mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-04-25 19:28:23 +00:00
229c8095be
``get_thread_history`` and ``get_thread_state`` in Gateway mode read
messages from ``checkpoint.channel_values["messages"]``. After
SummarizationMiddleware runs mid-run, that list is rewritten in-place:
pre-summarize messages are dropped and a synthetic summary-as-human
message takes position 0. The frontend then renders a chat history that
starts with ``"Here is a summary of the conversation to date:..."``
instead of the user's original query, and all earlier turns are gone.
The event store (``RunEventStore``) is append-only and never rewritten,
so it retains the full transcript. This commit adds a helper
``_get_event_store_messages`` that loads the event store's message
stream and overrides ``values["messages"]`` in both endpoints; the
checkpoint fallback kicks in only when the event store is unavailable.
Behavior contract of the helper:
- **Full pagination.** ``list_messages`` returns the newest ``limit``
records when no cursor is given, so a fixed limit silently drops
older messages on long threads. The helper sizes the read from
``count_messages()`` and pages forward with ``after_seq`` cursors.
- **Copy-on-read.** Each content dict is copied before ``id`` is
patched so the live store object (``MemoryRunEventStore`` returns
references) is never mutated.
- **Stable ids.** Messages with ``id=None`` (human + tool_result,
which don't receive an id until checkpoint persistence) get a
deterministic ``uuid5(NAMESPACE_URL, f"{thread_id}:{seq}")`` so
React keys stay stable across requests. AI messages keep their
LLM-assigned ``lc_run--*`` ids.
- **Legacy ``Command`` repr sanitization.** Rows captured before the
``journal.py`` ``on_tool_end`` fix (previous commit) stored
``str(Command(update={'messages': [ToolMessage(content='X', ...)]}))``
as the tool_result content. ``_sanitize_legacy_command_repr``
regex-extracts the inner text so old threads render cleanly.
- **Inline feedback.** When loading the stream, the helper also pulls
``feedback_repo.list_by_thread_grouped`` and attaches ``run_id`` to
every message plus ``feedback`` to the final ``ai_message`` of each
run. This removes the frontend's need to fetch a second endpoint
and positional-index-map its way back to the right run. When the
feedback subsystem is unavailable, the ``feedback`` field is left
absent entirely so the frontend hides the button rather than
rendering it over a broken write path.
- **User context.** ``DbRunEventStore`` is user-scoped by default via
``resolve_user_id(AUTO)``. The helper relies on the ``@require_permission``
decorator having populated the user contextvar on both callers; the
docstring documents this dependency explicitly so nobody wires it
into a CLI or migration script without passing ``user_id=None``.
Real data verification against thread
``6d30913e-dcd4-41c8-8941-f66c716cf359``: checkpoint showed 12 messages
(summarize-corrupted), event store had 16. The original human message
``"最新伊美局势"`` was preserved as seq=1 in the event store and
correctly restored to position 0 in the helper output. Helper output
for AI messages was byte-identical to checkpoint for every overlapping
message; only tool_result ids differed (patched to uuid5) and the
legacy Command repr at seq=48 was sanitized.
Tests:
- ``test_thread_state_event_store.py`` — 18 tests covering
``_sanitize_legacy_command_repr`` (passthrough, single/double-quote
extraction, unparseable fallback), helper happy path (all message
types, stable uuid5, store non-mutation), multi-page pagination,
summarize regression (recovers pre-summarize messages), feedback
attachment (per-run, multi-run threads, repo failure graceful),
and dependency failure fallback to ``None``.
Docs:
- ``docs/superpowers/plans/2026-04-10-event-store-history.md`` — the
implementation plan this commit realizes, with Task 1 revised after
the evaluation findings (pagination, copy-on-read, Command wrap
already landed in journal.py, frontend feedback pagination in the
follow-up commit, Standard-mode follow-up noted).
- ``docs/superpowers/specs/2026-04-11-runjournal-history-evaluation.md``
— the Claude + second-opinion evaluation document that drove the
plan revisions (pagination bug, dict-mutation bug, feedback hidden
bug, Command bug).
- ``docs/superpowers/specs/2026-04-11-summarize-marker-design.md`` —
design for a follow-up PR that visually marks summarize events in
history, based on a verified ``adispatch_custom_event`` experiment
(``trace=False`` middleware nodes can still forward the Pregel task
config via explicit signature injection).
Scope: Gateway mode only (``make dev-pro``). Standard mode
(``make dev``) hits LangGraph Server directly and bypasses these
endpoints; the summarize symptom is still present there and is
tracked as a separate follow-up in the plan.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>