| name | onboard-agent-asset |
| description | Triage and onboard new agent assets discovered in watched Claude, Codex, or Agents folders. Use when watch-dirs reports unmanaged files, when the user asks to onboard a skill/hook/agent asset, or when migrating a Claude/Codex skill into shared `.agents` ownership without letting the agent decide ownership silently. |
Onboard Agent Asset
Use this skill to turn a new or unmanaged agent asset into an intentional chezmoi-managed source file, ignore rule, or follow-up decision.
Guardrails
- Inspect first. Do not add, move, delete, or ignore until the asset is classified.
- Never decide ownership silently. Recommend, then ask before mutation unless the user already gave an explicit exact action.
- Preserve history with
git mv when the asset is already tracked in this repo.
- Treat
chezmoi status, chezmoi diff, and chezmoi cat as inspection commands.
- Treat
chezmoi apply as mutation.
- Never run global
chezmoi apply during onboarding.
- After classification, proceed automatically when there is no ownership or scope ambiguity and the change only touches the asset being onboarded.
- Ask before mutation only when classification, ownership, target path, generated/plugin status, or live-data side effects are unclear.
- Onboarding is not complete until the original live asset path still works for the tool or user that created it.
- When a source file, adapter, ignore rule, or executable-mode fix changes live state, finish automatically with path-scoped
chezmoi apply for only the affected target path(s).
- Apply only the exact live asset path(s) just onboarded or updated. Never apply a parent directory, watched directory, or global target.
- Commit exactly one onboarded skill or asset per commit. Stage only its source files, required adapters, and required ignore/watch updates.
- Never stage or commit unrelated dirty files, even when they are in the same repo or surfaced by
git status.
- Keep commits semantic:
- pure moves
- content cleanup
- adapters
- ignore/watch updates
- Do not onboard generated/plugin-managed assets into
.agents just because they look skill-shaped.
Classify
For each candidate path, inspect enough to choose one class:
- Shared self-authored skill: has
SKILL.md, appears written/owned by Flo, and is useful across agents.
- Tool-specific asset: depends on Claude-only or Codex-only config, slash commands, permissions, UI, or runtime behavior.
- External/plugin-managed asset: installed by a marketplace, plugin, GSD, npm/package manager, or another automated source.
- Support workspace: evals, test docs, iteration state, benchmarks, or scripts supporting a skill but not itself a skill.
- Not an agent asset: cache, log, generated file, local state, or unrelated file.
Actions
Recommend exactly one action per candidate:
- Shared skill:
- move canonical source to
dot_agents/skills/<name>
- add Claude adapter
dot_claude/skills/symlink_<name>.tmpl
- skip Codex adapter unless verification proves Codex cannot load
~/.agents/skills
- Tool-specific asset:
- keep under
dot_claude or dot_codex
- document why it is tool-specific if the reason is not obvious
- External/plugin-managed asset:
- add or confirm
.chezmoiignore entry
- keep watched directory coverage so new surprises still surface
- Support workspace:
- move to
dot_agents/skill-workspaces/<name> when shared
- add a tool adapter only when existing paths must keep working
- Not an asset:
- ignore or skip; do not force it into
.agents
Verification
Before reporting done:
git status --shortgit diff --cached --summary --find-renames for staged moves- show
chezmoi status
chezmoi cat or chezmoi diff for symlink adapters when relevant- verify every original live asset path still exists and resolves as intended after any path-scoped apply
codex debug prompt-input "probe" when Codex skill discovery or home instructions changed- report whether
chezmoi apply was run
- if anything was applied, name the exact target paths
- remind Flo about any skipped candidates that are good tests for this skill
