AI-authored model allocation

Agentlas separates workload judgment from runtime enforcement.

1. A parent or leader AI judges each child packet using structured features only: complexity, risk, context size, output size, tools, multimodal input, and fan-out.
2. It receives the host's live inventory and emits agentlas.model-allocation-decision.v1 with an exact advertised model ID, a provider-neutral cost tier (economy, balanced, or frontier), effort, and reason codes.
3. The host resolver checks the actual inventory, explicit user/agent/division/firm pins, context and tool support, and cost ceilings.
4. The host records agentlas.model-allocation-receipt.v1. Raw prompts and transcripts are forbidden in this receipt.

Every invocation has one host-owned role:

• plan, route, clarify, verify, and synthesize are orchestrator;
• execute is worker;
• unknown host stages stay on orchestrator instead of silently lowering quality.

The caller may propose an allocation decision, but it cannot change the host-owned stage. A decision whose phase differs from the canonical stage phase records phase_stage_mismatch and is not used as a valid tier decision.

Core contains no vendor model-name table and never infers cost or capability from a model ID. The host supplies those attributes with its live inventory. When more than one compatible model exists in a requested tier, the parent AI must return exactModelId; Core refuses to choose one by lexical order or a provider preference. A matching current model or a single unambiguous live candidate may be preserved without inventing a new model choice.

High-risk work does not automatically force Frontier. The receipt instead sets independentVerificationRequired; hosts must combine model choice with a separate verifier, narrower permissions, and operator approval for consequential side effects.

If a parent decision is absent or invalid, Core follows the role policy and then keeps the host's current compatible session as the final legacy fallback. A missing worker policy inherits the orchestrator policy for quality. An orchestrator policy never inherits a worker policy.

The allocation receipt always records role. Allocation happens before the runner call, so usage is null at this point. A later invocation or run receipt records observed input/output usage; Core must not write zero tokens as if they were observed.

MCP trust boundary

model_allocation_decisions may arrive from the parent/leader AI through the MCP route request. Cost ceilings and pins may not. The stdio MCP server ignores any legacy caller-supplied model_allocation_policy field.

The operator configures host guardrails when launching the MCP server with AGENTLAS_MODEL_ALLOCATION_POLICY_JSON. The preferred form is role scoped:

{
  "orchestrator": {
    "pinnedProvider": "codex",
    "pinnedModelId": "gpt-sol",
    "maxTier": "frontier",
    "maxEffort": "xhigh"
  },
  "worker": {
    "pinnedProvider": "local",
    "pinnedModelId": "luna",
    "maxTier": "economy",
    "maxEffort": "low"
  }
}

worker: {"inherit": true} explicitly inherits the orchestrator. Omitting the worker block has the same quality-first behavior. Legacy flat pinnedModelId, pinnedProvider, maxTier, maxEffort, and requiredCapabilities remain the orchestrator-compatible policy.

The names above are illustrative opaque host inventory values, not Core registrations. The same contract can point to Codex, Claude, Gemini, a BYOK provider, Ollama, or another local runtime. For example, a host may advertise fable or opus for orchestrator and sonnet or haiku for worker without adding provider-specific code to Core.

Fresh Codex and Claude Workforce sessions call model.resolve_allocation before each planner/manager, worker, synthesis, and verifier invocation. The tool accepts a host-owned stage, live session_inventory, and an optional parent decision. It does not accept model pins or cost ceilings in tool arguments; those remain operator environment policy.

For Codex, put the policy on the MCP server itself:

[mcp_servers.hephaestus-network.env]
AGENTLAS_MODEL_ALLOCATION_POLICY_JSON = '{"orchestrator":{"pinnedProvider":"codex","pinnedModelId":"gpt-5.6-sol"},"worker":{"pinnedProvider":"codex","pinnedModelId":"gpt-5.3-codex-spark"}}'

An outer shell export is not a portable substitute because the host controls which environment reaches its MCP child. The Agentlas installer preserves the operator-owned mcp_servers.hephaestus-network.env subtable when it refreshes the canonical command and arguments. A changed policy takes effect in a fresh host session.