* feat(auth): introduce backend auth module Port RFC-001 authentication core from PR #1728: - JWT token handling (create_access_token, decode_token, TokenPayload) - Password hashing (bcrypt) with verify_password - SQLite UserRepository with base interface - Provider Factory pattern (LocalAuthProvider) - CLI reset_admin tool - Auth-specific errors (AuthErrorCode, TokenError, AuthErrorResponse) Deps: - bcrypt>=4.0.0 - pyjwt>=2.9.0 - email-validator>=2.0.0 - backend/uv.toml pins public PyPI index Tests: 12 pure unit tests (test_auth_config.py, test_auth_errors.py). Scope note: authz.py, test_auth.py, and test_auth_type_system.py are deferred to commit 2 because they depend on middleware and deps wiring that is not yet in place. Commit 1 stays "pure new files only" as the spec mandates. * feat(auth): wire auth end-to-end (middleware + frontend replacement) Backend: - Port auth_middleware, csrf_middleware, langgraph_auth, routers/auth - Port authz decorator (owner_filter_key defaults to 'owner_id') - Merge app.py: register AuthMiddleware + CSRFMiddleware + CORS, add _ensure_admin_user lifespan hook, _migrate_orphaned_threads helper, register auth router - Merge deps.py: add get_local_provider, get_current_user_from_request, get_optional_user_from_request; keep get_current_user as thin str|None adapter for feedback router - langgraph.json: add auth path pointing to langgraph_auth.py:auth - Rename metadata['user_id'] -> metadata['owner_id'] in langgraph_auth (both metadata write and LangGraph filter dict) + test fixtures Frontend: - Delete better-auth library and api catch-all route - Remove better-auth npm dependency and env vars (BETTER_AUTH_SECRET, BETTER_AUTH_GITHUB_*) from env.js - Port frontend/src/core/auth/* (AuthProvider, gateway-config, proxy-policy, server-side getServerSideUser, types) - Port frontend/src/core/api/fetcher.ts - Port (auth)/layout, (auth)/login, (auth)/setup pages - Rewrite workspace/layout.tsx as server component that calls getServerSideUser and wraps in AuthProvider - Port workspace/workspace-content.tsx for the client-side sidebar logic Tests: - Port 5 auth test files (test_auth, test_auth_middleware, test_auth_type_system, test_ensure_admin, test_langgraph_auth) - 176 auth tests PASS After this commit: login/logout/registration flow works, but persistence layer does not yet filter by owner_id. Commit 4 closes that gap. * feat(auth): account settings page + i18n - Port account-settings-page.tsx (change password, change email, logout) - Wire into settings-dialog.tsx as new "account" section with UserIcon, rendered first in the section list - Add i18n keys: - en-US/zh-CN: settings.sections.account ("Account" / "账号") - en-US/zh-CN: button.logout ("Log out" / "退出登录") - types.ts: matching type declarations * feat(auth): enforce owner_id across 2.0-rc persistence layer Add request-scoped contextvar-based owner filtering to threads_meta, runs, run_events, and feedback repositories. Router code is unchanged — isolation is enforced at the storage layer so that any caller that forgets to pass owner_id still gets filtered results, and new routes cannot accidentally leak data. Core infrastructure ------------------- - deerflow/runtime/user_context.py (new): - ContextVar[CurrentUser | None] with default None - runtime_checkable CurrentUser Protocol (structural subtype with .id) - set/reset/get/require helpers - AUTO sentinel + resolve_owner_id(value, method_name) for sentinel three-state resolution: AUTO reads contextvar, explicit str overrides, explicit None bypasses the filter (for migration/CLI) Repository changes ------------------ - ThreadMetaRepository: create/get/search/update_*/delete gain owner_id=AUTO kwarg; read paths filter by owner, writes stamp it, mutations check ownership before applying - RunRepository: put/get/list_by_thread/delete gain owner_id=AUTO kwarg - FeedbackRepository: create/get/list_by_run/list_by_thread/delete gain owner_id=AUTO kwarg - DbRunEventStore: list_messages/list_events/list_messages_by_run/ count_messages/delete_by_thread/delete_by_run gain owner_id=AUTO kwarg. Write paths (put/put_batch) read contextvar softly: when a request-scoped user is available, owner_id is stamped; background worker writes without a user context pass None which is valid (orphan row to be bound by migration) Schema ------ - persistence/models/run_event.py: RunEventRow.owner_id = Mapped[ str | None] = mapped_column(String(64), nullable=True, index=True) - No alembic migration needed: 2.0 ships fresh, Base.metadata.create_all picks up the new column automatically Middleware ---------- - auth_middleware.py: after cookie check, call get_optional_user_from_ request to load the real User, stamp it into request.state.user AND the contextvar via set_current_user, reset in a try/finally. Public paths and unauthenticated requests continue without contextvar, and @require_auth handles the strict 401 path Test infrastructure ------------------- - tests/conftest.py: @pytest.fixture(autouse=True) _auto_user_context sets a default SimpleNamespace(id="test-user-autouse") on every test unless marked @pytest.mark.no_auto_user. Keeps