manta-coordinate

manta-coordinate

Purpose

Multiple clones share a single Manta Bus. Coordination is data-driven (locks, claims, broadcasts) — never social. This skill defines exactly which calls to make, when, and what to never do.

Allowed

• Lock before write: any path you intend to edit goes through manta.lock { clone_id, path } first. Renew via manta.renew_lock every 5 s while held. Release via manta.unlock the moment you stop editing.
• Claim before duplicating: before doing a piece of work that a sibling could also do (e.g. "summarize subdir X"), call manta.claim_work { item: "summarize:X", timeout_ms: 600000 }. If it returns a conflict, pick a different item.
• Filtered broadcast: only three event types are bus-traffic. Use them sparingly:
  • breakthrough — root cause / subproblem solved (sibling may unblock)
  • blocker — stuck (main agent intervenes)
  • dependency — discovered code that affects another clone's scope
• Drift report: every ~50 actions, call manta.drift_report { score: 0..1, evidence } so the main agent can spot scope drift early.
• Anchor sync: when you receive a contract_refresh event from the main, re-read manta.task_contract.read and re-ack via manta.ack_contract.

Cast-agnostic contract_refresh (Sec 5.7)

When you broadcast a manta.contract_refresh it fans out to every active clone, regardless of cast. Therefore the payload must be cast-agnostic — do not include per-cast approach hints, per-cast scope, or per-cast deliberations. For per-cast updates, use manta.task_contract.write per clone instead. Violation is a soft information leak across casts (research clone-C §6 "Contract refresh content").

Forbidden

• Holding a lock without renewing. Stale leases (15 s without renew_lock) are reaped by the orchestrator and emit lock_reap events visible in the post-mortem — a quality signal against you. Renew on time or release.
• Broadcasting status pings ("started reading routes/index.ts"). Local log only — bus traffic is for actionable events.
• Direct messages for opinions. manta.message is for round-table escalation (sibling proposes vs sibling proposes). Disagreements → broadcast blocker so the main can decide.
• Bypassing the work-claim board. Even if you "know" no sibling is doing something, claim it. Future-you (or a re-spawn) reads the claim log to reconstruct lineage.

Examples

You're cloning a refactor-wave with sibling B and C:

1. Each clone takes a region by claim:
   • You: manta.claim_work { item: "refactor:auth", timeout_ms: 600000 } → ok
   • B: claims refactor:billing
   • C: claims refactor:logging
2. You discover auth/index.ts imports billing/utils.ts — that's B's scope.
   • You broadcast: manta.broadcast { event_type: 'dependency', payload: { from: 'auth/index.ts', to: 'billing/utils.ts' } }.
   • You DO NOT edit billing yourself; you DO NOT message B asking permission.
3. You finish: manta.release_work { item: "refactor:auth" }, manta.unlock every held path.

Bad pattern:

• "I'll just touch billing/utils.ts since it's a one-liner." → outside scope. Hooks block. Fragility -1.
• "B's plan is wrong, I'll msg them." → anti-gossip violation. Escalate via blocker broadcast instead.