Expand description
What a harness’s configuration adds to one task’s runner invocation.
A configured harness carries three settings beyond its model: skill
bundles, MCP servers, and tools withheld from the model. Each reaches the
underlying CLI through whatever that CLI actually takes, and each runner
family is listed in [mechanism] with exactly what that is:
| Base | skills | mcpServers | disabledTools |
|---|---|---|---|
claude-code | --plugin-dir | --mcp-config | --disallowed-tools |
copilot | --plugin-dir | --additional-mcp-config | --excluded-tools |
copilot_headless | — | ACP session/new | — |
| every other base | — | — | — |
A setting a base has no way to apply refuses the task with
unsupported_harness_setting, naming the setting and the base. Running the
task without it would hand the client a result produced under a
configuration it did not ask for — a withheld tool offered anyway, a skill
never loaded — and nothing in the response could tell it so.
Only what is switched on counts: a skill or MCP server with enabled: false
is neither applied nor refused.
Structs§
- Applied
- Everything a harness’s configuration adds to one task’s run.
- Staged
Skills - Skill bundles written out for one task, removed from disk when dropped.
Enums§
- Setting
- One of the three harness settings a task run applies.
Constants§
- PLUGIN_
NAME - The name the staged skill bundles are loaded under.
Functions§
- apply
- Apply
harness’s settings to a run ofprovider, the tasktask_id.