Skip to content
cccskills
FreePublished on GitHubEnglish overview

check-pr

Inspect proposed changes in GitHub, GitLab, or Perforce for review feedback, failed automated checks, and incomplete descriptions. It separates actionable findings from informational or resolved comments, then helps fix issues and close addressed review threads when requested.

OriginalShow the original description

Check a GitHub, GitLab, or Perforce PR/MR/CL for review comments, failing checks, and PR-body gaps. Use when asked to inspect, fix, or prepare a change for submission.

Works withGitHubGitLab
AI rating80/ 100See how to install

Good for

  • Inspect a change before submission
  • Identify failed checks and required fixes
  • Organize human and bot review feedback
  • Fill description gaps and resolve addressed threads

Overview

What it does

This skill examines GitHub PRs, GitLab MRs, and Perforce CLs for review feedback, failed automated checks, and gaps in change descriptions. It groups findings into items that need action, reference information, and items already resolved. A table explains the reasons and next steps. When requested, it helps edit code or descriptions, apply the fixes, and resolve review threads that have been addressed.

When it helps

Use it before submitting a change or after review to identify the work that remains. It checks comments from people and bots, including summaries that have been edited. If you leave out the change number, it looks for the relevant change using the current branch or similar context.

Example requests

  • “List this PR’s failed checks and unresolved review findings.”
  • “Check the current MR before submission and identify gaps in its description.”
  • “Fix the review findings and resolve the threads that have been addressed.”

Notes

You need git and an authenticated gh, glab, or p4, depending on the platform. GitHub requires a successfully completed Greptile check for the current commit; older results cannot complete the inspection. GitLab and Perforce Greptile integrations have equivalent freshness checks. The license is MIT.

This overview was written by AI (OpenAI Codex) from the public SKILL.md. Check the original below for the exact behavior.

Objective evaluation

AI rated this skill on four points using only the contents of its SKILL.md. Stars and popularity are not taken into account.

Clear instructionsIs it clear when to use it and what it does?
AGoodIt explicitly targets review comments, failing checks, and description gaps in PRs, MRs, and CLs.
Concrete stepsAre steps, inputs, outputs, and examples spelled out?
AGoodPlatform-specific commands cover identifying changes, fetching details, and waiting for checks.
SafetyAre sending data, deleting files, and running commands disclosed?
BPartly lackingFixes are in scope, but the supplied excerpt lacks confirmation steps for applying changes or resolving threads.
Easy setupDoes it need few tools or API keys, and are they documented?
BPartly lackingAuthenticated CLIs are required explicitly, but installation steps for jq and Greptile are absent.

The total counts each point as A = 25, B = 15, C = 5. Rated by AI (OpenAI Codex) as of Oct 20, 2026. Treat it as a guide and read the original before installing.

Disagree? Request a re-evaluation

Tell us which point is wrong and why. No login needed. We will check it against the SKILL.md, redo the evaluation, and show the result here.

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

Files included (3)

  • SKILL.md14.3 KB
  • references/gitlab-api.md2.5 KB
  • references/graphql-queries.md2.0 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Check PR

Analyze a pull request (GitHub), merge request (GitLab), or shelved changelist (Perforce) for review comments, status checks, and description completeness, then help address any issues found.

Inputs

  • PR/MR/CL number (optional): If not provided, detect the PR/MR for the current branch, or the default pending changelist for p4.

Instructions

0. Detect platform

First check if the user is working in a Perforce depot by looking for a .p4config file or P4CLIENT/P4PORT environment variables:

# Check for Perforce environment
if p4 info >/dev/null 2>&1; then
  VCS="perforce"
else
  # Fall back to git remote detection
  REMOTE_URL=$(git remote get-url origin)
  if echo "$REMOTE_URL" | grep -qi "gitlab"; then
    VCS="gitlab"
  else
    VCS="github"
  fi
fi

For self-hosted GitLab instances whose hostname doesn't contain "gitlab", the user can override by passing --vcs gitlab as an input. For Perforce, the user can override by passing --vcs perforce.

