scode-model-routing specification
Dependencies: none
NOTE: This file is binding on the skill's text. It is deliberately sparse: it records only requirements that have been stated as such, not a description of everything the skill does. Absence of an entry means the behavior is not yet specified, not that it is unspecified on purpose. When the skill and this file disagree, that is a bug in one of them; fix the skill or change this file in the same change, never leave them apart.
Requirements
- The caller's task establishes its role. Harness, provider, model tier, and the inventory's
sotamark never disqualify a caller from orchestrating. Native capabilities and process constraints determine available launch mechanisms. A family with no suitable model may route across families or returnno suitable route; neither answer reclassifies the caller as a delegate. - The skill depends on no other skill, and its text refers to no other skill: no pointer upward or sideways into a consumer (a named skill or one of its files, "the gate", "the checkpoint protocol" as a thing the reader should go and read). Where the text needs a fact only a caller has, it names that fact as a request input and stops there.
- Loading the skill is side-effect free. Invoking it activates nothing for the session, writes nothing, and claims
nothing about later spawns; its
SKILL.mdsays so in its first paragraph and its description says it is loaded by other skills and inert alone. No harness-level switch turns off description-based selection: Codex'sallow_implicit_invocation: falseinagents/openai.yamlis also read by Muse Code, which then refuses to load the skill for the model at all (verified on Muse Code 1.0.2), and a skill that other skills load by name cannot carry it. The guarantee is inertness: an unsolicited load does nothing. - Routing answers a request of the shape
SKILL.mddescribes under "The routing request" with an answer of the shape under "The routing answer": a route (a model id with effort,orchestrator,inherit, orno suitable route), a launch mechanism (nativeorexternal), a verbatimlaunch preferencewhen the config file states one for an external route's model, theroute exhaustedandendpoint trustedfacts, the preference and cross-family divergence flags with reasons, and a one-line reason. A profile-table cell that readsnoneis never returned; it resolves inside routing. Routing is a pure function of the request and the local environment: it holds no state between requests and never decides whether to delegate or to escalate. - Routing never names or checks a CLI. Its mechanism answer is only
nativeorexternal, because that is all a model choice depends on (native-path bias, native-only process roles, writer resume, alias-only native model parameters); which program runs an external route, and whether one installed here can, is the launcher's decision. An external route that cannot launch comes back in the caller's list of models unreachable on a path (input 13), which routing treats as unavailable on that path for every unit; a demanded model unavailable on every path isno suitable route, never a substitute. The one exception to never naming a CLI is thelaunch preferencefield, which passes the user's own config-file words through unchanged. This keeps the CLI choice in one place instead of a second table that drifts. - Routing reads the model routing config file,
~/.scode-model-routing.md, itself when answering; it is not a request input. If that file is absent and~/.scode-galaxy-brainrc.mdis present, routing stops and tells the user to rename it rather than answering as if no config existed. Prose in this skill calls it "the model routing config file", never "the rc file". - The effort words attached to models in the inventory and profile tables are calibration defaults, not a whitelist of what a model may run. An explicit user demand for a model at an effort the tables do not list routes to that model at that effort (input 8); the tables' omission is never a reason to block, ask for confirmation, or substitute another model. What the launch mechanism can actually select (input 11) is the limit on a demanded effort.
- The inventory and profile tables name Claude models by harness alias alone (
haiku,sonnet,opus,fable), with no version number, because the alias is the launch argument and follows each new release. Only ids passed literally (GPT, Muse, GLM) carry a version there. The skill's former versioned labels (fable-5,sonnet-5.5, and so on) mean the plain alias. An explicit demand for a specific older Claude release is the one exception: routing answers it with that release's full model id, onexternalwhen the native mechanism takes only aliases. - A request that names a model by its line alone ("sol high", "sonnet medium") resolves to that line's model in the inventory, never to an older release of the line. A request that names a model by version (other than the former Claude labels) is an explicit demand even when the inventory does not list that version; absence from the inventory is not an availability fact.
SKILL.mdis the public surface and is kept lean, measured as what a session loads before its first delegation: everything needed to answer a routing request is inSKILL.md, and only the full inventory table, the calibration history, the size anchors, and config-file handling live in sidecars read on demand, each with its trigger named inSKILL.md.- A same-named project-local skill can shadow the installed one on every harness whose loader honors project-local skills. That is accepted; the consumer's name check proves identity, not revision.
- The skill is meant to work on modern Linux and macOS. Commands, paths, and tools it prescribes must be available on both; nothing may rely on one without an equivalent for the other. No other platform is of concern, and the skill's text need not accommodate one.
- Under GPT routing, idiomaticity, AI-slop, and docs/comment correctness reviews start at luna high. Data-flow and edge-input correctness lenses use the same focused-review route when the panel retains strong general correctness coverage. Its escalation and large-input route is sol high. Simplification remains mechanical review; standalone correctness, the remaining correctness lenses, security, test quality, and SPEC compliance remain critical review. This is a GPT profile default, subject to ordinary demand, availability, mechanism, and preference precedence; other families retain their existing review routes. The review charter's discovery scope is unchanged.