Skills system:
- Replace `CommandBadge.isSkill: boolean` with `kind: SkillKind` (agent_skill | slash_command | sdk_command | codepilot_command)
- handleSubmit dispatches by kind: agent_skill retains skill name in prompt, slash/sdk commands pass through to SDK, codepilot commands expand via COMMAND_PROMPTS
- Popover groups commands into Commands / Slash Commands / Agent Skills
- API `/api/skills` returns `kind` field; `displayOverride` persisted through full chain (frontend → API → DB)
- Forward SDK `system:init` slash_commands/skills metadata to frontend via SSE
MCP management:
- Fix remote server (SSE/HTTP) creation: validate `url` instead of `command`
- Fix dual-config read/write asymmetry: PUT routes servers by `_source` to correct file (~/.claude.json vs ~/.claude/settings.json)
- DELETE checks both config files; POST checks both for name collision
- McpManager preserves `_source` tags on edit/rename; JSON editor only manages settings.json servers
Completion processing:
- Extract core logic into `onboarding-processor.ts` and `checkin-processor.ts` (shared by route handlers and server-side detection)
- Remove frontend `detectAssistantCompletion` from ChatView — server-side path in `finally` block is now the single source of truth
- Replace hard-coded `fetch("http://localhost:PORT/...")` with direct function imports
- Both processors are internally idempotent (check state before processing)
Other:
- `settingSources` always includes 'user' so SDK loads ~/.claude/skills/
- Add 24 unit tests for SkillKind dispatch logic
- Add onboarding-completion extraction tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add "Assistant Workspace" feature that lets users designate a directory
(e.g. Obsidian vault) containing 4 Markdown files (soul.md, user.md,
memory.md, claude.md) to define AI personality, user profile, long-term
memory, and execution rules. These files are automatically injected into
the system prompt for assistant project sessions.
Core library (src/lib/assistant-workspace.ts):
- Workspace validation, initialization with template files
- File loading with 8000-char truncation, XML prompt assembly (40k limit)
- State management via .assistant/state.json
- Directory docs generation (README.ai.md + PATH.ai.md)
- Case-insensitive file name fallback (soul.md → SOUL.md)
System prompt injection (src/app/api/chat/route.ts):
- Scoped to sessions where working_directory === workspace path
- Workspace prompt prepended before session-level system prompt
- Onboarding/check-in instructions injected as <assistant-project-task>
- autoTrigger flag skips user message save and title update
Chat-based onboarding (10 questions) & daily check-in (3 questions):
- AI conducts conversation, outputs structured completion code blocks
- Frontend detects onboarding-complete/checkin-complete blocks
- Backend calls Claude API to generate file content from answers
- Fallback to raw answers if API call fails
Auto-trigger pipeline:
- Mount trigger (500ms delay) + focus fallback in MessageInput
- assistantTriggerFiredRef prevents duplicate triggers per mount
- hookTriggeredSessionId in state.json prevents cross-reload duplicates
- Daily check-in only triggers in the most recent workspace session
- hookTriggeredSessionId cleared via __clear__ sentinel after completion
API endpoints:
- GET/PUT /api/settings/workspace — path CRUD + file status
- POST /api/workspace/session — create (onboarding) or reuse (check-in)
- POST /api/workspace/onboarding — AI-generate soul.md + user.md
- POST /api/workspace/checkin — AI-update memory.md + user.md
- POST /api/workspace/hook-triggered — set/clear trigger guard
- GET /api/workspace/latest-session — latest session lookup
- POST /api/workspace/docs — refresh directory documentation
UI:
- Settings "Assistant" tab with workspace path, file status, controls
- Violet badge on workspace project group in ChatListPanel
- useAssistantWorkspace hook for state queries
Also includes:
- 102 passing tests (16 new regression tests for workspace logic)
- i18n keys for en.ts and zh.ts
- Handover documentation and execution plan
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add "View Plan" button with fullscreen modal to ExitPlanMode UI
When Claude writes a plan in plan mode, the plan content is written to a
file on disk but the ExitPlanMode UI only shows Approve/Reject buttons —
the user never sees the actual plan without manually finding the file.
Add a "View Plan" button that opens a fullscreen Dialog modal showing
the plan content rendered as markdown. The plan content is extracted from
the tool uses array (the Write tool's input for the plan file), so no
new API endpoints are needed.
Changes in StreamingMessage.tsx:
- extractPlanContent(): walks toolUses backwards to find the last Write
to a plans/*.md file and returns its content
- ExitPlanModeUI: accepts toolUses prop, renders "View Plan" button
(hidden if no plan content found), opens Radix Dialog modal with
MessageResponse markdown renderer
- Modal is 80vh tall, max-w-4xl, scrollable, with Close button
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add "Do this instead" feedback textbox to plan approval UI
When rejecting a plan, users can now type feedback and click "Do this
instead" (or press Enter) to send a custom deny message to the SDK.
This matches the Claude Code CLI behavior where users can provide
alternative instructions when rejecting a plan.
Previously, the deny message was hardcoded to "User denied permission"
— the API already supported custom messages but the client never sent
them.
Changes:
- StreamingMessage.tsx: add feedback textbox + "Do this instead" button
to ExitPlanModeUI, wire onDenyWithMessage callback
- ChatView.tsx: pass denyMessage through handlePermissionResponse
- MessageList.tsx: update onPermissionResponse type signature
- stream-session-manager.ts: add denyMessage param to
respondToPermission(), use it in the deny decision body
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: fetch plan content from disk so View Plan always shows latest version
Previously extractPlanContent read the Write tool's input.content, which
showed stale content when Claude later edited the plan via the Edit tool.
Now extractPlanFilePath finds the plan file path (matching both Write and
Edit tools, with Windows backslash support), and the View Plan button
fetches the file from disk on each click via /api/files/preview.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The stream idle timer (330s) fires during long-running Claude Code tool
executions because the SDK's keep_alive heartbeat messages are silently
dropped. This causes a false "Stream idle timeout" error even though
Claude is still actively working.
Root cause: claude-client.ts only handles 6 SDK message types in its
switch statement and has no default case, so keep_alive messages from
the SDK are never forwarded to the frontend. The idle timer in
stream-session-manager.ts never gets reset by markActive().
Fix: thread keep_alive through the existing 3-layer event pipeline:
- claude-client.ts: forward keep_alive as SSE event (via default case)
- useSSEStream.ts: add onKeepAlive callback to SSECallbacks interface
- stream-session-manager.ts: call markActive() on keep_alive to reset
the idle timer
No visible UI change — keep_alive only resets the timer silently.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
