Aireview

aireview - Codex Native Code Review

Positioning

aireview is the advanced review layer. Use normal code-review for fast baseline findings, then escalate to aireview when the diff is large, sensitive, cross-cutting, security-relevant, or release-blocking.

All review output should be in Chinese unless the user asks otherwise.

Native Subagent Protocol

Codex supports read-only child agents. Delegate with spawn_agent, collect with wait_agent, and clean up with close_agent.

spawn_agent -> wait_agent -> close_agent

No external model CLI is required. Deep mode is implemented with Codex child agents.

Quick Usage

aireview --diff
aireview --diff --staged
aireview HEAD~1..HEAD
aireview origin/feature-branch
aireview origin/feature-branch --quick
aireview origin/feature-branch --deep
aireview --pr 123

Review Modes

Quick

Use for small diffs. Main Codex performs a focused review:

• obvious correctness bugs
• security issues
• breaking API or compatibility changes
• missing critical tests

Standard

Use for ordinary PRs. Dispatch read-only perspectives:

Deep

Use for large or high-risk changes. Run multiple focused review passes:

1. explorer: trace architecture and integration impact
2. reviewer: lead with concrete bugs and risks
3. docs-researcher: verify version-sensitive claims
4. Main Codex: deduplicate, score confidence, and produce final findings

Workflow

Step 1: Gather Context

Run:

git status --short
git diff --stat
git diff
git diff --staged

For branch reviews:

git fetch --all --prune
git diff --stat main...<branch>
git diff main...<branch>

Read applicable repo instructions and changed files before judging.

Step 2: Dispatch Review Agents

Example prompts:

spawn_agent(agent_type="explorer", message="
Read-only. Trace the changed execution paths and cite files/symbols.
Focus on architecture impact, coupling, and hidden integration points.
")

spawn_agent(agent_type="reviewer", message="
Read-only. Review like an owner.
Lead with correctness, security, regressions, and missing tests.
Return only concrete findings with file/line references when possible.
")

spawn_agent(agent_type="docs-researcher", message="
Read-only. Verify any framework/API/version-sensitive claims against primary docs
or local package/config versions. Cite sources.
")

Step 3: Score Findings

For each finding:

• Severity: P0/P1/P2/P3
• Confidence: 0-100
• Evidence: file/line, command, test, or doc source
• Fix direction: concise and actionable

Filter out findings below 80 confidence unless the risk is severe and clearly labeled uncertain.

Step 4: Final Report

Lead with findings:

## Findings

1. [P1] Title
   - File: `path:line`
   - Evidence: ...
   - Impact: ...
   - Fix: ...
   - Confidence: 90

## Open Questions

## Test Gaps

## Summary

Guardrails

• Findings first; summary later.
• No style-only comments unless they hide a real bug.
• Do not rewrite code during review unless the user explicitly asks for fixes.
• Do not rely on agent agreement as evidence; cite files, commands, or primary docs.

Fallback

If child agents are unavailable, produce the same review in one response using:

[EXPLORER]
[REVIEWER]
[DOCS]
[SYNTHESIS]