Feishu CardKit and DingTalk AI Card updates run inside the serialized adapter message path. A hung platform update can prevent the later message_complete event from being handled, leaving mobile IM users stuck on a partial streaming card.
This bounds platform card API calls and gives Feishu a final patch fallback when CardKit finalization fails, so a bad intermediate frame cannot permanently block the chat turn.
Constraint: Feishu and DingTalk card APIs are external network calls without reliable SDK-level deadline guarantees
Rejected: Disable streaming cards entirely | loses the live IM experience for normal fast responses
Confidence: high
Scope-risk: narrow
Directive: Keep IM card platform calls bounded because adapter server messages are serialized per chat
Tested: bun test adapters/feishu/__tests__/streaming-card.test.ts
Tested: bun test adapters/dingtalk/__tests__/ai-card.test.ts
Tested: bun run check:adapters
Tested: bunx tsc -p adapters/tsconfig.json --noEmit
Not-tested: Full bun run verify was interrupted during the coverage lane after dependency setup
Deleting a desktop session removed the transcript but left IM adapter
chat mappings and live WebSocket state able to point at the old session.
The server now closes the active session socket and removes adapter
mappings after deletion, while adapters refresh the shared session store
before reads so a running process cannot reuse stale in-memory data.
Constraint: Adapter session mappings are shared through adapter-sessions.json across long-running IM processes
Rejected: Patch each adapter ensureSession path separately | shared store refresh fixes all current adapters and avoids drift
Confidence: high
Scope-risk: moderate
Directive: Do not cache adapter session mappings without invalidating after server-side session deletion
Tested: bun test common/__tests__/session-store.test.ts common/__tests__/ws-bridge.test.ts
Tested: bun test src/server/__tests__/websocket-handler.test.ts src/server/__tests__/sessions.test.ts --timeout 30000
Tested: cd adapters && bunx tsc --noEmit
Tested: bun run check:adapters
Tested: bun run check:server
Not-tested: bun run verify fails on pre-existing agent-utils coverage ratchet; changed-line coverage is 95.92% and affected lanes pass
Related: https://github.com/NanmiCoder/cc-haha/issues/305
DingTalk keeps AI card placement tied to the original card delivery time, so a card created before a permission request can later be updated with the final answer above the permission card. Finish and drop the active stream when permission interrupts a turn so authorized output creates a new card below the prompt.
Constraint: DingTalk does not reorder an existing delivered AI card when its streaming content is updated
Rejected: Reuse the pre-permission AI card after approval | final answers can remain visually above the permission request
Confidence: high
Scope-risk: narrow
Directive: Keep permission_request as a lifecycle boundary for DingTalk AI card streams
Tested: cd adapters && bun test dingtalk/__tests__/stream-state.test.ts dingtalk/__tests__/ai-card.test.ts
Tested: cd adapters && bunx tsc --noEmit
Tested: bun run check:adapters
Tested: git diff --check
Not-tested: Live DingTalk account end-to-end permission approval
The WeChat and DingTalk IM docs were behind the actual desktop adapter flow, especially around QR binding versus user pairing and how users discover slash commands when no platform menu exists. This records the setup path with screenshots and nudges paired users toward /help from the bot itself.
Constraint: DingTalk does not expose the same menu setup path documented for Feishu, so command discovery must work through chat text.
Rejected: Document commands only in VitePress | users may be on mobile when they need the commands.
Confidence: high
Scope-risk: narrow
Directive: Keep IM command docs aligned with adapters/common/format.ts when adding commands.
Tested: cd adapters && bun test common/ dingtalk/ wechat/
Tested: cd adapters && bunx tsc --noEmit
Tested: bun run check:docs
Tested: bun run check:adapters
Not-tested: Live WeChat or DingTalk message delivery with real platform accounts
Text-only IM channels forced users to copy long permission request IDs from mobile chat, which made approval slow and error-prone. This keeps the requestId-based authorization protocol intact while adding short replies for the single-pending-request case and preserving full command fallbacks for ambiguous cases.
