T1 of TASK-rectification-grounding-20260927 (recurrence of BUG-588).
- The attempt no longer streams range-changed / rescore-skipped /
compare-failed sentences; the finish whitelists and trims the model body,
then joins the server facts, and emits one final replace equal to the
persisted text.
- P3 whitelist (spoken-grounding.ts): a model sentence with a clock, clock
range or percentage that is not this turn's server fact is dropped whole;
the batch recap stands in when nothing is left.
- record-evidence-batch returns range_after_rescore (post-rescore
credible_range, representative minute, fit percent, delivers_range_this_turn);
the receipt fingerprint stays over the old shape.
- System prompt: range is said by the server; the delivery three sentences
only when the batch says this turn delivers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
Move-only (TASK-rectification-code-split-20260926 T3/T4). No behavior,
receipt, billing, Skill or scoring change.
- agent-run.ts: 1400 -> 135 lines; runV9AgentTurn 1046 -> 25 lines, no
nested functions. It keeps the public types, the budget constants and the
phase order: agent-run-prepare.ts (reliability, session, year re-ask, Skill
identity, delivered guard, reserve, turn row / replay),
agent-run-retry.ts (attempt loop), agent-run-attempt.ts (the old nested
streamAttempt), agent-run-finish.ts (billing settle, receipts, interview,
finalize), agent-run-support.ts (outcome types, retry classification, turn
receipt writers), agent-run-messages.ts (buildAgentMessages /
buildOpeningBrief, re-exported).
- One token changed with the move: the attempt passes its own
`previousErrorCode` argument to buildAgentMessages instead of reading the
enclosing `lastAttemptError`; the loop passes that same value (clears the
old unused-parameter warning).
- Tests: whole-source contracts read tests/rectification-agent-run-surface.ts;
behavioral runV9AgentTurn tests unchanged. agent-run growth caps added.
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
Occupation collect was wiping year-month into an unscored note, and idle gap copy never joined the evidence turn.
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>
Ask finance and health like other domains, put year-cued collect questions first, and let the classifier distinguish decline vs skip. Skill 10.0.22.
Co-authored-by: Cursor <cursoragent@cursor.com>
Host only says 记下了 after a real write; Mastra schema rejections fail closed. Ledger year keys no longer drop quality probes, dual-dasha agreement is per cluster, width uses cluster span, and public house tables follow the inference minute.
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>
Hour windows no longer drop later signature clusters. Credible range uses cluster coverage, and a mid-session spoken birth window gets a fixed reply without calling the model.
Co-authored-by: Cursor <cursoragent@cursor.com>
Engine by_time ledgers now include the full candidate set so posterior columns can look up fit and windows. Delivery turns keep up to three sentences after idle persist; the card subtracts repeated traits and folds the board chart.
Co-authored-by: Cursor <cursoragent@cursor.com>
Make the range card row-select, drop the composer 先这样 control and adopt status bar, keep delivery copy to three sentences with a folded verification report, and skip a second no-message agent run after a terminal delivery turn.
Co-authored-by: Cursor <cursoragent@cursor.com>
Rescored candidate sets were widening because new minutes stayed
neutral. Derive support/conflict from varga signs and dasha segments,
and announce range changes from facts rather than tool names.
Co-authored-by: Cursor <cursoragent@cursor.com>
Carry answered probe defs into the new candidate set, replay them by minute, and announce range changes on evidence turns.
Co-authored-by: Cursor <cursoragent@cursor.com>
Set-focus must not append a second narration, and a server-owned stem must not also appear as a question in the assistant body.
Co-authored-by: Cursor <cursoragent@cursor.com>
Engine asked_probe_keys no longer include varga split hashes that 400 the scorer, failed compares become visible and retry, user stop can still deliver a range on a stale snapshot, and holdout no longer reasks domains already in the ledger.
Co-authored-by: Cursor <cursoragent@cursor.com>
Intake stores how sure the user is; rectification now searches that range, offers a one-click widen when event fit is low at the edge, and trisects windows longer than two hours before the minute grid.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep askable cards after exhaustion, explain each probe, read the adopted credible range in reports and chat, and compare declared periods before the minute grid when the clock is unknown.
