Merge pull request #23 from mark3labs/fix/18-windows-session-dir-colon

fix(session): strip illegal characters from windows session dir (#18)

.kit/extensions/subagent-monitor.go

@@ -13,8 +13,6 @@
 // - No channels in maps (Yaegi panics on range over map[string]chan)
 // - All ctx.* calls guarded with nil checks
 // - Simple data structures only
-// - The extension runner serializes handler calls per-extension, so
-//   concurrent subagent events cannot race on this shared state.
 package main
 
 import (
@@ -45,8 +43,7 @@ const (
 )
 
 // ---------------------------------------------------------------------------
-// Package-level state — safe because the runner serializes all handler
-// invocations for the same extension (per-extension reentrant mutex).
+// Package-level state - all simple types
 // ---------------------------------------------------------------------------
 
 var (
@@ -285,8 +282,8 @@ func Init(api ext.API) {
 
 		submonPushWidget()
 
-		// Remove the entry — build a new slice to avoid aliasing bugs
-		newEntries := make([]*submonEntry, 0, len(submonEntries))
+		// Remove the entry immediately (no goroutine to avoid races)
+		newEntries := submonEntries[:0]
 		for _, en := range submonEntries {
 			if en.callID != e.ToolCallID {
 				newEntries = append(newEntries, en)

.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
+
+$@

README.md

@@ -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
 
@@ -162,6 +162,11 @@ mcpServers:
     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
@@ -626,6 +631,36 @@ 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:
@@ -646,7 +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
 
@@ -663,7 +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()

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

@@ -5,24 +5,24 @@ go 1.26.2
 require (
 	charm.land/bubbles/v2 v2.1.0
 	charm.land/bubbletea/v2 v2.0.6
-	charm.land/fantasy v0.19.0
+	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 v0.0.0-20260420095748-421e4a7fa8d7
+	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 v0.12.0
-	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 v0.49.0
+	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
@@ -37,21 +37,21 @@ require (
 	cloud.google.com/go/compute/metadata v0.9.0 // 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 v1.41.6 // indirect
-	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.9 // indirect
-	github.com/aws/aws-sdk-go-v2/config v1.32.16 // indirect
-	github.com/aws/aws-sdk-go-v2/credentials v1.19.15 // indirect
-	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.22 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.22 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.22 // indirect
-	github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.23 // indirect
-	github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.8 // indirect
-	github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.22 // indirect
-	github.com/aws/aws-sdk-go-v2/service/signin v1.0.10 // indirect
-	github.com/aws/aws-sdk-go-v2/service/sso v1.30.16 // indirect
-	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.20 // indirect
-	github.com/aws/aws-sdk-go-v2/service/sts v1.42.0 // indirect
-	github.com/aws/smithy-go v1.25.0 // 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
@@ -59,9 +59,9 @@ 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 v0.0.0-20260420102150-fe550f2efce5 // 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 v0.0.0-20260420102150-fe550f2efce5 // 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
@@ -69,30 +69,31 @@ require (
 	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 v0.3.15 // indirect
 	github.com/googleapis/gax-go/v2 v2.22.0 // indirect
 	github.com/gorilla/websocket v1.5.3 // indirect
-	github.com/kaptinlin/go-i18n v0.4.2 // indirect
-	github.com/kaptinlin/jsonpointer v0.4.18 // indirect
-	github.com/kaptinlin/jsonschema v0.7.8 // indirect
-	github.com/kaptinlin/messageformat-go v0.5.2 // 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 v0.276.0 // indirect
-	google.golang.org/genai v1.54.0 // indirect
-	google.golang.org/genproto/googleapis/rpc v0.0.0-20260420184626-e10c466a9529 // 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
 )
@@ -129,7 +130,7 @@ 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

