Configuration¶

Every field lives under the plugin entry in cordis.yml and is optional. Defaults target the public Kimi Code endpoint with fixed sampling. Settings reload live: a changed value reaches the next request without a restart.

plugins:
    llm: { }
    '@phillarmonic/dsh-llm-kimi':
        apiKeyEnv: KIMI_CODE_API_KEY
        reasoningEffort: low
        sendTemperature: false

Reference¶

Field	Default	Description
`apiKeyEnv`	`KIMI_CODE_API_KEY`	Credential reference (environment-variable name) resolved per request. Editable from the Settings card as API key reference.
`baseURL`	`https://api.kimi.com/coding/v1`	Endpoint base; `/chat/completions` is appended. Falls back to `$KIMI_BASE_URL` from a trusted environment layer.
`reasoningEffort`	`low`	Default thinking effort for the K3 family (`low`, `high`, `max`).
`sendTemperature`	`false`	Whether to forward an explicit sampling temperature. Kimi enforces fixed sampling, so this stays off by default.
`firstToolChoice`	unset	Force this tool on a conversation's first tool call (wire `tool_choice`), applied only while the history holds no earlier tool call and the request's tool list carries the name.
`maxTokens`	`131072`	Default per-request output cap; a model's own cap and explicit request values win.
`defaultContextWindow`	`1048576`	Context capacity used when the selected model has no exact value.
`imagePixelBudget`	`8294400`	Reject a request image whose intrinsic pixel count exceeds this budget. Default is Kimi's 4K ceiling (3840x2160).
`imageMaxBytes`	`1048576`	Reject a request image whose encoded byte length exceeds this budget. Keeps inline base64 within Kimi's per-message size limit.
`models`	four Kimi Code models	Advisory catalog shown by discovery consumers. See Models.
`streamIdleTimeoutMs`	`300000`	Maximum provider idle time while one stream read is outstanding.
`retryPolicy`	normal, five retries	Provider-owned model-request retry policy.

Reasoning effort¶

The K3 family (k3, k3-256k) accepts low, high, and max. An explicit per-request effort wins over the default. The always-thinking coding models (kimi-for-coding, kimi-for-coding-highspeed) have no effort selector, and the adapter never sends the field to them.

Kimi cannot disable thinking without routing to a weaker model, so the adapter never sends an off or none effort.

First tool choice¶

Some workflows declare a contract for the conversation's first tool call (a slash command that must open with one specific MCP call, for example). Models occasionally intend the right call but emit a neighboring tool name. Setting firstToolChoice pins that opening call: the adapter sends OpenAI-style tool_choice on requests whose history holds no earlier tool call and whose tool list carries the named tool. Sessions without that tool, and every request after the first call, stay unconstrained.

plugins:
    '@phillarmonic/dsh-llm-kimi':
        firstToolChoice: mcp__acme__acme_overview

The field also appears on the plugin's Settings card (web profile) alongside the endpoint, reasoning, and sampling fields.

Endpoint override¶

For a private or proxied endpoint, set baseURL directly, or export KIMI_BASE_URL from a trusted environment layer. An explicit baseURL in cordis.yml wins over the environment.

Image budgets¶

imagePixelBudget and imageMaxBytes are enforced inside the plugin from the durable attachment metadata, before any bytes are read. See Image input for the full flow.

Custom model catalog¶

Override models to advertise a different catalog to discovery consumers. Each entry keeps its supportsReasoning and supportsImage flags; requests to the endpoint remain unrestricted regardless of the advisory catalog. See Models.