existing 20+ persistence tests passing without modification - pyproject.toml [tool.pytest.ini_options]: register no_auto_user marker so pytest does not emit warnings for opt-out tests - tests/test_user_context.py: 6 tests covering three-state semantics, Protocol duck typing, and require/optional APIs - tests/test_thread_meta_repo.py: one test updated to pass owner_id= None explicitly where it was previously relying on the old default Test results ------------ - test_user_context.py: 6 passed - test_auth*.py + test_langgraph_auth.py + test_ensure_admin.py: 127 - test_run_event_store / test_run_repository / test_thread_meta_repo / test_feedback: 92 passed - Full backend suite: 1905 passed, 2 failed (both @requires_llm flaky integration tests unrelated to auth), 1 skipped * feat(auth): extend orphan migration to 2.0-rc persistence tables _ensure_admin_user now runs a three-step pipeline on every boot: Step 1 (fatal): admin user exists / is created / password is reset Step 2 (non-fatal): LangGraph store orphan threads → admin Step 3 (non-fatal): SQL persistence tables → admin - threads_meta - runs - run_events - feedback Each step is idempotent. The fatal/non-fatal split mirrors PR #1728's original philosophy: admin creation failure blocks startup (the system is unusable without an admin), whereas migration failures log a warning and let the service proceed (a partial migration is recoverable; a missing admin is not). Key helpers ----------- - _iter_store_items(store, namespace, *, page_size=500): async generator that cursor-paginates across LangGraph store pages. Fixes PR #1728's hardcoded limit=1000 bug that would silently lose orphans beyond the first page. - _migrate_orphaned_threads(store, admin_user_id): Rewritten to use _iter_store_items. Returns the migrated count so the caller can log it; raises only on unhandled exceptions. - _migrate_orphan_sql_tables(admin_user_id): Imports the 4 ORM models lazily, grabs the shared session factory, runs one UPDATE per table in a single transaction, commits once. No-op when no persistence backend is configured (in-memory dev). Tests: test_ensure_admin.py (8 passed) * test(auth): port AUTH test plan docs + lint/format pass - Port backend/docs/AUTH_TEST_PLAN.md and AUTH_UPGRADE.md from PR #1728 - Rename metadata.user_id → metadata.owner_id in AUTH_TEST_PLAN.md (4 occurrences from the original PR doc) - ruff auto-fix UP037 in sentinel type annotations: drop quotes around "str | None | _AutoSentinel" now that from __future__ import annotations makes them implicit string forms - ruff format: 2 files (app/gateway/app.py, runtime/user_context.py) Note on test coverage additions: - conftest.py autouse fixture was already added in commit 4 (had to be co-located with the repository changes to keep pre-existing persistence tests passing) - cross-user isolation E2E tests (test_owner_isolation.py) deferred — enforcement is already proven by the 98-test repository suite via the autouse fixture + explicit _AUTO sentinel exercises - New test cases (TC-API-17..20, TC-ATK-13, TC-MIG-01..07) listed in AUTH_TEST_PLAN.md are deferred to a follow-up PR — they are manual-QA test cases rather than pytest code, and the spec-level coverage is already met by test_user_context.py + the 98-test repository suite. Final test results: - Auth suite (test_auth*, test_langgraph_auth, test_ensure_admin, test_user_context): 186 passed - Persistence suite (test_run_event_store, test_run_repository, test_thread_meta_repo, test_feedback): 98 passed - Lint: ruff check + ruff format both clean * test(auth): add cross-user isolation test suite 10 tests exercising the storage-layer owner filter by manually switching the user_context contextvar between two users. Verifies the safety invariant: After a repository write with owner_id=A, a subsequent read with owner_id=B must not return the row, and vice versa. Covers all 4 tables that own user-scoped data: TC-API-17 threads_meta — read, search, update, delete cross-user TC-API-18 runs — get, list_by_thread, delete cross-user TC-API-19 run_events — list_messages, list_events, count_messages, delete_by_thread (CRITICAL: raw conversation content leak vector) TC-API-20 feedback — get, list_by_run, delete cross-user Plus two meta-tests verifying the sentinel pattern itself: - AUTO + unset contextvar raises RuntimeError - explicit owner_id=None bypasses the filter (migration escape hatch) Architecture note ----------------- These tests bypass the HTTP layer by design. The full chain (cookie → middleware → contextvar → repository) is covered piecewise: - test_auth_middleware.py: middleware sets contextvar from cookies - test_owner_isolation.py: repositories enforce isolation when contextvar is set to different users Together they prove the end-to-end safety property without the ceremony of spinning up a full TestClient + in-memory DB for every router endpoint. Tests pass: 231 (full auth + persistence + isolation suite) Lint: clean * refactor(auth): migrate user repository to SQLAlchemy ORM Move the users table into the shared persistence engine so auth matches the pattern of