F: close a cancelled task's pull requests (#111)

[mchwang]

Closes #111. Stacked on #110 (base feat/103-publish-pr); retarget to main once #110 merges.

What it does

When a task is cancelled, codeboost closes every open pull request it opened for the task, draft or ready. The codeboost/… branch is kept, so a person can reopen a closed PR.

Cancel stays immediate. It is never refused and never waits for GitHub. If a publish is running, it is aborted with PublishCancelled and recorded as stopped. Once it has settled, the close runs.

PullRequestPublisher.closeAll:

• It runs only for a cancelled task, and shares the per-task publish guard.
• It first settles a lost opening or update through the publish's recovery, finding the PR by its marker. So a PR that GitHub opened during the aborted opening is recorded and closed.
• Opened records are closed by number, because GitHub's list can lag behind a PR.
• An abandoned opening's PR is found by its marker, adopted, then closed. Two PRs carrying one marker are refused, and neither is closed.
• Every PR is attempted even when one fails. The problems are reported together: OpeningUnsettled, PullRequestMisplaced or a failure.
• beforeClose re-checks the closing flags right before each close.

GhPullRequestGateway.close:

• It reads the PR by number. A PR that is already closed counts as done.
• An open PR is closed only while it is still from the task branch, in this repository, with the marker on its first line. Otherwise the gateway refuses and leaves it open.
• It then sends PATCH state=closed and requires the answer to show the PR closed.

Outcome and retries. The outcome is recorded like a publish's, with action: 'close'. A close that is still owed runs:

• at startup;
• by itself once the settle time has passed, when an opening was still unsettled;
• on the new close-pull-requests runner action, whose replay reports the outcome.

Triggers.

• Cancel action: only the action's own cancel starts a close; a replayed or refused one starts nothing.
• Cancel that stops an attempt first: the close starts when the run ends.
• Retried attempt: the close starts when that attempt settles into cancelled.
• GET /api/runner shows publish.closable.

Also fixed. In this vitest, toMatchObject({ x: /regex/ }) passes whatever x is. One existing assertion, runner-shutdown.test.ts:232, used that form, and now uses expect.stringMatching.

Decisions (user, 2026-10-03): close every open PR the task opened; keep the branch; cancel now and close afterwards; a durable outcome, retried at startup and by an action.

Tests

test/runner-publishing.test.ts (11 new) uses a real local bare remote and a fake GitHub. It covers:

• a cancel in review closes the ready PR and keeps the branch;
• a needs-human cancel closes the draft PR;
• a cancel during a draft opening is immediate, records the publish as stopped, and recovers and closes the PR;
• a failed close stays owed, the action retries it, and the replay reports the outcome;
• shutdown during a close leaves it to the next start;
• a PR the list doesn't show is closed by number;
• a replayed or refused cancel starts nothing;
• a close refused while the opening settles retries by itself and closes a PR GitHub showed late;
• a task with no PR calls GitHub for nothing, and each action refuses the other's status.

test/publish.test.ts (8 new):

• gateway: the PATCH and its read-back; never closing a PR that isn't the task's (branch, fork, null head repository, marker, number); the re-check before the close; an edited description after the close does not fail it;
• publisher: only a cancelled task's PRs are closed; close by number; a moved PR is refused and left open; nothing is closed once shutdown begins between the read and the close.

Ten breaks were tried, and each failed a test: the cancel hook, the close after a stopped publish, the owed close, the close by number, the replay check, the retry timer, the re-check before the close, and three in the gateway.

Current local counts. Typecheck passes. The server-facing files pass (333 tests: start, shutdown, cli, production, lifecycle store, publishing and publish). A full run without the two Docker files had load timeouts, with the machine's load average about 166. Each of the 9 files affected passes when run alone.

Review rounds

Round	Source	Findings	Outcome
0	self	a vacuous regex assertion hid a wrong refusal for close-pull-requests; a leftover always-true PR check; a release hook that would have looped on a refused publish	fixed
1	independent review agent	close reported closed while a PR the list didn't show stayed open; doubled-marker advice told the person to close the task's PR; an OpeningUnsettled close never retried; a replayed cancel started a close; no re-check right before the PATCH; a landed close failed by an edited description; one PR's failure stopped the others; retry-started attempts not closed; closable stayed true; a wrong settled() comment; PublishCancelled → stopped untested; missing unit and race tests	fixed
1	same	a late abandoned PR after closed, or a PR a person reopens, is not looked at again; on upgrade, already-cancelled tasks' PRs are closed; the PATCH answer serves as the read-back	documented limits, not changed
2	/code-review high	the retry settle hook re-ran owed publishes; closable re-ran mode()	fixed
3	/code-review high	none	—

Review-lesson audit. Each finding falls under an existing AGENTS.md rule:

• read-back and "a later check must not fail a succeeded irreversible action" (the gateway close);
• "Re-read immediately before an irreversible action" (beforeClose);
• the idempotency rules (the replayed cancel);
• "A test MUST fail when the feature it tests is broken" (the vacuous regex).

The vacuous toMatchObject regex is a tool quirk. It could become a new AGENTS.md rule if you want one; it is not added here.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟡 Changes recommended

Repository-mismatched PR records can be skipped while closure is recorded as successful, suppressing automatic retries.

Review effort: Balanced
Findings: 1

Open (1)

• Refuse mismatched repository records instead of silently succeeding · New

What changed in this PR

Extends task publishing to close a cancelled task’s pull requests while keeping its branch and cancellation immediate.

Changes:

• Adds guarded PR closing, opening recovery and retries.
• Records close outcomes and exposes the retry action.
• Adds regression tests and lifecycle documentation.

File	Description
web/​server.ts	Wires cancellation, retries and close status.
test/​runner-shutdown.test.ts	Corrects a regex assertion.
test/​runner-publishing.test.ts	Adds cancellation and close integration tests.
test/​publish.test.ts	Tests publisher and gateway close guards.
runner/​store.ts	Persists and replays close outcomes.
runner/​publishing.ts	Coordinates background closes and retries.
runner/​publish.ts	Recovers and closes task PRs.
README.md	Documents cancellation and close retries.
github/​pull-requests.ts	Adds ownership-checked PR closing.
docs/​implementation/​runner-lifecycle.md	Documents shutdown tracking for closes.
docs/​implementation/​pull-request-opening.md	Documents close behavior and limits.
docs/​architecture.md	Adds cancellation closing to publishing responsibilities.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Irreversible GitHub actions span cancellation, recovery, retries and shutdown, with unresolved timing and regression coverage concerns.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Schedule recovery using remaining settle-window time · New
• Add abandoned-opening closure regression tests · New

Resolved since last review (1)

• Refuse mismatched repository records instead of silently succeeding

[Copilot]

Copilot review overview

🔵 Needs a closer look

Unresolved crash-recovery behavior and gaps in lifecycle regression coverage require correction and final human review.

Review effort: Balanced
Findings: 2 · 1

Open (3)

• Continue closing PRs after an earlier close failure · New
• Retry unsettled PR closes after restart despite prior success · New
• Add deferred cancellation and retry publishing regressions · New

Resolved since last review (2)

• Schedule recovery using remaining settle-window time
• Add abandoned-opening closure regression tests