ffyuuu/dsh-llm-longcat
dsh-llm-longcat adds the LongCat-2.0 provider to the DSH LLM seam: 1M-token context, thinking mode (its reasoning_content field is translated into harness ReasoningBlocks and replayed on tool-call turns), full function calling with raw-JSON arguments, streaming with usage-before-finish ordering, and per-request key resolution from the credential seam.
dsh-llm-longcat adds the LongCat-2.0 provider to the DeepSeek Harness LLM seam with 1M-token context, thinking mode and tool calling. It recognizes LongCat's reasoning_content field and translates it into harness ReasoningBlocks, keeps tool-call arguments as a raw JSON string end to end, and replays reasoning_content on tool-call turns as thinking-mode passback requires. Streaming uses SSE with the usage-before-finish ordering the harness relies on, and the key resolves per request from the credential seam or environment. Config options include apiKeyEnv (default LONGCAT_API_KEY), an optional baseURL ($LONGCAT_BASE_URL then the public API), a thinking deployment policy, reasoningEffort (off | high), maxTokens, defaultContextWindow and a retryPolicy.
Install
dsh plugin --profile default add github:ffyuuu/dsh-llm-longcatCopied verbatim from the README's Install block. The README also documents a commit-pinned form (dsh plugin --profile default add github:ffyuuu/dsh-llm-longcat#3dcb3b1b...) and notes the key may instead be stored through the Web UI's Models page. No npm package is documented, so the GitHub form is shipped.
Compatibility
The adapter registers LongCat as a model provider on the harness LLM seam and resolves the key per request from ctx.credentials or the environment, so no secret is stored in any config file. LongCat controls thinking with thinking: {type: enabled|disabled} and does not accept OpenAI's top-level reasoning_effort, so the adapter offers exactly two levels (off / high) rather than advertising a gradient that would collapse onto one value. A llm-longcat: section in $DSH_HOME/settings.yaml overrides any field - base URL, catalog, request defaults and idle budget - without a restart, while an in-flight stream keeps the facts it started with. The README also documents an upstream packaging state where installing the harness itself can fail on a dsh-typert-protocol version drift, with the workaround of pinning the @deepseek-ai/* packages through npm overrides.
Details
- Repo: ffyuuu/dsh-llm-longcat
- Category: Models & Providers
- Stars: 2
- Version: GitHub repo ffyuuu/dsh-llm-longcat (2*); installed from GitHub (no npm package documented). Declares >=0.1.0-rc.7. Supported model: LongCat-2.0 (1,048,576 context, 131,072 max output, text-only; thinking + tool calling). English README.
Alternatives
1624318455/dsh-plugin-adapter · aa2246740/dsh-antigravity-oauth · AmmarByFar/dsh-trustedrouter