Excuses (1) admitting a mistake comes with the fix in the same reply

[EdbertChan]

Summary

When the agent admits a mistake, it must now fix it in that same reply. A revert, a check run, a fix, or a named task id all count.

Before this, an admission could end the turn. The agent said "You're right, I did not follow that instruction," then marked the undo done without undoing anything.

A look at 81 sampled excuse-shaped replies found 10 that were never fixed. In 9 of 20 real excuses, the person had to push before anything happened.

This adds one rule to the correction principle skill, plus two new examples: one where the rule fires, one where the fix already came in the reply.

Review Claim

The always-on correction principle requires that an admission of fault carry an action in the same reply: a revert, a check run, a fix, or a named task or issue id.

Review Lane

behavior

Review Unit

corpus-lesson

Safety Invariant

Text-only change to one principle skill and its two example files. No hook, script, or install behavior changes.

Slice Rationale

One rule added to corpus/skills/principle-flag-your-own-corrections/SKILL.md plus its two examples in tests/. The lane is behavior since the rule changes what the agent does after admitting fault, even though only skill text changes. The hook that enforces this rule is a separate workflow, and it needs the rule written down first.

Non-goals

• No hook, script, or install change.
• No edit to cat-mode or any CLAUDE.md file.
• No overlap with the files of open PRs Four checker misfires from one session get fixtures and fixes #482, Reflect and automate-me hooks are off until one setting turns them on #500, or Wrong-check-reflect tests use the shared judge test base #528.
• "Want me to fix it?" stays allowed when the fix is out of scope or cannot be undone.

Test Plan

Test Plan

• python3 engine/skills/make-pr/scripts/preflight.py --base origin/main: exit 1. Every gate passes except the check that new rule text comes with code. This slice is text only, and the hook that enforces the rule is a separate workflow.
• python3 scripts/check_codify_has_code.py --base origin/main --allow-prose-only: exit 0. Run by hand, as preflight.py directs for a text-only change.

$ python3 engine/skills/make-pr/scripts/preflight.py --base origin/main
unit    corpus-lesson: 3 file(s)
declare Review Unit: corpus-lesson
gate    python3 scripts/check_codify_has_code.py --base origin/main
        fail  new rule prose with no code change (thrash-reflect-automate step 3 without step 2):
gate    python3 scripts/check_skills_three_harnesses.py
        ok      skills three-harness check
gate    python3 scripts/check_ecosystem_boundaries.py
        ok      ecosystem boundaries
gate    python3 scripts/run_skill_scenarios.py
        ok	all 17 scenario(s) behaved as declared
fail    preflight: fix the above before gh pr create
exit=1

$ python3 scripts/check_codify_has_code.py --base origin/main --allow-prose-only
ok      codify-has-code: prose-only change explicitly allowed (--allow-prose-only)
exit=0

• python3 scripts/check_skill_test_coverage.py --base origin/main --head HEAD: ok skill test coverage, exit 0
• python3 scripts/check_skill_file_refs.py: ok skill file refs, exit 0
• python3 scripts/check_no_dated_provenance.py: ok no dated provenance, exit 0
• python3 scripts/check_no_new_comments.py: ok no new comments, exit 0
• python3 scripts/run_skill_scenarios.py (run by preflight): ok all 17 scenario(s) behaved as declared
• bash scripts/scrub-handoff-artifacts.sh: scrub-handoff-artifacts-ok, exit 0

Revert Plan

Revert Plan

• Safe to revert? Yes
• Revert command: git revert <merge-sha>
• Post-revert steps: Re-run ./install.sh so installed skill copies match the reverted text.
• Data migration? No

🤖 Generated with Claude Code

---

Note

Low Risk
Text-only updates to one principle skill and two example files; no hooks, scripts, or runtime behavior changes.

Overview
Admissions of fault must now include a same-reply action, not just disclosure and reflect. The principle-flag-your-own-corrections skill adds a fourth obligation: name what was done in that turn—a revert, check run, fix, or task/issue id—and limits “Want me to fix it?” to out-of-scope or irreversible cases. A short jidoka analogy frames admission as detect-and-stop, with the reply carrying the fix so bad state does not roll forward.

The fires and stays silent harness examples are rewritten around incomplete vs complete corrections (e.g. saying undo was done without reverting vs reapplying the fix and reporting the check in the same message), and the fire case now explicitly includes the new fourth obligation.

^{Reviewed by Cursor Bugbot for commit 834de84. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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_d8b66d7f-2999-467d-93b3-67e1858f29c2)

[EdbertChan]

Backtest against real labeled cases with the real model judge (not the unit tests' stand-in) found this not fit to merge. Closing to start over with the backtest as the first step, before any code.

Judging 12 real admissions with this rule caught 4 of 5 excuses but flagged 3 of 6 non-excuses.