PR Autoloop
Automatically drive the agent PR maintenance loop to a terminal state after a PR is created, without a
human restarting each poll. This skill is shared by Codex and Claude Code; only the non-blocking wait
mechanism differs per agent and is absorbed by each agent guide (see "Per-Agent Wait Mechanism").
This is a bounded time-based loop with goal-based stop conditions. It polls an external system
(GitHub PR state) on an interval, reacts to changed CI/review/mergeability state, and stops only on
merge, escalation, or timeout. Do not use it as a generic autonomous coding loop; it exists for the
well-defined PR maintenance workflow.
Relationship to pr-maintainer
This skill is the autonomous looping layer. It does not define policy.
- All policy - draft gates, repair rules, repair-loop limit (4), Japanese review replies, design escalation,
merge conditions, post-merge worktree cleanup - comes from the skill and the repository's
PR maintenance ADR/policy, if one exists.
- This skill only specifies the contract for how to keep the loop running by itself until a terminal state.
If anything here appears to conflict with
,
wins on policy and this
skill wins only on "do not stall waiting for a human between polls".
When to Use
Use immediately when any of these is true:
- This agent just created a PR (), including a draft created only because workflow says so.
- A draft agent PR transitioned to ready for review (including a manual transition by the user).
- The user asks to continue/finish an agent-created PR through CI, review, repair, and merge.
First enforce the no-draft postcondition from
(ready the PR if it is reviewable), then
enter the loop. Do not stop at the PR URL when the PR is reviewable.
Autoloop Contract (agent-agnostic)
Treat maintenance as a self-paced poll loop. One loop cycle:
-
Gather state with
-
all four checks are mandatory every cycle; never skip based on CI status:
bash
# (a) PR fields - isDraft, mergeStateStatus, reviewDecision, head SHA
gh pr view "$PR" --repo "$REPO" \
--json number,isDraft,headRefOid,mergeStateStatus,reviewDecision,statusCheckRollup,labels
# (b) Check run statuses
gh pr checks "$PR" --repo "$REPO" \
--json name,state,bucket,link,startedAt,completedAt,workflow
# (c) Bot reactions on the PR issue (Codex signals: +1=clean, eyes=in-progress)
gh api --paginate "repos/$REPO/issues/$PR_NUM/reactions?per_page=100"
# (d) Inline review comments - detect P1/P2 badge findings from Codex
gh api --paginate "repos/$REPO/pulls/$PR_NUM/comments?per_page=100"
is
(e.g.
).
means GitHub merge preconditions pass; it does
not mean bot review is complete.
-
Classify the cycle into exactly one outcome:
-
continue - any of these: CI checks still running (step b has pending checks); no
chatgpt-codex-connector[bot]
reaction yet on the PR issue (step c is empty or only
); CI is green but Codex review/reaction has not posted yet → schedule the next cycle.
-
repair - any of these: CI checks failed (step b); Codex review state is
and step d contains P1/P2 badge findings → fix in the PR worktree, validate, push, reply in Japanese per
, then continue (count the repair against the limit).
-
escalate - repair limit reached, recurring findings in the same boundary, or a forbidden change is required → open the design issue, comment in Japanese, stop.
-
merge - all of the following are true (verified from step 1 data above):
- CI: all required checks pass (step b - no failed or pending).
- Draft: (step a).
- Codex reaction:
chatgpt-codex-connector[bot]
- No blocking findings: step d has no unaddressed P1/P2 Codex comments.
⚠️
with
alone does
not satisfy this gate.
→ squash merge, clean up the worktree, stop.
-
timeout - total elapsed reached the polling window with no terminal result → comment in Japanese that CI/review did not complete in the window, stop.
-
If the outcome is continue, wait the poll interval and start the next cycle automatically.
The stop condition is not a prose judgment such as "the PR looks good". It is the conjunction of
observable predicates gathered through
: CI is green, the PR is not draft, the expected bot review
artifact exists, no blocking findings remain, merge preconditions are clean, and the checked head SHA is
still current.
Loop parameters
- Poll interval: about 3 minutes between cycles.
- Polling window: at most 20 minutes total (about 6-7 cycles).
- Keep loop count and the last checked head SHA in session memory; track repair count with a todo.
- Never let the loop idle waiting for a human to manually re-trigger the next poll. The whole point of this
skill is that "continue" reschedules itself.
Per-Agent Wait Mechanism
The only agent-specific part is how you wait the poll interval without blocking your runtime. Do not
hardcode a single mechanism here; follow your agent guide:
- Claude Code → ScheduleWakeup self-pacing; fallback bounded loop; never .
- Codex → inline session polling (no ScheduleWakeup). Keep the loop inside the active session and
report the terminal state instead of relying on a platform routine to wake the agent later.
導入先の
/
に wait 機構のセクション (例: "pr-autoloop 実装") があれば
それを優先する。無ければ上のデフォルトに従い、どの機構を使ったか報告する。
Stop Conditions
Stop the loop (and report the final monitored state) on
merge,
escalate, or
timeout. Do not merge
on
continue, and do not treat an empty review/comment set as a clean review (it means "review not completed
yet" per
).
References
- skill - policy this skill executes.
- The repository's PR maintenance ADR/policy doc, if one exists (for example under ).
- 導入先の / - per-agent wait mechanism のプロジェクト上書き (任意)。