Problem

The agent.tool.responseLength telemetry event was only reporting the size of tool outputs, missing crucial information about the size of tool inputs. This made it difficult to get a complete picture of token consumption and costs associated with tool invocations.

Solution

Modified the sendInvokedToolTelemetry function to calculate and report both input and output token counts:

1. Added toolInput parameter to the function signature to accept the tool call input as a string
2. Parallel token counting using Promise.all to efficiently calculate both:
   • tokenCount: Number of tokens in the tool response (output)
   • inputTokenCount: Number of tokens in the tool input
3. Updated GDPR telemetry metadata to document the new inputTokenCount measurement field
4. Updated all call sites to pass the tool input:
   • In toolCalling.tsx: Pass this.props.toolCall.arguments (already a JSON string)
   • In chatVariables.tsx: Serialize the input object with JSON.stringify(input)

Example

When a tool is invoked with input like:

{"param1": "value1", "param2": "value2"}

The telemetry event now includes:

• toolName: Name of the tool being invoked
• tokenCount: Token count of the tool's response
• inputTokenCount: Token count of the tool's input (new)

This provides complete visibility into token consumption for both sides of tool invocations, enabling better cost tracking and optimization.

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.