kj-aiml/heli-harness/fix-loop

fix-loop

Trigger: test failures, CI failures, repeated failed fixes, or multiple findings that need closure.

Scope:

• Triage findings by severity and dependency order.
• Verify the premise for each item before editing.
• Fix one cause at a time.
• Run the smallest relevant check after each fix.
• Track failed attempts.
• Use the failure-class identity in diagnosis.json; do not share one attempt count across unrelated downstream failures.

Failure classes (do not conflate)

Rules:

• After two failed implementation attempts on the same issue, stop coding and write diagnosis.
• Do not retry expensive loops hoping for green.
• A same-class retry stays in fix-loop; a materially new class requires reroute and a new closest boundary. Expensive retries use heli diagnosis gate.
• Do not mix unrelated fixes unless the repo profile permits batching.
• Prefer script files over complex inline shell when friction repeats (see .heli-harness/templates/windows-shell-recipes.md).

Loop:

1. Confirm failure class (friction vs implementation vs capability).
2. Confirm failure still present.
3. Isolate cause.
4. Apply smallest fix (or rewrite command if friction).
5. Verify.
6. Record evidence.
7. Continue or stop with diagnosis.