threads_meta, runs, run_events, and feedback — one engine, one session factory, one schema init codepath. New files --------- - persistence/user/__init__.py, persistence/user/model.py: UserRow ORM class with partial unique index on (oauth_provider, oauth_id) - Registered in persistence/models/__init__.py so Base.metadata.create_all() picks it up Modified -------- - auth/repositories/sqlite.py: rewritten as async SQLAlchemy, identical constructor pattern to the other four repositories (def __init__(self, session_factory) + self._sf = session_factory) - auth/config.py: drop users_db_path field — storage is configured through config.database like every other table - deps.py/get_local_provider: construct SQLiteUserRepository with the shared session factory, fail fast if engine is not initialised - tests/test_auth.py: rewrite test_sqlite_round_trip_new_fields to use the shared engine (init_engine + close_engine in a tempdir) - tests/test_auth_type_system.py: add per-test autouse fixture that spins up a scratch engine and resets deps._cached_* singletons * refactor(auth): remove SQL orphan migration (unused in supported scenarios) The _migrate_orphan_sql_tables helper existed to bind NULL owner_id rows in threads_meta, runs, run_events, and feedback to the admin on first boot. But in every supported upgrade path, it's a no-op: 1. Fresh install: create_all builds fresh tables, no legacy rows 2. No-auth → with-auth (no existing persistence DB): persistence tables are created fresh by create_all, no legacy rows 3. No-auth → with-auth (has existing persistence DB from #1930): NOT a supported upgrade path — "有 DB 到有 DB" schema evolution is out of scope; users wipe DB or run manual ALTER So the SQL orphan migration never has anything to do in the supported matrix. Delete the function, simplify _ensure_admin_user from a 3-step pipeline to a 2-step one (admin creation + LangGraph store orphan migration only). LangGraph store orphan migration stays: it serves the real "no-auth → with-auth" upgrade path where a user's existing LangGraph thread metadata has no owner_id field and needs to be stamped with the newly-created admin's id. Tests: 284 passed (auth + persistence + isolation) Lint: clean * security(auth): write initial admin password to 0600 file instead of logs CodeQL py/clear-text-logging-sensitive-data flagged 3 call sites that logged the auto-generated admin password to stdout via logger.info(). Production log aggregators (ELK/Splunk/etc) would have captured those cleartext secrets. Replace with a shared helper that writes to .deer-flow/admin_initial_credentials.txt with mode 0600, and log only the path. New file -------- - app/gateway/auth/credential_file.py: write_initial_credentials() helper. Takes email, password, and a "initial"/"reset" label. Creates .deer-flow/ if missing, writes a header comment plus the email+password, chmods 0o600, returns the absolute Path. Modified -------- - app/gateway/app.py: both _ensure_admin_user paths (fresh creation + needs_setup password reset) now write to file and log the path - app/gateway/auth/reset_admin.py: rewritten to use the shared ORM repo (SQLiteUserRepository with session_factory) and the credential_file helper. The previous implementation was broken after the earlier ORM refactor — it still imported _get_users_conn and constructed SQLiteUserRepository() without a session factory. No tests changed — the three password-log sites are all exercised via existing test_ensure_admin.py which checks that startup succeeds, not that a specific string appears in logs. CodeQL alerts 272, 283, 284: all resolved. * security(auth): strict JWT validation in middleware (fix junk cookie bypass) AUTH_TEST_PLAN test 7.5.8 expects junk cookies to be rejected with 401. The previous middleware behaviour was "presence-only": check that some access_token cookie exists, then pass through. In combination with my Task-12 decision to skip @require_auth decorators on routes, this created a gap where a request with any cookie-shaped string (e.g. access_token=not-a-jwt) would bypass authentication on routes that do not touch the repository (/api/models, /api/mcp/config, /api/memory, /api/skills, …). Fix: middleware now calls get_current_user_from_request() strictly and catches the resulting HTTPException to render a 401 with the proper fine-grained error code (token_invalid, token_expired, user_not_found, …). On success it stamps request.state.user and the contextvar so repository-layer owner filters work downstream. The 4 old "_with_cookie_passes" tests in test_auth_middleware.py were written for the presence-only behaviour; they asserted that a junk cookie would make the handler return 200. They are renamed to "_with_junk_cookie_rejected" and their assertions flipped to 401. The negative path (no cookie → 401 not_authenticated) is unchanged. Verified: no cookie → 401 not_authenticated junk cookie → 401 token_invalid (the fixed bug) expired cookie → 401 token_expired Tests: 284 passed (auth + persistence + isolation) Lint: clean * security(auth): wire @require_permission(owner_check=True) on isolation routes Apply the