-
…-budget projection (conservative estimator, in-turn demotion, pre-send self-check, prompt_eval_count calibration)
-
…is 客户名称 — exact match silently dropped the prefill); deterministic tie-break prefers unique fkselect/名称 field, ambiguous cases skip
-
- single ReAct entry for all free text: 7 fixed tools (useSkill + 6); flow knowledge lives in 5 skill files (resources/skills/*.md, first line = brief for prompt index; xly.skills.dir overrides for hot reload) - useSkill returns skill body and logs skill_active; projection pins the active skill full text on the flow card until confirm/cancel/new skill - system prompt from template prompts/system.txt (domain map + skill index); bench reads the same files to avoid replica drift - delete IntentService/Intent/SlotFillService/LlmJsonClient/StateService/ OkHttpUtil and route()/handleCreate/handleWrite/deriveWriteAction/ ground/withState; dimension-split regex migrated into FormCollectTool prefill assist (model provably fails dimension splitting) - anti-fabrication guards now unconditional (zero-tools+digits retry, WRITE_CLAIM check); invariant layer and button endpoints untouched - bench_ext: --arch new (same-source prompt/skills) and --fifty (bench50 cases judged on full trajectory for both arches)
-
- chat:ledger becomes the only conversation store: model messages land as user/tool_call(payload)/tool_result/ai events via EventLogChatMemory (per-event rightPush, no more whole-blob read-modify-write races) - EventProjectionService: one renderer for both frontend history and LLM context (system+flow card / deterministic digest of expired turns / ~6000-char verbatim recent turns, current turn payload-exact) - pinned active-flow card (form pending / proposal pending / skill_active full text) unpinned on executed/cancelled/skill_done - anti-fab retry nudges logged as internal user events (LLM-visible, hidden from history); grounded suffixes never duplicate user events - delete ProjectedChatMemory; RedisChatMemoryStore demoted to legacy read-only compat (old convs age out via 30d TTL); conv list lazy cleanup - tests: projection correctness + concurrent stream/confirm append safety
-
First tests in the repo — they pin the audit fixes that are easiest to silently regress: no identity without introspection, no conversation across users, no silent numeric corruption, no tenant-less FK query.