arcasilesgroup Review
Reviews code changes with human-quality judgment — PR reviews, file reviews, diff analysis, architecture feedback — dispatching a full specialist roster through macro-agents with aggressive false-positive control. Use it for "review this", "look over my PR", or "is this merge-ready".
Workflow
Principles applied: §10.7 Clean Code (readability, naming, single-responsibility); §10.4 DRY (flag duplication and missed reuse). High-signal review with full specialist coverage and aggressive false-positive control. This SKILL.md owns the user-facing contract; reviewer agent files provide specialist lenses and validation stages.
- Step 0 — load contexts — read
.ai-engineering/manifest.ymlproviders.stacks; apply.ai-engineering/overrides/<stack>/conventions.mdper stack. - Detect target — PR number, file paths, or current diff.
- Dependency preflight — verify
review-context.md,review-validator.md, plus the.claude/agents/reviewer-*.mdfiles the selected mode + diff scope need (frontendconditional on UI work — React, hooks, animation, typography, forms, a11y). STOP and report the exact missing path — never paraphrase reviewer instructions inline. - Pre-review — dispatch
review-context.mdvia the Agent tool; serialize its output for every specialist. - Specialists —
normal= 3 macro-agents;--full= one agent per specialist. Both run the full roster — grouping controls cost only. - Validate — dispatch
review-validator.mdwith YAML finding blocks only (no reasoning chain). Code is read fresh; verdict CONFIRMED or DISMISSED per finding. - Emit — Findings / Risks / Recommendations / Self-Challenge, attributed by original specialist lens.
Dispatch threshold
Dispatch the ai-review agent for any narrative review over ≥ 1 changed file (PR, branch, diff, or path scope). Each specialist runs in its own context window via the Agent tool. .claude/agents/ai-review.md is the orchestrator handle; profiles, roster, output contract, and validator stage live here.
Specialist Roster
Spec-140 W3 collapsed the roster from 11 specialists to 6. reviewer-architecture (DRY/reuse/proportionality) and reviewer-maintainability (readability/naming) are absorbed into reviewer-correctness. reviewer-backend was deleted (this repo is a Python CLI, no separate backend tier).
| Specialist | Agent File | Focus |
|---|---|---|
correctness |
reviewer-correctness.md |
logic bugs, null handling, races, edge cases + absorbed architecture (DRY/reuse) + maintainability (readability/naming) |
security |
reviewer-security.md |
vulnerabilities, auth, data exposure, dependency risk |
testing |
reviewer-testing.md |
coverage, quality, edge cases, mocking patterns |
performance |
reviewer-performance.md |
query shape, complexity, hot paths, memory |
frontend |
reviewer-frontend.md |
React, hooks, a11y, TypeScript, animation, typography, forms (conditional; absorbs legacy design lens, D-127-10) |
compatibility |
reviewer-compatibility.md |
breaking changes, backwards compat, migrations |
normal macro-agent grouping: (1) correctness + testing + compatibility, (2) security + performance, (3) frontend (conditional on UI diff; merges with macro-agent-2 when not dispatched).
Output Contract
Group findings by severity first, then original specialist lens (keep attribution even in normal). Include not_applicable / low_signal outcomes where a specialist had little to contribute. Show which findings survived adversarial validation.
Stack-specific review guidance
For each stack in the diff, load .ai-engineering/overrides/<stack>/review.md (D-133-10). Greenfield mode (stacks=[]): use generic review criteria + hint "add a project file and run ai-eng doctor --fix".
Common Mistakes
- Treating the 3 macro-agents in
normalas reduced coverage — they are not. - Reporting by macro-agent instead of original specialist lens.
- Treating style preferences as blocking findings.
- Reading specialist agent files inline instead of dispatching via the Agent tool.
Examples
Example — review a PR before approval
User: "review PR #42"
/ai-review 42
Dispatches the 3 macro-agents (correctness/testing/compat, security/perf, conditional frontend) over the diff, aggregates findings with corroboration, runs the validator stage, and emits the Findings table with severity + remediation.
Integration
Called by: user directly, /ai-pr, /ai-build, /ai-autopilot (Phase 5). Dispatches: review-context, reviewer-*, review-validator agents. Read-only: never modifies code. See also: /ai-verify (evidence-backed gates), /ai-learn (extract review patterns post-merge).
Inline fallback. Agent-tool dispatch is the primary path. On a harness without a subagent/Agent-tool primitive, execute this skill by reading the specialist agent file(s) (review-context.md, the selected reviewer-*.md, review-validator.md) inline and running their steps in-context, sequentially — inline-sequential is the floor, not a substitute for the dispatch when it is available.
$ARGUMENTS