deer-flow/backend/packages/harness/deerflow/AGENTS.md

Request Trace Context (packages/harness/deerflow/trace_context.py)

DeerFlow's request-level correlation id — the X-Trace-Id header and the deerflow_trace_id key. Not Langfuse's trace id, not run_id, not the short subagent trace_id log label.

The ContextVar is the only source. Every path that reaches a run binds one first; downstream treats the id as a plain str, no if trace_id: guards.

Entry points and binders: Gateway HTTP — TraceMiddleware; scheduled occurrence — ScheduledTaskService._attempt_queued_run → launch_scheduled_thread_run; MCP task notification — launch_mcp_task_notification_run; IM inbound — ChannelManager._worker_loop; embedded / TUI / CLI turn — DeerFlowClient.stream().

Only the first is HTTP; the rest run outside ASGI, so the binding cannot live in middleware alone. Each scopes one unit of work, never a poller loop — a leaked binding on a reused worker task would tag later occurrences with the first id. ensure_trace_context inherits, keeping layered scheduled bindings and a manual trigger inside a Gateway request on one trace.

Every other carrier is a derived output, never read back as an input. worker._bind_trace_id stamps the runtime context and config["metadata"]; services.start_run stamps the run record; a caller-sent deerflow_trace_id (body.metadata, body.config.context) is replaced — honouring it would let the persisted run disagree with the header and the logs. _SERVER_OWNED_RUNTIME_CONTEXT_KEYS covers the embedded path and also rejects caller-supplied sandbox lease/scope identities, redact_config_secrets scrubs the kwargs echo (runs.kwargs_json), and build_run_config merges metadata onto a copy so the stamp cannot reach body.config. Callers pin an id with X-Trace-Id.

Accepted divergence: a crash-recovered scheduled launch reuses its run via the idempotency key without restamping — the record keeps the first attempt's id, the retry's logs a fresh one; restamping would rewrite an existing record. Not a bug. Thread metadata omits the key entirely — a thread spans many runs.

Do not open-code fallback chains. Two helpers own the resolution order:

• resolve_trace_id(*carriers) — first usable carrier, else ambient. For ids travelling as data in runtime.context; ContextVars do not survive a bare thread hop.
• ensure_trace_context(trace_id) — reuse the surrounding scope, else start a self-contained one. For boundary crossings (SubagentExecutor._aexecute, the memory trace_context_manager hook) and non-HTTP entry points; no argument mints a scoped id.

request_trace_context (HTTP) deliberately does not inherit: a crafted header must not fall back to the previous request's id.

get_current_trace_id() stays nullable only for the logging filter (pre-entry-point records render as trace_id=-); everything else uses ensure_trace_id()/resolve_trace_id().

DeerFlowClient.stream() binds per next() step and around inner.close(), never across a yield: a sync generator shares the caller's context, so a scope held across yields would leak the id and break on cross-context GC finalization.

logging.enhance.enabled gates log output only (trace_id field presence and format) — not the id, the header, or the run metadata — so TraceMiddleware reads no AppConfig; logging stays restart-required (STARTUP_ONLY_FIELDS["logging"]). X-Trace-Id is in CORS_EXPOSED_HEADERS (not safelisted). Unhandled-exception 500s keep the header — TraceMiddleware sends its own plain 500 (CORS-opaque, see its docstring) before re-raising; mid-stream failures propagate unchanged.

Tests: the tests/test_trace_* and tests/test_worker_trace_binding.py suites, test_gateway_services.py, test_run_metadata_secret_safety.py, plus the Langfuse suites in tracing/AGENTS.md.

Managed Lark CLI credentials (integrations/lark_cli.py)

Installed lark-shared guidance points to Capability Center > Plugins > Lark (/workspace/capabilities?tab=plugins&plugin=lark). Guidance changes bump the version marker; reinstalling the managed skill pack refreshes the stored text.

App registration and direct app switching replace the per-user Lark credential tree transactionally. Clear the old OAuth data before running lark-cli config init: on Linux that command writes the new app secret into the file-backed keychain under the data directory, so clearing the directory afterward would leave config.json with a dangling keychain reference. The transaction snapshot still supplies the previous OAuth data for logout and restores the complete old tree if any switch step fails.

Browser Progress Screenshots (community/browser_automation/)

Hidden per-action browser progress frames use JPEG at quality 80 to keep their storage and transfer cost bounded relative to lossless PNG. The explicit browser_screenshot tool remains PNG because it creates a user-requested artifact. New automatic capture entry points must reuse the shared progress encoding definition in tools.py so the byte encoding and .jpg suffix cannot drift.

Embedded Client (packages/harness/deerflow/client.py)

DeerFlowClient provides in-process access without HTTP/FastAPI, sharing Gateway's deerflow modules, config, data directories, and response schemas.

