Challenger Review

An additional independent review pass that strengthens or weakens candidate findings.

When to use

• Change is medium or high risk (yellow/red lane)
• Authoring was heavily AI-assisted
• Primary reviewer confidence is low
• Findings conflict or remain uncertain
• Configured to run automatically via enable_challenger_on setting

Key requirement

This skill must be usable in two configurations:

• Same-model challenger: same model family as primary reviewer, fresh context
• Cross-model challenger: different model family from primary reviewer

This allows apples-to-apples comparison between intra-model and cross-model review.

Inputs

• Evidence pack from pr-evidence-builder
• Raw diff

Do NOT read primary reviewer conclusions. The challenger always reviews independently.

Steps

1. Verify independence before starting. Confirm you have not read any files from the primary review output directory. Do not access findings from fresh-eyes-review. If any primary review content is in context, stop and restart with a clean context.

2. Review the evidence pack and raw diff independently. Approach the code as a separate critic with no knowledge of prior findings.

3. Produce candidate findings. Same output format as fresh-eyes-review. Each finding must meet the evidence threshold. If a finding does not meet the threshold, discard it rather than downgrading its severity.

4. For each finding, indicate whether it:

   • Confirms a likely issue (new independent evidence for something that may also appear in primary review)
   • Refutes a possible concern (evidence that something is actually safe)
   • Adds a novel finding (something the primary reviewer may have missed)

Output shape per finding

Same schema as fresh-eyes-review, with source set to "challenger". Example:

{
  "source": "challenger",
  "file": "src/api/handlers.ts",
  "line": 42,
  "severity": "high",
  "classification": "confirms",
  "title": "Unbounded query size allows denial-of-service",
  "evidence": "req.body.ids is passed directly to the IN clause with no length check — an array of 100k IDs will lock the table.",
  "recommendation": "Add a maximum length check (e.g., 100 items) before constructing the query."
}

Classification values:

• "confirms" — independent evidence supporting an issue that may also appear in the primary review
• "refutes" — evidence that a suspected concern is actually safe
• "novel" — finding not likely surfaced by the primary reviewer

Success criteria

• Improves precision or novelty, not just volume
• Reduces shared blind spots between review passes