Agentic Code Reviewer

A portable skill that runs 5 specialized review agents in parallel over your git diff, then an Opus Chairman pass that dedupes findings, re-rates them by evidence, and writes a verdict. Works on Claude Code, Codex, and Copilot CLI.

What it does

Five specialist reviewers — semantic, security, architecture, test-coverage, senior-dev — each look at the same git diff from their own angle and report only high-confidence findings (≥80). A final Chairman/Judge pass on Opus takes all five raw outputs, drops anything that can't cite concrete code, merges duplicates, resolves contradictions, re-rates severity by actual blast radius, and writes a 2-sentence top-line verdict.

On Claude Code there's an extra layer: a Stop hook that blocks the session from ending until the review has actually run (with a one-time escape hatch). On Codex and Copilot CLI you invoke the skill manually — see Claude Code exclusive: session-exit gate below.

The review council

Agent	Focus	Model
semantic-analyzer	Logic correctness, control flow, null handling, off-by-one, race conditions	Opus
security-scanner	OWASP Top 10, injection, secrets in code, auth/authorization gaps	Sonnet
architecture-reviewer	Module boundaries, system-level SOLID, missing abstractions, YAGNI	Opus
test-coverage-analyzer	Behavioral test gaps, missing edge cases, untested error paths	Sonnet
senior-dev-reviewer	Local DRY, naming, error handling, project conventions, dead code	Opus
chairman (judge)	Dedupe, drop no-evidence findings, resolve contradictions, re-rate severity, write verdict	Opus

How it works

1. Pre-flight. Guard for being inside a git repo. Run git diff --text HEAD (then plain git diff as fallback) and filter out lockfiles, minified assets, images, and build directories. If the filtered diff is empty, exit cleanly with No reviewable changes. If the diff exceeds 2000 lines or 50 files, print a cost/time warning.
2. Fan out. Dispatch all 5 reviewer agents in parallel using your platform's subagent tool (see the platform matrix). Each agent gets the same filtered diff and a strict prompt: report only findings with confidence ≥80, formatted as [SEVERITY] file:line — finding — reasoning.
3. Chairman pass. Send the diff + all 5 raw outputs to the chairman agent. It applies judge rules (evidence required, semantic dedupe, contradiction resolution, severity re-rated by blast radius, drop speculation) and emits the final report.
4. Print verdict. The Chairman's output is the final report — Verdict / CRITICAL / HIGH / NOTES / Summary. No further aggregation.
5. Completion signal. Touch /tmp/claude-code-review-${SESSION_ID}.done and emit <!-- AGENTIC-REVIEW-COMPLETE -->. On Claude Code these unblock the Stop hook; on other platforms they are harmless no-ops.

Platform support matrix

Feature	Claude Code	Codex	Copilot CLI
Parallel subagent fan-out	✅	✅ (requires multi_agent = true)	✅
Skill invocation	/agentic-code-reviewer + auto-trigger	skills load natively	skill tool
Session-exit auto-gate	✅	❌	❌
Subagent dispatch tool	Agent / Task	spawn_agent	task agent_type: general-purpose

Full tool-name mapping is in references/platform-tools.md.

Installation

Claude Code (primary)

Option A — one-line install (recommended):

curl -fsSL https://raw.githubusercontent.com/putchi/agentic-code-reviewer-skill/main/install.sh | bash

Or with an explicit platform flag to skip the prompt:

curl -fsSL https://raw.githubusercontent.com/putchi/agentic-code-reviewer-skill/main/install.sh | bash -s -- --platform claude

Option B — install from the published repo:

The repo is a self-hosting marketplace, so register it first, then install the plugin from it:

/plugin marketplace add putchi/agentic-code-reviewer-skill
/plugin install agentic-code-reviewer@agentic-code-reviewer-skill

Option C — local clone (for local development):

git clone git@github.com-secondary:putchi/agentic-code-reviewer-skill.git
cd agentic-code-reviewer-skill
./install.sh --platform claude

Verify with /plugin and confirm agentic-code-reviewer is listed.

Required tools: git, python3 (used by the Stop hook to parse hook JSON), bash.

