snow-The / dsh-ark-plan

dsh-ark-plan

Correctly activate DeepSeek v4 flash on the Volcano Ark plan API (https://ark.cn-beijing.volces.com/api/plan/v3) for DeepSeek Harness.

Why this exists: getting the ark route working is easy, but getting thinking effort = max to actually reach the model is not — it needs two things that are easy to miss, and missing either one silently degrades or breaks the setup.

What it does

1. Default config injection (cordis.patch.yml): a fresh profile gets a working ark route + reasoningEffort: max with zero manual YAML. Your own settings.yaml always wins over these defaults.
2. ark_plan_doctor — a self-check tool that reads your current config, flags the two known traps, and live-tests the endpoint to show the server-resolved model version and whether a thinking block comes back.

Install

# from this repo
dsh plugin --profile web add github:snow-The/dsh-ark-plan
# or local
dsh plugin --profile web add file:C:/path/to/dsh-ark-plan
# then restart dsh web

Set your API key once (in the GUI Models page, or):

# ~/.dsh/.credentials.yaml
ARK_API_KEY: <your volcano ark key>

That's it. The profile defaults to modlens-ark (vision wrapper from @liustack/modlens — install modlens if you want image input; without it, set agent-default-model.provider to ark).

Verify

Ask your agent to run ark_plan_doctor, or check the config yourself:

# expected effective values
llm-pi-ai.providers.ark.models[0].reasoningEfforts: { low: low, high: high, max: max }
agent-default-model.reasoningEffort: max

The doctor also live-probes the endpoint — expect serverModel = deepseek-v4-flash-ga-<date> (e.g. ...-ga-260731) and a returned reasoning block.

The traps this plugin encodes

Trap	Symptom if missed
reasoningEfforts must be declared on the ark model — otherwise pi-ai judges the model non-reasoning and its openai-responses adapter skips the whole reasoning branch. The request carries no effort, and the server silently uses its default (high).	effort=max never actually reaches the model
reasoningEffort: max must accompany reasoningEfforts — setting max without the model declaring it makes resolveReasoningLevel throw UNSUPPORTED_REASONING_EFFORT.	requests error out
contextWindow / maxTokens must be declared — the model's real spec is 1M context / 256K output, but pi-ai's defaults are contextWindow 262144 and maxTokens 32768. Without declaration DSH silently treats the model as a 256K/32K model. (Ark accepts max_output_tokens up to 393216; 262144 = the model's advertised 256K output.)	context shows 256K, output capped at 32K

Also encoded: the bare id deepseek-v4-flash is resolved by Ark to the latest GA snapshot (currently deepseek-v4-flash-ga-260731, i.e. the 0731 build).

Protocol note

The route speaks openai-responses (the same protocol family as the DeepSeek official API) pointed at Ark's plan URL (https://ark.cn-beijing.volces.com/api/plan/v3). Ark's plan API is OpenAI-Responses-compatible, so the DeepSeek-style provider settings work as-is with only the base URL swapped — verified live (thinking block returned with reasoning.effort=max, model resolves to deepseek-v4-flash-ga-260731).

How the defaults work

dsh-settings registers each entry's patch config as the base of its settings section, so plugin defaults only apply where the user has not configured the same keys. No overwriting of user settings.

License

MIT