Feature Request: Allow languageModelTools to specify required tool invocation behavior

[sgaabdu4]

Problem

Extension developers contributing tools via the languageModelTools contribution point have no way to ensure their tool is actually called by the LLM. The tool is registered and available, but the LLM (Copilot) decides autonomously whether to invoke it—often skipping the tool entirely even when instructed to call it.

This is problematic for tools that are required for a workflow to function correctly. Academic research shows LLM instruction compliance in agent scenarios is below 30%, making strong modelDescription text insufficient.

Current State

VS Code already has LanguageModelChatToolMode.Required in the API:

export enum LanguageModelChatToolMode {
    Auto = 1,
    Required = 2
}

However, this enum is only accessible to callers making LLM requests (extensions using lm.sendChatRequest()). Tool providers registering via languageModelTools or lm.registerTool() have no mechanism to influence how Copilot invokes the LLM with their tool.

Proposed Solutions

Option A: Required Invocation Mode

Allow tool contributors to declare required invocation behavior:

"languageModelTools": [
  {
    "name": "my_tool",
    "modelDescription": "...",
    "invocationMode": "required",
    "inputSchema": { ... }
  }
]

When a tool declares invocationMode: "required", Copilot would use LanguageModelChatToolMode.Required when sending requests that include this tool.

Option B: Priority Hints (Lighter Touch)

If hard requirements are too strong, a softer "priority" system could help:

"languageModelTools": [
  {
    "name": "my_tool",
    "modelDescription": "...",
    "priority": "high",  // or "normal", "low"
    "inputSchema": { ... }
  }
]

Copilot could use priority hints to:

• Boost the tool's prominence in the system prompt
• Add internal instructions emphasizing the tool's importance
• Weight tool selection when multiple tools are available

Option C: Both

Support both options—priority for soft hints and invocationMode: "required" for strict workflows where the extension knows the tool must be called.

Alternatives Considered

1. Stronger modelDescription: Already tried with maximum-strength wording. Models still skip the tool frequently.
2. Custom agent modes via .agent.md: Adds instructional text to the system prompt but cannot override Copilot's LLM invocation settings.
3. Injecting instructions via tool results: Adding reminder instructions in tool output; still not reliable.

Use Cases

1. Workflow continuation tools: Tools that must be called to proceed to the next step.
2. Approval gates: Tools requiring explicit confirmation before proceeding.
3. Mandatory checkpoints: Tools that enforce safety checks or logging.
4. Notification tools: Tools that must report results or status.

Environment

• VS Code Version: 1.112.0-insider
• OS: macOS

Related

• LanguageModelChatToolMode enum exists but is caller-side only
• languageModelTools contribution point lacks invocation control
• Similar pattern to MCP's proposed tool_choice parameter

[vs-code-engineering]

This feature request is now a candidate for our backlog. The community has 60 days to upvote the issue. If it receives 20 upvotes we will move it to our backlog. If not, we will close it. To learn more about how we handle feature requests, please see our documentation.

Happy Coding!