Skip to content

[routing]

Default routing strategy, model denylist, role aliases, provider failover and account selection.

Key Type Default Description
routing.strategy string "balanced" balanced | quality | cheap | fast | private
routing.unknown_output_limit_fail_closed_threshold integer 100000 An explicit-model request whose output limit is unknown is refused when it asks for more than this many output tokens. 0 or absent = 100000.
routing.auto_denylist array of string auto_denylist lists provider/model candidate IDs excluded from AUTO (selector/orchestration) candidate selection everywhere models are chosen automatically. Explicit-model requests bypass candidate selection and still route to a denied model; denied models stay discovered (catalog and /prism/models keep listing them). Entries match a full candidate ID case-insensitively; one trailing ‘’ is a prefix wildcard (“openrouter/”). Empty or absent = no denial. See routing. NewAutoDenylist for the exact semantics.
routing.role_aliases table of array of string role_aliases maps a role name, sent as a request’s model, to an ordered list of concrete model ids. triage, reasoning and cheap-toolcall are built in; any other key declares a new role (1-64 letters, digits, ’. ’, ‘_’ or ‘-’). Prism tries every provider of one target before the next target. An empty list leaves the role unconfigured.
routing.role_alias_params table of table of interface role_alias_params are default request parameters per role alias: [routing.role_alias_params. ] key = value. They fill parameters a request to that alias omits and never overwrite one it sends, so a fixed-shape client (an appliance that cannot send chat_template_kwargs or reasoning) can reach a model through its own alias while the model stays unchanged under its real name.
routing.conversation_failover string "plan" conversation_failover is the owner’s choice of what a conversation that continues gets when its model does not answer: “plan” (the default) walks the planned order, usually another model; “same_model” first tries the same model at its other providers, so the conversation keeps its model. The cache is lost either way: each provider caches on its own.
routing.allow_provider_failover bool false allow_provider_failover permits a provider-scoped request (“provider::model”) to be served by another provider that carries the same model when the named one cannot serve. Default false: the pin is the caller’s cost and provider choice, so a refusal is reported with its own reason instead of spending on an upstream the caller did not name. A request may override this per call with the X-Prism-Allow-Provider-Failover header.
routing.provider_chain array of string provider_chain is the initial seller order: for a model several providers serve, auto routing and a bare model id take it from the first available provider in the list, and the next ones back it up. A provider in the chain enters auto routing even for a model without a known price. The console’s routing page replaces it at runtime; empty means no chain.
routing.account_pool bool true account_pool makes every attempt pick its account (route) through the account pool, under policy “ranked”, which reproduces the router’s choice. Default true. Set false to keep the router’s route for every attempt; the pool then picks in shadow, counting disagreements in /prism/status and /metrics.
routing.app_profiles table of string app_profiles maps an X-Prism-App name to a profile from [profiles. ]; automatic requests from that app use the profile.