requesting-review

Files (2)

• SKILL.md
• code-reviewer.md · references/

SKILL.md (first lines)

# Requesting Review

Dispatch an `arsyn:code-reviewer` subagent to catch defects before they compound. The reviewer receives crafted context — never your session history — keeping it focused on the work product.

## When to Dispatch

**Required:** after each orchestrated task, after completing a feature, before merging to main.

**Worthwhile:** before refactoring (baseline audit), after resolving a tricky defect, when you want a second perspective.

## Procedure

**1. Determine diff range:**
```bash
BASE_SHA=$(git merge-base origin/main HEAD)
HEAD_SHA=$(git rev-parse HEAD)
```

**2. Dispatch subagent** using Task tool with `arsyn:code-reviewer` type. Fill the template at `references/code-reviewer.md`.

Placeholders: `{WHAT_WAS_IMPLEMENTED}`, `{PLAN_OR_REQUIREMENTS}`, `{BASE_SHA}`, `{HEAD_SHA}`, `{DESCRIPTION}`.

**3. Triage response:**

| Severity | Action |
|----------|--------|
| Critical | Fix immediately — blocks all progress |
| Important | Fix before moving to next task |
| Minor | Log for later or fix opportunistically |

If the reviewer is wrong, push back with technical reasoning — reference tests or code that proves correctness.

## Scenario: Post-Task Review in a Multi-Task Plan

```
[Finished Task 3: pagination for search results]

BASE_SHA=$(git log --oneline --grep="Task 2" | head -1 | awk '{print $1}')
HEAD_SHA=$(git rev-parse HEAD)

[Dispatch arsyn:code-reviewer]
  WHAT_WAS_IMPLEMENTED: Cursor-based pagination for /search endpoint
  PLAN_OR_REQUIREMENTS: Task 3 from docs/plans/search-overhaul.md
  BASE_SHA: b4e22f1
  HEAD_SHA: 9a3c8d0
  DESCRIPTION: Added cursor encoding, page-size limits, next/prev links

[Reviewer returns]
  Critical: Off-by-one in cursor decode skips first result
  Minor: Magic number 50 for default page size

→ Fix cursor bug, extract PAGE_SIZE_DEFAULT constant, proceed to Task 4.
```

## Workflow Integration

- **Subagent-driven development:** review after each task.
- **Plan execution:** review after each batch (~3 tasks).
- **Standalone work:** review before merge or when stuck.

## Gotchas

- Skipping review on "trivial" changes — trivial changes hide subtle regressions.
- Using session HEAD without verifying — stale SHAs produce empty diffs. Always re-derive.
- Ignoring Important-severity items to stay on schedule — they compound into Critical by the next task.
- Passing session history instead of the template — reviewer loses focus, burns context on irrelevant detail.

From renathoaz/arsyn, MIT.

More from renathoaz/arsyn