Paperclip에서 실제 브라우저부터 서버, 데이터베이스, runner까지 검증하는 E2E 평가를 추가·확장합니다. 테스트 데이터와 판정 조건, 증거를 기존 보고서에 연결합니다.
- 전체 사용자 작업 흐름의 평가 추가
- 테스트 데이터와 판정 조건 갱신
- 평가 보고서에 스크린샷과 산출물 추가
GitHub, GitLab, Perforce 변경의 검토 의견, 자동 검사 실패와 설명 누락을 확인하고, 할 일을 정리하며 요청에 따라 수정과 처리된 스레드 해결을 지원합니다.
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.
GitHub PR, GitLab MR, Perforce CL을 살펴보고 검토 의견, 자동 검사 실패와 변경 설명의 누락을 정리합니다. 조치가 필요한 항목, 참고 정보, 이미 해결된 항목으로 나눈 뒤 이유와 다음 조치를 표로 보고합니다. 요청에 따라 코드나 설명을 수정하고, 수정 사항 반영과 처리된 검토 스레드의 해결까지 지원합니다.
변경 사항을 제출하기 전에 점검하거나 검토 후 남은 작업을 파악할 때 유용합니다. 사람이 남긴 댓글뿐 아니라 봇의 지적과 편집된 요약도 확인할 수 있습니다. 번호를 생략하면 현재 브랜치 등의 정보를 바탕으로 대상을 찾습니다.
git과 대상 플랫폼에 맞게 인증된 gh, glab, p4가 필요합니다. GitHub에서는 현재 커밋을 대상으로 정상 완료된 Greptile 검사가 필수이며, 이전 결과로 점검을 완료하지 않습니다. GitLab과 Perforce의 Greptile 연동도 결과가 최신인지 확인합니다. 라이선스는 MIT입니다.
이 소개문은 공개된 SKILL.md를 바탕으로 AI(OpenAI Codex)가 작성했습니다. 정확한 사양은 아래 원문을 확인하세요.
AI가 SKILL.md의 내용만을 근거로 4가지 관점에서 판정했습니다. 스타 수나 인기도는 판정에 넣지 않았습니다.
총점은 각 항목을 A=25점, B=15점, C=5점으로 더한 값입니다. 2026년 10월 20일 기준으로 AI(OpenAI Codex)가 판정했습니다. 참고용이므로 설치 전에 원문도 확인해 주세요.
설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.
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.
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.
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 $P4CLIENT
Key field differences between platforms:
number, headRefName, headRefOidiid, source_branch, shashelved files for in-review CLsGitHub:
gh pr view <PR_NUMBER> --json title,body,state,reviews,comments,headRefName,headRefOid,statusCheckRollup
OWNER_REPO=$(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 numberStatus: pending, submitted, shelvedDescription: the CL description / commit messageFiles: list of files in the CLBefore 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.
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.
Once all checks are complete, evaluate these areas:
greptile-apps[bot] on GitHub, or the Greptile bot user on GitLab, linters, etc.)p4 review or external review toolsupdated_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.For each issue found, categorize as:
| Category | Meaning |
|---|---|
| Actionable | Code changes, test improvements, or fixes needed |
| Informational | Verification notes, questions, or FYIs that don't require changes |
| Already addressed | Issues that appear to be resolved by subsequent commits |
Present a summary table:
| Area | Issue | Status | Action Needed |
|---|---|---|---|
| Status Checks | CI build failing | Failing | Fix type error in src/api.ts |
| Review | "Add null check" - @reviewer | Actionable | Add guard clause |
| Description | TODO placeholder in test plan | Actionable | Fill in test plan |
| Review | "Looks good" - @teammate | Informational | None |
If there are actionable items:
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>
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.
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
Summarize:
아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.
로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.
아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.
Paperclip에서 실제 브라우저부터 서버, 데이터베이스, runner까지 검증하는 E2E 평가를 추가·확장합니다. 테스트 데이터와 판정 조건, 증거를 기존 보고서에 연결합니다.
Paperclip Runner의 동작 평가에 판정 조건, 실행 대상, 보고서 검증 데이터를 추가하고 증거와 실행 이력을 남겨 검증합니다.
실제 브라우저로 웹페이지와 앱을 조작하며 화면과 동작을 검증합니다. 스크린샷, 간단한 양식 입력, 오류 및 통신 기록으로 문제 재현과 결과 확인을 돕습니다.
Paperclip에서 할당받은 AgentMail 수신함으로 작업에 연결된 메일을 확인하고, 새 메일 전송과 답장, 전송 결과 확인까지 수행합니다.
구상이나 기존 저장소를 바탕으로 AI 에이전트의 역할, 보고 체계, 작업 인계를 정리하고 규격에 맞는 회사 또는 팀 패키지를 만듭니다.
에이전트가 도구, 권한, 작업 환경에 느낀 불만을 짧은 자유 형식 글로 Paperclip에 기록합니다. 해결책 없이 당시의 반응을 남기고 원래 작업을 이어갑니다.