watch-pr

作者: huggingface

监控PR提交后的CI检查与Greptile代码审查。轮询CI状态,通过ralph-loop自动修复失败项,等待Greptile审查,处理…

npx skills add https://github.com/huggingface/openenv --skill watch-pr

/watch-pr

Monitor a submitted PR until CI passes and code reviews are addressed.

EXECUTE THESE STEPS NOW

When this skill is invoked, you MUST execute these steps immediately. Do NOT just describe what will happen — actually do it.

Step 0: Resolve PR Number and Repo

Extract the PR number from $ARGUMENTS. If no argument was provided, detect from the current branch:

gh pr view --json number -q '.number'

If no PR is found, stop with: "No PR found for current branch. Create one with gh pr create or pass a PR number: /watch-pr 123"

Also resolve the repo identifier:

gh repo view --json nameWithOwner -q '.nameWithOwner'

Store as PR_NUMBER and REPO. Initialize counters:

  • CI_FIX_COUNT = 0 (max 5)
  • REVIEW_FIX_COUNT = 0 (max 3)

Report to the user:

## Watching PR #<PR_NUMBER>
Monitoring CI and reviews for https://github.com/<REPO>/pull/<PR_NUMBER>

Step 1: WAITING_CI — Poll CI Checks

Run the CI polling script with a 30-minute timeout:

bash .claude/hooks/ci-wait.sh <PR_NUMBER> 1800

Important: Set the Bash tool timeout to 600000ms (10 minutes). If the script exceeds this, re-invoke it with the remaining timeout: bash .claude/hooks/ci-wait.sh <PR_NUMBER> <REMAINING_SECONDS>.

Evaluate the exit code:

  • Exit 0 (all checks passed): Go to Step 3 (WAITING_REVIEW).
  • Exit 1 (checks failed): Go to Step 2 (CI_FAILED).
  • Exit 2 (timeout): Report to user: "CI checks did not complete within 30 minutes. Check manually." Stop.
  • Exit 3 (error): Report error and stop.

Step 2: CI_FAILED — Fix and Retry

Increment CI_FIX_COUNT. If CI_FIX_COUNT > 5, stop with:

CI has failed 5 times. Manual intervention required.
PR: https://github.com/<REPO>/pull/<PR_NUMBER>

2a. Identify failed checks and get logs:

# Get the head SHA for this PR
HEAD_SHA=$(gh pr view <PR_NUMBER> --json headRefOid -q '.headRefOid')

# List failed workflow runs for this commit
gh run list --commit "$HEAD_SHA" --json databaseId,name,conclusion --jq '.[] | select(.conclusion == "failure")'

For each failed run, fetch the failure logs:

gh run view <RUN_ID> --log-failed 2>&1 | tail -200

2b. Try ralph-loop first:

Invoke the ralph-loop plugin to fix the failures:

Skill tool: skill: "ralph-loop"

If ralph-loop successfully fixes the issues (tests pass locally), commit and push the changes, then go to Step 1.

2c. Fallback to inline fix (if ralph-loop unavailable or fails):

  1. Read the failure logs from step 2a
  2. Identify the root cause (test failure, lint error, build error, etc.)
  3. Read the relevant source files
  4. Fix the code
  5. Verify locally:
    bash .claude/hooks/lint.sh && bash .claude/hooks/test.sh
    
  6. Stage, commit, and push:
    git add <specific-files>
    git commit -m "fix: address CI failure (<brief description>)"
    git push
    

2d. Return to Step 1 (WAITING_CI).

After pushing the fix, CI will re-run. Go back to Step 1.


Step 3: WAITING_REVIEW — Poll for Greptile Review

All CI checks have passed. Now wait for Greptile's code review.

Poll every 120 seconds for up to 3 hours (max 90 polls). On each poll:

# Check for reviews from greptile-apps[bot]
REVIEW_COUNT=$(gh api repos/<REPO>/pulls/<PR_NUMBER>/reviews \
  --jq '[.[] | select(.user.login == "greptile-apps[bot]")] | length')
# Check for line-level comments from greptile-apps[bot]
COMMENT_COUNT=$(gh api repos/<REPO>/pulls/<PR_NUMBER>/comments \
  --jq '[.[] | select(.user.login == "greptile-apps[bot]")] | length')

Decision logic:

  • If both counts are 0: print "Waiting for Greptile review... (poll N/90)", sleep 120 seconds, and poll again.
  • If 3-hour timeout reached with no review: go to Step 5 (DONE) with note "Greptile review did not arrive within 3 hours."
  • If review exists, check if actionable:
    • Actionable (go to Step 4): The latest review state is CHANGES_REQUESTED, OR there are line-level comments from greptile-apps[bot].
    • Not actionable (go to Step 5): Review state is APPROVED or COMMENTED with no line-level comments.