1. Identify the PR/MR/CL

If a number was provided, use it. Otherwise, detect it:

GitHub:

gh pr view --json number,headRefName,headRefOid -q '{number: .number, branch: .headRefName, head: .headRefOid}'

GitLab:

glab mr view --output json | jq '.iid'

Perforce:

# List pending changelists for the current user/client
p4 changes -s pending -u $P4USER -c $P4CLIENT

Key field differences between platforms:

  • GitHub: number, headRefName, headRefOid
  • GitLab: iid, source_branch, sha
  • Perforce: changelist number (CL), shelved files for in-review CLs

2. Fetch PR/MR/CL details

GitHub:

gh pr view <PR_NUMBER> --json title,body,state,reviews,comments,headRefName,headRefOid,statusCheckRollup
OWNER_REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
gh api "repos/$OWNER_REPO/pulls/<PR_NUMBER>/comments"
gh api --paginate "repos/$OWNER_REPO/issues/<PR_NUMBER>/comments?per_page=100"

GitHub PRs are also issues, so general PR comments live on the issue comments endpoint. Greptile may edit a single general PR comment on each review cycle instead of creating a new review or comment. Always inspect the latest Greptile-authored general comment by updated_at, including any "Prompt to fix all with AI" section, before concluding that the PR is clear.

GitLab:

glab mr view <MR_IID> --output json
# Fetch discussions (inline diff comments are type "DiffNote"; general comments have null type)
glab api "projects/:fullpath/merge_requests/<MR_IID>/discussions"

For GitLab, paginate discussions if needed (add ?per_page=100&page=N).

Perforce:

# Get changelist description, files, and status
p4 describe -s <CL_NUMBER>

# Get shelved files (for in-review CLs)
p4 describe -S <CL_NUMBER>

# Get the diff of the shelved changelist
p4 diff2 //...@=<CL_NUMBER> //...@=<CL_NUMBER>

# List review comments (if using p4 review workflow)
p4 review -c <CL_NUMBER>

Key Perforce CL fields:

  • Change: changelist number
  • Status: pending, submitted, shelved
  • Description: the CL description / commit message
  • Files: list of files in the CL

3. Wait for pending checks

Before analyzing, ensure all status checks have completed. If any checks are PENDING or IN_PROGRESS (GitHub) / running or pending (GitLab), poll every 30 seconds until all checks reach a terminal state.

GitHub: poll statusCheckRollup from gh pr view.

GitLab:

glab api "projects/:fullpath/merge_requests/<MR_IID>/pipelines"

Pipeline statuses: running, pending, success, failed, canceled, skipped. Poll until no pipeline has running or pending status.

Perforce: Perforce doesn't have built-in CI checks natively. If the team uses a review tool (Swarm, etc.) or an external CI triggered by shelve events, check the relevant system. Otherwise, proceed to analysis immediately.

4. Require a fresh Greptile review for the current head

For GitHub PRs, do not treat an existing Greptile review, comment, or summary as current unless it is tied to the PR's exact current headRefOid. This is especially important after pushing a new commit to an existing PR: a Greptile review on an older commit is stale, even if the PR still has a Greptile comment or prior review.

Fetch the current head SHA immediately before the Greptile gate:

HEAD_SHA=$(gh pr view <PR_NUMBER> --json headRefOid -q .headRefOid)

Then inspect check-runs for that commit and require a completed Greptile run:

OWNER_REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
GREPTILE_CHECKS=$(gh api "repos/$OWNER_REPO/commits/$HEAD_SHA/check-runs?per_page=100" \
  --jq '[.check_runs[] | select(.name | test("greptile"; "i"))]')

