
 	// real-time progress in the TUI.
 	OnMCPServerLoaded func(serverName string, toolCount int, err error)
 
+	// MCPTaskMode overrides the per-server [MCPTaskMode] for task-augmented
+	// tools/call execution. Keys are MCP server names. Servers not present
+	// in the map fall back to the TasksMode field of MCPServerConfig (or
+	// MCPTaskModeAuto when that is empty). See the MCP Tasks spec for the
+	// underlying semantics:
+	// https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks
+	MCPTaskMode map[string]MCPTaskMode
+
+	// MCPTaskTimeout is the maximum wall-clock duration to wait for a
+	// task-augmented tool call to reach a terminal state. Independent of
+	// any per-call context deadline; whichever fires first wins. Zero
+	// means use the default (15 minutes).
+	MCPTaskTimeout time.Duration
+
+	// MCPTaskTTL is the TTL hint sent in TaskParams for every
+	// task-augmented tools/call. Zero omits the TTL and lets the server
+	// pick its own retention policy.
+	MCPTaskTTL time.Duration
+
+	// MCPTaskPollInterval is the fallback interval between tasks/get
+	// requests when the server does not suggest one. Zero means use the
+	// default (1 second).
+	MCPTaskPollInterval time.Duration
+
+	// MCPTaskMaxPollInterval caps the polling interval (a server-supplied
+	// pollInterval can otherwise grow without bound). Zero means use the
+	// default (5 seconds).
+	MCPTaskMaxPollInterval time.Duration
+
+	// MCPTaskProgress, if non-nil, is invoked once when a task is accepted
+	// and on every status transition observed by the polling loop. The
+	// final invocation always carries a terminal status. Implementations
+	// must not block; long work should run on a goroutine.
+	MCPTaskProgress MCPTaskProgressHandler
+
 	// CLI is optional CLI-specific configuration. SDK users leave this nil.
 	CLI *CLIOptions
 
@@ -971,14 +1133,29 @@ func InitTreeSession(opts *Options) (*session.TreeManager, error) {
 	return session.CreateTreeSession(sessionDir)
 }
 
+// viperInitMu serializes viper writes during [New]. Viper's global state
+// is not thread-safe, so concurrent calls (e.g. parallel subagent spawns)
+// must not overlap the Set/Get window. Note that this mutex only protects
+// the construction window — it does not isolate long-lived Kit instances
+// from each other. See the "Global viper state warning" on [Options].
+var viperInitMu sync.Mutex
+
 // New creates a Kit instance using the same initialization as the CLI.
 // It loads configuration, initializes MCP servers, creates the LLM model, and
 // sets up the agent for interaction. Returns an error if initialization fails.
-// viperInitMu serializes viper writes during kit.New(). Viper's global state
-// is not thread-safe, so concurrent calls (e.g. parallel subagent spawns)
-// must not overlap the Set()/Get() window.
-var viperInitMu sync.Mutex
-
+//
+// Global viper state warning: fields on [Options] are applied by calling
+// [viper.Set] on viper's process-global store. As a result, two Kits
+// constructed in the same process are NOT isolated: the second New
+// overwrites viper keys set by the first, and any downstream reader
+// (e.g. [Kit.SetModel], [Kit.GetThinkingLevel]) will observe the most
+// recent value. Callers that need multiple independent Kits should call
+// viper.Reset() between constructions, or avoid constructing more than
+// one Kit per process. Writes during New are serialized by [viperInitMu].
+//
+// TODO: refactor to use a per-call viper.New() instance so each Kit owns
+// its own isolated config store and Options stop leaking through the
+// global singleton.
 func New(ctx context.Context, opts *Options) (*Kit, error) {
 	if opts == nil {
 		opts = &Options{}
@@ -1039,6 +1216,47 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		}
 		viper.Set("stream", opts.Streaming)
 
+		// Generation parameter overrides. Each Options field, when set,
+		// is pushed into viper here so the existing downstream code
+		// (BuildProviderConfig, SetModel, modelSettings lookups) picks
+		// it up uniformly. Pointer-typed sampling params use viper.Set
+		// only when non-nil so that nil means "leave provider/per-model
+		// default in place" (BuildProviderConfig keys off viper.IsSet).
+		if opts.MaxTokens > 0 {
+			viper.Set("max-tokens", opts.MaxTokens)
+		}
+		if opts.ThinkingLevel != "" {
+			viper.Set("thinking-level", opts.ThinkingLevel)
+		}
+		if opts.Temperature != nil {
+			viper.Set("temperature", *opts.Temperature)
+		}
+		if opts.TopP != nil {
+			viper.Set("top-p", *opts.TopP)
+		}
+		if opts.TopK != nil {
+			viper.Set("top-k", *opts.TopK)
+		}
+		if opts.FrequencyPenalty != nil {
+			viper.Set("frequency-penalty", *opts.FrequencyPenalty)
+		}
+		if opts.PresencePenalty != nil {
+			viper.Set("presence-penalty", *opts.PresencePenalty)
+		}
+
+		// Provider overrides. TLSSkipVerify only takes effect when true —
+		// callers wanting to force-disable should use the config file or
+		// env var instead.
+		if opts.ProviderAPIKey != "" {
+			viper.Set("provider-api-key", opts.ProviderAPIKey)
+		}
+		if opts.ProviderURL != "" {
+			viper.Set("provider-url", opts.ProviderURL)
+		}
+		if opts.TLSSkipVerify {
+			viper.Set("tls-skip-verify", true)
+		}
+
 		// Resolve working directory for context/skill discovery.
 		cwd = opts.SessionDir
 		if cwd == "" {
@@ -1124,6 +1342,17 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		if pcErr != nil {
 			return fmt.Errorf("failed to build provider config: %w", pcErr)
 		}
+
+		// SDK last-resort max-tokens floor. When nothing — Options, env,
+		// config, nor a per-model default — supplied a value, we land on
+		// zero here (viper.GetInt returns 0 for unset keys). Apply the
+		// SDK default directly on the struct rather than via viper so
+		// viper.IsSet("max-tokens") stays false: downstream right-sizing
+		// can still raise this toward the model's known output ceiling,
+		// and per-model modelSettings[...].maxTokens can still win.
+		if providerConfig.MaxTokens == 0 && opts.MaxTokens == 0 {
+			providerConfig.MaxTokens = sdkDefaultMaxTokens
+		}
 		modelString = viper.GetString("model")
 		debug = viper.GetBool("debug")
 		noExtensions = opts.NoExtensions || viper.GetBool("no-extensions")
@@ -1136,8 +1365,11 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 	}
 	// ---- viperInitMu released — heavy I/O below runs concurrently ----
 
-	// Load MCP configuration. Use pre-loaded config if provided via CLI options.
-	if opts.CLI != nil && opts.CLI.MCPConfig != nil {
+	// Load MCP configuration. Use pre-loaded config if provided directly,
+	// via CLI options, or load from viper as a last resort.
+	if opts.MCPConfig != nil {
+		mcpConfig = opts.MCPConfig
+	} else if opts.CLI != nil && opts.CLI.MCPConfig != nil {
 		mcpConfig = opts.CLI.MCPConfig
 	}
 	if mcpConfig == nil {
@@ -1172,6 +1404,7 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 	afterTurn := newHookRegistry[AfterTurnHook, AfterTurnResult]()
 	contextPrepare := newHookRegistry[ContextPrepareHook, ContextPrepareResult]()
 	beforeCompact := newHookRegistry[BeforeCompactHook, BeforeCompactResult]()
+	prepareStep := newHookRegistry[PrepareStepHook, PrepareStepResult]()
 
 	// Build agent setup options, pulling CLI-specific fields when available.
 	// Pass the pre-built ProviderConfig and scalar viper snapshots so
@@ -1189,22 +1422,29 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		MaxSteps:          maxSteps,
 		StreamingEnabled:  streaming,
 		OnMCPServerLoaded: opts.OnMCPServerLoaded,
+		MCPTaskConfig: mcpTaskOptions{
+			perServer:       opts.MCPTaskMode,
+			defaultTTL:      opts.MCPTaskTTL,
+			pollInterval:    opts.MCPTaskPollInterval,
+			maxPollInterval: opts.MCPTaskMaxPollInterval,
+			timeout:         opts.MCPTaskTimeout,
+			progress:        opts.MCPTaskProgress,
+		}.toToolsConfig(),
 	}
 
-	// Set up OAuth handler for remote MCP servers.
+	// Set up OAuth handler for remote MCP servers. The SDK does not create
+	// a default handler: auto-construction would bind a local TCP port and
+	// (historically) shell out to a browser without the consumer asking,
+	// which is a surprise for library/daemon/web-app embedders. Consumers
+	// that want CLI behavior pass a [CLIMCPAuthHandler] explicitly; other
+	// consumers implement [MCPAuthHandler] themselves. If nil, remote MCP
+	// servers requiring OAuth will fail to connect with the underlying
+	// authorization-required error surfaced to the caller.
+	//
 	// The SDK MCPAuthHandler interface is structurally identical to
 	// tools.MCPAuthHandler, so any implementation satisfies both.
 	if opts.MCPAuthHandler != nil {
 		setupOpts.AuthHandler = opts.MCPAuthHandler
-	} else {
-		// Create a default handler that opens the system browser.
-		defaultHandler, authErr := NewDefaultMCPAuthHandler()
-		if authErr != nil {
-			// Non-fatal: OAuth just won't be available for remote servers.
-			log.Printf("WARN Failed to create OAuth handler; remote MCP servers requiring auth will fail: %v", authErr)
-		} else {
-			setupOpts.AuthHandler = defaultHandler
-		}
 	}
 
 	// Set up custom token store factory for MCP OAuth tokens.
@@ -1258,6 +1498,7 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		bufferedLogger:        agentResult.BufferedLogger,
 		authHandler:           setupOpts.AuthHandler,
 		opts:                  opts,
+		mcpConfig:             mcpConfig,
 		hasCustomSystemPrompt: hasCustomSystemPrompt,
 		beforeToolCall:        beforeToolCall,
 		afterToolResult:       afterToolResult,
@@ -1265,6 +1506,7 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		afterTurn:             afterTurn,
 		contextPrepare:        contextPrepare,
 		beforeCompact:         beforeCompact,
+		prepareStep:           prepareStep,
 	}
 
 	// Bridge extension events to SDK hooks.
@@ -1439,8 +1681,9 @@ type TurnResult struct {
 	Response string
 
 	// StopReason indicates why the turn ended. Derived from the LLM
-	// provider's finish reason: "stop", "length" (max tokens), "tool-calls",
-	// "content-filter", "error", "other", "unknown".
+	// provider's finish reason: FinishReasonStop, FinishReasonLength (max
+	// output tokens reached), FinishReasonToolCalls, FinishReasonContentFilter,
+	// FinishReasonError, FinishReasonOther, FinishReasonUnknown.
 	StopReason string
 
 	// SessionID is the UUID of the session this turn belongs to.
@@ -1582,14 +1825,30 @@ func (m *Kit) Subagent(ctx context.Context, cfg SubagentConfig) (*SubagentResult
 		tools = SubagentTools()
 	}
 
-	// Create child Kit instance.
+	// Create child Kit instance. Pass the parent's loaded MCP config to
+	// avoid re-reading viper (which races with concurrent subagent spawns).
+	// Streaming must be explicitly enabled — Options.Streaming defaults to
+	// false, and New() unconditionally writes viper.Set("stream", opts.Streaming).
+	// Without this, the subagent would (a) pollute viper global state for
+	// other concurrent callers and (b) potentially hit provider-level
+	// differences (e.g. Anthropic non-streaming timeouts with extended
+	// thinking).
 	childOpts := &Options{
 		Model:        model,
 		SystemPrompt: systemPrompt,
 		Tools:        tools,
 		NoSession:    cfg.NoSession,
 		Quiet:        true,
+		Streaming:    true,
+		MCPConfig:    m.mcpConfig,
 	}
+	// Propagate the parent's MCP task configuration so a child subagent
+	// invoking long-running MCP tools observes the same per-server modes,
+	// timeouts, and progress callback as the parent. Without this, child
+	// agents would silently fall back to MCPTaskModeAuto with default
+	// polling and no progress feedback even when the parent had configured
+	// custom values.
+	inheritMCPTaskOptions(childOpts, m.opts)
 	child, err := New(ctx, childOpts)
 	if err != nil {
 		return &SubagentResult{Elapsed: time.Since(start)}, fmt.Errorf("failed to create subagent: %w", err)
@@ -1695,21 +1954,21 @@ func (m *Kit) generate(ctx context.Context, messages []fantasy.Message) (*agent.
 		return sr, err
 	})
 
-	return m.agent.GenerateWithLoopAndStreaming(ctx, messages,
-		func(toolCallID, toolName, toolArgs string) {
+	return m.agent.GenerateWithCallbacks(ctx, messages, agent.GenerateCallbacks{
+		OnToolCall: func(toolCallID, toolName, toolArgs string) {
 			m.events.emit(ToolCallEvent{
 				ToolCallID: toolCallID, ToolName: toolName, ToolKind: toolKindFor(toolName),
 				ToolArgs: toolArgs, ParsedArgs: parseToolArgs(toolArgs),
 			})
 		},
-		func(toolCallID, toolName, toolArgs string, isStarting bool) {
+		OnToolExecution: func(toolCallID, toolName, toolArgs string, isStarting bool) {
 			if isStarting {
 				m.events.emit(ToolExecutionStartEvent{ToolCallID: toolCallID, ToolName: toolName, ToolKind: toolKindFor(toolName), ToolArgs: toolArgs})
 			} else {
 				m.events.emit(ToolExecutionEndEvent{ToolCallID: toolCallID, ToolName: toolName, ToolKind: toolKindFor(toolName)})
 			}
 		},
-		func(toolCallID, toolName, toolArgs, resultText, metadata string, isError bool) {
+		OnToolResult: func(toolCallID, toolName, toolArgs, resultText, metadata string, isError bool) {
 			evt := ToolResultEvent{
 				ToolCallID: toolCallID, ToolName: toolName, ToolKind: toolKindFor(toolName),
 				ToolArgs: toolArgs, ParsedArgs: parseToolArgs(toolArgs),
@@ -1723,17 +1982,17 @@ func (m *Kit) generate(ctx context.Context, messages []fantasy.Message) (*agent.
 			}
 			m.events.emit(evt)
 		},
-		func(content string) {
+		OnResponse: func(content string) {
 			m.events.emit(ResponseEvent{Content: content})
 		},
-		func(content string) {
+		OnToolCallContent: func(content string) {
 			m.events.emit(ToolCallContentEvent{Content: content})
 		},
 		// <think> tag filtering: models like Qwen/DeepSeek wrap reasoning inside
 		// <think>...</think> tags in the regular text stream. We intercept those
 		// spans here and re-route them as ReasoningDeltaEvent/ReasoningCompleteEvent
 		// so callers always receive clean, tag-free text and structured reasoning.
-		func() func(chunk string) {
+		OnStreamingResponse: func() func(chunk string) {
 			const (
 				thinkOpen  = "<think>"
 				thinkClose = "</think>"
@@ -1769,14 +2028,13 @@ func (m *Kit) generate(ctx context.Context, messages []fantasy.Message) (*agent.
 				}
 			}
 		}(),
-		func(delta string) {
+		OnReasoningDelta: func(delta string) {
 			m.events.emit(ReasoningDeltaEvent{Delta: delta})
 		},
-		func() {
+		OnReasoningComplete: func() {
 			m.events.emit(ReasoningCompleteEvent{})
 		},
-		func(toolCallID, toolName, chunk string, isStderr bool) {
-			// Emit tool output chunk event for streaming bash output
+		OnToolOutput: func(toolCallID, toolName, chunk string, isStderr bool) {
 			m.events.emit(ToolOutputEvent{
 				ToolCallID: toolCallID,
 				ToolName:   toolName,
@@ -1785,18 +2043,13 @@ func (m *Kit) generate(ctx context.Context, messages []fantasy.Message) (*agent.
 			})
 		},
 		// Persist step messages incrementally so that progress survives
-		// crashes and long-running turns don't lose work. Each step's
-		// messages are persisted as a unit: for tool-calling steps this is
-		// the assistant message (with tool_use parts) + tool-role message
-		// (with tool_result parts) as a pair; for the final step it's the
-		// assistant text/reasoning message alone.
-		func(stepMessages []fantasy.Message) {
+		// crashes and long-running turns don't lose work.
+		OnStepMessages: func(stepMessages []fantasy.Message) {
 			for _, msg := range stepMessages {
 				_, _ = m.session.AppendMessage(msg)
 			}
 		},
-		func(inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens int64) {
-			// Emit step usage event for real-time cost tracking
+		OnStepUsage: func(inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens int64) {
 			if viper.GetBool("debug") {
 				log.Printf("DEBUG Kit.generate emitting StepUsageEvent: input=%d output=%d cacheRead=%d cacheCreate=%d",
 					inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens,
@@ -1810,18 +2063,97 @@ func (m *Kit) generate(ctx context.Context, messages []fantasy.Message) (*agent.
 			})
 		},
 		// Password prompt handler for sudo commands
-		func(prompt string) (string, bool) {
-			// Emit event to TUI and wait for response via channel
+		OnPasswordPrompt: func(prompt string) (string, bool) {
 			responseCh := make(chan PasswordPromptResponse, 1)
 			m.events.emit(PasswordPromptEvent{
 				Prompt:     prompt,
 				ResponseCh: responseCh,
 			})
-			// Wait for response (TUI will send password or cancel)
 			resp := <-responseCh
 			return resp.Password, resp.Cancelled
 		},
-	)
+		// Tool call argument streaming
+		OnToolCallStart: func(toolCallID, toolName string) {
+			m.events.emit(ToolCallStartEvent{
+				ToolCallID: toolCallID,
+				ToolName:   toolName,
+				ToolKind:   toolKindFor(toolName),
+			})
+		},
+		OnToolCallDelta: func(toolCallID, delta string) {
+			m.events.emit(ToolCallDeltaEvent{
+				ToolCallID: toolCallID,
+				Delta:      delta,
+			})
+		},
+		OnToolCallEnd: func(toolCallID string) {
+			m.events.emit(ToolCallEndEvent{
+				ToolCallID: toolCallID,
+			})
+		},
+
+		// New callbacks for previously unwired Fantasy lifecycle events.
+		OnStepStart: func(stepNumber int) {
+			m.events.emit(StepStartEvent{StepNumber: stepNumber})
+		},
+		OnStepFinish: func(stepNumber int, hasToolCalls bool, finishReason string, usage fantasy.Usage) {
+			m.events.emit(StepFinishEvent{
+				StepNumber:   stepNumber,
+				HasToolCalls: hasToolCalls,
+				FinishReason: finishReason,
+				Usage:        usage,
+			})
+		},
+		OnTextStart: func(id string) {
+			m.events.emit(TextStartEvent{ID: id})
+		},
+		OnTextEnd: func(id string) {
+			m.events.emit(TextEndEvent{ID: id})
+		},
+		OnReasoningStart: func(id string) {
+			m.events.emit(ReasoningStartEvent{ID: id})
+		},
+		OnWarnings: func(warnings []string) {
+			m.events.emit(WarningsEvent{Warnings: warnings})
+		},
+		OnSource: func(sourceType, id, url, title string) {
+			m.events.emit(SourceEvent{
+				SourceType: sourceType,
+				ID:         id,
+				URL:        url,
+				Title:      title,
+			})
+		},
+		OnStreamFinish: func(usage fantasy.Usage, finishReason string) {
+			m.events.emit(StreamFinishEvent{
+				Usage:        usage,
+				FinishReason: finishReason,
+			})
+		},
+		OnError: func(err error) {
+			m.events.emit(ErrorEvent{Error: err})
+		},
+		OnRetry: func(attempt int, err error) {
+			m.events.emit(RetryEvent{Attempt: attempt, Error: err})
+		},
+		// PrepareStep hook — compose with steering (handled in agent layer)
+		// and then run SDK consumer hooks.
+		OnPrepareStep: func() agent.PrepareStepHandler {
+			if !m.prepareStep.hasHooks() {
+				return nil
+			}
+			return func(stepNumber int, messages []fantasy.Message) []fantasy.Message {
+				hookResult := m.prepareStep.run(PrepareStepHook{
+					StepNumber: stepNumber,
+					Messages:   messages,
+				})
+				if hookResult != nil && hookResult.Messages != nil {
+					return hookResult.Messages
+				}
+				return nil
+			}
+		}(),
+	})
 }
 
