deer-flow/backend/AGENTS.md

AGENTS.md

Project Overview

The backend runs a LangGraph-based super agent with sandbox execution, persistent memory, subagent delegation, and extensible tools in isolated per-thread environments.

Architecture:

• Gateway API (port 8001): REST API plus embedded LangGraph-compatible agent runtime
• Frontend (port 3000): Next.js web interface
• Nginx (port 2026): Unified reverse proxy entry point
• Provisioner (port 8002, optional in Docker dev): Started only when sandbox is configured for provisioner/Kubernetes mode

Runtime:

• make dev, Docker dev, and production all run the agent runtime in Gateway via RunManager + run_agent() + StreamBridge (packages/harness/deerflow/runtime/). Nginx exposes that runtime at /api/langgraph/* and rewrites it to Gateway's native /api/* routers.
• Gateway streams write_file and str_replace argument deltas in bounded batches for multi-mode messages-tuple consumers; single-mode message consumers retain the original per-chunk contract. Non-message frames flush pending batches, and values remains an optional complete-state snapshot rather than a prerequisite for batching.
• With stream_subgraphs, subgraph frames keep their namespace in the SSE event name (values|<ns>, LangGraph Platform style) instead of impersonating root frames — a delegated subagent inherits the parent checkpoint namespace, so publishing its values snapshot as bare values replaces the whole thread view in SDK clients (#4399). Root-only consumers (file-tool chunk batcher, subagent event persistence, LLM error-fallback detection) ignore namespaced frames. The web frontend does not request subgraph streaming; subtask progress rides root-namespace task_* custom events.
• Background subagent identity is deliberately split: the provider tool_call_id remains the correlation key for ToolMessage, task_* SSE events, persisted lifecycle events, frontend cards, and the public ExtensionData.scope_id contract (stored as SubagentResult.external_task_id), while SubagentExecutor.execute_async() generates a full server-side execution_id for SubagentResult.task_id, the process-wide registry, polling, cancellation, timeout handling, and cleanup. Provider IDs are not globally unique across parent runs, so they must never become registry ownership keys; scheduler closures retain their own SubagentResult rather than resolving ownership again through the mutable registry. Terminal subagent token usage travels in the current run's ToolMessage.additional_kwargs and is attributed from message state, never through a process-global provider-ID cache.
• Scheduled tasks dispatch through the normal Gateway run path. launch_scheduled_thread_run reads get_app_config() at dispatch and passes scheduler.recursion_limit (default 1000, matching the web UI; clamped by max_recursion_limit), so YAML changes apply on the next run without restarting Gateway.
• Run-history status filters are occurrence states, not task states. ScheduledTaskRunStatus in persistence/scheduled_tasks/model.py is the shared API/repository vocabulary and must match the active and terminal occurrence-status sets. Keep owner lookup before reading history, and apply SQL task/status predicates before pagination; omitted status preserves the existing response.
• The background scheduler is single-instance by default. scheduler.multi_instance=true opts into lease-aware recovery across Gateway instances and requires shared Postgres, run_ownership.heartbeat_enabled=true, and run_events.backend=db; otherwise startup rejects the configuration. Live scheduled runs are preserved when a peer starts; expired launch claims return to the durable queue, expired run leases are atomically taken over, stale launch writes are fenced by lease ownership, and the Postgres advisory-locked budget makes max_concurrent_runs a shared global cap for launching/running rows.
• The multi-process startup gate also fires on deployment.multi_instance: true; contract in docs/CONFIGURATION.md.
• Long-running MCP work uses a separate durable task runtime (McpTaskService + mcp_tasks, lease-based recovery) rather than keeping remote task IDs or status polling inside the Agent loop; only submit remains Agent-visible, the database is the source of truth, and ThreadState receives only a bounded current-thread projection. Full contract — including notification retries, dead-lettering, and the cancel endpoint's worker-stopped 503 (leases, cancellation fencing, delivery idempotency, management-tool exposure): packages/harness/deerflow/mcp/AGENTS.md.
• Scheduled-task dispatch permits one active occurrence per task via uq_scheduled_task_run_active (task_id WHERE status IN ('queued','launching','running')). Durable queued rows survive restarts; only lease-fenced launching may call Gateway launch; running references the durable run. Stable admission idempotency keys reuse that run after recovery. Reused-thread ConflictError returns launching to queued; other launch errors become failed. Atomic queue claims enforce max_concurrent_runs and the per-owner cap, excluding waiting rows; the budget count and its UPDATE are separate statements, so writers must serialize before the count (Postgres advisory lock; SQLite BEGIN IMMEDIATE, whose deferred transaction otherwise reserves the writer only at the UPDATE) or claims on distinct rows overshoot the cap. Repeated triggers coalesce; same-thread FIFO blocks behind older active rows. Queue admission, PATCH/resume, pause and delete lock the parent before the occurrence, freezing active task definitions. Pause/delete atomically cancel queued work but reject launching/running; PATCH/resume reject all active states. Only queued conflicts offer pause cancellation. Manual triggers may queue/run while paused.
• Scheduled-task lifecycle, recovery and queue timeout: persistence guide.
• POST /api/scheduled-tasks/preview-cron requires authenticated threads:read. Bounded cron previews call the shared scheduler calculator in asyncio.to_thread, preserving its DST semantics. Capture the optional aware reference once; return UTC and offset-bearing local occurrences without acquiring task/thread/run stores or dispatching work. This advisory API does not reserve execution.
• Gateway MCP/skill updates write extensions_config.json at runtime. Production Compose mounts it read-write (config.yaml stays :ro); Helm seeds a writable home-volume copy. Hold both extensions_config_write_lock and the sidecar advisory extensions_config_file_lock during read-modify-write to prevent lost updates across workers. atomic_write_extensions_config uses temp-file replacement; only mount-point EBUSY permits a non-atomic in-place overwrite (crashes can truncate). Warn about this fallback once per target; propagate other errors. get_extensions_config() revalidates reads by file signature; see config caching. Tests: tests/test_compose_extensions_config_writable.py, tests/test_extensions_config_atomic_write.py, tests/test_helm_extensions_config_writable.py.
• MCP cache reset scope and discovery fencing: MCP guide.

