feat(evals): screen each rule with and without it; pruning workflow for the audit skill
Adds --screen to the end-of-reply runner: each scenario runs with the rules and with the one rule its without: pattern names removed, and the output says whether the rule helps, makes no difference, or fails the scenario in both arms. Adds expect: reply and --scenarios <dir>, seven screening scenarios, and hr-agent-guidance-audit/references/prune-workflow.md (size, migrate, condense, screen, vendor guides, opt-ins, land and adopt), linked from the skill. No new skill, so the catalog size is unchanged.
Checks: node --test test/*.test.mjs 136 pass; bun test 462 pass; oxfmt --check . clean. The screen ran 4 runs per arm on Opus 5.5 and Sonnet 5.5 against the condensed core; results are in the plan, not shipped.
Summary by cubic
Adds --screen to the end-of-reply runner: each scenario runs with the rules and again with the one rule its without: pattern names removed, printing whether that rule helps, makes no difference, or fails the scenario in both arms. Also adds expect: reply, which grades a reply by the scenario's declared patterns alone (skipping the offer-layout check), and --scenarios <dir> for pointing at another scenario directory. Seven screening scenarios ship in evals/end-of-reply/screen/, each using required or forbidden patterns to demand the behavior the rule asks for, and the hr-agent-guidance-audit skill gains references/prune-workflow.md, the measured way to shrink always-loaded guidance (size, migrate, condense, screen, vendor guides, opt-ins, land and adopt). No new skill, so the catalog size is unchanged.
Summary by CodeRabbit
- New Features
- Added screening evaluations that compare results with and without a selected guidance rule, report pass counts, and indicate whether each rule helps, hurts, or makes no difference.
- Added options to select a scenario directory and evaluate replies against declared checks without requiring offer-layout validation.
- Added seven screening scenarios covering response guidance and tool-use behaviors.
- Added a workflow for measuring, refining, and testing always-loaded guidance.
- Documentation
- Documented screening setup, result interpretation, and recommended repeated runs across models.