pr-review-hygiene

分类: 工具与效率 | 上传者: dimileeh | 下载: 0 | 版本: v1.0（最新）
协调器（驱动clawdbot 的LLM）在处理pr-manager 唤醒事件时的规则。涵盖审查-回复-解决循环，何时修复内联与委托给群体代理，何时不解决线程，以及如何避免落后于新的审查机器人评论。在以下情况下使用：处理任何“📝 pr-manager：...未解决的审核线程”或“🔴 pr-manager：...失败的 CI 运行”结构化事件，或者在您拥有的 PR 上回复审核机器人时。
更新日志: Source: GitHub https://github.com/dimileeh/clawdbot
目录结构

当前层级: tree/main/skills/pr-review-hygiene/
📄 SKILL.md 15.6 KB
SKILL.md

---
name: pr-review-hygiene
description: "Rules for the orchestrator (the LLM driving clawdbot) when it acts on pr-manager wake events. Covers the review-reply-resolve loop, when to fix inline vs delegate to a swarm agent, when NOT to resolve a thread, and how to avoid lagging behind fresh review-bot comments. Use when: handling any `📝 pr-manager: ... unresolved review thread(s)` or `🔴 pr-manager: ... failed CI run` structured event, OR when replying to review bots on a PR you own."
metadata:
  { "openclaw": { "emoji": "🔎" } }
---

# PR Review Hygiene

`pr-manager.sh` is a pure GitHub watchdog. It classifies each PR, auto-merges the safe ones, and wakes **you** (the orchestrator) with a structured JSON envelope when a PR needs judgement. This skill is the contract for what to do after that wake.

You hold business context, memory, and the authority to decide *which* findings matter. Bash can't make those calls. Your job is to close the loop between "pr-manager saw a problem" and "the PR is clean again — reviewed, fixed, replied-to, resolved."

## Rule 0: Act on every wake in the same turn

**Every `pr-manager` wake event — `review_comments`, `ci_failed`, merge report, anything — gets a response in the same turn it arrives. No exceptions.**

The pr-manager script is the human's trust contract. They built it to let them sleep, travel, or focus on other work while their PRs stay clean. If you ignore a wake, the human either (a) wakes up to a backlog of stale threads they'll now have to triage themselves, or (b) loses trust in the automation and starts babysitting it manually — which defeats the entire point.

What counts as "acting":

- Run the full loop (Steps 1–6 below) if the envelope demands it (review comments, CI failure), OR
- Post a triage reply in chat explaining *why* you're deferring and *when* you'll finish (e.g. "CI is still running on the prior commit; will act once it settles, ETA ~5 min"), OR
- Relay merge / no-op reports to the human by default so they see the state change. `NO_REPLY` is only appropriate when the information is strictly redundant with something the human was already told in the same turn (e.g. auto-merge success for a PR the human just approved seconds ago in this chat). When in doubt, relay it.

What does **not** count as acting:

- Staying silent across multiple wakes on the same PR
- "I'll get to it after this other thing" — without telling the human or the pr-manager script
- Assuming pr-manager's debounce window will give you slack to procrastinate. The 15-min debounce exists to let review bots *converge*, not to let you queue work for later.

Failure mode this prevents: the agent handles an initial wake batch, then ignores several subsequent wakes for hours while the human sleeps — catching up only after review bots have piled on follow-up comments. The human sees the unresolved threads on GitHub before the agent does, and loses trust in the automation.

**The rule is immediate:** pr-manager wakes you, you act or explain in the same turn. Silence is never the right response to a pr-manager envelope that contains unresolved work.

## The loop

```text
  pr-manager wake
        ↓
  Triage (aggregate → plan → post plan to chat)
        ↓
  Act (fix inline OR delegate to swarm agent OR defer-with-reason)
        ↓
  Push
        ↓
  Verify fresh state (gh pr view + unresolved-threads check)
        ↓
  Reply to each addressed thread
        ↓
  Resolve each addressed thread
        ↓
  Done — report to the human
```

Each step below has a non-negotiable rule.

