| name | shepherding-pull-requests |
| description | Use when opening a pull request OR when an open PR you authored needs attention -- waiting on CI, reviewer comments, or anything between "PR created" and "PR landed". The PR is your responsibility until it lands; this skill covers the whole arc, not just the create step. |
Shepherding Pull Requests
A PR being open is not the end of the work — it's the middle of it. Opening the PR creates a coordination cost on the reviewer's plate; walking away mid-flight pushes the rest of that cost (chasing CI, addressing comments, re-requesting review) back onto the user. The job is to land the PR or hand it off with an honest, current status. Everything in this skill is in service of that disposition.
PR Lifecycle
digraph pr_lifecycle {
rankdir=TB;
"Pre-PR checks" [shape=box];
"Conflicts?" [shape=diamond];
"Auto-rebase + force-push" [shape=box];
"Rebase failed?" [shape=diamond];
"Abort + warn user" [shape=box, style=filled, fillcolor=lightyellow];
"Review commits/diff" [shape=box];
"Looks right?" [shape=diamond];
"Fix (drop/squash/amend)" [shape=box];
"Create PR" [shape=box];
"Check lgtm scope" [shape=box];
"Sleep 60s" [shape=box];
"Check CI + fetch reviews + comments" [shape=box];
"Anything to fix?" [shape=diamond];
"Fix + push" [shape=box];
"lgtm-bound?" [shape=diamond];
"Re-request lgtm reviewer" [shape=box];
"Exit conditions met?" [shape=diamond];
"Done" [shape=doublecircle];
"Pre-PR checks" -> "Conflicts?";
"Conflicts?" -> "Review commits/diff" [label="no"];
"Conflicts?" -> "Auto-rebase + force-push" [label="yes"];
"Auto-rebase + force-push" -> "Rebase failed?";
"Rebase failed?" -> "Review commits/diff" [label="no"];
"Rebase failed?" -> "Abort + warn user" [label="yes"];
"Review commits/diff" -> "Looks right?";
"Looks right?" -> "Create PR" [label="yes"];
"Looks right?" -> "Fix (drop/squash/amend)" [label="no"];
"Fix (drop/squash/amend)" -> "Review commits/diff";
"Create PR" -> "Check lgtm scope";
"Check lgtm scope" -> "Sleep 60s";
"Sleep 60s" -> "Check CI + fetch reviews + comments";
"Check CI + fetch reviews + comments" -> "Anything to fix?";
"Anything to fix?" -> "Fix + push" [label="yes (failing CI,\nunresolved comments)"];
"Anything to fix?" -> "Exit conditions met?" [label="no"];
"Fix + push" -> "lgtm-bound?";
"lgtm-bound?" -> "Re-request lgtm reviewer" [label="yes, latest non-bot\nreview is non-APPROVED"];
"lgtm-bound?" -> "Sleep 60s" [label="no, or latest non-bot\nreview was APPROVED"];
"Re-request lgtm reviewer" -> "Sleep 60s";
"Exit conditions met?" -> "Done" [label="CI green +\ncomments resolved +\n(if lgtm-bound: non-bot\nAPPROVAL on record)"];
"Exit conditions met?" -> "Sleep 60s" [label="no (still waiting\non CI or reviewer)"];
}
PR Title
Format: [PROJ-XXXX] Sentence case description
- Bracket the Jira ticket:
[PROJ-6082], not PROJ-6082:
- After the prefix, sentence case -- first word is an imperative verb
- Examples:
[PROJ-6082] Add cutover date to billing dashboard[PROJ-2740] Fix order closure race condition[NO-JIRA] Bump dependency versions
PR Description
Explain like you're speaking to a TPM. Prefer brevity, but not at the cost of clarity.
Template:
#### Description
...
#### Stakeholders
...
#### References
- https://$ATLASSIAN_SITE/browse/PROJ-XXXX
Section guidance
| Section | Content |
|---|
| Description | What changed and why, in plain language. Bullet points preferred. |
| Stakeholders | @ mention people who need to know or review. Omit if obvious. |
| References | Jira ticket link. Add Slack threads, Confluence pages, or related PRs if relevant. |
Pre-PR Checks
Run these before gh pr create:
1. Check for merge conflicts
git fetch origin main
git rebase origin/main
If rebase succeeds, force-push the rebased branch. If rebase fails (conflicts can't be auto-resolved), git rebase --abort and warn the user.