require_permission decorator to all 28 routes that take a {thread_id} path parameter. Combined with the strict middleware (previous commit), this gives the double-layer protection that AUTH_TEST_PLAN test 7.5.9 documents: Layer 1 (AuthMiddleware): cookie + JWT validation, rejects junk cookies and stamps request.state.user Layer 2 (@require_permission with owner_check=True): per-resource ownership verification via ThreadMetaStore.check_access — returns 404 if a different user owns the thread The decorator's owner_check branch is rewritten to use the SQL thread_meta_repo (the 2.0-rc persistence layer) instead of the LangGraph store path that PR #1728 used (_store_get / get_store in routers/threads.py). The inject_record convenience is dropped — no caller in 2.0 needs the LangGraph blob, and the SQL repo has a different shape. Routes decorated (28 total): - threads.py: delete, patch, get, get-state, post-state, post-history - thread_runs.py: post-runs, post-runs-stream, post-runs-wait, list_runs, get_run, cancel_run, join_run, stream_existing_run, list_thread_messages, list_run_messages, list_run_events, thread_token_usage - feedback.py: create, list, stats, delete - uploads.py: upload (added Request param), list, delete - artifacts.py: get_artifact - suggestions.py: generate (renamed body parameter to avoid conflict with FastAPI Request) Test fixes: - test_suggestions_router.py: bypass the decorator via __wrapped__ (the unit tests cover parsing logic, not auth — no point spinning up a thread_meta_repo just to test JSON unwrapping) - test_auth_middleware.py 4 fake-cookie tests: already updated in the previous commit (745bf432) Tests: 293 passed (auth + persistence + isolation + suggestions) Lint: clean * security(auth): defense-in-depth fixes from release validation pass Eight findings caught while running the AUTH_TEST_PLAN end-to-end against the deployed sg_dev stack. Each is a pre-condition for shipping release/2.0-rc that the previous PRs missed. Backend hardening - routers/auth.py: rate limiter X-Real-IP now requires AUTH_TRUSTED_PROXIES whitelist (CIDR/IP allowlist). Without nginx in front, the previous code honored arbitrary X-Real-IP, letting an attacker rotate the header to fully bypass the per-IP login lockout. - routers/auth.py: 36-entry common-password blocklist via Pydantic field_validator on RegisterRequest + ChangePasswordRequest. The shared _validate_strong_password helper keeps the constraint in one place. - routers/threads.py: ThreadCreateRequest + ThreadPatchRequest strip server-reserved metadata keys (owner_id, user_id) via Pydantic field_validator so a forged value can never round-trip back to other clients reading the same thread. The actual ownership invariant stays on the threads_meta row; this closes the metadata-blob echo gap. - authz.py + thread_meta/sql.py: require_permission gains a require_existing flag plumbed through check_access(require_existing=True). Destructive routes (DELETE/PATCH/state-update/runs/feedback) now treat a missing thread_meta row as 404 instead of "untracked legacy thread, allow", closing the cross-user delete-idempotence gap where any user could successfully DELETE another user's deleted thread. - repositories/sqlite.py + base.py: update_user raises UserNotFoundError on a vanished row instead of silently returning the input. Concurrent delete during password reset can no longer look like a successful update. - runtime/user_context.py: resolve_owner_id() coerces User.id (UUID) to str at the contextvar boundary so SQLAlchemy String(64) columns can bind it. The whole 2.0-rc isolation pipeline was previously broken end-to-end (POST /api/threads → 500 "type 'UUID' is not supported"). - persistence/engine.py: SQLAlchemy listener enables PRAGMA journal_mode=WAL, synchronous=NORMAL, foreign_keys=ON on every new SQLite connection. TC-UPG-06 in the test plan expects WAL; previous code shipped with the default 'delete' journal. - auth_middleware.py: stamp request.state.auth = AuthContext(...) so @require_permission's short-circuit fires; previously every isolation request did a duplicate JWT decode + users SELECT. Also unifies the 401 payload through AuthErrorResponse(...).model_dump(). - app.py: _ensure_admin_user restructure removes the noqa F821 scoping bug where 'password' was referenced outside the branch that defined it. New _announce_credentials helper absorbs the duplicate log block in the fresh-admin and reset-admin branches. * fix(frontend+nginx): rollout CSRF on every state-changing client path The frontend was 100% broken in gateway-pro mode for any user trying to open a specific chat thread. Three cumulative bugs each silently masked the next. LangGraph SDK CSRF gap (api-client.ts) - The Client constructor took only apiUrl, no defaultHeaders, no fetch interceptor. The SDK's internal fetch never sent X-CSRF-Token, so every state-changing /api/langgraph-compat/* call (runs/stream, threads/search, threads/{tid}/history, ...) hit CSRFMiddleware and got 403 before reaching the auth check. UI symptom: empty thread page with no error message; the SPA's hooks swallowed the rejection. - Fix: