Desktop sessions could show MCP configuration in the UI while the CLI SDK child process still started cold on the first user message. Start eligible desktop chat sessions as soon as their websocket is connected, keep init metadata cached but muted, and wait briefly for MCP startup in SDK print mode so turn one sees the tools.
Constraint: Desktop wraps the existing CLI SDK bridge rather than owning MCP startup directly.
Rejected: Prewarm every restored tab | synthetic tabs such as settings would start unnecessary CLI subprocesses.
Confidence: high
Scope-risk: moderate
Directive: Do not remove the synthetic-tab guard without validating restored settings and scheduled-task tabs in the browser.
Tested: bun test src/server/__tests__/conversations.test.ts src/server/__tests__/conversation-service.test.ts src/server/__tests__/mcp.test.ts
Tested: cd desktop && bun run test -- chatStore.test.ts mcpSettings.test.tsx
Tested: cd desktop && bun run lint
Tested: cd desktop && bun run build
Tested: agent-browser E2E new session prewarm plus first message reuse with mock CLI
Not-tested: Full desktop Vitest suite still has pre-existing English-vs-Chinese copy assertion failures.
The desktop model picker now stores a session-scoped provider/model selection instead of relying on the global active provider. That selection is replayed on connect, passed into the CLI startup path, and preserved across turns until the user changes it again.
To make that true end-to-end, the server now restarts the session process when runtime selection changes, injects provider-scoped env for third-party providers, and routes proxy traffic by provider id. The selector UI was also tightened so provider grouping stays visible while the actual model choice remains readable.
Constraint: Different providers can expose the same model id, so chat runtime selection cannot be derived from model id alone
Constraint: A desktop session reuses one CLI subprocess across turns, so runtime changes must restart that process to take effect
Rejected: Keep using Settings active provider as the chat selector | conflates defaults with live session state and breaks overlapping models
Rejected: UI-only runtime switching without server restart | later turns would continue using the old CLI subprocess configuration
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep provider defaults and session runtime overrides separate, and preserve provider-scoped proxy routing when extending model selection surfaces
Tested: cd desktop && bun run lint
Tested: cd desktop && bun run test src/stores/chatStore.test.ts
Tested: cd desktop && bun run test src/__tests__/generalSettings.test.tsx
Tested: bun test src/server/__tests__/conversation-service.test.ts
Tested: bun test src/server/__tests__/conversations.test.ts
Tested: bun -e "await import('./src/server/services/titleService.ts'); await import('./src/server/ws/handler.ts')"
Not-tested: Real third-party provider round-trip from the desktop UI against a live upstream account
The provider picker had two separate preset definitions and some defaults had
started to drift from the model IDs vendors document for Claude Code /
Anthropic-compatible usage. This change moves built-in presets to one
server-side JSON source, serves that source through the existing presets API,
and makes the desktop settings page consume the API instead of a duplicated
frontend constant. The preset defaults were then corrected against vendor docs
so the built-in values match documented model IDs and casing.
Constraint: Built-in provider defaults must match official vendor Claude Code or Anthropic docs
Rejected: Keep duplicated frontend and backend preset lists | values and casing drifted independently
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Update src/server/config/providerPresets.json and rerun provider-presets tests before changing built-in provider defaults again
Tested: bun test src/server/__tests__/provider-presets.test.ts; cd desktop && bun run lint
Not-tested: Manual desktop settings UI interaction after fetching presets from the API
Desktop MCP settings was synchronously probing every configured server from /api/mcp, which made the list page scale with connection latency and surface brittle behavior as installs accumulated more MCPs. Return lightweight snapshot rows from the list API, keep explicit status checks separate, and let the desktop UI refresh status in a constrained background lane while preserving project-aware server identity.
