Invoke user skills from a chat message (phase 1)

[andrii-novikov]

QuickApps version

latest

What is the problem this feature will solve?

A QuickApp agent can only use the skills its author attached to it, and the model alone decides when to use them. As a user I can't:

• make the agent use a skill for a specific message. Writing "use the code-review skill" is only a hint that the agent may ignore.
• use my own skills with an agent I don't own. Skills I wrote in DIAL, or that were shared with me or published, are unavailable unless the agent's author attaches each one to the agent.

What is the feature you are proposing to solve the problem?

Let the user pick one or more of their own skills when writing a message, and have the agent apply them to that message. In DIAL Chat this is a / picker listing the user's skills: their own, shared with them, and published.

Expected behaviour

• The user can pick any skill they have access to, whether or not it is attached to the agent.
• A picked skill is always applied to that message; the agent doesn't get to skip it.
• The user can see in the response that the skill was used.
• The skill stays available for the rest of the conversation, including the reference files bundled with it.
• If a picked skill can't be used (deleted, access lost, or invalid), the agent says so and still answers the message.
• If a picked skill has the same name as one of the agent's own, both stay usable and neither replaces the other: the picked one is applied to the message, and the agent keeps its own.
• Messages without a picked skill behave exactly as today.
• Available as a preview feature.

Not in this issue: picking the agent's own skills (the ones its author attached, and the built-in ones) is tracked separately. Until then the agent keeps using them on its own, and the user can ask for one in words.

Related: DIAL Core support for skills referenced from a message: epam/ai-dial-core#1955 (done). DIAL Chat: the / skill picker (to be filed).

What alternatives have you considered?

• Asking in words ("use my code-review skill"): works only for skills the agent already has, and the agent may still ignore it.
• Asking the agent's author to attach the skill: doesn't work for personal skills at scale, has to be repeated for every agent, and applies the skill to every user of that agent, not only to the one who wants it.