 // runTurn is the shared lifecycle for every prompt mode:
@@ -2235,6 +2567,35 @@ func (m *Kit) GetTools() []Tool {
 	return m.agent.GetTools()
 }
 
+// MaxTokens returns the effective max output tokens currently configured for
+// the agent. This is the value actually sent to the LLM provider on each
+// request, after CLI/env/config resolution, per-model overrides, model-aware
+// right-sizing, and any Anthropic thinking-budget adjustments.
+//
+// Returns 0 when the active provider suppresses the max_output_tokens
+// parameter (e.g. OpenAI Codex OAuth) or when no model is configured yet.
+// A non-zero value is the number that will cause a FinishReasonLength
+// truncation if the model tries to generate beyond it.
+func (m *Kit) MaxTokens() int {
+	if m.agent == nil {
+		return 0
+	}
+	return m.agent.GetMaxTokens()
+}
+
+// MaxOutputLimit returns the catalog-reported output ceiling for the current
+// model in tokens, or 0 when the model isn't in the registry (custom models,
+// new releases, Ollama, etc.). Pair with MaxTokens() to detect when the agent
+// is configured well below what the model supports and surface a hint to the
+// user.
+func (m *Kit) MaxOutputLimit() int {
+	info := m.GetModelInfo()
+	if info == nil {
+		return 0
+	}
+	return info.Limit.Output
+}
+
 // extractFileParts returns all FilePart entries from a message's Content.
 // Used to preserve image attachments when replacing user message text.
 func extractFileParts(msg fantasy.Message) []fantasy.FilePart {

pkg/kit/kit_test.go

@@ -5,6 +5,8 @@ import (
 	"os"
 	"testing"
 
+	"github.com/spf13/viper"
+
 	kit "github.com/mark3labs/kit/pkg/kit"
 )
 
@@ -54,6 +56,225 @@ func TestNewWithOptions(t *testing.T) {
 	}
 }
 
+// TestNewWithGenerationOptions verifies that the SDK-only generation
+// parameter overrides on Options propagate all the way through to the
+// agent without requiring any viper.Set workarounds in caller code.
+func TestNewWithGenerationOptions(t *testing.T) {
+	if os.Getenv("ANTHROPIC_API_KEY") == "" {
+		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
+	}
+
+	ctx := context.Background()
+
+	// MaxTokens override — keep ThinkingLevel off so Anthropic's thinking
+	// budget doesn't auto-bump MaxTokens above what we configured.
+	t.Run("MaxTokens", func(t *testing.T) {
+		defer resetViper()
+
+		const want = 12345
+		host, err := kit.New(ctx, &kit.Options{
+			Model:     "anthropic/claude-sonnet-4-5-20250929",
+			Quiet:     true,
+			MaxTokens: want,
+		})
+		if err != nil {
+			t.Fatalf("Failed to create Kit: %v", err)
+		}
+		defer func() { _ = host.Close() }()
+
+		if got := host.MaxTokens(); got != want {
+			t.Errorf("Options.MaxTokens=%d did not propagate; Kit.MaxTokens()=%d", want, got)
+		}
+		if !viper.IsSet("max-tokens") {
+			t.Error("viper.IsSet(\"max-tokens\") should be true after MaxTokens override")
+		}
+	})
+
+	// ThinkingLevel override — verified via the public getter, which
+	// reads back the configured (not provider-derived) level.
+	t.Run("ThinkingLevel", func(t *testing.T) {
+		defer resetViper()
+
+		const want = "high"
+		host, err := kit.New(ctx, &kit.Options{
+			Model:         "anthropic/claude-sonnet-4-5-20250929",
+			Quiet:         true,
+			ThinkingLevel: want,
+		})
+		if err != nil {
+			t.Fatalf("Failed to create Kit: %v", err)
+		}
+		defer func() { _ = host.Close() }()
+
+		if got := host.GetThinkingLevel(); got != want {
+			t.Errorf("Options.ThinkingLevel=%q did not propagate; Kit.GetThinkingLevel()=%q", want, got)
+		}
+	})
+
+	// Temperature override — pointer semantics let callers distinguish
+	// "explicitly 0.0" from "unset", which we assert by pushing a distinct
+	// value and reading it back off viper's merged state.
+	t.Run("Temperature", func(t *testing.T) {
+		defer resetViper()
+
+		want := float32(0.12345)
+		host, err := kit.New(ctx, &kit.Options{
+			Model:       "anthropic/claude-sonnet-4-5-20250929",
+			Quiet:       true,
+			Temperature: &want,
+		})
+		if err != nil {
+			t.Fatalf("Failed to create Kit: %v", err)
+		}
+		defer func() { _ = host.Close() }()
+
+		if !viper.IsSet("temperature") {
+			t.Fatal("viper.IsSet(\"temperature\") should be true after Temperature override")
+		}
+		if got := float32(viper.GetFloat64("temperature")); got != want {
+			t.Errorf("Options.Temperature=%v did not propagate; viper=%v", want, got)
+		}
+	})
+}
+
+// TestNewPreservesIsSetSemantics verifies that creating a Kit WITHOUT
+// populating the generation-param Options fields does NOT mark those
+// keys as explicitly set in viper. This is the precedence contract
+// that per-model defaults (ApplyModelSettings) and right-sizing
+// (rightSizeMaxTokens) rely on.
+//
+// Previously setSDKDefaults() used viper.SetDefault() for every param,
+// which caused viper.IsSet() to return true for all of them — silently
+// suppressing per-model defaults and pinning max-tokens at 4096 even
+// on models with much larger output limits.
+func TestNewPreservesIsSetSemantics(t *testing.T) {
+	if os.Getenv("ANTHROPIC_API_KEY") == "" {
+		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
+	}
+
+	defer resetViper()
+
+	ctx := context.Background()
+	host, err := kit.New(ctx, &kit.Options{
+		Model:      "anthropic/claude-sonnet-4-5-20250929",
+		Quiet:      true,
+		NoSession:  true,
+		SkipConfig: true, // isolate from any ~/.kit.yml values
+	})
+	if err != nil {
+		t.Fatalf("Failed to create Kit: %v", err)
+	}
+	defer func() { _ = host.Close() }()
+
+	// These keys must remain "unset" from viper's perspective so the
+	// downstream isExplicitlySet() checks allow per-model defaults to
+	// take effect.
+	checkKeys := []string{
+		"max-tokens",
+		"temperature",
+		"top-p",
+		"top-k",
+		"frequency-penalty",
+		"presence-penalty",
+		"thinking-level",
+	}
+
+	// With SkipConfig: true, InitConfig() is not invoked, so viper has
+	// no env-var bindings registered. Any IsSet() here would come purely
+	// from SDK-side SetDefault/Set calls — which is exactly what this
+	// test is guarding against.
+	for _, k := range checkKeys {
+		if viper.IsSet(k) {
+			t.Errorf("viper.IsSet(%q) == true when no Options field set it "+
+				"(SDK defaults must not corrupt IsSet semantics)", k)
+		}
+	}
+}
+
+// TestNewWithProviderOptions verifies that programmatic provider overrides
+// (API key, URL) take effect without env vars or config files, and that
+// Options.ProviderAPIKey *wins* over any pre-existing viper state.
+func TestNewWithProviderOptions(t *testing.T) {
+	if os.Getenv("ANTHROPIC_API_KEY") == "" {
+		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
+	}
+
+	ctx := context.Background()
+
+	t.Run("succeeds with API key from Options", func(t *testing.T) {
+		defer resetViper()
+
+		apiKey := os.Getenv("ANTHROPIC_API_KEY")
+		host, err := kit.New(ctx, &kit.Options{
+			Model:          "anthropic/claude-sonnet-4-5-20250929",
+			Quiet:          true,
+			NoSession:      true,
+			ProviderAPIKey: apiKey,
+		})
+		if err != nil {
+			t.Fatalf("Failed to create Kit with ProviderAPIKey option: %v", err)
+		}
+		defer func() { _ = host.Close() }()
+
+		if got := viper.GetString("provider-api-key"); got != apiKey {
+			t.Errorf("Options.ProviderAPIKey did not propagate to viper; got %q (len=%d)", got, len(got))
+		}
+	})
+
+	// Override precedence: even when viper already holds a different
+	// provider-api-key value (as it would if a config file or earlier
+	// Set() call populated one), Options.ProviderAPIKey must win.
+	t.Run("Options override beats pre-existing viper state", func(t *testing.T) {
+		defer resetViper()
+
+		viper.Set("provider-api-key", "sk-config-file-placeholder")
+
+		want := "sk-from-options-override"
+		// Use an OpenAI-flavored model so the validation path accepts
+		// the placeholder without attempting a real Anthropic handshake.
+		host, err := kit.New(ctx, &kit.Options{
+			Model:            "openai/gpt-4o-mini",
+			Quiet:            true,
+			NoSession:        true,
+			NoExtensions:     true,
+			DisableCoreTools: true,
+			ProviderAPIKey:   want,
+		})
+		// Creation may still fail if the model registry is strict, but
+		// we only care that the override reached viper before any
+		// provider handshake happened.
+		if host != nil {
+			defer func() { _ = host.Close() }()
+		}
+		_ = err
+
+		if got := viper.GetString("provider-api-key"); got != want {
+			t.Errorf("Options.ProviderAPIKey did not override pre-existing viper value; got %q, want %q", got, want)
+		}
+	})
+
+	// ProviderURL override must also reach viper.
+	t.Run("ProviderURL propagates", func(t *testing.T) {
+		defer resetViper()
+
+		const want = "https://custom.example.com/v1"
+		host, err := kit.New(ctx, &kit.Options{
+			Model:       "anthropic/claude-sonnet-4-5-20250929",
+			Quiet:       true,
+			NoSession:   true,
+			ProviderURL: want,
+		})
+		if err != nil {
+			t.Fatalf("Failed to create Kit with ProviderURL option: %v", err)
+		}
+		defer func() { _ = host.Close() }()
+
+		if got := viper.GetString("provider-url"); got != want {
+			t.Errorf("Options.ProviderURL did not propagate; got %q, want %q", got, want)
+		}
+	})
+}
+
 func TestSessionManagement(t *testing.T) {
 	if os.Getenv("ANTHROPIC_API_KEY") == "" {
 		t.Skip("Skipping test: ANTHROPIC_API_KEY not set")
@@ -81,3 +302,7 @@ func TestSessionManagement(t *testing.T) {
 		t.Error("Expected non-empty session ID")
 	}
 }
+
+// resetViper wipes viper's global state so a test case doesn't leak
+// viper.Set() calls into the next one. Used via defer in subtests.
+func resetViper() { viper.Reset() }

pkg/kit/mcp_tasks.go

@@ -0,0 +1,236 @@
+package kit
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/mark3labs/kit/internal/tools"
+	"github.com/mark3labs/mcp-go/mcp"
+)
+
+// MCPTaskStatus represents the lifecycle state of a task-augmented MCP
+// tool call. See https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks
+// for the underlying spec.
+type MCPTaskStatus string
+
+const (
+	// MCPTaskStatusWorking indicates the task is currently being processed.
+	MCPTaskStatusWorking MCPTaskStatus = MCPTaskStatus(mcp.TaskStatusWorking)
+	// MCPTaskStatusInputRequired indicates the server is waiting for client
+	// input before it can proceed (rare; typically surfaced via elicitation).
+	MCPTaskStatusInputRequired MCPTaskStatus = MCPTaskStatus(mcp.TaskStatusInputRequired)
+	// MCPTaskStatusCompleted indicates the task finished successfully.
+	MCPTaskStatusCompleted MCPTaskStatus = MCPTaskStatus(mcp.TaskStatusCompleted)
+	// MCPTaskStatusFailed indicates the task ended in error.
+	MCPTaskStatusFailed MCPTaskStatus = MCPTaskStatus(mcp.TaskStatusFailed)
+	// MCPTaskStatusCancelled indicates the task was cancelled before completion.
+	MCPTaskStatusCancelled MCPTaskStatus = MCPTaskStatus(mcp.TaskStatusCancelled)
+)
+
+// IsTerminal reports whether the status represents a final state — that is,
+// the task will not change again. Terminal states are completed, failed,
+// and cancelled.
+func (s MCPTaskStatus) IsTerminal() bool {
+	return mcp.TaskStatus(s).IsTerminal()
+}
+
+// MCPTaskMode controls when Kit augments tools/call requests with MCP task
+// metadata for a specific server.
+type MCPTaskMode string
+
+const (
+	// MCPTaskModeAuto augments tools/call with task metadata only when the
+	// server advertises tasks/toolCalls capability during initialize.
+	// This is the default and is safe to leave unconfigured for any
+	// existing MCP server.
+	MCPTaskModeAuto MCPTaskMode = MCPTaskMode(tools.MCPTaskModeAuto)
+	// MCPTaskModeNever forces every tools/call to be issued synchronously
+	// (no Task field), regardless of server capability.
+	MCPTaskModeNever MCPTaskMode = MCPTaskMode(tools.MCPTaskModeNever)
+	// MCPTaskModeAlways always opts into task augmentation, even when the
+	// server didn't advertise the capability. The server may still respond
+	// synchronously; this just expresses client intent unconditionally.
+	MCPTaskModeAlways MCPTaskMode = MCPTaskMode(tools.MCPTaskModeAlways)
+)
+
+// MCPTask is the SDK-level view of an MCP Task. Timestamps are best-effort
+// parsed from the server's ISO-8601 strings; they may be the zero time when
+// the server omitted them or used a non-RFC3339 format.
+type MCPTask struct {
+	// Server is the configured MCP server name this task lives on.
+	Server string
+	// TaskID is the server-assigned identifier for the task.
+	TaskID string
+	// Status is the current task lifecycle state.
+	Status MCPTaskStatus
+	// StatusMessage is an optional human-readable description provided by
+	// the server.
+	StatusMessage string
+	// CreatedAt is when the task was created on the server.
+	CreatedAt time.Time
+	// UpdatedAt is when the task was last updated on the server.
+	UpdatedAt time.Time
+	// TTL is how long the server intends to retain this task after creation.
+	// Zero means the server did not advertise a TTL.
+	TTL time.Duration
+	// PollInterval is the suggested time between status checks. Zero means
+	// the client should use its own default.
+	PollInterval time.Duration
+}
+
+// MCPTaskProgress is a single status update emitted while Kit is waiting
+// on a task-augmented tool call.
+type MCPTaskProgress struct {
+	// Server is the configured MCP server name.
+	Server string
+	// TaskID is the server-assigned identifier for the in-flight task.
+	TaskID string
+	// Status is the most recent task status observed.
+	Status MCPTaskStatus
+	// Message is the optional human-readable status message from the server.
+	Message string
+}
+
+// MCPTaskProgressHandler is called once when a task is accepted and again
+// on every observed status transition. The final invocation always carries
+// a terminal status. Implementations must not block; long work should be
+// dispatched on a goroutine.
+type MCPTaskProgressHandler func(MCPTaskProgress)
+
+// mcpTaskOptions carries SDK consumer configuration into the agent setup.
+// Stored on Options as a single value so the public surface stays compact;
+// individual fields are exposed via WithMCP* builder functions.
+type mcpTaskOptions struct {
+	perServer       map[string]MCPTaskMode
+	defaultTTL      time.Duration
+	pollInterval    time.Duration
+	maxPollInterval time.Duration
+	timeout         time.Duration
+	progress        MCPTaskProgressHandler
+}
+
+// toToolsConfig converts the SDK-level config to the internal tools-package
+// representation. Keeps the dependency arrow internal-only.
+func (o mcpTaskOptions) toToolsConfig() tools.MCPTaskConfig {
+	cfg := tools.MCPTaskConfig{
+		DefaultTTL:      o.defaultTTL,
+		PollInterval:    o.pollInterval,
+		MaxPollInterval: o.maxPollInterval,
+		Timeout:         o.timeout,
+	}
+	if len(o.perServer) > 0 {
+		cfg.PerServerMode = make(map[string]tools.MCPTaskMode, len(o.perServer))
+		for k, v := range o.perServer {
+			cfg.PerServerMode[k] = tools.MCPTaskMode(v)
+		}
+	}
+	if o.progress != nil {
+		h := o.progress
+		cfg.Progress = func(p tools.MCPTaskProgress) {
+			h(MCPTaskProgress{
+				Server:  p.Server,
+				TaskID:  p.TaskID,
+				Status:  MCPTaskStatus(p.Status),
+				Message: p.Message,
+			})
+		}
+	}
+	return cfg
+}
+
+// ListMCPTasks queries tasks/list on the named MCP server and returns the
+// active and recent tasks the server is willing to disclose. Returns an
+// error when the server isn't loaded, doesn't expose tasks/list, or the
+// underlying transport fails.
+func (m *Kit) ListMCPTasks(ctx context.Context, serverName string) ([]MCPTask, error) {
+	mgr, err := m.mcpToolManager()
+	if err != nil {
+		return nil, err
+	}
+	infos, err := mgr.ListServerTasks(ctx, serverName)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]MCPTask, len(infos))
+	for i, t := range infos {
+		out[i] = mcpTaskFromInternal(t)
+	}
+	return out, nil
+}
+
+// GetMCPTask queries tasks/get for a single in-flight task on the named
+// server. The returned MCPTask reflects the server's current view of the
+// task.
+func (m *Kit) GetMCPTask(ctx context.Context, serverName, taskID string) (MCPTask, error) {
+	mgr, err := m.mcpToolManager()
+	if err != nil {
+		return MCPTask{}, err
+	}
+	info, err := mgr.GetServerTask(ctx, serverName, taskID)
+	if err != nil {
+		return MCPTask{}, err
+	}
+	return mcpTaskFromInternal(info), nil
+}
+
+// CancelMCPTask issues tasks/cancel for an in-flight task on the named
+// server. Returns the post-cancel task state when the server responded
+// with one. Cancelling an already-terminal task is a no-op on most
+// servers.
+func (m *Kit) CancelMCPTask(ctx context.Context, serverName, taskID string) (MCPTask, error) {
+	mgr, err := m.mcpToolManager()
+	if err != nil {
+		return MCPTask{}, err
+	}
+	info, err := mgr.CancelServerTask(ctx, serverName, taskID)
+	if err != nil {
+		return MCPTask{}, err
+	}
+	return mcpTaskFromInternal(info), nil
+}
+
+// mcpToolManager returns the underlying MCP tool manager or an error when
+// no MCP servers are configured.
+func (m *Kit) mcpToolManager() (*tools.MCPToolManager, error) {
+	if m == nil || m.agent == nil {
+		return nil, fmt.Errorf("kit instance has no agent")
+	}
+	mgr := m.agent.GetMCPToolManager()
+	if mgr == nil {
+		return nil, fmt.Errorf("no MCP servers configured")
+	}
+	return mgr, nil
+}
+
+// mcpTaskFromInternal adapts the internal tools.MCPTaskInfo to the
+// SDK-level MCPTask type. Keeps the public surface independent of
+// internal package types.
+func mcpTaskFromInternal(t tools.MCPTaskInfo) MCPTask {
+	return MCPTask{
+		Server:        t.Server,
+		TaskID:        t.TaskID,
+		Status:        MCPTaskStatus(t.Status),
+		StatusMessage: t.StatusMessage,
+		CreatedAt:     t.CreatedAt,
+		UpdatedAt:     t.UpdatedAt,
+		TTL:           t.TTL,
+		PollInterval:  t.PollInterval,
+	}
+}
+
+// inheritMCPTaskOptions copies every MCP task-related field from parent
+// onto child. Used by Kit.Subagent so child instances observe the same
+// per-server modes, timeouts, and progress callback as their parent.
+// A nil parent is a no-op so callers don't have to guard at the call site.
+func inheritMCPTaskOptions(child, parent *Options) {
+	if child == nil || parent == nil {
+		return
+	}
+	child.MCPTaskMode = parent.MCPTaskMode
+	child.MCPTaskTimeout = parent.MCPTaskTimeout
+	child.MCPTaskTTL = parent.MCPTaskTTL
+	child.MCPTaskPollInterval = parent.MCPTaskPollInterval
+	child.MCPTaskMaxPollInterval = parent.MCPTaskMaxPollInterval
+	child.MCPTaskProgress = parent.MCPTaskProgress
+}

