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

Stream Bridge Heartbeats

Memory and Redis bridges take their default idle heartbeat cadence from the startup-only stream_bridge.heartbeat_interval_seconds setting. Keep the default on the bridge instance so SSE, /wait, and internal subscribers stay aligned; an explicit subscribe(..., heartbeat_interval=...) remains a per-subscription override.

Checkpoint Channel Modes (full / delta)

Checkpointer storage runs in one of two channel modes, selected by checkpoint_channel_mode in config.yaml (default full). delta mode adopts LangGraph 1.2's DeltaChannel for messages: checkpoints store a sentinel + per-step writes instead of the full message list, so storage/serde grows O(N) instead of O(N²) in turns. All checkpointer backends (memory/sqlite/postgres) serve both modes unchanged — the semantics live in the compiled graph's channel table, not in the saver.

Mode is process-frozen and restart-required. make_lead_agent and the embedded DeerFlowClient freeze the resolved mode (runtime/checkpoint_mode.py::freeze_checkpoint_channel_mode) before compiling the graph with the mode-matched schema (agents/thread_state.py::get_thread_state_schema, plus adapt_state_schema_for_mode / normalize_middleware_state_schemas for middleware state). Adapted middleware schemas are cached by schema, mode, and resolved snapshot frequency so a pre-freeze ephemeral graph cannot leave a stale default-frequency schema behind. A second, different mode or frequency in the same process raises CheckpointModeReconfigurationError. To switch: edit config, restart.

Delta snapshot cadence is configurable but frozen with the mode. database.checkpoint_delta.snapshot_frequency (default 10) sets the DeltaChannel snapshot cadence. It is frozen alongside the mode (freeze_checkpoint_snapshot_frequency; non-positive direct inputs raise ValueError, while a frozen-value mismatch raises CheckpointModeReconfigurationError), restart-required, and must match across every process sharing one checkpoint database — the cadence lives in each compiled graph's channel table and is deliberately NOT stamped into checkpoint metadata, so the mode-compatibility marker and full -> delta migration semantics are unchanged. Schema helpers resolve it explicit-arg -> frozen -> default, and every schema/graph cache (_delta_thread_state_schema, _adapt_state_schema_for_delta, the client agent-config key, the gateway accessor-graph cache) keys on the resolved value.

Compiled-graph cache cap is configurable and hot-reloadable. database.checkpoint_graph_cache.accessor_graph_max (default 64) bounds the gateway accessor-graph cache, which clears wholesale at the cap. The cap is re-read on every eviction check (resolve_checkpoint_graph_cache_max), so a config.yaml reload takes effect without a restart — a size change never affects graph semantics, only eviction timing.

Compatibility is asymmetric and fail-closed. Every checkpoint written in delta mode carries metadata marker deerflow_checkpoint_channel_mode: "delta" (injected via inject_checkpoint_mode; absence of marker = full, so pre-feature checkpoints need no migration). Before any state read/write, ensure_checkpoint_mode_compatible rejects a full-mode process opening a delta thread with CheckpointModeMismatchError (surfaced as HTTP 409 with the cause and thread id by the threads router; CheckpointModeReconfigurationError maps to 503) — a full-mode raw read of a delta blob would silently return empty/partial messages. The reverse direction is allowed: delta-mode processes read full checkpoints transparently (old full checkpoints seed the delta channel), so full → delta is the smooth migration path; delta → full requires materializing/converting the data first. Detection also honors upstream's counters_since_delta_snapshot.messages metadata, and an explicit config marker takes precedence over any ambient context value.

Never bypass CheckpointStateAccessor (runtime/checkpoint_state.py) for thread-state access. It is the single choke point binding graph + checkpointer + mode: it injects the mode marker into configs, runs the compatibility check before every get/update/history, and returns materialized state (delta checkpoints lack channel_values.messages — raw get_tuple reads see a sentinel). Gateway services.py builds and passes the accessor; thread-owned reads (state/history/regeneration) must use build_thread_checkpoint_state_accessor so the recorded assistant's middleware schema materializes every channel. history(limit) semantics: 0 means zero items (explicit empty), None means unlimited — do not pass limit=0 through to graph.get_state_history. Assistant metadata lookup is fail-closed for mutation accessors so a store outage cannot silently select the default schema and discard extension channels. In full mode the read path degrades to a raw checkpointer read (_RawCheckpointReadAccessor) when the agent factory cannot build the graph (bad model config, MCP outage) — full checkpoints carry complete channel_values, so reads don't need the graph; degraded snapshots take created_at from the standard checkpoint ts field, falling back to metadata only for compatibility. The delta gate still applies on the degraded path; next/tasks degrade to empty and thread status falls back to the stored status because task presence is not derivable, while delta mode has no fallback (materialization needs the channel table).

