Skip to main content
The stirrup Harness runs Stirrup, a Python agent library, inside the task Environment. It supplies the prepared task’s prompts, a code execution tool, and an image viewing tool, then collects the answer and trajectory.

Configure Stirrup

Select a package release with version. The default installation strategy installs that release when it is missing or the installed version differs. preinstalled requires the selected release to already be available to python3; install_always runs installation for every session. The Environment needs Python 3.12+, pip, and access to the package registry when installing. Shared parameters and their native equivalents control the same setting. For example, tool_timeout_seconds and shell_timeout set the code execution deadline. kwargs supplies native request defaults; explicit model.params values take precedence.

Parameters

Inspect the same parameter schema with agentcompass config docs harness stirrup.
ParameterNative parameterScopeTypeDefaultDescription
-SharedAppend the effective non-public rollout network policy to the harness input.
-SharedStirrup package version to install.
-SharedHow to prepare the stirrup package in the env.
SharedMaximum agent turns passed to stirrup.Agent.
SharedMaximum output tokens per model response.
SharedModel context window; stirrup summarizes history when usage nears it.
SharedPer-command wall-clock timeout in seconds for the code_exec tool.
SharedTimeout in seconds for one model request.
SharedTotal model attempts including the first; native max_retries counts additional retries.
-Harness-specificExpose Stirrup’s view_image tool. Set to true when the Benchmark’s modality includes image, where an explicit false is an error. Disable for text-only models, which reject image input.
Harness-specificAdditional model request parameters; explicit model.params values take precedence.
Harness-specificFraction of the context window used before summarizing conversation history.
Harness-specificRemaining turns at which the agent starts receiving a limit warning.
Harness-specificRun synchronous tool executors in a worker thread.
Harness-specificMove images from tool results into separate user messages.
Harness-specificAsk the agent to continue when an assistant response contains no tool calls.
Harness-specificSummarize and retry after a rejected context overflow.