pass an onRequest hook that injects X-CSRF-Token from the csrf_token cookie per request. Reading the cookie per call (not at construction time) handles login / logout / password-change cookie rotation transparently. The SDK's prepareFetchOptions calls onRequest for both regular requests AND streaming/SSE/reconnect, so the same hook covers runs.stream and runs.joinStream. Raw fetch CSRF gap (7 files) - Audit: 11 frontend fetch sites, only 2 included CSRF (login/setup + account-settings change-password). The other 7 routed through raw fetch() with no header — suggestions, memory, agents, mcp, skills, uploads, and the local thread cleanup hook all 403'd silently. - Fix: enhance fetcher.ts:fetchWithAuth to auto-inject X-CSRF-Token on POST/PUT/DELETE/PATCH from a single shared readCsrfCookie() helper. Convert all 7 raw fetch() callers to fetchWithAuth so the contract is centrally enforced. api-client.ts and fetcher.ts share readCsrfCookie + STATE_CHANGING_METHODS to avoid drift. nginx routing + buffering (nginx.local.conf) - The auth feature shipped without updating the nginx config: per-API explicit location blocks but no /api/v1/auth/, /api/feedback, /api/runs. The frontend's client-side fetches to /api/v1/auth/login/local 404'd from the Next.js side because nginx routed /api/* to the frontend. - Fix: add catch-all `location /api/` that proxies to the gateway. nginx longest-prefix matching keeps the explicit blocks (/api/models, /api/threads regex, /api/langgraph/, ...) winning for their paths. - Fix: disable proxy_buffering + proxy_request_buffering for the frontend `location /` block. Without it, nginx tries to spool large Next.js chunks into /var/lib/nginx/proxy (root-owned) and fails with Permission denied → ERR_INCOMPLETE_CHUNKED_ENCODING → ChunkLoadError. * test(auth): release-validation test infra and new coverage Test fixtures and unit tests added during the validation pass. Router test helpers (NEW: tests/_router_auth_helpers.py) - make_authed_test_app(): builds a FastAPI test app with a stub middleware that stamps request.state.user + request.state.auth and a permissive thread_meta_repo mock. TestClient-based router tests (test_artifacts_router, test_threads_router) use it instead of bare FastAPI() so the new @require_permission(owner_check=True) decorators short-circuit cleanly. - call_unwrapped(): walks the __wrapped__ chain to invoke the underlying handler without going through the authz wrappers. Direct-call tests (test_uploads_router) use it. Typed with ParamSpec so the wrapped signature flows through. Backend test additions - test_auth.py: 7 tests for the new _get_client_ip trust model (no proxy / trusted proxy / untrusted peer / XFF rejection / invalid CIDR / no client). 5 tests for the password blocklist (literal, case-insensitive, strong password accepted, change-password binding, short-password length-check still fires before blocklist). test_update_user_raises_when_row_concurrently_deleted: closes a shipped-without-coverage gap on the new UserNotFoundError contract. - test_thread_meta_repo.py: 4 tests for check_access(require_existing=True) — strict missing-row denial, strict owner match, strict owner mismatch, strict null-owner still allowed (shared rows survive the tightening). - test_ensure_admin.py: 3 tests for _migrate_orphaned_threads / _iter_store_items pagination, covering the TC-UPG-02 upgrade story end-to-end via mock store. Closes the gap where the cursor pagination was untested even though the previous PR rewrote it. - test_threads_router.py: 5 tests for _strip_reserved_metadata (owner_id removal, user_id removal, safe-keys passthrough, empty input, both-stripped). - test_auth_type_system.py: replace "password123" fixtures with Tr0ub4dor3a / AnotherStr0ngPwd! so the new password blocklist doesn't reject the test data. * docs(auth): refresh TC-DOCKER-05 + document Docker validation gap - AUTH_TEST_PLAN.md TC-DOCKER-05: the previous expectation ("admin password visible in docker logs") was stale after the simplify pass that moved credentials to a 0600 file. The grep "Password:" check would have silently failed and given a false sense of coverage. New expectation matches the actual file-based path: 0600 file in DEER_FLOW_HOME, log shows the path (not the secret), reverse-grep asserts no leaked password in container logs. - NEW: docs/AUTH_TEST_DOCKER_GAP.md documents the only un-executed block in the test plan (TC-DOCKER-01..06). Reason: sg_dev validation host has no Docker daemon installed. The doc maps each Docker case to an already-validated bare-metal equivalent (TC-1.1, TC-REENT-01, TC-API-02 etc.) so the gap is auditable, and includes pre-flight reproduction steps for whoever has Docker available. --------- Co-authored-by: greatmengqi <chenmengqi.0376@bytedance.com>
DeerFlow Backend
DeerFlow is a LangGraph-based AI super agent with sandbox execution, persistent memory, and extensible tool integration. The backend enables AI agents to execute code, browse the web, manage files, delegate tasks to subagents, and retain context across conversations - all in isolated, per-thread environments.