# A run only counts as a valid fresh pass when it has completed AND concluded cleanly.
# GitHub check-run conclusions: success, neutral, skipped, failure, timed_out,
# cancelled, action_required, stale. Treat success/neutral as clean;
# everything else (especially failure and action_required) must block.
FRESH_GREPTILE_COMPLETED=$(echo "$GREPTILE_CHECKS" \
  | jq '[.[] | select(.status == "completed")] | length')
FRESH_GREPTILE_CLEAN=$(echo "$GREPTILE_CHECKS" \
  | jq '[.[] | select(.status == "completed" and (.conclusion | IN("success","neutral")))] | length')
FRESH_GREPTILE_BLOCKING=$(echo "$GREPTILE_CHECKS" \
  | jq '[.[] | select(.status == "completed" and ((.conclusion | IN("success","neutral")) | not))] | length')

if [ "$FRESH_GREPTILE_COMPLETED" = "0" ]; then
  echo "Blocked: no completed Greptile review/check is tied to current PR head $HEAD_SHA."
  echo "Request a fresh Greptile review against this head before marking the PR check done."
  echo "Suggested trigger: gh pr comment <PR_NUMBER> --body \"@greptile review\""
  exit 1
fi

if [ "$FRESH_GREPTILE_BLOCKING" != "0" ] || [ "$FRESH_GREPTILE_CLEAN" = "0" ]; then
  echo "Blocked: Greptile completed on head $HEAD_SHA but did not conclude clean"
  echo "(conclusion was failure/action_required/timed_out/cancelled/stale, or no clean run exists)."
  echo "Address the findings, push, and re-run Greptile until it concludes success/clean on the new head."
  exit 1
fi

If a Greptile check exists for the current head but is still pending or in progress, wait for it with the same polling pattern used by greploop rather than proceeding from older review material. If no Greptile check appears for the current head after a reasonable wait, report the PR as blocked on a fresh Greptile review for HEAD_SHA and stop. Do not mark the check complete from PR comments, PR reviews, or Greptile summaries that cannot be associated with the current head SHA.

For GitLab installations with Greptile integration, apply the same freshness rule against the MR's current head SHA:

# 1. Get the MR's current head SHA
MR_SHA=$(glab mr view <MR_IID> --output json | jq -r '.sha // .diff_refs.head_sha')

# 2. Find the latest pipeline for that EXACT sha, then the Greptile job within it.
LATEST_PIPELINE_ID=$(glab api "projects/:fullpath/merge_requests/<MR_IID>/pipelines" \
  | jq -r --arg sha "$MR_SHA" '[.[] | select(.sha == $sha)] | sort_by(.id) | last | .id // empty')

if [ -n "$LATEST_PIPELINE_ID" ]; then
  GREPTILE_JOBS=$(glab api "projects/:fullpath/pipelines/$LATEST_PIPELINE_ID/jobs" \
    | jq --arg sha "$MR_SHA" '[.[] | select(.name | test("greptile"; "i"))
        | {name, status, pipeline_sha: $sha}]')
else
  GREPTILE_JOBS='[]'
fi

GREPTILE_JOB_SUCCESS=$(echo "$GREPTILE_JOBS" \
  | jq '[.[] | select(.status == "success")] | length')
GREPTILE_JOB_BLOCKING=$(echo "$GREPTILE_JOBS" \
  | jq '[.[] | select(.status != "success")] | length')

# 3. If Greptile integrates via MR notes instead of a CI job, require the newest
#    Greptile note to reference the current head sha and report a clean review.
GREPTILE_NOTES=$(glab api "projects/:fullpath/merge_requests/<MR_IID>/discussions?per_page=100" \
  | jq --arg sha "$MR_SHA" '[.[].notes[]
      | select(.author.username | test("greptile"; "i"))]
      | sort_by(.updated_at // .created_at)')
GREPTILE_NOTE_CLEAN=$(echo "$GREPTILE_NOTES" \
  | jq --arg sha "$MR_SHA" 'if length == 0 then 0
      elif ((last.body // "") | contains($sha) and test("Confidence Score:[[:space:]]*5/5|Confidence:[[:space:]]*5/5|\\b5/5\\b"; "i") and (test("Prompt To Fix|blocking issue|failed|action required"; "i") | not)) then 1
      else 0 end')

if [ "$GREPTILE_JOB_SUCCESS" = "0" ] && [ "$GREPTILE_NOTE_CLEAN" = "0" ]; then
  echo "Blocked: no successful Greptile job or completed-clean current-head Greptile note is tied to MR head $MR_SHA."
  exit 1
fi

if [ "$GREPTILE_JOB_BLOCKING" != "0" ]; then
  echo "Blocked: at least one Greptile job for MR head $MR_SHA did not succeed."
  exit 1
fi

