dsh-llm-gateway-compat

English | 中文

A community bundle compatible with DeepSeek Harness (DSH) 0.1.5-rc.1 or later. It stops empty streamed tool-call identity from wiping the call, turns the two most common request 400s into an llm-pi-ai compat write plus one retry, and can own Chat Completions routes that default to system / max_tokens.

This is not an official DeepSeek package. It is not endorsed by DeepSeek.

0.4.0 tracks the current harness APIs (ToolCallId, ctx.settings.installSection). It will not load against npm 0.0.1-rc.1.

What it does

Streamed tool-call identity (v0.1)

Wraps llm/stream so later SSE fragments with empty id / name cannot overwrite a nonempty value. Synthesizes compat_call_<index> when no id ever arrives. Official DeepSeek streams stay unchanged when identity is already stable.

Request dialect 400s (v0.2)

On agent/request-error, classifies developer-role and max_completion_tokens refusals, writes the matching field into official llm-pi-ai settings, retries the same step once, and injects a logged plugin notice. Generic 400s are not retried.

Chat Completions adapter (v0.3)

Optional routes under llm-gateway-compat.providers. Each route is a direct POST {baseURL}/chat/completions adapter with gateway-safe defaults:

• system prompt is always role: system
• output cap is always max_tokens
• empty tool-call id/name never overwrite, even if stream sanitizing is off
• extraBody for fields the harness vocabulary does not own (user, prompt_cache_key)
• extra headers, Authorization: Bearer or DashScope api-key
• thinking dialect: reasoning_content (default), thinking, think-tags, or none

Route ids must not collide with llm-deepseek or llm-pi-ai. Pick a new id such as dashscope-compat.

Install

From GitHub (ships built lib/, no install-time TypeScript build):

dsh plugin --profile web add github:snowshadow/dsh-llm-gateway-compat

Restart dsh web. Pin a commit (github:snowshadow/dsh-llm-gateway-compat#<sha>) so a later push cannot change what runs. Only add packages whose source you trust.

From a local checkout:

dsh plugin --profile web add /absolute/path/to/dsh-llm-gateway-compat

Config

Plugin switches (also live under $DSH_HOME/settings.yaml as llm-gateway-compat:):

key	default	meaning
enabled	true	master switch for stream wrapping and 400 recovery
diagnose	true	classify known gateway 400s and inject a YAML snippet
autoApplyCompat	true	persist the matching llm-pi-ai compat field and retry once
providers	{}	Chat Completions routes this plugin owns

One gateway route:

# $DSH_HOME/settings.yaml
llm-gateway-compat:
  providers:
    dashscope-compat:
      displayName: DashScope
      baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1
      apiKeyEnv: DASHSCOPE_API_KEY
      authHeader: bearer
      thinkingFormat: reasoning_content
      extraBody:
        user: harness
      models:
        - id: deepseek-v4-flash
          name: DeepSeek V4 Flash

Store DASHSCOPE_API_KEY on the Models page or in $DSH_HOME/.credentials.yaml, or export it in the environment that launches dsh. Include /v1 (or /compatible-mode/v1) in baseURL when the gateway requires it. Then select the dashscope-compat / deepseek-v4-flash route in the model picker.

Provider fields:

key	default	meaning
baseURL	required	origin plus path prefix; /chat/completions is appended
apiKeyEnv	required	credential reference: Models page / $DSH_HOME/.credentials.yaml, then that environment variable
authHeader	bearer	bearer or api-key
models	[]	advisory catalog; unlisted ids still resolve as text-only
extraBody	—	merged under harness-owned fields; max_completion_tokens is stripped
headers	—	extra request headers; User-Agent still comes from harness attribution
thinkingFormat	reasoning_content	history + stream reasoning dialect
includeUsage	true	send stream_options.include_usage

Develop

Typecheck and tests expect a sibling deepseek-harness checkout at ../deepseek-harness (0.1.5-rc.1 APIs). pnpm test / pnpm run build link @deepseek-ai/dsh-llm, dsh-settings, and cordis to that checkout.

pnpm install
pnpm test
pnpm run build

Known limitations

• Cannot recover a tool name the gateway never emitted.
• Image input is refused (UNSUPPORTED_CONTENT).
• think-tags is applied on replayed assistant history, not on partial streamed tags.
• No idle-stream watchdog; caller AbortSignal is honored.
• Auto-apply only writes supportsDeveloperRole: false and maxTokensField: max_tokens on llm-pi-ai routes.
• There is no Web settings card; edit settings.yaml or the profile patch.

License

MIT