Excuses (6) handing the user a command needs an attempt first

[EdbertChan]

Summary

Small checker scripts read each reply Claude writes. This adds one that notices when Claude asks the person to run something Claude never tried.

The problem: agents hand work back too early. One wrote "please run" a setup step without trying it, and the person answered "why don't you just run it for me".

The cause: the rule against this lived only in written guidance. Nothing read replies for it, so it kept happening.

The fix: a new checker asks a small model whether the reply hands work back with no earlier attempt. The answer arrives on the next turn.

Review Claim

A reply that asks the person to do a step the agent never attempted is flagged on the next turn, unless a refusal came first or only a human can do the step.

Review Lane

policy

Review Unit

engine-runtime

Safety Invariant

A hand-back that follows a permission denial, a sandbox or classifier refusal, or a step only a human can do (password, OAuth consent, hardware) is never flagged. If the judge cannot answer, the next turn says it could not judge; it never reports clean.

Slice Rationale

One new hook with its phrase dictionary, tests, and install wiring. The one-line pointer from the cat-mode hand-back rule to this hook is a separate corpus-lesson PR, because it is a different review unit.

Non-goals

• No change to any other hook's behavior.
• No edit to cat-mode or other skills (follow-up PR).
• Leaves 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 and Wrong-check-reflect tests use the shared judge test base #528 alone.

Test Plan

Test Plan

Built by Invoker workflow wf-1789281140471-18. Its only failing check, python3 scripts/check_install_effective.py, reads this machine's installed hooks rather than the branch, and fails on a clean origin/main checkout too (exit 1, 116 "bypasses the metrics runner" lines, 0 mentioning this hook).

Run on this branch:

• python3 -m unittest discover -s engine/hooks/handback-needs-attempt/tests → OK
• python3 scripts/check_hook_test_coverage.py → OK (38 hooks)
• python3 scripts/check_no_silent_hook_except.py → checked 205 files
• python3 scripts/check_rules_are_wired.py → OK (9 checkers)
• python3 scripts/check_no_new_comments.py --base origin/main → ok
• python3 engine/skills/make-pr/scripts/preflight.py --base origin/main --body-file <this body> → see below

unit    engine-runtime: 9 file(s)
neutral 1 file(s): docs/ecosystem.md
declare Review Unit: engine-runtime
gate    python3 scripts/check_hook_test_coverage.py engine/hooks/handback-needs-attempt
        check_hook_test_coverage: OK (1 hook(s) checked)
gate    python3 scripts/check_hook_test_coverage.py engine/hooks/llm-judge
        check_hook_test_coverage: OK (1 hook(s) checked)
gate    description_check /private/tmp/claude-501/-Users-edbertchan-Documents-GitHub-catstack/9b25a21a-0b29-4a9a-bcff-548fc74d2eef/scratchpad/pr-handback.md
        description clean
ok      preflight passed

Revert Plan

Revert Plan

• Safe to revert? Yes
• Revert command: git revert <merge sha>
• Post-revert steps: rerun ./install.sh so the hook link and settings entry are removed
• Data migration? No

🤖 Generated with Claude Code

https://claude.ai/code/session_016VgzyERsebhUHizFhFqP2t