[FEATURE] Multi-repo cloud sessions load each repository's CLAUDE.md in full, multiplying standing context

Summary

A cloud session with several repositories attached loads every repository's CLAUDE.md — with its @-imports — whole, into the opening context. A CLAUDE.md is usually not small, so the session's standing context is multiplied by the number of attached repositories, and every turn pays for it, whether or not the work touches those repositories.

Observed

With three repositories attached, all three CLAUDE.md files and their imports sat back to back in one system block: about 89 KB, roughly 22–25k tokens, against about 8k for a session with one of them. CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 is set by the platform, so this is the intended behaviour rather than a misconfiguration.

Ask

Load only the first source's CLAUDE.md in full. For each other attached repository, put a short reference in its place — its path and its CLAUDE.md's location — so the agent reads that file when the work moves into that repository, the way a nested CLAUDE.md loads on first read of its directory.