deer-flow/frontend/AGENTS.md

AGENTS.md

DeerFlow frontend guidance source of truth; CLAUDE.md imports @AGENTS.md.

Project Overview

DeerFlow's Next.js 16 frontend provides thread-based AI conversations with streaming responses, artifacts, skills, and tools through a LangGraph backend.

Stack: Next.js 16, React 19, TypeScript 5.8, Tailwind CSS 4, pnpm 10.26.2. Requires Node.js 22+ and pnpm 10.26.2+.

Core dependencies: @langchain/langgraph-sdk ^1.5.3 for orchestration/streaming, @langchain/core ^1.1.15, and @tanstack/react-query ^5.90.17 for server state. Registry UI: Shadcn, MagicUI, React Bits, Vercel AI SDK (see Code Style).

pnpm-workspace.yaml overrides vulnerable @xmldom/xmldom 0.9.x releases to 0.9.12 for GHSA-965w-775f-mr7g. Nextra pulls it in through MathJax and speech-rule-engine@4.1.2, which pins 0.9.8. Keep the override until the upstream dependency chain resolves a patched version without it; regenerate pnpm-lock.yaml and verify the docs build when changing this constraint.

Commands

Run from frontend/; host pnpm calls use ../scripts/pnpm.py.

• pnpm dev: Webpack development server; pnpm build / pnpm start: production build/server.
• pnpm check: ESLint + TypeScript (required before committing); pnpm lint / pnpm typecheck: individual checks; pnpm lint:fix: ESLint fixes.
• pnpm format / pnpm format:write: Prettier check/write.
• pnpm test: Rstest unit tests; pnpm test:e2e: Playwright Chromium E2E tests.

Unit tests under tests/unit/ mirror src/; import source via @/.

Webpack is the default development bundler. Use DEER_FLOW_DEV_BUNDLER=turbo with pnpm dev to opt in to Turbopack when diagnosing a local Next.js bundler issue.

On Windows pnpm dev binds 127.0.0.1 by default because Hyper-V/winnat excluded port ranges can reject Next's default 0.0.0.0 bind with EACCES (#2870). Pass pnpm dev -- --hostname 0.0.0.0 to listen on a LAN interface instead.

rstest.config.ts: node runs *.test.* pure logic; happy-dom runs *.dom.test.* components/hooks via @testing-library/react render/renderHook. DOM costs ~3x: keep the split. Effect ordering, unmount cleanup and store re-renders need DOM tests with real React.

Content/state suites may mock decorative FlipDisplay with tests/unit/helpers/flip-display.tsx, preserving children, classes and keyed remounts. Happy DOM 21's animation cancellation rejects finished; keep rejection reporting enabled. Real Motion behavior belongs in Playwright.

E2E tests live under tests/e2e/ and use Playwright with Chromium. They mock all backend APIs via page.route() network interception and test real page interactions (navigation, chat input, streaming responses). Config: playwright.config.ts. The real-backend auth contract in tests/e2e-real-backend/auth-disabled-contract.spec.ts and backend/tests/test_auth_me_permissions.py pin the complete route-permission list; update both when adding registered permissions (including projects:read/write/delete).

CI runs four E2E shards; only shard 1 runs auth recovery. Suite reports stay separate.

The dedicated run-history.ts hook replaces the unpaged runs hook. Show counts only after a successful history read, never during initial loading or errors. Scheduled run history uses task/page query keys and the existing live offset API. Fetch 51 rows to display 50 plus a next-page sentinel; never append pages. Only page zero polls or refreshes on focus/reconnect. Task switches reset to page zero, and consumed AbortSignals cancel obsolete reads. Live offsets are not snapshots; explicit mutations or navigation may observe newly inserted runs. Run status unmet identifies a finished occurrence whose scheduled goal was not satisfied; keep it distinct from execution failure. core/scheduled-tasks/goal-outcome.ts maps goal verdicts and host reason codes for run history; show known codes as localized labels; unknown codes, raw run errors and the evaluator's reason stay behind the run row's Details. Check-failure codes are "unchecked", not a miss. contracts/scheduled_goal_notes_contract.json pins the host strings it and run-error.ts match. Scheduled-task views read state through the pure core/scheduled-tasks helpers (status.ts, actions.ts, format.ts, describeTaskSchedule, errors.ts) so list, detail and chat card agree. "Is a run active?" is status === "running" or active_run_status (a recurring task stays enabled while it runs). Default views show no IDs, ISO times, cron strings or enum names; errors localize by detail.code (contracts/scheduled_task_errors_contract.json) with raw text only behind Details. tests/e2e/utils/readable.ts and tests/unit/helpers/readable.ts assert this. The tasks page (app/workspace/scheduled-tasks/page.tsx) only composes components/workspace/scheduled-tasks/*: list with status tabs, detail (Runs / Stops when / goal / Does / notes / History), outcome notice, create/edit/duplicate dialog (PATCH sends only changed fields, null clears) and the renew dialog that a limits_exhausted Resume opens. ?task_id= selects a task; ?thread_id= scopes the list to one chat. notification_group_connection_id adds one Feishu group without changing personal notifications. The selector requires features.scheduled_tasks.group_notifications (absent = false); otherwise a saved target is preserved but explicitly clearable. While the capability query is pending, show loading without a clear control. Reuse useChannelConnections: offer connected Feishu group bindings with matching nonempty metadata chat/workspace IDs. Show workspace names; disable unnamed groups and never display raw IDs. Loading/errors preserve selections. PATCH omits unchanged groups, clears with null, and leaves schedules untouched; Duplicate copies the group. BindFeishuGroup uses the connect API for a persistent, copyable /connect command and explicitly refreshes the shared query. The general connection poll stops early when a private binding already exists, so cannot wait for a new group. In chat, a schedule_task result becomes an assistant:scheduled-task group (core/scheduled-tasks/tool-result.ts, last card per task per turn; not a turn boundary) rendered by ScheduledTaskCard, which polls the task only while on screen. A human message with additional_kwargs.deerflow_scheduled_origin stays a human group with scheduledOrigin: ScheduledRunPrompt renders the origin's user-language parts, never the launched text, and it is not editable. Server-created threads: core/threads/origin.ts reads metadata.deerflow_origin (contracts/thread_origin_contract.json), then channel_source, then a legacy scheduled_task_id; provider names come from threads.origin.providers. core/threads/activity.ts polls /api/thread-activity (mount useThreadActivity once; unmounting resets its cursor; first poll only seeds; invalidate lists only on server-origin threads, truncation or a higher read_version) and useMarkThreadRead posts reads. core/scheduled-tasks/events.ts loads a chat's lifecycle events, not gated on its task list; placeTaskEvents puts each at the end of its after_run_id turn. The UI language is saved as the locale preference once per session and on each switch, only when the Gateway's preferences include locale. Thread rows (sidebar, /workspace/chats) show ThreadOriginIcon and, if unread === true and not the open thread, ThreadUnreadDot with threads.unreadLabel as row label. WorkspaceSidebar mounts useThreadActivity. Chat pages call useMarkOpenThreadRead once metadata loads (marks read on load, on an unread flip, on tab visible, and via its callback from onFinish). Chat pages pass useThreadScheduledTaskEvents data to MessageList (scheduledTaskEvents; the custom-agent page also passes its agent for the run link), which renders ScheduledTaskEventLine through VirtualMessageList.renderAfterGroup at the placeTaskEvents position, inside the group's measured row. Lines are history: never gated on the task list. Channel provider cards (Settings and the sidebar list) render ChannelScheduledUpdates from proactive_notifications, and nothing when the field is absent. Keep IM notification preferences viewer-fenced and separate from browser alerts; only confirmed server writes change switches.

Architecture

Model controls: core/models/use-model-control + HumanInputCard.