Project Structure (backend/ subtree; the root guide maps the full repo):

backend/
├── Makefile                    # Backend-only commands (dev, gateway, lint)
├── langgraph.json              # LangGraph Studio graph configuration
├── packages/
│   ├── extension-api/          # public, host-independent extension contracts (import: deerflow_extension_api.*)
│   └── harness/                # deerflow-harness package (import: deerflow.*)
│       └── deerflow/
│           ├── agents/         # LangGraph agent system (lead_agent/, middlewares/ — see Middleware Chain, memory/, thread_state.py)
│           ├── sandbox/        # Sandbox execution system
│           ├── subagents/      # Subagent delegation (builtins/, executor.py, registry.py)
│           ├── tools/builtins/ # Built-in tools (present_files, ask_clarification, view_image, review_skill_package)
│           ├── mcp/            # MCP integration (tools, cache, client)
│           ├── integrations/   # Managed first-party integration installers (e.g. Lark CLI skill pack)
│           ├── extensions/     # Python plugin loader, registry, placement, and isolation
│           ├── models/         # Model factory with thinking/vision support
│           ├── skills/         # Skills discovery, loading, parsing
│           ├── config/         # Configuration system
│           ├── community/      # Community tools (search/fetch/scrape, image search, AIO sandbox)
│           ├── reflection/     # Dynamic module loading (resolve_variable, resolve_class)
│           ├── utils/          # Utilities (network, readability)
│           └── client.py       # Embedded Python client (DeerFlowClient)
├── app/                        # Application layer (import: app.*)
│   ├── gateway/                # FastAPI Gateway API (app.py, routers/)
│   └── channels/               # IM platform integrations
├── scripts/benchmark/          # Standalone reproducible backend benchmarks
├── tests/                      # Test suite
└── docs/                       # Documentation

Important Development Guidelines

Documentation Update Policy

Every code change must keep docs accurate and current: update README.md for user-facing behavior and the relevant AGENTS.md for development changes. CLAUDE.md imports AGENTS.md; do not edit the shim.

Backend Benchmarks

scripts/benchmark/context_snapshot/: explicit run-live needs provider env vars; summarize and pytest are offline. See its README for the protocol.

Benchmarks in scripts/benchmark/ must be standalone and reproducible. Import production functions; never duplicate them or introduce an alternative runtime.