Architecture
┌──────────────────────────────────────┐
│ Nginx (Port 2026) │
│ Unified reverse proxy │
└───────┬──────────────────┬───────────┘
│ │
/api/langgraph/* │ │ /api/* (other)
▼ ▼
┌────────────────────┐ ┌────────────────────────┐
│ LangGraph Server │ │ Gateway API (8001) │
│ (Port 2024) │ │ FastAPI REST │
│ │ │ │
│ ┌────────────────┐ │ │ Models, MCP, Skills, │
│ │ Lead Agent │ │ │ Memory, Uploads, │
│ │ ┌──────────┐ │ │ │ Artifacts │
│ │ │Middleware│ │ │ └────────────────────────┘
│ │ │ Chain │ │ │
│ │ └──────────┘ │ │
│ │ ┌──────────┐ │ │
│ │ │ Tools │ │ │
│ │ └──────────┘ │ │
│ │ ┌──────────┐ │ │
│ │ │Subagents │ │ │
│ │ └──────────┘ │ │
│ └────────────────┘ │
└────────────────────┘
Request Routing (via Nginx):
/api/langgraph/*→ LangGraph Server - agent interactions, threads, streaming/api/*(other) → Gateway API - models, MCP, skills, memory, artifacts, uploads, thread-local cleanup/(non-API) → Frontend - Next.js web interface
Core Components
Lead Agent
The single LangGraph agent (lead_agent) is the runtime entry point, created via make_lead_agent(config). It combines:
- Dynamic model selection with thinking and vision support
- Middleware chain for cross-cutting concerns (9 middlewares)
- Tool system with sandbox, MCP, community, and built-in tools
- Subagent delegation for parallel task execution
- System prompt with skills injection, memory context, and working directory guidance
Middleware Chain
Middlewares execute in strict order, each handling a specific concern:
| # | Middleware | Purpose |
|---|---|---|
| 1 | ThreadDataMiddleware | Creates per-thread isolated directories (workspace, uploads, outputs) |
| 2 | UploadsMiddleware | Injects newly uploaded files into conversation context |
| 3 | SandboxMiddleware | Acquires sandbox environment for code execution |
| 4 | SummarizationMiddleware | Reduces context when approaching token limits (optional) |
| 5 | TodoListMiddleware | Tracks multi-step tasks in plan mode (optional) |
| 6 | TitleMiddleware | Auto-generates conversation titles after first exchange |
| 7 | MemoryMiddleware | Queues conversations for async memory extraction |
| 8 | ViewImageMiddleware | Injects image data for vision-capable models (conditional) |
| 9 | ClarificationMiddleware | Intercepts clarification requests and interrupts execution (must be last) |
Sandbox System
Per-thread isolated execution with virtual path translation:
- Abstract interface:
execute_command,read_file,write_file,list_dir - Providers:
LocalSandboxProvider(filesystem) andAioSandboxProvider(Docker, in community/) - Virtual paths:
/mnt/user-data/{workspace,uploads,outputs}→ thread-specific physical directories - Skills path:
/mnt/skills→deer-flow/skills/directory - Skills loading: Recursively discovers nested
SKILL.mdfiles underskills/{public,custom}and preserves nested container paths - File-write safety:
str_replaceserializes read-modify-write per(sandbox.id, path)so isolated sandboxes keep concurrency even when virtual paths match - Tools:
bash,ls,read_file,write_file,str_replace(bashis disabled by default when usingLocalSandboxProvider; useAioSandboxProviderfor isolated shell access)
Subagent System
Async task delegation with concurrent execution:
- Built-in agents:
general-purpose(full toolset) andbash(command specialist, exposed only when shell access is available) - Concurrency: Max 3 subagents per turn, 15-minute timeout
- Execution: Background thread pools with status tracking and SSE events
- Flow: Agent calls
task()tool → executor runs subagent in background → polls for completion → returns result
Memory System
LLM-powered persistent context retention across conversations:
- Automatic extraction: Analyzes conversations for user context, facts, and preferences
- Structured storage: User context (work, personal, top-of-mind), history, and confidence-scored facts
- Debounced updates: Batches updates to minimize LLM calls (configurable wait time)
- System prompt injection: Top facts + context injected into agent prompts
- Storage: JSON file with mtime-based cache invalidation
Tool Ecosystem
| Category | Tools |
|---|---|
| Sandbox | bash, ls, read_file, write_file, str_replace |