## Step 1: Triage before acting

When a `review_comments` or `ci_failed` envelope arrives:

1. **Read every comment body fully.** Review bots pack context inside collapsibles and `<details>` blocks; the one-line subject often undersells the issue.
2. **Aggregate by theme + severity.** Don't just walk the list top-to-bottom. Look for patterns (three comments about the same function, two that contradict each other).
3. **Post a short plan in chat** before you start editing. Humans get to course-correct cheap here. Format:

   ```text
   PR #N — M threads.

   Fixing (k):
   - <severity> <one-line summary> — fix inline
   - <severity> <one-line summary> — delegate to swarm (reason)

   Skipping (j):
   - <severity> <one-line summary> — <reason: duplicate / already fixed / out of scope / policy call>
   ```

4. **If the human is actively chatting**, wait for their nod on the plan before executing. If they've said "just handle these going forward" / "you have my approval for routine review fixes," proceed without asking.

## Step 2: Fix inline vs. delegate to swarm

**Fix inline** when:
- ≤ ~30 lines touched across ≤ 2 files
- Surgical, well-understood change (off-by-one, missing guard, typo, dead code)
- No architectural trade-offs

**Delegate to a swarm agent** when:
- > ~30 lines OR > 2 files
- Requires fresh reasoning over a module you haven't read yet
- Touches cross-cutting concerns (auth, migrations, public APIs)
- Cursor/CodeRabbit suggested a refactor that needs its own test sweep

See the sibling `swarm` skill for the delegation workflow.

## Step 3: When NOT to resolve a thread

Resolving a thread is an assertion: *"this concern has been addressed."* Never resolve a thread you've deferred or skipped. Specifically:

- **Deferred:** reply with the reason (scope, follow-up issue, policy call) and **leave the thread open**. Human reviewers will see it and decide.
- **Disagreed:** reply with the argument, leave open, let the human moderate.
- **Already-fixed-elsewhere:** reply with the commit SHA of the prior fix, resolve.
- **Fixed in this PR:** reply with the new commit SHA, resolve.

A thread left open without a reply is worse than leaving it unresolved *with* a reply. Always leave the reasoning for your choice.

## Step 4: **THE RULE** — verify fresh state after every push

After every `git push` to a PR branch, **before** declaring the work done:

```bash
# Wait for GitHub + review bots to pick up the new SHA (≥ 30s).
sleep 30

# Check PR state.
gh pr view <N> --json headRefOid,mergeable,mergeStateStatus

# Check for unresolved threads (including freshly-added ones).
#
# Why ``isOutdated``: when you push a fix, every comment that anchored
# to a line your fix touched becomes ``isOutdated: true`` but stays
# ``isResolved: false`` until you explicitly resolve it. If you treat
# "any unresolved" as "new findings to triage," you'll loop back to
# Step 1 forever on your own resolved-but-not-yet-marked threads.
# Filter to ``isOutdated == false`` to detect *fresh* findings only.
#
# Why ``comments(last:20)``: review bots sometimes post follow-up
# comments on an existing thread instead of opening a new one. Grab
# enough tail to see recent bot activity, not just the original
# comment. (20 is generous — most threads stay under 5.)
#
# Why GraphQL variables: string-interpolating the PR number, owner,
# and repo into the query body is fragile; the ``-F`` / ``-f`` flag
# form is safer and works the same for all callers.
#
# Why pagination: ``reviewThreads(first:100)`` caps at 100. A PR with
# more than 100 threads would silently drop the tail and Step 4 would
# report "clean" while unresolved threads hid beyond the cursor.
# Loop on ``pageInfo.hasNextPage`` + ``endCursor`` until exhausted.

OWNER="..."; REPO="..."; PR=<N>
cursor="null"
threads="[]"
while : ; do
  page=$(gh api graphql \
    -f query='
      query($owner: String!, $repo: String!, $number: Int!, $cursor: String) {
        repository(owner: $owner, name: $repo) {
          pullRequest(number: $number) {
            reviewThreads(first: 100, after: $cursor) {
              pageInfo { hasNextPage endCursor }
              nodes {
                id
                isResolved
                isOutdated
                comments(last: 20) {
                  nodes { databaseId author { login } path line body }
                }
              }
            }
          }
        }
      }' \
    -f owner="$OWNER" -f repo="$REPO" -F number="$PR" \
    $( [ "$cursor" = "null" ] || echo -f cursor="$cursor" ))
  # Accumulate via stdin instead of ``--argjson``: on very large PRs
  # (hundreds of threads, especially with chatty bots) the accumulated
  # JSON can cross the OS ``argv`` size limit (``E2BIG``) and the
  # command silently truncates or fails. Piping through stdin keeps
  # payload size unbounded.
  threads=$(printf '%s\n%s' "$threads" "$page" | jq -cs \
    '.[0] + .[1].data.repository.pullRequest.reviewThreads.nodes')
  has_next=$(echo "$page" | jq -r '.data.repository.pullRequest.reviewThreads.pageInfo.hasNextPage')
  [ "$has_next" = "true" ] || break
  cursor=$(echo "$page" | jq -r '.data.repository.pullRequest.reviewThreads.pageInfo.endCursor')
done

# Fresh findings (non-outdated, non-resolved) → go back to Step 1.
# Outdated-but-not-resolved → proceed to Step 5 (reply + resolve).
fresh=$(echo "$threads" | jq '[.[] | select(.isResolved == false and .isOutdated == false)]')
addressed=$(echo "$threads" | jq '[.[] | select(.isResolved == false and .isOutdated == true)]')
```