pkg/kit/mcp_tasks_test.go

@@ -0,0 +1,165 @@
+package kit
+
+import (
+	"testing"
+	"time"
+
+	"github.com/mark3labs/kit/internal/tools"
+)
+
+func TestMCPTaskStatusIsTerminal(t *testing.T) {
+	cases := []struct {
+		s    MCPTaskStatus
+		want bool
+	}{
+		{MCPTaskStatusWorking, false},
+		{MCPTaskStatusInputRequired, false},
+		{MCPTaskStatusCompleted, true},
+		{MCPTaskStatusFailed, true},
+		{MCPTaskStatusCancelled, true},
+		{MCPTaskStatus("unknown"), false},
+	}
+	for _, tc := range cases {
+		if got := tc.s.IsTerminal(); got != tc.want {
+			t.Errorf("MCPTaskStatus(%q).IsTerminal() = %v, want %v", tc.s, got, tc.want)
+		}
+	}
+}
+
+func TestMCPTaskOptionsToToolsConfig(t *testing.T) {
+	called := 0
+	o := mcpTaskOptions{
+		perServer: map[string]MCPTaskMode{
+			"alpha": MCPTaskModeAlways,
+			"beta":  MCPTaskModeNever,
+		},
+		defaultTTL:      30 * time.Second,
+		pollInterval:    250 * time.Millisecond,
+		maxPollInterval: 2 * time.Second,
+		timeout:         5 * time.Minute,
+		progress:        func(p MCPTaskProgress) { called++ },
+	}
+	cfg := o.toToolsConfig()
+
+	if cfg.DefaultTTL != 30*time.Second {
+		t.Errorf("DefaultTTL = %v, want 30s", cfg.DefaultTTL)
+	}
+	if cfg.PollInterval != 250*time.Millisecond {
+		t.Errorf("PollInterval = %v, want 250ms", cfg.PollInterval)
+	}
+	if cfg.MaxPollInterval != 2*time.Second {
+		t.Errorf("MaxPollInterval = %v, want 2s", cfg.MaxPollInterval)
+	}
+	if cfg.Timeout != 5*time.Minute {
+		t.Errorf("Timeout = %v, want 5m", cfg.Timeout)
+	}
+	if cfg.PerServerMode["alpha"] != tools.MCPTaskModeAlways {
+		t.Errorf("PerServerMode[alpha] = %q, want always", cfg.PerServerMode["alpha"])
+	}
+	if cfg.PerServerMode["beta"] != tools.MCPTaskModeNever {
+		t.Errorf("PerServerMode[beta] = %q, want never", cfg.PerServerMode["beta"])
+	}
+
+	// Progress conversion: invoking the internal handler must call our
+	// SDK-level callback with the converted struct.
+	if cfg.Progress == nil {
+		t.Fatal("Progress callback was lost in conversion")
+	}
+	cfg.Progress(tools.MCPTaskProgress{
+		Server: "alpha",
+		TaskID: "t1",
+		Status: "working",
+	})
+	if called != 1 {
+		t.Errorf("expected SDK progress handler to be invoked once, got %d", called)
+	}
+}
+
+func TestMCPTaskFromInternal(t *testing.T) {
+	in := tools.MCPTaskInfo{
+		Server:        "srv",
+		TaskID:        "t-1",
+		Status:        "working",
+		StatusMessage: "phase 1",
+		CreatedAt:     time.Date(2026, 5, 4, 12, 0, 0, 0, time.UTC),
+		UpdatedAt:     time.Date(2026, 5, 4, 12, 0, 1, 0, time.UTC),
+		TTL:           5 * time.Minute,
+		PollInterval:  500 * time.Millisecond,
+	}
+	out := mcpTaskFromInternal(in)
+
+	if out.Server != "srv" || out.TaskID != "t-1" {
+		t.Errorf("identity fields not copied: %+v", out)
+	}
+	if out.Status != MCPTaskStatusWorking {
+		t.Errorf("Status = %q, want working", out.Status)
+	}
+	if out.StatusMessage != "phase 1" {
+		t.Errorf("StatusMessage = %q, want phase 1", out.StatusMessage)
+	}
+	if out.TTL != 5*time.Minute || out.PollInterval != 500*time.Millisecond {
+		t.Errorf("durations not copied: %+v", out)
+	}
+}
+
+func TestKitMCPTasksWithoutAgentReturnsError(t *testing.T) {
+	// A nil/zero Kit must not panic — task RPCs should surface a clear
+	// error instead. Useful for SDK consumers that try task ops on a Kit
+	// constructed without MCP servers.
+	var k *Kit
+	ctx := t.Context()
+	if _, err := k.ListMCPTasks(ctx, "any"); err == nil {
+		t.Error("ListMCPTasks on nil Kit should error")
+	}
+	if _, err := k.GetMCPTask(ctx, "any", "id"); err == nil {
+		t.Error("GetMCPTask on nil Kit should error")
+	}
+	if _, err := k.CancelMCPTask(ctx, "any", "id"); err == nil {
+		t.Error("CancelMCPTask on nil Kit should error")
+	}
+}
+
+func TestSubagentPropagatesMCPTaskOptions(t *testing.T) {
+	// Exercises the helper Kit.Subagent uses to copy MCP task options
+	// onto child Options. Calling the real helper (rather than
+	// duplicating its body in the test) means any new field added to
+	// the propagation list is picked up automatically by the
+	// equivalence assertion below.
+	parent := &Options{
+		MCPTaskMode: map[string]MCPTaskMode{
+			"build": MCPTaskModeAlways,
+			"chat":  MCPTaskModeNever,
+		},
+		MCPTaskTimeout:         30 * time.Minute,
+		MCPTaskTTL:             45 * time.Minute,
+		MCPTaskPollInterval:    750 * time.Millisecond,
+		MCPTaskMaxPollInterval: 4 * time.Second,
+		MCPTaskProgress:        func(MCPTaskProgress) {},
+	}
+
+	child := &Options{}
+	inheritMCPTaskOptions(child, parent)
+
+	if child.MCPTaskMode["build"] != MCPTaskModeAlways || child.MCPTaskMode["chat"] != MCPTaskModeNever {
+		t.Errorf("MCPTaskMode not propagated: got %+v", child.MCPTaskMode)
+	}
+	if child.MCPTaskTimeout != 30*time.Minute {
+		t.Errorf("MCPTaskTimeout = %v, want 30m", child.MCPTaskTimeout)
+	}
+	if child.MCPTaskTTL != 45*time.Minute {
+		t.Errorf("MCPTaskTTL = %v, want 45m", child.MCPTaskTTL)
+	}
+	if child.MCPTaskPollInterval != 750*time.Millisecond {
+		t.Errorf("MCPTaskPollInterval = %v, want 750ms", child.MCPTaskPollInterval)
+	}
+	if child.MCPTaskMaxPollInterval != 4*time.Second {
+		t.Errorf("MCPTaskMaxPollInterval = %v, want 4s", child.MCPTaskMaxPollInterval)
+	}
+	if child.MCPTaskProgress == nil {
+		t.Error("MCPTaskProgress not propagated")
+	}
+
+	// Nil parent is a no-op rather than a panic.
+	inheritMCPTaskOptions(&Options{}, nil)
+	inheritMCPTaskOptions(nil, parent)
+}

pkg/kit/oauth.go

@@ -5,18 +5,18 @@ import (
 	"fmt"
 	"net"
 	"net/http"
-	"os/exec"
-	"runtime"
 	"sync"
 	"time"
 )
 
 // MCPAuthHandler handles OAuth authorization for MCP servers.
 // Implementations control the user experience — opening a browser, showing a
-// prompt, displaying a URL, etc.
+// prompt, displaying a URL, posting to a message bus, etc.
 //
-// The default implementation ([DefaultMCPAuthHandler]) opens the system browser
-// and starts a local HTTP callback server to receive the authorization code.
+// [DefaultMCPAuthHandler] provides the transport mechanics (port reservation
+// and callback server) but performs no user-facing I/O on its own; consumers
+// wire presentation via [DefaultMCPAuthHandler.OnAuthURL] or implement
+// MCPAuthHandler from scratch.
 type MCPAuthHandler interface {
 	// RedirectURI returns the OAuth redirect URI that the callback server
 	// will listen on. This is called during MCP transport setup — before any
@@ -37,23 +37,44 @@ type MCPAuthHandler interface {
 	HandleAuth(ctx context.Context, serverName string, authURL string) (callbackURL string, err error)
 }
 
-// DefaultMCPAuthHandler opens the system browser and starts a local HTTP
-// callback server to receive the OAuth authorization code. It eagerly reserves
-// a TCP port on construction so [RedirectURI] is stable for the lifetime of
-// the handler.
+// DefaultMCPAuthHandler provides the transport mechanics of an OAuth flow —
+// reserving a local TCP port and running a one-shot HTTP callback server —
+// without making any user-experience decisions. It performs no browser opens,
+// no printing, no TUI calls; consumers attach presentation by setting
+// [DefaultMCPAuthHandler.OnAuthURL] or by wrapping the handler.
 //
-// Create instances with [NewDefaultMCPAuthHandler] (random port) or
-// [NewDefaultMCPAuthHandlerWithPort] (explicit port).
+// The handler eagerly reserves a TCP port on construction so [RedirectURI] is
+// stable for the lifetime of the handler. Create instances with
+// [NewDefaultMCPAuthHandler] (random port) or [NewDefaultMCPAuthHandlerWithPort]
+// (explicit port). Always call [DefaultMCPAuthHandler.Close] when done to
+// release the port.
 type DefaultMCPAuthHandler struct {
 	listener net.Listener
 	port     int
 	mu       sync.Mutex // guards listener lifecycle
+
+	// OnAuthURL, if set, is invoked exactly once per [HandleAuth] call with
+	// the authorization URL the user must visit. This is where consumers
+	// plug in their UX: open a browser, print to stderr, post to a TUI
+	// stream, render a QR code, etc. The handler performs no I/O on the
+	// URL itself; if OnAuthURL is nil the URL is silently dropped and the
+	// user has no way to complete the flow.
+	//
+	// OnAuthURL is called synchronously before the handler blocks on the
+	// callback. It must not block indefinitely — long-running work should
+	// be dispatched to a goroutine.
+	OnAuthURL func(serverName, authURL string)
 }
 
 // NewDefaultMCPAuthHandler creates a handler that listens on a random
 // available port on localhost. The port is reserved immediately so
 // [RedirectURI] returns a stable value. Call [DefaultMCPAuthHandler.Close]
 // when the handler is no longer needed to release the port.
+//
+// The returned handler has no OnAuthURL hook configured and will therefore
+// appear to hang on HandleAuth until the context deadline fires. Set
+// OnAuthURL before using the handler, or use a higher-level wrapper such
+// as [CLIMCPAuthHandler].
 func NewDefaultMCPAuthHandler() (*DefaultMCPAuthHandler, error) {
 	listener, err := net.Listen("tcp", "localhost:0")
 	if err != nil {
@@ -88,9 +109,9 @@ func (h *DefaultMCPAuthHandler) Port() int {
 	return h.port
 }
 
-// HandleAuth opens the system browser to authURL and waits for the OAuth
-// callback on the local server. It returns the full callback URL including
-// query parameters (code, state, etc.).
+// HandleAuth invokes [OnAuthURL] with the authorization URL (if configured)
+// and waits for the OAuth callback on the local server. It returns the full
+// callback URL including query parameters (code, state, etc.).
 //
 // If the context has no deadline, a default 2-minute timeout is applied.
 // The callback server is started for each HandleAuth call and shut down
@@ -136,19 +157,13 @@ func (h *DefaultMCPAuthHandler) HandleAuth(ctx context.Context, serverName strin
 		Handler: mux,
 	}
 
-	// Start serving on the pre-reserved listener. We need to create a new
-	// listener on the same port because http.Server.Serve takes ownership
-	// and closes the listener when done. The original listener is kept open
-	// to reserve the port; we create a second listener via SO_REUSEADDR
-	// semantics (Go's default on most platforms) or, more reliably, we
-	// temporarily release and re-acquire.
-	//
-	// Strategy: use the held listener directly for Serve. After Serve
-	// returns (due to Shutdown), re-acquire the listener to keep the port
-	// reserved for future HandleAuth calls.
+	// Start serving on the pre-reserved listener. http.Server.Serve takes
+	// ownership and closes the listener when Shutdown is called, so we
+	// re-acquire a fresh listener on the same port in the deferred cleanup
+	// below to keep the port reserved for subsequent HandleAuth calls.
 	h.mu.Lock()
 	serveListener := h.listener
-	h.listener = nil // Serve will close it
+	h.listener = nil
 	h.mu.Unlock()
 
 	if serveListener == nil {
@@ -184,10 +199,11 @@ func (h *DefaultMCPAuthHandler) HandleAuth(ctx context.Context, serverName strin
 		}
 	}()
 
-	// Open the system browser.
-	if err := openBrowser(authURL); err != nil {
-		// Browser open is best-effort; the user can still navigate manually.
-		_ = err
+	// Surface the authorization URL to the consumer. This is the single
+	// presentation seam: the SDK itself does not open browsers, print,
+	// or otherwise touch the user's environment.
+	if h.OnAuthURL != nil {
+		h.OnAuthURL(serverName, authURL)
 	}
 
 	// Wait for the callback, a server error, or context cancellation.
@@ -214,22 +230,6 @@ func (h *DefaultMCPAuthHandler) Close() error {
 	return nil
 }
 
-// openBrowser opens the default system browser to the given URL. This is a
-// best-effort operation — errors are returned but callers typically ignore
-// them since the user can navigate manually.
-func openBrowser(url string) error {
-	switch runtime.GOOS {
-	case "linux":
-		return exec.Command("xdg-open", url).Start()
-	case "windows":
-		return exec.Command("rundll32", "url.dll,FileProtocolHandler", url).Start()
-	case "darwin":
-		return exec.Command("open", url).Start()
-	default:
-		return fmt.Errorf("unsupported platform: %s", runtime.GOOS)
-	}
-}
-
 // oauthSuccessHTML is the HTML page returned to the browser after a
 // successful OAuth callback.
 const oauthSuccessHTML = `<!DOCTYPE html>

pkg/kit/oauth_cli.go

@@ -5,32 +5,49 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"os/exec"
+	"runtime"
 )
 
-// CLIMCPAuthHandler wraps a [DefaultMCPAuthHandler] and prints status messages
-// to a writer (typically stderr) so the user knows what's happening during
-// OAuth authorization. This is the handler used by the CLI/TUI binary.
+// CLIMCPAuthHandler is the MCP OAuth handler for CLI/TUI consumers. It wraps
+// a [DefaultMCPAuthHandler] and layers standard CLI behavior on top of the
+// underlying transport mechanics:
 //
-// For TUI integration, set NotifyFunc to route messages through the TUI's
-// event system instead of (or in addition to) the writer.
+//   - Opens the authorization URL in the system browser
+//   - Prints status messages (or routes them to a TUI via [NotifyFunc])
+//
+// Non-CLI consumers (web apps, daemons, custom TUIs) should not use this
+// handler; implement [MCPAuthHandler] directly or configure a
+// [DefaultMCPAuthHandler] with a custom OnAuthURL instead.
 type CLIMCPAuthHandler struct {
 	inner *DefaultMCPAuthHandler
 	w     io.Writer
 