Knowledge evidence is parsed in core/citations/sources.ts from version-1 ToolMessage artifacts. Message rows receive sources from preceding loaded tool messages. #knowledge-source-ID links resolve to KnowledgeCitation and an expandable source panel with actual excerpts. Missing URLs/pages remain absent. Web citations and old history remain compatible. Never infer evidence from assistant labels or numbered legacy text. Shared records use the RetrievalSource type (native RAGFlow retains KnowledgeSource). Accept raw and #user-content-knowledge-source-ID fragments without disabling sanitizer clobber protection. Validate immutable tool messages through createKnowledgeSourceReader to reuse per-object parses. Cross-component field bounds live in contracts/knowledge_sources.v1.json. Completed SubtaskCard reports read shared evidence only from the matching task ToolMessage after its invocation in the current subagent group, selected by findSubtaskResultMessage. Prompts/earlier steps do not borrow final evidence. Remount report Markdown when validated source IDs change: Streamdown caches unchanged text blocks independently of renderer components.

Frontend (Next.js) ──▶ LangGraph SDK ──▶ LangGraph Backend (lead_agent)
                                              ├── Sub-Agents
                                              └── Tools & Skills

Users create threads (conversations), send messages, set thread-scoped /goal completion conditions, and receive streamed responses, agent artifacts (files/code), todos and goal state updates.

Source Layout (src/)

AuthProvider publishes the query owner (user ID, confirmed logout, or unknown auth). QueryClientProvider owns a mount-local client and remounts its consumers when that owner changes, clearing the old cache and cancelling old queries. Confirmed logout unmounts query consumers immediately; public and unknown-auth routes still render with an independent client. Keep auth outside the query provider. Superseded auth refreshes cannot restore an earlier owner.

See project structure for the directory map and src/AGENTS.md for component and domain ownership.

Code Style

core/utils/markdown.ts reads web-fetch titles from the first nonblank line. Match zero to three literal spaces before a single # followed by an ASCII space or tab, without trimming indentation; mixed space/tab code blocks must fall back to the URL. Strip optional closing hashes with a backwards scan so long fetched titles cannot cause quadratic regex backtracking. Closing hashes require a preceding space/tab (or an empty title) and only ASCII spaces/tabs afterward, apart from the terminal CR in CRLF input. Non-ASCII trailing whitespace must not turn literal hashes into closing syntax. Keep this local to title extraction rather than changing the shared streamdown fence parser.

Custom Agent display_name is an optional Unicode UI label, edited in AgentSettingsDialog. Use it with a fallback to name for gallery/chat text; keep name for React identity, URLs, requests, and runtime agent_name. The 100-code-point budget uses [...value.trim()].length, matching Pydantic; do not use HTML maxLength, which counts UTF-16 code units instead.

Custom Agent tool access is edited from the configured names returned by GET /api/agent-tool-groups. Preserve the null / [] / explicit-list contract (inherit all / disable all / allowlist), retain missing selected names as removable entries, and submit tool_groups only when the user changes it. Mark missing names unavailable only after a successful catalog load; loading and errors preserve neutral labels and removable selections. When a changed allowlist excludes knowledge, send knowledge_scope: null in the same update so a saved retrieval default cannot break run admission. Gate knowledge selection on both draft and persisted tool access: catalog reads authorize against the saved agent, so new access must be saved first. Restored tool selections remain unchanged and must not clear a knowledge default.

Custom Agent portability uses the versioned deerflow.custom-agent JSON document through core/agents/api.ts. Keep file parsing client-side only for previewing the proposed local name; the Gateway is authoritative for schema, name, model, and conflict validation. Export downloads must never synthesize runtime state from browser caches.

New-Agent saves need the latest setup_agent success and API read-back. Failed/exhausted saves restore Save. Fence superseded or unmounted callbacks. DOM regression: new-agent-save.dom.test.tsx.

