Skip to content

Feature request: make provider skill handles discoverable #506

@guibor

Description

@guibor

Checklist

  • I agree to communicate with the author myself (not AI-generated).
  • I've read the README's Filing issues section.
  • I'm running the latest versions (fill in below).
    • agent-shell version: 20260329.1008
    • acp.el version: 20260325.1158
    • ACP package (e.g. claude-agent-acp) version: codex-acp 0.11.1
    • Agent CLI (e.g. claude, gemini) version: codex-cli 0.118.0
  • For requesting new agent support, I'm including a link to the ACP-capable agent or related ACP package.
  • For issues, I'm including ACP traffic (as per README).

Request

Could agent-shell make provider-defined skill handles / commands discoverable in the same way native agent UIs do?

Problem

In native agent UIs, skills are often discoverable from a prefix trigger:

  • Cursor: typing / shows available commands/skills.
  • Codex: typing $ shows available skills.

In agent-shell, those provider-specific handles are not discoverable, so unless the user already knows the exact handle names, they are effectively hidden.

This feels adjacent to #486, but a bit broader. That issue is about Claude-installed plugins not appearing. This request is about provider-defined skill/handle discovery in general.

Why this matters

A big part of the UX value of skills is discoverability. In the native tools, the prefix trigger teaches the user what is available. In agent-shell, the same capabilities may exist on the provider side, but there is no equivalent discovery surface.

That makes skills much harder to use from Emacs, especially when switching between providers with different conventions.

Expected behavior

One of these would already help a lot:

  • a completion popup when typing a provider-specific trigger such as / or $
  • a buffer section similar to "Available /commands" for skills / handles
  • an extension point so ACP adapters can advertise discoverable skills/handles and the UI can render them consistently

Concrete examples

  • With Cursor-style agents, typing / would ideally reveal the available commands.
  • With Codex, typing $ would ideally reveal the available skills.

Even if the exact rendering is provider-specific, having some discoverable list inside agent-shell would make these features much easier to use.

Question

Is this something agent-shell could support directly, or is the missing piece more likely in the ACP adapters so they can advertise skills/handles in a way agent-shell can consume?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions