Skip to content

feat(evals): mcp-evals agent evaluation harness#9

Closed
yellowcap wants to merge 9 commits into
mainfrom
evals
Closed

feat(evals): mcp-evals agent evaluation harness#9
yellowcap wants to merge 9 commits into
mainfrom
evals

Conversation

@yellowcap

Copy link
Copy Markdown
Member

Splits the evals harness out of the evals-framework branch into its own PR.

What

Adds the mcp-evals package — an evaluation harness that runs the in-process mcp-agent against eval cases and scores it:

  • Loads cases from a shared Google Sheet (zero-auth CSV export) or a local CSV.
  • Scores tool use, dataset surfacing, and answer quality (Mistral judge); writes summary/detailed CSVs and a pass-rate table.
  • mcp-evals run CLI, example cases, and tests.
  • .github/workflows/evals.yml: manual workflow_dispatch only — evals are slow, nondeterministic, and billed, so deliberately not in PR/push CI. It serves the chosen toolset locally and uploads result CSVs as an artifact.
  • README "Evaluations" section + design docs.

Notes

  • Root pyproject.toml gains the mcp-evals workspace member; uv.lock regenerated.
  • Dropped an accidentally-committed LibreOffice lock file (.~lock.gold.csv#).
  • Companion PR: CDS EQC tooling (branch cds-eqc-tooling).

ciaransweet and others added 9 commits June 11, 2026 21:42
* fix: address missing/incorrect steps in readme for initial deploy (and applied them)

* fix: unblock workflows for initial deploy
Add the mcp-evals package: loads eval cases from a Google Sheet or CSV,
runs the in-process mcp-agent against each query, and scores tool use,
dataset surfacing, and answer quality (Mistral judge). Includes the CLI,
example cases, tests, a manual workflow_dispatch CI workflow, and docs.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants