pr-qa（质量保证）

PR QA Orchestrator

End-to-end pull request quality assurance. Takes a developer's branch or existing PR through to a merge-ready state by automating creation, multi-persona review, secrets/CVE scanning, Greptile-driven feedback loops, and systematic comment resolution.

When to Use

• The task is to submit a PR, finish PR QA, get a PR merge-ready, or clean up review feedback end to end.
• The workflow needs PR creation, scans, Greptile or review-bot loops, comment handling, or final merge-readiness checks.
• The request may start from a local branch, an existing draft PR, or an open PR with unresolved comments.
• The user wants end-to-end assurance across security, architecture, dependency risk, review feedback, and thread resolution.

Do NOT use when:

• The task is only to summarize changes into a PR body.
• The user wants a lightweight code review with no PR orchestration.
• The request concerns someone else’s PR where you should review rather than operate the full workflow.

Response format

Always structure the final response with these top-level sections, in this order:

1. Summary — state the task, scope, and main conclusion in 1-3 sentences.
2. Decision / Approach — state the key classification, assumptions, or chosen path.
3. Artifacts — provide the primary deliverable(s) for this skill. Use clear subheadings for multiple files, commands, JSON payloads, queries, or documents.
4. Validation — state checks performed, important risks, caveats, or unresolved questions.
5. Next steps — list concrete follow-up actions, or write None if nothing remains.

Rules:

• Do not omit a section; write None when a section does not apply.
• If files are produced, list each file path under Artifacts before its contents.
• If commands, JSON, SQL, YAML, or code are produced, put each artifact in fenced code blocks with the correct language tag when possible.
• Keep section names exactly as written above so output stays predictable across skills.

Workflow

Phase 1: Pre-flight Checks

Verify the environment is ready before any work begins.

# Verify gh CLI is authenticated
gh auth status

# Verify current branch has a remote tracking branch
git status -sb

# Check for uncommitted changes
git diff --stat
git diff --cached --stat

# Identify base branch (usually main or master)
git remote show origin | grep 'HEAD branch'

If there are uncommitted changes, warn the user and ask whether to commit them first or proceed with what's already committed.

Detect the project ecosystem by checking for: package.json (Node), requirements.txt/pyproject.toml (Python), go.mod (Go), Cargo.toml (Rust), pom.xml/build.gradle (Java). This determines which CVE scanner to use in Phase 2.

Phase 2: Security and Dependency Scan

Secrets Scanning

Run gitleaks scoped to the PR diff only:

gitleaks detect --source . --log-opts "origin/main..HEAD" --report-format json --report-path /tmp/gitleaks-report.json

If gitleaks is not installed, prompt the user to install it. Try these in order:

1. bun add -g gitleaks (preferred)
2. pnpm add -g gitleaks
3. npm install -g gitleaks
4. brew install gitleaks (macOS)

HARD GATE: If critical secrets are found (API keys, passwords, private keys, tokens), STOP and alert the user. Do NOT create the PR until secrets are removed from the diff and git history.

CVE Audit

Run the ecosystem-appropriate vulnerability scanner on dependency changes:

Flag any critical or high severity CVEs introduced by the PR's dependency changes. Medium/low CVEs are informational only.

Phase 3: Deep Self-Review

Perform a comprehensive review of the diff from four expert perspectives. Before reviewing, scan the repo root for architectural guidance documents:

# Look for design docs that define how the codebase should be architected
ls -la README.md ARCHITECTURE.md PRD.md DESIGN.md docs/*.md docs/plans/*.md 2>/dev/null

Read any found design documents to inform the review.

