Hephaestus Agent Workforce Network

The active host LLM is the per-turn orchestrator. Hub supplies content and qualification evidence, exact immutable releases, and BYOM directives; it does not select the final team or run a server-side LLM.

1. Create a redacted agentlas.workforce-work-order.v1 with substantive role

slots, skills/knowledge, MCP tools, artifacts, runtime/language/authority, cardinality, and handoff/review edges. Keep private context local.

1. Call Hub MCP workforce.search_candidates.
2. As the host LLM, author agentlas.workforce-selection.v1 from exact content

and eval evidence. Do not use lexical top-1, popularity, ratings, history, revenue, or local callability as semantic fit.

1. Call workforce.validate_selection, revise on rejection, then call

workforce.prepare_execution. Require exact release version, package hash, content digest, and directive bundle; never silently substitute.

1. Bind the prepared plan to the stable current goal/task id with

workforce.bind_goal. On every later turn read workforce.goal_context, reuse the incumbent roster plus local skills when sufficient, recruit only a real additive gap, and record the posture with workforce.record_goal_turn.

1. Before each bound planner/manager, worker, synthesis, or verifier call,

advertise the host's real sessions and call model.resolve_allocation with the host-owned stage. Use the exact provider/model/effort receipt. Pins and ceilings come only from AGENTLAS_MODEL_ALLOCATION_POLICY_JSON; missing worker policy inherits orchestrator. usage: null before invocation is not worker-execution proof.

1. Run only useful bound manager/planner, workers, synthesis, and verifier as

distinct model invocations with explicit artifact handoffs and nested Team graphs.

Keep the roster across turns, sessions, restarts, compaction, and lease expiry. Release it only with workforce.complete_goal(explicitCompletion=true) after explicit whole-goal completion/cancellation. A 24-hour lease controls only the next Hub charge; standby is durable availability, not a continuously running model. Memory/Experience accrue on actual invocations.

If this OpenClaw host cannot call the Workforce MCP tools or create separate child invocations, report the last truthful state instead of calling the legacy router. Execution success requires planner parse success without fallback, every child/handoff receipt, synthesis, and a passing verifier.