Introduction

When using Volcengine Ark’s Agent Plan and Coding Plan in DSH, you need to first align provider routing, protocols, compatibility switches, credential order, and model limits. dsh-llm-volcengine bundles these into a DSH plugin; after installation, the two providers will directly enter the model selector.

What is this

dsh-llm-volcengine is maintained by Badakonpro and is licensed under MIT. It registers two LLM provider routes for DeepSeek Harness (DSH): volcengine-plan and volcengine-coding, with verified thinking intensity compatibility. The plugin uses a self-contained pi-ai-backed LlmAdapter, allowing the model directory and compatibility switches to interact with pi-ai without relying on the dsh-llm-pi-ai settings compat schema.

Core Features

  • Registers volcengine-plan and volcengine-coding LLM provider routes.
  • The Agent Plan route uses openai-responses; specifically, Kimi K2.6 / K2.7 Code uses openai-completions.
  • The Coding Plan route uses openai-completions, setting supportsDeveloperRole: false, supportsStore: false, supportsStrictMode: false, and maxTokensField: "max_tokens".
  • Exposes thinking levels when supported by the endpoint: low / medium / high / xhigh / max.
  • Supports parsing API keys via the DSH credential service or ambient environment.
  • Supports configuring agentPlanApiKeyEnv, codingPlanApiKeyEnv, and defaultReasoning via cordis.patch.yml.

Installation and Usage

Execute the following in the DSH profile root directory (e.g., ~/.dsh/profiles/web):

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

After installation, restart DSH or reload the profile to activate the new bundle layer.

After activation, two providers will appear in the model selector:

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

Typical Usage

Credentials

The plugin parses API keys via the DSH credential service or ambient environment. The default priority order is:

  • volcengine-plan: ARK_AGENT_PLAN_API_KEY, VOLCENGINE_ARK_PLAN_API_KEY, ARK_CODE_API_KEY
  • volcengine-coding: ARK_CODING_PLAN_API_KEY, VOLCENGINE_CODING_API_KEY, HUOSHAN_API_KEY

You can save API keys via the web Models page (writing to the DSH credential store) or by exporting environment variables.

cordis.patch.yml

The following fields can be configured in cordis.patch.yml:

agentPlanApiKeyEnv: ARK_AGENT_PLAN_API_KEY
codingPlanApiKeyEnv: ARK_CODING_PLAN_API_KEY
defaultReasoning: high

defaultReasoning can be set to off | minimal | low | medium | high | xhigh | max, with a default value of high.

Applicable Scenarios and Notes

Suitable for scenarios where you need to integrate Volcengine Ark Agent Plan / Coding Plan into DSH, unifying provider routing, credential parsing, and thinking levels.

Please note:

  • The plugin runs with the current DSH process permissions; you should check the source code and MIT license before installing.
  • package.json declares peerDependencies: @deepseek-ai/dsh-credentials, @deepseek-ai/dsh-launch-environment, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-llm-pi-ai, @earendil-works/pi-ai.
  • The Coding Plan gateway rejects the developer role and store, and requires the max_tokens field; the plugin handles these request items via the Coding Plan route’s compatibility switches.
  • Some Coding Plan models do not have thinking control: minimax-m2.7, minimax-m3, doubao-seed-code.
  • Different models have their own maxTokens limits. The README lists verified limits per model; when selecting a model, refer to the limit specified in the README for that model.

Links

The value of dsh-llm-volcengine lies in bundling the protocols, compatibility switches, credential order, and thinking levels of Volcengine Ark’s two plan routes into a DSH plugin.

GitHub: https://github.com/Badakonpro/dsh-llm-volcengine

Directory Page Link: https://www.skillhub.cn/plugins/Badakonpro/dsh-llm-volcengine (from plugin link, not verified in current fetched materials). The community directory is an independent site; do not interpret it as the official DeepSeek / Fantuan app store.