What problem does it solve? Coordinating unattended development cycles across a multi-host fleet (Linux, macOS, Windows, forge containers) is error-prone: agents lose uncommitted work, skip daily maintenance, overspend rate-limited token budgets, or exit without proof that commits reached the remote. This Skill provides a single top-level loop that classifies the host, runs daily maintenance gates, drains claimable plan work, executes e2e smoke gates, and enforces a strict commit-push-attest exit contract. ## Core Features & Use Cases - Invocation Modes: Full mode runs the complete cycle; Smoke Mode gives fast, cheap verification with a machine-greppable MO-SMOKE: PASS/FAIL verdict without touching the repo. - Terminal Attestation: Full mode must emit a derived MO-FULL: marker via scripts/mo-full-attest.sh self, proving local HEAD converged with the remote before exit. - Work Salvage: scripts/salvage-dirty-worktree.sh pushes dirty or stranded work to a salvage ref before any refusal, preventing loss of uncommitted work. - Token Budget Governance: Caps sub-agent counts, tiers models by task class, and logs token spend per cycle. - Use Case: An operator launches ./repeat --prompt "Use the /meta-orchestration skill" on a fleet host; the skill detects the host type, runs the daily maintenance gate once, claims and completes plan work, runs the build gate, commits, pushes, and emits a verified attestation marker. ## Quick Start Run the meta-orchestration skill on this host to execute a full unattended development cycle and end with a verified push attestation.
