The dashboard stamps its liveness every 5 s; job processes expire leases only while
it serves, and for missing heartbeats only after it served the whole idle window.
After a gap every lease's heartbeat, claim and run start move forward by the gap.
A token's drops count at most once in 10 minutes.
The time limit counts from the first running heartbeat; preparing may last 15 min.
A claim repeated with its request id within 5 minutes gets the same reply (kept in
the dashboard's memory; after a restart the orphaned lease goes back unused).
Ended and requeued leases keep their secret hash, so every lease route answers 410.
ok:false requeues with the next attempt and skips that login; a repeated accepted
submit is accepted again. Cool-off is per login, counts drops and 5 failures in a
row within 30 min, and is cleared by login.
codex exec runs in the helper's bubblewrap sandbox with its own CODEX_HOME login (harness codex), its shell and
the other tools that reach past the task turned off, and reads through read-only file tools that the helper
serves to the run over localhost MCP, confined to the mounts. One Codex call at a time per helper, so refreshes
of its login never race. Codex has no tier defaults: the config names its models.
An author's claim skips a judging call when another login's helper made an empty claim within the idle window that covered
the call's tier and repos, fits the role rules, has cached write permission and authored nothing in the call. Otherwise the
author's helper gets it. The same rule applies to the next_task count. Empty claims record their tiers and repos.
The runner sums the time in which any of a job's calls waited for a helper and leaves at most 12 hours of it out of the
deadline and no-progress stop. A done reply is cleared once the caller reads it. A lease records its token and ends
(returned to the queue, no attempt used) when the token may no longer run tasks or its login is no longer a maintainer
or operator. The sync stores each commit's authors as GitHub lists them, Co-authored-by trailers included, and judging
calls skip them; classify.haiku, security.verify and security.fixcheck judge too. Claims refuse helpers older than the
required version, skip calls the login released in the last 10 minutes and check write permission with the 1-minute
cache. Lease expiry is compare-and-set, and maxWaitMs counts queue time only.
A remote call's prompt marks the phrases that name Claude's tools; the bundle and the run's prompt fill them for
the claiming harness, with exactly the server's text for claude. Calls with the reference server get the
reference input flag and a sentence that names reference_search and reference_around instead of /reference.
The helper keeps a session when the bundle's keepSession says so, submits
usage per model id (explorer included) and the CLI's total cost, and gives a
lease back through the release route when the fault is its own: too old,
wrong platform, missing harness or token, file setup failure, or stopped by
the user. The work loop stops when it can take no call and pauses after a
release. Submit carries costUsd; the server stores it with the call's usage
and records it as the run's cost.
The inference setting (server | user-local) is set with `autocraft inference` or the operator-only
GET|PUT /api/settings/inference, which also holds the accepted model families per tier. In user-local mode
askClaude queues each Claude turn in remote_calls with a task bundle (prompt parts, claude arguments, checkout
and workspace mounts, repo docs inline, per-lease MCP servers) and polls for the result; follow-up turns go to
the helper that holds the session. Helpers claim with an agent token that may run tasks, in the model queue's
order filtered by tier, repo, write permission and author exclusion, and use a per-call lease secret for the
bundle, heartbeat, release and submit routes. Lost leases return to the queue up to the prompt's attempts.
Queue time does not count toward a job's deadline or no-progress stop, and accounts are not consulted.
- Lanes keep the order they first arrived in; a lane the server stops sending stays as ended until the reader returns to the tab. Quiet steps are left out before the queued LIMIT.
- Open integration batches are lanes, with the build of their combined tree.
- The latest five finished Now feed events sit under the board.
- Account labels on model calls are for staff only, and a lone call's progress line is blanked when it names a hidden repo, on the board and in the Now feed.
- The merge queue row shows its wait label, conflict note and operator approval as text.
- The live query and the item status list follow their topics; the board does not poll.
- Workable tables switch to a wider layout only with room to spare, so a page scrollbar cannot flip them back and forth.
- Security findings get their own icon; item icons name their kind and state without explanations; the board's reading tip is gone.
- The dashboard spec describes the Live board, the feed list under it and the item peek.
GET /api/live gives running, queued and lately ended jobs (and calls or build tasks of their own) with their model calls and build steps, and the median and 90th percentile of earlier runs of the same work. The board draws them on a shared axis: solid done, pulsing running, dashed the expected rest, a red line where 90% of earlier runs were done. It replaces the Live feed section on Home.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GET /api/items/status rows add author, closes and checks (review, CI, Verify, mergeability, size, our build) for PRs the plan holds a queue view for, veiled as in the merge queue. GET /api/items/:repo/:n/peek returns an item's title, author, labels, status and closes for hover cards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A take PR whose CI fix ended with a failed comment, or with its last unsent one, waits on a person.
- The item bar reads the newest row past unsent comments that are still retried, and counts a fix-up's unsent comments per source and carry-over. Reposting an ask reads the review's own fix-up rows.
- The CI fix keeps a stored patch while its unsent comments are not final and leaves the timing to the plan.
- A later round's patch that needs no lint task is stored when its comment is not sent.