| name | interrogate |
| description | Adversarial multi-reviewer synthesis for contested or high-risk plans and changes. Use for multi-model review, independent angles, challenge, stress-test, or blind-spot requests. Do not use for routine single-reviewer checks. |
Interrogate
Spawn independent reviewers across diverse model families to adversarially
review code changes. Each model receives the same intent, diff, rubric, and
code-quality lens. Agreement across models is high-confidence signal; lone-model
findings provide exploratory context.
The deliverable is a synthesized lead verdict. Do NOT auto-apply changes.
Workflow
1. Determine Scope
Identify target changes from context:
- Explicit file paths or diffs.
- Feature branches:
git diff main...HEAD (or target base branch).
- Recent workspace edits and surrounding context files.
2. State Intent
Derive what the change accomplishes from user input, commit history, and PR
descriptions. Write one clear paragraph. Reviewers challenge whether the code
achieves this intent well, not the intent itself.
3. Spawn Reviewers
Launch parallel reviewers using distinct model families or subagents.
Each reviewer receives:
- Stated intent.
- Changeset diff and surrounding context.
- Review rubric from rubric.md.
- Code-quality lens from
code-quality-review.md.
- Prompt template from reviewer-prompt.md.
4. Synthesize Findings
- Collect structured findings from all reviewers.
- Identify consensus findings (raised by 2+ models independently).
- Identify lone-model findings and assess confidence.
- Deduplicate overlapping issues across models.
- Note explicit disagreements between reviewers.
5. Apply Lead Judgment
Act as lead reviewer using lead-judgment.md:
- Act on: Core issues affecting correctness, security, or maintainability.
- Consider: Valid tradeoffs worth user attention.
- Noted: Low-priority or premature concerns.
- Dismissed: False positives or nitpicks lacking context.
Output Format
Present the synthesized verdict:
- Intent: Stated goal of the changes.
- Reviewers: Model names and findings count.
- Act On: Critical issues with model sources and impact rationale.
- Consider: Tradeoffs and potential improvements.
- Noted: Low-priority observations.
- Dismissed: Filtered findings with rejection reasons.
- Agreement Map: Consensus and divergence summary across models.