Constraint: IM authorization must still resolve through the existing requestId permission_response path
Rejected: Replace request IDs with global numeric IDs | concurrent permission prompts would make numeric IDs ambiguous across chats
Confidence: high
Scope-risk: moderate
Directive: Do not allow short numeric replies when more than one permission request is pending in the chat
Tested: cd adapters && bun test common/ telegram/ feishu/ dingtalk/ wechat/
Tested: cd adapters && bunx tsc --noEmit
Tested: bun run check:adapters
Not-tested: Real WeChat or Telegram account end-to-end message delivery
WeChat development turns were reaching the local CLI, but the adapter treated iLink business failures as successful sends and only emitted a short-lived typing indicator. The adapter now validates sendmessage/sendtyping ret codes, retries text replies without stale context tokens, keeps typing alive during long-running tool work, and surfaces queue failures back to the chat.
Constraint: WeChat uses separate sendmessage and sendtyping APIs with business-level ret codes inside HTTP 200 responses.
Rejected: Only fixing the typing indicator | the transcript showed the agent completed the task, so the reply delivery path also needed hardening.
Confidence: high
Scope-risk: narrow
Directive: Do not treat WeChat HTTP 200 responses as successful until the iLink ret/errcode body has been checked.
Tested: bun test adapters/wechat/__tests__/protocol.test.ts adapters/wechat/__tests__/typing.test.ts
Tested: cd adapters && bunx tsc --noEmit
Tested: bun run check:adapters
Tested: real bound WeChat getconfig, typing, text send, cancel typing smoke
Not-tested: Full inbound WeChat user-message loop without a fresh user-triggered message
This captures the pending worktree fixes before applying them to the
current local main. The changes tighten IM adapter path and credential
handling, preserve retry behavior for failed desktop notifications, and
make Azure/OpenAI provider auth and stop reasons reflect actual runtime
state.
Constraint: Worktree was detached from an older local main with pending uncommitted fixes
Rejected: Merge the detached HEAD directly | would also replay unrelated stale history
Rejected: Leave notification dedupe as fire-and-forget | failed sends consumed retry keys
Confidence: high
Scope-risk: broad
Directive: Keep adapter absolute-path matching constrained to configured work roots
Tested: git diff --check
Not-tested: full quality gate before local main integration
DingTalk can receive interactive card callbacks, but the card UI still depends on a published template, so the adapter now supports a template-backed card path with text commands as the reliable fallback. Telegram and WeChat use the same allow-once, allow-always, and deny semantics so manual authorization behaves the same across platforms.
Constraint: DingTalk button rendering requires an operator-provided interactive card template id
Rejected: Treat the existing AI streaming card template as a permission card | it cannot guarantee visible action buttons
Confidence: high
Scope-risk: moderate
Directive: Do not remove the text approval fallback unless DingTalk card template provisioning is guaranteed
Tested: bun test adapters/common/__tests__/permission.test.ts adapters/dingtalk/__tests__/permission-card.test.ts adapters/telegram/__tests__/telegram.test.ts adapters/common/__tests__/config.test.ts src/server/__tests__/adapters.test.ts
Tested: bunx tsc -p adapters/tsconfig.json --noEmit
Tested: bun run check:adapters
Tested: cd desktop && bun run build
Tested: bun run check:server
Tested: bun run check:docs
Tested: git diff --check
Not-tested: bun run quality:pr is blocked by existing CLI core approval policy; see artifacts/quality-runs/2026-05-03T13-19-56-232Z/report.md
WeChat and DingTalk were using different pairing, attachment, and response
state paths, which made the new IM channels behave differently from Feishu
and Telegram. Align the shared pairing model, wire inbound media into the
existing attachment bridge, and map platform response capabilities to their
real APIs: WeChat block streaming plus typing, DingTalk AI Card streaming.
Constraint: WeChat iLink exposes typing and block streaming, but no editable message/card streaming API
Constraint: DingTalk streaming depends on the AI Card create/deliver/stream/finalize lifecycle
Rejected: Fake DingTalk typing with standalone markdown | it would add chat noise instead of platform state
Rejected: Auto-pair WeChat after QR login | it bypasses the shared IM pairing model
Confidence: high
Scope-risk: moderate
Directive: Keep WeChat streaming as block-send unless iLink adds editable messages; keep DingTalk streaming on AI Card APIs
Tested: bun run check:adapters; bun run check:server; cd desktop && bun run test src/stores/adapterStore.test.ts; cd desktop && bun run build; bunx tsc -p adapters/tsconfig.json --noEmit; git diff --check
Not-tested: Live WeChat and DingTalk platform smoke with real production credentials
DingTalk uses QR registration to store client credentials, then reuses the existing IM pairing and session bridge. The merged main implementation keeps the existing WeChat QR binding path intact while adding DingTalk as a peer IM platform. The default IM workdir now falls back to the local user working directory so a newly bound chat can start immediately even when recent-project history is empty.
