delivery-wait

Delivery wait

Run scripts/delivery_wait.py whenever a workflow must await external GitHub state. The harness passively waits for the command to finish; it must not simulate waiting with model heartbeats.

Running a wait

Hosts limit how long one command may run, and a delivery transition can take longer. These rules keep a wait armed for as long as the transition needs:

• If the host can run a command in the background and resume the session when it exits, run the wait that way. Then end the turn, or continue only work independent of the awaited state. If that capability is absent, use the fallback below; never hold the turn open with a blocking wait.
• Set each invocation's --deadline inside the host's limit for one command, and keep the time the caller gives up on the transition separately. A timed-out result before that time means: run the wait again with the same --state. That resumes one wait; it is not a heartbeat.
• Run the helper as the whole command, with nothing piped or chained after it, so the exit status the host reports is the helper's own. A host notice that the command finished says only that it ended; the result envelope says what happened.
• Keep one running wait per state file. Never detach a wait from the host, for example with setsid, nohup or &: a detached process cannot resume the session.
• While a wait runs, do not inspect the state it covers. Act on its result.

The address-feedback CodeRabbit adapter's run follows the same rules. It keeps no state file: run it again with the same arguments, and it re-reads the head's state from GitHub.

Long commands and delegated lanes

Apply the same turn discipline to long builds, tests and other commands, in coordinators and workers alike. Start commands expected to outlast a short check through the host's tracked background execution. Keep the task or session ID, output location, exact revision, completion condition and resume action in the handoff. A launch acknowledgment is pending work, never passing evidence. Short commands and bounded diagnostic checks can stay in the foreground.

Confirm that the command survives the owning lane's turn and that completion can resume that lane or its coordinator. Some hosts stop a foreground worker's commands when it returns; have a surviving owner start the command instead. Respect background-command and monitor deadlines. A monitor must emit only meaningful changes or terminal results; streaming unchanged status wakes the model just as polling does. Keep one owner and one active wait per transition.

When only a command, worker or external transition remains, record the pending state and end the turn. Continue independent work while it runs, then inspect the returned exit status and result before resuming dependent work. Do not use foreground gh pr checks --watch, sleep loops, repeated gh pr view, or repeated output/status reads to keep a lane alive. A worker with pending work returns its handle and next transition to its coordinator, not a completion claim; do not create another worker just to poll it.

If the host cannot notify or resume the owner, make one bounded status check, record the missing capability and an explicit next check time or event, and return pending. Use the caller's cadence when given; otherwise check no sooner than five minutes later, with a transition deadline. Each resumed turn checks once and parks again if unchanged. This is a fallback only: do not replace a working notification path with scheduled polling. If commands cannot survive a turn at all, return the required execution capability to the caller rather than launching an untracked process or claiming the check ran.

Inputs

The caller supplies the repository, expected identity, absolute deadline, and an optional state path. The command reconciles any saved state with fresh GitHub state, waits while nothing relevant changes, and exits with one result. Use --json from workflow skills; human output is for direct terminal use. Skills default state files beneath .agent/waits/, which the consuming repository must gitignore; another harness may supply any private state path through --state.

Invocation

wait <kind> --repo <owner/repo> --deadline <ISO 8601> [--interval <seconds>] [--state <path>] [--json] runs a wait; inspect <kind> takes the same identity flags without --deadline, --interval, or --state. Each kind requires its own identity flags:

• checks-terminal: --pr <number>, --head <sha>, and --all-workflows, one --check <name> per expected check context, or both. A run with neither exits with an operational error instead of waiting.
  • --all-workflows needs no check names. It finds every GitHub Actions workflow run for the exact head and every job in each run's latest attempt. It waits while there is no run, or while any run or job is queued, in progress, waiting, pending or requested. It reports changed as soon as a run or job ends in failure, cancelled, timed out, action required or another non-success conclusion. It reports satisfied only when every run and job succeeded, was skipped or was neutral, and two consecutive observations one interval apart show the same census. Jobs added by later matrix or needs stages keep their run in progress. A run started after another run finishes shows up in the second observation. inspect cannot confirm a stable census, so it reports waiting instead of satisfied for an all-success snapshot.
  • --all-workflows does not cover commit statuses or check runs from other apps, such as CodeRabbit or Vercel. Add a --check for each one the caller needs; the wait then also requires those contexts to succeed.
  • A named check passes when it ends in success, skipped or neutral. The wait reports changed as soon as one named check ends any other way, even while the rest still run. Check runs that share a name count once per source: their app and, for GitHub Actions, their workflow and triggering event. A workflow that runs on both push and pull_request therefore needs both runs to pass. From one source only the newest check suite counts, so a re-run, a run cancelled by a newer one or a re-check after an edit replaces the older run, whichever started first. A commit status counts by its latest state.
  • --check alone judges only the named contexts, whatever else is still running for the head. Use it only when the complete expected set is known, such as the repository's required status checks. Checks visible at wait time are not that set: workflows add jobs in stages.
• pr-merged: --pr <number> and --head <sha>. It reports changed when the pull request is closed without merging.
• workflow-terminal: --run-id <id> and --head <sha>. Use it for a dispatched or other run whose checks do not appear on the pull request.
• tag-target: --tag <name> and --head <sha>.
• release-assets: --tag <name>, --head <sha>, and one --asset <name> per required asset.
• wake-at (wait only): --deadline alone. It keeps no state file, so --state has no effect.

There is no --checkpoint flag. The state path is --state, and the command derives a default path from the kind and identity when it is omitted.

Both inspect and wait accept --rest-page-size <1..100> (default: 100). Use a smaller value when GitHub fails to serve large REST pages. It applies to workflow runs, jobs, and the REST fallback for checks and statuses. All pages are still read and the complete workflow census is still required. The option does not change GraphQL queries or the saved wait identity, so an existing wait can resume with a different REST page size.

Results

--json prints one envelope whose state is the result. Only satisfied establishes the awaited fact. changed means a check or workflow reached a non-success terminal result or other relevant state moved; inspect it before deciding. invalidated means the expected head, ref, or saved identity no longer matches, timed-out means the deadline passed, and cancelled means the wait was interrupted. These exit 0; operational-error exits 2. inspect can also report waiting.

Use inspect for a single authoritative delivery snapshot and wait to await a transition. Supported predicates are exact-head checks, pull-request merge, workflow completion, tag target, release assets, and an absolute wake time. A changed expected head or ref invalidates the wait instead of being followed silently. GitHub remains authoritative; checkpoints contain normalized identities and observations, no credentials, raw comment bodies, or raw API payloads.

The helper uses authenticated gh transport. GraphQL is primary. A REST fallback is used only when GraphQL is rate-limited and REST preserves the same fact, or when GitHub exposes the fact only through REST.