Repository navigation
auto-claude 1.3.0: wait on compute without ending the run - #41
Merged
Merged
Conversation
Measured in ArtifactNet on vela (2026-11-09 to 11-15). Milestones carried
prose start conditions ("(start only after ~/data/.../cache.log shows
DONE)"). The driver could not read them, so it started each line at once
and the model deferred it. Once every line was deferred, the final check
failed and a round was spent. On 11-09 three rounds went in 25 min. On 11-15
the run stopped at round 3/3, and nothing resumed it when the cache
finished three hours later. The guard refused Monitor and said the driver
would prompt the session, which was not true for compute.
Changes:
- New queue fields after the check: `:: after: <cmd>` must exit 0 before
the line starts, `:: alive: <cmd>` detects a dead job, `:: within: 12h`.
The driver polls `after` every 3 min from the repository root (fetching
first for origin/ refs) and builds other lines meanwhile.
- Lines that still wait keep the run going: no final check, no round spent,
not idle. Waiting time is left out of maxHours, up to maxWaitHours (72).
- When `alive` fails twice, or a wait passes its `within` (12 h without
one), the model is prompted once with the last poll output. A wait past
twice its `within` is held.
- New tool mcp__auto-claude__wait_until { until, alive?, within, reason }.
At the next stop it puts the open milestone back to wait on its own
branch. A stacked milestone keeps the commit it was stacked on, so the
restarted branch is still rebased off its parent.
- validatePlan refuses parenthesised prose start conditions. The plan
prompt, POLICY, the loop policy, the kickoff and the wait refusal now
point to `:: after:` / wait_until for compute, never to defer_goal.
- /auto-claude status lists each wait; the status line piece shows ·waitN.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
The ArtifactNet run on vela (session 287dfa7e) stopped because of how it handled compute waits, not because anything hung:
(start only after … shows DONE). The driver could not read them, so it started each line at once, and the model deferred it to[!].What
:: after:/:: alive:/:: within:go after the check, withbaselast. The driver pollsafterevery 3 min from the repository root and builds other lines meanwhile.maxHours, up tomaxWaitHours(72).alivefails twice, or a wait passes itswithin(12 h by default), the model is asked once. A wait past twice itswithinis held.wait_untiltool. It puts the open milestone back to wait on its own branch. A stacked milestone keeps itsonto.validatePlanrefuses parenthesised prose start conditions. Every text that steers the model now points compute waits towait_until/:: after:, never todefer_goal.Tests
scripts/ci.shis green: auto-claude 149 pass (was 135), tsc OK, and the shared copies are in sync.afterpasses.ontohand-off fails the stacked test.