Constraint: MCP list must stay responsive even with many configured servers
Rejected: Probe all servers from the list view without limits | still fans out with server count and can overload slow installs
Rejected: Keep servers permanently unchecked until detail view | misses the desired loading feedback when the MCP page opens
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep /api/mcp as a configuration snapshot endpoint; do not reintroduce per-row live connect work on list load
Tested: bun test src/server/__tests__/mcp.test.ts; cd desktop && bun run test -- mcpSettings.test.tsx; cd desktop && bun x tsc --noEmit --ignoreDeprecations 5.0
Not-tested: Real desktop interaction against an environment with dozens of live MCP servers
Desktop plugin details now route into the shared Skills, Agents, and MCP management surfaces instead of maintaining separate read-only drilldowns. This also extends the desktop/server skill aggregation so plugin-provided skills appear in the shared list, groups MCP entries by source, and preserves detail-view back navigation based on where the user entered the page.
The implementation keeps plugin detail as the high-level capability hub while pushing real inspection into the existing management pages. Disabled plugins no longer expose false navigation paths into shared views, and the agent-browser regression script was expanded to exercise the new end-to-end flows.
Constraint: Shared Agents data only includes enabled plugin agents, so disabled plugins cannot deep-link into agent detail
Rejected: Keep duplicating full Skills/Agents/MCP detail inside Plugin detail | creates divergent UI flows and stale data paths
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: If a detail view can be opened from multiple entry points, keep the return target in store state rather than hardcoding a single back destination
Tested: desktop vitest for plugins/skills/agents/mcp; desktop tsc --noEmit; desktop vite build; server skills API test; agent-browser web regression on plugin->skill/mcp and plugin->agent back navigation
Not-tested: packaged desktop app regression after rebuilding the Tauri bundle
Merged the desktop MCP management work into local main and retained
existing plugin settings affordances while resolving router and settings-tab
conflicts. The resulting flow keeps MCP settings global-only for speed and
uses slash-command cards to route users into concrete MCP or skill targets.
Constraint: local main already had plugin settings and API routes that had to remain available
Rejected: Favor the worktree version wholesale during merge | would have dropped local plugin tab and router support on main
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Keep slash-command entry flows and settings tabs additive during future merges; do not collapse MCP and plugin navigation into one another
Tested: Conflict resolution review during merge; source commit 66edded validated with bun test src/server/__tests__/mcp.test.ts; cd desktop && bun run test -- pages.test.tsx mcpSettings.test.tsx; cd desktop && bun x tsc --noEmit --ignoreDeprecations 5.0
Not-tested: Re-running full desktop/manual verification from the merged main worktree after merge completion
Related: 66edded
Desktop MCP management now has a working server API, a global-only settings surface,
and slash-command entry points that surface MCP and skills from the composer before
routing users into the right settings view.
Constraint: Project-scoped MCP browsing in settings was too slow and noisy because it scanned multiple workdirs
Rejected: Keep project MCP aggregation on the settings homepage | duplicated entries and poor responsiveness
Rejected: Route /mcp directly on Enter without an intermediate card | removed the user's ability to choose a specific target first
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Keep settings focused on global MCP; add project-scoped MCP affordances in the chat-context slash surfaces instead of re-expanding the settings homepage
Tested: bun test src/server/__tests__/mcp.test.ts; cd desktop && bun run test -- pages.test.tsx mcpSettings.test.tsx; cd desktop && bun x tsc --noEmit --ignoreDeprecations 5.0
Not-tested: Manual IAB verification after this final commit/merge cycle
The desktop app could read plugin-produced skills and agents, but it had no
plugin control plane of its own. This adds a dedicated Settings tab backed by
server-side plugin APIs so installed plugins can be inspected, enabled,
disabled, updated, reloaded, and uninstalled from the WebUI.
The implementation also teaches browser-based desktop dev sessions to honor a
custom backend URL, which made it possible to run isolated worktree ports for
real UI automation. During verification, the long-lived desktop server kept a
stale installed-plugin snapshot after external CLI mutations, so cache clearing
now resets that session-level plugin installation state as well.
