A catalogue of agent skills that measures its own cost: how many tokens each installed skill costs every session, and what happens to routing when two descriptions compete.
Measured
Every figure below is produced by a command rather than asserted. The validator fails if the taxonomy changed and the registry did not.
- Skills in the catalogue
- 260
- Runnable
- 234
- Need an MCP server
- 26
- Upstream sources
- 7
- Clusters
- 25
- Profiles
- 8
- Tokens per session, full set
- 22.3k
- Tokens per session, profile
- 3.2k
Capabilities
- Measuring a catalogue's cost
How many tokens an agent pays merely for having skills installed, and why that bill cannot be read off the skill count.
- Routing conflicts between skills
What happens when two descriptions overlap, why a "percentage of skills without a trigger" metric cannot be published, and how to measure conflict so the number cannot inflate itself.
Questions
How much does a 260-skill catalogue cost?
22.3k tokens per session for the descriptions alone, whatever the task. Under the infra-ops profile, 49 skills remain and the bill drops to 3.2k.
What about the 26 blocked skills?
They require an MCP server that is not connected. Their descriptions are prefixed with [needs MCP: …] so the agent routes around them instead of picking one and failing. The prefix clears itself when the server is connected.
How do you know two skills conflict?
By word-set overlap of the descriptions at a Jaccard threshold of 0.25, which yields 24 pairs out of 260. A pair counts as resolved only when a guard or trigger on one side names the other skill explicitly.
What does the gate check before a commit?
That the installed copy matches upstream, that the description matches the registry, skill availability, conflicting pairs, broken cluster and profile references, orphans in the install directory, and whether the registry is stale relative to the taxonomy.