Replay checkpoint lookup prefers lineage and degrades only for an explicitly missing legacy parent link. Branch and regenerate paths first walk parent_config, which prevents a global chronological scan from selecting a sibling created by regeneration. CheckpointParentMissingError alone enables the bounded newest-first history fallback in app/gateway/checkpoint_lineage.py; cycles, dangling/non-addressable parents, target mismatches, and depth exhaustion raise CheckpointLineageIntegrityError and fail closed instead of selecting a sibling. The compatibility scans request 400 raw checkpoints so up to 200 duration-only entries do not consume the effective branch-history budget; the fallback scans oldest-to-newest internally, skips duration-only checkpoints, and accepts only checkpoints with an addressable id as the replay base. A source history with no discoverable pre-user checkpoint preserves the historical single-checkpoint branch behavior instead of rejecting the branch; regeneration remains unavailable for that inherited response. Existing single-checkpoint branches are not mutated by regenerate preparation, and no raw checkpoint tuple is copied across threads because delta state depends on ancestry and pending writes. Regenerate source-run lookup uses the current thread's exact event, then the server-stamped run_id on the copied human message, then verified RunManager content matching; it does not read parent-thread events. When an interrupted response was streamed but never checkpointed, regeneration accepts only the latest visible human message's server-stamped run_id after verifying that it belongs to the same thread and still has interrupted status. Storage or checkpoint-mode failures are not treated as a missing base and still fail closed.

A delta-mode run cannot fork; runtime/runs/worker.py linearizes the resume instead. Resuming from an older checkpoint (regenerate, or any client-supplied checkpoint) forks the lineage, and delta state for a fork is not materializable: BaseCheckpointSaver.get_delta_channel_history — and the bespoke overrides in InMemorySaver/PostgresSaver — collect every pending_writes entry stored on each on-path ancestor, but a shared parent also carries the writes of the sibling child that was abandoned. Those writes replay into the fork, so the run starts from a message list still containing the answer it was supposed to replace (#4458: regenerating in a branched thread showed the superseded assistant message beside the new one after a reload; reproduced on postgres, sqlite, and the in-memory saver). Write-to-child ownership belongs to the upstream delta contract, so DeerFlow does not reimplement the walk: _linearize_delta_checkpoint_resume materializes the requested checkpoint's complete state and writes every channel onto the current head (which has no siblings) through the state mutation graph, using Overwrite for reducer channels and resetting newer head-only channels to their schema default (or None when no constructible default exists); it then drops the checkpoint_id selector and lets the run proceed linearly, while the abandoned turn stays in history as the rewritten head's ancestry. The worker holds _checkpoint_thread_lock across _capture_rollback_point and the optional linear rewrite, making the rollback snapshot and rewrite atomic with graph streaming and the preceding run's duration-metadata checkpoint write. Capture preserves the complete real pre-run state; cancel-with-rollback then linearly replaces the current delta head with that captured state rather than forking the now-shared pre-run checkpoint, so the abandoned turn is restored without replaying the resume sibling's writes. The worker also recomputes the current-run message boundary from the rewritten state and fails closed (an unreadable resume checkpoint raises rather than falling back to the corrupt fork). full mode keeps forking — its checkpoints carry complete channel_values and need no replay — so LangGraph branching semantics are unchanged there. Root namespace only; subgraph namespaces are left alone.

