Providers

Every run asks for agent-default unless its class picks otherwise: GLM-5.3 on Z.ai for public work, Claude for internal work. Two more providers are available for public data only, and selectable per class. Nothing changes by default. Each run’s cost is metered at the provider’s list price, like any other run. Mistral is not recommended as an implementer: it can spend a run’s whole token budget on a small change. Their state is on the status page.

Alias	Model	Tier	Processed in
`kimi-frontier`	Kimi K3.1 Pro (`k3.1-pro`, Kimi Code subscription)	frontier	China
`mistral-standard`	Mistral Medium 3.5 (`mistral-medium-2604`)	standard	EU
`mistral-light`	Mistral Small 4 (`mistral-small-2603`)	light	EU

Public data only. Both are processors approved for public repositories and tasks, not for internal data. Internal tasks never get them, held three times: the gateway routes them only from its public listener, only the public capability profiles list their aliases, and the factory applies a class’s choice to public tasks only.

Switch a class on

No class names a provider by default. One key on the class’s loops row, in tooling/base/agent-factory/helm-values-configmap.yaml: models.<role>.<tier> is the alias that role runs on for a public task of that tier.

loops:
  docs: {scope: scoper, …, models: {implementer: {standard: mistral-standard}, reviewer: {frontier: kimi-frontier}}}

Rule	Why
Each alias serves its own tier only; the factory refuses one on another tier at config load	The run caps stay the tier’s, and the gateway’s hourly token limit stays above a run’s cap spent on cache reads
A role or tier without an entry keeps today’s model	No default changes
`loops.<class>.judge` may name `mistral-light`	The judge’s field is already an alias
Each run is priced at the provider’s list price, cache reads included (`meter.prices`)	The run’s dollar cap and spend per identity stay true

Mistral reasons with reasoning_effort: "none". agentgateway v1.6.0 parses a Chat Completions message’s content as a string only, and Mistral answers with a list of chunks when it reasons, so the mistral backend forces reasoning_effort to "none" (infrastructure/base/agent-gateway/mistral.yaml). Kimi is unaffected.

A run never switches model mid-conversation, and there is no failover to another provider: a provider’s 429 ends the run, as Z.ai’s does.

Owner prerequisites

The agents’ own keys, in OpenBao: bao kv put -mount=agents moonshot api_key=- and bao kv put -mount=agents mistral api_key=-. Set a spend cap for each key on its provider’s console. A smoke test through the gateway is in runbook 14.

Runtime RunLore