Question: VS Code chat (via codemie proxy connect vscode) shows ~94k tokens used on a fresh session vs ~19k in terminal CLI

[pdemchuk]

Description

Comparing context usage on two freshly-started sessions:

1. Terminal, via codemie-claude CLI: fresh session, ran /model and /context.
   Result: 19.3k / 967k tokens (2%) — System prompt 5.9k, Memory files 4.3k,
   Skills 8.8k, Messages 233 tokens, Autocompact buffer 33k.

2. VS Code chat, connected via codemie proxy connect vscode: fresh session,
   sent a single one-line message ("Hey").
   Result: 67.0k / 200k tokens (34%) used, with Messages alone at 60.9k tokens,
   plus a 33k autocompact buffer reserved on top — roughly 94k of the 200k window
   consumed before any real conversation took place.

Expected Behavior

Fixed overhead (system prompt, memory files, skills) should be comparable in
absolute terms between the CLI and the VS Code proxy integration, and "Messages"
should reflect only the actual conversation content, not grow disproportionately
on a one-line exchange.

Actual Behavior

The VS Code chat session's "Messages" category is ~60.9k tokens after a single
one-line user message, which is far larger than would be expected from the
message content alone. This suggests something (e.g. repeated system-reminder
context, or per-turn re-injection of skills/agent listings) may be counted
under "Messages" in the VS Code integration rather than as a stable one-time
category, the way the CLI's /context output separates them.

Question

Is this expected behavior for the codemie proxy connect vscode integration?
Is there a way to confirm whether per-turn context (e.g. available-skills or
available-agents listings) is being re-sent on every message rather than cached,
and if so, is that something CodeMie's proxy controls or is it inherent to how
VS Code's chat/agent request routing works?

Environment

Field	Value
OS	macOS (Darwin)
Interface	VS Code native chat, connected via CodeMie proxy