Build a fixed baseline

Use one repository commit, one prompt, one model/provider, one permission profile, a cleared or recorded cache state, and a fixed success criterion. Record wall time, time to first visible output, model tokens if exposed, tool calls, retries, changed files, and final test result.

Locate the expensive layer

Provider latency appears before or during model responses. Context overhead grows with repository scanning, long histories, loaded Skills, and tool descriptions. Tool latency appears between model turns. Retries and unstable subagents create repeated work. Cache behavior can make the second run incomparable to the first.

  • Provider/model: response latency and reasoning behavior.
  • Context: repository, conversation, Skills, MCP tool schemas.
  • Tools: shell, browser, MCP, file search, and external APIs.
  • Orchestration: subagents, retries, scheduler, and loop strategy.
  • Cache: prefix stability and warm-versus-cold state.

Apply the remedy to the measured cause

Reduce context only when context is the source. Disable unused MCP servers or Skills only when their tool/schema load is material. Choose a faster model only after provider and orchestration overhead are separated. Split tasks when a broad prompt forces repeated exploration. Preserve a stable prompt prefix if the provider cache depends on it.

Comparisons need the same provider boundary

When comparing DSH with OpenCode or another harness, use the same provider and model where possible. Otherwise, a faster result may describe the model route rather than the harness. Record cache state and repeated runs.

Controlled results are not published yet

HarnessBase will publish the table only after completing fixed-task runs across relevant modes and retaining raw timing, token, tool, and failure observations. Until then, community reports remain observations rather than measurements.

Sources and evidence boundary

Primary sources establish product facts. HarnessBase labels direct testing separately and does not convert community observations into verified results.

Target query cluster: DeepSeek Harness slow · DeepSeek Harness too many tokens · DeepSeek Harness token usage