Constraint: Local main already carries WeChat IM, so the merge keeps WeChat config, QR binding, sidecar args, and unbind behavior intact while adding DingTalk.
Rejected: Keep empty defaultProjectDir as project-picker-only | newly bound IM users can hit a dead end with no recent projects.
Confidence: high
Scope-risk: moderate
Directive: Keep IM platform unions synchronized across config, pairing, sidecar args, desktop settings, and docs.
Tested: bun test common/ dingtalk/ wechat/; bun test src/server/__tests__/adapters.test.ts; cd adapters && bunx tsc --noEmit; bun run check:policy; bun run check:server; bun run check:desktop; bun run check:adapters; bun run check:native; bun run check:docs
Not-tested: quality:pr full gate was blocked by existing local main CLI-core diff requiring allow-cli-core-change maintainer approval; live post-fix DingTalk second-message delivery was not repeated after the merge.
WeChat needs a QR-paired path instead of bot-token setup, so the adapter layer now includes the iLink protocol calls, desktop pairing UI, server-side bind/unbind APIs, and shared IM command behavior. Empty project history falls back to the user's default work directory so mobile /new works without pre-opening a desktop project.
Constraint: Tencent iLink login returns a URL that the desktop UI must render as a QR image locally
Constraint: IM adapters should keep /new, /projects, status, permission, and default workdir behavior consistent across WeChat, Feishu, and Telegram
Rejected: Require users to paste absolute project paths for first WeChat sessions | mobile onboarding should work from the default user working directory
Confidence: high
Scope-risk: moderate
Directive: Do not change WeChat polling back to overlapping intervals; getupdates is a long-poll endpoint and must remain serialized
Tested: Real WeChat QR scan, inbound /status, outbound reply, and unbind E2E
Tested: bun run check:adapters
Tested: bun run quality:pr
Not-tested: Re-scan live WeChat after the default-workdir fallback tweak; covered by adapter config tests and PR gate
Three security improvements for adapters.json storage:
- Use crypto.randomBytes for temp filename instead of predictable
Date.now(), preventing TOCTOU symlink race during atomic write.
- Enforce 0o600 permissions on adapters.json via writeFileSync mode
option to prevent credential leakage via permissive umask.
- Enforce 0o700 permissions on ~/.claude/ config directory via
mkdirSync mode option.
adapters.json contains sensitive credentials including Feishu app
secrets and Telegram bot tokens.
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
Feishu cards used to sit on "正在思考中..." for the entire turn and
then dump everything at message_complete, while Telegram showed the
reasoning and tool calls progressively. Two underlying causes:
1. handleServerMessage ignored `thinking`, `content_start{tool_use}`
and `tool_use_complete` entirely, so nothing was fed into the card
between turns of text output.
2. StreamingCard.performFlush silently set `cardKitStreamActive=false`
on any non-rate-limit error, after which all middle frames were
skipped and only finalize touched the card — exactly the
"long wait → all at once" symptom.
StreamingCard now tracks accumulatedReasoningText and a toolSteps
list, with appendReasoning / startTool / completeTool methods that
trigger throttled flushes like appendText. renderedText composes
tools → reasoning → answer in plain markdown (no blockquotes / lists
that historically tripped Feishu's parser). consecutiveStreamFailures
threshold (3) keeps a single transient error from killing the stream.
terminalText() is used by finalize so the final card holds only the
answer text, matching the Desktop UI's clean post-completion view.
handleServerMessage wires the new events to the existing per-chat
StreamingCard via the streamingCards map (read-only — never
auto-creates cards from thinking/tool events to avoid empty cards
for /clear-style commands).
