🤖 AI Code Review System

GitHub Actions based AI code review, version 1.0.0. The default setup uses OpenAI gpt-5.6-luna, low reasoning effort, Flex Processing, and a concise output contract that only reports actionable issues backed by diff evidence.

Quick Start

1. Put this project in your repository.
2. Add GitHub Actions secrets:
   • GH_TOKEN: GitHub PAT with repository read access and issue write access.
   • OPENAI_KEY: OpenAI API key.
3. Enable GitHub Actions.
4. Push code. Review results are created as Issues with the ai-code-review label.

OPENAI_BASE_URL is optional and defaults to https://api.openai.com/v1. Set it only for self-hosted or OpenAI-compatible providers.

Default Behavior

• Push workflow reviews commits in the current repository.
• Scheduled workflow scans every repository accessible by GH_TOKEN.
• No default_repo setting is needed.
• No repository list is needed; unspecified means monitor all accessible repositories.
• Set projects.enabled_repos only when you want an allowlist.
• If no actionable issue is found, the output is exactly: 未發現需要修改的問題。

Minimal Config

config.json can be as small as:

{
  "review": {
    "response_language": "zh-TW"
  }
}

You can also remove config.json; built-in defaults will be used.

Built-In Model Defaults

{
  "model": {
    "name": "gpt-5.6-luna",
    "fallback_models": [],
    "api_mode": "responses",
    "reasoning_effort": "low",
    "verbosity": "low",
    "service_tier": "flex",
    "flex_fallback_to_auto": true,
    "max_retries": 3,
    "retry_backoff_seconds": 1.0,
    "max_tokens": 16384,
    "timeout": 900
  },
  "projects": {
    "enabled_repos": ["*"]
  }
}

temperature is intentionally not part of the default config. GPT-5.6 review behavior is controlled with reasoning_effort, verbosity, and a strict output contract.

Flex Processing

service_tier defaults to flex, which is a good fit for asynchronous GitHub Actions review: lower cost, slower responses, and occasional resource unavailability. By default a Flex resource-unavailable response is retried with service_tier: "auto"; disable this with flex_fallback_to_auto: false when cost is more important than completion. To always use standard project routing, use:

{
  "model": {
    "service_tier": "auto"
  }
}

Output Language

review.response_language defaults to zh-TW. Any structurally valid BCP 47 language tag is accepted, including en-US, zh-Hant-TW, sr-Latn-RS, and es-419. Code, identifiers, paths, and severity labels remain unchanged.

Repository Allowlist

Only set this when you do not want all accessible repositories monitored:

{
  "projects": {
    "enabled_repos": [
      "owner/repo1",
      "owner/repo2"
    ]
  }
}

Review Output

Each actionable finding includes:

• 位置: file and line/hunk
• 證據: concrete diff evidence
• 影響: concrete failure mode
• 修法: exact implementation direction
• 驗證: command or test case
• AI_AGENT_FIX_PROMPT: direct repair instruction for an AI coding agent

If there is no issue, the review does not invent one.

Local Validation

python3 scripts/test_config.py
python3 -m unittest discover -s tests -v
python3 scripts/benchmark_review.py

The workflows use current actions/checkout@v7 and actions/setup-python@v7, upgrade pip, and install the latest compatible Requests 2.x release. See PERFORMANCE.md for the reproducible before/after benchmark and bottleneck analysis.

Advanced Config

See CONFIG.md.