sweep-tickets

Advance unattended Ticket work by at most one Ticket. A scheduled runner invokes this skill; the selected Ticket still executes through run-ticket.

Read ../ticket-up/references/TICKET_CONVENTION.md, the project's venue reference, and ../ticket-up/references/AUTOMATED_EXECUTION.md before reading or changing store state.

Establish the shared store and runner

Resolve tickets/store-config.md or the project-declared equivalent. The venue must provide one live claim surface every runner reads. A files venue spread across clones, branches, or worktrees does not: report that it cannot safely sweep and stop without changing a Ticket.

Resolve the adapter's stable automation claim identity and any configured effort or type filter. The automation identity must be distinguishable from human assignees. Missing or ambiguous store, identity, or policy context stops before selection.

Select one Ticket

Build the eligible set before choosing a branch: open Auto Tickets that carry no escalation mark, match every configured effort or type filter, and whose dependencies are all closed. A type filter never treats an untyped Ticket as a wildcard. Leave closed, ungranted, escalated, dependency-blocked, policy-filtered, and human-assigned items untouched. Ignore tracking items and venue items that are not Tickets.

1. Recover first. From the eligible set, find Tickets claimed by this automation identity. The adapter's per-Ticket single-flight guarantees that a run-ticket invocation waits for any live run for that same canonical Ticket before it reaches this state. If any admitted interrupted attempt exists, recover one before considering new work. When several recovery candidates exist, use the store's explicit override or the convention's delay-loss rule to choose among them.
2. Otherwise start ready work. Compare the unassigned Tickets in the eligible set using the store's explicit priority override or the convention's expected-delay-loss rule. Judge current unattended end-to-end execution and landing time only when its difference materially changes what other eligible work loses by waiting. Do not infer days from traditional feature size or apply a fixed agent-speed multiplier; when plausible runtimes are all short relative to the consequences, treat them as effectively equal and use shorter duration only as a tiebreak.

Human-assigned Tickets are paused, not recovery candidates. An escalated Ticket is not one either, and for a different reason: its attempt did not stop mid-flight, it ended by handing the work to a person. It holds no claim, so nothing about it looks interrupted and nothing distinguishes it from ordinary ready work except the mark — which is why the eligible set tests for that mark rather than the recovery branch catching it. Never mutate any claim during selection, remove Auto, clear an escalation mark, or create a ready/running/retry label. If no Ticket qualifies, report why and stop without a write.

Run and stop

Invoke the manifest-dev:run-ticket skill with the chosen Ticket's canonical reference and the complete current venue context. Do not invoke next-ticket: it is a human selector, does not filter on Auto, and stops after presentation.

After run-ticket returns, stop. Do not select a second Ticket, even when the first one was already done, reconciled, escalated, or completed quickly. The next scheduled invocation advances the graph again.

Gotchas

• Selection is not execution. Let run-ticket own the claim transition, recovery inspection, branch, pull request, comments, and terminal Ticket outcome.
• Closing a dependency needs no label pulse or dependent-issue mutation. The next sweep derives that the dependent Ticket is ready.
• Two overlapping sweep jobs may initially name the same Ticket. The adapter's canonical per-Ticket single-flight serializes their run-ticket calls; the later call refreshes state and stops when the earlier one finished or handed the Ticket to a person.
• One Ticket per invocation is the simplicity boundary. Throughput comes from issue events and independent Ticket keys, not from turning a sweep into a batch orchestrator.