Metadata-Version: 2.5
Name: inspect_steward
Version: 0.2.5
Summary: An agent that supervises evaluations, monitoring progress, diagnosing problems, and adjusting runtime state on your behalf.
Project-URL: Source Code, https://github.com/meridianlabs-ai/inspect_steward
Project-URL: Issue Tracker, https://github.com/meridianlabs-ai/inspect_steward/issues
Author: Meridian Labs
License: MIT License
License-File: LICENSE
Classifier: Operating System :: MacOS
Classifier: Operating System :: POSIX :: Linux
Requires-Python: >=3.12
Requires-Dist: apprise>=1.9
Requires-Dist: click>=8.0
Requires-Dist: httpx>=0.27
Requires-Dist: inspect-ai>=0.3.263
Requires-Dist: inspect-scout>=0.5.1
Requires-Dist: platformdirs>=4.0
Requires-Dist: psutil>=5.9
Requires-Dist: python-dotenv>=1.0
Requires-Dist: pyyaml>=6.0
Provides-Extra: flow
Requires-Dist: inspect-flow>=0.12.0; extra == 'flow'
Provides-Extra: hawk
Requires-Dist: hawk[cli,runner]>=2.4.0; (python_version >= '3.13') and extra == 'hawk'
Requires-Dist: uv>=0.5; (python_version >= '3.13') and extra == 'hawk'
Description-Content-Type: text/markdown

# Inspect Steward

Welcome to Inspect Steward, an agent that supervises evaluations on your behalf.

Steward plays the role of the human sitting in front of a running eval. It:

-   Watches progress as a run unfolds
-   Diagnoses problems as they surface
-   Adjusts runtime state in response

A steward acts on behalf of someone who isn't there. That's the relationship this package models: you start a run and step away, and the steward minds it.

Concretely, Steward runs each task of an eval set in its own process — so one crash costs one task rather than the run, and CPU-bound work actually runs in parallel — and reconciles the log directory against what your definition asks for on a schedule. A timer guarantees that happens with nobody watching; what needs judgement waits for someone who can exercise it.

```bash
pip install inspect-steward

steward init my-sweep     # create a workspace
steward tasks evalset.py  # see what a definition resolves to, before running it
steward launch            # start it, arm the timer, and return
steward status            # ask how it's going, whenever you like
```

Steward runs on macOS and Linux. Windows is not supported: a run has to outlive the process that started it, and the detached-process, control-socket, and process-table mechanisms that achieve it are POSIX-only.

Learn more about using Steward at <https://meridianlabs-ai.github.io/inspect_steward>.