Block completion if, for MR_SHA, there is (a) no Greptile job or note at all (missing), (b) the newest Greptile job/note is tied to a different sha (stale), or (c) the Greptile job status is not success. A completed Greptile result for a different SHA is stale and must block completion.

For Perforce installations with Greptile integration, apply the same rule using the CL's current shelved-revision identity and the Greptile webhook/review artifact tied to it; a Greptile result for an earlier shelf is stale and must block completion.

5. Analyze the PR/MR

Once all checks are complete, evaluate these areas:

A. Status Checks

  • Are all CI checks passing?
  • If any are failing, identify which ones and the failure reason.

B. PR/MR Description

  • Is the description complete and follows team conventions?
  • Are all required sections filled in?
  • Are there TODOs or placeholders that need updating?

C. Review Comments

  • Inline code review comments that need addressing
  • Look for bot review comments (e.g. from greptile-apps[bot] on GitHub, or the Greptile bot user on GitLab, linters, etc.)
  • Human reviewer comments
  • Perforce: review comments from p4 review or external review tools

D. General Comments

  • Discussion comments on the PR/MR
  • For GitHub, check the issue comments endpoint and use updated_at to catch bot comments edited in place. Greptile's latest edited summary can contain actionable items even when there are no new inline comments.
  • Bot comments (deploy previews, etc.) - usually informational
  • Perforce: CL description should include a clear summary, affected files rationale, and testing notes

6. Categorize issues

For each issue found, categorize as:

CategoryMeaning
ActionableCode changes, test improvements, or fixes needed
InformationalVerification notes, questions, or FYIs that don't require changes
Already addressedIssues that appear to be resolved by subsequent commits

7. Report findings

Present a summary table:

AreaIssueStatusAction Needed
Status ChecksCI build failingFailingFix type error in src/api.ts
Review"Add null check" - @reviewerActionableAdd guard clause
DescriptionTODO placeholder in test planActionableFill in test plan
Review"Looks good" - @teammateInformationalNone

8. Fix issues (if requested)

If there are actionable items:

  1. Switch to the PR/MR's branch (git) or ensure files are open in the correct CL (Perforce) if not already.
  2. Ask the user if they want to fix the issues.
  3. If yes, make the fixes, then:

GitHub/GitLab: commit and push:

git add <files>
git commit -m "address review feedback"
git push

Perforce: open files for edit, make changes, and re-shelve:

p4 edit <file>
# make changes
p4 shelve -f -c <CL_NUMBER>

9. Resolve review threads

After addressing comments, resolve the corresponding review threads.

Perforce - Perforce does not have a native "resolve thread" concept. Instead, mark comments as addressed by updating the CL description or by responding in the review tool being used (Swarm, etc.). If using p4 review:

# Mark files as reviewed after addressing feedback
p4 review -c <CL_NUMBER>

GitHub - fetch unresolved thread IDs (paginate if needed - see the GraphQL reference):

gh api graphql -f query='
query($cursor: String) {
  repository(owner: "OWNER", name: "REPO") {
    pullRequest(number: PR_NUMBER) {
      reviewThreads(first: 100, after: $cursor) {
        pageInfo { hasNextPage endCursor }
        nodes {
          id
          isResolved
          comments(first: 1) {
            nodes { body path }
          }
        }
      }
    }
  }
}'

If hasNextPage is true, repeat with -f cursor=ENDCURSOR to get remaining threads.

