mirror of https://github.com/garrytan/gstack.git
c48972c0c8
The skill-llm-eval.test.ts judge test for 'document-release/SKILL.md workflow' slices from '# Document Release:' to '## Important Rules' in the combined SKILL.md + sections/release-body.md content. The marker was in SKILL.md (after Step 1.5), causing the judge to see only the stub (Steps 0-1.5) instead of the full 9-step workflow — resulting in scores of clarity=3, completeness=2, actionability=3 (all below the >=4/>=3/>=4 thresholds). Fix: move ## Important Rules from SKILL.md.tmpl to the end of release-body.md.tmpl (after Step 9 and CODEX_DOC_REVIEW). Regenerate. The judge now sees all 9 steps (~26K chars vs ~3K before) and should score at or above the required thresholds. Confirmed fix: runWorkflowJudge combines SKILL.md + sections/release-body.md, searches for '## Important Rules' in the combined string — which now appears at offset 69382 (after release-body content ends at Step 9). |
||
|---|---|---|
| .. | ||
| sections | ||
| SKILL.md | ||
| SKILL.md.tmpl | ||