recovery-loop-3strike

name: recovery-loop-3strike description: "Recover from failed execution or verification with a bounded three-step loop: retry, focused fix-spec, then honest handoff. Use when an agent is likely to spin on the same failure, when flaky environment issues blur the signal, or when you need a disciplined stop condition instead of endless optimistic retries." license: MIT metadata: abvx_status: experimental abvx_origin: adapted

Recovery Loop 3-Strike

Bound retries so failure handling becomes diagnostic instead of repetitive.

Use When

• a verification step fails and the cause is not yet clear;
• the task is long enough that repeated blind retries are likely;
• you need an explicit escalation path for autonomous or semi-autonomous work.

Core Rules

• Failure one is for learning, not denial.
• Each retry must add new evidence or a narrower hypothesis.
• After two failed passes, stop improvising and write a focused fix plan.
• After the third failure, hand off honestly instead of pretending the loop is still productive.

Workflow

1. First failure:
   • capture the exact failing signal;
   • state the most likely cause;
   • retry once with one concrete probe or adjustment.
2. Second failure:
   • write a compact fix spec for the same phase or task slice;
   • constrain the change to the verified failure mode;
   • execute the fix and re-run verification.
3. Third failure:
   • stop;
   • mark the task or phase blocked;
   • hand off with probe history, attempted fixes, and best next action.

Required Failure Record

For each strike, preserve:

• failing command or acceptance criterion;
• observed output or symptom;
• hypothesis tested;
• action taken;
• result.

Anti-Patterns

• Do not spend all three strikes on the same unmodified action.
• Do not escalate to a large rewrite unless the evidence points there.
• Do not hide a blocker behind optimistic language.

Final Report

State which strike resolved the issue, or provide a clean blocked handoff with evidence.