Wholesale state replacement uses a state-only mutation graph + Overwrite. update_state values pass through channel reducers (add_messages merge in full, append in delta), so replacing reducer values requires Overwrite rather than an ordinary update. Full-mode rollback and context compaction replace messages; delta resume and delta rollback replace every materialized channel and reset current-head-only channels to their schema default (or None). These writes go through build_state_mutation_graph(as_node, mode, state_schema), and state_schema MUST be the thread's effective schema (graph_state_schema(assistant_graph)), because the base-ThreadState fallback silently discards written channels contributed by custom AgentMiddleware.state_schema. Channels absent from a full-mode fork write inherit the parent's channel blobs, so middleware channels survive rollback/compaction (locked by test_rollback_preserves_middleware_contributed_channels and test_compact_thread_context_preserves_middleware_contributed_channels). The compiled mutation graph has one no-op node (entry = finish) whose checkpoint machinery (channels/versions/metadata) is identical to the agent graph's but schedules no pending tasks, so the restored/compacted head stays idle instead of re-triggering the agent. Never hand-write checkpoints via checkpointer.aput for this; raw writers elsewhere must preserve checkpoint parentage — severed ancestry breaks delta replay (see runtime/runs/worker.py writer parenting and checkpoint_patches.py).

Run rollback flow (runtime/runs/worker.py): _capture_rollback_point materializes the complete pre-run state via the accessor and captures raw pending_writes via aget_tuple into an immutable RollbackPoint before the run starts — capture failure disables rollback (fail-closed), never restores partial state. In full mode, cancel-with-rollback forks from the pre-run checkpoint via the mutation graph and inherits non-message channels from that parent. In delta mode, forking is unsafe once the cancelled path has attached sibling writes to the pre-run checkpoint, so rollback replaces every captured channel on the current head, using Overwrite for reducers and schema defaults for current-head-only channels. Both modes reattach only the captured pre-run pending writes to the restored checkpoint. Edit replay runs (metadata.replay_kind="edit") also restore the pre-run checkpoint on failed, timed-out, or interrupted completion and publish the restored values snapshot to the stream before end, so clients do not remain on a transient edited branch when the replay did not produce a successful replacement.

Message feed seq stamping (#4666): a checkpoint carries no position of its own and loses messages to summarization, so a client merging a values frame with the seq-ordered run_events feed cannot place a checkpoint-kept message once the feed's loaded page window no longer reaches back to it. RunEventStore.get_message_seqs(thread_id, identities) resolves the seq the store already assigned, keyed by runtime/events/message_identity.py::message_identity — the backend half of the identity rule frontend/src/core/threads/hooks.ts::messageIdentity applies (tool messages by tool_call_id; X / X__user human copies collapse to one). The two halves must stay in sync: a mismatch is silent, degrading placement rather than raising. runs/worker.py::_MessageSeqStamper attaches the result as additional_kwargs.deerflow_seq on root values frames only — subgraph frames are not part of the thread feed's ordering, and nothing is written back to the checkpoint. The run-scoped cache makes the compaction frame the only one that costs a lookup, and the stamper soft-resolves the user id once at build time — like the worker's write paths beside it — so a launch path that never inherits the auth contextvar (a null-owner scheduled task) still stamps instead of the db store's strict AUTO default raising per frame. A resolved seq is cached for the run (earliest-seq-wins makes it final), but a miss is not: a message this run produces reaches a frame before RunJournal flushes it, so it misses and is persisted moments later. Misses are re-asked when RunJournal.feed_generation — bumped once per successful event-store write, never while the buffer merely fills — shows the feed gained rows, which keeps the retry bounded by writes rather than by frames and makes a failed lookup cost one generation instead of the run. REST reads (GET /threads/{id}/state, POST /threads/{id}/history) stamp through events/message_seq.py::stamp_messages_with_seq, the request-scoped counterpart: everything a checkpoint still holds is already persisted, so one batched lookup resolves the whole list. The db store prefilters candidate rows in SQL (a LIKE clause per wanted raw id, wildcards escaped; an id json.dumps would escape falls the set back to the full scan) so a wanted identity absent from the feed — a message still streaming — does not force a full fetch-and-decode of every message row's tool outputs on long threads. deerflow_seq is server-owned display metadata: the gateway strips it from client input, because a welded-in seq goes stale when a fork re-seeds the feed (#4380).