| Built-in | present_files, ask_clarification, view_image, task (subagent) |
| Community | Tavily (web search), Jina AI (web fetch), Firecrawl (scraping), DuckDuckGo (image search) |
| MCP | Any Model Context Protocol server (stdio, SSE, HTTP transports) |
| Skills | Domain-specific workflows injected via system prompt |
Gateway API
FastAPI application providing REST endpoints for frontend integration:
| Route | Purpose |
|---|---|
GET /api/models |
List available LLM models |
GET/PUT /api/mcp/config |
Manage MCP server configurations |
GET/PUT /api/skills |
List and manage skills |
POST /api/skills/install |
Install skill from .skill archive |
GET /api/memory |
Retrieve memory data |
POST /api/memory/reload |
Force memory reload |
GET /api/memory/config |
Memory configuration |
GET /api/memory/status |
Combined config + data |
POST /api/threads/{id}/uploads |
Upload files (auto-converts PDF/PPT/Excel/Word to Markdown, rejects directory paths) |
GET /api/threads/{id}/uploads/list |
List uploaded files |
DELETE /api/threads/{id} |
Delete DeerFlow-managed local thread data after LangGraph thread deletion; unexpected failures are logged server-side and return a generic 500 detail |
GET /api/threads/{id}/artifacts/{path} |
Serve generated artifacts |
IM Channels
The IM bridge supports Feishu, Slack, and Telegram. Slack and Telegram still use the final runs.wait() response path, while Feishu now streams through runs.stream(["messages-tuple", "values"]) and updates a single in-thread card in place.
For Feishu card updates, DeerFlow stores the running card's message_id per inbound message and patches that same card until the run finishes, preserving the existing OK / DONE reaction flow.
Quick Start
Prerequisites
- Python 3.12+
- uv package manager
- API keys for your chosen LLM provider
Installation
cd deer-flow
# Copy configuration files
cp config.example.yaml config.yaml
# Install backend dependencies
cd backend
make install
Configuration
Edit config.yaml in the project root:
models:
- name: gpt-4o
display_name: GPT-4o
use: langchain_openai:ChatOpenAI
model: gpt-4o
api_key: $OPENAI_API_KEY
supports_thinking: false
supports_vision: true
- name: gpt-5-responses
display_name: GPT-5 (Responses API)
use: langchain_openai:ChatOpenAI
model: gpt-5
api_key: $OPENAI_API_KEY
use_responses_api: true
output_version: responses/v1
supports_vision: true
Set your API keys:
export OPENAI_API_KEY="your-api-key-here"
Running
Full Application (from project root):
make dev # Starts LangGraph + Gateway + Frontend + Nginx
Access at: http://localhost:2026
Backend Only (from backend directory):
# Terminal 1: LangGraph server
make dev
# Terminal 2: Gateway API
make gateway
Direct access: LangGraph at http://localhost:2024, Gateway at http://localhost:8001
Project Structure
backend/
├── src/
│ ├── agents/ # Agent system
│ │ ├── lead_agent/ # Main agent (factory, prompts)
│ │ ├── middlewares/ # 9 middleware components
│ │ ├── memory/ # Memory extraction & storage
│ │ └── thread_state.py # ThreadState schema
│ ├── gateway/ # FastAPI Gateway API
│ │ ├── app.py # Application setup
│ │ └── routers/ # 6 route modules
│ ├── sandbox/ # Sandbox execution
│ │ ├── local/ # Local filesystem provider
│ │ ├── sandbox.py # Abstract interface
│ │ ├── tools.py # bash, ls, read/write/str_replace
│ │ └── middleware.py # Sandbox lifecycle
│ ├── subagents/ # Subagent delegation
│ │ ├── builtins/ # general-purpose, bash agents
│ │ ├── executor.py # Background execution engine
│ │ └── registry.py # Agent registry
│ ├── tools/builtins/ # Built-in tools
│ ├── mcp/ # MCP protocol integration
│ ├── models/ # Model factory
│ ├── skills/ # Skill discovery & loading
│ ├── config/ # Configuration system
│ ├── community/ # Community tools & providers
│ ├── reflection/ # Dynamic module loading
│ └── utils/ # Utilities
├── docs/ # Documentation
├── tests/ # Test suite
├── langgraph.json # LangGraph server configuration
├── pyproject.toml # Python dependencies
├── Makefile # Development commands
└── Dockerfile # Container build
Configuration
Main Configuration (config.yaml)
Place in project root. Config values starting with $ resolve as environment variables.
