sandbase-harness/docs/deepseek-v4.md

DeepSeek V4

SandBase Harness can use DeepSeek V4 through DeepSeek's OpenAI-compatible API. DeepSeek V4 supports a context window of up to 1 million tokens. The runtime compacts long sessions automatically; it does not currently expose a separate context-window setting.

Configure the provider

Set your API key in the environment:

export DEEPSEEK_API_KEY="<your DeepSeek API key>"

In Dashboard Settings > Models, switch to JSON and configure:

{
  "vendor": "openai_compatible",
  "base_url": "https://api.deepseek.com/v1",
  "api_key": "${DEEPSEEK_API_KEY}",
  "options": {
    "reasoning_effort": "max"
  }
}

Save and activate the settings. reasoning_effort is forwarded as reasoning_effort in each OpenAI-compatible chat-completions request. Use max with DeepSeek V4 Pro for the strongest coding and agent performance.

Create and run an agent

Create an agent whose model is deepseek-v4-pro (or deepseek-v4-flash for lower latency), then start a session from the Console. For a YAML seed definition:

name: deepseek-coder
model: deepseek-v4-pro
system: You are a careful coding agent. Inspect the repository, make focused changes, and verify them.
tools:
  - type: agent_toolset_20260401

From a tagged source checkout, build the runtime and start it with node dist/index.js start, then open http://127.0.0.1:3000/dashboard, select the agent, and send the first task.

Do not commit API keys to the workspace. The ${DEEPSEEK_API_KEY} reference is resolved only by the runtime.