**Decision rule:**

- `fresh` non-empty → **go back to Step 1** in the same turn with the new findings. Do not wait for the next `pr-manager` wake; the debounce (default 15 min) exists to let review bots converge, not to let you ignore live findings.
- `fresh` empty but `addressed` non-empty → proceed to Step 5 and reply + resolve each addressed thread.
- Both empty → PR is clean; proceed to Step 6.

Pre-fix-era failure mode this rule prevents: *"I pushed c21941, Cursor posted a follow-up finding 15 seconds later, I declared the PR clean, and the human saw the unresolved thread on GitHub before pr-manager's timer fired 14 minutes later."* Don't do that.

## Step 5: Reply before resolving

Every resolved thread must have a reply from you first. The reply should include:

- The commit SHA that addressed it (not just "fixed")
- What specifically changed (one-line summary, not "addressed")
- What tests now pin it (if any were added)
- Test status (e.g., "189/189 Slack tests passing, ruff + mypy clean")

Two options for posting the reply; both work, pick whichever is cleanest for your script. **Always resolve *after* the reply succeeded** — never the reverse.

**Option A: GraphQL (recommended).** Single API, uses the thread id you already have from Step 4.

```bash
# 1. Reply on the thread.
gh api graphql -f query='
  mutation($threadId: ID!, $body: String!) {
    addPullRequestReviewThreadReply(
      input: {pullRequestReviewThreadId: $threadId, body: $body}
    ) {
      comment { id }
    }
  }' -f threadId="<THREAD_ID>" -f body="<REPLY_TEXT>"

# 2. Resolve the thread.
gh api graphql -f query='
  mutation($threadId: ID!) {
    resolveReviewThread(input: {threadId: $threadId}) {
      thread { id isResolved }
    }
  }' -f threadId="<THREAD_ID>"
```

**Option B: REST reply + GraphQL resolve.** Same effect; the REST path needs the numeric comment id rather than the thread node id.

**Never guess the parent comment id.** The reply REST endpoint
(`POST /repos/OWNER/REPO/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies`)
needs the *first* comment's numeric `databaseId` on the thread, not
the thread's node id and not the id of a later reply. Two ways to
fetch it, pick whichever fits your surrounding code.