Constraint: Desktop WebUI needed an isolated backend URL instead of the hard-coded 127.0.0.1:3456 fallback
Constraint: Reuse existing plugin operations and loaders instead of rebuilding plugin lifecycle logic in the desktop layer
Rejected: Fold plugin management into Skills or Adapters | mixed unrelated lifecycles and hid plugin-specific health/actions
Rejected: Expose only read-only plugin status in desktop | did not satisfy enable-disable-reload-uninstall verification needs
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep desktop plugin actions routed through the shared plugin operation layer and clear installed-plugin session caches when plugin state changes externally
Tested: cd desktop && bun run lint
Tested: cd desktop && bun run test -- src/__tests__/pluginsSettings.test.tsx
Tested: bun test src/server/__tests__/plugins.test.ts src/server/__tests__/skills.test.ts
Tested: Browser automation against isolated ports 15120/38456 covering discord plugin list/detail/disable/apply/enable/update/uninstall flows
Not-tested: Full desktop session runtime parity with CLI /reload-plugins AppState refresh beyond the new desktop API path
Desktop chat was rendering AskUserQuestion twice: once as an inline question card and again as a generic permission request. That split also dropped the structured answers on the floor because the websocket permission_response shape only carried allow/deny state.
This change keeps AskUserQuestion on the permission pipeline end-to-end. The desktop websocket contract now carries toolUseId and updatedInput, AskUserQuestion submits answers through permission_response, and the generic permission card is suppressed for that tool so the user sees a single question flow.
Constraint: AskUserQuestion answers must round-trip through updatedInput.answers for the CLI tool contract to complete
Rejected: Leave AskUserQuestion as a plain chat reply in desktop | the tool never receives structured answers and the pending approval UI remains stuck
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep AskUserQuestion bound to the permission-response path unless the desktop protocol grows a separate structured elicitation channel
Tested: desktop lint; vitest src/components/chat/AskUserQuestion.test.tsx src/stores/chatStore.test.ts; bun test src/server/__tests__/conversations.test.ts
Not-tested: Manual desktop click-through against a live plan-mode session after bundling
Desktop sessions were failing before the actual fetch because Anthropic's domain preflight can be unreachable on restricted networks, and the next runtime path was missing turndown for HTML-to-Markdown conversion.
This change defaults desktop sessions to skip the preflight unless the user explicitly overrides it, exposes that behavior as a desktop General setting, seeds new settings JSON with the desktop-safe default, and adds regression coverage for both the runtime default and the UI toggle. It also adds the missing turndown dependency so successful fetches can continue through HTML reduction instead of failing at module resolution.
Constraint: Desktop must keep an escape hatch for users who want upstream preflight restored explicitly
Rejected: Force skipWebFetchPreflight globally for every session | would silently change CLI and non-desktop behavior
Rejected: UI-only toggle without runtime default | existing desktop users would still fail until they manually opened settings
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep desktop-specific WebFetch behavior scoped to desktop session detection and explicit user settings; do not broaden it to general CLI flows without separate validation
Tested: bun test src/tools/WebFetchTool/utils.test.ts; cd desktop && bun run lint; cd desktop && bunx vitest run src/__tests__/generalSettings.test.tsx; runtime import verification for turndown via node
Not-tested: End-to-end desktop packaging smoke test against a freshly built DMG/app bundle
Desktop sessions were missing a visible request_access approval path and could
mis-detect their own app window as an unapproved frontmost target, which caused
Computer Use clicks to fail even after opening the intended app. On macOS, text
entry was also split across inconsistent clipboard and keystroke paths, making
Electron inputs unreliable for Chinese and short strings.
This change adds a desktop approval bridge over the existing session websocket,
renders a dedicated desktop approval modal, threads the real desktop bundle id
into the Computer Use executor, and switches macOS clipboard typing onto the
native pasteboard plus system paste shortcut path. It also makes tool error
results expandable in the desktop chat UI so frontmost-gate failures are fully
visible during debugging.
