Skills Catalog Management

The always-on cost of skills is the rendered ## Skills block inside the system prompt — not the size of every SKILL.md on disk. Skill bodies load on demand.

Measuring the block

hermes prompt-size [--json] [--platform cli]

Offline, no API call. Reports the skills index characters/bytes, plus per-skill index cost. Before/after measurements are comparable only with the same platform and working directory.

Measured on the Agent Host (cli platform, 2026-09-12):

State	Skills index (chars)
Before reform (YEU .agents/skills in external_dirs + 29 global symlinks)	25 008
external_dirs removed (77 bmad skills out of the global catalog)	16 510
29 Matt Pocock symlinks removed from ~/.hermes/skills	13 649

The reform cut the rendered block by 45 %; zero bmad skills remain in a session outside the YEU repo.

Source scopes

Source	Scope	Note
~/.hermes/skills	Profile-global	Only genuinely cross-task skills belong here
skills.external_dirs	Profile-global, read-only	Repo directories belong to project scope, not here
<git-root>/.hermes/skills, <git-root>/.agents/skills	Project-scoped	Load only when the session working directory is inside the repo
skills.trusted_project_dirs	Trust barrier	hermes skills trust <path> / hermes skills untrust <path>; untrusted repos never inject skills
Skill body	On demand	skill_view or the skill command

Project skills shadow same-named profile skills and are tagged [project] in the index.

Guardrails and approvals

Setting	Effect
skills.guard_agent_created: true	Security scan on every agent-created skill write; not an approval workflow
skills.write_approval: true	Stages skill mutations for review: /skills pending, /skills diff <id>, /skills approve <id>, /skills reject <id>
skills.creation_nudge_interval: 0	Disables the periodic skill-review nudge; existing skills are untouched
memory.write_approval: true	Same gate for memory: /memory pending, /memory approve <id>, /memory reject <id>
memory.nudge_interval: 0	Disables periodic memory nudges
Curator (enabled)	Manages created_by: agent skills only: marks stale, archives with backups, never deletes

Disabling a skill

skills.disabled in config.yaml removes names from the rendered index; the skill files stay on disk and remain loadable explicitly.

Reversible change procedure

1. Back up the sources (~/.agents/, config, symlink manifest).
2. Move rather than delete (symlinks to a backup directory, cp -a for canonical sources).
3. Change one source at a time; start a fresh session or measure with hermes prompt-size after each step.
4. Verify absence outside the trusted repo and presence inside it (project skills require the session working directory inside the repo).