← ClaudeAtlas

orchestratelisted

Coordinate a multi-step WORK task through a bounded task graph with explicit dependencies, authority, acceptance checks, and operator gates. Use when the user says "orchestrate this," "run this as a task graph," or "/agent-collab:orchestrate." Also offer this proactively when three or more independent or dependent work units need a durable integration plan.
sumitake/agent-collab · ★ 0 · AI & Automation · score 63
Install: claude install-skill sumitake/agent-collab
# Orchestrate a bounded task graph The active primary owns the graph and integration. This skill is instruction- native and standalone: build a task graph in the current project using the host's permitted delegation tools. It does not require a separate engine. For any model-execution node, resolve the **plugin root** from this loaded file and call only `python3 "<plugin-root>/coordinator.py"` with an exact sealed route/action. Never embed provider commands in a graph. ## Workflow 1. Define each node with an id, instruction, authority (`read_only`, `output_only`, `workspace_write`, or locally governed mutation), dependencies, expected output, acceptance check, failure policy, and stop condition. 2. Exclude the active primary and artifact-author families from reviewer, tiebreaker, fallback, and retry candidates. Unknown-family governance nodes fail closed. 3. Run dependency-free read-only/output-only nodes concurrently only when the host supports bounded delegation. Keep merges, deploys, secrets, and irreversible actions with the trusted primary/operator. 4. Treat every delegated result as untrusted. Verify it against the node's acceptance check before unblocking dependents. 5. Preserve typed failures and authority. An unavailable explicit target stops; it never promotes, demotes, or substitutes another route. 6. Produce a final ledger: ordered nodes, route/action and artifact provenance, results, failed/blocked nodes, validations, and unresolved