Skill usage and context cost

The catalog records skill use so a team can distinguish a useful grant from instructions that load repeatedly without helping. Usage is evidence of recorded activity, not a billing estimate for a whole task.

Inspect this conversation

Show skill usage for this channel over the last 30 days. Identify granted skills with no recorded use and explain any inferred records.

skill_usage_report defaults to 30 days and accepts 1–365 days. It returns one total per skill, the exact time window, recorded user/conversation attribution, inferred-read provenance, and shared grants never used in that range. Missing historical identities remain unknown rather than being guessed.

Explore the administrator view

Admin → Skills → Usage supports skill and conversation views and filters for the available record dimensions. The catalog’s main table shows Usage 30d, while a skill detail shows its longer recorded history. Use attribution and timestamps to understand what produced a count.

A package description is always present when granted; detailed files are read when the agent applies it. show_channel_skills estimates always-on context across the effective profile and flags overlapping triggers or missing dependencies.

Tune the profile

The default context soft cap is 6,000 tokens, configurable under Skills → Sync settings → Catalog settings → Context soft cap. It is a warning threshold, not a hard grant limit. Organization grants contribute to every conversation, so account for them when previewing a template.

Deactivate unused optional conversation skills, remove personal grants you no longer need, or refine overly broad trigger descriptions through the correct authoring path. A dependency cannot be removed while its parent still requires it. Mandatory organization packages require an administrator to change their grant.

Related: grants, discovery, usage and costs, and usage control.