Token usage metadata shows zeros despite high input token consumption from Zed editor

[matifali]

Description

When using Zed editor with AI Bridge, the input token count is significantly higher than expected, but the Token Usage Metadata display shows all zeros.

Problem

• Sending a simple message like "Hello from Zed" results in 6353 input tokens
• Zed's "Text" mode: ~6000+ input tokens
• Zed's "Agent" mode: ~11000+ input tokens
• Token Usage Metadata in AI Bridge displays all zeros, making it impossible to track/debug token consumption

Expected Behavior

Token Usage Metadata should accurately reflect actual token consumption so users can understand what's contributing to high token input counts.

Questions

• Where in AI Bridge can users see what's consuming the extra input tokens?
• Why does the Token Usage Metadata show zeros when tokens are clearly being consumed?

Context

This may be related to Zed injecting additional context (system prompts, editor context, etc.) that isn't being tracked properly by the metadata display or AI Bridge in general.

[matifali]

Turns out the extra tokens are injected by the Coder's MCP tool injection.
In Agent Mode, the extra tokens come from a long system prompt from Zed.

As a follow-up, I suggest we capture where the Input tokens come from and their breakdown (prompt, system, injected, etc.) in the token metadata.

[matifali]

Closign in favour of #120