-	// NotifyFunc, when set, is called with status messages instead of writing
-	// to the writer. This allows the TUI to display system messages in the
-	// chat stream. If nil, messages are written to w.
+	// NotifyFunc, when set, is called with status messages instead of
+	// writing to the writer. This allows the TUI to display system
+	// messages in the chat stream. If nil, messages are written to w.
 	NotifyFunc func(serverName, message string)
 }
 
 // NewCLIMCPAuthHandler creates a CLI auth handler that prints status messages
-// to stderr and delegates the actual OAuth flow to a [DefaultMCPAuthHandler].
+// to stderr, opens the authorization URL in the system browser, and delegates
+// the callback-server mechanics to a [DefaultMCPAuthHandler].
 func NewCLIMCPAuthHandler() (*CLIMCPAuthHandler, error) {
 	inner, err := NewDefaultMCPAuthHandler()
 	if err != nil {
 		return nil, err
 	}
-	return &CLIMCPAuthHandler{inner: inner, w: os.Stderr}, nil
+	h := &CLIMCPAuthHandler{inner: inner, w: os.Stderr}
+	// Wire the CLI presentation policy into the inner handler's hook.
+	// This is the one place in the codebase where OAuth triggers a
+	// browser open; the SDK core remains I/O-free.
+	inner.OnAuthURL = func(serverName, authURL string) {
+		h.notify(serverName, fmt.Sprintf("🔐 MCP server %q requires authentication. Opening browser...", serverName))
+		h.notify(serverName, fmt.Sprintf("   If the browser doesn't open, visit:\n   %s", authURL))
+		// Browser open is best-effort; the user can still navigate manually.
+		_ = openBrowser(authURL)
+	}
+	return h, nil
 }
 
 // RedirectURI returns the OAuth redirect URI from the inner handler.
@@ -38,17 +55,15 @@ func (h *CLIMCPAuthHandler) RedirectURI() string {
 	return h.inner.RedirectURI()
 }
 
-// HandleAuth prints status messages and delegates to the inner handler.
+// HandleAuth delegates to the inner handler (which invokes OnAuthURL, runs
+// the callback server, and returns the full callback URL) and emits a final
+// success or failure notification.
 func (h *CLIMCPAuthHandler) HandleAuth(ctx context.Context, serverName string, authURL string) (string, error) {
-	h.notify(serverName, fmt.Sprintf("🔐 MCP server %q requires authentication. Opening browser...", serverName))
-	h.notify(serverName, fmt.Sprintf("   If the browser doesn't open, visit:\n   %s", authURL))
-
 	callbackURL, err := h.inner.HandleAuth(ctx, serverName, authURL)
 	if err != nil {
 		h.notify(serverName, fmt.Sprintf("✗ Authentication failed for %q: %v", serverName, err))
 		return "", err
 	}
-
 	h.notify(serverName, fmt.Sprintf("✓ Authenticated with %q", serverName))
 	return callbackURL, nil
 }
@@ -66,3 +81,20 @@ func (h *CLIMCPAuthHandler) notify(serverName, message string) {
 	}
 	_, _ = fmt.Fprintln(h.w, message)
 }
+
+// openBrowser opens the system default browser at url. Intentionally
+// unexported: browser opening is CLI policy, not SDK surface. Consumers
+// that need similar behavior for their own UX should bring their own
+// helper (or use a third-party package like github.com/pkg/browser).
+func openBrowser(url string) error {
+	switch runtime.GOOS {
+	case "linux":
+		return exec.Command("xdg-open", url).Start()
+	case "windows":
+		return exec.Command("rundll32", "url.dll,FileProtocolHandler", url).Start()
+	case "darwin":
+		return exec.Command("open", url).Start()
+	default:
+		return fmt.Errorf("unsupported platform: %s", runtime.GOOS)
+	}
+}

pkg/kit/tools.go

@@ -2,6 +2,7 @@ package kit
 
 import (
 	"context"
+	"strings"
 
 	"charm.land/fantasy"
 
@@ -52,6 +53,22 @@ func ErrorResult(content string) ToolOutput {
 	return ToolOutput{Content: content, IsError: true}
 }
 
+// ImageResult creates a [ToolOutput] that returns an image to the LLM.
+// The data is the raw image bytes and mediaType is the MIME type
+// (e.g. "image/png", "image/jpeg"). The optional text content accompanies
+// the image and is visible to the LLM alongside it.
+func ImageResult(content string, data []byte, mediaType string) ToolOutput {
+	return ToolOutput{Content: content, Data: data, MediaType: mediaType}
+}
+
+// MediaResult creates a [ToolOutput] that returns non-image binary media
+// (e.g. audio, video) to the LLM. The data is the raw bytes and mediaType
+// is the MIME type (e.g. "audio/wav", "video/mp4"). The optional text
+// content accompanies the media.
+func MediaResult(content string, data []byte, mediaType string) ToolOutput {
+	return ToolOutput{Content: content, Data: data, MediaType: mediaType}
+}
+
 // toolCallIDKey is the context key for the tool call ID.
 type toolCallIDKey struct{}
 
@@ -63,9 +80,35 @@ func ToolCallIDFromContext(ctx context.Context) string {
 	return s
 }
 
+// toolOutputToResponse converts a [ToolOutput] into the underlying
+// framework's ToolResponse, inferring the response Type from Data/MediaType
+// so that binary content (images, audio, etc.) is forwarded to the LLM
+// instead of being silently dropped.
+func toolOutputToResponse(result ToolOutput) fantasy.ToolResponse {
+	resp := fantasy.ToolResponse{
+		Content:   result.Content,
+		IsError:   result.IsError,
+		Data:      result.Data,
+		MediaType: result.MediaType,
+	}
+	// Infer response type from binary data so the downstream framework
+	// creates a media content block instead of a plain-text one.
+	if len(result.Data) > 0 && result.MediaType != "" {
+		if strings.HasPrefix(result.MediaType, "image/") {
+			resp.Type = "image"
+		} else {
+			resp.Type = "media"
+		}
+	}
+	if result.Metadata != nil {
+		resp = fantasy.WithResponseMetadata(resp, result.Metadata)
+	}
+	return resp
+}
+
 // NewTool creates a custom [Tool] with automatic JSON schema generation from
 // the TInput struct type. The handler receives a typed input (deserialized
-// from the LLM's JSON arguments) and returns a [ToolResult].
+// from the LLM's JSON arguments) and returns a [ToolOutput].
 //
 // Struct tags on TInput control the generated schema:
 //
@@ -77,6 +120,11 @@ func ToolCallIDFromContext(ctx context.Context) string {
 // The tool call ID is injected into the context and can be retrieved with
 // [ToolCallIDFromContext].
 //
+// Binary results: When [ToolOutput.Data] and [ToolOutput.MediaType] are set,
+// the response type is automatically inferred so the LLM receives the binary
+// content (e.g. an image) instead of only the text. Use [ImageResult] or
+// [MediaResult] for convenience.
+//
 // Example:
 //
 //	type WeatherInput struct {
@@ -84,7 +132,7 @@ func ToolCallIDFromContext(ctx context.Context) string {
 //	}
 //
 //	tool := kit.NewTool("get_weather", "Get weather for a city",
-//	    func(ctx context.Context, input WeatherInput) (kit.ToolResult, error) {
+//	    func(ctx context.Context, input WeatherInput) (kit.ToolOutput, error) {
 //	        return kit.TextResult("72°F, sunny in " + input.City), nil
 //	    },
 //	)
@@ -96,16 +144,7 @@ func NewTool[TInput any](name, description string, fn func(ctx context.Context,
 			if err != nil {
 				return fantasy.NewTextErrorResponse(err.Error()), nil
 			}
-			resp := fantasy.ToolResponse{
-				Content:   result.Content,
-				IsError:   result.IsError,
-				Data:      result.Data,
-				MediaType: result.MediaType,
-			}
-			if result.Metadata != nil {
-				resp = fantasy.WithResponseMetadata(resp, result.Metadata)
-			}
-			return resp, nil
+			return toolOutputToResponse(result), nil
 		},
 	)
 }
@@ -121,16 +160,7 @@ func NewParallelTool[TInput any](name, description string, fn func(ctx context.C
 			if err != nil {
 				return fantasy.NewTextErrorResponse(err.Error()), nil
 			}
-			resp := fantasy.ToolResponse{
-				Content:   result.Content,
-				IsError:   result.IsError,
-				Data:      result.Data,
-				MediaType: result.MediaType,
-			}
-			if result.Metadata != nil {
-				resp = fantasy.WithResponseMetadata(resp, result.Metadata)
-			}
-			return resp, nil
+			return toolOutputToResponse(result), nil
 		},
 	)
 }

pkg/kit/tools_test.go

@@ -117,3 +117,149 @@ func TestToolOutput_BinaryData(t *testing.T) {
 		t.Errorf("MediaType = %q, want %q", r.MediaType, "image/png")
 	}
 }
+
+// TestImageResult verifies the ImageResult convenience constructor.
+func TestImageResult(t *testing.T) {
+	data := []byte{0x89, 0x50, 0x4E, 0x47}
+	r := kit.ImageResult("here is the image", data, "image/png")
+	if r.Content != "here is the image" {
+		t.Errorf("Content = %q, want %q", r.Content, "here is the image")
+	}
+	if len(r.Data) != 4 {
+		t.Errorf("Data len = %d, want 4", len(r.Data))
+	}
+	if r.MediaType != "image/png" {
+		t.Errorf("MediaType = %q, want %q", r.MediaType, "image/png")
+	}
+	if r.IsError {
+		t.Error("ImageResult should not set IsError")
+	}
+}
+
+// TestMediaResult verifies the MediaResult convenience constructor.
+func TestMediaResult(t *testing.T) {
+	data := []byte{0xFF, 0xFB, 0x90, 0x00}
+	r := kit.MediaResult("audio clip", data, "audio/mpeg")
+	if r.Content != "audio clip" {
+		t.Errorf("Content = %q, want %q", r.Content, "audio clip")
+	}
+	if len(r.Data) != 4 {
+		t.Errorf("Data len = %d, want 4", len(r.Data))
+	}
+	if r.MediaType != "audio/mpeg" {
+		t.Errorf("MediaType = %q, want %q", r.MediaType, "audio/mpeg")
+	}
+	if r.IsError {
+		t.Error("MediaResult should not set IsError")
+	}
+}
+
+// TestNewTool_BinaryImageResponse verifies that NewTool correctly infers the
+// response type for image data so binary content is forwarded to the LLM
+// (issue #17).
+func TestNewTool_BinaryImageResponse(t *testing.T) {
+	type Input struct {
+		Path string `json:"path"`
+	}
+
+	imgData := []byte{0x89, 0x50, 0x4E, 0x47} // PNG magic bytes
+
+	tool := kit.NewTool("read_image", "Read an image file",
+		func(ctx context.Context, input Input) (kit.ToolOutput, error) {
+			return kit.ImageResult("Here is the image", imgData, "image/png"), nil
+		},
+	)
+
+	// Run the tool and inspect the raw ToolResponse via the AgentTool interface.
+	resp, err := tool.Run(context.Background(), kit.LLMToolCall{
+		ID:    "call_1",
+		Name:  "read_image",
+		Input: `{"path": "test.png"}`,
+	})
+	if err != nil {
+		t.Fatalf("Run() error: %v", err)
+	}
+
+	// The Type field must be "image" so the downstream framework creates a
+	// media content block instead of discarding the binary data.
+	if resp.Type != "image" {
+		t.Errorf("ToolResponse.Type = %q, want %q", resp.Type, "image")
+	}
+	if len(resp.Data) != 4 {
+		t.Errorf("ToolResponse.Data len = %d, want 4", len(resp.Data))
+	}
+	if resp.MediaType != "image/png" {
+		t.Errorf("ToolResponse.MediaType = %q, want %q", resp.MediaType, "image/png")
+	}
+	if resp.Content != "Here is the image" {
+		t.Errorf("ToolResponse.Content = %q, want %q", resp.Content, "Here is the image")
+	}
+}
+
+// TestNewTool_BinaryMediaResponse verifies type inference for non-image media.
+func TestNewTool_BinaryMediaResponse(t *testing.T) {
+	type Input struct{}
+
+	tool := kit.NewTool("get_audio", "Get audio",
+		func(ctx context.Context, input Input) (kit.ToolOutput, error) {
+			return kit.MediaResult("audio clip", []byte{0xFF, 0xFB}, "audio/mpeg"), nil
+		},
+	)
+
+	resp, err := tool.Run(context.Background(), kit.LLMToolCall{
+		ID:    "call_2",
+		Name:  "get_audio",
+		Input: `{}`,
+	})
+	if err != nil {
+		t.Fatalf("Run() error: %v", err)
+	}
+	if resp.Type != "media" {
+		t.Errorf("ToolResponse.Type = %q, want %q", resp.Type, "media")
+	}
+}
+
+// TestNewTool_TextResponseTypeNotSet verifies that text-only responses do NOT
+// get an inferred type (preserving existing behavior).
+func TestNewTool_TextResponseTypeNotSet(t *testing.T) {
+	type Input struct{}
+
+	tool := kit.NewTool("echo", "Echo",
+		func(ctx context.Context, input Input) (kit.ToolOutput, error) {
+			return kit.TextResult("hello"), nil
+		},
+	)
+
+	resp, err := tool.Run(context.Background(), kit.LLMToolCall{
+		ID: "call_3", Name: "echo", Input: `{}`,
+	})
+	if err != nil {
+		t.Fatalf("Run() error: %v", err)
+	}
+	// Text responses should not have Type set (the framework treats "" as text).
+	if resp.Type != "" {
+		t.Errorf("ToolResponse.Type = %q, want empty string for text responses", resp.Type)
+	}
+}
+
+// TestNewParallelTool_BinaryImageResponse mirrors the NewTool binary test for
+// NewParallelTool.
+func TestNewParallelTool_BinaryImageResponse(t *testing.T) {
+	type Input struct{}
+
+	tool := kit.NewParallelTool("snap", "Take a snapshot",
+		func(ctx context.Context, input Input) (kit.ToolOutput, error) {
+			return kit.ImageResult("snapshot", []byte{0xFF, 0xD8}, "image/jpeg"), nil
+		},
+	)
+
+	resp, err := tool.Run(context.Background(), kit.LLMToolCall{
+		ID: "call_4", Name: "snap", Input: `{}`,
+	})
+	if err != nil {
+		t.Fatalf("Run() error: %v", err)
+	}
+	if resp.Type != "image" {
+		t.Errorf("ToolResponse.Type = %q, want %q", resp.Type, "image")
+	}
+}

pkg/kit/types.go

@@ -157,6 +157,18 @@ type LLMTextPart = fantasy.TextPart
 // LLMReasoningPart is a reasoning/chain-of-thought content part.
 type LLMReasoningPart = fantasy.ReasoningPart
 
+// LLMToolCall represents the raw tool invocation passed to a [Tool]'s Run
+// method. It carries the call ID, tool name, and the JSON-encoded input
+// arguments from the LLM. This is the execution-layer call object — distinct
+// from [ToolCall] (a message content part).
+type LLMToolCall = fantasy.ToolCall
+
+// LLMToolResponse represents the raw response returned from a [Tool]'s Run
+// method. Most SDK consumers should use [ToolOutput] with [NewTool] /
+// [NewParallelTool] instead — this alias is provided for advanced use cases
+// that need to call Tool.Run() directly (e.g. testing).
+type LLMToolResponse = fantasy.ToolResponse
+
 // LLMToolCallPart represents an LLM-initiated tool invocation within a message.
 type LLMToolCallPart = fantasy.ToolCallPart
 
@@ -172,6 +184,40 @@ type LLMToolResultOutputContentText = fantasy.ToolResultOutputContentText
 // LLMToolResultOutputContentError is an error-valued tool result output.
 type LLMToolResultOutputContentError = fantasy.ToolResultOutputContentError
 
+// LLMToolResultOutputContentMedia is a media-valued tool result output
+// (images, audio, etc.) carrying base64-encoded data and a MIME type.
+type LLMToolResultOutputContentMedia = fantasy.ToolResultOutputContentMedia
+
+// LLMToolResultContentType classifies the kind of a tool result output
+// ("text", "error", or "media").
+type LLMToolResultContentType = fantasy.ToolResultContentType
+
+// Tool result content type constants.
+const (
+	// LLMToolResultContentTypeText represents text output.
+	LLMToolResultContentTypeText = fantasy.ToolResultContentTypeText
+	// LLMToolResultContentTypeError represents error text output.
+	LLMToolResultContentTypeError = fantasy.ToolResultContentTypeError
+	// LLMToolResultContentTypeMedia represents media (binary) output.
+	LLMToolResultContentTypeMedia = fantasy.ToolResultContentTypeMedia
+)
+
+// LLMToolInfo describes a tool's name, description, and JSON-Schema parameters.
+type LLMToolInfo = fantasy.ToolInfo
+
+// LLMProviderOptions carries provider-specific key/value option maps, keyed
+// by provider name (e.g. "anthropic"). Use this when configuring or
+// inspecting provider-specific tool behaviour.
+type LLMProviderOptions = fantasy.ProviderOptions
+
+// LLMProviderMetadata carries provider-specific metadata returned alongside
+// LLM responses, keyed by provider name.
+type LLMProviderMetadata = fantasy.ProviderMetadata
+
+// LLMPrompt is an ordered sequence of [LLMMessage] values forming a complete
+// prompt for the LLM.
+type LLMPrompt = fantasy.Prompt
+
 // LLMMessageRole identifies the participant role in an LLM conversation.
 type LLMMessageRole = fantasy.MessageRole
 
@@ -198,6 +244,12 @@ var NewLLMUserMessage = fantasy.NewUserMessage
 // prompt strings.
 var NewLLMSystemMessage = fantasy.NewSystemMessage
 
+// newLLMTextErrorResponse creates a tool-error response (internal helper).
+var newLLMTextErrorResponse = fantasy.NewTextErrorResponse
+
+// newLLMTextResponse creates a plain-text tool response (internal helper).
+var newLLMTextResponse = fantasy.NewTextResponse
+
 // ==== Compaction Types (internal/compaction/) ====
 
 // CompactionResult contains statistics from a compaction operation.

scripts/acp_smoke_test.py

@@ -3,10 +3,13 @@
 ACP smoke test — drives `kit acp` over JSON-RPC 2.0 stdio.
 
