feat: invoke Agent Skills via slash commands

[NTaylorMullen]

What would you like to be added?

Enable the ability to invoke Agent Skills directly via slash commands (e.g., /<skill-name>). This should also support passing optional arguments that are then submitted as a follow-up prompt to the Gemini model (e.g., /<skill-name> hello world).

Additionally, for these explicit client-initiated skill activations, the CLI should bypass user confirmation prompts if the policy engine returns an ASK_USER decision, treating the slash command as implicit user consent.

Why is this needed?

Users want a faster way to trigger specific skills without relying on the model to select them or navigating through the /skills menu. Supporting follow-up prompts allows for a more streamlined "trigger and query" workflow. Bypassing confirmation for these explicit commands reduces friction for power users.

Additional context

This issue is being created retroactively to track the implementation provided in PR #21758.

[SaschaHeyer]

+1

[bayov]

Generally, I think it would be best to consolidate skills and commands altogether.

Something like (ignoring bikeshedding):

---
name: pdf-processing
description: Extract PDF text, fill forms, merge files. Use when handling PDFs.
command: process-pdf
command_only: true
---

• Optional command Frontmatter field simply makes the skill usable as a custom /process-pdf command.
• Optional command_only Frontmatter field controls whether the agent is even allowed to autonomously discover and use the skill. When enabled, the skill description is not added to agent's context, and the agent cannot load it.

Basically, these two fields would allow controlling whether a SKILL.md defines a command, an implicitly usable tool, or both.

[Jorge-D-Robles]

I'd like the option to also have a /command that automatically loads my custom agents as a follow-up to this feature.
You could do something like /agent , with the UI then showing at the bottom next to the input text box, showing the user the selected agent.
It could be loaded like this:
--start--
System prompt, you are gemini CLI, etc
---end of system prompt---
---Agent prompt---
full custom agent prompt. I.E. this is a code reviewer agent, focusing on making sure the Angular code written follows our best practices
---end of agent prompt--

Essentially, it'd be part of the prompt that's loaded up even when you clear or compress the context window. I've been getting away with doing a variation of this by running gemini CLI with a custom system prompt (that includes whatever relevant system prompt is already defaulted) and linking it with a bash alias in .bashrc, but it's clunky and means I can't just switch my main agent at will.

This would unlock the ability to create full on custom agent workflows that rely on their own subagents, etc. I tried to create my own system with Gemini CLI that worked pretty well, and my friend and I also made our own CLI harness for Gemini to experiment further with this idea. Would love to discuss further on chat or here!

[fpuga]

I’m not sure if this is feasible, but Anthropic already uses a property in the frontmatter of SKILL.md that serves the same purpose as the proposed command_only.

Using disable-model-invocation as the property name would be a huge win for user consistency across platforms.

Related Claude doc

Updade

• Cursor implements the same flag.
• Code has a similar feature (allow_implicit_invocation) but it seem to be a bit of discussion around it.