go.sum

@@ -2,8 +2,8 @@ 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 v2.0.6 h1:UHN/91OyuhaOFGSrBXQ/hMZD8IO1Uc4BvHlgHXL2WJo=
 charm.land/bubbletea/v2 v2.0.6/go.mod h1:MH/D8ZLlN3op37vQvijKuU29g3rqTp+aQapURFonF9g=
-charm.land/fantasy v0.19.0 h1:fnNXkIJ/xcIW3sdVtWxjtQGpWWe8pDGhBCWSHkgbrd0=
-charm.land/fantasy v0.19.0/go.mod h1:V9cCIUMZB9g3Bq40aKEY8xBNzDd48EdfHp2OMS0uzWs=
+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=
@@ -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 v1.41.6 h1:1AX0AthnBQzMx1vbmir3Y4WsnJgiydmnJjiLu+LvXOg=
-github.com/aws/aws-sdk-go-v2 v1.41.6/go.mod h1:dy0UzBIfwSeot4grGvY1AqFWN5zgziMmWGzysDnHFcQ=
-github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.9 h1:adBsCIIpLbLmYnkQU+nAChU5yhVTvu5PerROm+/Kq2A=
-github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.9/go.mod h1:uOYhgfgThm/ZyAuJGNQ5YgNyOlYfqnGpTHXvk3cpykg=
-github.com/aws/aws-sdk-go-v2/config v1.32.16 h1:Q0iQ7quUgJP0F/SCRTieScnaMdXr9h/2+wze1u3cNeM=
-github.com/aws/aws-sdk-go-v2/config v1.32.16/go.mod h1:duCCnJEFqpt2RC6no1iK6q+8HpwOAkiUua0pY507dQc=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.15 h1:fyvgWTszojq8hEnMi8PPBTvZdTtEVmAVyo+NFLHBhH4=
-github.com/aws/aws-sdk-go-v2/credentials v1.19.15/go.mod h1:gJiYyMOjNg8OEdRWOf3CrFQxM2a98qmrtjx1zuiQfB8=
-github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.22 h1:IOGsJ1xVWhsi+ZO7/NW8OuZZBtMJLZbk4P5HDjJO0jQ=
-github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.22/go.mod h1:b+hYdbU+jGKfXE8kKM6g1+h+L/Go3vMvzlxBsiuGsxg=
-github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.22 h1:GmLa5Kw1ESqtFpXsx5MmC84QWa/ZrLZvlJGa2y+4kcQ=
-github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.22/go.mod h1:6sW9iWm9DK9YRpRGga/qzrzNLgKpT2cIxb7Vo2eNOp0=
-github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.22 h1:dY4kWZiSaXIzxnKlj17nHnBcXXBfac6UlsAx2qL6XrU=
-github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.22/go.mod h1:KIpEUx0JuRZLO7U6cbV204cWAEco2iC3l061IxlwLtI=
-github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.23 h1:FPXsW9+gMuIeKmz7j6ENWcWtBGTe1kH8r9thNt5Uxx4=
-github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.23/go.mod h1:7J8iGMdRKk6lw2C+cMIphgAnT8uTwBwNOsGkyOCm80U=
-github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.8 h1:HtOTYcbVcGABLOVuPYaIihj6IlkqubBwFj10K5fxRek=
-github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.8/go.mod h1:VsK9abqQeGlzPgUr+isNWzPlK2vKe9INMLWnY65f5Xs=
-github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.22 h1:PUmZeJU6Y1Lbvt9WFuJ0ugUK2xn6hIWUBBbKuOWF30s=
-github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.22/go.mod h1:nO6egFBoAaoXze24a2C0NjQCvdpk8OueRoYimvEB9jo=
-github.com/aws/aws-sdk-go-v2/service/signin v1.0.10 h1:a1Fq/KXn75wSzoJaPQTgZO0wHGqE9mjFnylnqEPTchA=
-github.com/aws/aws-sdk-go-v2/service/signin v1.0.10/go.mod h1:p6+MXNxW7IA6dMgHfTAzljuwSKD0NCm/4lbS4t6+7vI=
-github.com/aws/aws-sdk-go-v2/service/sso v1.30.16 h1:x6bKbmDhsgSZwv6q19wY/u3rLk/3FGjJWyqKcIRufpE=
