wrong-check-reflect: only the person's own reflect request suppresses the hook

[EdbertChan]

user_already_asked_reflect matched any user-shaped line mentioning reflect.
The harness files its own injections as type: "user" rows, so two of them
matched:

• diu-stop's Stop-hook feedback, whose text tells the agent to reflect
• the reflect skill's own injected body, so running /reflect disarmed the
  hook that asks for it

Read-confirmed on a real transcript: both carry isMeta: true on a
type: "user" row, which is the field
engine/skills/reflect/scripts/token_audit.py:313 already keys off. This reads
the record rather than a prose prefix, which could only ever catch wordings
somebody had already seen. agentId and isSidechain are excluded the same
way.

Chesterton's fence checked. git log -S "user_already_asked_reflect" origin/main returns three commits: 5b3b2fb, 30f62e5 "Add
wrong-check-reflect hook for first-person check admissions (#57)", and
e9a9750 (#462). 30f62e5 is the one that introduced the function, and its
message asks for the follow-up "once per transcript" -- do not nag twice.
Matching any mention anywhere in the file is far wider than that. This
narrows it back to the original intent; the turn scoping and the
once-per-reply key are the commit before this one.

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

Depends-On: #786

---

Note

Medium Risk
Changes when the reflect-enforcement hook is suppressed on Stop; misclassification could either nag after the user already asked for reflect or skip needed follow-ups, though behavior is covered by new tests.

Overview
Fixes wrong-check-reflect so harness-injected transcript rows no longer count as the user asking for /reflect.

user_already_asked_reflect used to treat every user-shaped line in the current turn as the person speaking, so diu-stop Stop-hook feedback and the reflect skill’s injected body (both filed as type: "user" with isMeta) could mention “reflect” and silence the hook—including the case where running /reflect disarmed the detector that should follow up.

The change adds _is_meta_line (aligned with token_audit via isMeta, plus agentId / isSidechain and a "Stop hook feedback" prefix fallback), tags those rows as meta when parsing the transcript, and only scans real user lines for the reflect regex. README documents the behavior; tests cover hook feedback, skill body, and genuine user /reflect in the same turn.

^{Reviewed by Cursor Bugbot for commit bd6a6fd. 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

[EdbertChan]

Open question for the reviewer: one existing scenario now fails, and it contradicts itself

bash scripts/test/run_all_tests.sh on this stack:

Ran 644 tests in 1472.932s
FAILED (failures=1)
  'wrong-check-reflect: expected NO judge job, queued one'

The failing case is tests/scenarios/self-correction.json, scenario admission-skipped-when-user-already-said-reflect:

"situation": "Pins the documented skip: ALREADY_REFLECT_RE suppresses the follow-up when the user's own message already asked for /reflect ...",
"user": "\"Claim I made was wrong\" is a trigger for /reflect",
"expect_no_enqueue": ["wrong-check-reflect"]

The situation and the user field disagree. The situation says the user asked for a reflect. The user text does not ask for anything — it quotes the phrase that triggers a reflect while describing the rule. That is a mention, not a request. This PR makes the hook fire on it, which is why the scenario goes red.

This is the same mention-vs-use defect that PR #785 fixes in the marker gate, showing up in a fixture instead of in code.

Not resolved here, deliberately — a scenario whose label and data contradict each other is a judgment call for the owner, not something to silently re-point. Recommendation:

1. Change this scenario's user to an actual request (e.g. please /reflect on this) so it keeps pinning the legitimate skip: do not nag when a reflect is genuinely in flight.
2. Add a second scenario using the current text verbatim, pinning that a mention does not suppress.

Separately: scripts/test/run_all_tests.sh printed FAILED (failures=1) and still exited 0. A check that failed reported success. Worth its own fix; not in this stack.

🤖 Generated with Claude Code