EvalView - AI Agent Testing

Version updated for https://github.com/hidai25/eval-view to version v0.7.1.

  • This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

EvalView is a GitHub Action designed to detect and manage behavioral regressions in AI agents, such as unexpected changes in tool usage, decision-making, or output quality, that standard tests and health checks might miss. It automates the process of tracking output drift, classifying changes, and providing actionable insights to help developers address issues and maintain consistent agent behavior. With features like confidence-based drift detection, change classification, and auto-healing for certain failures, EvalView enables teams of any size to implement robust regression testing and ensure reliable AI performance.

What’s Changed

What’s new in 0.7.1

Added

  • TOML test cases — TestCaseLoader now understands .toml alongside .yaml/.yml (#209). Same schema, different on-disk encoding for teams that already author config in TOML. Uses stdlib tomllib on Python 3.11+ and falls back to tomli on 3.9/3.10. Example in examples/test_case_toml.toml.
  • CSV log import — evalview generate --from-log now accepts CSV alongside JSONL, OpenAI, and EvalView capture formats (#216). Header row identifies columns, with aliases matching the JSONL parser. Tool cells accept JSON list, comma-, semicolon-, or pipe-separated forms. Example fixture and README under examples/log-import/.

Changed

  • Chat-mode command validator now derives its allowlist from the Click registry instead of a hand-maintained set.
  • evalview inspect is now a hidden alias that forwards to evalview visualize (which is the strict superset).
  • evalview view is no longer hidden — surfaced under the “Inspect & Visualize” section of --help.
  • Daily dogfood workflow now updates a single rolling issue instead of filing a new one each failing day.
  • parse_csv’s warn callback type tightened from Any to Optional[Callable[[str], None]] (#221).

Removed

  • evalview quickstart deprecated shim (use evalview demo or evalview init).
  • requirements.txt (use pyproject.toml + uv.lock).

Internal

  • Four 1k+ files split into focused submodules following the PR #202 mixin pattern (#215).
  • Root-directory polish pass (#218): tighter .gitignore, install.sh → scripts/install.sh, 39 tracked root entries.

Docs

  • AGENT_INSTRUCTIONS.md → AGENTS.md, and guides/ moved under docs/ (#219).
  • docs/README.md index expanded; per-directory READMEs in demo/, demo-agent/, demo-tests/.

Fixed

  • Hermetic test suite — git clone && make test goes green on first run with no env setup beyond uv.
  • Lint-clean main — drops unused imports left after the quickstart removal.

PyPI: https://pypi.org/project/evalview/0.7.1/ npm: https://www.npmjs.com/package/evalview/v/0.7.1 Full changelog: https://github.com/hidai25/eval-view/blob/main/CHANGELOG.md

🤖 Co-authored with Claude Code