dsh-llm-longcat

LongCat adapter for the DeepSeek Harness LLM seam.

Adds LongCat-2.0 as a model provider: 1M context, thinking mode, tool calling.

Features

• Thinking mode — recognizes LongCat's reasoning_content field and translates it into harness ReasoningBlocks
• Tool calling — full function-calling support, with arguments kept a raw JSON string end to end
• Multi-turn — replays reasoning_content on tool-call turns, as thinking-mode passback requires
• Streaming — SSE with the usage-before-finish ordering the harness relies on
• Credential seam — the key resolves per request from ctx.credentials or the environment; no secret in any config file

Supported models

Model	Context	Max output	Notes
LongCat-2.0	1,048,576	131,072	text-only; thinking + tool calling

Facts from GET /openai/v1/models/LongCat-2.0, the only documented endpoint that reports supported_parameters. Tool calling is not mentioned on the chat-completions doc page and is only visible there.

Install

dsh plugin --profile default add github:ffyuuu/dsh-llm-longcat
export LONGCAT_API_KEY=...   # create one at https://longcat.chat/platform/api_keys

Installing a bundle lets the package's install scripts run on your machine, outside the sandbox the agent runs under. Pin a commit so a later push cannot change what executes:

dsh plugin --profile default add github:ffyuuu/dsh-llm-longcat#3dcb3b1b5870ba52baab053453bdbb28826e5f13

Then pick LongCat-2.0 in the model selector. The key may also be stored through the Web UI's Models page instead of the environment.

If dsh itself will not install

At the time of writing, installing the harness can fail before any plugin is reached, with either ETARGET … dsh-typert-protocol@^0.1.0-rc.8 or an npm heap exhaustion. That is an upstream packaging state, not this plugin: @deepseek-ai/dsh published 0.1.0-rc.8 while several packages it depends on stopped at 0.1.0-rc.7, and because the manifests use caret ranges, ^0.1.0-rc.7 still resolves up into the missing rc.8. npm then backtracks over an unsatisfiable graph until it runs out of memory.

Pinning every @deepseek-ai/* package to an exact 0.1.0-rc.7 through npm overrides avoids the drift. Nothing in this plugin needs changing either way — it declares >=0.1.0-rc.7 and works against whichever of those the host ends up with.

Config

- id: llm-longcat
  name: dsh-llm-longcat
  config:
    apiKeyEnv: LONGCAT_API_KEY   # default; resolved per request, never a literal key
    baseURL: https://api.longcat.chat/openai/v1  # optional; $LONGCAT_BASE_URL then the public API
    thinking: enabled            # optional deployment policy; `disabled` locks every request to off
    reasoningEffort: high        # optional; off | high — LongCat's switch is binary
    maxTokens: 131072            # optional per-request output cap
    defaultContextWindow: 1048576
    streamIdleTimeoutMs: 300000  # optional; five-minute default
    retryPolicy:                 # optional; omission uses bounded normal defaults
      mode: normal
      maxRetries: 3
    models:
      - id: LongCat-2.0
        contextWindow: 1048576

…