Menu
Evaluate Codex workflow outputs with repository scorecards and hard gates.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Improve, audit, simplify, rewrite, or migrate prompts, tool descriptions, agent instructions, and prompt stacks for GPT-5.6 Sol or the GPT-5.6 family. Use for outcome-first prompt design, autonomy boundaries, tool routing, PTC, grounding, verbosity, reasoning effort, and prompt evals.
Apply a deep, design-minded engineering workflow for complex work that needs careful planning, iteration, and simplification.
Use when preparing or running a Claude CLI agentic engineering review through claude -p for repository audits, architecture analysis, system design, security review, large refactors, specs, RFCs, or evidence-grounded long-form engineering analysis.
Use when preparing or running a ZCode GLM-5.2 agentic builder workflow through zcode --prompt for fast implementation, iterative code generation, focused fixes, and validation on an existing repository.
Apply adversarial opposite-analysis to plans, specs, architecture, code changes, and claims. Use when the user asks for adversarial review, opposing analysis, contrarian review, red-team reasoning, or Z.ai and MiniMax cross-checks through the Ralph MCP router.
Review and adjudicate Bugbot, Cursor, Seer, and similar automated PR feedback with local evidence before accepting, fixing, or dismissing findings.
| name | evaluate |
| description | Evaluate Codex workflow outputs with repository scorecards and hard gates. |
Use this skill when a change needs measured evidence. The result must point to a fixture, command, or run artifact, and it must stay separate from the implementation being judged.
Use scripts/evals/run_scorecard.py for single metric files. Use a suite-specific script under scripts/evals when a workflow needs train/holdout splits, mutation checks, or JSONL logging.
Scorecards and fixtures are read-only inputs while the run is active. Generated reports belong under .ralph-codex/reports/evals, which keeps source files clean.
Hard gates are blocking. The required gates are tests pass, no secret leak, eval files unchanged, no scope violation, and no eval gaming. A high weighted score cannot override a failed gate.
When train and holdout data exist, use train for iteration and holdout for the decision. Do not tune the scorecard or fixture after seeing holdout results.
Return the scorecard id, version, score, gate status, decision, report path, and residual risk. If a tool was skipped or unavailable, state that directly.