fix(skill_eval): conflict guard + descriptionOverride normalization (reworks #28)

[antongulin]

Reworks #28 (credit @mxl for the original report and implementation — co-authored on the commit).

What changed

• Installed-skill conflict guard: skill_eval and skill_optimize_loop now fail fast with a clear error when a skill with the same base name is already visible to opencode (installed skill or skills.paths entry), since it can steal triggers from the synthetic eval skill and produce false negatives.
• Unlike fix(skill_eval): add conflict guard and normalize descriptionOverride #28, the guard runs opencode debug skill through the shared runProcess helper — 10s timeout (no hang if the CLI stalls) and no Bun globals in compiled output (package.test.mjs enforces this; direct Bun.spawn was why fix(skill_eval): add conflict guard and normalize descriptionOverride #28's package tests failed after rebuild).
• Parsing is extracted into a pure findSkillConflicts() with unit tests (conflict found, missing location, no match, invalid JSON, non-array).
• descriptionOverride normalization: empty/whitespace-only strings are treated as omitted in both tools.
• triggerOnly gating fix (follow-up to fix(skill_eval): add trigger-only mode #29): the early success return now respects triggerOnly: false — a triggered run that ends in a failed process throws again, matching the flag's documented behavior.
• Rebuilt dist.

Verification

• npm run build clean
• npm test: 24/24 pass
• npm run test:ts: 49/49 pass
• grep -c '\bBun\b' dist/skill-creator.js → 0

Closes #28.

🤖 Generated with Claude Code

[github-actions]

🏹 Robin

✅ Review done. I flagged 4 things worth a look.

Push fixes whenever you like, then comment /robin for another pass.

[github-actions]

🏹 Robin

Heads up: this is a point-in-time review. Push fixes freely, then comment /robin whenever you want another pass.

⚠️ 1 Medium | 🔵 2 Low | 💡 1 Suggestions

Summary

The diff adds a conflict guard to prevent installed skills from shadowing eval-created synthetic skills, which is a solid defensive measure. The normalizeDescriptionOverride helper and the findSkillConflicts unit tests are well done. The main concern is a behavioral change in runSingleQuery that alters the early-return condition, which could affect eval results in non-triggerOnly mode if not carefully validated.

---

Robin — the Robin Hood of code review. Free for every PR.