Run delivery receipts (runtime/journal.py + runs/worker.py): RunJournal records each non-empty artifact update once per tool Command for the terminal run.delivery event. When a command contains multiple messages, a unique tool name resolved from matching ToolMessage entries supplies attribution; additional command messages do not duplicate artifact paths or counts. If multiple different tool names resolve for one flat artifact update, the paths remain counted but unattributed because the command does not carry a per-path mapping. RunJournal callbacks set run_inline=True: they do only in-memory bookkeeping or schedule async writes, and staying on the run's event-loop thread serializes parallel tool callbacks before terminal delivery recording and flushing. Each worker creates a separate journal per run before cancellable/fallible preflight work, so checkpoint compatibility failures and cancellation while waiting for prior finalization still emit a zero-delivery receipt. The worker flushes ordinary journal events, idempotently persists the run-scoped receipt, and only then persists the staged terminal run status. A receipt failure is retried on a short bounded schedule while the owning worker still knows the real outcome and holds the lease. Delivery candidates are every regular file created or modified under /mnt/user-data/outputs; internal process-feedback files are excluded (the scanner's EXCLUDED_DIR_NAMES plus the configured tool_output.storage_subdir), so a run that only externalized oversized tool outputs does not fail delivery. At least one candidate must be covered by a path attributed by the journal to present_files; presenting only an unrelated pre-existing path does not satisfy delivery. Receipts for such runs add produced_paths, presented_paths, matched_paths, verification, stage, and satisfied to the Slice 1 fact fields. Missing a matching presentation becomes a run error; a successful presentation is also downgraded to error if its receipt cannot be durably verified. Runs without changed outputs preserve ordinary chat behavior and the original receipt shape. Orphan recovery first atomically claims an expired lease, then uses the same singleton write to backfill a zero-delivery receipt — a stale recovery scan cannot overwrite a live run's later detailed receipt, an event-store outage does not undo the terminal takeover, and an existing detailed receipt is preserved when a worker crashed after writing it. Event stores serialize put_if_absent with ordinary thread writers: memory and JSONL provide the documented single-process guarantee, while the DB store adds per-thread in-process locks and PostgreSQL advisory locks for cross-process writers. Moving journal construction ahead of preflight is receipt-only on early failure paths: a separate boundary flag preserves the previous completion-data semantics, so checkpoint incompatibility or cancellation while waiting for an older finalizing run does not persist an empty completion snapshot. Worker tests pin one accumulated receipt across multiple goal-continuation _stream_once calls; journal tests drive LangChain's real async callback dispatcher against a single journal to pin serialized, deduplicated parallel tool callbacks.

Targeted run-event attribution (runtime/events/store/): RunEventStore.find_latest_ai_message_run_ids() has a complete-or-error contract. Its default implementation walks list_messages() backward in 1000-row pages, preserves the first page's high-watermark through the exclusive before_seq cursor, and raises when a full page has no safe progressing seq. Memory and database stores use that bounded path; the JSONL store overrides it with one complete thread-log read because each JSONL page would otherwise rescan every run file. The default and JSONL paths share the public normalize_message_ids() and match_ai_message_run_id() helpers from events/store/base.py. Database owner filtering is inherited on every page. Callers may use a missing key as proof that no valid AI event exists only after an ordinary return, never after an exception. A caller that crosses a run or checkpoint-write admission boundary must repeat the complete audit after admission; a pre-admission exact hit can be superseded by a later event just as a pre-admission miss can become an exact hit.

Gateway POST /api/threads/{id}/history uses that lookup to migrate legacy AI messages. An exhaustive miss preserves the human-boundary fallback; an incomplete lookup removes unproven synthesized IDs. Its metadata-only write-on-read cache stores run_message_ids for every audited AI ID (including exhaustive misses) plus required run_durations; duration presence alone does not prove attribution. Historical body.before reads write the audit to the head, and the merge may retain IDs no longer in materialized history, which readers ignore. Migration must acquire the durable checkpoint_write reservation, then repeat the whole message audit and batch-reload required run rows before persisting. Post-admission exact hits replace foreground exact or boundary mappings, and recomputed final durations replace foreground snapshots. Successful workers keep their durable run row active through the final duration checkpoint write, so a peer migration cannot enter during terminalization. The first RunManager.list_by_thread() hydration page uses a 100-row floor or the number of required IDs, whichever is larger; missing exact runs use targeted get() calls.

