dsh-llm-volcengine

Volcengine Ark Agent Plan and Coding Plan providers for DeepSeek Harness (DSH), with verified thinking-effort compatibility.

A DSH profile bundle that registers two LLM provider routes through a self-contained pi-ai-backed LlmAdapter, so model catalogs and compatibility switches reach pi-ai without depending on the dsh-llm-pi-ai settings compat schema. Thinking levels low / medium / high / xhigh / max are exposed where the endpoint honors them.

Why

A hand-declared Volcengine Ark route in settings.yaml runs into several gotchas that this bundle resolves once and for all:

Install

From the root of a DSH profile (e.g. ~/.dsh/profiles/web):

dsh plugin --profile web add dsh-llm-volcengine

Then restart DSH (or reload the profile) so the new bundle layer is composed. The two providers appear in the model selectors as:

• volcengine-plan/<model>
• volcengine-coding/<model>

Credentials

The bundle resolves an API key by trying each candidate credential reference (left to right) through the DSH credential service, then Ambient environment:

Store a key through the web Models page (it writes the DSH credential store) or export the env var. If you already have ARK_CODE_API_KEY / HUOSHAN_API_KEY configured, the bundles pick them up as fallbacks.

Override the first tried reference per route through the bundle config in cordis.patch.yml:

- id: llm-volcengine
  name: dsh-llm-volcengine
  config:
    agentPlanApiKeyEnv: ARK_AGENT_PLAN_API_KEY
    codingPlanApiKeyEnv: ARK_CODING_PLAN_API_KEY
    defaultReasoning: high

defaultReasoning is one of off | minimal | low | medium | high | xhigh | max (default: high).

Provider routes

volcengine-plan — Volcengine Ark Agent Plan

• Endpoint: https://ark.cn-beijing.volces.com/api/plan/v3
• Protocol: openai-responses for most models; openai-completions for Kimi K2.6 / K2.7 Code (mixed-api provider)
• reasoning.effort maps to the selected thinking level

volcengine-coding — Volcengine Ark Coding Plan

• Endpoint: https://ark.cn-beijing.volces.com/api/coding/v3
• Protocol: openai-completions with supportsDeveloperRole: false, supportsStore: false, supportsStrictMode: false, maxTokensField: "max_tokens"
• DeepSeek/GLM accept reasoning_effort; Kimi uses the qwen enable_thinking toggle; MiniMax/Doubao-seed-code expose no thinking control (reasoning is auto-captured)

How it works

The bundle inserts a single plugin row (id: llm-volcengine). On apply it:

1. Builds two pi-ai Provider objects directly with createProvider, passing the full compat block on each Model. Because the models are constructed in code (not through dsh-llm-pi-ai settings resolution), the compat fields the Coding Plan gateway requires reach pi-ai regardless of the installed dsh-llm-pi-ai compat schema.
2. Wraps them in the exported PiAiAdapter from @deepseek-ai/dsh-llm-pi-ai, which already implements the harness LlmAdapter contract (stream/resolveModel/listModels) and thinking-level clamping against each model's thinkingLevelMap.
3. Registers the adapter for both routes with ctx.llm.registerAdapter, so the providers join the model selectors and request routing like any built-in route.

Agent Plan models share a mixed-api provider (one createProvider with an api map) so Kimi K2.6/K2.7 Code dispatch to openai-completions while the rest use openai-responses.

Acknowledgements

Model catalogs, max output tokens, and compatibility switches are sourced from the community-verified Volcengine Ark provider extensions for pi:

• pi-provider-volcengine-agent-plan — the Agent Plan Responses path, tier gating, and Kimi routing decisions.
• pi-provider-volcengine-codingplan — the Coding Plan compat switches and model caps.
• pi-provider-volcengine-ark — per-model thinking formats for the coding endpoint.

This bundle adapts those compat findings to the DeepSeek Harness LLM seam.

License

MIT