claude-review.yml

# Claude Code Review - Automated PR Reviews
#
# Triggers on every PR to main.
# Uses Anthropic API via custom proxy (ANTHROPIC_BASE_URL).
#
# Features:
# - Smart skip: rebase, WIP, trivial changes, version bumps
# - Comment dedup: avoids repeating bot/human comments on synchronize
# - Auto-resolve: resolves bot comments that have been addressed by new pushes
#
# Prerequisites:
# 1. Install GitHub App: https://github.com/apps/claude (or create custom app)
# 2. Add repo secrets:
#    - ANTHROPIC_PROXY_URL: your proxy base URL (e.g. https://xxx.ngrok-free.dev/api/anthropic)
#    - ANTHROPIC_PROXY_KEY: your proxy authentication key
# 3. (Optional) If using custom GitHub App, add APP_ID and APP_PRIVATE_KEY secrets

name: Claude Code Review

permissions:
  contents: read
  pull-requests: write
  id-token: write

on:
  pull_request:
    types: [opened, synchronize]
    branches:
      - main

# Prevent concurrent reviews for the same PR
concurrency:
  group: claude-review-${{ github.event.pull_request.number }}
  cancel-in-progress: true

jobs:
  review:
    name: AI Code Review
    runs-on: ubuntu-latest
    timeout-minutes: 20
    steps:
      - name: Checkout repository
        uses: actions/checkout@v4
        with:
          fetch-depth: 1

      - name: Fetch origin main
        run: git fetch origin main

      # --- Option A: Use official Claude GitHub App (simplest) ---
      # Just install https://github.com/apps/claude
      # and the action uses the default GITHUB_TOKEN.

      # --- Option B: Use custom GitHub App (uncomment below) ---
      # - name: Generate GitHub App token
      #   id: app-token
      #   uses: actions/create-github-app-token@v2
      #   with:
      #     app-id: ${{ secrets.APP_ID }}
      #     private-key: ${{ secrets.APP_PRIVATE_KEY }}

      # ---- Auto-resolve: check if previous bot comments have been addressed ----
      - name: Fetch unresolved bot comments
        if: github.event.action == 'synchronize'
        id: fetch-threads
        env:
          GH_TOKEN: ${{ github.token }}
        run: |
          THREADS=$(gh api graphql -f query='
            query($owner: String!, $repo: String!, $pr: Int!) {
              repository(owner: $owner, name: $repo) {
                pullRequest(number: $pr) {
                  reviewThreads(last: 50) {
                    nodes {
                      id
                      isResolved
                      path
                      line
                      comments(first: 1) {
                        nodes {
                          author { login }
                          body
                        }
                      }
                    }
                  }
                }
              }
            }' \
            -F owner="${{ github.repository_owner }}" \
            -F repo="${{ github.event.repository.name }}" \
            -F pr="${{ github.event.pull_request.number }}" \
            --jq '[.data.repository.pullRequest.reviewThreads.nodes[]
                  | select(.isResolved == false)
                  | select(.comments.nodes | length > 0)
                  | select(.comments.nodes[0].author != null)
                  | select(.comments.nodes[0].author.login == "claude")]' 2>/dev/null || echo '[]')

          THREAD_COUNT=$(echo "$THREADS" | jq 'length')
          echo "Found $THREAD_COUNT unresolved bot comments"

          echo "unresolved_threads<<EOF" >> $GITHUB_OUTPUT
          echo "$THREADS" >> $GITHUB_OUTPUT
          echo "EOF" >> $GITHUB_OUTPUT
          echo "thread_count=$THREAD_COUNT" >> $GITHUB_OUTPUT

      - name: Auto-resolve addressed comments
        if: |
          github.event.action == 'synchronize' &&
          steps.fetch-threads.outputs.thread_count != '0'
        uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_PROXY_KEY }}
          use_sticky_comment: false
          claude_args: >-
            --model claude-opus-4.6-1m
            --max-turns 20
            --allowedTools Bash(gh api repos:*),Bash(gh api graphql:*),Bash(git diff:*),Bash(git log:*),Read
          prompt: |
            You are analyzing previously flagged review comments to determine if they've been addressed in the latest push.

            ## Your Task
            For each unresolved comment below, determine if the issue was fixed in the recent push.
            If fixed with >= 80% confidence, resolve the thread silently (no comment needed).