Key sections:
models- LLM configurations with class paths, API keys, thinking/vision flagstools- Tool definitions with module paths and groupstool_groups- Logical tool groupingssandbox- Execution environment providerskills- Skills directory pathstitle- Auto-title generation settingssummarization- Context summarization settingssubagents- Subagent system (enabled/disabled)memory- Memory system settings (enabled, storage, debounce, facts limits)
Provider note:
models[*].usereferences provider classes by module path (for examplelangchain_openai:ChatOpenAI).- If a provider module is missing, DeerFlow now returns an actionable error with install guidance (for example
uv add langchain-google-genai).
Extensions Configuration (extensions_config.json)
MCP servers and skill states in a single file:
{
"mcpServers": {
"github": {
"enabled": true,
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {"GITHUB_TOKEN": "$GITHUB_TOKEN"}
},
"secure-http": {
"enabled": true,
"type": "http",
"url": "https://api.example.com/mcp",
"oauth": {
"enabled": true,
"token_url": "https://auth.example.com/oauth/token",
"grant_type": "client_credentials",
"client_id": "$MCP_OAUTH_CLIENT_ID",
"client_secret": "$MCP_OAUTH_CLIENT_SECRET"
}
}
},
"skills": {
"pdf-processing": {"enabled": true}
}
}
Environment Variables
DEER_FLOW_CONFIG_PATH- Override config.yaml locationDEER_FLOW_EXTENSIONS_CONFIG_PATH- Override extensions_config.json location- Model API keys:
OPENAI_API_KEY,ANTHROPIC_API_KEY,DEEPSEEK_API_KEY, etc. - Tool API keys:
TAVILY_API_KEY,GITHUB_TOKEN, etc.
LangSmith Tracing
DeerFlow has built-in LangSmith integration for observability. When enabled, all LLM calls, agent runs, tool executions, and middleware processing are traced and visible in the LangSmith dashboard.
Setup:
- Sign up at smith.langchain.com and create a project.
- Add the following to your
.envfile in the project root:
LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=lsv2_pt_xxxxxxxxxxxxxxxx
LANGSMITH_PROJECT=xxx
Legacy variables: The LANGCHAIN_TRACING_V2, LANGCHAIN_API_KEY, LANGCHAIN_PROJECT, and LANGCHAIN_ENDPOINT variables are also supported for backward compatibility. LANGSMITH_* variables take precedence when both are set.
Langfuse Tracing
DeerFlow also supports Langfuse observability for LangChain-compatible runs.
Add the following to your .env file:
LANGFUSE_TRACING=true
LANGFUSE_PUBLIC_KEY=pk-lf-xxxxxxxxxxxxxxxx
LANGFUSE_SECRET_KEY=sk-lf-xxxxxxxxxxxxxxxx
LANGFUSE_BASE_URL=https://cloud.langfuse.com
If you are using a self-hosted Langfuse deployment, set LANGFUSE_BASE_URL to your Langfuse host.
Dual Provider Behavior
If both LangSmith and Langfuse are enabled, DeerFlow initializes and attaches both callbacks so the same run data is reported to both systems.
If a provider is explicitly enabled but required credentials are missing, or the provider callback cannot be initialized, DeerFlow raises an error when tracing is initialized during model creation instead of silently disabling tracing.
Docker: In docker-compose.yaml, tracing is disabled by default (LANGSMITH_TRACING=false). Set LANGSMITH_TRACING=true and/or LANGFUSE_TRACING=true in your .env, together with the required credentials, to enable tracing in containerized deployments.
Development
Commands
make install # Install dependencies
make dev # Run LangGraph server (port 2024)
make gateway # Run Gateway API (port 8001)
make lint # Run linter (ruff)
make format # Format code (ruff)
Code Style
- Linter/Formatter:
ruff - Line length: 240 characters
- Python: 3.12+ with type hints
- Quotes: Double quotes
- Indentation: 4 spaces
Testing
uv run pytest
Technology Stack
- LangGraph (1.0.6+) - Agent framework and multi-agent orchestration
- LangChain (1.2.3+) - LLM abstractions and tool system
- FastAPI (0.115.0+) - Gateway REST API
- langchain-mcp-adapters - Model Context Protocol support
- agent-sandbox - Sandboxed code execution
- markitdown - Multi-format document conversion
- tavily-python / firecrawl-py - Web search and scraping
Documentation
- Configuration Guide
- Architecture Details
- API Reference
- File Upload
- Path Examples
- Context Summarization
- Plan Mode
- Setup Guide
License
See the LICENSE file in the project root.
Contributing
See CONTRIBUTING.md for contribution guidelines.