dsh-llm-longcat

When plugging non-OpenAI models into DeepSeek Harness, the details that usually bite — SSE frame ordering, reasoning-token parsing, tool-call assembly — demand a hand-rolled adapter for every vendor. dsh-llm-longcat wraps LongCat-2.0's 1,048,576-token context window, its reasoning_content thinking stream, and full function-calling support into Harness's standard LLM seam, so an agent can switch models from the selector and start chatting without anyone touching SSE usage-before-finish ordering or the explicit-null continuation-delta edge cases.

The adapter makes protocol-level choices deliberately: thinking mode is a binary toggle (enabled / disabled) rather than a fake low-medium-high gradient; tool-call arguments stay a raw JSON string end-to-end to avoid mid-stream deserialization; reasoning_content is replayed into history only on turns that carried tool calls and dropped otherwise to save tokens; cached prompt tokens are subtracted from input count to honour Harness's disjoint-accounting convention. Credentials resolve per request from ctx.credentials or the environment — no secret written to disk.

Built for developers already running multi-model agents in Harness who want LongCat's million-token context for document analysis or long-horizon reasoning. Image input and stop sequences are not supported, and reasoning has exactly two levels — all stated upfront in the README.