Does disable-model-invocation in Custom Agents Actually Reduce Context Window Usage and Token Impact?

[r-shakeri]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

VS Code

Body

Hi,

I’m trying to better understand how disable-model-invocation: true affects context loading for Agent Skills.

From the documentation and other sources, I understand that:

Skills are discovered at startup
Their metadata (name + description) is used for matching and be included in the model context
Full skill content is loaded only when the skill is activated
However, it’s unclear how disable-model-invocation interacts with this process.

My question is:

If a skill has disable-model-invocation: true, and we explicitly invoke which skills an agent should use,
is all skills metadata (e.g. name and description) still included in the model’s initial context / skill discovery list,
or are they completely excluded from the context unless explicitly invoked?

In other words:

Does this flag only disable automatic invocation?
Or does it also prevent the skill from contributing to the context (e.g. metadata injection / token usage)?
This is especially important for understanding context window usage and token impact when working with many skills.

Thanks!

[AniXDex]

Hi @r-shakeri ,

My understanding is that disable-model-invocation: true only disables automatic invocation by the model.

The skill should still be discovered/registered, with its metadata (name/description) available, while the full skill content is only loaded when explicitly invoked.

I haven't seen documentation confirming whether that metadata is excluded from the model's context or token budget, so a GitHub team member may need to clarify that part.

If this helps answer your question, please consider marking it as the accepted answer. 🙂

[r-shakeri]

Thanks for the reply, But I really need to understand that specific part, because it's a key factor in deciding whether I can add more skills without worrying about unnecessarily consuming the context window.

[ashish-da]

No — disable-model-invocation does not reduce context window usage or token impact directly.

What disable-model-invocation Actually Does
It prevents the agent from calling/invoking sub-models or nested model calls within a custom agent pipeline. It is a control flag, not a token optimization tool.

[jasonmarie8]

Good question this flag usually only stops automatic invocation, while skill metadata can still be present for discovery, so it won’t fully remove context/token usage unless explicitly excluded. Pretty similar to how you optimize strategy in Pakrummy game, where control matters more than presence.

[tanvishinde017]

Hi @r-shakeri,

My understanding is that disable-model-invocation: true only prevents the model from automatically selecting or invoking the skill.

The skill itself is still discovered by VS Code/Copilot, and its metadata (such as name and description) remains available so that the system knows the skill exists. The full contents of the skill are typically loaded only when the skill is explicitly invoked.

So, based on the current documentation, I would interpret the behavior as:

• ✅ Skill discovery still occurs.
• ✅ Metadata remains available.
• ✅ Explicit invocation should still work.
• ❌ Automatic model-driven invocation is disabled.
• ❓ Whether metadata contributes to token usage in the model context may depend on implementation details that aren't clearly documented.

I haven't seen documentation explicitly stating that disable-model-invocation removes the skill entirely from the discovery/context phase, so I would currently assume it affects invocation behavior rather than visibility.

Hopefully someone from the Copilot team can confirm the exact token and context implications, since that's an important consideration when managing large numbers of skills.