| name | evaluator |
| description | Evaluate-phase skill for pi-forge. Use during the Evaluate phase to make a holistic, freeform judgment of the work against the brief, write artifacts/scorecard.md with an approve|iterate recommendation plus justification and risks, and record the verdict via forge_metric to inform the G2 gate. |
Evaluator (Evaluate phase)
Make a holistic judgment of the work. The output recommendation informs the human's G2 decision. Drive transitions only through forge_* tools.
Steps
- Read
artifacts/brief.json (acceptance + eval criteria), artifacts/test-report.md, metrics.json, and the implemented code/diff.
- Judge holistically across: correctness; alignment to the brief; code quality; security; maintainability; UX (if applicable); estimated cost. Use the quantitative inputs (test pass rate, coverage) as evidence, but the verdict is qualitative.
- Write
artifacts/scorecard.md with: a clear recommendation — approve or iterate; a justification; and the top risks/gaps. If iterate, list the specific things to fix.
- Record it:
forge_metric eval {"recommendation": "approve|iterate", "summary": "<one line>"} and forge_artifact scorecard artifacts/scorecard.md.
- Return control to the orchestrator, which opens G2 (forge_gate, select). The human's verdict — guided by your recommendation — routes to deliver (approve) or improve (iterate). Do not advance yourself.
CXAS path (when project.type = cxas)
Incorporate the cxas lint findings (rule severities) and the cxas run eval pass/fail (from the test report / metrics) into the holistic judgment and the approve|iterate recommendation.
Rules
- Be specific and honest: tie every judgment to evidence in the brief, the report, or the code.
- The recommendation is advisory; the human's G2 verdict is authoritative.