-github.com/aws/aws-sdk-go-v2/service/sso v1.30.16/go.mod h1:CudnEVKRtLn0+3uMV0yEXZ+YZOKnAtUJ5DmDhilVnIw=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.20 h1:oK/njaL8GtyEihkWMD4k3VgHCT64RQKkZwh0DG5j8ak=
-github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.20/go.mod h1:JHs8/y1f3zY7U5WcuzoJ/yAYGYtNIVPKLIbp61euvmg=
-github.com/aws/aws-sdk-go-v2/service/sts v1.42.0 h1:ks8KBcZPh3PYISr5dAiXCM5/Thcuxk8l+PG4+A0exds=
-github.com/aws/aws-sdk-go-v2/service/sts v1.42.0/go.mod h1:pFw33T0WLvXU3rw1WBkpMlkgIn54eCB5FYLhjDc9Foo=
-github.com/aws/smithy-go v1.25.0 h1:Sz/XJ64rwuiKtB6j98nDIPyYrV1nVNJ4YU74gttcl5U=
-github.com/aws/smithy-go v1.25.0/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 v0.0.0-20260420095748-421e4a7fa8d7 h1:PbFxahSfyADcQOp+7WxbeqN3wX37KA/Rk+EXOW1xS9Q=
-github.com/charmbracelet/ultraviolet v0.0.0-20260420095748-421e4a7fa8d7/go.mod h1:3YdTxlnV/L0bQ3VN8WOSw8doF7LZV/xawUQ4MuAPDvo=
+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 v0.0.0-20260420102150-fe550f2efce5 h1:3ElWZRQqSRqML2P/r2TmuSkdXPMDI+Jg3f0bGA6Ekg4=
-github.com/charmbracelet/x/exp/charmtone v0.0.0-20260420102150-fe550f2efce5/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 v0.0.0-20260420102150-fe550f2efce5 h1:QqpW1CPNAnOpM3Nj0X7IT2IFlR90bLdAkO5+A3Hwbi4=
-github.com/charmbracelet/x/exp/slice v0.0.0-20260420102150-fe550f2efce5/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=
@@ -124,8 +124,8 @@ 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 v0.12.0 h1:GoIC6RrkPMBIVQ3ckSkl+bO/ERV/IRK6clBdZmx4Uf4=
-github.com/coder/acp-go-sdk v0.12.0/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=
@@ -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=
@@ -167,8 +167,8 @@ 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=
@@ -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 v0.4.2 h1:52gGOx4ZwbLEiOyDMNA1ax2WktKlrKsmV6Ydf9Tw3/I=
-github.com/kaptinlin/go-i18n v0.4.2/go.mod h1:IACLIi+sHn3pGyryFMiqr2N1CJry4OKFD0MAEneEVQk=
-github.com/kaptinlin/jsonpointer v0.4.18 h1:EDUXT4WKpOKguU7oaFv6VaNatN7uHFe6dEYHX0+OFxs=
-github.com/kaptinlin/jsonpointer v0.4.18/go.mod h1:ndmfvrqrEDSbV3F7yGaOuDvr29WrxYU1aqkvef9L2do=
-github.com/kaptinlin/jsonschema v0.7.8 h1:aHv28bYtfLfUXYI/10Phb1nvVyLXNz1lmu73vtKmlOY=
-github.com/kaptinlin/jsonschema v0.7.8/go.mod h1:cz7SK0jTHdabKdQp+SwBKKmOeZ55txuNo72Jx9Sbb2w=
-github.com/kaptinlin/messageformat-go v0.5.2 h1:E+D5oQVRepHgyMiLWRHnPXYFbqBDI4Sek7/CTIAByj4=
-github.com/kaptinlin/messageformat-go v0.5.2/go.mod h1:NKjwS6e9u7DRhAK+vydjDDwJ7UbdHhYjk/yk2WPuZPs=
+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 v0.49.0 h1:7Ssx4d7/T86qnWoJIdye7wEEvUzv39UIbnZb/FqUZMY=
-github.com/mark3labs/mcp-go v0.49.0/go.mod h1:BflTAZAzXlrTpiO44gmjMu89n2FO56rJ9m31fp4zd5k=
-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 v0.276.0 h1:nVArUtfLEihtW+b0DdcqRGK1xoEm2+ltAihyztq7MKY=
-google.golang.org/api v0.276.0/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 v0.0.0-20260420184626-e10c466a9529 h1:XF8+t6QQiS0o9ArVan/HW8Q7cycNPGsJf6GA2nXxYAg=
-google.golang.org/genproto/googleapis/rpc v0.0.0-20260420184626-e10c466a9529/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

