Skip to content

ci: run daily Luna agent evals - #331

Merged
jlitola merged 12 commits into
mainfrom
jlitola/agent-eval-ci-pipeline
Aug 31, 2026
Merged

ci: run daily Luna agent evals#331
jlitola merged 12 commits into
mainfrom
jlitola/agent-eval-ci-pipeline

Conversation

@jlitola

@jlitola jlitola commented Aug 31, 2026

Copy link
Copy Markdown
Contributor

Summary

  • add bounded workload concurrency and schema-v3 execution metadata
  • add a concise CI report with tool-call, token, cost, duration, and failure evidence
  • run Luna discovery and intent scenarios daily, manually, or from an exact same-repository PR label event
  • securely forward GITHITS_API_TOKEN to Codex MCP by variable name without persisting its value
  • keep Braintrust persistence and baseline comparison out of this increment

Validation

  • 3,634 repository tests pass
  • 124 focused harness tests and 60 focused suite/reporter tests pass
  • typecheck, format, lint, build, actionlint, and diff checks pass
  • Luna preflight complete
  • Opus review clean after findings were fixed

Live validation

Scheduled/manual default-branch validation remains pending merge because the workflow does not exist on main yet.

Run workload executions through a bounded in-process pool, persist the execution bound in run and suite artifacts, and preserve legacy suite readers and comparisons.
Render schema-validated suite telemetry as a compact absolute report and preserve failure diagnostics before returning a failing status.
Run bounded Luna discovery and intent suites in clean GitHub Actions jobs, retain raw evidence, and render an absolute CI summary.
Document bounded workload pools, manifest ordering, and workflow-run evidence links after the CI concurrency implementation.
Write a non-secret scenario marker before checkout so artifact uploads remain non-empty when setup or authentication fails.
Prevent child execution for suites with no selected workloads, fail zero-execution CI evidence, and leave enough timeout margin for the six intent waves.
Keep operational documentation consistent with the 40-minute scenario job timeout while preserving the separate five-minute workload timeout.
Align Phase 3 failure enumerations and suite preflight documentation with zero-workload validation.
@jlitola jlitola added the agent-eval Authorize paid agent evals for the current same-repository PR head SHA label Aug 31, 2026
Pass the GitHits token through Codex MCP env_vars by name so stdio MCP children inherit authentication without persisting secret values.
Keep Codex MCP env_vars in the server table before its literal environment subtable so generated TOML remains valid.
@jlitola jlitola added agent-eval Authorize paid agent evals for the current same-repository PR head SHA and removed agent-eval Authorize paid agent evals for the current same-repository PR head SHA labels Aug 31, 2026
Document the failed and corrected label-triggered runs, measured CI timings and costs, and the remaining post-merge scheduled/manual validation.
Keep live label evidence distinct from the trusted manual and scheduled default-branch checks that remain pending after merge.
@jlitola jlitola added agent-eval Authorize paid agent evals for the current same-repository PR head SHA and removed agent-eval Authorize paid agent evals for the current same-repository PR head SHA labels Aug 31, 2026
@jlitola
jlitola marked this pull request as ready for review August 31, 2026 10:34
@jlitola
jlitola merged commit e1599b7 into main Aug 31, 2026
13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

agent-eval Authorize paid agent evals for the current same-repository PR head SHA

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant