| name | self-critique |
| description | Before claiming "done / fixed / tests pass", run smelltest's model-free structural kernel on your own output — re-grading the claim against the actual git diff (substantive added lines, test-tampering) instead of trusting your sense of completion. Use whenever you are about to write "done", "implemented", "fixed", or "tests pass". |
Self-critique before you claim "done"
Your sense that it's finished is systematically overconfident. Re-grade the claim against the
structure of what actually changed.
Run the kernel on yourself
Easiest — let the extractor build the evidence from the transcript + git:
node "${CLAUDE_PLUGIN_ROOT}/src/cli.ts" --latest
Or pipe a hand-built evidence object (the diff is parsed into hunks — the kernel counts
substantive added lines itself):
{ "finalMessage": "<your about-to-send message>",
"diff": { "available": true, "isEmpty": false, "filesTouched": ["src/x.ts"],
"hunks": [{ "file": "src/x.ts", "addedLines": ["..."], "removedLines": [] }] },
"scope": { "filesRead": ["src/x.ts"], "filesEdited": ["src/x.ts"] } }
echo '<evidence>' | node "${CLAUDE_PLUGIN_ROOT}/src/cli.ts" --stdin
Act on the verdict
- WARN
done.no_substance → you claimed completion but the diff added 0 substantive lines
(comment/stub/import don't count). Make the real change, then re-run.
- WARN
tests.tampered → the diff weakened the test suite under a "tests pass" claim.
Never edit tests to pass — fix the code.
- note
scope.blind_edit → say it out loud — you edited a file you never read.
- No flags → say "structural checks did not flag this", never "verified". A self-graded
claim never earns a check.
