usage-window.py stores a calibrated window quota and emits an explicit
verdict, so a run can no longer park on an uncalibrated raw token count.
Milestone assess is scoped; full suite moves to the gated push points.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YUXS63P1WCdC6bEKWcnAE
Parallel yolo/nightshift/blitz sessions each ran the full suite and OOMed
the box. gate.sh caps concurrency, memory and build parallelism; skills now
run scoped checks in the loop and one gated full run per push.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YUXS63P1WCdC6bEKWcnAE
A night of autonomous work needs a review gate, especially on a live repo.
Output is now one branch and one PR rather than commits on the default
branch.
Separates committing from pushing, which had been conflated. Local commits
are what protect against an agent dying mid-task, and they cost no CI, so
they stay at every milestone. Pushing is what triggers CI, so it becomes a
deliberate act: at the end, when parking, or on request. A night of
milestone pushes is a night of CI runs on work that was half-finished at
the time, which trains the user to ignore the build.
Subagents are told to commit locally and explicitly not to push.
Covers both forges: gh for GitHub, the API for Gitea (no tea CLI on most
setups), and reporting a compare URL rather than faking success if auth
fails.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SSXDKsik8yYrezStVaDXt
Takes an issue, ticket or plain description and works it for hours as an
architect: freeze the interface contract first, then delegate implementation
to subagents against frozen signatures, review, commit every milestone.
Checks the rolling 5-hour window before each dispatch and parks rather than
letting agents die mid-task. The usage script sums every Claude Code session
on the machine, and is explicit that it is a floor — it cannot see other
machines, claude.ai, or direct API use — so the reactive backstop of an
agent dying on a limit error is documented as the stronger signal.
Encodes what worked in practice: disjoint file ownership per agent, asking
agents to report disagreement (their pushback was the highest-value output),
sabotage-testing every safety net, and asking what a passing check is
structurally unable to see.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SSXDKsik8yYrezStVaDXt
naps62
naps62-yolo