deepseek-harness-friendli

A DeepSeek Harness LLM adapter for FriendliAI serverless inference. Friendli speaks the OpenAI-compatible chat-completions API, so this plugin implements the harness LlmAdapter contract over it and pulls its model catalog from Friendli at runtime.

What you get

• Streaming chat completions against https://api.friendli.ai/serverless/v1, translated into the harness StreamChunk protocol (text, reasoning, and tool-call blocks).
• Dynamic model discovery — the catalog comes from GET /models at runtime; nothing is hand-maintained. Deprecated entries are dropped.
• Reasoning support for controllable models via parse_reasoning + include_reasoning, with an on/off thinking toggle mapped to Friendli's chat_template_kwargs.enable_thinking.
• Stable error codes — HTTP and transport failures become LlmError with codes (AUTH, RATE_LIMIT, MODEL_NOT_FOUND, …); model-list failures are distinct from inference failures.
• AbortSignal forwarding and mandatory attributionHeaders() on every request, per the harness adapter contract.

Install

Until this is published to npm, install from GitHub:

pnpm add github:Lee-Si-Yoon/deepseek-harness-friendli

Peer dependencies (provided by your harness composition): @deepseek-ai/dsh-llm, @deepseek-ai/cordis, @deepseek-ai/schemastery.

Authenticate

The token is read from the environment at each request — never inline it in a config file or commit it.

export FRIENDLI_TOKEN="flp_..."

Get a token from the Friendli dashboard.

Configure

Add the adapter to your cordis.yml and point an agent at a Friendli model id. See examples/cordis.yml for a complete fragment.

- id: deepseek-harness-friendli
  name: 'deepseek-harness-friendli'
  config:
    apiKeyEnv: FRIENDLI_TOKEN                       # default
    baseURL: https://api.friendli.ai/serverless/v1  # default
    providers: [friendli]                           # default
    thinking: enabled                               # optional; on/off for controllable models

Config field	Default	Meaning
apiKeyEnv	FRIENDLI_TOKEN	Environment variable holding the bearer token
baseURL	serverless endpoint	Endpoint base; /chat/completions and /models are appended
providers	[friendli]	Provider route(s) this adapter serves
thinking	model default	enabled / disabled for controllable-reasoning models; ignored by always-reasoning models
modelCacheTtlMs	60000	How long the fetched model catalog is cached

Use a model

Reference a model by the exact id returned by GET /models:

- id: agent-loop
  name: '@deepseek-ai/dsh-agent-loop'
  config:
    agents:
      - id: main
        provider: friendli
        model: zai-org/GLM-5.2

Dynamic model discovery

Available models are fetched from Friendli Serverless at runtime, so the list reflects the current catalog rather than a snapshot baked into this package. The set of models — and whether a given request is accepted — depends on the account, region, permissions, and model lifecycle. A model appearing in the public /models response does not guarantee every request for it will succeed; permission and lifecycle checks still apply at inference time and surface as an LlmError.

Reasoning

Friendli parses reasoning model-agnostically. This adapter always requests parse_reasoning: true + include_reasoning: true, so reasoning tokens arrive on delta.reasoning_content and become harness reasoning blocks.

Models fall into two kinds (Friendli docs):

• Always-reasoning (e.g. MiniMaxAI/MiniMax-M2.5) — reason regardless; thinking config has no effect.
• Controllable (e.g. zai-org/GLM-5.2) — thinking: enabled|disabled maps to chat_template_kwargs.enable_thinking. The adapter advertises an on/off reasoning toggle only for models whose /models entry declares a toggle capability. Discrete effort levels are not exposed, because the serverless reasoning API is documented only for the on/off enable_thinking switch — offering effort levels would promise choices the wire cannot honor.

Troubleshooting

Symptom	Meaning
MISSING_CREDENTIAL	FRIENDLI_TOKEN (or your apiKeyEnv) is unset or blank
AUTH (401)	Invalid or expired token
FORBIDDEN (403)	Account lacks access to the model
MODEL_NOT_FOUND (404)	Model id unavailable or deprecated
RATE_LIMIT (429)	Friendli rate limit reached
SERVER (5xx)	Temporary provider-side failure
MODEL_LIST_FAILED	Runtime GET /models discovery failed (distinct from an inference failure)

Develop

pnpm install
pnpm run lint:all   # typecheck + tests in one pass
pnpm run build

Publish to the community plugin store

There is no official DeepSeek plugin hub. The community DSH Plugin Store crawls every GitHub repo tagged with the dsh-plugin topic hourly — no PR or review. This repo carries that topic, so once it is public it is picked up automatically. The store performs no vetting; review any third-party plugin before installing.

License

MIT © Lee-Si-Yoon