Claude Code ignores loaded CLAUDE.md content when context is under 1/3 capacity

Resolved 💬 3 comments Opened Apr 24, 2026 by asalcedo29 Closed Apr 28, 2026

Bug: Claude Code ignores loaded CLAUDE.md content when context is under 1/3 capacity

Repo: https://github.com/anthropics/claude-code/issues

Description

Claude Code does not reliably reference CLAUDE.md content when answering questions, even when the context window is well under capacity (observed at 33% usage). Instead of consulting the documentation already loaded in context, it reasons from scratch and generates incorrect conclusions.

Observed Behavior

During a session, CLAUDE.md contained detailed documentation about MCP token flows, gateway URLs, auth chains, and file paths. At 33% context usage (no compaction), Claude Code:

  1. Claimed "we don't have access to the MCP server" — it was configured in appsettings.json, documented in CLAUDE.md
  2. Claimed "MCP needs a different auth flow — this is a real piece of work" — the auth was already implemented in dv-auth.ts, documented in CLAUDE.md
  3. Wrote hardcoded workaround code instead of using existing infrastructure
  4. Required multiple rounds of user correction before consulting the documentation that was already in its context

This wasted ~10 conversation turns before the user pointed Claude to information it already had loaded.

Expected Behavior

Claude Code should reference CLAUDE.md content that is loaded in its context window before generating answers about repository infrastructure. When the context is under 1/3 capacity, there is no compaction or decay excuse — the content is there.

Impact

  • Incorrect claims presented with confidence
  • Wasted user time correcting wrong assumptions
  • Unnecessary code written that had to be replaced
  • User loses trust in whether Claude Code is working from project context or making things up

Environment

  • Model: Opus 4.6 (1M context)
  • Context usage at time of failure: ~33%
  • CLAUDE.md size: ~400 lines with detailed infrastructure docs

Notes

This appears to be a behavioral pattern — defaulting to generating reasoning rather than looking up answers from loaded documentation. It is not a context window or compaction issue.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