• Imports: Enforced ordering (builtin → external → internal → parent → sibling), alphabetized, newlines between groups. Use inline type imports: import { type Foo }.
• Unused variables: Prefix with _.
• Class names: Use cn() from @/lib/utils for conditional Tailwind classes.
• Path alias: @/* maps to src/*.
• Components: ui/ and ai-elements/ are generated from registries (Shadcn, MagicUI, React Bits, Vercel AI SDK) — don't manually edit these.

Sidebar customization: components/workspace/workspace-sidebar-frame.tsx. Rail/drawer mount one tree; useWorkspaceSidebarExpanded preserves desktop SidebarProvider.open. Keep polling/delete-dialog provider outside that tree. Reuse stock Sidebar; close before link navigation and on pathname changes, preserving page focus. Dismissal restores the opener or remounted rail trigger. WorkspaceSidebarTrigger owns surface icons/ARIA; keep generated UI unchanged. Test frame DOM behavior and tests/e2e/sidebar.spec.ts. Agents headers/actions wrap with the rail; sidebar E2E tests cover bounds.

Scheduled-task list search filters the current authorized query result by title or prompt, composing with status/type filters and thread scope. Selection must derive from the filtered list so hidden tasks cannot remain actionable. Keep literal matching in core/scheduled-tasks/search.ts; clearing search retains other filters.

Single-run schedule edits retain the mounted task's original run_at while its wall time and timezone match. The parent echoes edits through initial; retain a stable snapshot and reset the parent draft during render before remounting with a task key when switching tasks. Use the resolved timezone consistently for the snapshot and displayed wall time. Component and scheduled-task E2E tests cover DST folds and timestamp precision.

The scheduled-task schedule input initializes Custom cron from the current preset and its current fields on every transition into the custom editor. Do not substitute a daily schedule or restore stale parts.raw after preset edits. Component tests cover emitted specs; browser tests pin unchanged schedules through the task form's PATCH payload on desktop and mobile.

Environment

Scheduled-task interval forms preserve the initial every_seconds on mount, timezone changes, and untouched blur. The backend's configurable interval minimum can be lower than the UI's default 60-second floor. Apply that UI floor only after an explicit amount/unit edit so editing metadata or duplicating a task cannot silently change its cadence. Component regressions live in tests/unit/components/workspace/scheduled-task-schedule-input.dom.test.tsx.

Backend API URLs are optional; an nginx proxy is used by default:

NEXT_PUBLIC_BACKEND_BASE_URL=http://localhost:8001
NEXT_PUBLIC_LANGGRAPH_BASE_URL=http://localhost:8001/api

Leave these unset for the standard make dev / Docker flow, where nginx serves the public /api/langgraph/* prefix and rewrites it to Gateway's native /api/* routes.

make build-static creates a standalone read-only demo and copies .next/static and public into the output. In static mode, core/api/static-response.ts resolves Gateway REST reads with the bundled capability catalog and safe installation projections from existing same-origin /mock/api fixtures; writes and unknown API routes fail locally. The homepage client counter calls /github-stars, outside the Gateway proxy. That dynamic route reads the server-only GITHUB_OAUTH_TOKEN at runtime, caches GitHub data for one hour, and returns 204 when the count is unavailable. Start the standalone server from frontend/ with node --env-file=.env .next/standalone/server.js to load the current credentials.

To reach a dev server on anything other than localhost — a LAN address, or a proxied hostname — list the host in DEER_FLOW_DEV_ALLOWED_ORIGINS (comma-separated; a full URL is reduced to its host). It feeds Next's allowedDevOrigins, which gates /_next/*, fonts, and HMR. Without it those requests get a 403 and the page renders server-side but never hydrates, so nothing on it — including the login form — responds. Development only; production builds ignore it.

One-time schedule input uses validZonedLocalToUtcIso to reject wall times that do not round-trip in the selected timezone. Invalid input emits an empty spec and localized inline feedback; both create and edit must block submission. Keep this UI validation separate from the API payload. Preserve the original instant when wall time and timezone match the mounted snapshot; validate changed inputs, and restore the exact original timestamp when those edits are reverted.