Constraint: Desktop sessions run the CLI over the SDK websocket path, so Ink tool JSX dialogs are not visible there
Constraint: macOS IME and Electron text inputs are unreliable with pyautogui.write and generic hotkey synthesis
Rejected: Reuse CLI setToolJSX dialogs in desktop mode | no transport for mid-call Ink UI over the SDK bridge
Rejected: Keep shell pbcopy/pbpaste for clipboard typing | inconsistent with NSPasteboard path and less reliable for Chinese text
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Keep desktop Computer Use approvals and macOS text-entry behavior on a single bridge/path; avoid reintroducing separate CLI-only and desktop-only codepaths for the same action
Tested: python3 -m unittest runtime/test_helpers.py
Tested: bun test src/utils/computerUse/permissions.test.ts src/server/__tests__/conversation-service.test.ts
Tested: cd desktop && bun run test ComputerUsePermissionModal chatStore
Tested: cd desktop && bun run test chatBlocks
Tested: cd desktop && bun run lint
Not-tested: End-to-end manual Computer Use interaction against a live Electron target app on macOS
This introduces a persisted light/dark appearance setting, maps the desktop shell onto semantic theme tokens, and reworks the highest-traffic chat/settings surfaces so the new dark mode is usable without regressing the original light theme.
The same pass tightens markdown rendering for chat replies by improving inline code, table overflow handling, and safe external-link behavior so dark-mode content stays legible in real conversations.
Constraint: Preserve the existing light theme while adding a user-selectable dark theme in Settings > General
Constraint: Avoid introducing new dependencies for styling or markdown handling
Rejected: Replacing the light palette with a single dual-purpose palette | would risk broad visual regressions across the existing desktop UI
Rejected: Implementing dark mode only for shell chrome | leaves chat markdown, diffs, and permission flows visually broken
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: New desktop UI should use semantic theme variables instead of hard-coded color literals so both themes stay aligned
Tested: bun run lint; bun run test; bun run build; browser review of theme switching, provider/permission states, and chat surfaces
Not-tested: Prototype-style pages with remaining hard-coded colors (scheduled/session control mock surfaces) were not fully normalized in this change
This bundles the pending desktop/server team-session fixes with the local adapter recovery changes already in the worktree. The team path now keeps teammate membership stable under concurrent spawns, surfaces real teammate identities in the desktop UI, and allows direct interaction with member transcripts. The adapter changes recover automatically when stale thinking signatures invalidate an existing session.
Constraint: Team config writes can happen concurrently while multiple reviewers spawn in parallel
Constraint: Desktop member views must follow mailbox/transcript semantics rather than hijacking teammate runtime sessions
Rejected: Keep relying on config.json alone for member discovery | in-process teammates can be lost after concurrent writes
Rejected: Open teammate sessionIds as normal desktop sessions | would attach a second CLI instead of the running teammate
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Preserve locked team-file mutation for any future teammate registration path and keep teammate labels sourced from member names before agent types
Tested: bun test src/server/__tests__/teams.test.ts src/server/__tests__/team-watcher.test.ts
Tested: cd desktop && bun run test --run src/stores/chatStore.test.ts src/pages/ActiveSession.test.tsx
Tested: cd desktop && bun run lint
Tested: cd desktop && bun run build
Not-tested: Manual end-to-end validation against a live Agent Teams run in the desktop app
This folds together the desktop-side fixes needed before broader rollout.
Session resume no longer deadlocks waiting on init, Mermaid and inline image
output render inside chat, task and sub-agent state stay visible during
execution, local build/release paths are safer, and Feishu/Telegram now expose
lightweight mobile commands (/help, /status, /clear) without adding a new
adapter-specific protocol.
