AI Native Model Router

AI Native Model Router

Use this Skill as a deterministic route/v1 resolver with additive route/v2 assurance routing. It owns project-local selection and returns a machine-readable RouteDecision; it does not execute a host, call a provider, load credentials, or claim that a route was enforced.

Version 0.4.0 removes the bundled route/v1 Profile. The only bundled Profile is gpt5.6, and it uses route/v2. The route/v1 schemas and APIs remain available for explicitly selected safe project Profiles. Retired Profile IDs are rejected before configuration, discovery, or Profile loading; they are not aliases or supported discovery names.

Contract navigation

• Read interface for the v1 and v2 RouteRequest and RouteDecision contracts.
• Read configuration for selection precedence, version pairing, digest checks, safe project profiles, and collision handling.
• Read provider evidence before relying on a dated identity or fallback claim.
• Use resolve_route.py for validation and resolution. The Phase 1 config helper remains router_config.py.
• The package contract is in provider-catalog.json, the immutable bundled gpt5.6 profile, and the v1 config, v1 request, v1 decision, v2 config, v2 request, and v2 decision schemas.

Operating rules

1. Every request explicitly selects a profile. The bundled gpt5.6 profile is default_active: false with explicit-owner-selection; file presence never activates a profile. Require explicit task selection for writer.high-volume-deterministic.
2. Preserve route/v1 and its validator.independent independent-validation semantics. Mixed v1/v2 request, config, and profile versions fail closed.
3. route/v2 uses validator.assurance: same-model review is allowed and is reported through same_model_as_writer; it is not independent validation. R1 is gpt-5.6-luna:max -> gpt-5.6-terra:max -> gpt-5.6-sol:high. R2 is gpt-5.6-terra:max -> gpt-5.6-sol:high. R3 requires gpt-5.6-terra:max + gpt-5.6-sol:high and must never degrade to one Validator. For R3, apply this precedence: missing or unaccepted route-bound evidence for any unavailable required route => decision_status: blocked, even when the other required route is unknown; otherwise, any required route with unknown availability => decision_status: unknown; otherwise, accepted unavailability => decision_status: blocked. R3 never selects or degrades to one route.
4. route-bound evidence may be only model-not-found, authenticated-provider-outage, quota-exhaustion, or repeated-bounded-transport-failure, keyed to the unavailable route. Treat unknown as unknown; unknown remains unknown after the R3 precedence above; never infer availability. A valid Validator rejection is not a route failure. A substantive Validator rejection does not trigger escalation; create a new candidate ID before revalidation.
5. Resolve a fallback only for an unavailable route with complete accepted evidence and an available next route. Return enforcement_status: not-executed always; this is not an execution receipt.
6. Keep provider execution, authentication, and transport outside this package. No secret, endpoint, credential, transport command, or price value belongs in configuration, profiles, catalog data, prompts, or evidence.

CLI

python scripts/resolve_route.py list-profiles
python scripts/resolve_route.py validate-profile --profile <id>
python scripts/resolve_route.py resolve --request <json-path-or-json> [--config <path>] [--injected-config <path-or-json>]

CLI success is JSON on stdout. Errors are JSON-free diagnostics on stderr with a nonzero exit code. Ordering is deterministic.