wrong-check-reflect: one shot per reply, and scope the reflect scan to the turn

[EdbertChan]

Two independent lockouts kept this hook from speaking on Claude Code.

One: already_prompted was keyed on the transcript path, so the Stop of the
reply BEFORE a correction spent the session's single shot. The correction
itself then hit an already-prompted key. The key is now the transcript path
plus a hash of the reply text, so each distinct reply gets its own chance and
the same reply is still judged only once.

Two: user_already_asked_reflect scanned the whole transcript. One /reflect
typed at the start of a session switched the detector off for every later
reply. The scan is now scoped to the user messages of the turn that produced
the reply being judged, which is the case the skip was written for.

An unreadable transcript now says so on stderr instead of silently reading as
"the user did not ask".

The harness-parity test asserts the same reply is judged under both Stop and
stop. It passes before this change too: this hook compares no event name
anywhere, so event-name case cannot be what split its hits by harness. The
test pins that, rather than leaving the parity unstated.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com

Depends-On: #780

---

Note

Medium Risk
Changes when reflect enforcement fires across multi-turn sessions; behavior is hook-local with expanded tests but affects safety/reflection workflows when CATSTACK_REFLECT_ENFORCEMENT is on.

Overview
Fixes two wrong-check-reflect lockouts that stopped the hook from enqueueing LLM judge jobs when it should.

Once per reply: Prompted state is keyed by transcript path plus a hash of the assistant reply (reply_key), not the transcript alone. A Stop on an earlier reply no longer consumes the session’s only shot, so a later correction reply can still be judged; identical reply text is still deduped.

/reflect skip scoped to the current turn: user_already_asked_reflect only scans user messages in the turn that produced the reply being judged, so an early /reflect no longer disables the hook for the rest of the session. Same-turn /reflect still suppresses double follow-up.

Unreadable transcripts emit a stderr hook error (“unchecked”) via _transcript_roles instead of failing open as “user did not ask.” README and tests cover correction-after-prior-reply, session-long /reflect, dedup, harness Stop/stop parity, and read failures.

^{Reviewed by Cursor Bugbot for commit dad472d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[EdbertChan]

This pull request is part of a Mergify stack:

#	Pull Request	Link
1	unverified-tag-ledger: read the turn's tools from the transcript, not a key Claude Code never sends	#780
2	wrong-check-reflect: one shot per reply, and scope the reflect scan to the turn	#781	👈
3	reflect: catch an evidence-order correction from the ledger transition, not from wording	#782
4	unverified-tag-ledger: gate the next-prompt reminder, default to stale only	#783
5	diu-stop: a file citation buys silence only when it is backed	#784
6	diu-stop: the marker gate reads prose, so a mention is not a use	#785
7	cat-mode: show the escape-hatch tag in the form the gate accepts	#786
8	wrong-check-reflect: only the person's own reflect request suppresses the hook	#787
9	principle-subagent-inherits-scope: a brief carries decisions, not just facts and a question	#788
10	wrong-check-reflect: only a real /reflect invocation counts as already asked	#789
11	llm-judge: claude is the only default runner	#793

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit dad472d. Configure here.}

[cursor]

Reflect scan splits mid-turn replies

Medium Severity

user_already_asked_reflect opens the turn after the previous non-empty assistant line. One Claude turn often writes earlier assistant text (a preamble before tools) before the reply being judged, so a /reflect in that turn's user prompt sits outside the window and the skip does not run.

 

^{Reviewed by Cursor Bugbot for commit dad472d. Configure here.}