Test coverage: realistic event-stream regression (thinking → tool_use
→ text → finalize), first-frame-failure recovery, finalize drops
intermediate state, plus per-method tests for the new APIs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Re-implements the inbound and outbound attachment paths on top of the
StreamingCard rewrite (ae586b0) instead of the legacy MessageBuffer
pipeline. Replaces the dangling extractText reference and ports the
behavior originally drafted on feat/im-attachments to the new structure:
Inbound (user → Claude):
- handleMessage now uses extractInboundPayload to detect text and any
PendingDownloads (image/file/file_archive/post-embedded media)
- safeMessageId captured outside enqueue to keep TS narrowing happy
- All command branches gated on `!hasAttachments` so attachment-bearing
messages bypass /new, /clear, /stop, /projects, etc.
- Per-message Promise.allSettled download via FeishuMediaService;
per-attachment checkAttachmentLimit; partial-failure user hint
- AttachmentRef[] forwarded through bridge.sendUserMessage(...)
- effectiveText guard prevents empty content from being sent to Claude
when all attachments are rejected and the user supplied no text
Outbound (Claude → user):
- content_delta handler feeds the streaming text into per-chat
ImageBlockWatcher and dispatches each new PendingUpload via
dispatchOutboundImage (fire-and-forget)
- dispatchOutboundImage handles all three source kinds (base64/path/url),
applies the same checkAttachmentLimit, dedupes by fingerprint, and
publishes the image as an independent im.message.create({msg_type:'image'})
alongside the streaming card text
- Card-embedded `{tag:'img', img_key}` rendering is intentionally deferred
to a follow-up PR — the MVP keeps streaming text and outbound images as
separate messages
Lifecycle: clearTransientChatState and startNewSession now also delete
imageWatchers and uploadedImageKeys for the chat so a new session never
inherits stale fingerprints.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- attachment-store.ts: pin Dirent<string>[] and pass `encoding: 'utf8'`
so newer @types/node doesn't infer Dirent<NonSharedBuffer>
- feishu/media.ts: drop Readable.from(buffer) and pass Buffer directly
to im.image.create / im.file.create — Lark SDK already accepts Buffer
and the stream wrapper trips the stricter Buffer | ReadStream union
used on the main branch's tsconfig
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Lists inbound/outbound attachment coverage, local staging directory,
size limits, and the file/module additions from the im-attachments
feature branch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Implements TDD Task 9: wraps grammY bot.api.getFile/sendPhoto/sendDocument
with local attachment staging via AttachmentStore.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Address two production issues in the AttachmentStore:
1. TOCTOU race in resolvePath() — concurrent downloads with same filename in
the same millisecond could collide. Fixed by appending a 6-char random suffix
to the timestamp-based path, ensuring millisecond-level collisions never occur.
2. Orphan .part files from crashed writes never cleaned up. Fixed by extending
gc() to recognize .part files and clean them with a shorter grace period
(10 minutes by default) rather than the normal retention window.
Added two test cases covering heavy collision pressure and .part cleanup.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace legacy im.message.patch streaming path with a Feishu CardKit
Schema 2.0 pipeline mirroring openclaw-lark's implementation. Desktop
now renders real formatted markdown (headings, tables, code blocks)
instead of literal text, and streaming output updates per chat without
cross-talk.