@@ -185,6 +185,26 @@ func (a *Agent) ListSessions(_ context.Context, _ acp.ListSessionsRequest) (acp.
 	}, 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.CloseSessionResponse{}, nil
+	}
+
+	log.Debug("acp: close session", "session", sessionID)
+	sess.cancelPrompt()
+	a.registry.remove(sessionID)
+	return acp.CloseSessionResponse{}, nil
+}
+
+// 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) {

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.
@@ -126,6 +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.
@@ -160,6 +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{}
@@ -258,6 +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.
@@ -277,6 +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{})
 
@@ -423,13 +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, 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,
@@ -444,6 +534,31 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 	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
 	// fantasy agent with the full tool set. This is a no-op when no MCP
@@ -451,13 +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.
@@ -477,9 +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 ||
-		onToolCallStart != nil || onToolCallDelta != nil || onToolCallEnd != 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
@@ -488,9 +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{
@@ -503,8 +624,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onToolCallStart != nil {
-					onToolCallStart(id, toolName)
+				if cb.OnToolCallStart != nil {
+					cb.OnToolCallStart(id, toolName)
 				}
 				return nil
 			},
@@ -512,8 +633,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onToolCallDelta != nil {
-					onToolCallDelta(id, delta)
+				if cb.OnToolCallDelta != nil {
+					cb.OnToolCallDelta(id, delta)
 				}
 				return nil
 			},
@@ -521,8 +642,39 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onToolCallEnd != nil {
-					onToolCallEnd(id)
+				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
 			},
@@ -532,8 +684,8 @@ func (a *Agent) GenerateWithLoopAndStreaming(ctx context.Context, messages []fan
 				if ctx.Err() != nil {
 					return ctx.Err()
 				}
-				if onReasoningDelta != nil {
-					onReasoningDelta(delta)
+				if cb.OnReasoningDelta != nil {
+					cb.OnReasoningDelta(delta)
 				}
 				return nil
 			},
@@ -543,8 +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
 			},
@@ -554,8 +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
 			},
@@ -568,13 +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
@@ -586,14 +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
@@ -607,8 +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)
 				}
 
@@ -618,65 +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
@@ -702,8 +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)
@@ -723,8 +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
@@ -915,6 +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()
 		})

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

@@ -923,7 +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
@@ -1241,7 +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",
@@ -1262,11 +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

@@ -534,9 +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)
@@ -547,7 +547,7 @@ func TestRecordStepUsage_updatesTracker(t *testing.T) {
 		OutputTokens:     45,
 		CacheReadTokens:  5,
 		CacheWriteTokens: 2,
-	}, nil)
+	}, nil, nil)
 
 	usage.mu.Lock()
 	defer usage.mu.Unlock()
@@ -559,9 +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)
 	}
 }
 

internal/app/events.go

@@ -210,6 +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/config/config.go

@@ -38,6 +38,23 @@ type MCPServerConfig struct {
 	// 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.
@@ -68,6 +85,7 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		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
@@ -80,6 +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
@@ -96,6 +115,7 @@ func (s *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		s.OAuthClientSecret = newConfig.OAuthClientSecret
 		s.OAuthScopes = newConfig.OAuthScopes
 		s.NoOAuth = newConfig.NoOAuth
+		s.TasksMode = newConfig.TasksMode
 		return nil
 	}
 
@@ -116,6 +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
@@ -324,6 +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/extensions/api.go

@@ -1094,6 +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.
@@ -1301,6 +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
@@ -2253,6 +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

@@ -96,6 +96,35 @@ 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.
@@ -109,6 +138,8 @@ func AllEventTypes() []EventType {
 		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) != 24 {
-		t.Fatalf("expected 24 event types, got %d", len(all))
+	if len(all) != 32 {
+		t.Fatalf("expected 32 event types, got %d", len(all))
 	}
 }
 

internal/extensions/loader.go

@@ -618,6 +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/symbols.go

@@ -172,6 +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/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/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_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
 
@@ -374,6 +391,7 @@ func OpenTreeSession(path string) (*TreeManager, error) {
 		return nil, fmt.Errorf("failed to open session file for append: %w", err)
 	}
 	tm.file = f
+	tm.writer = bufio.NewWriter(f)
 
 	return tm, nil
 }
@@ -427,6 +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
@@ -451,6 +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
@@ -465,6 +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
@@ -479,6 +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
@@ -494,6 +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
@@ -510,6 +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
@@ -541,6 +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
@@ -926,11 +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
@@ -1090,13 +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.
@@ -1148,12 +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
@@ -1274,15 +1350,44 @@ func (tm *TreeManager) buildTreeNodeDepth(id string, depth int, visited map[stri
 // --- 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/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
@@ -262,7 +263,9 @@ 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 oauthEnabled && IsOAuthError(err) {
 			if flowErr := p.oauthFlow.RunAuthFlow(ctx, serverName, err); flowErr != nil {
@@ -270,7 +273,7 @@ func (p *MCPConnectionPool) createConnection(ctx context.Context, serverName str
 				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
 			}
@@ -280,15 +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))
@@ -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/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/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, width, 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 {
@@ -338,7 +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

internal/ui/model.go

@@ -1881,6 +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
@@ -2072,6 +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
@@ -3696,6 +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 {

internal/ui/render/blocks.go

@@ -45,7 +45,7 @@ func UserBlock(content string, width int, ty *herald.Typography, theme style.The
 // 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)
+	accentStyle := style.GetCachedStyles().FileTokenAccent
 	return fileTokenPattern.ReplaceAllStringFunc(text, func(token string) string {
 		return accentStyle.Render(token)
 	})
@@ -75,8 +75,8 @@ func ReasoningBlock(content string, duration int64, width int, ty *herald.Typogr
 	if width > 4 {
 		contentStr = wrapText(contentStr, width-4)
 	}
-	mutedStyle := lipgloss.NewStyle().Foreground(theme.Muted)
-	contentRendered := mutedStyle.Render(ty.Italic(contentStr))
+	cs := style.GetCachedStyles()
+	contentRendered := cs.Muted.Render(ty.Italic(contentStr))
 
