Skip to main content

/work picks ready lanes only, with a reliable claim lease

· 3 min read
Carefree Investments LLC
Anchor maintainers

Bare /work now selects only from ready lanes (bugs/, features/) and never scans in-progress/. In-progress plans are owned by a required lease — so several agents can share one project without ever stealing each other's work.

The collision it fixes​

/work used to try to be helpful by resuming "your" in-progress plans first. To do that it had to reason about ownership of everything under in-progress/ — and ownership was fragile. Leases were optional (interactive /work often just did a git mv), and the lease TTL was a short 1 hour. So an in-progress plan that a live agent was actively working — but hadn't leased, or whose hour had lapsed — looked stalled to a second agent's /work, which could then silently reclaim it. Two agents, one plan.

What changed​

  • Bare /work is ready-lanes-only. It never scans in-progress/ to resume or reclaim. Resuming your own work is now an explicit named target (/work in-progress/<slug>.md), and only if you hold the lease.

  • Claiming is atomic and required. Moving a ready plan into in-progress/ writes its lease under .plans/.leases/ in the same step. The lease is what marks the plan yours; a bare git mv with no lease is no longer how you claim.

  • Long TTL + heartbeat. The lease TTL is now 24h, and a live agent extends it with a heartbeat, so slow-but-alive work never looks orphaned:

    python scripts/work_once.py --heartbeat in-progress/foo.md --agent-id worker-1

  • No silent reclaim. A foreign, unleased, or expired-lease in-progress plan is never auto-taken. Recovering genuinely-abandoned work (a crashed agent past the 24h TTL) is an explicit choice:

    python scripts/work_once.py --recover --path in-progress/foo.md --agent-id worker-2

A deterministic picker for smaller models​

Less-reliable models shouldn't have to reason about lanes and leases by hand. There is now a single command that prints exactly one next ready plan — or a clear "nothing ready" (exit 1) — and, with --claim, performs the move-and-lease itself:

python scripts/plan_select.py --next                       # print the next ready plan
python scripts/plan_select.py --next --claim --agent-id w1  # claim it atomically
python scripts/plan_select.py --next --json                 # machine-readable

Multi-agent by default​

The project-orchestrator MCP moves with the same rules: plans_claim gains a recover flag, there's a new plans_heartbeat tool, and plans_complete now moves an agent-finished plan to review-needed/ for human sign-off rather than straight to completed/. Unique --agent-id per worker, required leases, and no silent reclaim are what let a fleet pull from one backlog safely.