Wall-time conversion matches the scheduler's ZoneInfo(fold=0) policy: choose the first occurrence of a repeated time and the pre-transition offset for a skipped time. Stored one-time schedules use that instant for display; the form's round-trip validation must still reject newly entered skipped times.

Resources

• LangGraph Documentation
• LangChain Core Concepts
• TanStack Query Documentation
• Next.js App Router

Contributing

Follow src/ ownership, add types/error handling and unit/E2E coverage, run pnpm check, and update this guide for architecture, command, or convention changes.

Route asset budgets are enforced with pnpm perf:check. The command measures /login from a normal production build, then builds in static-demo mode for the fixture-backed workspace routes. It starts the production server on temporary local ports, measures the unique JavaScript and CSS files referenced by representative routes, writes the detailed result to .next/performance-results.json, and compares totals with performance-budgets.json. Fix route ownership or split points when a budget fails; do not raise a ceiling without documenting and reviewing the measured regression.

Chat archive is a thread metadata flag (deerflow_archived === true), independent of run status. Sidebar and Chats explicitly request the Gateway's optional archived filter through searchThreadsByArchive; the SDK drops this extension, so use the authenticated REST fetcher. Static demos retain SDK fixture queries. core/threads/archive.ts waits for the write, cancels stale reads, merges only the owned flag into metadata snapshots, then restarts metadata reads and resets list pagination. Keep both default and Custom Agent header restore controls in sync. Pin/archive responses must not merge unrelated metadata flags: out-of-order organization requests can otherwise roll back each other's confirmed state. Run-created optimistic snapshots have no archive flag: refresh archive-filtered lists from the server instead of inserting those snapshots into either view.

Delimited artifact preview

CSV/TSV previews share artifact-table-preview.tsx between the panel and standalone viewer. Papa Parse runs only inside delimited-preview.worker.ts; use-delimited-preview.ts bounds input before transfer, cancels stale work, and enforces a five-second timeout. The parser detects the first record separator outside quoted fields and passes it explicitly to Papa Parse, so embedded newlines in an incomplete quoted field cannot corrupt newline detection. It retains at most 202 logical records and 50 columns, discarding an incomplete final record from truncated input. UI pagination displays at most 200 data rows in pages of 50. Keep the table mounted but inactive when switching to source so header/pagination state survives; changing file identity resets it. Pending write_file content stays in source mode until success.

For a truncated sample whose first separator is LF or CRLF, strip a terminal CR before parsing so a split CRLF separator cannot make a quoted final field invalidate the whole preview, including LF-first files with later CRLF records. The terminal record remains incomplete and is discarded; complete CR-only files and malformed quotes retain their existing behavior.

Custom skill export is admin-only and disabled in static demos. The lazy skill-export-dialog.tsx must abort requests and ignore stale callbacks on close or user/skill changes. core/skills/export.ts owns the revision-bound Blob download; HTTP 409 requires explicit preview refresh. Keep file lists paginated and diagnostics localized. Browser handoff does not prove the file was saved to disk.

Sidebar rows request deletion through ThreadDeleteDialogProvider, hosted in WorkspaceSidebar outside the virtualized flat/project lists. Keep the selected thread snapshot and retry UI alive when a partial deletion removes its row. Focus Cancel on open and after a failed deletion has re-enabled the actions; block dismissal while deletion is pending. Show the error message when available, with a localized fallback, and log the rejection for debugging. The shared delete helper accepts remote 404 (not 403) before retrying local cleanup, and onDeleted runs only after both deletion steps succeed.

Capability Center

skill-diagnostics.tsx mounts only for administrators outside static demos. Its user-keyed diagnostics query is separate from useSkills() and every picker. Reload posts /api/skills/reload before invalidating both normal skills and diagnostics; failures retain the warning and show localized retry feedback. Never interpret raw parser exception wording in the UI. Keep the scope note visible even when diagnostics are empty. It must describe the YAML-syntax-only check and its exclusions (missing frontmatter, metadata validation, unreadable and non-UTF-8 files); it is not a package-validity check.