Terminal run cleanup explicitly breaks graph-scoped references while preserving the existing RunRecord grace period. Every agent.astream() iterator is closed in _stream_once, including abort/exception/early-break paths. A close failure after an abort is warning-only and cannot replace the user-requested interrupted outcome; normal-completion close failures still surface, and an in-flight stream exception remains authoritative over a secondary close failure. Journal construction and cancellable preflight work (including MCP task projection and the prior-finalization wait) live inside the worker's guarded body, so cancellation before agent startup still terminalizes the run and closes its stream. run_agent() wraps the complete terminal-finalization sequence in an outer teardown guard, so cancellation or failure from any terminal-stage await cannot skip RunJournal.close(), removal of the journal, __pregel_runtime, and internal runtime-context values from every runnable config, or release of local graph/payload references. That guard schedules bridge cleanup, run-record cleanup, and cyclic GC even when interruption happens before the terminal stream marker or terminal publication itself fails, so neither a cancelled observer nor a delivery-backend outage can strand process-local run state. RunJournal.flush() clears its _pending_progress_task after awaiting or cancelling it; ordinary close() detaches the event store/progress reporter and clears callback bookkeeping only after that flush succeeds, preserving the buffer for retry on a transient store failure. A fenced worker instead calls close(flush=False), which cancels pending journal work and detaches without initiating another event-store write after lease ownership is lost; its final detach runs even if a second cancellation interrupts pending-task shutdown. RunManager.cleanup(run_id) retains the process-local RunRecord, completed task, and request payload for its default 300-second local join/status window before releasing them. Durable history remains in RunStore; StreamBridge data keeps its separate 60-second late-subscriber window, and both cleanup coroutines run in a fresh empty contextvars.Context. A contextless full cyclic-GC pass, coalesced to at most once every 10 seconds and dispatched through the default executor, bounds the lifetime of unreachable LangGraph callback/loop cycles without synchronously walking the heap in the event-loop timer; passes taking at least 100 ms are logged at INFO because CPython GC may still impose interpreter-level pauses.

Where things live:

• runtime/checkpoint_mode.py — mode + snapshot-frequency freeze, marker injection, delta detection, compatibility gate, both error types
• runtime/checkpoint_state.py — CheckpointStateAccessor, build_state_mutation_graph, RollbackPoint
• checkpoint_patches.py (package root) — checkpoint-machinery patches: delta-history folding for InMemorySaver (delegating to the base walk), stable message IDs across materialization, upstream first-write drop fix, and BinaryOperatorAggregate unwrapping an Overwrite first write into an empty (MISSING) channel — Union-typed reducer channels (sandbox/goal/todos/promoted) have no constructible default, so a replace-style write into a fresh branch thread or a never-written channel stored the wrapper literally and crashed the next consumer (#4380; probe-guarded, stands down if upstream fixes it)
• agents/thread_state.py — ThreadState/DeltaThreadState, delta_messages_field / DELTA_MESSAGES_FIELD (DeltaChannel at the configured snapshot_frequency, default 10), schema adaptation helpers
• runtime/context_compaction.py — compaction via accessor + mutation graph (reference consumer)
• runtime/checkpoint_cache/ + runtime/checkpointer/cached_saver.py — delta-mode checkpoint history cache; checkpoint state reads MUST go through CheckpointStateAccessor, and the checkpointer may be a CachedHistorySaver wrapper — never rely on concrete saver types
• Tests: tests/test_checkpoint_mode.py (freeze/detect/gate), tests/test_checkpoint_state.py (accessor/mutation graph), tests/test_delta_channel_checkpointers.py (saver parity), tests/test_threads_checkpoint_mode.py, tests/test_gateway_checkpoint_mode.py (dual-mode e2e parity), tests/test_context_compaction.py (mutation-graph write, no scheduling), tests/test_run_worker_rollback.py, tests/test_cached_history_saver.py + tests/test_cached_history_saver_integration.py (history cache)

