Skip to content

Fix recurring nightly save_memory eval output regression #2651

Description

@acoliver

Summary

The Evals Nightly workflow continues to fail all save_memory attempts after the reporting fix from issue 2605.

Example run: https://github.com/vybestack/llxprt-code/actions/runs/29981538815

The model successfully calls save_memory with the exact expected fact, but the eval compares the exact expected assistant answer against TestRig's complete human-readable process result. Successful plain-text runs can append stderr diagnostics to stdout, so a correct answer is reported as an output hundreds of characters long and fails the exact comparison.

Expected behavior

  • The eval validates the assistant response independently from process diagnostics.
  • A successful save_memory call followed by the exact answer $blue$ passes.
  • Raw stdout, stderr, exit status, timeout state, and tool telemetry remain available in eval artifacts.
  • JSON output remains intact when Podman telemetry is present.

Acceptance criteria

  • Evals use the CLI's structured JSON output contract.
  • The response field is schema-validated before assertions run.
  • TestRig captures raw stdout and stderr separately without changing existing plain-text callers.
  • Podman telemetry filtering preserves legitimate CLI JSON and user output.
  • Deterministic FakeProvider integration coverage exercises save_memory and the exact answer.
  • Process failures and timeouts preserve diagnostic captures.

Related to issue 2605.

Metadata

Metadata

Assignees

Labels

bugci/cdIssues with github and the workflow scripts and CI CD environment.

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions