Skip to content

auto-claude 1.1.0: never rest while work is queued - #38

Merged
unohee merged 2 commits into
mainfrom
feat/auto-claude-keep-working
Nov 11, 2026
Merged

unohee merged 2 commits into
mainfrom
feat/auto-claude-keep-working

Conversation

@unohee

@unohee unohee commented Nov 11, 2026

Copy link
Copy Markdown
Contributor

Why

In 1.0.0 the run was strictly serial: while a PR waited on CI (10–90 min), the session sat idle. The watchdog's first wake came after 3 min, and nothing stopped the model from waiting inside a turn (long sleeps, CI watches, cron/wake-ups).

What

  • Stacked look-ahead. While a PR ships, the next milestone starts on its branch, with the start commit recorded. At most 2 milestones are built ahead. They ship FIFO. When the one ahead lands (squashed) or is held, the driver runs git rebase --onto <default> <recorded commit> and pushes with --force-with-lease. A branch that does not rebase gets a fix goal that rebases it by hand. Fix goals keep their own base.
  • Deploy safety. While a deploy repair is under way, nothing new is stacked and nothing stacked ships. The repair ships first.
  • One channel to the model. A diagnosis prompt waits until the milestone being built is done. A ship error is held only after the model was asked and had 10 min to act; with no answer for 60 min it is held anyway.
  • Faster wakes. An idle milestone is prompted at 1, 3 and 6 min and blocked at 12. The planning reminder comes after 1 min.
  • No waiting inside a turn. During a run the guard refuses sleep of 60 s or more (or a computed duration), gh run watch, gh pr checks --watch, Monitor, CronCreate and ScheduleWakeup.

Verification

  • bash scripts/ci.sh: green (auto-claude 127 tests, tsc).
  • New integration tests:
    • stacking → rebase → lease push → done;
    • a rebase conflict produces a fix goal;
    • a broken deploy is repaired before a stacked milestone merges;
    • a ship error is not held before the model was asked.
  • Real-git check of squash merge followed by rebase --onto, with the parent fixed after the child started: testing/stack_rebase_261111_v1.sh (local, excluded).
  • CodeQL javascript security-and-quality: 0 findings.
  • Not covered: stacking has not been exercised in a real session. The E2E fixture uses local CI, which has no busy window.

…e a PR waits on CI, wake at 1 min, refuse in-turn waits

- startBase: with a PR shipping, the next milestone starts on its branch (up to MAX_AHEAD=2 ahead), its start commit recorded; never while a deploy is broken; fix goals keep their own base
- stacked ship items wait their turn; when the one ahead lands or is held, git rebase --onto <default> <recorded commit> drops its commits; a branch that does not rebase gets a fix goal that does it by hand
- final check waits for every unfinished ship, stacked ones included
- the diagnose prompt waits until the milestone being built is done (no prompt mid-milestone)
- watchdog wakes an idle milestone at 1, 3, 6 min, blocks at 12; planning reminder after 1 min
- guard refuses long sleeps (>=60 s or computed), gh run watch, gh pr checks --watch, Monitor, CronCreate, ScheduleWakeup during a run
…hip, lease-push rebased branches, ask before holding

- while a deploy repair is under way nothing new is stacked and nothing stacked ships; the repair ships first and clears on merge, hold or block
- a stacked branch rebased by the driver is pushed with --force-with-lease
- a repeating ship error is held only after the model was asked and had 10 min (held unasked after 60 min)
@unohee
unohee merged commit 59fad96 into main Nov 11, 2026
1 check passed
@unohee
unohee deleted the feat/auto-claude-keep-working branch November 11, 2026 13:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant