What problem does it solve? Agent harnesses inject the full skills/tools catalog (names + descriptions) into context every turn and into every subagent, so an auto-minted catalog of hundreds of skills multiplies token cost on fan-out and can overflow small-context agents entirely. Over-cap descriptions are silently truncated mid-word, killing trigger phrases with no warning. ## Core Features & Use Cases - Catalog audit and curation methodology: Classify skills by description intent (episodic traps vs real procedures), apply conservative asymmetric bias, and use levers like skillOverrides, disable-model-invocation, and the native skillListingBudgetFraction/skillListingMaxDescChars budget across Claude Code, Cursor, Codex, Copilot CLI, and Gemini CLI. - Publish-time description gate: scripts/check_skill_descriptions.py fails CI when a SKILL.md description exceeds the 1536-char cap, lists trigger phrases lost to truncation, detects line-wrap corruption, and diffs trigger surfaces between revisions with --compare. - Measurement and reporting: Measure per-turn and per-subagent catalog overhead via /doctor and cache-token analysis, then render an interactive HTML recap of a skillOverrides treatment with scripts/render_treatment_report.py. - Use Case: A team running an auto-skill-minting loop finds subagents carrying ~30k tokens of base context; use this skill to hide episodic trap lessons globally, scope irrelevant skills per-project, and gate all published descriptions against the cap. ## Quick Start Ask the agent to audit your skills catalog for context cost and run scripts/check_skill_descriptions.py against your skills directory to find over-cap descriptions and dead trigger phrases.
