Lazyllm Review

LazyLLM Review Skill

Installation (minimal)

Core package plus MCP (required for DeepWiki and other MCP-based tools):

pip install lazyllm "mcp>=1.7.0"

Or use the bundled extra:

pip install "lazyllm[agent-advanced]"

For GitHub App authentication (JWT):

pip install PyJWT cryptography

Note: After installing, check the version:

pip show lazyllm | grep Version

If the version is ≤ 0.7.6, the lazyllm review and lazyllm review-local commands are not available yet. In that case, clone the latest code directly from GitHub and add the project root to PYTHONPATH:

git clone https://github.com/LazyAGI/LazyLLM.git
export PYTHONPATH=/path/to/LazyLLM:$PYTHONPATH

Replace /path/to/LazyLLM with the actual clone path (e.g. ~/LazyLLM). No installation needed — the lazyllm review / lazyllm review-local commands will work immediately.

Model configuration

API Key lookup order

The review tool calls lazyllm.OnlineChatModule(source=..., model=...). Keys are resolved in this order:

1. --base-url flag paired with an environment variable (see below)
2. Environment variable LAZYLLM_<SOURCE>_API_KEY (e.g. LAZYLLM_CLAUDE_API_KEY, LAZYLLM_OPENAI_API_KEY, LAZYLLM_QWEN_API_KEY)
3. lazyllm config file field <source>_api_key

Recommended: set an environment variable

# Claude (default source)
export LAZYLLM_CLAUDE_API_KEY=sk-ant-xxx

# OpenAI
export LAZYLLM_OPENAI_API_KEY=sk-xxx

# Qwen
export LAZYLLM_QWEN_API_KEY=sk-xxx

# SiliconFlow / PPIO or any OpenAI-compatible proxy
export LAZYLLM_OPENAI_API_KEY=your-key

Specifying model and base-url

# Claude (default)
lazyllm review-local --source claude --model claude-opus-4-6

# OpenAI
lazyllm review-local --source openai --model gpt-4o

# Custom proxy / base-url
lazyllm review-local --source openai --model gpt-4o \
  --base-url https://your-proxy.com/v1/

Scenario 1: Review your own code (review-local)

See local-review.md for the full workflow.

Quick commands

# Diff against main, including uncommitted changes (default)
lazyllm review-local --base main --output review.json

# Committed changes only (exclude working-tree changes)
lazyllm review-local --base main --no-uncommitted --output review.json

# Specify repo path
lazyllm review-local --repo-path /path/to/repo --base develop --output review.json

How the AI tool should process review.json

1. Read review.json. Structure:

   • instructions: embedded AI processing guide — follow it directly
   • pr_info: branch info and PR summary
   • stats: issue counts (total / critical / normal / suggestion)
   • issues[]: each issue has id, file, line, severity, category, problem, suggestion

2. Process by severity: critical → normal → suggestion

3. For each issue:

   • Judge Reliable (valid concern) vs Not Reliable (misunderstanding / already fixed / invalid constraint)
   • Reliable → apply a fix at the indicated file + line
   • Not Reliable → record rejection reason, skip

4. Generate a final report using the template below

Final report template

# Review Fix Report

## Summary
- Total issues: N, Fixed: X, Rejected: Y

## Fixed issues
| ID | File | Line | Severity | Fix description |
|----|------|------|----------|-----------------|
| 1  | foo.py | 42 | critical | Added None guard to prevent null dereference |

## Rejected issues (Not Reliable)
| ID | Reason |
|----|--------|
| 3  | Already fixed in the previous commit |

## Unresolved issues
(If any, explain why)

Scenario 2: Review someone else's PR (review)

See pr-review.md for the full workflow.

GitHub authentication

The tool resolves a token in this order:

1. Environment variable GITHUB_TOKEN or GH_TOKEN
2. gh CLI already logged in (gh auth token)
3. GitHub App (app_id + installation_id + PEM private key)

Recommended: gh CLI

gh auth login          # one-time browser auth
lazyllm review --pr 42 --repo owner/repo

Or set a token directly

export GITHUB_TOKEN=ghp_xxx
lazyllm review --pr 42 --repo owner/repo

Post review comments to GitHub

Add --post to publish comments back to the PR:

lazyllm review --pr 42 --repo owner/repo --post

Without --post, results are printed locally only.

Common flags

Resuming from a checkpoint

# Restart from r1 (keeps clone/arch/spec cache)
lazyllm review --pr 42 --repo owner/repo --resume-from r1

# Full restart
lazyllm review --pr 42 --repo owner/repo --clear-checkpoint

Stage order: clone → arch → spec → pr_summary → r1 → r2 → r3 → final → upload

Scenario 3: Fix bugs from PR review comments (fix-bug)

See fix_bug.md for the full workflow.

Use this scenario when the user says: "帮我修复 review 评论", "fix review comments", "处理 PR 上的 review 意见", "根据 review 修 bug", "fix bug from review", "回复 review 评论"。

Quick steps

# 1. Fetch all review comments (line-level)
gh api repos/{OWNER}/{REPO}/pulls/{PR}/comments

# 2. Fetch overall reviews
gh api repos/{OWNER}/{REPO}/pulls/{PR}/reviews

# 3. Reply to a comment after fixing
gh api --method POST \
  repos/{OWNER}/{REPO}/pulls/{PR}/comments \
  --field in_reply_to={comment_id} \
  --field body="已修复：{简要说明}"

AI processing rules

1. Must use gh CLI to fetch comments — never browse the PR page directly (collapsed comments will be missed)
2. For every comment, judge Reliable (fix it) vs Not Reliable (reject with reason)
3. Apply code fixes at the indicated file + line
4. Reply to every comment (fixed or rejected) via gh api
5. Generate a final fix report grouped by category (correctness / performance / architecture / style / maintainability)

Other platforms

# GitLab
export GITLAB_TOKEN=glpat-xxx
lazyllm review --pr 10 --repo owner/repo --backend gitlab

# Gitee
export GITEE_TOKEN=xxx
lazyllm review --pr 5 --repo owner/repo --backend gitee

Reference docs

• Local review full workflow: local-review.md
• PR review full workflow: pr-review.md
• Fix bugs from PR review comments: fix_bug.md