feat(skills): skill-usage observability + parametrized catalog test#73
Merged
Conversation
…ized catalog test Two pieces of skill-usage evaluation, both small: 1. Daily-maintenance §1 'Skill usage scan' subsection. Aggregates yesterday's audit log for read_file calls on src/skills/<name>/skill.md paths → per-skill discovery counts. Operator decision rule: consistent zero-reads with obvious triggers = catalog issue (refine frontmatter), obsolete skill (delete), or genuinely missed (note + watch). Counts land in §4 journal synthesis under a new '### Skill usage' sub-section so the trend is queryable. 2. Parametrized 'every skill in src/skills/<name>/' must appear in the assembled system prompt's catalog. Added at the eval-harness structural layer. Trigger: PR #70 shipped exa-search without any catalog-presence test — under the prior single-skill (test_skill_creator_visible_in_catalog), a new skill could ship invisible. Parametrize fixes it for every future skill automatically. Plus .gitignore entry for mining/ so the blog scratch from session-jsonl mining doesn't keep leaking into PRs (the entry on the ask_operator branch in PR #71 hasn't merged yet). Together (1) + (2) cover the operator's catalog-presence and discovery-observability questions on SAM-43. The deeper Opus-as-judge ('did Sam apply the right skill') stays as a separate follow-up. Tests: 9 skills defended by the parametrize. 16 eval tests pass (was 8). Full suite: 143 passed locally.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this enables
Two pieces of skill-usage evaluation, addressing the SAM-43 questions on catalog verification + audit-log-based usage measurement:
1. Skill usage scan in daily-maintenance §1
New subsection that aggregates yesterday's audit log for
read_filecalls onsrc/skills/<name>/skill.mdpaths → per-skill discovery counts. Counts land in §4 journal synthesis under a new### Skill usagesub-section so the trend is queryable over time.Decision rule when a skill has consistent zero reads: catalog issue (refine frontmatter), obsolete skill (delete), or genuinely missed (note + watch). Sam-as-LLM makes the call.
2. Parametrized catalog-presence test
test_every_skill_visible_in_catalogintests/eval/test_structural.pyparametrizes over everysrc/skills/<name>/directory. Asserts the skill name appears in the assembled system prompt. Adding a new skill automatically gets defended — no manual test addition needed.Trigger: PR #70 shipped exa-search without any catalog-presence test. Under the prior single-skill check, a new skill could ship invisible. This parametrize fixes that for every future skill.
Consequences
What this doesn't cover
How to verify
pytest tests/eval/test_structural.py— 16 passed (was 8); 9 new parametrized cases, one per skill.### Skill usageblock with per-skill counts.Bonus
.gitignoreaddsmining/so blog-scratch files from session-jsonl mining don't keep leaking into PRs (the entry on PR #71 hasn't merged yet).Tier
Tier 1 (skill prose + tests). No runtime changes.
Closes the catalog-verification + skill-usage-observability part of SAM-43.