Agent Policy

---
name: agent-policy
description: >-
  Author, locate, register, or activate source-owned agent instructions, skills, hooks, and modules.
grounded: 2026-10-09
written: 2026-10-09
metadata:
  kind: playbook
---

# Agent policy

Keep one source for each rule. Use `agents status`, `agents inspect <id>`, and
`agents path <id>` to resolve the live catalog and owner; do not infer source
locations from projection paths or remembered repository layouts.

Use the nearest bootstrap for universal boundaries, a skill for a triggered
workflow, a module for a group of registered units, and hook code for
mechanical enforcement. Permission and resolved activity are distinct.

## SKILL.md and references

One skill per idea, with a plain name. Keep trigger descriptions short and
`SKILL.md` complete for ordinary execution. The skill's `references/` retain
constraints, rationale, examples, alternatives, and clearly labeled exploratory
ideas. Read them for a named detail or when revising `SKILL.md`; do not load
them routinely. Split long notes by topic and link the relevant file from
`SKILL.md` or `references/notes.md`. Update both layers when an adopted rule
changes.

## Change and activate

1. Read owner instructions; edit an owned worktree, never a projection or pin.
2. Follow `skill-creator`. Register each identity once; preserve explicit
   module membership and hook events. Do not add aliases or duplicate rules.
3. Run the relevant source/catalog check. Local skill edits need no NixOS
   rebuild.
4. Land required owner and catalog commits and update clean main checkouts.
   Run `agents sync`; change permission only when requested or needed.
5. Verify the active ID, owner revision, and provider projection resolve to
   that published source. Unlanded or ambiguous authority blocks activation.

Size global policy, repository AGENTS.md files and skills by measurement, not line caps: `agents sync` fails a change only when replay conformance drops or it adds startup tokens with no measured gain against south:agents/operating-point.toml. Put command detail in CLI help, `help TOPIC` or indexed docs; skills keep prerequisite knowledge.

Use `agents --help` and [catalog notes](references/notes.md) for catalog/projection detail.