wrong-check-reflect: only a real /reflect invocation counts as already asked

[EdbertChan]

The suppression matched any prose containing the word, so a sentence about
the hook switched the hook off. "Claim I made was wrong" is a trigger for /reflect describes when the detector fires; it asks for nothing, and
nothing is in flight to avoid duplicating.

Match the harness's own record of an invocation -- the
<command-name>/reflect</command-name> envelope -- instead. That envelope
starts with <command-, which the meta filter drops, so the one
unambiguous signal was the one being discarded while vague prose was kept.
The window now includes meta rows for this check.

Also stop returning False when the transcript holds no assistant row. A
Stop payload can carry the reply before its row lands, and a session's
first turn has none at all; in both cases every row belongs to this turn,
and bailing out meant a real request went unhonoured.

Erring toward asking is the safe direction here: this detector recorded
1,682 Claude Code invocations and spoke 0 times.

tests/scenarios/self-correction.json: the scenario that pinned this said
the user "already asked for /reflect" while its user field only mentioned
the word. Its intent and its data disagreed. The request case now uses a
real invocation; the mention case is its own scenario and must still fire.

Ran 27 tests in 16.245s OK (wrong-check-reflect)
Ran 6 tests in 0.045s OK (scenario suite)
check_hook_test_coverage: OK (1 hook(s) checked)

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

Depends-On: #788

---

Note

Low Risk
Narrower suppression logic in a stop-hook detector with expanded tests; may enqueue slightly more often but avoids missing real reflect requests and false silencing.

Overview
Fixes false suppression when users only mention /reflect in prose (e.g. describing hook triggers). The hook now treats “already asked for reflect” only when the transcript records a real command envelope (<command-name>/reflect</command-name> or <command-message>reflect</command-message>), not bare words like please /reflect in text.

Turn-window logic no longer bails out when there is no assistant row yet (first turn or Stop before the reply is written). The scan runs through the end of the file and includes meta rows so a genuine invocation is not missed while prose about the hook no longer disarms the detector.

Tests and self-correction.json scenarios are split: real invocation still suppresses enqueue; mention-only text must still enqueue wrong-check-reflect.

^{Reviewed by Cursor Bugbot for commit 721426e. 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