• Pin every external dataset by immutable revision and SHA-256. Callers provide the local dataset path; evaluation commands must not silently download data.
• Never commit upstream dataset text, credentials, complete provider requests, or response headers. Committed manifests may contain stable IDs and source locators. Synthetic cases must identify themselves as synthetic.
• Read provider credentials and endpoints from named environment variables. Version model IDs, inference parameters, prompts, retry rules, clocks, and random seeds in the evaluation config.
• Public raw results may contain case IDs, policy decisions, model hypotheses, grades, and non-secret response metadata. Keep dataset questions, reference answers, memory content, and full provider payloads in ignored local run directories.
• Use fixed clocks and deterministic ordering for offline selection. Results must record the config, manifest, prompt, dataset, and git revisions used.

scripts/benchmark/deermem_eviction/ evaluates the production select_facts_for_capacity() implementation used by DeerMem. It compares only the historical confidence policy and PR #4789's opt-in hybrid-v1; do not add another eviction strategy to this evaluation. Run its offline checks from backend/:

PYTHONPATH=. uv run python -m scripts.benchmark.deermem_eviction validate-contracts
PYTHONPATH=. uv run python -m scripts.benchmark.deermem_eviction validate --dataset "$LONGMEMEVAL_ORACLE_PATH"
PYTHONPATH=. uv run python -m scripts.benchmark.deermem_eviction run-policy \
  --dataset "$LONGMEMEVAL_ORACLE_PATH" \
  --output-dir /tmp/deermem-eviction-policy-run
PYTHONPATH=. uv run pytest tests/test_bench_deermem_eviction_*.py -q

The offline test suite must not require network access, provider credentials, or the LongMemEval dataset. Small LongMemEval-shaped fixtures must be synthetic and generated by tests.

scripts/benchmark/concurrency/ measures multi-process contention on the users table (N separate OS processes, not asyncio tasks) for SQLite vs Postgres -- the scenario CONFIGURATION.md requires Postgres for. worker.py connects directly via SQLAlchemy (skipping the ~8.5s Alembic bootstrap the orchestrator already ran once) and mirrors the app's per-connection SQLite PRAGMAs; run_concurrency_bench.py seeds a disposable per-run Postgres schema, synchronises workers on a READY/GO barrier before timing, and exits non-zero on any crash, short op count, or errors > 0. Postgres runs need a throwaway database via --pg-url; nothing here touches public. Run from backend/:

uv run python scripts/benchmark/concurrency/run_concurrency_bench.py \
  --backend sqlite --workers 2,4,8,16 --ops-per-worker 50 --read-ratio 0.7
uv run pytest tests/test_bench_concurrency.py tests/test_bench_worker.py -q

Commands

Root directory (for full application):

make check      # Check system requirements
make install    # Install all dependencies (frontend + backend)
make extension-install SOURCE=...  # Install and enable a trusted Python extension
make extension-upgrade SOURCE=...  # Replace an installed extension and keep its config
make extension-list                # List configured Python extensions
make extension-enable NAME=...     # Enable an installed extension
make extension-disable NAME=...    # Disable an extension without uninstalling it
make extension-remove NAME=...     # Remove a managed extension
make detect-thread-boundaries  # Inventory backend executor/thread/event-loop boundaries
make dev        # Start all services (Gateway + Frontend + Nginx), with config.yaml preflight
make start      # Start production services locally
make stop       # Stop all services

Backend directory (for backend development only):

make install            # Install backend dependencies
make dev                # Gateway API, reload (port 8001)
make gateway            # Gateway API only (port 8001)
make test               # four offline workers; TEST_JOBS=1 for serial
make test-live          # live tests (real APIs)
make test-blocking-io   # strict Blockbuster gate on tests/blocking_io/
make test-shard SPLITS=4 GROUP=2  # one duration-aware shard
make test-shard-durations  # refresh baseline
make lint               # ruff lint
make format             # ruff format
make migrate-rev MSG="..."  # Autogenerate a new alembic revision (see Schema Migrations section)

The backend make dev target pre-creates and excludes DEER_FLOW_HOME (default: backend/.deer-flow) and backend/sandbox from Uvicorn's reload watcher. Do not replace it with a bare uvicorn --reload: agent tasks write Python and other runtime files below DEER_FLOW_HOME, which would otherwise restart the Gateway during an active run.

Configuration checks and operator migration scripts reuse the installed backend environment with uv run --no-sync --project backend from the repository root. --project preserves the caller's working directory and relative runtime/config selectors; --no-sync retains installed extras such as PostgreSQL drivers.

More specific AGENTS.md files in backend code directories contain the subsystem sections split from this file. Follow the nearest file in the directory tree.