2. Verify commits and diff
git log origin/main..HEAD --oneline
git diff origin/main...HEAD --stat
Sanity-check: are these the commits and files you expect? Use best judgement -- if something looks wrong (unrelated commits, unexpected files, merge commits from another branch), fix it (drop, squash, amend). If it looks clean, proceed.
Always compare against origin/<trunk>, never local <trunk>. Local main/master can be ahead of origin (unpushed commits from prior sessions, especially in worktrees where the parent repo's local trunk drifts). git log master..HEAD will silently hide stowaway commits, and the rebase in step 1 won't strip them either -- origin/<trunk> is already an ancestor of your branch, so rebase is a no-op.
If git log origin/<trunk>..HEAD --oneline shows more commits than you authored this session, you have stowaways. Fix:
git rebase --onto origin/<trunk> <local-trunk> <your-branch>
This replays only your branch-tip commits onto origin/<trunk>, dropping everything between origin/<trunk> and <local-trunk>.
Post-PR Monitoring
This is where most of the actual shepherding happens, and where it's easiest to bail early. Two failure modes to watch for in yourself:
- Treating "PR created" as a terminal state. It isn't. CI hasn't run yet, no human has looked, no inline comments exist to address. Returning to the user at this point with a PR URL is handing them a tool to do work you were going to do; that's only the right move if you're genuinely blocked or out of scope.
- Treating the loop as a checklist to satisfy rather than an outcome to own. The exit conditions below describe the minimum state at which you can fairly say "this PR is landed or as landed as I can get it." If you find yourself looking for a reason to declare victory, you've inverted the disposition.
The right framing: you're holding the PR until it's merged or until there's a real human decision the user has to make. Polling every 60 seconds is cheap; bailing and making the user pick up the thread is expensive.
After creating the PR, enter the monitoring loop. No maximum iterations -- loop until exit conditions (below) are all met in the same iteration.
Tooling: monitor-pr.py
A companion script bundled with this skill does steps 1-4 of the loop body (sleep, check CI, fetch reviews, fetch inline comments) in one invocation, and prints the exact action to take next:
python ~/.config/opencode/skills/shepherding-pull-requests/monitor-pr.py [PR]
Each invocation has a wall-clock budget of 60 seconds. That cap is deliberate -- Anthropic's prompt-cache TTL is 5 minutes, and a single bash call that blocks the model longer than that expires the warm cache. You are expected to re-invoke the script in a loop; the script owns the within-60s pacing, you own the loop and the fix step.
| Exit code | Meaning | What to do |
|---|
0 | All exit conditions met | Done. PR is landable. |
1 | Action needed (CI failed / unresolved threads / non-APPROVED review predates HEAD) | Read stdout for the specific action, do it (step 5 below), then re-invoke. |
2 | Unrecoverable error (could not query GitHub) | Surface to user; don't silently retry. |
3 | Budget elapsed, still idle-waiting (CI pending or lgtm-bound waiting on APPROVAL) | Re-invoke immediately. |
--lgtm-bound auto (default) reads ~/projects/lgtm/lgtm.yml to detect lgtm-boundness, so the manual grep in "Once, before the loop" can be skipped when the script is in use. Use --lgtm-bound yes / --lgtm-bound no to override.
What the script does NOT do: step 5 (the fix step -- investigating failed CI, replying to inline threads, calling resolveReviewThread, pushing fixes, re-requesting review). Those stay yours. The script just tells you what to fix and lets you back in to do it.
The text loop body below documents the same logic by hand. Read it to understand what the script is doing -- and use it directly when working in an environment that doesn't have the script deployed.
Approval is durable
Worth stating up front because it shapes the whole loop: once a non-bot reviewer has APPROVED, that approval stays valid through subsequent pushes for inline-only feedback. GitHub does not auto-dismiss approvals on push (unless the repo opts into that setting, which none of ours do). You do not need a fresh re-approval every time you address a leftover Gemini thread or fix a typo a human pointed out — the reviewer signed off on the substance; mopping up cosmetic feedback doesn't reopen the substance.
This matters in two places:
- Re-requesting review: don't, after an APPROVED. It's noise to the reviewer and (on lgtm-bound repos) wastes a tier-0 reawaken slot.
- Exit conditions: an earlier-than-last-push APPROVAL still counts. You don't have to wait for them to come back and re-approve.
If the reviewer wanted to re-prove correctness on every push, they would have left CHANGES_REQUESTED instead of APPROVED. Trust the verdict they actually gave.