 Protocol flow:
-  1. session/new  → get sessionId
-  2. session/set_model → set opencode/kimi-k2.5
-  3. session/prompt → "What is 2+2? Answer in one sentence."
-  4. Collect session updates until done
+  1. initialize        → negotiate capabilities
+  2. session/new       → get sessionId
+  3. session/list      → verify session listing works
+  4. session/set_config_option → set model
+  5. session/prompt    → "What is 2+2? Answer in one sentence."
+  6. Collect session/update notifications until prompt response
+  7. session/cancel    → verify cancel is accepted (no-op since prompt is done)
 """
 
 import json
@@ -21,9 +24,24 @@ MODEL   = os.environ.get("MODEL", "opencode/kimi-k2.5")
 CWD     = os.path.expanduser("~")
 TIMEOUT = 60  # seconds to wait for the prompt to complete
 
+# Request ID counter — initialize=1, session/new=2, etc.
+_next_id = 0
 
-def rpc(method, params, req_id):
-    return json.dumps({"jsonrpc": "2.0", "id": req_id, "method": method, "params": params}) + "\n"
+
+def next_id():
+    global _next_id
+    _next_id += 1
+    return _next_id
+
+
+def rpc_request(method, params):
+    """Build a JSON-RPC 2.0 request with auto-incrementing ID."""
+    return json.dumps({"jsonrpc": "2.0", "id": next_id(), "method": method, "params": params}) + "\n"
+
+
+def rpc_notification(method, params):
+    """Build a JSON-RPC 2.0 notification (no id)."""
+    return json.dumps({"jsonrpc": "2.0", "method": method, "params": params}) + "\n"
 
 
 def send(proc, line):
@@ -32,7 +50,7 @@ def send(proc, line):
     proc.stdin.flush()
 
 
-def read_responses(proc, collected, done_event):
+def read_responses(proc, collected, done_event, prompt_id):
     """Read newline-delimited JSON from stdout until process exits."""
     for raw in proc.stdout:
         raw = raw.strip()
@@ -50,32 +68,49 @@ def read_responses(proc, collected, done_event):
         if "result" in msg:
             result = msg["result"]
             print(f"← RESP  id={msg.get('id')}  result={json.dumps(result)[:200]}", flush=True)
-            # Prompt complete when we get a stopReason on id=3
-            if msg.get("id") == 3 and "stopReason" in result:
+            # Prompt complete when we get a stopReason on the prompt request ID
+            if msg.get("id") == prompt_id and "stopReason" in result:
                 done_event.set()
         elif "error" in msg:
             print(f"← ERROR id={msg.get('id')}  {json.dumps(msg['error'])}", flush=True)
             # If it's the prompt call that errored, unblock
-            if msg.get("id") == 3:
+            if msg.get("id") == prompt_id:
                 done_event.set()
         elif "method" in msg:
             # Notification / session update
             m = msg.get("method", "")
             p = msg.get("params", {})
-            if m in ("session/update", "session/updated"):
+            if m == "session/update":
                 update = p.get("update", {})
-                stype = update.get("sessionUpdate") or update.get("type", "?")
+                stype = update.get("sessionUpdate", "?")
                 content = update.get("content", {})
+                text = content.get("text", "")
                 if stype == "agent_thought_chunk":
-                    print(f"  [thinking] {content.get('text','')}", end="", flush=True)
+                    print(f"  [thinking] {text}", end="", flush=True)
                 elif stype == "agent_message_chunk":
-                    print(f"  [response] {content.get('text','')}", end="", flush=True)
+                    print(f"  [response] {text}", end="", flush=True)
+                elif stype in ("tool_call", "tool_call_update"):
+                    title = update.get("title", update.get("toolCallId", "?"))
+                    status = update.get("status", "?")
+                    print(f"\n  [{stype}] {title} ({status})", flush=True)
                 else:
                     print(f"\n  [update/{stype}] {json.dumps(update)[:200]}", flush=True)
             else:
                 print(f"\n← NOTIF {m}  {json.dumps(p)[:200]}", flush=True)
 
 
+def wait_for_response(collected, req_id, timeout=5.0, label="response"):
+    """Block until we have a response for the given request ID."""
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        for msg in collected:
+            if msg.get("id") == req_id and ("result" in msg or "error" in msg):
+                return msg
+        time.sleep(0.1)
+    print(f"\n✗ FAIL: timed out waiting for {label} (id={req_id})", flush=True)
+    return None
+
+
 def main():
     print(f"Starting: {KIT_BIN} acp -m {MODEL}", flush=True)
 
@@ -91,8 +126,13 @@ def main():
     collected = []
     done_event = threading.Event()
 
-    reader = threading.Thread(target=read_responses, args=(proc, collected, done_event), daemon=True)
-    reader.start()
+    # We'll set the prompt_id once we know it
+    prompt_id_holder = [None]
+
+    # Start reader thread — prompt_id will be set before prompt is sent
+    class ReaderThread(threading.Thread):
+        def run(self):
+            read_responses(proc, collected, done_event, prompt_id_holder[0])
 
     stderr_lines = []
     def read_stderr():
@@ -105,16 +145,55 @@ def main():
 
     time.sleep(0.3)  # let the process initialise
 
-    # 1. session/new
-    send(proc, rpc("session/new", {"cwd": CWD, "mcpServers": []}, 1))
+    # ── Step 1: initialize ──────────────────────────────────────────────
+    init_id = next_id()
+    send(proc, json.dumps({
+        "jsonrpc": "2.0",
+        "id": init_id,
+        "method": "initialize",
+        "params": {
+            "protocolVersion": 1,
+            "clientCapabilities": {
+                "fs": {"readTextFile": False, "writeTextFile": False},
+            },
+            "clientInfo": {"name": "acp-smoke-test", "version": "1.0.0"},
+        },
+    }) + "\n")
+
+    # Start a simple reader for the initialize response
+    reader = threading.Thread(target=read_responses, args=(proc, collected, done_event, None), daemon=True)
+    reader.start()
+
     time.sleep(1.0)
 
-    session_id = None
-    for msg in collected:
-        if msg.get("id") == 1 and "result" in msg:
-            session_id = msg["result"].get("sessionId")
-            break
+    init_resp = wait_for_response(collected, init_id, timeout=5, label="initialize")
+    if not init_resp or "error" in init_resp:
+        print(f"\n✗ FAIL: initialize failed: {init_resp}", flush=True)
+        proc.terminate()
+        sys.exit(1)
 
+    result = init_resp["result"]
+    proto_ver = result.get("protocolVersion")
+    agent_info = result.get("agentInfo", {})
+    print(f"\n✓ Initialized: protocol_version={proto_ver} agent={agent_info.get('name', '?')} v{agent_info.get('version', '?')}", flush=True)
+
+    # ── Step 2: session/new ─────────────────────────────────────────────
+    new_session_id = next_id()
+    send(proc, json.dumps({
+        "jsonrpc": "2.0",
+        "id": new_session_id,
+        "method": "session/new",
+        "params": {"cwd": CWD, "mcpServers": []},
+    }) + "\n")
+    time.sleep(1.0)
+
+    session_resp = wait_for_response(collected, new_session_id, timeout=10, label="session/new")
+    if not session_resp or "error" in session_resp:
+        print(f"\n✗ FAIL: session/new failed: {session_resp}", flush=True)
+        proc.terminate()
+        sys.exit(1)
+
+    session_id = session_resp["result"].get("sessionId")
     if not session_id:
         print("\n✗ FAIL: did not get sessionId from session/new", flush=True)
         proc.terminate()
@@ -122,31 +201,102 @@ def main():
 
     print(f"\n✓ Got sessionId: {session_id}", flush=True)
 
-    # 2. session/set_model (model already set via -m flag, but exercise the RPC)
-    send(proc, rpc("session/set_model", {"sessionId": session_id, "modelId": MODEL}, 2))
+    # ── Step 3: session/list ────────────────────────────────────────────
+    list_id = next_id()
+    send(proc, json.dumps({
+        "jsonrpc": "2.0",
+        "id": list_id,
+        "method": "session/list",
+        "params": {},
+    }) + "\n")
     time.sleep(0.5)
 
-    # 3. session/prompt
+    list_resp = wait_for_response(collected, list_id, timeout=5, label="session/list")
+    if not list_resp:
+        print("\n⚠ WARN: session/list timed out (non-fatal)", flush=True)
+    elif "error" in list_resp:
+        print(f"\n⚠ WARN: session/list returned error: {list_resp['error']} (non-fatal)", flush=True)
+    else:
+        sessions = list_resp["result"].get("sessions", [])
+        print(f"\n✓ session/list returned {len(sessions)} session(s)", flush=True)
+
+    # ── Step 4: session/set_config_option (model) ───────────────────────
+    # Uses the new session/set_config_option method (replaces the old session/set_model).
+    # The model is already set via -m flag, but we exercise the RPC to verify it works.
+    config_id = next_id()
+    send(proc, json.dumps({
+        "jsonrpc": "2.0",
+        "id": config_id,
+        "method": "session/set_config_option",
+        "params": {
+            "sessionId": session_id,
+            "configId": "model",
+            "value": MODEL,
+        },
+    }) + "\n")
+    time.sleep(0.5)
+
+    config_resp = wait_for_response(collected, config_id, timeout=5, label="session/set_config_option")
+    if not config_resp:
+        print("\n⚠ WARN: session/set_config_option timed out (non-fatal)", flush=True)
+    elif "error" in config_resp:
+        print(f"\n⚠ WARN: session/set_config_option returned error: {config_resp['error']} (non-fatal)", flush=True)
+    else:
+        print(f"\n✓ session/set_config_option accepted", flush=True)
+
+    # ── Step 5: session/prompt ──────────────────────────────────────────
+    prompt_id = next_id()
+    prompt_id_holder[0] = prompt_id
+
+    # Re-wire the reader to know the prompt ID (the existing thread is already running)
+    # Since we can't change it mid-flight easily, we check the collected list instead.
+
     prompt_params = {
         "sessionId": session_id,
         "prompt": [{"type": "text", "text": "What is 2+2? Answer in one sentence."}],
     }
-    send(proc, rpc("session/prompt", prompt_params, 3))
+    send(proc, json.dumps({
+        "jsonrpc": "2.0",
+        "id": prompt_id,
+        "method": "session/prompt",
+        "params": prompt_params,
+    }) + "\n")
 
-    # Wait for finished update or timeout
-    if not done_event.wait(timeout=TIMEOUT):
-        print(f"\n✗ FAIL: timed out after {TIMEOUT}s waiting for finished update", flush=True)
+    # Wait for finished update or timeout — poll collected list
+    deadline = time.time() + TIMEOUT
+    prompt_resp = None
+    while time.time() < deadline:
+        for msg in collected:
+            if msg.get("id") == prompt_id and ("result" in msg or "error" in msg):
+                prompt_resp = msg
+                break
+        if prompt_resp:
+            break
+        time.sleep(0.2)
+
+    if not prompt_resp:
+        print(f"\n✗ FAIL: timed out after {TIMEOUT}s waiting for prompt response", flush=True)
         proc.terminate()
         sys.exit(1)
 
-    # Check we got a successful prompt response
-    prompt_resp = next((m for m in collected if m.get("id") == 3), None)
-    if prompt_resp and "error" in prompt_resp:
+    if "error" in prompt_resp:
         print(f"\n✗ FAIL: prompt returned error: {prompt_resp['error']}", flush=True)
         proc.terminate()
         sys.exit(1)
 
-    print("\n✓ SMOKE TEST PASSED", flush=True)
+    stop_reason = prompt_resp["result"].get("stopReason", "?")
+    print(f"\n✓ Prompt completed: stopReason={stop_reason}", flush=True)
+
+    # ── Step 6: session/cancel (no-op, prompt already done) ─────────────
+    # This is a notification (no id), so no response expected.
+    send(proc, rpc_notification("session/cancel", {"sessionId": session_id}))
+    time.sleep(0.3)
+    print("✓ session/cancel sent (no-op)", flush=True)
+
+    # ── Summary ─────────────────────────────────────────────────────────
+    # Count session updates received
+    update_count = sum(1 for m in collected if m.get("method") == "session/update")
+    print(f"\n✓ SMOKE TEST PASSED  ({update_count} session updates received)", flush=True)
     proc.terminate()
     proc.wait(timeout=5)
 

skills/kit-extensions/SKILL.md

@@ -55,7 +55,7 @@ The `Init` function receives an `ext.API` object for registering handlers, and e
 
 ## Lifecycle Events
 
-Kit provides 18 lifecycle events. Each handler receives an event struct and a `Context`.
+Kit provides 21 lifecycle events. Each handler receives an event struct and a `Context`.
 
 ### Session Events
 
@@ -93,7 +93,7 @@ api.OnAgentEnd(func(e ext.AgentEndEvent, ctx ext.Context) {
     // e.Response string
     // e.StopReason string — "error" (on failure), "completed" (when LLM returns
     //   empty stop reason), or the raw LLM provider value passed through
-    //   (e.g. "stop", "end_turn", "max_tokens", "tool_use").
+    //   (e.g. "stop", "length" (max output tokens hit), "tool-calls", "content-filter").
     //   To detect errors, check e.StopReason == "error".
     //   Do NOT compare against "completed" for success — instead check != "error".
 })
@@ -136,6 +136,37 @@ api.OnToolResult(func(e ext.ToolResultEvent, ctx ext.Context) *ext.ToolResultRes
 })
 ```
 
+### Tool Call Input Streaming Events
+
+These events fire during the LLM's tool argument generation phase, **before** the tool call is fully parsed and before `OnToolCall` fires. They enable UIs to show tool activity immediately rather than waiting for the full argument JSON to finish streaming.
+
+```go
+// Fires when the LLM begins generating tool call arguments.
+// The tool name is known but the full argument JSON is still streaming.
+api.OnToolCallInputStart(func(e ext.ToolCallInputStartEvent, ctx ext.Context) {
+    // e.ToolCallID string — stable ID for correlating tool lifecycle events
+    // e.ToolName string — name of the tool being called
+    // e.ToolKind string — "execute", "edit", "read", "search", "agent"
+    ctx.PrintInfo("Tool starting: " + e.ToolName)
+})
+
+// Fires for each streamed fragment of tool call arguments.
+// Useful for live-previewing artifact content or showing a progress indicator.
+api.OnToolCallInputDelta(func(e ext.ToolCallInputDeltaEvent, ctx ext.Context) {
+    // e.ToolCallID string
+    // e.Delta string — JSON fragment of tool arguments
+})
+
+// Fires when tool argument streaming is complete, before the tool call
+// is parsed and execution begins. Transition UI from "generating args"
+// to "executing".
+api.OnToolCallInputEnd(func(e ext.ToolCallInputEndEvent, ctx ext.Context) {
+    // e.ToolCallID string
+})
+```
+
+**Full tool lifecycle order**: `OnToolCallInputStart` → `OnToolCallInputDelta` (repeated) → `OnToolCallInputEnd` → `OnToolCall` → `OnToolExecutionStart` → `OnToolOutput` (optional, repeated) → `OnToolExecutionEnd` → `OnToolResult`
+
 ### Input Events
 
 ```go

skills/kit-sdk/SKILL.md

@@ -80,6 +80,23 @@ host, err := kit.New(ctx, &kit.Options{
     Quiet:     true, // suppress debug output
     Debug:     true, // enable debug logging
 
+    // Generation parameters — override env/config/per-model defaults.
+    // Leaving a field at its zero/nil value lets the precedence chain
+    // resolve a value (KIT_* env → .kit.yml → modelSettings/customModels →
+    // 8192 floor for MaxTokens, provider defaults for samplers).
+    MaxTokens:        16384,             // 0 = auto-resolve; non-zero suppresses right-sizing
+    ThinkingLevel:    "medium",          // "off", "none", "minimal", "low", "medium", "high" ("" = default)
+    Temperature:      ptrFloat32(0.2),   // pointer so explicit 0.0 != unset
+    TopP:             nil,                // nil = leave provider/per-model default
+    TopK:             nil,                // nil = leave provider/per-model default
+    FrequencyPenalty: nil,
+    PresencePenalty:  nil,
+
+    // Provider configuration — override env/config without viper.Set workarounds.
+    ProviderAPIKey: "sk-...",                    // "" = use config / provider env var
+    ProviderURL:    "https://proxy.internal/v1", // "" = provider default endpoint
+    TLSSkipVerify:  false,                       // true only; can't force-disable via Options
+
     // Session
     SessionDir:  "/path/to/project",  // base dir for session discovery (default: cwd)
     SessionPath: "/path/to/session.jsonl", // open specific session file
@@ -108,7 +125,12 @@ host, err := kit.New(ctx, &kit.Options{
     AutoCompact:       true,                        // auto-compact near context limit
     CompactionOptions: &kit.CompactionOptions{...}, // nil = defaults
 
