@phillarmonic/dsh-llm-kimi
用于 DeepSeek Harness llm 能力接口的 Kimi K3 连接器插件。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:phillarmonic/dsh-llm-kimi#a132c76baf2aaa42089cc30fb6bd79d8729d9e2cLlm Kimi 以 @phillarmonic/dsh-llm-kimi 发布，当前版本为 0.3.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A Kimi K3 connector plugin for the DeepSeek Harness llm capability seam.
It registers a direct-fetch LlmAdapter for the kimi-code provider route, streaming Kimi Code chat completions (an OpenAI-compatible dialect) as harness StreamChunks. Connection facts resolve per request, so a changed base URL, catalog, or key reaches the next request without a restart.
Install as a DeepSeek Harness bundle so dsh plugin registers it automatically:
dsh plugin --profile <name> add @phillarmonic/dsh-llm-kimi
dsh --profile <name>
Or add it manually when you compose cordis.yml directly:
pnpm add @phillarmonic/dsh-llm-kimi
The harness packages are peer dependencies; install them alongside the plugin if they are not already present:
pnpm add @deepseek-ai/cordis @deepseek-ai/dsh-llm @deepseek-ai/dsh-credentials \
@deepseek-ai/dsh-settings @deepseek-ai/dsh-launch-environment @deepseek-ai/dsh-timeout \
@deepseek-ai/dsh-attachment @deepseek-ai/schemastery
Add the plugin to your cordis.yml. Every field is optional; the defaults target the public Kimi Code endpoint.
plugins:
llm: {}
'@phillarmonic/dsh-llm-kimi':
apiKeyEnv: KIMI_CODE_API_KEY
reasoningEffort: low
Then export your key (from the Kimi Code console at api.kimi.com) or store it through the harness credentials service:
export KIMI_CODE_API_KEY=sk-...
Select the provider and a model when you run a task, for example provider kimi-code with model k3.
| Field | Default | Description |
|---|---|---|
apiKeyEnv | KIMI_CODE_API_KEY | Credential reference (environment-variable name) resolved per request. |
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. |
streamIdleTimeoutMs | 300000 | Maximum provider idle time while one stream read is outstanding. |
retryPolicy | normal, five retries | Provider-owned model-request retry policy. |
| Model id | Context | Reasoning effort | Image input |
|---|---|---|---|
k3 | 1,048,576 | low / high / max | yes |
k3-256k | 262,144 | low / high / max | yes |
kimi-for-coding | 262,144 | always on (no effort levels) | yes |
kimi-for-coding-highspeed | 262,144 | always on (no effort levels) | yes |
All four Kimi Code models accept images. When a request carries image content, the adapter reads each image through the harness attachment service (ctx.attachments) and inlines it as a base64 image_url data URL. The attachment service is optional: a request with images fails loudly when it is not mounted, and text-only requests never touch it.
Budgeting is enforced inside this plugin from the durable attachment metadata, before any bytes are read. An image whose pixel count exceeds imagePixelBudget, or whose encoded byte length exceeds imageMaxBytes, is rejected rather than pushing the request past Kimi's per-message size limit. A request that sends image content to a model configured without image support is rejected too.
temperature, top_p, or n unless sendTemperature is enabled.reasoning_content. The adapter replays the harness reasoning block onto that field so multi-turn tool sessions stay valid.off or none effort. The K3 family exposes low, high, and max; the coding models keep thinking on with no effort selector.pnpm install
pnpm run typecheck
pnpm run test
pnpm run build
MIT. See LICENSE.