void-autopilot

The autopilot orchestrate step creates one worktree and branch per ticket before any worker starts, then runs disjoint-footprint tickets in parallel lanes up to the plan's concurrency limit. Tickets that collide on files, lockfiles, or migrations are sequenced one at a time.

Use `mode autopilot 6h` with a duration, not a ticket count. The budget comes from the programme's autopilot.chainBudget declaration as a ceiling, or a two-hour fallback when absent, and the chain step decides between units whether to continue or stop.

Only when the programme declares mergeGate: union-reviewed with a non-deploying target branch. The grant refuses on production downstreams, human-gated tickets, unprotected bases, sensitive paths like migrations and lockfiles, unread or contradicted union reviews, and stale readings.

Yes. Both runtimes read the same OrchestrationPlan and return the same WorkerResult schema. Claude executes via its Workflow primitive while Codex uses native subagents with pre-created worktrees, since subagents share filesystem access.

A worktree isolates the working tree, index, and HEAD, but refs like refs/stash, tags, notes, and remotes are one shared namespace across all worktrees. Two workers stashing concurrently can take each other's entries, so the plan denies the whole class of shared git state writes.

The failed unit is journaled as unit-blocked with its cause, its branch and worktree are preserved, and the remaining tickets continue through reconciliation and verification. A unit is taken exactly once and never reproposed onto its own pull request.