-    // MCP OAuth
+    // MCP OAuth — both fields are opt-in. If MCPAuthHandler is nil,
+    // remote MCP servers that require OAuth will fail to connect with
+    // an authorization-required error instead of silently opening a
+    // browser. CLI consumers use NewCLIMCPAuthHandler; other embedders
+    // implement MCPAuthHandler or configure DefaultMCPAuthHandler.
+    MCPAuthHandler: mcpAuthHandler,             // nil = OAuth disabled
     MCPTokenStoreFactory: func(serverURL string) (kit.MCPTokenStore, error) {
         return myCustomStore(serverURL), nil  // custom OAuth token storage
     },
@@ -118,12 +140,34 @@ host, err := kit.New(ctx, &kit.Options{
         "docs": mcpSrv,  // *server.MCPServer from mcp-go — no subprocess needed
     },
 })
+
+// Tiny helper to take the address of a literal for pointer fields.
+func ptrFloat32(v float32) *float32 { return &v }
 ```
 
 **Critical distinction**: `Tools` replaces ALL default tools (core + MCP + extension). `ExtraTools` adds tools alongside the defaults. Use `Tools` to restrict the agent's capabilities; use `ExtraTools` to extend them.
 
 **In-process MCP servers** bypass subprocess spawning entirely. Pass `*server.MCPServer` instances from mcp-go via `InProcessMCPServers` or call `AddInProcessMCPServer()` at runtime.
 
+### Generation & provider Options (cheat sheet)
+
+| Field | Type | Empty/nil means | Notes |
+|-------|------|-----------------|-------|
+| `MaxTokens` | `int` | Auto-resolve (env → config → per-model → 8192 floor) | Non-zero suppresses `rightSizeMaxTokens` |
+| `ThinkingLevel` | `string` | Auto-resolve (→ `"off"`) | Valid: `"off"`, `"none"`, `"minimal"`, `"low"`, `"medium"`, `"high"` |
+| `Temperature` | `*float32` | Leave provider/per-model default | Pointer so explicit `0.0` ≠ unset |
+| `TopP` | `*float32` | Leave provider/per-model default | |
+| `TopK` | `*int32` | Leave provider/per-model default | |
+| `FrequencyPenalty` | `*float32` | Leave provider/per-model default | OpenAI-family |
+| `PresencePenalty` | `*float32` | Leave provider/per-model default | OpenAI-family |
+| `ProviderAPIKey` | `string` | Use config / provider env var | Overrides pre-existing viper state |
+| `ProviderURL` | `string` | Use provider default endpoint | Same base URL flag as `--provider-url` |
+| `TLSSkipVerify` | `bool` | — | Only effective when `true`; cannot force-disable via Options |
+
+These fields eliminate the old `viper.Set("max-tokens", 16384)` dance many
+downstream embedders used to do before calling `kit.New()`. Everything is
+now discoverable via godoc on `kit.Options`.
+
 ---
 
 ## Prompt Methods
@@ -208,6 +252,25 @@ unsub := host.OnToolCall(func(e kit.ToolCallEvent) {
 })
 defer unsub()
 
+host.OnToolCallStart(func(e kit.ToolCallStartEvent) {
+    // Fires when the LLM begins generating tool call arguments.
+    // e.ToolCallID, e.ToolName, e.ToolKind
+    // Use this to show a "running" indicator immediately — before the
+    // full argument JSON finishes streaming (eliminates "dead air").
+})
+
+host.OnToolCallDelta(func(e kit.ToolCallDeltaEvent) {
+    // Fires for each streamed fragment of tool call arguments.
+    // e.ToolCallID, e.Delta (JSON fragment)
+    // Useful for live-previewing artifact content or progress indicators.
+})
+
+host.OnToolCallEnd(func(e kit.ToolCallEndEvent) {
+    // Fires when tool argument streaming is complete, before execution.
+    // e.ToolCallID
+    // Transition UI from "generating args" to "executing".
+})
+
 host.OnToolResult(func(e kit.ToolResultEvent) {
     // e.ToolCallID, e.ToolName, e.ToolKind, e.ToolArgs, e.ParsedArgs
     // e.Result, e.IsError, e.Metadata (*ToolResultMetadata)
@@ -218,7 +281,7 @@ host.OnToolOutput(func(e kit.ToolOutputEvent) {
     // Streaming bash output chunks
 })
 
-host.OnStreaming(func(e kit.MessageUpdateEvent) {
+host.OnMessageUpdate(func(e kit.MessageUpdateEvent) {
     fmt.Print(e.Chunk) // real-time text streaming
 })
 
@@ -233,8 +296,64 @@ host.OnTurnStart(func(e kit.TurnStartEvent) {
 host.OnTurnEnd(func(e kit.TurnEndEvent) {
     // e.Response, e.Error, e.StopReason
 })
+
+host.OnStepStart(func(e kit.StepStartEvent) {
+    // e.StepNumber — which LLM call step (1-based)
+})
+
+host.OnStepFinish(func(e kit.StepFinishEvent) {
+    // e.StepNumber, e.HasToolCalls, e.FinishReason, e.Usage (LLMUsage)
+})
+
+host.OnWarnings(func(e kit.WarningsEvent) {
+    for _, w := range e.Warnings {
+        log.Printf("warning: %s", w)
+    }
+})
+
+host.OnError(func(e kit.ErrorEvent) {
+    log.Printf("agent error: %v", e.Error)
+})
+
+host.OnRetry(func(e kit.RetryEvent) {
+    log.Printf("retrying (attempt %d): %v", e.Attempt, e.Error)
+})
+
+host.OnTextStart(func(e kit.TextStartEvent) {
+    // e.ID — content block ID
+})
+
+host.OnTextEnd(func(e kit.TextEndEvent) {
+    // e.ID — content block ID
+})
+
+host.OnReasoningStart(func(e kit.ReasoningStartEvent) {
+    // e.ID — reasoning block ID
+})
+
+host.OnSource(func(e kit.SourceEvent) {
+    // e.SourceType, e.ID, e.URL, e.Title
+})
+
+host.OnStreamFinish(func(e kit.StreamFinishEvent) {
+    // e.Usage (LLMUsage), e.FinishReason
+})
+
+// Additional typed subscribers for previously generic-only events:
+host.OnMessageStart(func(e kit.MessageStartEvent) {})
+host.OnMessageEnd(func(e kit.MessageEndEvent) { /* e.Content */ })
+host.OnReasoningDelta(func(e kit.ReasoningDeltaEvent) { /* e.Delta */ })
+host.OnReasoningComplete(func(e kit.ReasoningCompleteEvent) {})
+host.OnToolExecutionStart(func(e kit.ToolExecutionStartEvent) { /* e.ToolCallID, e.ToolName, e.ToolKind, e.ToolArgs */ })
+host.OnToolExecutionEnd(func(e kit.ToolExecutionEndEvent) { /* e.ToolCallID, e.ToolName, e.ToolKind */ })
+host.OnToolCallContent(func(e kit.ToolCallContentEvent) { /* e.Content */ })
+host.OnStepUsage(func(e kit.StepUsageEvent) { /* e.InputTokens, e.OutputTokens, e.CacheReadTokens, e.CacheWriteTokens */ })
+host.OnCompaction(func(e kit.CompactionEvent) { /* e.Summary, e.OriginalTokens, e.CompactedTokens, ... */ })
+host.OnSteerConsumed(func(e kit.SteerConsumedEvent) { /* e.Count */ })
 ```
 
+> **Rename note:** `OnStreaming` has been renamed to `OnMessageUpdate`. The old `OnStreaming` name is kept as a deprecated alias for one release cycle.
+
 ### Generic subscriber (receives all events)
 
 ```go
@@ -259,6 +378,9 @@ unsub := host.Subscribe(func(e kit.Event) {
 | `message_start` | `MessageStartEvent` | *(none)* |
 | `message_update` | `MessageUpdateEvent` | `Chunk` |
 | `message_end` | `MessageEndEvent` | `Content` |
+| `tool_call_start` | `ToolCallStartEvent` | `ToolCallID`, `ToolName`, `ToolKind` |
+| `tool_call_delta` | `ToolCallDeltaEvent` | `ToolCallID`, `Delta` |
+| `tool_call_end` | `ToolCallEndEvent` | `ToolCallID` |
 | `tool_call` | `ToolCallEvent` | `ToolCallID`, `ToolName`, `ToolKind`, `ToolArgs`, `ParsedArgs` |
 | `tool_execution_start` | `ToolExecutionStartEvent` | `ToolCallID`, `ToolName`, `ToolKind`, `ToolArgs` |
 | `tool_execution_end` | `ToolExecutionEndEvent` | `ToolCallID`, `ToolName`, `ToolKind` |
@@ -270,6 +392,39 @@ unsub := host.Subscribe(func(e kit.Event) {
 | `reasoning_delta` | `ReasoningDeltaEvent` | `Delta` |
 | `step_usage` | `StepUsageEvent` | `InputTokens`, `OutputTokens`, `CacheReadTokens`, `CacheWriteTokens` |
 | `steer_consumed` | `SteerConsumedEvent` | `Count` |
+| `step_start` | `StepStartEvent` | `StepNumber` |
+| `step_finish` | `StepFinishEvent` | `StepNumber`, `HasToolCalls`, `FinishReason`, `Usage` |
+| `text_start` | `TextStartEvent` | `ID` |
+| `text_end` | `TextEndEvent` | `ID` |
+| `reasoning_start` | `ReasoningStartEvent` | `ID` |
+| `warnings` | `WarningsEvent` | `Warnings` |
+| `source` | `SourceEvent` | `SourceType`, `ID`, `URL`, `Title` |
+| `stream_finish` | `StreamFinishEvent` | `Usage`, `FinishReason` |
+| `error` | `ErrorEvent` | `Error` |
+| `retry` | `RetryEvent` | `Attempt`, `Error` |
+| `password_prompt` | `PasswordPromptEvent` | `Prompt`, `ResponseCh` |
+
+**Tool call streaming lifecycle**: `ToolCallStartEvent` → `ToolCallDeltaEvent` (repeated) → `ToolCallEndEvent` → `ToolCallEvent` → `ToolExecutionStartEvent` → `ToolOutputEvent` (optional, repeated) → `ToolExecutionEndEvent` → `ToolResultEvent`
+
+**PasswordPromptEvent** (for sudo password handling):
+```go
+// PasswordPromptEvent fires when a sudo command needs a password.
+// The TUI should display a password prompt and send the result back via ResponseCh.
+type PasswordPromptEvent struct {
+    // Prompt is the message to display to the user.
+    Prompt string
+    // ResponseCh receives the password from the TUI.
+    // The TUI must send exactly one value: (password, false) for submit
+    // or ("", true) for cancel.
+    ResponseCh chan<- PasswordPromptResponse
+}
+
+// PasswordPromptResponse carries the password prompt result.
+type PasswordPromptResponse struct {
+    Password  string
+    Cancelled bool
+}
+```
 
 ### Tool kind constants
 
@@ -332,6 +487,20 @@ host.OnAfterTurn(kit.HookPriorityNormal, func(h kit.AfterTurnHook) {
 })
 ```
 
+### PrepareStep — intercept/replace messages before each LLM call
+
+```go
+host.OnPrepareStep(kit.HookPriorityNormal, func(h kit.PrepareStepHook) *kit.PrepareStepResult {
+    // h.StepNumber  — which step in the current turn (1-based)
+    // h.Messages    — []kit.LLMMessage being sent to the LLM
+    // Return nil to pass through unchanged, or replace messages:
+    modified := filterSensitiveMessages(h.Messages)
+    return &kit.PrepareStepResult{Messages: modified}
+})
+```
+
+`PrepareStep` fires before every LLM API call within a turn (including tool-call loop iterations). Unlike `ContextPrepare` (which operates on the full context window once per turn), `PrepareStep` runs per-step and sees the messages that include the latest tool results.
+
 ### ContextPrepare — filter/inject context window
 
 ```go
@@ -404,6 +573,8 @@ host, _ := kit.New(ctx, &kit.Options{
 |----------|-------------|
 | `kit.TextResult(content)` | Successful text result |
 | `kit.ErrorResult(content)` | Error result (LLM sees it as a tool error) |
+| `kit.ImageResult(content, data, mediaType)` | Image result with binary data (e.g. `"image/png"`) |
+| `kit.MediaResult(content, data, mediaType)` | Non-image media result (e.g. `"audio/mpeg"`) |
 
 **ToolOutput fields** (for advanced use):
 
@@ -760,9 +931,65 @@ err = host.SubscribeMCPResource(ctx, "myserver", "file:///path/to/file")
 err = host.UnsubscribeMCPResource(ctx, "myserver", "file:///path/to/file")
 ```
 
+### MCP OAuth Authorization
+
+When a remote MCP server requires OAuth, Kit runs the full authorization flow
+(dynamic client registration → PKCE → user consent → token exchange → token
+persistence) but delegates the **user-facing step** — displaying the
+authorization URL and receiving the callback — to an `MCPAuthHandler`.
+
+The SDK ships three building blocks:
+
+| Building block | When to use |
+|---|---|
+| **No handler** (`Options.MCPAuthHandler = nil`) | Default. OAuth is disabled; 401s from remote MCP servers surface as errors. Correct for library, daemon, and web-app embedders that don't want side effects. |
+| **`kit.NewCLIMCPAuthHandler()`** | CLI/TUI apps. Opens the system browser, prints status to stderr (or via `NotifyFunc`), runs a localhost callback server. This is what the `kit` binary uses. |
+| **`kit.NewDefaultMCPAuthHandler()` + `OnAuthURL`** | Custom UX. Get the transport mechanics (port reservation + callback server) from the SDK; wire your own presentation in the `OnAuthURL(serverName, authURL)` closure. |
+| **Implement `kit.MCPAuthHandler` directly** | Full control. No localhost binding — e.g. return the URL from an HTTP endpoint and have the consumer POST the callback URL back. |
+
+**CLI-style embedder (browser + stderr):**
+
+```go
+authHandler, err := kit.NewCLIMCPAuthHandler()
+if err != nil {
+    log.Fatal(err)
+}
+defer authHandler.Close() // release the reserved port
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPAuthHandler: authHandler,
+})
+```
+
+**Custom UX embedder (TUI modal, QR code, web redirect, etc.):**
+
+```go
+authHandler, _ := kit.NewDefaultMCPAuthHandler()
+authHandler.OnAuthURL = func(serverName, authURL string) {
+    // Render the URL however you like — no browser or terminal assumptions.
+    myUI.ShowAuthPrompt(serverName, authURL)
+}
+defer authHandler.Close()
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPAuthHandler: authHandler,
+})
+```
+
+**Important:** `DefaultMCPAuthHandler` with no `OnAuthURL` set will silently
+drop the authorization URL and block until the 2-minute callback timeout
+fires. Always set `OnAuthURL`, or use a higher-level wrapper like
+`CLIMCPAuthHandler`.
+
 ### MCP OAuth Token Storage
 
-For remote MCP servers that use OAuth, you can provide a custom token store:
+Once authorization succeeds, the resulting access/refresh tokens are persisted
+by an `MCPTokenStore`. By default tokens are written to
+`$XDG_CONFIG_HOME/.kit/mcp_tokens.json` (fallback `~/.config/.kit/mcp_tokens.json`),
+keyed by server URL, with `0600` file permissions.
+
+Provide a custom store for encrypted storage, database persistence, or
+in-memory-only flows:
 
 ```go
 host, _ := kit.New(ctx, &kit.Options{
@@ -772,7 +999,7 @@ host, _ := kit.New(ctx, &kit.Options{
 })
 ```
 
-The `MCPTokenStore` interface requires `GetToken`/`SetToken`/`DeleteToken` methods. Return `kit.ErrMCPNoToken` from `GetToken` when no token is stored. When nil (default), tokens are persisted to `$XDG_CONFIG_HOME/.kit/mcp_tokens.json`.
+The `MCPTokenStore` interface requires `GetToken`/`SetToken`/`DeleteToken` methods. Return `kit.ErrMCPNoToken` from `GetToken` when no token is stored.
 
 ---
 
@@ -943,18 +1170,41 @@ kit.Config, kit.MCPServerConfig
 // Provider types
 kit.ProviderConfig, kit.ProviderResult, kit.ModelInfo, kit.ModelCost, kit.ModelLimit
 
-// LLM types — concrete Kit-owned structs (no external library dependency)
+// LLM types — clean aliases (no external library dependency in consumer code)
 kit.LLMMessage      // {Role LLMMessageRole, Content string}
+kit.LLMMessagePart  // interface for message content parts
 kit.LLMMessageRole  // "user" | "assistant" | "system" | "tool"
 kit.LLMUsage        // {InputTokens, OutputTokens, TotalTokens, ReasoningTokens,
                      //  CacheCreationTokens, CacheReadTokens}
 kit.LLMResponse     // {Content, FinishReason, Usage}
 kit.LLMFilePart     // {Filename, Data []byte, MediaType}
+kit.LLMTextPart     // plain-text content part
+kit.LLMReasoningPart // reasoning/chain-of-thought content part
+kit.LLMToolCall     // {ID, Name, Input string} — execution-layer tool call (for Tool.Run)
+kit.LLMToolResponse // {Type, Content, Data, MediaType, IsError, ...} — raw tool response
+kit.LLMToolCallPart    // LLM-initiated tool invocation within a message
+kit.LLMToolResultPart  // tool result within a message
+kit.LLMToolResultOutputContent      // interface for tool result output
+kit.LLMToolResultOutputContentText  // text tool result
+kit.LLMToolResultOutputContentError // error tool result
+kit.LLMToolResultOutputContentMedia // media tool result {Data, MediaType, Text}
+kit.LLMToolResultContentType        // "text" | "error" | "media"
+kit.LLMToolInfo          // {Name, Description, Parameters, Required, Parallel}
+kit.LLMProviderOptions   // provider-specific option maps (keyed by provider name)
+kit.LLMProviderMetadata  // provider-specific response metadata
+kit.LLMPrompt            // []LLMMessage — ordered prompt sequence
+kit.LLMFinishReason      // "stop" | "length" | "tool-calls" | ...
 
 // Compaction types
 kit.CompactionResult, kit.CompactionOptions
 
 // MCP OAuth types
+kit.MCPAuthHandler         // interface: RedirectURI() + HandleAuth(ctx, server, authURL) for OAuth UX
+kit.DefaultMCPAuthHandler  // SDK-provided transport mechanics (port + callback server); set OnAuthURL hook
+kit.CLIMCPAuthHandler      // CLI wrapper around DefaultMCPAuthHandler: opens browser, prints status
+kit.NewDefaultMCPAuthHandler()         // random port, no UX side effects
+kit.NewDefaultMCPAuthHandlerWithPort() // fixed port (useful when registering a stable redirect URI)
+kit.NewCLIMCPAuthHandler()             // CLI handler: browser + stderr + localhost callback
 kit.MCPTokenStore        // interface for custom OAuth token storage
 kit.MCPToken             // OAuth token struct (access, refresh, expiry)
 kit.MCPTokenStoreFactory // func(serverURL string) (MCPTokenStore, error)
@@ -1017,7 +1267,7 @@ for {
 ### Pattern: Streaming output to terminal
 
 ```go
-host.OnStreaming(func(e kit.MessageUpdateEvent) {
+host.OnMessageUpdate(func(e kit.MessageUpdateEvent) {
     fmt.Print(e.Chunk)
 })
 response, _ := host.Prompt(ctx, "Write a poem")

test_messages.txt

@@ -1,9 +0,0 @@
-1. Hello, world!
-
-2. Testing one, two, three.
-
-3. This is a quick test message.
-
-4. Sample text for verification.
-
-5. All systems operational.

www/pages/cli/commands.md

@@ -10,9 +10,10 @@ description: Complete reference for all Kit CLI subcommands.
 For OAuth-enabled providers like Anthropic.
 
 ```bash
-kit auth login [provider]    # Start OAuth flow (e.g., anthropic)
-kit auth logout [provider]   # Remove credentials for provider
-kit auth status              # Check authentication status
+kit auth login [provider]          # Start OAuth flow (e.g., anthropic)
+kit auth login [provider] --set-default  # Set provider's default model as system default
+kit auth logout [provider]       # Remove credentials for provider
+kit auth status                    # Check authentication status
 ```
 
 ## Model database
@@ -66,7 +67,7 @@ These commands are available inside the Kit TUI during an interactive session:
 | `/servers` | Show connected MCP servers |
 | `/model [name]` | Switch model or open model selector |
 | `/theme [name]` | Switch color theme or list available themes |
-| `/thinking [level]` | Set thinking level (off, minimal, low, medium, high) |
+| `/thinking [level]` | Set thinking level (off, none, minimal, low, medium, high) |
 | `/compact [focus]` | Summarize older messages to free context |
 | `/clear` | Clear conversation |
 | `/clear-queue` | Clear queued messages |

www/pages/cli/flags.md

@@ -52,14 +52,14 @@ These flags control Kit's behavior. When a prompt is passed as a positional argu
 
 | Flag | Short | Default | Description |
 |------|-------|---------|-------------|
-| `--max-tokens` | — | `4096` | Maximum tokens in response |
+| `--max-tokens` | — | `8192` | Base cap for output tokens. Auto-raised per-model up to 32768 when the model's catalog ceiling is higher and no explicit value is set. |
 | `--temperature` | — | `0.7` | Randomness 0.0–1.0 |
 | `--top-p` | — | `0.95` | Nucleus sampling 0.0–1.0 |
 | `--top-k` | — | `40` | Limit top K tokens |
 | `--stop-sequences` | — | — | Custom stop sequences (comma-separated) |
 | `--frequency-penalty` | — | `0.0` | Penalize frequent tokens (0.0–2.0) |
 | `--presence-penalty` | — | `0.0` | Penalize present tokens (0.0–2.0) |
-| `--thinking-level` | — | `off` | Extended thinking level: off, minimal, low, medium, high |
+| `--thinking-level` | — | `off` | Extended thinking level: off, none, minimal, low, medium, high |
 
 ## System
 

www/pages/configuration.md

@@ -18,7 +18,7 @@ Create `~/.kit.yml`:
 
 ```yaml
 model: anthropic/claude-sonnet-latest
-max-tokens: 4096
+max-tokens: 8192
 temperature: 0.7
 stream: true
 ```
@@ -28,7 +28,7 @@ stream: true
 | Key | Type | Default | Description |
 |-----|------|---------|-------------|
 | `model` | string | `anthropic/claude-sonnet-latest` | Model to use (provider/model format) |
-| `max-tokens` | int | `4096` | Maximum tokens in response |
+| `max-tokens` | int | `8192` | Base cap for output tokens. Auto-raised per-model up to 32768 when the model's catalog ceiling is higher and no explicit value is set. Use [`modelSettings[provider/model].maxTokens`](#per-model-settings) to override per-model. |
 | `temperature` | float | `0.7` | Randomness 0.0–1.0 |
 | `top-p` | float | `0.95` | Nucleus sampling 0.0–1.0 |
 | `top-k` | int | `40` | Limit top K tokens |
@@ -37,7 +37,7 @@ stream: true
 | `compact` | bool | `false` | Enable compact output mode |
 | `system-prompt` | string | — | System prompt text or file path |
 | `max-steps` | int | `0` | Maximum agent steps (0 = unlimited) |
-| `thinking-level` | string | `off` | Extended thinking: off, minimal, low, medium, high |
+| `thinking-level` | string | `off` | Extended thinking: off, none, minimal, low, medium, high |
 | `provider-api-key` | string | — | API key for the provider |
 | `provider-url` | string | — | Base URL for provider API |
 | `tls-skip-verify` | bool | `false` | Skip TLS certificate verification |
@@ -83,6 +83,16 @@ mcpServers:
   search:
     type: remote
     url: "https://mcp.example.com/search"
+
+  pubmed:
+    type: remote
+    url: "https://pubmed.mcp.example.com"
+    noOAuth: true  # skip OAuth for public servers
+
+  builds:
+    type: remote
+    url: "https://builds.mcp.example.com"
+    tasksMode: always  # always run tools/call as async tasks (Phase 1 MVP)
 ```
 
 ### MCP server fields
@@ -95,9 +105,35 @@ mcpServers:
 | `url` | string | URL for remote servers |
 | `allowedTools` | list | Whitelist of tool names to expose |
 | `excludedTools` | list | Blacklist of tool names to hide |
+| `noOAuth` | bool | Skip OAuth for this server (for public servers that don't require auth) |
+| `tasksMode` | string | When to augment `tools/call` with MCP task metadata: `auto` (default — only when the server advertises task support), `never`, or `always`. See [MCP tasks](#mcp-tasks-long-running-tools). |
 
 A legacy format with `transport`, `args`, `env`, and `headers` fields is also supported.
 
+### MCP tasks (long-running tools)
+
+Kit advertises [MCP task support](https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks)
+during `initialize` so servers can respond to `tools/call` with a
+`CreateTaskResult` (a task ID + `working` status) instead of blocking until
+the operation finishes. Kit then polls `tasks/get` / `tasks/result` until the
+task reaches a terminal state, and best-effort `tasks/cancel`s on context
+cancellation.
+
+This avoids HTTP/SSE proxy timeouts on long builds, deploys, and batch jobs,
+and lets the user/agent abort cleanly with Ctrl-C.
+
+**Per-server `tasksMode`:**
+
+| Value | Behaviour |
+|-------|-----------|
+| `auto` (default) | Augment `tools/call` with task metadata only when the server advertised `tasks/toolCalls` capability. Servers that don't advertise it run synchronously, exactly as before. |
+| `never` | Always issue `tools/call` synchronously, regardless of server capability. |
+| `always` | Always opt into task augmentation, even when the server didn't advertise the capability. The server may still respond synchronously — this just expresses client intent unconditionally. |
+
+Defaults are safe: any existing MCP server keeps its previous behaviour
+bit-for-bit. SDK consumers can also override the mode programmatically and
+plug in a progress callback — see [SDK options](/sdk/options#mcp-tasks).
+
 ## Custom models
 
 Define custom models in your `.kit.yml` for use with the `custom` provider. This is useful for self-hosted models or API endpoints not in the built-in database:
@@ -175,10 +211,24 @@ modelSettings:
 | `thinkingLevel` | string | Thinking level override |
 | `systemPrompt` | string | Per-model system prompt (used when no explicit prompt is set) |
 
-Settings from `modelSettings` and `customModels.params` act as model-level defaults — explicit CLI flags and global config values always take precedence.
+Settings from `modelSettings` and `customModels.params` act as model-level defaults — explicit CLI flags, `KIT_*` environment variables, global config values, and SDK `Options.*` fields all take precedence over them.
 
 When switching models via `/model` or `SetModel()`, if the new model has a per-model system prompt and no custom global prompt was set, the per-model prompt automatically replaces the previous one.
 
+### Precedence summary
+
+For the generation and provider parameters documented above, the resolved value at runtime comes from the first source that sets it:
+
+1. CLI flag (e.g. `--max-tokens`, `--temperature`, `--provider-api-key`)
+2. SDK `Options.X` when embedding Kit as a library (`kit.Options.MaxTokens`, `Temperature`, `ProviderAPIKey`, etc.)
+3. `KIT_*` environment variable (`KIT_MAX_TOKENS`, `KIT_TEMPERATURE`, ...)
+4. `.kit.yml` / `.kit.yaml` / `.kit.json` (project-local, then global)
+5. Per-model defaults (`modelSettings[provider/model]` / `customModels[...].params`)
+6. Provider-level defaults (e.g. Anthropic's own temperature default)
+7. SDK last-resort floor — currently an 8192 output-token ceiling matching the CLI `--max-tokens` default, auto-raised per-model up to 32768 when the model's catalog ceiling is higher
+
+See the [SDK options reference](/sdk/options) for the full list of `kit.Options` fields that map to these keys.
+
 ## Theme configuration
 
 ```yaml

www/pages/development.md

@@ -37,7 +37,7 @@ internal/acpserver/  - ACP (Agent Client Protocol) server
 internal/clipboard/  - Cross-platform clipboard operations
 internal/compaction/ - Conversation compaction and summarization
 internal/config/     - Configuration management
-internal/core/       - Built-in tools (bash, read, write, edit, grep, find, ls)
+internal/core/       - Built-in tools (bash with sudo password prompt, read, write, edit, grep, find, ls)
 internal/extensions/ - Yaegi extension system
 internal/kitsetup/   - Initial setup wizard
 internal/message/    - Message content types and structured content blocks

www/pages/extensions/capabilities.md

@@ -7,7 +7,7 @@ description: All extension capabilities — lifecycle events, tools, commands, w
 
 ## Lifecycle events
 
-Extensions can hook into 23 lifecycle events:
+Extensions can hook into 26 lifecycle events:
 
 | Event | Description |
 |-------|-------------|
@@ -17,6 +17,9 @@ Extensions can hook into 23 lifecycle events:
 | `OnAgentStart` | Agent loop started |
 | `OnAgentEnd` | Agent loop completed |
 | `OnToolCall` | Tool call requested by the model |
+| `OnToolCallInputStart` | LLM began generating tool call arguments (tool name known, args streaming) |
+| `OnToolCallInputDelta` | Streamed JSON fragment of tool call arguments |
+| `OnToolCallInputEnd` | Tool argument streaming complete, before execution begins |
 | `OnToolExecutionStart` | Tool execution beginning |
 | `OnToolOutput` | Streaming tool output chunk (for long-running tools) |
 | `OnToolExecutionEnd` | Tool execution completed |

www/pages/index.md

@@ -13,7 +13,7 @@ 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, subagent with no MCP overhead
+- **Built-in Core Tools** — bash (with interactive sudo password prompt), read, write, edit, grep, find, ls, subagent with no MCP overhead
 - **Smart @ Attachments** — Binary files auto-detected via MIME type, MCP resources via `@mcp:server:uri`
 - **MCP Integration** — Connect external MCP servers for expanded capabilities (tools, prompts, and resources)
 - **Extension System** — Write custom tools, commands, widgets, and UI modifications in Go

www/pages/sdk/callbacks.md

@@ -20,7 +20,7 @@ unsub2 := host.OnToolResult(func(event kit.ToolResultEvent) {
 })
 defer unsub2()
 
-unsub3 := host.OnStreaming(func(event kit.MessageUpdateEvent) {
+unsub3 := host.OnMessageUpdate(func(event kit.MessageUpdateEvent) {
     fmt.Print(event.Chunk)
 })
 defer unsub3()
@@ -41,6 +41,32 @@ unsub6 := host.OnTurnEnd(func(event kit.TurnEndEvent) {
 defer unsub6()
 ```
 
+## Tool call argument streaming
+
+For tools with large arguments (e.g., `write` with a full file body), the `ToolCallEvent` only fires after the full argument JSON finishes streaming — which can take 5-10+ seconds of "dead air." These three events fire during argument generation so UIs can show activity immediately:
+
+```go
+host.OnToolCallStart(func(event kit.ToolCallStartEvent) {
+    // Fires as soon as the LLM begins generating tool arguments.
+    // event.ToolCallID, event.ToolName, event.ToolKind
+    fmt.Printf("⏳ %s generating arguments...\n", event.ToolName)
+})
+
+host.OnToolCallDelta(func(event kit.ToolCallDeltaEvent) {
+    // Each streamed JSON fragment of the tool arguments.
+    // event.ToolCallID, event.Delta
+    // Useful for live-previewing content or showing byte progress.
+})
+
+host.OnToolCallEnd(func(event kit.ToolCallEndEvent) {
+    // Tool argument streaming complete — execution about to begin.
+    // event.ToolCallID
+    fmt.Printf("✓ Arguments ready, executing...\n")
+})
+```
+
+**Full tool lifecycle**: `ToolCallStartEvent` → `ToolCallDeltaEvent` (repeated) → `ToolCallEndEvent` → `ToolCallEvent` → `ToolExecutionStartEvent` → `ToolOutputEvent` (optional) → `ToolExecutionEndEvent` → `ToolResultEvent`
+
 ## Hook system
 
 Hooks can **modify or cancel** operations. Unlike events (read-only), hooks are read-write interceptors.
@@ -90,6 +116,24 @@ host.OnAfterTurn(kit.HookPriorityNormal, func(h kit.AfterTurnHook) {
 })
 ```
 
+### PrepareStep — intercept messages between steps
+
+The most powerful hook — fires between steps within a multi-step agent turn, after any steering messages are injected and before messages are sent to the LLM. Can replace the entire context window.
+
+```go
+host.OnPrepareStep(kit.HookPriorityNormal, func(h kit.PrepareStepHook) *kit.PrepareStepResult {
+    // h.StepNumber — zero-based step index within the turn
+    // h.Messages   — current context window (includes any steering)
+    
+    // Example: transform tool results with images into user messages
+    modified := transformImageToolResults(h.Messages)
+    return &kit.PrepareStepResult{Messages: modified}
+    // Return nil to pass through unchanged
+})
+```
+
+Use cases: transforming tool results (e.g., image data for vision models), dynamic tool filtering per step, mid-turn context injection, custom stop conditions.
+
 ### Hook priorities
 
 ```go
@@ -100,6 +144,44 @@ kit.HookPriorityLow    = 100 // runs last
 
 Lower values run first. First non-nil result wins.
 
+## All event types
+
+| Event | Typed Subscriber | Description |
+|-------|-----------------|-------------|
+| `TurnStartEvent` | `OnTurnStart` | Agent turn started |
+| `TurnEndEvent` | `OnTurnEnd` | Agent turn completed |
+| `MessageStartEvent` | `OnMessageStart` | New assistant message begins |
+| `MessageUpdateEvent` | `OnMessageUpdate` | Streaming text chunk from LLM |
+| `MessageEndEvent` | `OnMessageEnd` | Assistant message complete |
+| `ToolCallStartEvent` | `OnToolCallStart` | LLM began generating tool call arguments |
+| `ToolCallDeltaEvent` | `OnToolCallDelta` | Streamed JSON fragment of tool call arguments |
+| `ToolCallEndEvent` | `OnToolCallEnd` | Tool argument streaming complete |
+| `ToolCallEvent` | `OnToolCall` | Tool call fully parsed, about to execute |
+| `ToolExecutionStartEvent` | `OnToolExecutionStart` | Tool begins executing |
+| `ToolExecutionEndEvent` | `OnToolExecutionEnd` | Tool finishes executing |
+| `ToolResultEvent` | `OnToolResult` | Tool execution completed with result |
+| `ToolCallContentEvent` | `OnToolCallContent` | Text content alongside tool calls |
+| `ToolOutputEvent` | `OnToolOutput` | Streaming output chunk from tool (e.g., bash) |
+| `ResponseEvent` | `OnResponse` | Final response received |
+| `ReasoningStartEvent` | `OnReasoningStart` | LLM begins reasoning/thinking |
+| `ReasoningDeltaEvent` | `OnReasoningDelta` | Streaming reasoning/thinking chunk |
+| `ReasoningCompleteEvent` | `OnReasoningComplete` | Reasoning/thinking finished |
+| `StepStartEvent` | `OnStepStart` | New LLM call begins within a turn |
+| `StepFinishEvent` | `OnStepFinish` | Step completes (with usage, finish reason, tool call info) |
+| `StepUsageEvent` | `OnStepUsage` | Per-step token usage |
+| `StreamFinishEvent` | `OnStreamFinish` | Per-step stream completes (with usage + finish reason) |
+| `TextStartEvent` | `OnTextStart` | LLM begins text content generation |
+| `TextEndEvent` | `OnTextEnd` | LLM finishes text content generation |
+| `WarningsEvent` | `OnWarnings` | LLM provider returned warnings |
+| `SourceEvent` | `OnSource` | LLM referenced a source (e.g., web search) |
+| `ErrorEvent` | `OnError` | Agent-level error during streaming |
+| `RetryEvent` | `OnRetry` | LLM request retried after transient error |
+| `CompactionEvent` | `OnCompaction` | Conversation compacted |
+| `SteerConsumedEvent` | `OnSteerConsumed` | Steering messages injected into turn |
+| `PasswordPromptEvent` | — | Sudo command needs password (respond via `ResponseCh`) |
+
+> **Note:** `OnStreaming` is a deprecated alias for `OnMessageUpdate` and will be removed in a future release.
+
 ## Subagent event monitoring
 
 Monitor real-time events from LLM-initiated subagents (when the model uses the `subagent` tool):

www/pages/sdk/options.md

@@ -22,6 +22,20 @@ host, err := kit.New(ctx, &kit.Options{
     Quiet:        true,
     Debug:        true,
 
+    // Generation parameters (override env/config/per-model defaults)
+    MaxTokens:        16384,              // 0 = auto-resolve; non-zero suppresses right-sizing
+    ThinkingLevel:    "medium",           // "off", "none", "minimal", "low", "medium", "high"
+    Temperature:      ptrFloat32(0.2),    // pointer so explicit 0.0 != unset
+    TopP:             nil,                 // nil = provider/per-model default
+    TopK:             nil,
+    FrequencyPenalty: nil,
+    PresencePenalty:  nil,
+
+    // Provider configuration
+    ProviderAPIKey: "sk-...",                      // "" = use config / provider env var
+    ProviderURL:    "https://proxy.internal/v1",  // "" = provider default endpoint
+    TLSSkipVerify:  false,                         // only effective when true
+
     // Session
     SessionPath:  "./session.jsonl",
     SessionDir:   "/custom/sessions/",
@@ -51,7 +65,11 @@ host, err := kit.New(ctx, &kit.Options{
     // Session (advanced)
     SessionManager: myCustomSession,    // custom SessionManager implementation
 
-    // MCP OAuth
+    // MCP OAuth — both opt-in. Leave MCPAuthHandler nil to disable
+    // OAuth entirely (remote MCP 401s bubble up as errors). CLI apps
+    // pass kit.NewCLIMCPAuthHandler(); custom UX embedders implement
+    // MCPAuthHandler or configure DefaultMCPAuthHandler + OnAuthURL.
+    MCPAuthHandler: authHandler,                  // nil = OAuth disabled
     MCPTokenStoreFactory: func(serverURL string) (kit.MCPTokenStore, error) {
         return myStore(serverURL), nil
     },
@@ -65,6 +83,8 @@ host, err := kit.New(ctx, &kit.Options{
 
