dsh-llm-longcat

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

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

A llm-longcat: section in $DSH_HOME/settings.yaml overrides any field
without a restart: base URL, catalog, request defaults, and idle budget all
take effect on the next request, while an in-flight stream keeps the facts it
started with.

Reasoning is binary, deliberately

LongCat controls thinking with thinking: {type: enabled|disabled} and does
not accept OpenAI's top-level reasoning_effort — its
supported_parameters lists the former and omits the latter. There is
therefore no low/medium/high gradient to map, and this adapter offers exactly
two levels rather than advertising controls that would collapse onto the same
two request bodies:

off serializes an explicit disabled rather than omitting the field —
omitting it would hand the decision to LongCat's server-side default, which is
not what selecting Off should mean. Requesting low, medium, or max fails
with UNSUPPORTED_REASONING_EFFORT before any network I/O.

Wire-format notes

• Tool-call deltas repeat id and name as explicit null. LongCat sends
  them on the opening delta and then null (not omitted) on every
  continuation, so a naive !== undefined guard blanks the assembled call's
  name. Verified on live traffic; pinned by a regression test.
• Streaming only, with stream_options.include_usage always on. Usage may
  arrive attached to the finish chunk or as a trailing usage-only chunk; both
  are deferred to [DONE] so usage always precedes finish.
• The first thinking-mode delta can be an empty string — it must not open a
  reasoning block.
• Reasoning passback: on assistant turns that carried tool calls,
  reasoning_content is serialized back into history; on tool-call-free turns
  it is dropped (ignored anyway — saves tokens).
• Assistant content is always a string, never null: the message is durable
  session history, and a null there would make later turns replay a body the
  endpoint can reject.
• Cache accounting: prompt_tokens_details.cached_tokens maps to
  cacheReadTokens and is subtracted out of inputTokens to keep the
  harness's disjoint-count convention.

Errors

Non-2xx responses throw LlmError with stable codes. LongCat documents a
dedicated 402 for exhausted token quota and puts insufficient_quota on
403, where most OpenAI-compatible providers use 429 — both are classified
as QUOTA before the auth and rate-limit buckets, so a depleted balance is
never reported as a bad key or retried as a transient rate limit.

A completed stream that opened no content blocks becomes a finish error with
EMPTY_RESPONSE, which the shipped retry policy treats as retryable.

Tests

npm run typecheck   # against the published @deepseek-ai/dsh-llm types
npm test            # 30 unit tests over serialize + translate
npm run build       # emits lib/ and lib/types/
npm run test:e2e    # real API, needs LONGCAT_API_KEY, spends a few hundred tokens

test:e2e drives the built adapter's own serialize → SSE → translate pipeline
against api.longcat.chat, so it verifies what the plugin actually sends
rather than a hand-written approximation. It is what caught the null-name
delta bug.

Limitations

• No image input. LongCat-2.0 reports modality: text->text, so image
  content is refused before sending, naming the model.
• No stop sequences. stop is absent from supported_parameters; passing
  one fails with UNSUPPORTED_OPTION rather than silently running past it.
• Reasoning is binary — no low/medium/high gradient exists to map.

License

MIT