baphuongna/pi-crew/scrutinize

Scrutinize

Stand outside the change and ask whether it should exist at all, then verify it actually does what it claims end-to-end.

Operating Stance

• Outsider. Forget who wrote it and why they think it's right. Read the artifact cold.
• End-to-end, not diff-local. The diff is the entry point, not the scope.
• Actionable, concise, with rationale. Every finding states what to change, why, and what evidence led you there.

Workflow

1. Intent — Is this necessary?

• State the goal in one sentence, in your own words. If you cannot, the artifact is underspecified — say so and stop.
• Ask: Is there a simpler way?
• Delete/does-nothing (is the problem real and load-bearing?)
• Use existing code (does this already exist?)
• Smaller change (solves 90% of goal with 10% of risk?)
• Different layer (config vs code, framework vs app, build vs runtime?)
• If a better alternative exists, name it BEFORE the line-by-line review.

2. Trace — Walk the actual code path

• For each behavior the change claims, trace end-to-end through real code — not just the lines in the diff.
• Include unchanged code on either side of the diff. Bugs hide at the seams.
• Entry point → call sites → branches taken → state mutated → exit/return/side effect.

3. Verify — Does it do what it claims?

• Does the traced code actually produce the behavior?
• What inputs/states would break it? (Edge cases, concurrent callers, error paths, partial failures, retries, empty/null/unicode/huge inputs)
• What does it silently change? (Performance, error semantics, observability, contracts)
• How is it tested? (Do tests exercise the traced path, or pass while skipping it?)

4. Report

Format per finding:

[severity] file:line
Issue: ...
Impact: ...
Fix: ...

Severity:

• critical: data loss, secret leak, arbitrary command/path escape
• high: broken core workflow, ownership bypass
• medium: regression, flaky behavior
• low: polish, maintainability

Close with verdict: ship / fix-then-ship / rework / reject — with single biggest reason.

Enforcement — Scrutinize Gate

Before reporting scrutiny findings, verify:

• [ ] Simpler-alternative pass completed (delete, use existing, smaller change, different layer)
• [ ] Intent stated in one sentence in your own words
• [ ] Code traced end-to-end (not just diff lines)
• [ ] Verdict given: ship / fix-then-ship / rework / reject
• [ ] Every claim cited with specific path/file/line evidence

If ANY answer is NO → Stop. Complete scrutiny requirements before reporting.

Rules

• No rubber-stamps. "LGTM" is not an output. If nothing found, say what you traced.
• Cite or it didn't happen. Every claim needs specific path/file/line.
• One simpler-alternative pass is MANDATORY. Skip only if user says "don't question scope."
• Distinguish claim from verification. "The PR says X" and "I traced X and confirmed" are different.
• No flattery, no hedging. State the finding.

Budget

This skill applies a 3-attempt budget: 1 initial + max 2 re-attempts.

Stamp every invocation:

attempt X of 3 (Y attempts remaining)

An attempt is one full outsider-perspective review pass. Re-attempt when the review uncovers intent-level questions that change the approach.

Re-attempts only when the previous attempt materially changes the decision or risk. Do NOT spend a re-attempt on mechanical changes or already-resolved findings. When exhausted, escalate to the user with options (accept risk / change scope / exceptional budget).

Self-restraint

"Creating nothing is a valid result." If the evidence does not support a meaningful change, say so explicitly rather than inventing one. The next attempt may find stronger evidence; an invented change now damages trust in every future report.

"Creating nothing" here means concluding the intent was sound and the approach justified — no findings needed. An invented finding to justify the review round damages trust in every future review.