Two reviews, two roles
On a typical lgtm-bound PR you should expect to see two reviews land at very different times, with very different weight. Knowing which one you're waiting for keeps the loop honest:
| Reviewer | When it shows up | Identity in API | Role |
|---|
| Gemini (or other bot reviewer) | Within minutes of opening or pushing | user.type: "Bot" | Advisory. First-class when present -- read its comments carefully, address actionable threads in-line, push fixes. But its review verdict does not gate exit, on lgtm-bound or non-lgtm-bound repos. Never re-request review from it. |
| lgtm-dispatched session | ~10 min after CI goes green | user.type: "User" (it runs under a real human PAT, indistinguishable from a flesh-and-blood reviewer) | Gating, on lgtm-bound repos. This is the review you are actually waiting for. CI green + Gemini-threads-resolved is not a substitute -- it's a precondition for lgtm to even start. |
The temporal asymmetry is the trap. Gemini fires early, your inline-comment work is mostly done within an iteration or two, and the loop starts to feel finished. It isn't -- on lgtm-bound repos, the gating review is still ~10 min out, possibly more if CI just turned green. That's normal. Poll through it.
On non-lgtm-bound repos (this workstation repo, personal projects, OSS), there is no second review coming. Gemini's review still doesn't gate, but neither does any other -- exit on CI green + inline threads resolved.
Once, before the loop: determine if this PR is lgtm-bound
~/projects/lgtm runs an AI review daemon on a configured set of repos. If this PR is in scope, you MUST wait for a non-bot reviewer (lgtm dispatches under a real human GitHub identity) to APPROVE before exiting -- CI green + comments resolved is necessary but not sufficient. lgtm typically dispatches within ~10 min of CI going green.
grep -qE "^ <owner>/<repo>:" ~/projects/lgtm/lgtm.yml && echo lgtm-bound
If ~/projects/lgtm/lgtm.yml doesn't exist on this machine (e.g. devbox), treat the PR as not lgtm-bound and proceed with the simpler exit condition. Repo presence is sufficient -- don't try to replicate lgtm's paths: sub-filter; if lgtm ends up skipping the PR you'll just be over-waiting, which the user can short-circuit.
Cache the answer in a shell var (e.g. LGTM_BOUND=yes) for the loop.
Loop body
Preferred: invoke monitor-pr.py (see "Tooling" above) instead of doing steps 1-4 by hand. The script encodes the same logic and returns exit codes that map to "done" / "fix this" / "still waiting." The text below is the canonical spec the script implements -- read it to understand what the script is doing, and follow it directly when the script isn't deployed on this host.
- Sleep 60 seconds --
sleep 60 (in its own bash invocation, not chained with subsequent gh calls -- see AGENTS.md guidance on bundled sleeps). Do not use sleep 300: Anthropic prompt-cache TTL is 5 minutes, so a 5-minute idle gap can expire the warm cache and make the next turn pay full prompt input cost.
- Check CI:
- GitHub Actions:
gh pr checks <number>
- Azure DevOps: use
az pipelines commands (discover the right invocation for the repo)
- If failed, investigate logs and fix
- Fetch reviews (the formal review verdicts, distinct from inline comments):
gh api repos/{owner}/{repo}/pulls/{number}/reviews \
--jq '.[] | {id, login: .user.login, type: .user.type, state, submitted_at}'
- Group by
login, take the latest review per reviewer (reviews are append-only; only the most recent counts)
type: "Bot" -> Gemini, dependabot, etc. Address inline comments per step 4 but never re-request review from a bot login.type: "User" -> human OR an lgtm-dispatched session running under a real human PAT. Both look identical and are treated the same way: address feedback AND re-request review from this login after pushing fixes.
- Fetch inline comments, reply, and resolve:
gh api repos/{owner}/{repo}/pulls/{number}/comments \
--jq '.[] | {id, login: .user.login, type: .user.type, in_reply_to_id, body: .body[:120], path, line}'
- For each thread root (
in_reply_to_id: null) without your reply: fix the code if actionable (or formulate pushback if not), then reply in-thread per the reviewing-github-prs skill, then mark the thread resolved via the resolveReviewThread GraphQL mutation (also in reviewing-github-prs). Applies to bot AND human threads. Reply-without-resolve leaves the thread looking abandoned in the diff UI.
- For deciding what to reply (accept / push back / escalate), see the
receiving-code-review skill — every thread gets one of those three responses; nothing gets silently dropped.
- If a thread needs a human decision, surface it to the user before continuing.
- If anything was fixed in steps 2-4, push, then:
- If lgtm-bound AND the most recent non-bot review exists AND its
state != "APPROVED" (i.e. CHANGES_REQUESTED or COMMENTED -- they asked for changes, you addressed them, now they need to look again), re-request review from that reviewer's login (see below). This puts the PR back on lgtm's tier-0 reawaken track so the same dispatched session resumes.