Security Engineer Perspective

• Input validation on all user-facing endpoints
• Authentication/authorization checks on protected resources
• SQL injection, XSS, CSRF, SSRF, path traversal risks
• Cryptographic misuse (weak algorithms, hardcoded keys, missing salts)
• Sensitive data exposure in logs, errors, or responses
• Race conditions in concurrent operations

Application Architect Perspective

• Consistency with existing codebase patterns and abstractions
• Proper use of existing utilities (no reinventing what exists)
• Separation of concerns and appropriate layering
• Error handling strategy (consistent with project conventions)
• Configuration management (no hardcoded values that should be configurable)
• Missing or inadequate tests for new functionality
• Adherence to project design docs (README, PRD, ARCHITECTURE.md)

Network Engineer Perspective

• API endpoint design (RESTful conventions, proper status codes)
• Timeout and retry configuration
• Connection pooling and resource cleanup
• TLS/certificate handling
• DNS and service discovery patterns

Scalability Engineer Perspective

• N+1 query patterns in database access
• Missing pagination on list endpoints
• Unbounded memory growth (loading full datasets)
• Missing caching where appropriate
• Blocking operations in async contexts
• Missing indexes for new query patterns

Produce structured findings with severity levels:

For each finding, provide:

1. File and line location
2. Description of the issue
3. Why it matters (which persona flagged it)
4. Concrete fix with a diff snippet

Auto-fix boundary: Automatically fix formatting, linting issues, obvious null checks, missing error handling, and trivial bugs. Present fixes for business logic, API contracts, database changes, or architectural decisions to the user for approval.

Phase 4: Create Or Reuse PR

Generate a structured PR description:

## Summary

[1-2 sentence overview of what this PR does]

## Motivation

[Why this change is needed — link to issue if applicable]

## Changes

[Group by concept, not file-by-file. Each group has a heading and explanation]

### [Concept 1]

- What changed and why

### [Concept 2]

- What changed and why

## Testing Instructions

[Copy-pasteable steps to verify the change works]

## Rollback Plan

[How to revert if something goes wrong]

## Reviewer Notes

[Areas that need careful review, known limitations, follow-up work]

Determine the current PR state before acting:

1. No PR exists yet — push the branch and create a draft PR
2. A draft PR already exists — reuse it and keep it draft through the QA loop
3. An open non-draft PR already exists — reuse it, skip draft creation, and only call gh pr ready later if the PR is actually in draft state

If no PR exists yet, create the draft PR:

# Push any auto-fix commits first
git push origin HEAD

# Create draft PR
gh pr create --draft --title "<concise title>" --body "<structured description>"

If a PR already exists, capture its PR number from gh pr view or the user input and reuse it for all subsequent phases.

Phase 5: Greptile Review Loop

This phase runs up to 5 cycles. Each cycle:

Treat this phase as a required loop, not a one-time check. After every code fix push that is meant to satisfy Greptile, request Greptile again and restart the polling step from the beginning. Do not leave Phase 5 until the latest Greptile round has been polled to completion and either returned clean or hit the cycle limit.

Step 1: Request Greptile Review

# Comment @greptile on the PR to trigger review and capture the exact comment ID
COMMENT_ID=$(gh api repos/{owner}/{repo}/issues/<PR_NUMBER>/comments -f body='@greptile' --jq '.id')

Step 2: Poll for Greptile Response

Greptile responds with an eyes emoji when it starts reviewing (~10 minutes to complete). Poll every 20 seconds and compare against the state that existed before the current @greptile request:

# Snapshot existing Greptile comment IDs before starting this cycle
INLINE_BEFORE=$(gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments --jq '[.[] | select(.user.login == "greptile-inc[bot]" or .user.login == "greptile[bot]") | .id]')
ISSUE_BEFORE=$(gh api repos/{owner}/{repo}/issues/<PR_NUMBER>/comments --jq '[.[] | select(.user.login == "greptile-inc[bot]" or .user.login == "greptile[bot]") | .id]')

# Check reactions on the comment (eyes = reviewing, thumbs-up = done clean)
gh api repos/{owner}/{repo}/issues/comments/$COMMENT_ID/reactions --jq '.[].content'

# Check for new review comments from Greptile
gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments --jq '[.[] | select(.user.login == "greptile-inc[bot]" or .user.login == "greptile[bot]")]'

# Also check issue-level comments
gh api repos/{owner}/{repo}/issues/<PR_NUMBER>/comments --jq '[.[] | select(.user.login == "greptile-inc[bot]" or .user.login == "greptile[bot]")]'

# Check current review-thread state for Greptile-authored comments
gh api graphql -f query='
  query($owner: String!, $repo: String!, $pr: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $pr) {
        reviewThreads(first: 100) {
          nodes {
            id
            isResolved
            comments(first: 10) {
              nodes {
                databaseId
                author { login }
              }
            }
          }
        }
      }
    }
  }
' -f owner="{owner}" -f repo="{repo}" -F pr=<PR_NUMBER>

Polling strategy:

• Check every 20 seconds
• Timeout after 15 minutes per cycle
• Capture the Greptile comment IDs that existed before the cycle and keep polling until the current cycle produces new Greptile output or an explicit clean signal
• One empty poll is never success; lack of change means keep waiting
• If you push fixes for this cycle, return to Step 1 and restart this entire polling step for the next cycle
• If timeout: warn user and ask whether to continue waiting or proceed

Step 3: Process Greptile Findings

When Greptile responds with findings:

1. Fetch PR-level comments (general review comments): gh api repos/{owner}/{repo}/issues/<PR_NUMBER>/comments These are useful for Greptile summary comments and status updates, but they are not review threads and cannot be closed with resolveReviewThread.
2. Fetch ALL inline code comments (file-specific review comments): gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments --jq '[.[] | select(.user.login == "greptile-inc[bot]" or.user.login == "greptile[bot]")]' IMPORTANT: Greptile often leaves multiple inline comments across different files. You MUST iterate over every inline comment individually — do not stop after reading the first one. Each comment targets a specific file and line range, and each must be addressed on its own merits.
3. Fetch review threads so comment IDs can be mapped to thread IDs and resolution state: gh api graphql -f query=' query($owner: String!, $repo: String!, $pr: Int!) {repository(owner: $owner, name: $repo) {pullRequest(number: $pr) {reviewThreads(first: 100) {nodes {id isResolved comments(first: 10) {nodes {databaseId author {login} path line}}}}}}} ' -f owner="{owner}" -f repo="{repo}" -F pr=<PR_NUMBER>
4. For EACH inline comment, extract:

- id — the comment ID (needed for replying and resolving) - path — the file the comment targets - line / original_line — the line number - body — the suggestion or finding text - diff_hunk — surrounding diff context

1. Categorize each finding by severity and addressability
2. Address every finding individually:

- Code fix needed: Make the change, commit with message referencing the finding - Explanation needed: Reply to the comment explaining why the current approach is correct - Dismiss with justification: Reply explaining why the suggestion doesn't apply

1. Reply to each inline comment to create a paper trail: # Reply to an inline review comment gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies \ -f body="Fixed — <brief description of what was changed>"
2. Resolve each addressed comment thread using GraphQL: # First, get the GraphQL node ID for the review thread # Each pull request review comment belongs to a thread that can be resolved gh api graphql -f query=' query($owner: String!, $repo: String!, $pr: Int!) {repository(owner: $owner, name: $repo) {pullRequest(number: $pr) {reviewThreads(first: 100) {nodes {id isResolved comments(first: 1) {nodes {databaseId}}}}}}} ' -f owner="{owner}" -f repo="{repo}" -F pr=<PR_NUMBER> # Then resolve the thread by its GraphQL node ID gh api graphql -f query=' mutation($threadId: ID!) {resolveReviewThread(input: {threadId: $threadId}) {thread {isResolved}}} ' -f threadId="<THREAD_NODE_ID>" Map comment IDs to thread IDs: The GraphQL query returns threads with their first comment's databaseId. Match this against the REST API comment id to find the correct thread to resolve. Only resolve threads you have actually addressed — do not bulk-resolve without processing each one. Verify the close actually stuck: After each resolveReviewThread mutation, re-fetch reviewThreads and confirm the targeted thread now shows isResolved: true. If it is still false, treat the thread as unresolved and keep it in the next processing pass.
3. Acknowledge Greptile issue-level summary comments with a follow-up PR comment describing what was fixed or why no change was needed. These summary comments cannot be resolved via GraphQL, so do not mistake them for closed review threads.
4. Push all fixes:

git add -A && git commit -m "address Greptile review findings (cycle N)" && git push origin HEAD

