unity-review-code-pr

Fetch a GitHub PR diff, assess change size, then review adaptively — quick single-pass for minor PRs, parallel subagents for large ones. Post inline comments via GitHub API. Never stop until the review is confirmed posted on GitHub.

Workflow

1. Fetch PR data — follow references/pr-review-workflow.md steps 1-4
2. Assess PR size — see references/size-assessment.md for classification rules
3. Route by size:
   • Minor → step 4a (quick review)
   • Large → step 4b (deep review)
     4a. Quick review — review all changed files yourself in a single pass; load unity-standards checklists, check all 6 criteria inline; produce findings as [{path, line, severity, title, body}]
     4b. Deep review — spawn 7 parallel subagents per unity-standards/references/review/parallel-review-criteria.md; collect via background_output; aggregate (deduplicate by path+line, keep highest severity, sort file→line)
4. Map lines — use right-side file line numbers (step 5 from references/pr-review-workflow.md)
5. Build + Submit — build review JSON per unity-standards/references/review/pr-submission.md, submit via gh api
6. Verify — confirm review posted (step 7 from references/pr-review-workflow.md), retry on failure until confirmed

Inline Comment Format

Every finding MUST use this template. The suggestion block is required for MEDIUM severity and above — show the corrected code, even if multi-line.

Template:

**{🔴|🟠|🟡|🔵|⚪} Title** — `SEVERITY`

{1-3 lines: what is wrong, why it matters}

```suggestion
{corrected code replacing the commented line(s)}
```

Critical/High (🔴🟠) — include cause, impact, and suggestion block.
Medium (🟡) — problem + suggestion block.
Low/Style (🔵⚪) — compact; suggestion encouraged but optional.

Severity scale: 🔴 CRITICAL → 🟠 HIGH → 🟡 MEDIUM → 🔵 LOW → ⚪ STYLE

Severity Calibration

Minimum floors — do not rate below these:

• GetComponent/Find* called every frame (Update/FixedUpdate/LateUpdate) → MEDIUM
• Missing event unsubscription (+= without matching -= in OnDestroy/OnDisable) → HIGH
• Coroutine not stopped on lifecycle exit (OnDisable/OnDestroy) → HIGH
• async void on any method (swallows exceptions) → MEDIUM
• Exposed internal collection (returning List<T> instead of IReadOnlyList<T>) → MEDIUM
• Resources.Load<T>() called at runtime outside initialization (Awake/Start/constructor) → MEDIUM

Rules

• Always fetch full file content — diff hunks lack context for lifecycle and event cleanup analysis
• Prioritize finding coverage: identify all issues first, then format with suggestion blocks. Do not let formatting effort reduce the number of issues found.
• Every finding at MEDIUM severity or above must include a suggestion code block
• Do not post duplicate comments for the same line
• See unity-standards/references/review/pr-submission.md for review body format, inline comment examples, and line/side/batching rules

Reference Files

• references/pr-review-workflow.md — PR fetch, line mapping, submit, verify
• references/size-assessment.md — PR size classification and routing rules

Load on demand via read_skill_file("unity-review-code-pr", "references/{file}").

Standards

Load unity-standards for all review criteria. Load on demand via read_skill_file("unity-standards", "references/review/<file>"):

• review/logic-checklist.md — correctness, edge cases, state, data flow
• review/unity-lifecycle-risks.md — order-of-execution, null timing, scene load, cleanup pairs
• review/serialization-risks.md — missing fields, type changes, prefab overrides
• review/performance-checklist.md — allocations, Update, physics, rendering
• review/architecture-checklist.md — coupling, SOLID, assembly boundaries, event coupling
• review/concurrency-checklist.md — threading, race conditions, async/await, main thread rule
• review/security-checklist.md — input validation, injection, data exposure
• review/pr-submission.md — output template: JSON payload, event decision, batching, gh CLI
• review/parallel-review-criteria.md — subagent delegation, criteria table, prompt template