Repeat Failure Analysis

---
name: repeat-failure-analysis
description: "`analysis-agent`/`task-agent`/`review-agent`: use when repeated failure needs a new hypothesis or proof path; skip an initial failure with verified cause and a different action."
---

# repeat-failure-analysis

## Registry Trigger

**Use when**

- the same path, cause, patch shape, or validator has failed twice
- a repair repeats an approach already contradicted by current evidence

**Do not use when**

- the first failure has a verified cause and a materially different next action
- prior attempts are unavailable and the current repository must be inspected first

## Skill Role

Analyze repeated failure from observable source, commands, diffs, findings, and current-task evidence. Do not create a persistent memory subsystem.

## Inputs

- attempted hypotheses and their observable results
- current diff, failing output, changed paths, and acceptance
- known owner, adjacent patterns, and validation entry points

## High-Value Rules

- Classify observed facts, rejected hypotheses, and untested hypotheses separately.
- After two failures, do not repeat the same path without new evidence.
- Search the same pattern and affected consumers before applying a local repair.
- Change at least one material dimension: hypothesis, owner, patch location,
  instrumentation, reproduction, or validation method.
- Select the smallest feasible reproducer that can falsify the next hypothesis without losing the suspected mechanism.

## Anti-Patterns

- Renaming the same patch is not a different approach.
- A green unrelated command does not disprove the observed failure.
- Previous conversation summaries are navigation hints, not source truth.

## Execution Checklist

1. List the failed attempts on the rejected path and the evidence each produced.
2. State why the prior path is rejected or still uncertain.
3. Inspect the owner and same-pattern occurrences.
4. Choose one falsifiable next hypothesis and a different proof path.
5. Return the bounded next action or a concrete blocker.

## Stop Conditions

- Stop when no new hypothesis or evidence distinguishes another attempt.
- Escalate when the next probe is destructive, privileged, production-facing,
  or outside the allowed scope.

## Output Contract

- observed facts, rejected hypotheses, same-pattern impact, next hypothesis,
  changed execution path, validation method, proof limits, and blocker if any

## Targeted References

| Path | Type | Load when | Do not load when | Required by | Required output |
|---|---|---|---|---|---|
| [repeat failure](references/repeat-failure-checklist.md) | decision-checklist | the same path cause patch or validator has failed twice | this is the first verified failure or the next path is materially different | analysis-agent, task-agent, review-agent | checklist-result, residual-risk |