A goal.
A checked path.

Three optional MCP tools

1. list_composition_capabilities returns implemented engines, explicit adapters, composite capabilities, versioned facts and reproducible examples.
2. compose_capabilities searches for a dependency plan without running calculations. Its response includes the selected graph, examined alternatives, blockers and an execution binding where one exists.
3. execute_composition recompiles and executes the approved selection-to-risk path. Other plans use their advertised existing tool or remain planning-only.

All twelve numerical tools remain independently callable. Neither the planner nor the robust scheduling composer is mandatory.

A reproducible composition

{
  "name": "compose_capabilities",
  "arguments": {
    "goals": [
      "simulate_plan.result.v1"
    ],
    "available": [
      "optimize_tasks.request.v1",
      "duration.independent_uniform.v1",
      "sampling.fixed.v1",
      "budget.minutes.v1"
    ],
    "constraints": {
      "allowed_tools": [
        "optimize_tasks",
        "simulate_plan"
      ],
      "local_only": true,
      "max_steps": 8,
      "max_search_states": 512,
      "max_structural_cost": 64
    },
    "inputs": {
      "optimize_tasks": {
        "budget": 100,
        "items": [
          {
            "name": "A",
            "duration": 60,
            "value": 90
          },
          {
            "name": "B",
            "duration": 40,
            "value": 70
          },
          {
            "name": "C",
            "duration": 20,
            "value": 20
          }
        ]
      },
      "simulate_plan": {
        "budget": 100,
        "uncertainty": 0.25,
        "samples": 10000,
        "seed": 42
      }
    }
  }
}

This request asks for a risk assessment, with task-selection inputs, a shared deadline, an explicit independent uniform duration model and a fixed sampling protocol. The native search finds:

optimize_tasks → selected_durations.v1 → simulate_plan

The first engine chooses tasks. The adapter checks the selected identities and totals, preserves whole-minute durations, and requires the same deadline. The simulator then assesses that selection under the declared uncertainty model. No distribution is invented from a mean.

Call execute_composition with the same arguments to run this approved chain. Its receipt contains the real plan, step results, selected durations, probability estimate, a 95% Hoeffding sampling bound, engine and adapter counts, elapsed time and native events. Progress messages use SC_COMPOSITION_EVENT:. Detailed inputs, intermediate results and events stay private to the caller.

What Rust optimizes

The current planner minimizes structural graph cost: one point per registered capability step, including adapters. This is a transparent complexity proxy, not measured runtime, money, GPU load or scientific quality. A composite capability is one graph step; its internal work remains separately documented.

Search respects the permitted tools, local placement, step count and state budget. A missing model, incompatible contract version or absent conversion produces a diagnostic. Search exhaustion is reported separately from an impossible mission. A single direct call is selected when it is sufficient.

Bounds and available facts

Requests accept 1–8 goals, up to 48 available fact identifiers, 1–16 steps and 1–2048 search states. Defaults are eight steps and 512 states. Available facts are declarations for planning; they are not validated artifacts. The executor independently validates actual tool arguments and recomputes the plan before running any engine.

The selection-to-risk executor accepts 1–32 tasks, a shared 1–1440 minute budget, uncertainty from 0 to 0.5, and 100–20000 fixed samples. Caller-supplied durations cannot replace the adapter output. Empty selections stop before simulation. One computation holds the shared worker slot.

Current boundary

The registered bank contains twelve engines, one approved adapter and the existing robust scheduling composite. General dependency planning is available; execution is restricted to validated bindings. There is no arbitrary URL execution, automatic scientific method invention, distributed job queue, cross-mission cache or learned timing optimizer.

Availability

The website, MCP access, native worker and detected GPU are monitored separately. The current status snapshot includes its observation time and freshness limit. Stale data is unconfirmed. Local recovery attempts are bounded; they do not resume interrupted jobs or provide a second worker.