01Safety
0.0 / 30
What changed in the harness
Selection accuracy 100→100, token cost up 8%, unconfirmed writes 0%→0%.
Category breakdown
Where the score comes from.
Earned points across the four signals Gradable measures. Safety and Legibility are scored out of 30; Economics and Discoverability are scored out of 20.
02Legibility
29.1 / 30
03Economics
20.0 / 20
04Discoverability
13.8 / 20
Highest-impact fix
Estimated gain +30 pointsAdd explicit identity and permission preflight tools
Expose machine-readable principal/tenant confirmation and a non-mutating permission check so agents can verify both before destructive actions.
Description evidence
Defects and rewrites.
2 defects found across the exposed tool descriptions. Suggested rewrites make purpose, inputs, boundaries, and returns easier for an agent to understand.
| Tool | Defect types | Suggested rewrite |
|---|---|---|
search_skillselion |
no_return_description |
Search Skillselion, a curated directory of Claude Code agent skills, MCP servers and plugin marketplaces, ranked by real community signal (installs + GitHub stars). Use to browse trusted skills/MCPs/marketplaces for the task at hand (e.g. "postgres", "code review", "playwright"). Returns a ranked list of matching listings (skill ids, MCP/plugin names, plus relevance and community-signal context) so you can feed an id to load_skill; omit type to search all listing types and use limit to cap results. |
top_skillselion |
no_return_description |
List the top entries on Skillselion (the leaderboard), optionally by type. Skills rank by installs; MCP servers and marketplaces by GitHub stars. Returns the ranked leaderboard entries — the names/ids and ranking metadata for the top listings in the chosen category — so you can see what the community trusts most or pick a starting point to load. Use type to restrict to skill, mcp, or marketplace, and limit to cap how many entries come back. |
Selection evidence
Confusable tool pairs.
2 pairs where similar names or overlapping descriptions may send an agent toward the wrong tool.
| Tool A | Tool B | Confidence | Why they collide |
|---|---|---|---|
load_skill |
synthesize_skills |
medium | A task phrased as a domain question, e.g. "what's the current best way to do e2e testing in my repo", sits on the fuzzy boundary: load_skill covers 'tasks a skill likely covers with a proven recipe', while synthesize_skills covers 'broad best-practices across several skills'. The definitions cross-reference each other, so an agent deciding between 'one skill's recipe' vs 'merged field guidance' can plausibly pick wrong for a single vaguely-broad query. |
search_skillselion |
top_skillselion |
low | search_skillselion requires a semantic query while top_skillselion is a query-less leaderboard; the only shared link is the brand token 'skillselion'. Confusion would need a task like 'list the best postgres skills', but 'list/top' strongly signals the leaderboard and 'for <topic>' strongly signals search, and both descriptions make the distinction explicit. |
Compare the field