DeepSeek Harness (dsh)

dsh workers run DeepSeek's @deepseek-ai/dsh CLI in headless JSON mode. The worker image already includes it. This page covers the current route and model selection behavior; see Harness Providers for adapter internals.

Configure the worker

HARNESS_PROVIDER=dsh
DEEPSEEK_API_KEY=...

DSH_BINARY optionally selects another trusted dsh executable. The worker accepts either DEEPSEEK_API_KEY or OPENROUTER_API_KEY at startup, but the selected model decides which key a run needs.

Pick the route

The prefix is the route selector. A DeepSeek key does not turn an openrouter/... model into a direct DeepSeek request.

Use the DeepSeek API for tiered tasks

Every built-in dsh tier currently defaults to openrouter/...; ultra even defaults to an Anthropic model. A worker with only DEEPSEEK_API_KEY therefore passes its credential check but a tiered task needs OpenRouter. Set a worker tier map to use DeepSeek directly:

MODEL_TIER_MAP='{"smol":"deepseek-v4-flash","regular":"deepseek-v4-flash","smart":"deepseek-v4-pro","ultra":"deepseek-v4-pro"}'

For an untiered task, set MODEL_OVERRIDE=deepseek-v4-flash (or another bare DeepSeek id). A task carrying modelTier resolves its tier before MODEL_OVERRIDE, so use MODEL_TIER_MAP for tiered work.

Runtime behavior

• Reasoning effort is passed to dsh when the selected model supports it.
• The adapter inserts the swarm MCP client for each task. Tool names appear as mcp__agent-swarm__<tool>.
• dsh currently has no live steering and no native session resume.
• dsh reports step token usage, not billed USD. The swarm prices the usage from its dsh pricing rows and derives context from each step's token counts.