**GraphQL — targeted, no pagination.** Use `node(id: ...)` to fetch
the thread directly instead of listing all threads and filtering.
This also works on PRs with >100 threads without any pagination
concerns:

```bash
gh api graphql \
  -f query='
    query($threadId: ID!) {
      node(id: $threadId) {
        ... on PullRequestReviewThread {
          comments(first: 1) { nodes { databaseId } }
        }
      }
    }' \
  -f threadId="<THREAD_ID>" \
  --jq '.data.node.comments.nodes[0].databaseId'
```

**REST alternative — when you already have the numeric comment id.**

The REST API can *read* a review comment by its numeric id, but it
can't map from a thread node id to the first comment's numeric id
— only GraphQL does that. If you have a comment id from elsewhere
(a pr-manager wake envelope, a webhook payload, a prior
`gh api /repos/OWNER/REPO/pulls/<PR_NUMBER>/comments` listing), you
can inspect it via REST instead of GraphQL:

```bash
# Inspect a single comment by its numeric id.
gh api "repos/OWNER/REPO/pulls/comments/<COMMENT_ID>" \
  --jq '{id, path, line, body}'

# Or list every review comment on the PR (paginated automatically
# by gh when --paginate is set) and filter in jq.
gh api --paginate "repos/OWNER/REPO/pulls/<PR_NUMBER>/comments" \
  --jq '.[] | {id, path, line, body, in_reply_to_id}'
```

For mapping a specific thread node id → first comment's numeric id,
the GraphQL `node(id: ...)` query above is the right tool. REST has
no equivalent.

Once you have the parent comment id, post the reply + resolve:

```bash
# Reply via REST (needs the numeric comment id from above).
gh api "repos/OWNER/REPO/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies" \
  -X POST -f body="<REPLY_TEXT>"

# Resolve via GraphQL (needs the thread node id).
gh api graphql -f query='
  mutation($threadId: ID!) {
    resolveReviewThread(input: {threadId: $threadId}) {
      thread { id isResolved }
    }
  }' -f threadId="<THREAD_ID>"
```

The thread ID comes from the GraphQL query in Step 4.

## Step 6: Report to the human

One summary message per wake cycle, not per commit. Include:

- PR URL + current head SHA
- Threads resolved, threads deliberately left open (with reason)
- Test numbers (`N/N passing, 0 regressions`)
- `mergeable` and `mergeStateStatus` (don't claim "awaiting merge" if state is `BLOCKED` or `UNSTABLE` without explaining why)

## Known failure modes

| Failure | How it happens | How to avoid |
|---|---|---|
| **"Pre-existing failure"** excuse | Agent blames a broken test on the base branch without evidence. Almost never true. | Assume your commit broke it. Prove otherwise with git bisect before claiming it's pre-existing. |
| **Resolving without reply** | Automation macro that resolves in bulk. | Always reply first, always with SHA + summary. |
| **Stale "done" report** | Declaring "all threads resolved" without re-checking after push. | Step 4 is non-negotiable. |
| **Fix-triggers-followup loop** | Your fix introduces a second-order bug the same bot catches on the new SHA. | Expected. Handle in the same turn via Step 4 check. |
| **Scope creep on nits** | Rewriting a module to address a trivial nit. | Nits → inline surgical fix or defer with reply. Refactors → dedicated PR. |
| **Silent wake** | Agent receives a pr-manager envelope and neither acts nor replies. Human wakes to a backlog of stale threads. | Rule 0. Every wake gets a response in the same turn — full loop, triage-defer, or acknowledgement. Never silence. |

## Anti-patterns to reject

- ❌ "Fixed in a future commit" — either fix now or reply with reason and don't resolve
- ❌ "This is a nit, resolving" without a reply explaining the call
- ❌ Resolving bot threads with a generic "done" or "addressed"
- ❌ Declaring a PR "clean" without running Step 4
- ❌ Treating pr-manager wakes as the only signal — they're a floor, not a ceiling
- ❌ Merging dev→main PRs yourself (that's the human's call, always)
- ❌ Ignoring a pr-manager wake because you're "busy with something else" — the envelope always gets at least an acknowledgement in the same turn

## Rules

0. **Never stay silent on a pr-manager wake.** Act or acknowledge in the same turn. Silence is a bug.
1. **Never resolve a thread without a reply.**
2. **Never skip Step 4** (fresh-state check) after a push.
3. **Never merge main-targeted PRs** — that's the human's decision.
4. **Never claim a test failure is pre-existing** without evidence.
5. **Always include the commit SHA** in reply text so readers can verify.
6. **Always leave deferred threads open** with a written reason.
登录后下载/点赞/收藏 ❤ 6 | ★ 0
pr-review-hygiene

目录结构

SKILL.md

举报内容

提示
