model-routing

name	model-routing
description	Resolve one bounded model, effort, transport, and budget decision through railyard/model-routing/v1. Use before model-specific task, subagent, provider-review, or steering actions.

Model routing

railyard:model-routing is the only public model, effort, budget, and transport-policy entrypoint. It returns a frozen decision or receipt; it never creates a task, invokes a provider, runs a browser, or executes a command on a provider's behalf.

../../references/model-routing.md is the normative contract — commands, request shapes, catalog schema, budget rules, CE override clauses, and state rules live there, and its transport phase incorporates the provider-task policy. Do not invoke a second router or copy a model table. For per-harness session defaults and cross-harness invocation, see ../../references/harness-model-invocation.md.

Activation

Resolve SKILL_DIR from the activated skill path (never an installed-cache path or source checkout), then:

ROUTER="$SKILL_DIR/../../scripts/model-routing.mjs"
printf '%s\n' '<request JSON>' | node "$ROUTER"

Every request carries exact "contractVersion":"railyard/model-routing/v1". Requests are content-free: no prompts, task titles, paths, source, files, tokens, endpoints, or command text.

Lifecycle

1. Classify the bounded destination work: role, categorical work shape, adapter/dispatch kind, scope, and privacy. Runtime and transport facts are router-owned, never caller JSON.
2. resolve reads one immutable policy snapshot. No-config defaults: Sol high/max for orchestration and review, Luna at max for implementation, Terra at max only as the router-attested Luna substitute. When the selected policy attributes providers to a harness, include harness:"claude" or harness:"codex" in the request. A mismatched provider needs a specific crossHarnessReason; the resolver copies it into the decision. A catalog, request, or environment variable cannot nominate Terra or mark Luna unavailable.
3. For configured work-starting actions, admit with a stable caller-generated requestId, a frozen artifact digest, and every applicable scope; then claim-dispatch immediately before the one carrier dispatch. A claim is one-way and cannot authorize a retry spawn.
4. The owning workflow invokes only the selected fixed adapter/carrier and reconciles its receipt through the fixed importer. Model output and caller-authored JSON are not receipts; adapters outside the fixed set return transport_unsupported.
5. status, inspect-claim, local-only refresh, and learning inspect|clear|disable|enable go only through this contract; refresh never probes a remote provider.

Read the decision's disclosure (and any fallback/settlement disclosure) as the content-free R28 record; do not reconstruct it from provider output.

For a budget-neutral status/narrowing message, use resolve with budgetEffect:"none", a stable actionId, and the exact prior route binding; scope-expanding steering uses adjust_active against the active reservation. Unknown or changed work class blocks inheritance (prior_route_unknown) rather than silently reusing a route. The reference defines the binding fields and the closed action-receipt schema.

Fixed transports

The full adapter/carrier tables are in the reference. Operative rules:

• Native task creation may carry contextFork:"none" or "1"–"999"; everything else is rejected.
• A visible task create needs a one-use mint-task-authority receipt from the fixed in-process user-turn attestor; public stdin and CODEX_* variables cannot mint it.
• Fable/Opus review runs only through the supported Compound Engineering Claude -p adapter; this skill never builds a parallel Claude runner.
• Oracle review is oracle-browser on chatgpt_current_pro only; lifecycle is the separate oracle-homebrew-lifecycle carrier. Oracle API is unsupported here.
• GLM is Codex-only (below). The glm-5-2-scout/glm-5-2-engineer carrier rows remain fail-closed (transport_unsupported); they are never a Codex selector or native-subagent value.

Harness defaults and GLM

The router's frozen no-config route is harness-independent. The owner catalog uses per-harness role tiers: hard implementation is Sol max/Fable high/max, medium or long-running implementation is Terra max/Sonnet medium, mechanical implementation is Luna max/Haiku low, and plain Codex implementation defaults to Luna max. Effort is part of every route. Escalate deliberately. The harness-model-invocation reference has the table, the current rate data, and why sticker rates settle almost nothing (meters differ, operating points differ, cache rates dominate).

Dispatch rule — explicit model and effort on every subagent, thread, or worker, no exceptions. Harness children inherit the session model when the dispatch omits one; on a premium-tier session (Fable, Sol max) that silently runs workers at the top tier — the inversion this router exists to prevent, and it burns the premium meter without consent. Every child dispatch therefore names its model and effort from the table above, using the harness's exact parameters: Claude Code Agent sets model; Codex spawn_agent sets model + reasoning_effort (it has no provider field — a non-OpenAI child requires its thread already on that provider); Codex thread/start sets model + config.model_reasoning_effort plus modelProvider for any non-OpenAI model; codex exec uses -m plus -c model_provider= and -c model_reasoning_effort=. A child deliberately run on the session's own tier is a named escalation with its reason stated at dispatch. Silent inheritance is a routing violation. Cross-harness dispatch is additionally opt-in only: never a silent default, because each harness meters separately.

Dispatch banner — every dispatch prompt makes the child announce its route. The dispatcher composes the line (the worker cannot introspect its own model or effort) and instructs the child to echo it verbatim as the first line of its first message, then proceed without waiting; a model/effort change mid-thread gets a ▸ route change: line the same way, and is N/A on harnesses that cannot change either mid-thread. The canonical format, per-harness fields, and mid-thread table are in the harness-model-invocation reference.

Claude Code cannot invoke GLM-5.2. A Claude session cannot both authenticate to Z.ai and keep its account-bound capabilities, so the route does not exist — do not rebuild it. GLM work goes through Codex: codex exec -m glm-5.2 -c model_provider=zai_litellm, available whenever the provider config and local proxy are present; a failed command is the availability check. GLM's case is subscription headroom (Z.ai Coding Plan credits, a meter that never converts to USD) and provider diversity, not price. Benchmarks are not a selection criterion.

Cross-harness: Claude reaches Codex models through the codex plugin's codex:rescue skill (or a direct codex CLI invocation); Codex reaches Claude models through claude -p.

Work contract and CE overrides

Use build-work-contract for a carrier-neutral execution envelope: supply the seven semantic digests plus the selected carrier/model/effort, preserve the invariant digest, and apply the returned source-owned presentation overlay (lean brief for GPT/Sol, complete spec for Opus, autonomy/pause boundaries for Fable, standards-plus-plan for GLM, self-contained one-shot for Oracle). Direct user and repository instructions outrank the overlay. Changed semantic inputs against a frozen digest return invariant_contract_mutation.

Compound Engineering is never edited. A frozen decision may replace only a named CE execution mechanism at a seam in the closed registry, using the runtime replacement clause in the reference; workflow, persona, legitimacy, writer, review, and terminal authority stay with CE. If the exact carrier/seam is not attested, return transport_unsupported or the disclosed fallback.

Completion truth

offline_implementation_ready, host_capability_attested, and live_carrier_verified are distinct; offline tests prove only the first, and the public CLI cannot mint positive capability evidence from JSON. Never call an optional route live-verified without its separately authorized minimal canary or an equivalently bound trusted receipt.