- Validate stage prefixes before forwarding model tokens. Invalid prefixes
may use the existing repair budget; once output starts, validation failure
aborts the turn without appending a regenerated reply. Bound the token
queue and close inner streams when consumption stops.
- Replace delete-time lock eviction with reference-counted session locks
covering holders and waiters. Serialize set/get with chat/stream/delete
and release cancelled waiters without replacing a live lock.
- Store person synonyms under workflow-native driver keys and expose
consistent alias views to single/group reads and graph prompt injection.
- Separate committed field patches from public formUpdate visibility.
Commit callbacks keep repository fields synchronized even when updates
are hidden or the consumer closes after graph completion.
- Disable OpenAI SDK automatic retries. Own bounded retry attempts in the
application, use 1/2/4 second backoff, retry streams only before text is
emitted, enforce attempt timeouts, and close upstream/client resources.
- Translate FastGPT SDK max_retries (total attempts) from the configured
extra-retry count, and verify the real SDK loop uses four attempts.
- Document the per-LLM budget and streaming failure semantics without
claiming the whole multi-call HTTP turn has the same timeout.
Validation: 267 offline tests passed, including actual SDK requests through
MockTransport, stream repair/abort, timeout/close, lock races, aliases and
formUpdate visibility; existing four dependency deprecation warnings.
Replaces the 'complete answer sliced into chunks' stopgap with token-
level streaming. Graph nodes are untouched: they still consume the
batched LLMService.complete interface and validate full replies.
- agent/llm: OpenAILLMService.stream_tokens() iterates the provider
stream with the same message construction, temperature and extra_body
forwarding as complete(); identical error classification (401 ->
LLMAuthError, 429 -> LLMRateLimitError, timeout via asyncio.timeout
around the whole iteration); deliberately no retries (tokens already
emitted cannot be recalled)
- backends/langgraph: _TokenPassThroughLLM wraps the injected LLM for
stream turns - dialogue calls push tokens to a queue bridge while
accumulating the full text for node validation; calls marked with the
【字段提取】 marker take the silent inner.complete path so extraction
JSON never leaks into the client stream; LLMs without stream_tokens
fall back to single-chunk push (fake/test compatibility)
- stream(): graph runs in a background task; the generator yields
TextDelta per token and FormUpdate after the turn; graph-task errors
re-raise with the existing ChatBackend error mapping; generator close
(client disconnect) cancels the graph task and the in-flight LLM call
- fixed-answer turns with no LLM call (welcome) fall back to emitting
the full response text, preserving previous behavior
- endpoint untouched: its buffering parser already handles arbitrary
chunk boundaries (state tags split across tokens, stage-first order)
Verified: 241 unit tests (11 new streaming tests: token incrementality,
extraction bypass, FormUpdate ordering, error mapping on both paths,
disconnect cancellation, fallback chunking, SSE protocol end-to-end
with per-token text_delta events)
FastGPT maintains full history server-side; LangGraph truncated the
injected history at a hardcoded 50 messages, so long conversations
behaved differently across backends.
- Config.AGENT_HISTORY_MESSAGES (env): max messages injected into the
model prompt (recent N, both roles counted); default 50 unchanged;
larger values support long conversations; 0 = no truncation, matching
FastGPT's full-history semantics. Only the injected window is
affected - dialogue stays fully persisted in the checkpoint.
- validate(): rejects negative values; 0 and positive accepted
- conversation_history reads the config (import-time class attribute,
monkeypatched in tests); explicit None-check keeps 0 falsy-safe
- .env.example documents the knob
Verified: 225 unit tests (6 new history-window tests)
Closes the error-code gap between LangGraph and FastGPT backends: the
endpoint's 401/429 branches previously could never trigger under the
LangGraph backend because all provider failures collapsed into LLMAPIError
(500).
