Interrogate

Interrogate

Spawn independent reviewers on mixed capability tiers to adversarially review a change. The signal comes from diverse priors, not assigned personas. Agreement is high-confidence. Lone findings are worth reading at lower weight.

The deliverable is a synthesized verdict. Do not auto-apply changes.

Companion to review-dispatch (front door), code-review (correctness gate), grok-review (one external engine), and full-code-review. Use this skill when the ask is a multi-reviewer adversarial pass.

Authorized Scope

Apply this engine only within the user's requested task and existing explicit authorization. Loading or delegating to it grants no additional authority. Preserve report-only restrictions and the caller's target, host, provider, and cost limits. Existing approval satisfies a gate only for the same actions and scope; obtain approval before expanding them. Forward these limits to delegates.

Contract

Inputs:

• A diff, branch, or PR to challenge

Outputs:

• A lead verdict bucketed Act on / Consider / Noted / Dismissed, plus an agreement map

Creates/Modifies:

• None. Report only.

External Side Effects:

• Read-only git / gh to gather the diff

Confirmation Required:

• None

Delegates To:

• Named, not fired: review-dispatch when the human wants the catalog review front door instead

Steps

1. Scope. Use the pointed files or diff. On a feature branch, git diff <trunk>...HEAD including the working tree.
2. Intent. Write one paragraph of what the code is trying to accomplish. Reviewers challenge whether the work achieves that intent, not whether the intent is correct. If intent is unclear, ask before spawning.
3. Spawn reviewers. Launch all reviewers in one message, read-only, across at least two capability tiers and, when possible, different families. Each gets the same filled template from references/reviewer-prompt.md, the rubric, and the code-quality lens.
4. Synthesize. Parse findings. Consensus (2+ independent) is highest signal. Deduplicate. Note disagreements.
5. Lead judgment. You are a pragmatic lead, not an aggregator. Read references/lead-judgment.md. Categorize every finding.

Output format

Intent

The stated intent paragraph.

Reviewers

One bullet per reviewer: label, capability tier, finding count. Never name a concrete model.

Act On / Consider / Noted / Dismissed

Each finding: description, which reviewers raised it, one-line rationale.

Agreement Map

Where reviewers agreed, where they diverged, and what that pattern means.

Interrogate procedure

Read interrogate procedure when running this workflow. Apply the authorized scope and mode of this entry point to every step. Resolve other skills through this distribution’s active catalog; resolve resources relative to the installed skill directory.