Agentic Code Reviewer

Agentic Code Review

This skill launches a fully backgrounded code review of the current git diff: 5 specialist reviewer processes in parallel, a synthesizer judge pass, and a local review UI for triaging findings. Runtime review execution is owned by the process-based orchestrator in scripts/; the rest of this document is the architecture reference for that runtime.

Runtime Model

Codex should treat requests like run code-review on this repo, run the code-reviewer skill, or run the agentic-code-reviewer skill as requests to launch this skill.

/code-review runs:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/orchestrator.sh" --repo "$(pwd)"

/code-review and /pr-review run the same launcher with --pr "$ARGUMENTS". Requests like code-review PR #42 or review PR 42 with the agentic code reviewer should launch this skill's PR mode — results open in the review UI, never inline in the terminal. (Note: Anthropic's separate official code-review plugin also reviews PRs, inline with a confidence threshold; that is a different tool, not this skill.)

On Codex, CLAUDE_PLUGIN_ROOT is not set. Resolve the installed skill root before launching:

SKILL_ROOT="${CLAUDE_PLUGIN_ROOT:-$HOME/.codex/skills/agentic-code-reviewer}"
bash "${SKILL_ROOT}/scripts/orchestrator.sh" --repo "$(pwd)"

The launcher validates required tools, creates .claude/review-runs//, starts scripts/orchestrator.py with nohup, then prints a compact status line every 20 seconds until the review UI is ready. Set ACR_STATUS_POLL=0 to return immediately without polling.

In Claude Code and in Codex installs with hooks enabled, the Stop hook is controlled by stopHookMode in ~/.claude/agentic-code-reviewer/settings.json. The default is prompt, including upgraded installs whose older settings files do not have this key. Claude Code reads the hook from the plugin manifest; Codex reads the user-level entry installed in ~/.codex/hooks.json. hooks/code-review-gate.sh delegates to scripts/review-gate.py, which uses hook JSON cwd when present, hashes the current reviewable diff, and then follows the resolved mode: prompt stops the turn and waits for the user's next yes/no/skip reply, disabled exits silently, and auto launches or reuses a fast review with a 3 minute hook budget. Codex may require reviewing or trusting the hooks through /hooks.

Claude Code named subagents and Codex spawn_agent are not used as the review execution primitive. The five reviewers run as independent provider subprocesses through scripts/run-reviewer.sh, one output file per reviewer:

• agents/semantic-analyzer.json
• agents/security-scanner.json
• agents/architecture-reviewer.json
• agents/test-coverage-analyzer.json
• agents/senior-dev-reviewer.json

The synthesizer runs only after all five reviewer result files exist and parse. It writes synthesis.json, and the Bun review UI is launched with --run-dir .

Runtime provider selection is automatic:

• Claude Code uses claude --print --output-format json
• Codex uses codex exec --json --sandbox read-only

Override with ACR_PLATFORM, ACR_REVIEW_PROVIDER, ACR_CLAUDE_BIN, ACR_CODEX_BIN, ACR_MODEL_BALANCED, ACR_MODEL_FAST, ACR_MODEL_JUDGE, or ACR_CODEX_REASONING_*.

Run Directory

Each run writes:

.claude/review-runs//
  run.json
  context.json
  diff.txt
  orchestrator.log
  review-gate.json
  READY
  ui.pid
  prompts/*.prompt.md
  agents/*.json
  synthesis.json
  decisions.json

Reviewer failures, timeouts, and invalid JSON are recorded as explicit status: "failed" reviewer JSON files so synthesis can still proceed with partial evidence.

Continuation

When auto mode launches or watches a run, server auto-resume is disabled and the hook owns continuation. Implement / Done / confirmed Dismiss write final decisions; Save decisions is non-final. If any decision is actionable and the diff hash still matches, the hook wakes the host agent with instructions to run the resume reader and act on the result. If the diff changed while review was running, the hook marks the run stale and allows the session to finish without using outdated findings. A user-confirmed prompt launch behaves like a manual launch and leaves server auto-resume enabled when host session metadata is available.

Manual fallback:

/review-resume 

The resume command reads synthesis.json and decisions.json, validates the decision actions, and prints deterministic implementation instructions. The host agent then implements only findings marked ask_claude_to_implement or accept_fix, skips ignore, answers ask_claude_to_explain, and records create_follow_up_task items for the user.

Decision Actions

Each finding has exactly one action:

• accept_fix
• ignore
• create_follow_up_task
• ask_claude_to_explain
• ask_claude_to_implement

Per-finding comments, line annotations, and a global comment are preserved in decisions.json.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: putchi
• Source: putchi/agentic-code-reviewer-skill
• License: MIT

Install and usage instructions live in the source repository linked above.