- agent/llm: LLMAuthError / LLMRateLimitError as LLMAPIError subclasses;
OpenAILLMService raises them from openai AuthenticationError (401, no
retry) and RateLimitError (429, after bounded retries exhausted)
- backends/langgraph: _complete_turn translates LLMAuthError ->
ChatBackendAuthenticationError and LLMRateLimitError ->
ChatBackendRateLimitError; stream reuses complete() so the mapping
applies to both paths; other errors (timeout etc.) propagate unchanged
- endpoint untouched: existing 401/429 response branches now fire for
both backends
Verified: 212 unit tests (7 new error-mapping tests)
- merge main (b0ff2e0..c7ca781): workflow/20261022 moves dynamic context
out of fixed prompts - collection/extraction prompts carry no inline vars;
reference time and current form travel in the user message <turn_context>
block; photo/confirm data kept at prompt tails
- prompts: all 8 files regenerated from 20261022 (byte-identity tests
point at the new export)
- collect: user message = <turn_context>(本轮参考时间) + <user_input>;
extraction message adds 当前表单; single trusted clock per turn threads
through reply/extraction/repair
- photo: raw input only (20261022 photo nodes use system_text, no wrapper)
- workflow fake llm: unwraps <user_input> before event parsing
- fields: accident_time normalized to extraction format (YYYY年M月D日H点m分,
strip zero padding) - fixes form assertion mismatch with padded output
- evals/run.py: trust_env=False - bypass Windows system proxy (registry
proxy at 192.168.0.4:7897 was intercepting localhost and returning 503)
- verified: 172 unit tests; ALL golden suites pass on real model
(46/46): collection_skip 12/12, photo_skip 4/4, no_false 2/2,
button_guidance 12/12, non_motor_vehicle 14/14, time_confirmation 2/2
- time_context.py: server-side observation clocks for relative-time
evidence (anti client-forgery); resolves X分钟前/小时前 against the
observation timestamp of the turn that produced them, never re-basing
history on the current clock; flags historical references without a
reliable baseline
- state: dialogue_timestamps keyed by dialogue message index, persisted
with checkpoint; welcome seeds index 0
- collect: single clock per turn shared by reply/extraction/repair;
time_evidence injected into runtime context; reply_validator hook
rejects time claims lacking evidence (one shared repair budget);
model still interprets facts and chooses actions
- prompts: collection prompt restructured (shorter, state-bound-to-action
table, gatekeeper principle, time source rules); extraction prompt
aligned; workflow@20261021 json updated to match
- evals: prompt_probes / prompt_compare / time_boundary probe scripts
(supplementary checks, original cases untouched)
- docs: collection-prompt-optimization.md (method + baselines),
time-boundary-fix.md (problem, design, verification)
- gitignore: editor artifact dirs (.agents/.claude/.vscode, skills-lock)
Verified: 171 unit tests pass (16 new time-context tests)
- prompts/collection.md: versioned collection system prompt extracted
from workflow node yQaBxfrWjbRr with source/version/porting notes
- llm.py: LLMService protocol; FakeLLMService (scripted, records
calls); OpenAILLMService (async, wait_for timeout with cancellation,
retry only on timeout/429/connection, errors normalized)
- collect.py: run_collect reference impl - one LLM call produces
state prefix + body, controlled repair (max 1 retry) for missing
prefix or illegal stage proposal, whitelist guard, CollectReplyError
without raw model content; dialogue history + confirmed fields
injected into prompt so the model skips known info (作业4-5)
- fields.py / clock.py: three-state field candidates (unknown/
candidate/confirmed), per-type validation (count not inferred from
description, future time rejected), injectable Clock
- graph.py: welcome node - first message returns fixed 1001 greeting
without LLM (mirrors workflow is_first_message -> Assigned Reply#3),
route empty stage to welcome
- endpoints/delete_session: backend switch - langgraph clears graph
checkpoint, fastgpt path unchanged
- config: AGENT_LLM_PROVIDER fake|openai with validation
- untrack .env (secrets), move template to root .env.example
Verified:
- pytest 103 passed; smoke turns 1/2/1
- Real LLM e2e: POST /chat multi-turn context works
- Golden eval collection_skip_known_info: 12/12 cases pass
- Golden eval no_false: 8/10 cases pass; 4 confirm/photo cases fail
at final step - model routes on short reply '对' (prompt-adherence
gap, photo/confirm stages are assignment 5 scope); recorded in
evals/results/20261020-144856
Known limits: field extraction from model reply not wired (stub);
photo/confirm nodes still scaffold; clarify counting is approximate.
Architecture (mirrors workflow/20261016 chatNode structure):
- events.py: client markers parsed to typed ClientEvent (sources
verified against workflow JSON); plain text like '拍好了' is not an event
- transitions.py: LEGAL_TRANSITIONS whitelist drafted from chatNode
prompts and Condition edges (16 entries, each with source_ref);
is_legal_transition rejects unlisted/terminal-violating moves
- graph.py: multi-node graph (collect/single_photo/dual_photo/...
matching workflow chatNodes), route_by_stage conditional edges,
_fake_llm_propose simulates LLM stage-code proposals (assignment 4
replaces with real LLM); code only keeps context bookkeeping
(no_reply_count) and validation
- state.py: expanded domain fields (pending_question, accident/
person fields, clarify/no-reply counts, turn_patch, state_version)
Key finding: 单车拍照#6/双车拍照#7 are browser_addon branches ending
at 0000, so 2005/2015 allow both 300x (direct) and 0000.
Verified: pytest 87 passed (24 transition/event tests newly green).
Known limit: stage proposals come from a fake LLM; per-stage scripts
and real LLM integration are assignments 4-5.
Eric Wang