audit(token_audit): count naming the shortcut, and typing the fix command

[EdbertChan]

Summary

The same-type intervention detector matched only told-you, accusation, agent-blame, restated-ask, and proof-challenge.

So a session where the user twice named the shortcut the agent had taken, and typed the fix-the-process command, scored zero and reported that no automate-me was owed. The audit said no while the user was asking for exactly that.

Two kinds join the set. One catches naming the shortcut ("cheap way out", "why would you", "straight up", "that's weird"), the other catches a typed /automate-me, /reflect, or /thrash.

Reaching for the fix-the-process command is itself the intervention.

Before and After

A two-message session where the user says "cheap way out", then "why would you ... straight up ... thats wierd", and types /reflect, run through python3 engine/skills/reflect/scripts/token_audit.py codex session.jsonl on the base (54530d6) and on this PR (e8454f8):

Before:
  [1] 2026-09-09T05:46:08.000Z ['multi-question-marks']: 'why would you submit a rule that would ignore files straight up??? thats wierd'
frustration-flagged user messages: 1/2; interruptions: 0; peak window 2026-09-09T05:46:08.000Z -> 2026-09-09T05:46:08.000Z
intervention-must-automate: no (count=0) no repeated intervention class (one correction is not automate-me)

After:
  [0] 2026-09-09T04:40:47.000Z ['cheap-way-out', 'explicit-invocation']: 'since you took the cheap way out and needed a stop hook, add that hook and /reflect'
  [1] 2026-09-09T05:46:08.000Z ['cheap-way-out', 'multi-question-marks']: 'why would you submit a rule that would ignore files straight up??? thats wierd'
frustration-flagged user messages: 2/2; interruptions: 0; peak window 2026-09-09T04:40:47.000Z -> 2026-09-09T05:46:08.000Z
intervention-must-automate: yes (count=3) same-type complaint / iteration: cheap-way-outx2, cheap-way-out+explicit-invocation

Review Claim

Naming the shortcut twice, or typing the fix-the-process command, counts toward the same-type intervention that makes automate-me mandatory.

Review Lane

behavior

Review Unit

tooling-policy

Safety Invariant

Two entries are appended to the pattern list and two names to the kind set; nothing existing is edited, so every previously flagged session flags the same way. Product blame and plain questions still score zero, asserted by a negative fixture alongside the positive one.

Slice Rationale

The detector change is separable from the skill prose in the slice below it, and either can be reverted alone.

Non-goals

Does not change what the flag causes to happen, only when it fires. Does not touch the frustration score itself or the cost audit.

Test Plan

Test Plan

• python3 -m unittest discover -s engine/skills/reflect/scripts/tests — 172 tests, OK (2 new: the real shape fires, product blame and a plain question stay silent)
• Re-audited the session that motivated this: intervention-must-automate: yes (count=3) where it previously read no (count=0)
• python3 scripts/check_no_new_comments.py --base origin/main — ok

Revert Plan

Revert Plan

• Safe to revert? Yes
• Revert command: git revert <sha>
• Post-revert steps: None
• Data migration? No

🤖 Generated with Claude Code

https://claude.ai/code/session_01KU2pPKob4MJ1NqjsfTNyYJ

---

Note

Medium Risk
Changes live PreToolUse/Stop routing for subagent publishing and how CAT-UNVERIFIED tags behave; mis-tuned command parsing or liveness caching could block or allow the wrong path, but tests and fail-open/UNCHECKED branches are documented.

Overview
Replaces agent-routing-guard (PreToolUse on Agent prompt regex) with publish-act-guard, which blocks actual publishing shell commands (git push, gh pr create|merge|ready, mutating gh api on pulls, mergify stack push, create-pr.mjs, safe-stack-push.mjs --execute) only when the caller is a subagent and invoker-cli query capacity says a live owner is reachable. Liveness failures are UNCHECKED and allow the command with a named reason; install merges the new hook and strips the retired Agent hook entry.

Adds unverified-tag-ledger: Stop records well-formed {{CAT-UNVERIFIED: …}} tags, refuses a turn that tags without running a verification tool (unless stop_hook_active), and UserPromptSubmit reminds on outstanding claims; it opts out of SubagentStop inheritance. pr-schema-gate no longer treats validator exit 0 as clean when output says UNCHECKED/SKIPPED/not installed.

Policy and reflect: cat-mode and named-constraints add hook/classifier block is a stop, not a puzzle; CLAUDE.core.md tightens evidence for cache/offline reads and stale readings after restarts. reflect step 5 requires fanout_complete.py (complete / incomplete / unchecked exit codes) before applying Accepted findings. token_audit counts cheap way out / why would you and typed /reflect|automate-me|thrash toward intervention-must-automate.

^{Reviewed by Cursor Bugbot for commit 0d53f86. 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	cat-mode: rejoin wrapped prose so the file is back under its own cap	#382
2	hook(diu-stop): scope the unverified marker to its own paragraph	#383
3	skill(reflect): do not apply findings from a fan-out that never returned	#384
4	audit(token_audit): count naming the shortcut, and typing the fix command	#385	👈

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_e37bedbb-23a6-4755-98f3-4a91bd38efcd)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_ce859959-1e3a-4f4c-9afc-5742617a75de)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_d7e45ac9-b31a-4fea-932a-5a7336830c0d)

[mergify]

Queued — the merge queue status continues in this comment ↓.

[EdbertChan]

@Mergifyio queue

[mergify]

Merge Queue Status

• ✅ Entered queue — 2026-09-12 18:17 UTC · Rule: admin-bypass · triggered by @EdbertChan with the @mergifyio queue command
• ✅ Checks passed · on draft merge queue: checking #385 on main (d94bde8) #495
• ✅ Merged — 2026-09-12 18:22 UTC · at 814e208994ef31fd2d7b6a565e31bb0be684ebce · squash

This pull request spent 5 minutes 16 seconds in the queue, including 4 minutes 28 seconds running CI.

Required conditions to merge

• check-success = lint
• check-success = test