Codex (CLI + App)

Option A — one-line install (recommended):

curl -fsSL https://raw.githubusercontent.com/putchi/agentic-code-reviewer-skill/main/install.sh | bash -s -- --platform codex

Or install for both Claude Code and Codex at once:

curl -fsSL https://raw.githubusercontent.com/putchi/agentic-code-reviewer-skill/main/install.sh | bash -s -- --platform both

Option B — from a local clone:

git clone git@github.com-secondary:putchi/agentic-code-reviewer-skill.git
cd agentic-code-reviewer-skill
./install.sh --platform codex

After installing, enable multi-agent support — add this to ~/.codex/config.toml:

[features]
multi_agent = true

Without this, parallel subagent dispatch will fail.

Copilot CLI

Manual install: clone the repo and copy skills/agentic-code-reviewer/, agents/, and references/ into your Copilot CLI skills directory.

git clone git@github.com-secondary:putchi/agentic-code-reviewer-skill.git

Usage

• Claude Code: run /agentic-code-reviewer, or let the Stop-hook gate prompt the agent to run it for you when you try to end a session with unreviewed changes.
• Codex: skills load natively — tell the agent run the agentic-code-reviewer skill.
• Copilot CLI: invoke the skill via the skill tool: skill agentic-code-reviewer.

On every platform, empty diffs exit cleanly (No reviewable changes — nothing to review.) and diffs above 2000 lines / 50 files print a cost and timing warning before fanning out.

Output format

The Chairman emits a fixed structure. Severity buckets stay in the report even if empty (they read _None._).

## Code Review Results

### Verdict
[Two sentences. First: ship-readiness assessment. Second: the single most important issue.]

### CRITICAL
- [CRITICAL] file:line — finding — reasoning — EVIDENCE: `<code excerpt>` (dim: semantic, security, ...)

### HIGH
- [HIGH] file:line — finding — reasoning — EVIDENCE: `<code excerpt>` (dim: architecture, ...)

### NOTES
- [NOTE] file:line — finding — reasoning — EVIDENCE: `<code excerpt>` (dim: senior-dev)

### Summary
- X critical, Y high, Z notes retained; W findings dropped (D no-evidence, M merged, C contradictions resolved).

Claude Code exclusive: session-exit gate

This part runs only on Claude Code; it has no Codex or Copilot CLI equivalent.

The Stop hook (hooks/code-review-gate.sh) does the following on every Stop event:

• Runs git diff HEAD (then git diff) to see if any code was changed this session.
• If there are changes and /tmp/claude-code-review-${SESSION_ID}.done is missing, returns {"decision":"block"} with a system message instructing the model to run the skill. It also touches /tmp/claude-code-review-${SESSION_ID}.blocked.
• The blocked sentinel is the one-time escape hatch: if it already exists, the next Stop is allowed through. You are never held hostage — you can always end the session by trying twice.
• Stale .done and .blocked sentinels older than 1 day are auto-cleaned on every invocation.
• Backward-compat fallback: if for some reason the .done file can't be written, the hook still honors a transcript grep for AGENTIC-REVIEW-COMPLETE.

On Codex and Copilot CLI you are responsible for invoking the skill before ending the session.

What it does NOT do

• Does not auto-fix code — it only reports findings.
• Does not block commits or pushes — only gates the Stop event in the current Claude Code session.
• Does not review binary, lockfile, or build-artifact diffs (filtered out before fan-out).
• Does not report findings below 80% confidence.

Costs and timing

For a large diff (>2000 lines or >50 files) the run fans out to 5 reviewers + 1 Chairman pass and takes roughly 30–90 seconds at an estimated ~$0.30–$0.80. Smaller diffs are proportionally cheaper. The Chairman runs on Opus; the reviewers are a mix of Opus and Sonnet (see the review council).

Project layout

.
├── .claude-plugin/plugin.json     # Claude Code manifest
├── agents/                        # 5 reviewers + chairman (prompts are portable)
│   ├── semantic-analyzer.md
│   ├── security-scanner.md
│   ├── architecture-reviewer.md
│   ├── test-coverage-analyzer.md
│   ├── senior-dev-reviewer.md
│   └── chairman.md
├── commands/agentic-code-reviewer.md   # Claude Code slash command
├── hooks/                              # Claude Code exclusive (Stop-event gate)
│   ├── hooks.json
│   └── code-review-gate.sh
├── references/
│   └── platform-tools.md          # Claude Code / Codex / Copilot CLI mapping
└── skills/agentic-code-reviewer/SKILL.md

License

MIT — see LICENSE.