review-lens

name	review-lens
allowed-tools	Bash(git:*) Bash(gh:*) Bash(find:*) Read Write Edit Task
description	Code review in quick and deep modes, with confidence-scored findings. Use for diff review, guideline audits, re-reviewing fixes, or applying findings before a pull request. Not for acceptance-criteria verification, visual design review, or commit and branch mechanics.

Review Lens

Code review with anti-hallucination diff annotation and confidence-scored findings, in two modes — a fast walkthrough-plus-findings pass by default, or a multi-material fan-out on demand. Runs before a pull request: the report goes to the chat, with optional fix application and a saved report.

Triggers

• Quick review (default — "review", "review my changes", "check my diff", "review against main") → quick-review.md
• Deep review ("deep review", "full review", "thorough review") → deep-review.md
• Re-review ("re-review", "check fixes", "are the issues resolved") → the Re-Review section of the mode that ran (default quick)

Shared rules live in common.md; guideline discovery lives in guidelines-audit.md. Neither is a direct trigger.

Modes

Both modes share the same [L<n>] diff annotation, the confidence ≥ 80 bar, the what-not-to-report rules, and the output template (all in common.md). They differ in depth and cost:

• Quick (default) — two agents run in parallel by role: a Haiku walkthrough describes the change, a Sonnet findings pass catches issues across every scope. Fast, with reasoning where it counts.
• Deep — fans out by material (diff, guideline files, git history, prior PRs) with an independent confidence judge. Thorough, for risky or wide-reaching changes.

Default to quick. Reach for deep only when the user asks for depth or the change is risky.

Deep Mode: Material Fan-Out

The fan-out divides by source of material, not by concern — each agent reads something the others don't, so each adds new signal. The concerns (security, bugs, data-loss, performance) are a checklist inside one bug-scan, not separate agents. Sonnet is used only for the two reasoning passes; everything else is Haiku.

Agent	Model	Reads
setup	inline	annotate [L<n>], size gate, gather guideline paths, walkthrough
bug-scan	Sonnet	the diff — security/bugs/data-loss/performance checklist
compliance	Haiku	the project's guideline files
git-history	Sonnet	git log / blame of the changed files
prior-PRs	Haiku	past PRs + review comments (when gh + GitHub remote)
judge	Haiku	every finding — scores confidence, drops < 80

Guidelines

• Annotate the diff with [L<n>] markers before reviewing — the line allowlist is the anti-hallucination guard in both modes
• Only report findings with confidence ≥ 80
• Default to quick; reserve the deep fan-out for risky or wide diffs
• Guideline discovery reads the project's files — including .claude/rules/*.md — never ~/.claude (personal global settings)
• Suggest fixes freely (they are text); apply to the working tree only with explicit confirmation
• The review runs pre-PR — output goes to the chat (and optional CODE_REVIEW.md), never posted to a pull request

Anti-Pattern: Confidence Inflation

Reporting findings below 80 confidence buries real issues under noise. The rubric is calibrated: <80 means speculation or style preference, not a real bug. When unsure, drop down — gather more context, re-read the diff — instead of pushing a low-confidence finding through.