tree: 66bfa23ac1c0e6a223acdf9cf703abf00ccc1222
  1. step-reporting/
  2. step-scope-selection/
  3. README.md
tools/skill-evals/evals/ci-runner-audit/README.md

ci-runner-audit evals

Behavioral evals for the ci-runner-audit skill.

Suites (6 cases total)

SuiteStepCasesWhat it covers
step-scope-selectionScope selection and command choice4explicit repo, ambiguous Apache project, full-org scan, prompt injection ignored
step-reportingReporting discipline2high-confidence vs broad candidates, CI-risk language instead of security overclaiming

Run

# All cases
PYTHONPATH=tools/skill-evals/src python3 -m skill_evals.runner \
    tools/skill-evals/evals/ci-runner-audit/

# Single suite
PYTHONPATH=tools/skill-evals/src python3 -m skill_evals.runner \
    tools/skill-evals/evals/ci-runner-audit/step-scope-selection/fixtures/

# Single case
PYTHONPATH=tools/skill-evals/src python3 -m skill_evals.runner \
    tools/skill-evals/evals/ci-runner-audit/step-scope-selection/fixtures/case-4-injection-ignored

What the suites cover

step-scope-selection

Given a maintainer request, the model determines whether the scan scope is explicit enough to run immediately or whether it must ask a scope question first. The suite also checks that a prompt-injection attempt in user-supplied text is flagged and ignored.

step-reporting

Given mock TSV output, the model determines how to report findings. The suite asserts that setup-action mismatches are high-confidence, broad macOS candidates are marked false-positive-prone, and runner findings are described as CI breakage / portability risks rather than security vulnerabilities.