To check review state:

gh api repos/<REPO>/pulls/<PR_NUMBER>/reviews \
  --jq '[.[] | select(.user.login == "greptile-apps[bot]")] | last | .state'

Step 4: REVIEW_RECEIVED — Address Greptile Comments

Increment REVIEW_FIX_COUNT. If REVIEW_FIX_COUNT > 3, stop with:

Greptile has requested changes 3 times. Manual intervention required.
PR: https://github.com/<REPO>/pull/<PR_NUMBER>

4a. Fetch the review body:

gh api repos/<REPO>/pulls/<PR_NUMBER>/reviews \
  --jq '[.[] | select(.user.login == "greptile-apps[bot]")] | last | .body'

4b. Fetch all line-level comments:

gh api repos/<REPO>/pulls/<PR_NUMBER>/comments \
  --jq '[.[] | select(.user.login == "greptile-apps[bot]")] | .[] | {id: .id, path: .path, line: .line, body: .body}'

4c. Address each comment:

For each line-level comment:

  1. Read the file at the specified path and line
  2. Understand the suggestion
  3. If it aligns with project principles: apply the fix
  4. If it conflicts with .claude/docs/PRINCIPLES.md or .claude/docs/INVARIANTS.md: do NOT apply it. Reply explaining why:
    gh api repos/<REPO>/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies \
      -f body="Not applied: <reason based on project principles>"
    
  5. For applied fixes, reply to acknowledge:
    gh api repos/<REPO>/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies \
      -f body="Fixed in latest push."
    

4d. Human approval checkpoint:

Before committing review fixes, present a summary to the user for approval:

## Greptile Review Changes Summary

| # | Comment | Action | File |
|---|---------|--------|------|
| 1 | <brief description> | Applied / Declined | <path> |
| 2 | ... | ... | ... |

Approve these changes before pushing? (y/n)

Use the AskUserQuestion tool to get explicit approval. If the user declines, stop and let them handle the review manually.

4e. Verify, commit, and push:

After user approval:

bash .claude/hooks/lint.sh && bash .claude/hooks/test.sh

If local checks pass:

git add <specific-files>
git commit -m "fix: address Greptile review comments"
git push

4f. Return to Step 1 (WAITING_CI).

CI will re-run after the push. Go back to Step 1.


Step 5: DONE — Final Report

## Watch PR Complete

### PR
https://github.com/<REPO>/pull/<PR_NUMBER>

### CI Status: PASSED
- CI fix iterations: <CI_FIX_COUNT>

### Greptile Review
| Status | Details |
|--------|---------|
| Received | YES / NO (timed out) |
| Actionable comments | N |
| Comments addressed | N |
| Comments declined | N (with reasons) |
| Review fix iterations | <REVIEW_FIX_COUNT> |

### Final Status: READY FOR HUMAN REVIEW / NEEDS ATTENTION

If final status is READY (CI green, reviews addressed), report:

PR is ready for human review.

If final status is NEEDS ATTENTION (hit iteration limits), explain what remains.


Safeguards

LimitValueBehavior when exceeded
CI fix iterations5Stop, report failures, ask user
Greptile wait timeout3 hoursContinue without review
Review fix iterations3Stop, report outstanding comments

When to Use

  • After /pre-submit-pr creates and pushes a PR
  • After pushing fixes to an existing PR
  • When you want automated CI monitoring and review handling

When NOT to Use

  • Before a PR exists (run /pre-submit-pr first)
  • For draft PRs that aren't ready for review
  • When you want manual control over CI fixes

Workflow Integration

/work-on-issue #42  →  Start from GitHub issue
    ↓
/write-tests        →  Create failing tests (Red)
    ↓
/implement          →  Make tests pass (Green)
    ↓
/update-docs        →  Fix stale docs across repo
    ↓
/simplify           →  Refactor (optional)
    ↓
/pre-submit-pr      →  Validate before PR
    ↓
/watch-pr           →  Monitor CI + Greptile review  ← THIS SKILL

来自 huggingface 的更多技能

sync-models
huggingface
将chat-ui的模型配置与HuggingFace路由器同步——为新模型添加描述,标记支持推理的模型,为32B及以上的模型启用artifacts……
custom-blocks
huggingface
Use when the user has written (or wants to write) a `ModularPipelineBlocks` subclass in a local Python file and needs to package it into a Hub-uploadable…
self-review
huggingface
Use before opening a PR, or whenever asked to self-review a diffusers contribution. Applies the same rubric as the `@claude` CI (checks the diff against…
hf-cloud-sagemaker-production-defaults
huggingface
创建一个SageMaker端点(实时或异步),默认启用自动扩缩、CloudWatch告警和标签功能。每当要创建…时,使用此技能。
hf-cloud-serving-image-selection
huggingface
为SageMaker模型部署选择合适的服务容器,并查找其当前镜像URI。在即将部署模型到……时使用此技能。
Hugging Face Cli
huggingface
Execute Hugging Face Hub operations using the `hf` CLI. Use when the user needs to download models/datasets/spaces, upload files to Hub repositories, create repos, manage local cache, or run compute jobs on HF infrastructure. Covers authentication, file transfers, repository creation, cache operations, and cloud compute.
Hugging Face Datasets
huggingface
在 Hugging Face Hub 上创建和管理数据集。支持初始化仓库、定义配置/系统提示、流式更新行数据,以及基于 SQL 的数据集查询与转换。设计为与 HF MCP 服务器协同工作,实现完整的数据集工作流。
Hugging Face Evaluation
huggingface
在Hugging Face模型卡片中添加和管理评估结果。支持从README内容中提取评估表格、从Artificial Analysis API导入分数,以及使用vLLM/lighteval运行自定义模型评估。兼容model-index元数据格式。