Agent Conversation:

• chat(message, thread_id) — synchronous, accumulates streaming deltas per message-id and returns the final AI text
• stream(message, thread_id) — subscribes to LangGraph stream_mode=["values", "messages", "custom"] and yields StreamEvent:
  • "values" — state snapshot (title, messages, artifacts, summary_text). Always forward summary_text (current summary or None), including unchanged values/resets. Never re-emit AI text delivered via messages; serialized ToolMessage entries retain non-None native artifact
  • "messages-tuple" — AI text deltas (concatenate per id); emit tool calls/results once each, preserving non-None native result artifact
  • "custom" — forwarded from StreamWriter; DeerFlow-built-in custom events are dual-emitted through deerflow.utils.custom_events, so astream_events(version="v2") consumers also receive one on_custom_event with name=payload["type"] and the unchanged payload as data
  • "end" — stream finished (carries cumulative usage counted once per message id)
• Custom-event invariant — use emit_custom_event / aemit_custom_event, never StreamWriter alone. Built-in payloads require a non-empty string type; typeless payloads stay writer-only, absent from astream_events. The writer runs first and is authoritative for Gateway/Web UI/embedded clients; best-effort callbacks must not break it. Async graph hooks must await the async helper, never dispatch synchronously on a running event loop.
• Lazy graph creation uses create_agent() + build_middlewares().
• Cache graphs by storage user_id and the unordered set of named-agent mcp_plugins. stream() materializes user_id before worker/loop boundaries in every auth mode.
• Supports checkpointer parameter for state persistence across turns
• reset_agent() reloads AgentConfig and rebuilds the graph. Every run's metadata carries mcp_plugins for delegation, including cache hits.
• Streaming design: Gateway/client parallel paths, LangGraph stream_mode, per-id deduplication, and regression tests

Gateway Equivalent Methods (replaces Gateway API):

Category	Methods	Return format
Models	list_models(), get_model(name)	{"models": [...]}, {name, display_name, ...}
MCP	get_mcp_config(), update_mcp_config(servers)	{"mcp_servers": {...}}
Skills	list_skills(), get_skill(name), update_skill(name, enabled), install_skill(path)	{"skills": [...]}
Goals	get_goal(thread_id), set_goal(thread_id, objective, max_continuations=8), clear_goal(thread_id)	{"goal": {...}} or {"goal": None}
Memory	get_memory(), reload_memory(), get_memory_config(), get_memory_status()	dict
Uploads	upload_files(thread_id, files), list_uploads(thread_id), delete_upload(thread_id, filename)	{"success": true, "files": [...]}, {"files": [...], "count": N}
Artifacts	get_artifact(thread_id, path) → (bytes, mime_type)	tuple

Gateway differences: Upload takes local Path, not UploadFile, rejects directories before copying, and reuses one conversion worker inside an active event loop. Artifacts return (bytes, mime_type), not HTTP Response. Gateway alone deletes .deer-flow/threads/{thread_id} after LangGraph thread deletion; the client has no equivalent. update_mcp_config() and update_skill() invalidate the cached agent.

Tests: tests/test_client.py is offline, including TestGatewayConformance. tests/test_client_live.py requires root config.yaml, valid API credentials, and opt-in via make test-live or DEER_FLOW_RUN_LIVE_TESTS=1. It calls real APIs (possible costs) and may create local sandboxes, artifacts, and files. Marked live, it is excluded from make test and skipped in default CI.

Gateway Conformance Tests (TestGatewayConformance): Parse every dict-returning client method's output through its Gateway Pydantic model so missing required fields raise ValidationError in CI. Covers: ModelsListResponse, ModelResponse, SkillsListResponse, SkillResponse, SkillInstallResponse, McpConfigResponse, UploadResponse, MemoryConfigResponse, MemoryStatusResponse.

AIO Sandbox Network Policy

Restricted AIO keeps sandboxes internal; a per-sandbox, ICC-disabled sidecar handles egress and its token-authenticated API relay. Parse headers strictly; reject policy-denied names before DNS and try all validated answers. Claim the oldest unsurfaced denial; subagent/non-interactive runs drain and deny. Approvals never replay tools; policy labels fence reuse. CONNECT/SNI cannot inspect encrypted authority. Discovery and enumeration are read-only, including on a policy or network-mode mismatch; only the provider may replace it after the orphan grace, local teardown reservation, and cross-instance teardown lease. Destroy the sandbox, sidecar, and both networks together.

E2B Mount Uploads

E2B uploads host mounts during sandbox creation using binary file objects. Per-mount limits: 100 MiB/file, 512 MiB total, 2,000 files. The full creation pass shares a 512 MiB / 2,000-file budget across skill projections and mounts.

The pass has a cooperative deadline controlled by mount_upload_deadline_seconds (default: 120 seconds). The provider checks it before each mount, during directory preflight, and before each SDK write. The deadline does not interrupt active filesystem or E2B SDK calls.

The provider checks mount limits before upload. It rechecks each opened file descriptor against its preflight size before SDK upload.

For policy-scoped turns, clearing the four managed remote skill categories and uploading their prepared projection is one per-user/thread/skills-root critical section, shared with acquire and release. The provider snapshots that canonical root at startup and carries it through warm-pool identity and E2B metadata; a VM from another root is never adopted. A second policy sync cannot reset the remote tree until the first upload pass has completed.

An invalid mount does not block later mounts.

Each successful upload logs its source, destination, file count, byte count, and elapsed time.

A stopped pass logs its limit reason and elapsed time. It reports attempted and completed upload totals separately.

After creation, E2BSandbox.mount_upload_result holds a MountUploadResult. result.truncated is true only for resource-limit stops (deadline, file count, bytes), not logged mount failures (missing paths, SDK errors). A provider-level map preserves creation results within the Gateway process; None on a reclaimed sandbox means unavailable.

Workspace Snapshot Cancellation (workspace_changes/recorder.py)

After _prepare_capture() hands off roots, cancellation must drain text scans (include_text=True) before removing the cache the worker may still access. Metadata scans (include_text=False) own no cache: cancel promptly, let the worker continue, and consume/log its outcome in a completion callback. Prepare-stage cancellation retains its handoff/reclaim path. Regressions in tests/blocking_io/test_workspace_changes_cancellation.py must cover prompt metadata cancellation and text-cache drain/cleanup.