 	// Build label based on duration
 	if duration > 0 {
@@ -86,14 +86,14 @@ func ReasoningBlock(content string, duration int64, width int, ty *herald.Typogr
 		} 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)
@@ -194,7 +194,7 @@ 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

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)
@@ -476,9 +475,8 @@ func (s *StreamComponent) renderReasoningBlock(reasoning string) string {
 	if s.width > 4 {
 		content = lipgloss.NewStyle().Width(s.width - 4).Render(content)
 	}
-	theme := GetTheme()
-	mutedStyle := lipgloss.NewStyle().Foreground(theme.Muted)
-	parts = append(parts, mutedStyle.Render(s.ty.Italic(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
@@ -494,8 +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)
 	}
 

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

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:

pkg/kit/events.go

@@ -58,6 +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"
 )
 
 // ---------------------------------------------------------------------------
@@ -379,6 +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 {
@@ -517,7 +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)
@@ -555,6 +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

@@ -356,4 +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

@@ -66,6 +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
@@ -731,7 +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 == "" {
@@ -1034,6 +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
 
@@ -1368,6 +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
@@ -1385,6 +1422,14 @@ 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. The SDK does not create
@@ -1461,6 +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.
@@ -1781,14 +1827,28 @@ func (m *Kit) Subagent(ctx context.Context, cfg SubagentConfig) (*SubagentResult
 
 	// 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)
@@ -1894,21 +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),
@@ -1922,17 +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>"
@@ -1968,14 +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,
@@ -1984,18 +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,
@@ -2009,37 +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 — fire as the LLM generates tool arguments
-		func(toolCallID, toolName string) {
+		// Tool call argument streaming
+		OnToolCallStart: func(toolCallID, toolName string) {
 			m.events.emit(ToolCallStartEvent{
 				ToolCallID: toolCallID,
 				ToolName:   toolName,
 				ToolKind:   toolKindFor(toolName),
 			})
 		},
-		func(toolCallID, delta string) {
+		OnToolCallDelta: func(toolCallID, delta string) {
 			m.events.emit(ToolCallDeltaEvent{
 				ToolCallID: toolCallID,
 				Delta:      delta,
 			})
 		},
-		func(toolCallID string) {
+		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:

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/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.

skills/kit-sdk/SKILL.md

@@ -281,7 +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
 })
 
@@ -296,8 +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
@@ -336,6 +392,16 @@ 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`
@@ -421,6 +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
@@ -493,6 +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):
 
@@ -1088,13 +1170,30 @@ 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
@@ -1168,7 +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")

www/pages/configuration.md

@@ -88,6 +88,11 @@ mcpServers:
     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
@@ -101,9 +106,34 @@ mcpServers:
 | `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:

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()
@@ -116,6 +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
@@ -128,19 +146,41 @@ Lower values run first. First non-nil result wins.
 
 ## All event types
 
-| Event | Description |
-|-------|-------------|
-| `ToolCallStartEvent` | LLM began generating tool call arguments (tool name known, args streaming) |
-| `ToolCallDeltaEvent` | Streamed JSON fragment of tool call arguments |
-| `ToolCallEndEvent` | Tool argument streaming complete, before execution begins |
-| `ToolCallEvent` | Tool call fully parsed and about to execute |
-| `ToolResultEvent` | Tool execution completed with result |
-| `ToolOutputEvent` | Streaming output chunk from tool (e.g., bash stdout/stderr) |
-| `MessageUpdateEvent` | Streaming text chunk from LLM |
-| `ResponseEvent` | Final response received |
-| `TurnStartEvent` | Agent turn started |
-| `TurnEndEvent` | Agent turn completed |
-| `PasswordPromptEvent` | Sudo command needs password (respond via `ResponseCh`) |
+| 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
 

www/pages/sdk/options.md

@@ -169,6 +169,12 @@ when embedding Kit as a library.
 | `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
 
@@ -248,6 +254,79 @@ 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

www/pages/sdk/overview.md

@@ -101,8 +101,10 @@ 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.
 
@@ -141,7 +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)
 })
 ```
@@ -213,6 +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);