Skill Tester

@rules/test-matrix.md @rules/scenario-design.md @rules/repair-workflow.md @rules/skill-maker-handoff.md @rules/evidence-reporting.md @references/prompt-pack-template.md

Skill Tester

Establish a reproducible baseline, repair only authorized target-owned defects, and prove the current behavior.

Default all user-facing reports, reusable test packs, validation notes, and handoffs to Korean. Preserve file paths, commands, schema keys, API names, and machine-readable fields in their required form. Use English only when the user requests it or an existing target artifact requires it.

• Prove that an existing skill triggers for its intended requests, stays inactive for neighboring work, and follows a safe, complete workflow.
• Test its core contract, direct resources, deterministic helpers, runtime degradation, bilingual behavior, and safety boundaries with observable scenarios.
• When the user asks to fix, strengthen, add, edit, or delete target-owned content, make the smallest evidence-backed repair and rerun the affected baseline.

Use skill-tester for evidence-led testing of an existing skill or skill folder, including a bounded test-and-repair pass.

Use skill-maker to create a new reusable skill or perform a broad structural refactor. Use autoresearch-skill for repeated score-driven optimization. Use an application QA workflow for product behavior rather than a skill package.

Do not use this skill when no target skill can be inferred after local inspection, the user requests only a document review, or requested deletion reaches outside the target skill's proven ownership.

Positive requests:

• "Test skills/git-maker/ for trigger precision and workflow regressions before release."
• "이 스킬이 제대로 켜지고 안전하게 동작하는지 엣지 케이스까지 검증해줘."
• "Validate this skill, fix its broken support link, and rerun the same checks."

Negative requests:

• "Create a Codex skill for reviewing SQL migrations." Route to skill-maker.
• "내 웹앱 결제 플로우를 실제 브라우저에서 QA 해줘." Route to application QA.

Boundary requests:

• "Review this skill and fix any issues you find." Test first, then make only bounded target-owned repairs; hand broad restructuring to skill-maker.
• "Keep optimizing this skill until its benchmark improves." Test the baseline, then route the repeated measured loop to autoresearch-skill.

Minimum input is a target skill path or pasted skill content. Infer its intended job from discovery metadata and local context first. If neither target nor intent is safely inferable, ask one focused question and do not fabricate findings.

For a repair or deletion pass, require an explicit user request to fix, strengthen, add, edit, prune, or delete. Treat only the named target skill and its proven owned resources as writable.

Load support files only for their stated purpose:

• Read rules/test-matrix.md to select smoke, targeted, standard, or thorough coverage and the smallest fast gate.
• Read rules/scenario-design.md to create executable positive, negative, boundary, edge, adversarial, workflow, and regression scenarios.
• Read rules/repair-workflow.md before any target edit, addition, or deletion.
• Read rules/skill-maker-handoff.md when findings require a broad skill structure refactor; use its packet to hand work to skill-maker, then recheck the returned target with the unchanged cases.
• Read rules/evidence-reporting.md before declaring a verdict or handing work off.
• Use references/prompt-pack-template.md only when the user asks for a reusable test pack; use its Korean sibling by default for Korean artifacts.
• Run node skills/skill-tester/scripts/validate-skill-tester.js --root skills/skill-tester --evals skills/skill-tester/assets/evals/skill-tester-cases.jsonl --json when this package changes.
• Run node skills/skill-tester/scripts/validate-skill.mjs <target-skill> for a quick target check and node skills/skill-tester/scripts/validate-skills-corpus.mjs --root skills --only <skill-name> --json for repository-skill structure.

The core owns trigger, authority, repair boundary, loop, and stop logic. Rules own recurring decisions; the template is an output resource; assets/evals/skill-tester-cases.jsonl is the machine-readable regression fixture; scripts are deterministic, local-only validators.

Assessment selects no loop. A requested repair uses exactly one bounded cycle: baseline -> diagnose -> minimal repair -> rerun unchanged affected cases -> decision. Keep the repair only if all critical guards pass and it reduces the named defect without new regression. If a critical check still fails, a required capability is absent, or another repair would broaden the scope, stop and hand off or block. Never use self-grading, an altered baseline, or "keep improving" as acceptance evidence.

Report in Korean with this minimum shape:

## Skill Test Report

**Target**: `skills/example/`
**Risk / mode**: targeted / assess | repair
**Verdict**: ship | caveated ship | iterate | block

### Baseline and current results
| Case / check | Baseline | Current | Evidence | Result |

### Findings and repairs
- **[severity] [taxonomy] Title**
  - Evidence / impact / minimal repair or handoff.

### Trace and safety
- Read-before-edit, side-effect boundary, fallback, and post-repair rerun evidence.

### Remaining risk
- ...

Use trigger-miss, trigger-overreach, scope-conflict, workflow-gap, resource-drift, validation-gap, edge-case-gap, runtime-gap, or safety-gap consistently. A reusable prompt pack must use the linked template and include its scenario matrix, binary oracle, trace assertions, baseline/current results, and untested risks.

Before completion, confirm:

• Target, intended behavior, risk, repair authorization, and excluded paths are recorded.
• Target core and direct resources were read before findings or edits.
• Scenarios cover the needed positive, negative, boundary, edge, regression, and—at standard or thorough risk—adversarial or workflow behavior.
• Expected behavior is observable and has an oracle; tool or delegation work also has trace assertions.
• Static checks were run when an executable capability and target path exist; unavailable required checks are disclosed as caveats or blockers.
• Any repair is target-owned, minimal, linked to a baseline finding, and rechecked with unchanged affected cases.
• Deletions meet the safe-deletion gate in rules/repair-workflow.md and every remaining local reference resolves.
• Korean and English behavior are compared for localized targets; file-pair existence alone is insufficient.
• Report maps Claim -> Risk -> Evidence -> Verification -> Result -> Caveat and names a final decision.