Constraint: Desktop releases must publish updater artifacts from non-draft GitHub releases
Constraint: IM commands need short, phone-friendly responses and low operational complexity
Rejected: Add a dedicated IM command API surface | re-used existing slash commands and session/task REST endpoints to keep adapters thin
Rejected: Wait for task_update push events in WebUI | added low-risk polling because the current frontend ignores that event path
Confidence: medium
Scope-risk: broad
Reversibility: clean
Directive: Keep IM command replies terse and mobile-first, and merge local fallback slash commands when server-provided lists are partial
Tested: cd desktop && bun x vitest run src/components/chat/MermaidRenderer.test.tsx src/components/markdown/MarkdownRenderer.test.tsx
Tested: cd desktop && bun x vitest run src/components/chat/composerUtils.test.ts src/pages/ActiveSession.test.tsx src/stores/chatStore.test.ts
Tested: cd desktop && bun run lint
Tested: bun test src/server/__tests__/conversations.test.ts --test-name-pattern "SDK init arrives only after the first user turn" --timeout 60000
Tested: cd adapters && bun test common/ feishu/ telegram/
Tested: cd adapters && bunx tsc --noEmit
Not-tested: Full GitHub Actions release run on all three desktop platforms
Not-tested: Local DMG packaging end-to-end on Apple Silicon
Not-tested: Real Feishu/Telegram device sessions against a live adapter process
Add a protocol-translating reverse proxy that allows using OpenAI-compatible
API providers (DeepSeek, OpenRouter, Groq, etc.) with Claude Code.
The proxy intercepts Anthropic Messages API requests from the CLI, transforms
them to OpenAI Chat Completions or Responses API format, forwards to the
upstream provider, and transforms streaming/non-streaming responses back.
Key features:
- Request transform: Anthropic Messages → OpenAI Chat/Responses
- Response transform: OpenAI → Anthropic (streaming SSE + non-streaming)
- Provider-agnostic reasoning support (reasoning_content, thinking_blocks,
reasoning fields from DeepSeek, OpenAI o-series, GLM-5, Groq, etc.)
- Event queue pattern for correct Anthropic SSE event ordering
- Two-step test: ① connectivity check ② full proxy pipeline validation
- Desktop UI: API format selector, two-step test results display
- License attribution for cc-switch (MIT, Jason Young)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Improve the desktop Skills browser so SKILL.md metadata renders cleanly and the settings view uses space like a real document browser. Add coverage for the new detail, markdown, and i18n behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Server: GET /api/skills (list) and GET /api/skills/detail (tree + files)
- Desktop: type definitions, API client, Zustand store
- i18n: EN/ZH translation keys for Skills tab
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The desktop chat view flattened Agent tool activity, which made sub-agent work
hard to follow and separated key evidence from the main conversation. This
change threads parent tool linkage through the server bridge and desktop store,
renders dispatched sub-agents as grouped cards with nested tool activity, and
moves long final outputs into a markdown preview dialog so the main transcript
stays readable on narrow layouts.
Constraint: Existing sessions and live websocket events both needed to preserve parent-child relationships
Rejected: Add brand-new subagent websocket event types | unnecessary protocol expansion when parent linkage already existed upstream
Rejected: Inline full sub-agent markdown in the card body | too cramped for narrow desktop chat layouts
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Keep Agent card summaries compact; route long-form sub-agent output through the preview dialog unless the main chat layout is widened substantially
Tested: cd desktop && bun run test -- MessageList.test.tsx chatStore.test.ts
Tested: cd desktop && bun run lint
Tested: bun test src/server/__tests__/sessions.test.ts -t should\ reconstruct\ parent\ agent\ tool\ linkage\ from\ parentUuid\ chains
Not-tested: Full end-to-end visual verification against live CLI sessions with sub-agent text/thinking nested inline
After the first assistant response, derive a quick placeholder title from
the user message, then asynchronously call the provider's Haiku model to
generate a polished 3-7 word title. Titles update again at message 3 with
fuller conversation context. Updates push to frontend in real-time via
WebSocket `session_title_updated` event.
- extractTitle now reads `ai-title` JSONL entries (priority: custom > ai > first message)
- New titleService with deriveTitle + generateTitle using active provider config
- Handler tracks per-session message count and triggers generation on result
- Frontend sessionStore receives live title updates for sidebar + header
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Implement IM adapters allowing users to chat with Claude Code from Telegram
and Feishu/Lark. Includes persistent session management (chatId→sessionId
mapping), project selection via /projects command, and a web UI settings page
for configuring bot tokens, allowed users, and default project directory.