Checkpoint channel benchmark: scripts/benchmark/checkpoint/bench_channels.py runs paired full/delta message-only StateGraphs in a fresh child process per case, using sync InMemorySaver or SqliteSaver so reducer, serialization, and saver costs stay separate from Gateway/async scheduling. It reports deterministic correctness digests, write windows/percentiles, warm and graph-rebuilt cold reads, logical checkpoint/write bytes, SQLite DB/WAL/SHM footprint, reducer replay time, and peak RSS as versioned JSONL. The controller alternates mode order and rejects performance data when paired modes materialize different state. Its default 1 GiB estimated cumulative full-payload cap skips both modes of an oversized pair when full is selected, including every delta cadence in a --snapshot-frequencies sweep; intentional --modes delta diagnostics bypass this full-payload cap, so size those runs explicitly. Use --allow-large-cases only on a provisioned machine. Duplicate CSV matrix values are ignored with a warning; use --repetitions for repeated samples. Summarize paired successful repetitions with scripts/benchmark/checkpoint/summarize_channels.py (all ratios are delta/full). --profile-dir /tmp/checkpoint-profiles writes one cProfile artifact per case for attribution. Profiled rows carry profiled: true, and the summarizer automatically excludes them from baseline summaries with a warning. Storage-size collection relies on saver-specific diagnostic layouts; if those layouts change, the timing/correctness row remains successful while storage fields become null and storage_stats_error records the diagnostic failure. Example:

cd backend
PYTHONPATH=. uv run python scripts/benchmark/checkpoint/bench_channels.py \
  --backends sqlite --updates 100,500,999,1000,1001 --payload-bytes 128 \
  --repetitions 7 --output /tmp/checkpoint-bench.jsonl
PYTHONPATH=. uv run python scripts/benchmark/checkpoint/summarize_channels.py \
  /tmp/checkpoint-bench.jsonl

The production-shaped layer lives in scripts/benchmark/checkpoint/bench_production.py: per-case child processes run graph-level ainvoke turns through the real lead-agent graph (scripted deterministic model, real AsyncSqliteSaver), then measure GET /threads/{id}/state and POST /threads/{id}/history through the real Gateway route stack in the same event loop (httpx ASGITransport), split into cold/warm accessor-graph-cache samples. It sweeps snapshot_frequency (config: checkpoint_delta.snapshot_frequency, process-frozen like the mode), pairs every delta frequency against the same full row, and fails both rows of a pair when materialized or wire digests diverge. Each case must have more than the two discarded warm-up turns, and SQLite DB/WAL/SHM sizes are captured while the saver is still open so they represent the online storage footprint. Summarize with scripts/benchmark/checkpoint/summarize_production.py (ratios are delta/full; it also emits snapshot_write_spike and cache_effect_ms, the decision inputs for the production snapshot-frequency and accessor-cache defaults). Harness tests live in tests/test_bench_checkpoint_production.py and tests/test_summarize_checkpoint_production.py; timing thresholds are not CI gates. The matrix test pins that every (repetition, turns) group contains both modes and that their execution order flips between consecutive groups, including across repetition boundaries.

Operational limits learned from the first runs (the default matrix is too large to run blindly):

• The default --timeout-seconds 900 is insufficient for delta mode at snapshot_frequency=1000 once turns reach 500 (measured: delta-500 takes ~1100-1200s; delta-2000 takes ~45min). Pass an explicit --timeout-seconds for any large matrix, and treat the turns=2000 corner as practical only at small snapshot frequencies.
• Full-mode 2000-turn runs produce a ~33GB sqlite DB. Point TMPDIR at real disk, not tmpfs (the benchmark uses tempfile.TemporaryDirectory, which honors TMPDIR), or the run dies mid-case.
• The history route clamps limit to 100 (le=100 on ThreadHistoryRequest.limit), so --history-limits values above 100 are measured and reported by their effective (clamped) limit.

Example:

cd backend
PYTHONPATH=. uv run python scripts/benchmark/checkpoint/bench_production.py \
  --turns 10,100,500,1000,2000 --payload-bytes 128 \
  --snapshot-frequencies 10,50,100,500,1000 \
  --repetitions 7 --output /tmp/production-bench.jsonl
PYTHONPATH=. uv run python scripts/benchmark/checkpoint/summarize_production.py \
  /tmp/production-bench.jsonl