/workspace/capabilities owns Plugins and Skills navigation. Plugins composes the MCP manager and a lazily loaded Lark configuration dialog; installation, OAuth, mutation permissions, and cache ownership remain in the existing hooks. Skill display names/summaries are presentation metadata; runtime names and full descriptions remain unchanged. Public, custom, integration, and legacy sources must stay distinct. Community currently offers archive import, not a remote marketplace. Screenshot E2E fixtures are demo data. backend/packages/harness/deerflow/capabilities/builtin.json owns localized catalog manifests. Refresh the generated demo snapshot with pnpm catalog:sync after changing the catalog; unit tests enforce equality with the source. Demo business projections derive provider IDs from the catalog adapter metadata. The sync script uses decoded filesystem paths for formatter configuration lookup. plugin-catalog.ts only resolves localized text and explicit installation metadata; never infer provider identity from server display names. plugin-directory.tsx groups rows and applies search/category/installed filters. core/capabilities consumes catalog and safe status projections; MCP secrets and raw settings remain in the administrator-only editor. plugin-adapters.tsx registers integration-specific settings flows once, independent of catalog size. The business form uses manifest credential fields without asking for an MCP URL; its backend adapter generates bundled DingTalk/WeCom notification or HubSpot CRM connections. These also appear in MCP discovery, so deduplicate projections by installation ID. Keep their labels as configuration, not package installation. Keep installation, enabled state, configured credentials, and verified authorization distinct. Agent mcp_plugins uses stable installation IDs; null means all, [] means none. The settings dialog submits only selections changed from its opening snapshot, preserving concurrent updates on unrelated saves and treating restored selections as unchanged. It is runtime selection, not a replacement authorization policy. See docs/capability-center.md for the complete contract and extension example. PluginIcon is shared by recommendations, configured entries, and the editor; brand assets and their provenance live in public/images/plugins/. Brand icons require explicit catalog metadata; a custom server name never selects a brand. Ambiguous installation IDs remain visible but cannot be selected for an Agent. The icon picker accepts local PNG/JPEG/WebP up to 2 MiB, checks the signature, decodes and contains the image in a 128px PNG, and stages changes until the existing targeted MCP save. presentation.icon is a bounded PNG data URL carried by the API's existing extra metadata support; it must never enter transport parameters. Preserve sibling presentation fields and masked credentials; cancel/reset/unmount must fence stale image-decoding results. Uploaded remote URLs and SVG are never rendered. Existing shared-MCP administrator checks remain authoritative; this adds no personal scope.

This is NOT the Next.js you know

This version has breaking changes — APIs, conventions, and file structure may all differ from your training data. Read the relevant guide in node_modules/next/dist/docs/ (resolved from this file's directory; in monorepos the next package may not be visible from the repo root) before writing any code. Heed deprecation notices.

This block is written and re-added by next dev — verify at node_modules/next/dist/server/lib/generate-agent-files.js. Removing it from a diff only re-creates the uncommitted change; committing it with your work keeps the tree clean.

Shared model settings

Settings → Models (?settings=models) offers administrator-only catalog management through /api/managed-models. YAML entries are read-only. core/models/management.ts whitelists editable fields so source metadata and has_api_key never get posted. Draft credentials stay in editor state, never query cache or browser storage; blank keeps the saved key, explicit removal sends an empty key. Saving invalidates both the admin catalog and MODELS_QUERY_KEY. Editor unmount aborts probes and fences late callbacks. Static demos and non-admin users must not query the management API.

