Interrogate
Review the requested diff or artifact against its intended outcome. Deliver a
verdict; apply fixes only when implementation is also requested. Requires the
artifact and surrounding context; execution strengthens uncertain findings.
- Define scope and intent. Identify the requested revisions, working-tree
changes, and relevant constraints. Derive the intended behavior from the user's
request, specification, or PR. Resolve uncertainty that would change the verdict
before classifying behavior as a bug.
- Challenge the change. Inspect correctness, failure paths, boundary validation,
compatibility, state ownership, test coverage, and maintainability where relevant.
For each proposed finding, name the trigger, reachable path, observable consequence,
and precise location. Distinguish introduced failures from pre-existing issues.
- Check the findings. Trace real callers and guards, and run a small meaningful
reproducer where affordable. A hypothetical input is not a reachable bug without
supporting evidence. A suggested abstraction needs a concrete benefit in this
scope. Keep valuable rationale comments; comment removal is not a quality metric.
- Judge and prioritize. Deduplicate and classify supported findings as act on,
consider, or contextual. Dismiss incorrect or purely stylistic claims with a brief
reason when they were material to the review. A single verified severe finding
outweighs reviewer agreement. Avoid quotas on either findings or dismissals.
- Deliver the verdict. Lead with actionable findings ordered by impact, each
with a source location, trigger, consequence, and evidence or verification gap.
Explain material tradeoffs and disagreements. If there are no supported findings,
say so and state the review's coverage limits without inventing issues.
If the task requests independent reviewers and the harness supports them, give
reviewers the same intent and context before they see each other's conclusions.
Verify their reports yourself. Multiple opinions are optional evidence-gathering;
consensus is not proof, and sequential passes by one agent are not independent.
1---2name: interrogate3description: Adversarially review a change and deliver a prioritized verdict grounded in reachable failures.4---56# Interrogate78Review the requested diff or artifact against its intended outcome. Deliver a9verdict; apply fixes only when implementation is also requested. Requires the10artifact and surrounding context; execution strengthens uncertain findings.11121. **Define scope and intent.** Identify the requested revisions, working-tree13 changes, and relevant constraints. Derive the intended behavior from the user's14 request, specification, or PR. Resolve uncertainty that would change the verdict15 before classifying behavior as a bug.162. **Challenge the change.** Inspect correctness, failure paths, boundary validation,17 compatibility, state ownership, test coverage, and maintainability where relevant.18 For each proposed finding, name the trigger, reachable path, observable consequence,19 and precise location. Distinguish introduced failures from pre-existing issues.203. **Check the findings.** Trace real callers and guards, and run a small meaningful21 reproducer where affordable. A hypothetical input is not a reachable bug without22 supporting evidence. A suggested abstraction needs a concrete benefit in this23 scope. Keep valuable rationale comments; comment removal is not a quality metric.244. **Judge and prioritize.** Deduplicate and classify supported findings as act on,25 consider, or contextual. Dismiss incorrect or purely stylistic claims with a brief26 reason when they were material to the review. A single verified severe finding27 outweighs reviewer agreement. Avoid quotas on either findings or dismissals.285. **Deliver the verdict.** Lead with actionable findings ordered by impact, each29 with a source location, trigger, consequence, and evidence or verification gap.30 Explain material tradeoffs and disagreements. If there are no supported findings,31 say so and state the review's coverage limits without inventing issues.3233If the task requests independent reviewers and the harness supports them, give34reviewers the same intent and context before they see each other's conclusions.35Verify their reports yourself. Multiple opinions are optional evidence-gathering;36consensus is not proof, and sequential passes by one agent are not independent.
inprealpha