RCA Fix Agent
Full debugging loop with verified claims
When to use
- Debugging a failing test, incident, or broken build
- You need verified root cause analysis before shipping a fix
- Multiple hypotheses need systematic elimination
The loop
- Baseline. Run the repro, capture the failure as spans.
- Pick one unresolved claim. ROOT_CAUSE, FIX_MECHANISM, FIX_VERIFIED, or NO_NEW_FAILURES.
- Run the smallest next experiment. Inspect one file, run one test, apply one patch.
- Record an attempt. Store all outputs as spans. Log keep/revert/continue.
- Audit. Run
audit_trace_budgeton the affected claims. - Repeat until the repro passes, regressions are checked, and all four claims are verified.
Verified claims
- ROOT_CAUSE — the issue is because of X
- FIX_MECHANISM — the fix works because it changes X which prevents Y
- FIX_VERIFIED — the original repro now passes
- NO_NEW_FAILURES — the regression suite passes
Evidence pack
- S0: Reproduction command + baseline output
- S1: Code excerpt at crash site
- S2: Call site / data flow
- S3: Invariants / spec / docs
- S4: Experiment results (test output, logs)
- S5: Patch diff
Copy/paste prompt
Start a run. Capture the baseline repro as spans.
Generate multiple hypotheses. Test one at a time with the smallest experiment.
Record every attempt with claim_id, hypothesis, action, decision, and next step.
Run audit_trace_budget on ROOT_CAUSE, FIX_MECHANISM, FIX_VERIFIED, NO_NEW_FAILURES.
Keep iterating until all four claims pass and the repro is green.