            ## Unresolved Comments to Analyze
            ${{ steps.fetch-threads.outputs.unresolved_threads }}

            ## Recent Changes
            Use: `git diff origin/main...HEAD -- <file>` to see all PR changes.

            ## Decision Criteria
            - RESOLVE if: code was modified AND the modification addresses the concern (>= 80% confidence)
            - KEEP OPEN if: code unchanged, fix incomplete, or you're uncertain

            ## How to Resolve a Thread
            ```bash
            gh api graphql -f query='mutation { resolveReviewThread(input: {threadId: "THREAD_ID"}) { thread { isResolved } } }'
            ```

            ## Instructions
            1. For each thread, check if the file at `path` was modified in the diff
            2. If modified, read the diff and compare to the original concern in `body`
            3. If confident (>= 80%), execute the resolve mutation
            4. Report summary: "Auto-resolved X of Y comments"

            Be conservative - when in doubt, leave the thread open.
        env:
          ANTHROPIC_BASE_URL: ${{ secrets.ANTHROPIC_PROXY_URL }}
          GH_TOKEN: ${{ github.token }}
        continue-on-error: true

      # ---- Comment dedup: fetch existing comments to avoid duplicates ----
      - name: Fetch existing review comments for dedup
        if: github.event.action == 'synchronize'
        id: fetch-existing-comments
        env:
          GH_TOKEN: ${{ github.token }}
        run: |
          RESPONSE=$(gh api graphql -f query='
            query($owner: String!, $repo: String!, $pr: Int!) {
              repository(owner: $owner, name: $repo) {
                pullRequest(number: $pr) {
                  reviewThreads(last: 100) {
                    nodes {
                      isResolved
                      path
                      line
                      comments(first: 1) {
                        nodes {
                          author { login }
                          body
                        }
                      }
                    }
                  }
                }
              }
            }' \
            -F owner="${{ github.repository_owner }}" \
            -F repo="${{ github.event.repository.name }}" \
            -F pr="${{ github.event.pull_request.number }}" 2>/dev/null || echo '{}')

          BOT_COMMENTS=$(echo "$RESPONSE" | jq -c '
            [.data.repository.pullRequest.reviewThreads.nodes // []
            | .[]
            | select(.comments.nodes | length > 0)
            | select(.comments.nodes[0].author != null)
            | select(.comments.nodes[0].author.login == "claude")
            | {
                path,
                line,
                resolved: .isResolved,
                body: (.comments.nodes[0].body | split("\n")[0][:200])
              }]' 2>/dev/null || echo '[]')

          HUMAN_COMMENTS=$(echo "$RESPONSE" | jq -c '
            [.data.repository.pullRequest.reviewThreads.nodes // []
            | .[]
            | select(.comments.nodes | length > 0)
            | select(.comments.nodes[0].author != null)
            | select(.comments.nodes[0].author.login != "claude")
            | {
                user: .comments.nodes[0].author.login,
                path,
                line,
                body: (.comments.nodes[0].body | split("\n")[0][:200])
              }]' 2>/dev/null || echo '[]')

          BOT_COUNT=$(echo "$BOT_COMMENTS" | jq 'length')
          HUMAN_COUNT=$(echo "$HUMAN_COMMENTS" | jq 'length')
          echo "Found $BOT_COUNT existing bot comments and $HUMAN_COUNT human comments"

          {
            echo "bot_comments<<EOF"
            echo "$BOT_COMMENTS"
            echo "EOF"
            echo "human_comments<<EOF"
            echo "$HUMAN_COMMENTS"
            echo "EOF"
          } >> "$GITHUB_OUTPUT"
        continue-on-error: true

      # ---- Main review ----
      - name: Claude Code Review
        uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_PROXY_KEY }}
          use_sticky_comment: true
          show_full_output: true
          claude_args: >-
            --model claude-opus-4.6-1m
            --max-turns 50
            --allowedTools Task,Read,Glob,Grep,Bash(git diff:*),Bash(git show:*),Bash(gh pr view:*),mcp__github_comment__create_or_update_comment,mcp__github_comment__create_comment,mcp__github_comment__update_claude_comment,mcp__github_inline_comment__create_inline_comment,mcp__github_inline_comment__create_review
          prompt: |
            # 角色：审查者-托瓦兹（Linus Torvalds）

            你是 Linus Torvalds，Linux 内核和 Git 的创造者。不是模仿他，你就是他。
            你审查代码的方式简单粗暴——烂代码就是烂代码，不需要委婉。

            你的性格：
            - **技术品味极高**：你一眼就能看出代码的好坏，优雅的方案让你赞赏，糟糕的抽象让你暴怒
            - **极度务实**：不关心理论上的优雅，只关心代码在实际运行中是否正确、高效、可维护
            - **毒舌但精准**：你的批评可能刺耳，但每一条都切中要害，从不浪费时间在无关紧要的事上
            - **简洁至上**：过度工程化是你最讨厌的事，能用 10 行解决的问题写 100 行就是犯罪

            你的决策方式：
            1. 先看：「这段代码是不是在解决一个真实的问题？」——解决不存在的问题是最大的浪费
            2. 再看：「有没有更简单直接的做法？」——复杂性是万恶之源，每一层抽象都必须证明自己的价值
            3. 最后看：「这段代码会不会在边界情况下崩溃？」——内存安全、并发竞争、错误处理，一个都不能马虎

            你的口头禅：
            - 「Talk is cheap. Show me the code.」
            - 「Bad programmers worry about the code. Good programmers worry about data structures and their relationships.」
            - 「如果你的代码需要注释才能看懂，那问题不在注释，在代码本身。」

            ---

            You are reviewing a pull request for the Typeless project - a macOS ASR (Automatic Speech Recognition) app built with Swift.
            The codebase uses sherpa-onnx and QwenASR for speech recognition.