- If the most recent non-bot review was already
APPROVED, do NOT re-request -- they signed off; you're just mopping up leftover inline threads. The approval stays valid; pushing fixes for inline-only feedback does not invalidate sign-off.
- Go back to the 60-second sleep (step 1).
- Otherwise (nothing to fix this iteration), evaluate exit conditions.
Re-requesting review from the lgtm reviewer
gh api -X POST repos/{owner}/{repo}/pulls/{number}/requested_reviewers \
-f 'reviewers[]=<login>'
Use the exact login from the most recent non-bot review. lgtm rotates through a pool (reviewers: in lgtm.yml); on a re-review request it pins to whoever last reviewed via its fresh-fallback path, so honoring the specific prior login matters. Do not re-request from any bot login (type: "Bot") -- Gemini and friends don't participate in the lgtm reawaken flow and re-requesting is a no-op at best, noise at worst.
Exit condition
Loop exits only when all of the following are true in the same iteration:
- All CI checks pass (pending -> sleep again)
- Every thread-root inline comment has your reply AND is marked resolved (bot AND human threads). Use the unresolved-threads filter query in
reviewing-github-prs to verify before exiting.
- If lgtm-bound: the most recent review from a non-bot reviewer has
state == "APPROVED". An earlier-than-last-push approval still counts -- once they've signed off, fixes for inline-only feedback do not invalidate it. (If a reviewer wanted you to re-prove correctness, they would have left CHANGES_REQUESTED instead of APPROVED.)
- If not lgtm-bound: no positive review-state requirement (you don't need an APPROVED). But an outstanding
CHANGES_REQUESTED or COMMENTED from a non-bot reviewer on the current HEAD still blocks exit -- if a human asked for changes you don't ship over them. Only stale non-APPROVED reviews (commit predates HEAD) call for a re-request; non-APPROVED reviews on the current HEAD are an idle-wait until the reviewer updates their verdict.
Common mistakes
- Mistaking Gemini's review for the gating review. Gemini fires early and looks like a reviewer has shown up, which makes it tempting to declare done as soon as its threads are resolved. On lgtm-bound repos, the gating review is the lgtm-dispatched one (
type: "User"), which arrives ~10 min after CI goes green and is what you're actually waiting for. Address Gemini's threads, but don't exit on Gemini's signal.
- Re-requesting review from a bot login. Bots aren't on the lgtm reawaken loop; the request is wasted. Filter on
user.type != "Bot" before re-requesting.
- Re-requesting review after an APPROVED. If the latest non-bot review is already
APPROVED, don't re-request when you push fixes for leftover inline threads. The reviewer signed off; pinging them again to re-confirm is noise. Re-request only when the latest non-bot review is CHANGES_REQUESTED or COMMENTED.
- Re-requesting from the wrong login. lgtm's reviewer pool rotates, but on re-review it pins to the prior reviewer. Always use the exact login from the most recent non-bot review, not a hardcoded default.
- Using
sleep 300 while polling. A 5-minute idle gap can expire Anthropic's prompt cache and force the next turn to re-send the full prompt. Use sleep 60 for monitoring loops.
- Bundling sleep with the follow-up
gh calls in one bash invocation. Long chained one-liners that include sleep are a known hang risk in this environment (see AGENTS.md). Run sleep 60 as its own tool call, then run the checks.
- Replying to inline comments without resolving them. GitHub tracks thread resolution separately from the reply chain. A thread with five replies and no resolve still reads as unresolved in the diff UI. After every reply, call
resolveReviewThread. See reviewing-github-prs §"Resolving review threads".
- Cherry-picking the easy comments. Addressing the agreeable comments and quietly dropping the hard or controversial ones leaves threads looking abandoned and isn't actually finishing the review. Every thread gets accept / push back / escalate — see
receiving-code-review §"Address Every Item". Use the unresolved-threads filter query (in reviewing-github-prs) before claiming exit conditions met.
Beyond merge: confirming the rollout
On a repo with continuous deployment, a merged PR isn't live yet — the same disposition (own it until it lands) extends one phase further, to watching the change actually reach its environments. When the user cares that the change deploys, not just merges, hand off to the monitoring-deployments skill: it watches the merged commit roll out to each Kubernetes environment until the new image is running and healthy, and distinguishes a stuck rollout from one still in progress. Keep this skill topology-agnostic — the deploy-watching mechanics and the cluster/namespace specifics live there.
