Quality Review

Overview

Dispatches one parallel Sonnet agent per bullet point in the ## Implementation Quality section of the nearest AGENTS.md. Each agent independently discovers what changed and checks the codebase against exactly one criterion. A triage pass then removes false positives, duplicates, and contradictions before the final report.

Workflow

1. PARSE    → Extract each bullet point from ## Implementation Quality in AGENTS.md
2. DISPATCH → One Task agent per criterion (all in parallel, single message)
3. TRIAGE   → Review all raw findings: drop false positives, duplicates, and contradictions
4. REPORT   → Compile cleaned findings into summary table

Step 1: Parse Criteria from AGENTS.md

Read the project's AGENTS.md and extract every bullet point under ## Implementation Quality. Stop at the next ## heading — do not include bullets from any other section. Each - line becomes one criterion.

Do not hardcode criteria — always read from the current project's AGENTS.md so the skill stays in sync with the project's actual standards.

Step 2: Dispatch Parallel Agents

CRITICAL: All agents must be launched in a SINGLE message with multiple Task tool calls. Do not loop sequentially.

Use subagent_type: "Explore" and model: "sonnet" on every Task call.

Agent prompt template for each criterion (set model: "sonnet" on every Task call):

You are a focused code reviewer responsible for checking ONE specific quality criterion.

Criterion:
{criterion_text}

Your job:
1. Discover what changed — start with `git diff origin/main...HEAD` or `git diff main...HEAD`.
2. Feel free to explore the broader codebase or search the web for anything — do as much research as needed to make a confident judgment.
3. Check whether the changes comply with your assigned criterion.
4. Report ALL violations you find — do not stop at the first one.

Output format (exactly this structure, nothing else):
**{criterion_short_name}**: PASS | WARN | FAIL
- `file:line` — description of violation  (repeat for every violation found)
(omit bullet lines entirely if PASS)

Replace {criterion_short_name} with a 2-5 word label derived from the criterion.

Step 3: Triage Raw Findings

After all agents complete, review their combined output before building the report table. Apply each filter below and silently drop any finding that fails:

Filter	Rule
False positive	The flagged code is correct per the criterion when its full context is understood (e.g., a "one-time abstraction" that is actually reused elsewhere, framework-generated boilerplate the author didn't write, or a WARN that the criterion explicitly permits).
Duplicate	Two or more findings point to the same file+line for the same root cause, regardless of which criterion reported it. Keep only the most specific one.
Contradiction	A finding is itself in tension with another principle in AGENTS.md — e.g., flagging missing abstraction under DRY when adding it would violate YAGNI, or flagging missing error handling when the code correctly follows "fail fast." Drop the finding if following its recommendation would violate a different criterion.

After filtering, re-evaluate each criterion's overall verdict:

• If all its violations were dropped → change verdict to PASS.
• If only FAIL violations were dropped but WARNs remain → change verdict to WARN.

Do not modify the verdict of a finding you decide to keep.

Step 4: Compile Report

Output a table with one row per violation — if a criterion has multiple violations, give each its own row. Criteria with no violations get a single PASS row.

## Quality Review

| Criterion | Verdict | Finding |
|-----------|---------|---------|
| Engineered enough / YAGNI | WARN | `src/bar.ts:10` — abstraction added for single use case |
| Engineered enough / YAGNI | WARN | `src/baz.ts:88` — second violation of same criterion |
| DRY | PASS | |
| ... | | |

**Result: N criteria checked. X passed, Y warnings, Z failures.**

Each violation gets its own row, even when multiple violations share a criterion. List FAILs first, then WARNs, then PASSes so issues surface immediately.

If all pass: ✓ All N criteria passed.

Common Mistakes

Mistake	Fix
Launching agents sequentially	All Task calls in ONE message — that's the whole point
Hardcoding criteria	Always read from AGENTS.md — criteria drift over time
Passing the diff to agents	Don't — agents discover changes themselves via git
Agents checking multiple criteria	Each agent gets exactly one criterion
Skipping PASS rows in output	Include all criteria so nothing appears missed
Including TypeScript safety rules	Only parse ## Implementation Quality — stop at the next ## heading
Skipping triage	Always run Step 3 — subagents can't see each other's output and will produce overlapping findings
Dropping findings without justification	Each dropped finding must match a specific triage filter; do not drop findings just because they seem minor