Admin user management

Settings → Users (admin-only nav entry; ?settings=users) suspends and restores accounts through core/auth/admin-users.ts → /api/v1/admin/users. The page reuses userSchema for rows — both endpoints parse against it, so a malformed payload fails the load instead of rendering a phantom row (disabled optional — absent means not disabled) — and PATCHes only {disabled}; role stays a read-only column. Like model management, the page also gates on isStaticWebsiteOnly() (the demo user is nominally an admin, but there is no Gateway to query), and the settings dialog falls back to the first nav section when a deep link or stale store activates a section the current user's nav does not offer. The Gateway's rejection detail (last-active-admin 409) renders verbatim. A session that turns out to be disabled maps account_disabled (see AUTH_ERROR_CODES) to a login redirect with a stated reason (login surface only — sign-up keeps the backend message so a suspended account cannot be probed) on all three surfaces: SSR AuthResult tag (consumed exhaustively by the workspace/artifacts/auth layouts), the shared fetcher's 401 interception, and AuthProvider.refreshUser. Direct auth-page visits with a live suspended-session cookie get the SSR verdict via AuthDisabledContext, so the login form states the known reason without waiting for a deliberately generic login failure.

Other sign-in bounces keep path, query and fragment via currentReturnPath(). The workspace layout cannot read the URL, so unauthenticated renders WorkspaceLoginRedirect (never children), not redirect().

Model reasoning capabilities

/api/models projects a per-model reasoning contract (issue #5073) beside the deprecated supports_thinking / supports_reasoning_effort booleans. core/models/reasoning.ts is the only place that interprets it: it falls back to the booleans for older Gateways, clamps the chat mode (getResolvedMode never yields flash for a required-thinking model), lists the effort options the composer and the sidecar render, and maps mode presets and remembered values through the contract's aliases/default (resolveReasoningEffort). Effort values are open strings (ReasoningEffortValue), so provider-specific tokens such as max flow through local settings and account preferences unchanged; the custom agent dialog offers only the intersection with the per-agent low/medium/high schema and blocks new "off" for required-thinking models. Do not read the booleans in components directly. Legacy resolveReasoningEffort keeps only advertised generic values, dropping provider tokens after a switch; backend requests stay unchanged. Agent settings omit untouched reasoning, show invalid saved values for clearing, and require a known model to save a model switch.

Full-stack plugin UI

Follow invariants for plugin loaders/pages/actions: viewer-bound auth, native actions, i18n and mount/dispose/abort fencing. Shadow DOM is not a sandbox. Bundled extensions need typed en/zh core/extensions/catalog.ts entries with their namespace. Catalog entries are not proof of installation or activation.

Composer references

components/workspace/mentions/ owns cursor-local @ detection, the picker, and atomic references. Render labels as DOM text, never HTML; canonical tokens preserve draft positions. Submission expands tokens to @label and sends up to 16 unique skill IDs in additional_kwargs.skill_references; the backend checks each against the user registry and agent allowlist. The web slash picker lists only /goal and /compact; skill selection uses @. Legacy typed slash text continues through normal message submission for backend and channel compatibility. Project files require confirmed additional_kwargs.files. Conversation context is reconciled from tokens against the current capability and limit. Only successful discovery may flatten references; pending/errors preserve IDs and block reference sends, with retry on failure. Polish restores whole labels longest-first in one pass. Confirmed thread creation seeds metadata before migration; references stay locked during attachment. Deleting an object removes its context; undo restores references already attached to the draft. Labels grant no read authority. IME, caret, thread changes and attachment fences apply to editor and picker. Commit mention text before restoring focus synchronously; a delayed callback must not overwrite subsequent typing or caret movement. The @ and attachment buttons share the picker.

Web model provider presets

Settings → Models obtains provider form metadata from the admin-only backend /api/managed-models/providers; do not copy endpoints, protocols or adapter maps into the UI. New models need only provider, API key and arbitrary model ID, with server-generated stable names. Custom endpoint fields and legacy edits stay supported. Provider/endpoint changes clear test results and saved-key reuse; omitted keys retain only the same connection. Capability overrides belong in Advanced settings. Successful saves/toggles invalidate both managed-model queries and MODELS_QUERY_KEY so existing chat/Agent selectors refresh. Keep locale text and administrator/static-mode restrictions aligned.