1. If this is cycle < 5 and there were code changes, new Greptile findings, or any Greptile thread remained unresolved: go to Step 1 for the next cycle
2. If the latest Greptile cycle returned thumbs-up or no findings and all Greptile review threads from that cycle are resolved: the Greptile loop is complete

Step 4: Escalation (after 5 cycles)

If 5 cycles complete without a clean Greptile review:

• Present all remaining unresolved findings to the user
• Ask for manual decision on each: fix, dismiss, or defer to post-merge

Phase 6: Comment Resolution

Fetch ALL open review conversations, not just Greptile:

# Get all PR review comments (inline)
gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments

# Get all issue-level comments (top-level)
gh api repos/{owner}/{repo}/issues/<PR_NUMBER>/comments

# Get review threads with resolution status via GraphQL
gh api graphql -f query='
  query($owner: String!, $repo: String!, $pr: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $pr) {
        reviewThreads(first: 100) {
          nodes {
            id
            isResolved
            comments(first: 10) {
              nodes {
                databaseId
                body
                author { login }
                path
                line
              }
            }
          }
        }
      }
    }
  }
' -f owner="{owner}" -f repo="{repo}" -F pr=<PR_NUMBER>

Use the GraphQL response as the source of truth for which threads are still unresolved (isResolved: false). This catches threads from Greptile, human reviewers, and any other bot.

For each unresolved thread:

1. Read ALL comments in the thread (not just the first) to understand the full conversation context
2. If it requires a code change: make the fix, push, reply confirming the fix
3. If it's a question: reply with the answer
4. If it's resolved by a previous commit: reply linking to the commit
5. Mark the thread as resolved after addressing it: gh api graphql -f query=' mutation($threadId: ID!) {resolveReviewThread(input: {threadId: $threadId}) {thread {isResolved}}} ' -f threadId="<THREAD_NODE_ID>"
6. Immediately re-query reviewThreads and confirm the same thread now reports isResolved: true. If not, leave it on the unresolved list and keep working it in the next pass.

For issue-level Greptile comments:

• Post a follow-up PR comment summarizing the fix, explanation, or dismissal
• Do not treat issue-level comments as resolvable review threads
• Consider them closed only after the follow-up is posted and the next Greptile cycle returns clean

Do NOT resolve threads you haven't addressed. Each resolution must follow a reply that demonstrates the finding was handled — either by a code fix, an explanation, or a justified dismissal.

Phase 7: Final Verification

Re-run all checks on the final state:

# Re-run secrets scan
gitleaks detect --source . --log-opts "origin/main..HEAD"

# Check CI status
gh pr checks <PR_NUMBER>

# Verify no unresolved review threads
gh api graphql -f query='
  query($owner: String!, $repo: String!, $pr: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $pr) {
        reviewThreads(first: 100) {
          nodes { isResolved }
        }
      }
    }
  }
' -f owner="{owner}" -f repo="{repo}" -F pr=<PR_NUMBER> --jq '.data.repository.pullRequest.reviewThreads.nodes | map(select(.isResolved == false)) | length'

# Verify branch is up to date
git fetch origin && git log HEAD..origin/main --oneline

If the base branch has moved ahead, rebase and re-push:

git fetch origin main && git rebase origin/main && git push origin HEAD --force-with-lease

Phase 8: Mark Ready for Review

If the PR is already open and not in draft state, skip gh pr ready and simply report that it remains open after QA completion. Only run gh pr ready when the current PR is actually draft.

gh pr ready <PR_NUMBER>

Present a summary to the user:

• Total findings discovered across all review phases
• Findings fixed automatically vs. fixed with approval vs. deferred
• Greptile cycles completed
• CI check status
• Any remaining items that need manual attention

Checklist

• gh auth status confirms authentication
• Branch is pushed with commits ahead of base
• Secrets scan completed with no critical findings
• CVE audit completed, critical/high CVEs addressed
• Self-review completed from all four perspectives
• Auto-fixes committed and pushed
• Draft PR created via gh pr create --draft when no PR already existed
• Existing draft or open PR reused when the request started from an existing PR
• Greptile review requested via @greptile comment
• Greptile re-requested and re-polled after every fix push until the latest round completed
• ALL Greptile inline comments individually addressed (not just the first)
• Greptile findings fixed or escalated after 5 cycles
• Each addressed Greptile thread verified resolved via GraphQL resolveReviewThread and isResolved: true
• Greptile issue-level summary comments acknowledged with follow-up comments
• Final secrets re-scan clean
• CI checks passing
• Branch up to date with base
• PR marked ready via gh pr ready if the PR was draft
• Existing open non-draft PR left open without forcing gh pr ready

Example

Input: Developer on branch feature/add-rate-limiting with 3 commits adding rate limiting middleware to an Express API.

Phase 2 output (secrets scan):

gitleaks: 1 finding
  File: src/config.ts:15
  Rule: generic-api-key
  Match: RATE_LIMIT_API_KEY = "sk-proj-abc123..."

ACTION REQUIRED: Remove hardcoded API key before PR creation.
  Fix: Move to environment variable, add to .env.example

Phase 3 output (self-review findings):

[CRITICAL] Security: Rate limiter uses client IP from X-Forwarded-For without validation
  File: src/middleware/rateLimiter.ts:23
  Persona: Security Engineer
  Fix: Validate X-Forwarded-For against trusted proxy list

[HIGH] Architecture: Rate limit config is hardcoded, should use existing ConfigService
  File: src/middleware/rateLimiter.ts:5-8
  Persona: Application Architect
  Fix: Inject ConfigService, move values to config/rate-limiting.yaml

[MEDIUM] Scalability: In-memory rate limit store won't work with multiple instances
  File: src/middleware/rateLimiter.ts:12
  Persona: Scalability Engineer
  Fix: Use Redis store (project already has Redis dependency)

Phase 5 output (Greptile cycle 1):

Greptile review requested... polling every 20s
[20s] Checking... eyes reaction detected, review in progress
[40s] Checking... still reviewing
...
[200s] Greptile review complete. 2 inline findings:

1. src/middleware/rateLimiter.ts:45 - "Consider using sliding window instead
   of fixed window for smoother rate limiting behavior"
   -> Addressing: switching to sliding window algorithm

2. src/routes/api.ts:12 - "Rate limiter should be applied after auth middleware,
   not before"
   -> Addressing: reordering middleware chain

Pushing fixes... requesting Greptile cycle 2...
[cycle 2] Greptile returned thumbs-up. No findings.
[cycle 2] Re-checking reviewThreads... all targeted Greptile threads now show isResolved=true.

Common Mistakes

Quick Reference

Key Principles

1. Draft-first, always — Never create a PR as ready-for-review. The draft state is your working space for QA. Only mark ready after ALL checks pass. This prevents premature review notifications and wasted reviewer time.
2. Block on secrets, warn on everything else — A leaked secret is irreversible. Stop immediately if gitleaks finds credentials. For all other findings (CVEs, architecture issues, style), present them ranked by severity but let the workflow continue.
3. Fix forward, don't suppress — When Greptile or self-review finds an issue, fix the code. Don't dismiss findings without a concrete justification. If a finding is genuinely a false positive, reply to the comment explaining why — don't just ignore it.
4. Scope scans to the PR diff — Run gitleaks on origin/main..HEAD, not the full repo. Run CVE audits on changed dependencies, not the full lockfile. This keeps results relevant and fast.
5. Exhaust automation before asking the human — Run all scans, do the self-review, go through Greptile cycles. Only escalate to the user for decisions that require business context (architectural trade-offs, feature scope, intentional risk acceptance).
6. Poll to the latest state, not the first response — A Greptile cycle is only complete when the newest request after the newest fix push has been polled to completion and its addressed review threads are confirmed isResolved: true.