Co-authored-by: Cursor <cursoragent@cursor.com>
Focuses now carry asked_turn_id so GET rebuilds stem and options on the
same turn. Agent writes spokenPrompt; the live question slot is gone.
Co-authored-by: Cursor <cursoragent@cursor.com>
The live question slot disappeared on refresh because it was never written to assistant_message. Attach the current collect_spoken prompt to that turn before finalize so chat history keeps it.
Co-authored-by: Cursor <cursoragent@cursor.com>
Persist the interview slot before the opening turn, keep the stem in the question slot, and hide that slot while a run is busy so the same ask does not appear twice.
Co-authored-by: Cursor <cursoragent@cursor.com>
Choice path already wrote spokenFollowupForUser into the body; free-text
dropped that stem and only filled empty answers, so a new collect_spoken
focus stayed invisible after “记下了”.
Co-authored-by: Cursor <cursoragent@cursor.com>
The standalone collect prompt bar reused the choice-card chrome and
duplicated the question. Visibility now comes from the spoken reply,
with an empty-body fallback that posts the persisted focus prompt.
Co-authored-by: Cursor <cursoragent@cursor.com>
Closing a discriminator used to leave GET without a card after refresh.
Write the next dated question in the same request, skip childhood career
and move probes, and do not continue a read-only turn when that question
is already persisted.
Co-authored-by: Cursor <cursoragent@cursor.com>
Python and TypeScript now share a four-option probe contract, persist Focus before asking, and pick the highest-value renderable probe instead of preferring low-gain career events over D24.
Co-authored-by: Cursor <cursoragent@cursor.com>
Exam-quality cards may still jump ahead of adoption, but career years stay on method rotation. Server stamps only period and family; spoken questions remain model-authored.
Co-authored-by: Cursor <cursoragent@cursor.com>
Recorded-year quality probes were spoken-only, so the interview had no
choice card. Compare also re-scored after batch until the 105s attempt
aborted the turn.
Co-authored-by: Cursor <cursoragent@cursor.com>
Conflict probes were jumping after one dated event, so the interview asked
another domain before method collection. Spoken replies now follow the
stamped choice prompt instead of a topic denylist.
Co-authored-by: Cursor <cursoragent@cursor.com>
Give the interview a hidden reasoning channel so planning leaves the spoken reply, publish terminal text-delta as-is, and stop regex or Case templates from replacing the model.
Co-authored-by: Cursor <cursoragent@cursor.com>
Aborting the turn on a second identical public tool-call failed staging
after evidence and compare had already succeeded. Skip the duplicate
receipt instead and let maxSteps bound real loops.
Co-authored-by: Cursor <cursoragent@cursor.com>
Mastra intermediate text-delta was published as answer.delta, then set-focus domain errors reset the attempt and replayed evidence. Publish only the terminal no-tool step, persist the next probe on the server, and ground batch quotes in the source turn.
Co-authored-by: Cursor <cursoragent@cursor.com>
Clicking A/B/C/D or stop must persist the answer, close the probe, and
update posteriors in one idempotent transaction instead of sending the
option text as a chat message.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep provider thinking on a separate channel so process talk is not billed as the spoken reply (BUG-359).
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep provider thinking off, classify CoT as 思考, and stop remounted sessions from firing a second opening.
Co-authored-by: Cursor <cursoragent@cursor.com>
Disable provider thinking so Flash CoT cannot fill max_tokens, raise the
spoken budget to 16384, emit a server-owned step tree, and continue once
when the body ends on length.
Co-authored-by: Cursor <cursoragent@cursor.com>
Enumerate evidence kinds so education cannot be proposed as a kind, and stream Chinese thinking on a separate channel that collapses when the reply arrives.
Co-authored-by: Cursor <cursoragent@cursor.com>
Rectification dropped tool.activity started events and treated length finishes as completed. Share generation settings with consultation, keep the activity line through streaming, and name multi-domain chart calculation.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use the quiet overlay scrollbar on chat and the house board, replace the blocking mobile dialog with a bottom sheet, and emit agent deltas before billing so users can read progress and stop a run.
Co-authored-by: Cursor <cursoragent@cursor.com>
