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.
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.
- A CI fix comment that did not reach GitHub leaves its linting or formatting row (a try) in place and gets its own row, timed from the failure. After the backoff the stored patch or formatter result is posted again with no new model or lint run; a patch that needed no lint task is kept as a linting row without one. The stored patch is not dropped during the backoff.
- A carry-over's unsent ask is not a try: its own fixups row times the same backoff, and the carry-over is held after the last one.
- The item bar shows the last unsent comment as failed.
- A PR stays assigned while an unsent comment waits for its retry.
A write that fails before any request goes out (the GitHub App's installation token could not be minted) returns sent: false. The fix-up, conflict and CI fix steps record such a comment as comment-unsent instead of the final comment-failed; the planner asks again 10 minutes after the first and an hour after the second, and a third unsent comment is final. Existing comment-failed rows with an installation-token reason become comment-unsent.
Fast-tracked items have a weight from 1 to 10 (default 5): heavier items
go first in every stage that orders by the work rank, and equal weights
keep the order added. The weight is a new fast_track column; the scores
(fastScoreSql, workScores) put it above the order added, so every planner
and the merge order follow it without other changes. The CLI takes
`fasttrack add <item> --weight N` and `fasttrack weight <item> N`, and
POST /api/fast-track/weight sets it (audited as fasttrack-weight).
"Add to merge queue" (POST /api/queue/add {item, weight}, operators)
starts a pre-approved merge at the PR's current head with no fix-up, a
row in `review` as a carried Merge answer uses, so the PR merges only
after a sure take review there, a clean pre-merge review and green CI; a
take-fix review, a defect or a red head asks the Merge? question. It also
fast-tracks the PR with the weight.
The Merge queue tab's add box now adds to the merge queue with a weight,
the fast-track marker shows the weight, and operators change it on the
board's cards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A confidential environment is ready from its worker's first call after the flag, or the environment, was turned on (confidential_since, confidential_online_at); label, slot and key changes do not move the 4-hour build limit of private tasks. Until one is ready the fix job ends before the model call, the finding notes the wait and advanceSecurityFixes continues it once one is ready, and the confidential-env health check warns while security fixes or security build tasks wait; build-queue leaves those tasks to it.
build_envs.confidential is set on the Build environments page or with envs add/set --confidential, never by the worker. claim() hands fix-check and scan-test tasks only to confidential environments and gives those no PR build, integration batch or PR/issue task; a signing worker skips integration batches too. While no enabled environment is confidential the tasks stay queued with the reason "Waiting for a confidential build environment.", the security fix's check tool answers at once, and the 4-hour round and scan-test limit starts only once a confidential environment can take the tasks.
- History counts per GitHub account: a PR row with an id counts for that id, an id-less row for the
account the login is bound to, and for nobody while the login is unbound. The sync no longer binds a
login that has id-less merged or closed PRs; the backfill binds it from the history's own author.
A failed binding pass shows as the author-ids health warning.
- An allow override stores the account it was given to (looked up when set) and counts only for it;
a deny holds for any account under the login, in any letter case. The outside-author limits read
the same per-account standing.
- The pipeline's own Fix, main CI fix and security fix PRs keep their own tier only while the head is
a commit the pipeline pushed to its own branch (own_commits), or reached from one only through its
own later pushes; otherwise a person is asked.
- Removing or demoting an operator revokes their org setup links, and a link works only while its
creator is an operator. Removal drops Discord links and codes in any letter case.
- Decision alerts to maintainers need a repo permission read within a day.
- A cached author permission is not used for another account under the login, nor once its refresh
has failed too often or for too long.
- Pushes to new branches and new PRs need the grant of a step that opens the pipeline's own PRs;
security fix writes carry their own grant kind.