 ## Options fields
 
+### Core
+
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
 | `Model` | `string` | config default | Model string (provider/model format) |
@@ -74,24 +94,253 @@ host, err := kit.New(ctx, &kit.Options{
 | `Streaming` | `bool` | `true` | Enable streaming output |
 | `Quiet` | `bool` | `false` | Suppress output |
 | `Debug` | `bool` | `false` | Enable debug logging |
+
+### Generation parameters
+
+These fields override the corresponding values from `.kit.yml` / `KIT_*`
+environment variables. Leaving a field at its zero/nil value lets the
+precedence chain resolve a value (`KIT_*` env → config file → per-model
+defaults from `modelSettings`/`customModels` → an 8192 SDK floor for
+`MaxTokens` (matching the CLI `--max-tokens` default) and provider-level
+defaults for samplers).
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `MaxTokens` | `int` | auto-resolved | Max output tokens per response. `0` = auto-resolve; non-zero suppresses automatic right-sizing (same semantics as `--max-tokens`). |
+| `ThinkingLevel` | `string` | auto-resolved | Reasoning effort: `"off"`, `"none"`, `"minimal"`, `"low"`, `"medium"`, `"high"`. `""` falls through to config/env/per-model/`"off"`. |
+| `Temperature` | `*float32` | — | Sampling randomness. Pointer type so explicit `0.0` is distinguishable from "unset". |
+| `TopP` | `*float32` | — | Nucleus sampling cutoff. `nil` leaves provider/per-model default. |
+| `TopK` | `*int32` | — | Top-K sampling limit. `nil` leaves provider/per-model default. |
+| `FrequencyPenalty` | `*float32` | — | OpenAI-family frequency penalty. `nil` leaves provider default. |
+| `PresencePenalty` | `*float32` | — | OpenAI-family presence penalty. `nil` leaves provider default. |
+
+Pointer-typed samplers are populated via a tiny helper:
+
+```go
+func ptrFloat32(v float32) *float32 { return &v }
+```
+
+These fields eliminate the need for `viper.Set()` calls before `kit.New()`
+when embedding Kit as a library.
+
+### Provider configuration
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `ProviderAPIKey` | `string` | — | API key used to authenticate with the provider. `""` falls back to config / provider-specific env var (e.g. `ANTHROPIC_API_KEY`). When set, overrides any pre-existing viper state. |
+| `ProviderURL` | `string` | — | Override the provider endpoint (e.g. LiteLLM, vLLM, Azure OpenAI, internal proxy). `""` = provider default. |
+| `TLSSkipVerify` | `bool` | `false` | Disable TLS certificate verification on the provider HTTP client. Only effective when `true`; to force-disable, use config file or env var instead. For self-signed dev certs only. |
+
+### Session
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
 | `SessionPath` | `string` | — | Open a specific session file |
 | `SessionDir` | `string` | — | Base directory for session discovery |
 | `Continue` | `bool` | `false` | Resume most recent session |
 | `NoSession` | `bool` | `false` | Ephemeral mode (no persistence) |
+| `SessionManager` | `SessionManager` | — | Custom session backend (advanced) |
+
+### Tools & extensions
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
 | `Tools` | `[]Tool` | — | Replace the entire default tool set |
 | `ExtraTools` | `[]Tool` | — | Additional tools alongside core/MCP/extension tools |
 | `DisableCoreTools` | `bool` | `false` | Use no core tools (0 tools, for chat-only) |
-| `SkipConfig` | `bool` | `false` | Skip .kit.yml file loading |
-| `AutoCompact` | `bool` | `false` | Auto-compact when near context limit |
-| `CompactionOptions` | `*CompactionOptions` | — | Configuration for auto-compaction |
+| `NoExtensions` | `bool` | `false` | Disable Yaegi extension loading |
+| `NoContextFiles` | `bool` | `false` | Disable automatic AGENTS.md loading |
+
+### Skills & configuration
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `SkipConfig` | `bool` | `false` | Skip `.kit.yml` file loading (viper defaults + env vars still apply) |
 | `Skills` | `[]string` | — | Explicit skill files/dirs to load |
 | `SkillsDir` | `string` | — | Override default skills directory |
 | `NoSkills` | `bool` | `false` | Disable skill loading entirely |
-| `NoExtensions` | `bool` | `false` | Disable Yaegi extension loading |
-| `NoContextFiles` | `bool` | `false` | Disable automatic AGENTS.md loading |
-| `SessionManager` | `SessionManager` | — | Custom session backend (advanced) |
-| `MCPTokenStoreFactory` | `func` | — | Custom OAuth token storage for MCP servers |
+
+### Compaction & MCP
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `AutoCompact` | `bool` | `false` | Auto-compact when near context limit |
+| `CompactionOptions` | `*CompactionOptions` | — | Configuration for auto-compaction |
+| `MCPAuthHandler` | `MCPAuthHandler` | — | OAuth handler for remote MCP servers. `nil` disables OAuth (servers returning 401 fail with the authorization-required error). See [MCP OAuth](#mcp-oauth-authorization) below. |
+| `MCPTokenStoreFactory` | `func` | — | Custom OAuth token storage for MCP servers (default: JSON file in `$XDG_CONFIG_HOME/.kit/mcp_tokens.json`). |
 | `InProcessMCPServers` | `map[string]*MCPServer` | — | In-process mcp-go servers (no subprocess) |
+| `MCPTaskMode` | `map[string]MCPTaskMode` | — | Per-server override for task-augmented `tools/call`. Keys are server names; missing entries fall back to the `tasksMode` field of the matching `MCPServerConfig`. See [MCP Tasks](#mcp-tasks). |
+| `MCPTaskTimeout` | `time.Duration` | `15m` | Maximum wall-clock to wait for a task to reach a terminal state. Independent of any per-call context deadline. |
+| `MCPTaskTTL` | `time.Duration` | — | TTL hint sent in `TaskParams` for every task-augmented call. Zero omits the field and lets the server pick. |
+| `MCPTaskPollInterval` | `time.Duration` | `1s` | Fallback interval between `tasks/get` requests when the server does not suggest one. |
+| `MCPTaskMaxPollInterval` | `time.Duration` | `5s` | Cap on the polling interval (a server-supplied `pollInterval` can otherwise grow without bound). |
+| `MCPTaskProgress` | `MCPTaskProgressHandler` | — | Optional callback invoked once when a task is accepted and on every observed status transition. The final invocation always carries a terminal status. |
+
+## MCP OAuth Authorization
+
+When a remote MCP server (SSE or Streamable HTTP) requires OAuth, Kit runs
+the full authorization flow (dynamic client registration → PKCE → user
+consent → token exchange → token persistence) but delegates the **user-facing
+step** — displaying the authorization URL and receiving the callback — to
+an `MCPAuthHandler`.
+
+The SDK is deliberately inert when `MCPAuthHandler` is `nil`: it does **not**
+auto-construct a default handler, bind a local TCP port, or open a browser.
+This keeps library, daemon, and web-app embedders free of surprise I/O.
+Consumers opt in by passing a handler explicitly.
+
+| Building block | When to use |
+|---|---|
+| `MCPAuthHandler = nil` (default) | OAuth disabled. Remote MCP servers requiring auth fail with a clear error. Correct for libraries, daemons, and web apps. |
+| `kit.NewCLIMCPAuthHandler()` | CLI/TUI apps. Opens the system browser, prints status to stderr (or via `NotifyFunc`), runs a localhost callback server. Used by the `kit` binary. |
+| `kit.NewDefaultMCPAuthHandler()` + `OnAuthURL` | Custom UX. Use the SDK's port reservation and callback server; plug in your own presentation via the `OnAuthURL(serverName, authURL)` closure. |
+| Implement `kit.MCPAuthHandler` directly | Full control. No localhost binding — e.g. return the URL from an HTTP endpoint and have the consumer POST the callback URL back. |
+
+**CLI-style embedder:**
+
+```go
+authHandler, err := kit.NewCLIMCPAuthHandler()
+if err != nil {
+    log.Fatal(err)
+}
+defer authHandler.Close() // release the reserved port
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPAuthHandler: authHandler,
+})
+```
+
+**Custom UX embedder (TUI modal, QR code, web redirect, etc.):**
+
+```go
+authHandler, _ := kit.NewDefaultMCPAuthHandler()
+authHandler.OnAuthURL = func(serverName, authURL string) {
+    // No browser or terminal assumptions — render however you like.
+    myUI.ShowAuthPrompt(serverName, authURL)
+}
+defer authHandler.Close()
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPAuthHandler: authHandler,
+})
+```
+
+**Fully custom handler (no local port binding at all):**
+
+```go
+type WebAuthHandler struct {
+    redirectURI string
+    callbacks   chan string
+}
+
+func (h *WebAuthHandler) RedirectURI() string { return h.redirectURI }
+
+func (h *WebAuthHandler) HandleAuth(ctx context.Context, serverName, authURL string) (string, error) {
+    // Push the URL to the user's existing browser session via your web app,
+    // then block on the callback that your HTTP handler pushes onto the channel.
+    h.pushToUserSession(serverName, authURL)
+    select {
+    case callbackURL := <-h.callbacks:
+        return callbackURL, nil
+    case <-ctx.Done():
+        return "", ctx.Err()
+    }
+}
+```
+
+::: warning
+`DefaultMCPAuthHandler` with no `OnAuthURL` set will silently drop the
+authorization URL and hang until the 2-minute callback timeout fires. Always
+set `OnAuthURL`, or use a higher-level wrapper like `CLIMCPAuthHandler`.
+:::
+
+## MCP Tasks
+
+The [MCP Tasks utility](https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks)
+turns a synchronous `tools/call` into a pollable async job: the server
+returns a `taskId` with status `working` immediately, and the client polls
+`tasks/get` / `tasks/result` until the task reaches a terminal state.
+
+Kit advertises task support during `initialize` and, by default, augments
+`tools/call` with task metadata only when the server advertises
+`tasks/toolCalls` capability — so any existing MCP server keeps its previous
+synchronous behaviour bit-for-bit. Long-running tools (builds, deployments,
+batch jobs, sub-agent runs) get HTTP/SSE timeout-resistance and clean
+cancellation "for free" once both sides opt in.
+
+### Per-server mode
+
+```go
+import "time"
+
+host, _ := kit.New(ctx, &kit.Options{
+    MCPTaskMode: map[string]kit.MCPTaskMode{
+        "build-server": kit.MCPTaskModeAlways, // force task-augmented calls
+        "chat-server":  kit.MCPTaskModeNever,  // force synchronous calls
+        // any server not in the map honours its `tasksMode` config field
+        // (default "auto")
+    },
+})
+```
+
+| Mode | Behaviour |
+|---|---|
+| `MCPTaskModeAuto` (default) | Augment `tools/call` with `TaskParams` only when the server advertised `tasks/toolCalls`. |
+| `MCPTaskModeNever` | Always issue `tools/call` synchronously, ignoring server capability. |
+| `MCPTaskModeAlways` | Always opt in, even when the server didn't advertise the capability. The server may still respond synchronously. |
+
+### Progress callbacks
+
+```go
+host, _ := kit.New(ctx, &kit.Options{
+    MCPTaskTimeout:  15 * time.Minute,        // total wall-clock cap
+    MCPTaskTTL:      30 * time.Minute,        // server retention hint
+    MCPTaskProgress: func(p kit.MCPTaskProgress) {
+        log.Printf("%s/%s: %s %s", p.Server, p.TaskID, p.Status, p.Message)
+    },
+})
+```
+
+The handler fires once when a task is accepted and again on every observed
+status transition. The final call always carries a terminal status
+(`MCPTaskStatusCompleted`, `MCPTaskStatusFailed`, or `MCPTaskStatusCancelled`).
+Do not block in the handler — dispatch long work on a goroutine.
+
+### Inspecting and cancelling tasks
+
+```go
+tasks, _ := host.ListMCPTasks(ctx, "build-server")
+for _, t := range tasks {
+    fmt.Printf("%s: %s (%s)\n", t.TaskID, t.Status, t.StatusMessage)
+}
+
+t, _ := host.GetMCPTask(ctx, "build-server", taskID)
+if !t.Status.IsTerminal() {
+    _, _ = host.CancelMCPTask(ctx, "build-server", taskID)
+}
+```
+
+`Kit.ListMCPTasks`, `Kit.GetMCPTask`, and `Kit.CancelMCPTask` work against any
+loaded MCP server that advertises the corresponding capability.
+`MCPTaskStatus.IsTerminal()` is the canonical check for completion.
+
+Context cancellation also works end-to-end: cancelling the `ctx` passed to a
+tool execution triggers a best-effort `tasks/cancel` before the call returns.
+
+## Precedence
+
+For any given generation or provider field, the effective value is resolved
+in this order (highest priority first):
+
+1. `Options.X` (SDK caller)
+2. `KIT_X` environment variable
+3. `.kit.yml` (project-local then `~/.kit.yml`)
+4. Per-model defaults (`modelSettings[provider/model]` or `customModels[...].params`)
+5. Provider-level defaults (e.g. Anthropic's own temperature default)
+6. SDK last-resort floor (currently: `MaxTokens = 8192`, matching the CLI `--max-tokens` default)
+
+Sampling params that remain `nil` after the SDK resolution step are left out
+of the provider call entirely, so the LLM library applies its own default.
 
 ## Tool configuration
 

www/pages/sdk/overview.md

@@ -101,11 +101,34 @@ Return values:
 |--------|-------------|
 | `kit.TextResult(s)` | Successful text result |
 | `kit.ErrorResult(s)` | Error result (LLM sees it as a tool error) |
+| `kit.ImageResult(s, data, mediaType)` | Image result with binary data (e.g. `"image/png"`) |
+| `kit.MediaResult(s, data, mediaType)` | Non-image media result (e.g. `"audio/mpeg"`) |
 
-For advanced use, return a `kit.ToolOutput` struct directly with `Data`, `MediaType`, and `Metadata` fields.
+Binary data (images, audio, etc.) in `ToolOutput.Data` is automatically forwarded to the LLM when `MediaType` is set. For advanced use, return a `kit.ToolOutput` struct directly with `Data`, `MediaType`, and `Metadata` fields.
 
 Use `kit.NewParallelTool` for tools that are safe to run concurrently. Use `kit.ToolCallIDFromContext(ctx)` to retrieve the LLM-assigned call ID for logging or tracing.
 
+## Generation & provider overrides
+
+SDK consumers can configure generation parameters and provider endpoints
+entirely in-code via `Options`, without touching `.kit.yml` or `viper.Set()`:
+
+```go
+host, _ := kit.New(ctx, &kit.Options{
+    Model:          "anthropic/claude-sonnet-4-5-20250929",
+    MaxTokens:      16384,             // 0 = auto-resolve (env → config → per-model → floor)
+    ThinkingLevel:  "high",            // "off" | "none" | "minimal" | "low" | "medium" | "high"
+    Temperature:    ptrFloat32(0.2),   // nil = provider/per-model default
+    ProviderAPIKey: os.Getenv("MY_SECRET"), // overrides pre-existing viper state
+    ProviderURL:    "https://proxy.internal/v1",
+})
+
+func ptrFloat32(v float32) *float32 { return &v }
+```
+
+See [Options](/sdk/options#generation-parameters) for the full field reference,
+including `TopP`, `TopK`, `FrequencyPenalty`, `PresencePenalty`, and `TLSSkipVerify`.
+
 ## Event system
 
 Subscribe to events for monitoring:
@@ -120,7 +143,7 @@ host.OnToolResult(func(event kit.ToolResultEvent) {
     fmt.Println("Tool result:", event.Name)
 })
 
-host.OnStreaming(func(event kit.MessageUpdateEvent) {
+host.OnMessageUpdate(func(event kit.MessageUpdateEvent) {
     fmt.Print(event.Chunk)
 })
 ```
@@ -192,6 +215,33 @@ resources := host.ListMCPResources()
 content, _ := host.ReadMCPResource(ctx, "server", "file:///path")
 ```
 
+## MCP tasks (long-running tools)
+
+Kit advertises [MCP task support](https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/tasks)
+during `initialize`, so cooperating servers can return a `taskId` immediately
+and let Kit poll `tasks/get` / `tasks/result` until the operation completes.
+This avoids HTTP/SSE proxy timeouts on long tools and gives you clean
+cancellation via context.
+
+```go
+host, _ := kit.New(ctx, &kit.Options{
+    MCPTaskMode: map[string]kit.MCPTaskMode{
+        "build-server": kit.MCPTaskModeAlways,
+    },
+    MCPTaskProgress: func(p kit.MCPTaskProgress) {
+        log.Printf("%s: %s", p.TaskID, p.Status)
+    },
+})
+
+// Inspect / cancel in-flight tasks
+tasks, _ := host.ListMCPTasks(ctx, "build-server")
+_, _    = host.CancelMCPTask(ctx, "build-server", tasks[0].TaskID)
+```
+
+Defaults to `MCPTaskModeAuto` per server, so any existing MCP server keeps
+its previous synchronous behaviour. See [SDK options → MCP Tasks](/sdk/options#mcp-tasks)
+for the full surface.
+
 ## Context and compaction
 
 Monitor and manage context usage:

www/public/session/index.html

@@ -1968,35 +1968,42 @@ a:hover { text-decoration: underline; }
 
   function renderEditBody(input, result) {
     let html = '';
-    const oldText = input.old_text || '';
-    const newText = input.new_text || '';
+    const edits = input.edits || [];
 
     html += '<div class="tool-input">';
     html += '<div class="tool-input-label">Changes</div>';
-    html += '<div class="diff-display">';
 
-    // Render old text lines as removed
-    if (oldText) {
-      const oldLines = oldText.split('\n');
-      for (const line of oldLines) {
-        html += `<div class="diff-line removed">- ${escapeHtml(line)}</div>`;
+    for (const edit of edits) {
+      const oldText = edit.old_text || '';
+      const newText = edit.new_text || '';
+
+      html += '<div class="diff-display">';
+
+      // Render old text lines as removed
+      if (oldText) {
+        const oldLines = oldText.split('\n');
+        for (const line of oldLines) {
+          html += `<div class="diff-line removed">- ${escapeHtml(line)}</div>`;
+        }
       }
-    }
 
-    // Separator
-    if (oldText && newText) {
-      html += '<div class="diff-separator">───</div>';
-    }
-
-    // Render new text lines as added
-    if (newText) {
-      const newLines = newText.split('\n');
-      for (const line of newLines) {
-        html += `<div class="diff-line added">+ ${escapeHtml(line)}</div>`;
+      // Separator
+      if (oldText && newText) {
+        html += '<div class="diff-separator">───</div>';
       }
+
+      // Render new text lines as added
+      if (newText) {
+        const newLines = newText.split('\n');
+        for (const line of newLines) {
+          html += `<div class="diff-line added">+ ${escapeHtml(line)}</div>`;
+        }
+      }
+
+      html += '</div>';
     }
 
-    html += '</div></div>';
+    html += '</div>';
 
     if (result && result.is_error) {
       html += renderToolOutput(result.content, true);