Key changes:
- adapters/: Telegram and Feishu adapter scripts with shared common modules
(WsBridge, MessageBuffer, SessionStore, HttpClient, config, formatting)
- Backend: adapterService + REST API (GET/PUT /api/adapters) with secret masking
- Frontend: AdapterSettings page in Settings tab with i18n support
- DirectoryPicker: use React Portal for dropdown to fix overflow clipping
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add 7 frequency modes (every N min/hours, daily, weekdays, specific days, monthly, custom cron) with progressive disclosure UI
- Add "Run Now" button with confirmation popover and fire-and-forget API
- Add execution logs panel (TaskRunsPanel) with auto-polling and accordion behavior
- Add task edit mode with cron reverse-parsing (parseCron) to populate form
- Add server-side extractAssistantText to store meaningful AI responses instead of raw NDJSON
- Fix session linking: pass --session-id to CLI subprocess so "View conversation" navigates to actual content
- Fix MACRO undefined error by adding --preload to Bun.spawn
- Add confirmation popovers for all destructive actions (run/disable/delete)
- Add DayOfWeekPicker component for specific-days scheduling
- Add cronDescribe utility with i18n support and unit tests
- Display task creation time and last run time
- Add ~50 i18n keys (en/zh) for all new UI elements
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The CLI subprocess runs in --print (non-interactive) mode, which disabled
V2 task tools (TaskCreate/TaskUpdate). Only TodoWrite was available, but
the frontend only tracked V2 tool names for task refresh — so the Tasks
bar never appeared for new WebApp sessions.
Changes:
- Set CLAUDE_CODE_ENABLE_TASKS=1 in CLI subprocess env to enable V2 tasks
- Add TodoWrite to TASK_TOOL_NAMES for backward compat with V1 sessions
- Parse TodoWrite input.todos directly into TaskBar state (no disk read)
- Extract last TodoWrite from history on session load for V1 sessions
- Inline completed task summary into message flow when user continues chat
(sticky TaskBar converts to scrollable inline summary)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Settings: new Providers tab with full CRUD, activation, and connectivity
test; Model tab shows active provider name; General tab simplified
- Tasks: new CLI Tasks page displaying task lists from ~/.claude/tasks/
with status, owner, and dependency (blocks/blockedBy) visualization
- NewTaskModal: add Advanced options (model, permission mode, working dir)
- Backend: fix TaskService to parse CLI V2 task format; extend /api/tasks
with /lists endpoint for grouped queries
- Fix ModelInfo.context type from number to string
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The desktop app now keeps the composer stable while turns are active,
reduces low-signal tool noise in the transcript, restores project context
under the composer after session creation, and relies on the CLI's own
permission requests instead of injecting broader desktop-side Bash asks.
This also brings in the supporting desktop app source tree and the server
routes/session metadata needed for git info, filesystem browsing, session
resume, slash commands, and SDK-backed permission bridging so the UI can
operate as a coherent feature instead of a partial patch.
Constraint: Desktop transcript needs to stay usable during long multi-tool sessions without hiding file-change diffs
Constraint: Permission prompts must mirror CLI behavior closely enough that read-only commands do not get desktop-only prompts
Rejected: Keep rendering Read/Bash bodies inline | too noisy and unlike the intended transcript model
Rejected: Commit only the touched desktop files | would leave the newly introduced desktop app incomplete in git history
Confidence: medium
Scope-risk: broad
Reversibility: messy
Directive: Treat non-writing tools as summary-first transcript events; do not re-expand them by default without validating the UX against long sessions
Tested: cd desktop && bun run lint
Tested: cd desktop && bun run test -- --run
Tested: bun test src/server/__tests__/conversations.test.ts
Not-tested: Manual visual regression against the exact screenshots in a live desktop session
Not-tested: Full root TypeScript check (repository still has unrelated extracted-native parse failures)