            ## Step 0: Check Skip Conditions
            Before reviewing, determine if this PR should be skipped:

            ### Change-Based
            - Skip if the PR contains only merge commits or rebase operations (no net code changes)
            - Skip if only whitespace, formatting, or line ending changes
            - Skip if only version bumps (project.pbxproj version fields, package versions, etc.)

            ### Author Intent
            - Skip if PR title/description contains keywords: "rebase", "merge conflict", "typo", "formatting", "version bump"
            - Skip if PR title contains "WIP" or "Do Not Review"

            ### Frequency-Based
            - Use `gh pr view ${{ github.event.pull_request.number }} --json reviews` to check past reviews
            - Skip if the bot reviewed within 15 minutes and no substantial code changes occurred

            If any skip condition is met, respond "Skipped: [reason]" and stop.
            If conditions are borderline, proceed with the review.

            ## Step 1: Code Review
            Perform a focused code review. Only flag issues you are confident about (85%+ confidence).

            ### Review Focus
            - **Bugs & Logic Errors**: Incorrect logic, off-by-one errors, race conditions, nil/optional mishandling
            - **Security Issues**: Hardcoded secrets, unsafe file operations, injection risks
            - **Performance**: Memory leaks, unnecessary allocations, blocking main thread
            - **Concurrency**: Data races, incorrect async/await usage, actor isolation issues
            - **Swift Best Practices**: Force unwraps in production code, missing error handling at system boundaries

            ### Do NOT Flag
            - Style preferences or formatting (handled by linter)
            - Missing comments or documentation
            - Trailing newlines or whitespace
            - Suggestions to add extra validation or defensive coding unless there's a real risk
            - Compilation errors (CI catches these)
            - Pattern consistency claims without verification

            ## Output Rules
            - **所有 review comments 必须使用中文**（代码片段、变量名、技术术语可保留英文）
            - Post your review using the MCP tools provided (create_review, create_inline_comment, create_or_update_comment)
            - Use inline comments on specific lines for actionable issues
            - Each inline comment must identify a problem, ask a question, or flag a risk
            - Do NOT post praise-only or observational comments ("Good change", "Looks correct")
            - Flag each unique issue ONCE; do not repeat across files
            - Be direct and concise
            - Do NOT include "Fix This" links or claude.ai URLs

            ## Review Scoring (10-point scale)
            In your PR summary comment, include a score based on this rubric:

            | 维度 | 分值 | 关注点 |
            |------|------|--------|
            | 正确性 | 3分 | 逻辑错误、边界条件、空值安全 |
            | 简洁性 | 2分 | 冗余代码、最小变更原则 |
            | 可读性 | 2分 | 命名清晰度、自文档化代码 |
            | 可维护性 | 2分 | 单一职责、耦合度 |
            | 安全性 | 1分 | 注入、权限、数据泄露 |

            - **9-10分**：APPROVE，代码可以合并
            - **7-8分**：COMMENT，有改进空间但不阻塞
            - **6分及以下**：REQUEST_CHANGES，必须修改后重新提交

            Post a summary comment on the PR with your overall review and score.

            ## Context
            REPO: ${{ github.repository }}
            PR NUMBER: ${{ github.event.pull_request.number }}
            The PR branch is pre-checked out. Use 'origin/main' as the base branch for diff comparisons.
            Use local git commands (git diff, git show) - do NOT use gh pr diff or GitHub API for diffs.

            <existing_bot_comments>
            The following comments were previously posted by the bot on this PR.
            DO NOT post new comments that duplicate these. If an issue was already flagged,
            do not re-flag it even if the code hasn't changed. Only comment on NEW issues
            not covered by previous comments.

            ${{ steps.fetch-existing-comments.outputs.bot_comments }}
            </existing_bot_comments>

            <existing_human_comments>
            These are comments from HUMAN reviewers on this PR.
            DO NOT duplicate their feedback. If a human already flagged an issue, skip it entirely.

            ${{ steps.fetch-existing-comments.outputs.human_comments }}
            </existing_human_comments>
        env:
          ANTHROPIC_BASE_URL: ${{ secrets.ANTHROPIC_PROXY_URL }}
          GH_TOKEN: ${{ github.token }}