/manage

name: manage
description: Lifecycle rules for user-created custom skills — versions and drafts, application modes, publishing, rollback, deletion, scopes, and approvals. Load before creating, listing, reading, publishing, rolling back, changing the mode of, or deleting a custom skill.

Managing custom skills

These rules apply to custom skills saved with this extension's `create` tool. The agent's own authored skills have no lifecycle versions.

`create`, `list`, and `publish` are always available. `read`, `set_mode`, and `remove` are deferred: find them with `eve__search` (for example "read custom skill") and call them through `eve__tool`.

• A clean first version goes live immediately. Every later save is an

immutable draft: it changes neither the live version nor how it applies. The live version changes only when `publish` succeeds.

• Each version is saved with an application mode, and it applies in that

mode once published. `set_mode` changes how the live version applies.

• The live (`active`) version is what applies. Versions above it are drafts;

versions below it are superseded. To roll back, `publish` the earlier version.

• Every `create` call needs `applicationMode`:
• `always` for a persona, writing preference, or standing behavior. It is

injected on every turn.

• `on-demand` for a task procedure. It is listed in the prompt and loads

when relevant.

• `deferred` for a rarely needed procedure. It stays out of the prompt

until a skill search finds it.

• An `on-demand` or `deferred` description is a routing condition that says

when to load the skill, such as "Use when the user asks for a deploy checklist." If `create`, `publish`, or `set_mode` rejects a description or instructions, save a new version that follows the error.

• On-demand and deferred skills load through the skill loader under their

advertised name, which carries the mount prefix and, for shared scopes such as `channel`, a scope prefix. `list` and `create` return that `loadName`. Lifecycle tools take the bare name plus `scope`.

• Changes apply from the next message, not the current turn.
• Use `read` to inspect an exact draft before publishing it; it also lists

every version.

• `publish`, `set_mode`, and `remove` take each skill's `id` as

well as its name, from `list`, `read`, or `create`. If a call says the skill was replaced, or that the scope changed while it was checked, list the skills again and call it with the current id.

• Some changes need approval, such as publishing, rolling back, deleting, and

anything that goes live in a shared scope. Call the tool directly with the exact target: the approval card shows the change and is the only confirmation, so never ask first in prose or with a question tool.

• When a live skill is not applying, check its application mode before

anything else. Never publish a version that is already live; switch an `on-demand` skill with persistent behavior to `always` with `set_mode`.

• To delete several skills, call `list` once, then `remove` once with every

skill's name and id.

• `list` names the scopes available here. Default to the personal scope. Use a

shared scope such as `channel` only when the user asks for a team or channel skill. A shared-scope change affects everyone in that scope. A read-only scope's skills can be loaded but not changed.

• After an approval or denial, report the actual tool result. Never call a

draft live before `publish` succeeds.