`/context` command injects output into conversation history, burning the context it measures

Open 💬 2 comments Opened Jun 26, 2026 by taylorsegell

Problem

/context is designed to show token usage, but its output is injected into conversation history as an assistant message. This means running /context consumes the very resource it's reporting on.

Observed behavior:

  • Session start: ~14% context used (baseline overhead)
  • After two /context calls: ~26% context used
  • ~12% of a 200k context window (~24k tokens) consumed just from running the diagnostic command twice

The full MCP tool table and skill listing in /context all output is particularly large — 200+ tools each with token counts serialized into message history.

Expected behavior

/context should render client-side only (e.g., printed to terminal stdout / rendered in the UI status bar) and never be injected into the conversation history.

Same principle as git status — it reports state without modifying state.

Suggested fix

Implement /context as a local terminal command that reads session token state and renders it in the UI without entering the conversation. The output should be ephemeral — visible to the user but not serialized into the context window.

Environment

  • Claude Code CLI (Mac)
  • Model: claude-sonnet-4-6
  • Context window: 200k tokens

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