Skill evaluation execution

Traceability

• Spec ID: skill-evaluation-execution
• Status: Implemented

Intent

Add a project-local execution protocol for evaluating the Better Harness Skill. The protocol should combine the repository's Agent Customize evidence boundary with Plugin Eval's deterministic JSON result, static budget analysis, and benchmark workflow without presenting static checks as demonstrated task lift.

Acceptance Scenarios

• AC-1: references/agent-customize/skill-eval.md defines the canonical command sequence for target resolution, Better Harness static checks, Plugin Eval analysis, benchmark preparation, and final evidence normalization.
• AC-2: Agent Customize routes existing-Skill evaluation requests to skill-eval.md, while skill-review.md remains the owner of Gates, D1-D8, and E0-E3 ceilings.
• AC-3: The execution protocol treats Plugin Eval JSON as its machine source of truth, labels static budgets as estimates, and prevents static results from exceeding E0.
• AC-4: The protocol matches the installed Plugin Eval script contract: starter benchmark scenarios are inspected before a real Codex run, and no unsupported dry-run command is prescribed.
• AC-5: The Better Harness Skill is evaluated with the documented static route; generated benchmark configuration is kept local-only, and the reported result distinguishes observed output from unrun dynamic evidence.
• AC-6: Focused Skill-routing and documentation-link tests pass, and the generated Better Harness documentation graph is current.

Non-goals

• Implement a second scoring engine or copy Plugin Eval into this repository.
• Put Skill evaluation fields into ordinary Agent Customize coverage rows.
• Claim E1-E3, task lift, repeatability, or production safety from a static run.
• Execute generic generated benchmark scenarios as if they were project acceptance cases.

Plan and Tasks

1. Add the execution protocol beside the existing Agent Customize Skill rubric.
2. Route Agent Customize and existing-Skill evaluation requests to the new protocol, preserving the separate rubric owner.
3. Add focused assertions for routing, evidence boundaries, and the current benchmark contract.
4. Regenerate the documentation graph and run the static evaluation against skills/better-harness.

Test and Review Evidence

• AC-1..AC-4: node --test test/better-harness-skill.test.mjs
• AC-2, AC-6: node scripts/doc-link-graph/cli.mjs skills/better-harness and node --test test/doc-link-graph.test.mjs
• AC-3, AC-5: run agent-lint with agent-assets-review, then run Plugin Eval start, analyze, explain-budget, and init-benchmark against skills/better-harness.
• Risk: Plugin Eval is an optional external CLI. Record the resolved version and entrypoint; if unavailable, keep the result partial instead of inventing it.
• Risk: benchmark performs real Codex executions. Do not run generated starter scenarios until workspace, prompts, verifiers, permissions, and stop conditions have been reviewed.

Implemented evidence:

• agent-lint reported one entrypoint, five resolved references, and zero findings for skills/better-harness.
• Plugin Eval produced a static 58/D/high result with estimated budgets of 70 trigger, 2,994 invoke, and 12,391 deferred tokens. Its broken-link check cited