What problem does it solve? Running autonomous coding tasks against shared repositories risks corrupting the primary checkout, losing work to crashed sessions, and shipping unverified changes. This Skill provides a fail-closed execution loop that claims tasks atomically from a Supabase queue, implements them only in validated isolated worktrees, and hands finished branches to a canonical merge/release train. ## Core Features & Use Cases - Atomic task claiming: Uses a single SQL CTE with FOR UPDATE SKIP LOCKED to claim up to 5 queued tasks, prioritized by kind (recovery, toolchain-repair, bugfix, build, canary) and dependency readiness. - Fail-closed worktree isolation: Acquires a server-side branch lease, allocates a dedicated git worktree per task, and never mutates the primary repository checkout; failures requeue the task instead of falling back. - Single-writer branch delivery: Commits real code changes, pushes only to agent/{slug} branches, records the commit SHA, and marks tasks DONE only when a remotely addressable artifact exists. - Use Case: An orchestrator queue contains a bugfix task for a web app. The executor claims it, fetches pre-enriched context from the runner pipeline, implements the fix in an isolated worktree, pushes agent/bugfix-slug, and lets the merge train handle integration and deployment. ## Quick Start Ask the agent to run the cowork executor loop to claim queued tasks from Supabase and implement them in isolated worktrees.
