Session limit burns at 2-3% per message on trivially light workloads (research + .md editing)

Problem

Session limit is being consumed at ~2-3% per message during extremely light usage — simple Q&A and editing .md files. No code generation, no large file reads, no complex tool chains.

Observed behavior:

• Started a fresh session
• Asked a few conversational questions + minor .md file edits
• Hit 92% → 95% within ~3 messages of light Q&A
• Limit resets on a time window (~7pm America/Costa_Rica)

What we ruled out

We audited the entire local setup to identify possible overhead causes. None of these are the issue:

• Memory files: Minimal. One reference .md file (~83 lines). MEMORY.md index is 3 lines. Not bloated.
• CLAUDE.md: Does not exist — neither globally (~/.claude/CLAUDE.md) nor at the project level.
• MCP servers: None configured.
• Hooks: None configured.
• Project settings: Minimal (settings.json only has DISABLE_AUTOUPDATER=1 and skipAutoPermissionPrompt).
• Auto mode: Was just enabled this session — same behavior occurred the previous day without it.
• Conversation length: Short. A handful of messages.
• File sizes: Working directory contains only .md files, nothing large.

Environment

• Claude Code CLI (latest, auto-updates enabled)
• macOS (Darwin 24.6.0)
• Claude Max subscription
• Shell: zsh
• Project: plain markdown research notes repo (no code)

Expected behavior

Light research conversations and .md edits should consume a small fraction of the session limit. 2-3% per conversational message with zero tool-heavy operations is disproportionate and makes the session limit effectively unusable for sustained work.

Request

Please clarify:

1. What exactly is being measured by "session limit" (tokens? weighted API calls? something else?)
2. Why the per-message cost is so high on trivially light workloads
3. Whether there is overhead being charged that is outside the user's control (e.g. system prompt size, injected reminders, model tier)