Architecture

Harness / App Split

The backend is split into two layers with a strict dependency direction:

• Harness (packages/harness/deerflow/): Publishable agent framework package (deerflow-harness). Import prefix: deerflow.*. Contains agent orchestration, tools, sandbox, models, MCP, skills, config — everything needed to build and run agents.
• App (app/): Unpublished application code. Import prefix: app.*. Contains the FastAPI Gateway API and IM channel integrations (Feishu, Slack, Telegram, DingTalk).

Dependency rule: App imports deerflow, but deerflow never imports app. This boundary is enforced by tests/test_harness_boundary.py which runs in CI.

Import conventions:

# Harness internal
from deerflow.agents import make_lead_agent
from deerflow.models import create_chat_model

# App internal
from app.gateway.app import app
from app.channels.service import start_channel_service

# App → Harness (allowed)
from deerflow.config import get_app_config

# Harness → App (FORBIDDEN — enforced by test_harness_boundary.py)
# from app.gateway.routers.uploads import ...  # ← will fail CI

Package import hygiene: the deerflow.agents and deerflow.subagents package roots expose heavyweight graph/executor entrypoints lazily. The deerflow.agents:make_lead_agent LangGraph Server entrypoint is a concrete thin module-level function because the server resolves graph factories directly from the module dictionary; the wrapper keeps the lead-agent and skill-cache imports inside the function so importing the package remains lightweight. Internal modules that only need lightweight types, config, or registries should import the concrete submodule instead of adding eager package-root imports that pull in the tool graph or subagent executor during state/schema imports.

ThreadMetaStore.search() keeps JSON filter semantics identical across memory, SQLite, and PostgreSQL: missing differs from null, bool differs from int, and float filters accept integer or real JSON numbers through json_value_matches.

Gateway Run-Context Trust Boundary

Gate server-owned run context on both client feeds (body.context and body.config): merge_run_context_overrides admits it only for internal=True, while strip_internal_context_keys scrubs both destinations. Treat disable_clarification like non_interactive. Before run/state writes, _normalize_input_messages rejects canonical external system/developer roles; only AUTH_SOURCE_INTERNAL run input may retain them.

Development Workflow

Test-Driven Development (TDD) — MANDATORY

Every new feature or bug fix MUST be accompanied by unit tests. No exceptions.

• Write tests in backend/tests/ following the existing naming convention test_<feature>.py
• Run both offline targets before and after your change: make test and make test-blocking-io
• Tests must pass before a feature is considered complete
• For lightweight config/utility modules, prefer pure unit tests with no external dependencies
• If a module causes circular import issues in tests, add a sys.modules mock in tests/conftest.py (see existing example for deerflow.subagents.executor)

# Four duration-balanced offline shards
make test

# Run strict blocking-I/O tests
make test-blocking-io

# Live API tests: require config.yaml, credentials, and explicit opt-in
make test-live

# Run a specific test file
PYTHONPATH=. uv run pytest tests/test_<feature>.py -v

Keep live tests opt-in via DEER_FLOW_RUN_LIVE_TESTS=1; guard POSIX-only markers with os.name for Windows collection.

Jina logging tests use dummy keys; the Jina redirect regression runs offline via httpx.MockTransport and keeps POST across temporary redirects; InfoQuest connect/read timeout is 30s, separate from crawl timeouts (tests/test_jina_client.py, tests/test_infoquest_http_timeout.py).

Running the Full Application

Run make dev from the repo root to start all services at http://localhost:2026.

All startup modes:

	Local Foreground	Local Daemon	Docker Dev	Docker Prod
Dev	./scripts/serve.sh --dev make dev	./scripts/serve.sh --dev --daemon make dev-daemon	./scripts/docker.sh start make docker-start	—
Prod	./scripts/serve.sh --prod make start	./scripts/serve.sh --prod --daemon make start-daemon	—	./scripts/deploy.sh make up

Action	Local	Docker Dev	Docker Prod
Stop	./scripts/serve.sh --stop make stop	./scripts/docker.sh stop make docker-stop	./scripts/deploy.sh down make down
Restart	./scripts/serve.sh --restart [flags]	./scripts/docker.sh restart	—

Running Backend Services Separately

From the backend directory:

# Gateway API
make gateway

Direct access (without nginx):

• Gateway: http://localhost:8001

Frontend Configuration

The frontend uses environment variables to connect to backend services:

• NEXT_PUBLIC_LANGGRAPH_BASE_URL - Defaults to /api/langgraph (through nginx)
• NEXT_PUBLIC_BACKEND_BASE_URL - Defaults to empty string (through nginx)

When using make dev from root, the frontend automatically connects through nginx.