Then resolve threads that have been addressed or are informational:

gh api graphql -f query='
mutation {
  resolveReviewThread(input: {threadId: "THREAD_ID"}) {
    thread { isResolved }
  }
}'

Batch multiple resolutions into a single mutation using aliases (t1, t2, etc.).

GitLab - fetch unresolved discussions (see the GitLab API reference):

glab api "projects/:fullpath/merge_requests/<MR_IID>/discussions?per_page=100"

Filter for discussions where "resolved": false. Collect each discussion's id.

Resolve each discussion individually (GitLab has no batch resolution):

glab api --method PUT \
  "projects/:fullpath/merge_requests/<MR_IID>/discussions/<DISCUSSION_ID>" \
  --field resolved=true

Repeat for each unresolved discussion ID.

10. Multiple PRs/MRs/CLs

If checking a chain of PRs/MRs/CLs, process them sequentially.

Perforce - to check multiple changelists at once:

p4 changes -s pending -u $P4USER -c $P4CLIENT -l

Output format

Summarize:

  • PR/MR/CL title or description and current state
  • Platform detected (GitHub / GitLab / Perforce)
  • Status checks summary (passing/failing/pending) - or N/A for Perforce
  • Total issues found
  • Actionable items with descriptions
  • Items that can be ignored with reasons
  • Recommended next steps

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

add-product-e2e-eval

FreeSummaryAI rating 90

Add or extend Paperclip end-to-end evaluations that exercise real browser workflows across the server, database, and runner. Define expected outcomes, check independent evidence, and connect results to the existing reporting pipeline for local or Daytona execution.

  • Adding an evaluation for a complete user workflow
  • Updating test data and outcome checks
  • Including screenshots and artifacts in evaluation reports
paperclipai/paperclip105.7KUpdated Oct 21, 2026

add-runner-eval

FreeSummaryAI rating 90

Add or extend Paperclip Runner evaluations that test real sessions against a mock control service. Define observable pass conditions, validate cases and run selections, and retain source revisions, evidence, and costs for authorized live runs.

  • Adding a focused Runner protocol evaluation
  • Testing grading rules with valid and invalid evidence
  • Updating evaluation selections and campaign membership
paperclipai/paperclip105.7KUpdated Oct 21, 2026

agent-browser

FreeSummaryAI rating 90

Use a real browser to verify web pages and apps through short, supervised interactions. Capture screenshots, inspect errors and network requests, and document reproducible steps for UI bugs or expected behavior checks.

  • Capture screenshots to verify UI changes
  • Read content rendered by JavaScript
  • Reproduce UI bugs and record errors
paperclipai/paperclip105.7KUpdated Oct 21, 2026

agentmail

FreeSummaryAI rating 100

Read email linked to assigned Paperclip tasks, send new messages or replies through an assigned AgentMail inbox, and check delivery outcomes. Follow task permissions and preserve retry identifiers to avoid duplicate sends.

  • Reading email context attached to an assigned task
  • Sending a new email from an assigned inbox
  • Replying to a specific message in a conversation
paperclipai/paperclip105.7KUpdated Oct 21, 2026

company-creator

FreeSummaryAI rating 100

Create an AI agent company or team package from an idea or an existing repository. Define roles, reporting lines, skills, and work handoffs, then generate structured files that follow the Agent Companies specification.

  • Build an agent company from an idea
  • Turn a repository into an agent team package
  • Define agent roles and work handoffs
paperclipai/paperclip105.7KUpdated Oct 21, 2026

complain

FreeSummaryAI rating 70

Records an agent’s brief, unfiltered complaints about tools, permissions, or working conditions in Paperclip. It preserves the reaction without requiring a diagnosis or solution, then lets the agent continue its main task.

  • Recording frustration with agent tools
  • Capturing complaints about permissions or environments
  • Preserving reactions without proposing solutions
paperclipai/paperclip105.7KUpdated Oct 21, 2026

See all skills by paperclipai

Report a problem with this skill