New modules:
- card-errors.ts: structured Lark SDK error parsing for code 230020
(rate limit -> skip frame) and 230099/11310 (table limit -> disable
CardKit streaming and fall back to settings+update at finalize)
- flush-controller.ts: mutex + needsReflush + long-gap batching
throttler for per-chat flush coordination
- cardkit.ts: thin wrapper over client.cardkit.v1.* (card.create +
im.message.{create,reply} + cardElement.content + card.settings +
card.update), with monotonic sequence and CardKitApiError
- streaming-card.ts: per-chat StreamingCard state machine covering
idle -> creating -> streaming -> finalizing -> completed/aborted,
including patch fallback when CardKit create/send fails
index.ts rewrite:
- replaces MessageBuffer/chatStates/buffers maps with a single
streamingCards Map<chatId, StreamingCard>
- content_start{text}/content_delta create or reuse the card; finalize
runs on message_complete; abort renders a red error card
- /clear and other silent commands stay out of the streaming-card path
so they don't leave empty cards
- all command handlers + normal chat are wrapped in the per-chat
enqueue() serial queue, preventing reply reordering when commands
are fired rapidly
- createSessionForChat() now first resets any stale WS session before
calling connectSession, fixing /projects pick_project leaving the
old session bound to the previous workDir
- normal messages pre-create the card before sendUserMessage so users
see a "☁️ 正在思考中..." loading indicator immediately
- content_start{tool_use} no longer finalizes the current card, letting
one user turn's full text stay in a single card
markdown-style.ts:
- default cardVersion changed from 1 to 2 (Schema 2.0)
- headings H2~H6 demoted to H5, H1 to H4, only when source has H1~H3
- Schema 2.0 <br> padding around headings, tables, code blocks
- stripInvalidImageKeys() drops non-img_* image references
- FEISHU_CARD_TABLE_LIMIT=3 + sanitizeTextForCard() wraps 4th+ table
in a fenced code block to avoid 230099/11310
Tests: 182 pass, 0 fail, 383 expect() calls across 6 files. tsc clean.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Carries forward ~800 lines of in-flight work that existed in the working
tree before the IM-attachment branch was cut. Kept as a single isolated
commit so later attachment work has a clean diff baseline and so this
WIP can be cherry-picked back to main on its own merits.
Contents:
- ws-bridge: per-chat handler chain serialization; sendPermissionResponse
now accepts optional `rule: 'always'` for persistent permits
- feishu/index.ts: permission card rewritten to Schema 2.0 with icon,
cross-dir warning, and ♾️ 永久允许 button; optimizeMarkdownForFeishu
applied to streaming card
- feishu/markdown-style.ts: new Feishu-specific markdown downgrader
- tests: corresponding bun:test coverage
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Resolve bot open_id via /open-apis/bot/v3/info — the SDK's
contact.user.get({user_id: 'me'}) path is invalid on Feishu and
returned 99992351 on every startup.
- Stream replies via interactive cards instead of post messages:
im.message.patch only accepts msg_type=interactive, so the old
path-rich-text placeholder caused "This message is NOT a card"
(230001) on every flush. Added buildStreamingCard helper.
- Add project picker card: one column_set row per project with
[info (name + branch + path) | select button], Schema 2.0,
mobile-friendly path display (~ substitution + middle-truncation),
header subtitle with project count. pick_project actions route
through the existing handleCardAction.
Card patterns verified against openclaw-lark's Schema 2.0 usage in
src/tools/oauth-cards.ts and src/card/builder.ts. 34 unit tests
cover the new card structure and action routing.
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 matchProject() to http-client for fuzzy project matching by
index or name
- Extract startNewSession() in both adapters to share reset + match +
create logic
- Project selection replies now go through startNewSession(), creating
a clean new session instead of continuing the old one
- Support: /new 2 (by index), /new backend (by name), /new (default)
- Add usage hint in /projects list output
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When all buffered text has already been flushed via periodic timer,
buf.complete() calls flush(true) which returns early on empty buffer,
skipping the onFlush callback that cleans up placeholder state.
This causes subsequent messages to keep editing the old completed
message instead of creating a new one.
Add explicit placeholder cleanup in message_complete handler as
a safety net, matching the same pattern used for tool_use finalization.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix WebSocket race condition: wait for connection to open before
sending first message, preventing silent message loss after pairing
or project selection
- Fix response text appearing above tool calls: finalize placeholder
before tool_use blocks so post-tool text gets a new message
- Remove noisy tool_use and tool_result messages from IM output;
thinking indicator is preserved, details visible in Desktop
- Fix /new command using default project dir instead of always showing
project picker
- Fix duplicate projects in list by deduping on realPath instead of
projectPath
- Improve formatToolUse with human-readable summaries for common tools
- Add send failure feedback to users
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove .claude/ and docs/superpowers/ from git tracking as they are
local-only config and planning files. Also remove stale screenshots
and update docs/channel content.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rate limit: max 5 failed pairing attempts per user per 5 minutes
- Default closed: reject all users when no allowedUsers/pairedUsers configured
- Extract isAllowedUser() to common/pairing.ts to eliminate duplication
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>
