T2 of TASK-rectification-grounding-20260927 (product decision P1).
The rectification 「重试（重新生成）」 was a blind agent.generate rewrite that
persisted unchecked text (invented range / fit rate reproduced). Removed:
the regenerate route, regenerate-turn.ts, the regeneration agent and its
read-only tool set, the client regenerate action/state and the
regenerating/canRegenerate props. ChatMessageActions renders the regenerate
button only when onRegenerate is passed; ordinary consultation is unchanged.
The DB function regenerate_agentic_rectification_turn is kept (AGENTS §7.6,
retire in a later round). DESIGN.md / VOICE.md updated; source-contract
tests follow with 原值/新值/原因 notes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
- D2: each intent-classifier attempt is capped at 10 s (same session model,
thinking untouched); a hang takes the existing retry -> classifier_unavailable
path. Success is timed too (RectificationClassifierDiagnostic).
- D3: a typed message builds the NDJSON stream first; the first line is
turn.progress "received", then the classifier and deterministic replies run
inside the stream. Preflight rejections become turn.rejected (old status,
code, message) and the client handles them like the old HTTP rejection.
Stage lines 收到，正在对照你的档案… / 正在记下这件事… / 正在重新对照盘面… /
正在准备下一个问题… are driven by existing tool events and engine calls, are
transient (live row only) and never persisted. VOICE / DESIGN updated.
- D4: RectificationRunDiagnostic records per-step start/end, provider token
usage incl. reasoning tokens, classifier timing and per-engine-call
durations (AsyncLocalStorage scope per turn); RectificationTurnDiagnostic
for deterministic turns. No user text, birth data or model text.
- D5: /v5/versions memo (30 s, complete identities only, per transport); the
exit gate skips its second persistNextInterviewIfIdle when the run's own
call found the next focus already active (provably identical).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
Hospital records, family estimates and period-only windows now keep distinct copy on the board and range card. Scoring still uses the reported clock as the search-window centre. Hospital minutes outside the current range are stated with the offset and no preference.
Refresh persist failures no longer count as attempts. tied_first completes with a range before stillNeedNarrowing. Delivery narration and cards share DELIVERY_OUTCOMES. The question gap gets a delivered terminal so the unavailable copy does not appear after a range is given.
Distinguish followups that cannot stamp a probe now go to persistExhaustionCollect instead of repeating the card stem in the assistant body. Choice focuses without asked_turn_id hang on the last assistant turn, and persisted_question with a live choice_card renders the existing card. Representative-time inconsistency is warn-only (BUG-676 investigating).
Keep targeted existence questions as A-D cards. Recover collect-schema
stock by question-id prefix, surface the stem when persist fails, and
send spoken targeted existence to the repair exit instead of a naked prompt.
GET dropped the A–D card when the recast had no choice_frame, so a refresh left a disabled card and a collect-wait. Project from the persisted copy, rebuild the keep frame, and send a dead choice to repair-exit.
Co-authored-by: Cursor <cursoragent@cursor.com>
Empty engine refresh now records an already_answered attempt so GET can leave wait-to-narrow; collect kinds map onto the table CHECK; remaining candidates drive probe refresh.
Co-authored-by: Cursor <cursoragent@cursor.com>
When dated choice probes are exhausted after the training gate, stop treating yearless D9/D10 cards as the next discriminator and persist a range carrier in the same answer transaction.
Co-authored-by: Cursor <cursoragent@cursor.com>
Stop domain-wheel collecting and age-band years in prompts. Ask until the training gate, then discriminate until convergence, then deliver a range plus a concrete follow-up. Reserve holdout only with four dated events.
Co-authored-by: Cursor <cursoragent@cursor.com>
Evidence turns that already recorded a batch no longer fail the whole run when the model emits no text; unchanged ranges now name which clock spans lead or lag, and the timeline no longer says 收窄.
Co-authored-by: Cursor <cursoragent@cursor.com>
Adopted reverse-verify repeated already-asked collect stems, treated 「这题跳过」 as stopping the case, and promised holdout/OOS checks that never ran. Collect spoken stems no longer prefix a year.
Co-authored-by: Cursor <cursoragent@cursor.com>
Opening a rectification Case used to switch sessions first and read the
turns afterwards, so the reader saw a plain transcript, then an empty
panel, then a remount when the turns arrived (and again after the first
turn settled, because the panel key carried a ready/loading suffix).
The surface hook now reads turns and snapshot in one Case request under
the home reveal budget and only then makes the session active; the
sidebar defers the switch the same way; a session selected at bootstrap
(deep link, refresh) is hydrated during the prepare phase so the reveal
shows the surface itself; the panel key is the session/Case binding only,
later turns fill an empty transcript as a prop update, and unmounting
aborts any stream or snapshot read. The homepage card and the sidebar row
say 正在打开 statically while the Case opens — no spinner after the reveal.
BUG-505
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JUei7K13cYxLHE3Axe4A45