Key Features

Web Search Recency

Optional web_search time_range=day|week|month|year mappings: DDG → d|w|m|y; Brave → pd|pw|pm|py; Tavily/SearXNG → unchanged; Sofya → unchanged in freshness; Serper → qdr:d|w|m|y in tbs. Omission preserves request shape; Serper null also omits tbs. Image search is unchanged. DDGS 9.14.1 recency uses only enabled Brave/DuckDuckGo/Yahoo engines honoring timelimit: auto/all selects this set; drop incompatible engines and fall back to this set if empty. Re-check on DDGS upgrades.

Tavily Fetch

Title fallback: result URL, then request URL.

File Upload

Outlines use ATX syntax (1–6 hashes, space/tab separator, ≤3 leading spaces), strip closing hashes and skip fenced code; closing markers use a linear suffix scan — never unanchored whitespace regex on unbounded headings.

• Endpoint: POST /api/threads/{thread_id}/uploads; PDF/PPT/Excel/Word via markitdown, one conversion worker per request; directories rejected, keeping uploads all-or-nothing; duplicate filenames in one request get _N suffixes.
• Files live in thread-isolated users/{user_id}/threads/{thread_id}/user-data/uploads (IM channels thread the owner via the user_id= kwarg — see IM Channels → Owner-scoped file storage; HTTP/embedded callers resolve get_effective_user_id()); per-thread upload-companions/ holds source mtime/ctime.
• Gateway HTTP uploads stage .upload-*.part files hidden from upload listings, agent context, and sandbox listings/searches; publication is atomic after size validation. Startup sweep keeps lone .part files under 24h (possibly in flight on another replica); st_nlink > 1 means a published alias, removed at any age so the multi-link check can still replace the file.
• Upload/list/delete handlers offload filesystem work through deerflow.utils.file_io.run_file_io (ContextVar-preserving); non-mounted sandbox uploads use SandboxProvider.acquire_async() and offload read_bytes() + sandbox.update_file() together. Mounted uploads skip acquire/sync; AIO remote/provisioner requires accurate sandbox.thread_data_mounts: true, omission keeps backend auto-detection.
• UploadsMiddleware caps outline titles at 200 characters and previews at 2000 including markers. Titles use original_user_content; attachment-only titles use a sanitized, bounded filename or count.

See docs/FILE_UPLOAD.md for details.

Plan Mode

config.configurable.is_plan_mode=True enables TodoList write_todos for multi-step tasks: one in_progress task, real-time updates. See usage.

Run Interaction Policy

Interaction-sensitive changes must follow policy.

Context Summarization

Automatic conversation summarization when approaching token limits, configured under the summarization key in config.yaml (triggers: tokens, messages, or fraction of max input); recent messages are kept while older ones are summarized.

• Manual compaction uses POST /api/threads/{id}/compact, reuses the same DeerFlowSummarizationMiddleware, writes a new checkpoint with updated messages/summary_text, and bumps only those channel versions. The route uses the shared reserve_checkpoint_write() boundary (also used by manual state updates); its short-lived checkpoint_write operation shares the durable active-thread uniqueness constraint with run admission, preventing worker-local and cross-worker checkpoint-write races.
• Cache only first-candidate no-ops; never suppress primary retries with fallback results. See reuse and telemetry for cache and counter contracts.

See docs/summarization.md for details.

Vision Support

For models with supports_vision: true: ViewImageMiddleware processes images and adds view_image_tool. Images reach the model as base64 in a hidden message carrying a reserved ID prefix plus a server-owned metadata marker (Gateway strips the marker from untrusted input; the middleware requires both to recognize its own message). The injection happens inside wrap_model_call, so payloads never enter graph state — checkpoints keep only lightweight viewed_images metadata, and client-chosen IDs survive; the middleware also sweeps its own message out of every request before rebuilding it, so a payload stranded in an older checkpoint by an interrupted run stops being resent.

Code Style

• Uses ruff for linting and formatting
• Line length: 240 characters
• Python 3.12+ with type hints
• Double quotes, space indentation

Documentation

See docs/ directory for detailed documentation:

• CONFIGURATION.md - Configuration options
• ARCHITECTURE.md - Architecture details
• API.md - API reference
• SETUP.md - Setup guide
• FILE_UPLOAD.md - File upload feature
• PATH_EXAMPLES.md - Path types and usage
• summarization.md - Context summarization
• plan_mode_usage.md - Plan mode with TodoList