[Feature Request] Larger context window (2M–2.5M tokens) for long-lived codebases

Summary

Request for a context window larger than 1M tokens (ideally 2M–2.5M) in Claude Code, at least as an opt-in option for heavy users.

Problem

1M tokens works well for most projects, but for a production app that has been developed with Claude for nearly 2 years, it is no longer enough to hold the context a session needs.

In practice this means:

• Constant compaction. Long sessions hit the limit and auto-compact, losing details that were important earlier in the session.
• Re-explaining context. After compaction, I often have to re-establish architecture decisions, conventions, and recent changes.
• Workarounds have a ceiling. I already keep a curated CLAUDE.md with a reading guide, use subagents to keep the main context lean, and maintain an external memory system. These help a lot, but at this project size the 1M window is still the bottleneck.

Model quality is no longer the main limitation for this kind of work. Context size is.

Proposed solution

• An extended context option (2M–2.5M tokens) for Opus/Fable-class models in Claude Code, even if it's opt-in, beta, or limited to certain plans.
• Alternatively, if a larger window isn't feasible soon: smarter compaction options (e.g. pinning files or messages that must survive compaction).

Environment

• Platform: Linux
• Terminal: Warp
• Claude Code version: 2.1.283
• Feedback ID: b26080cb-eb70-4071-88ff-a426b5427005

Thanks for considering it!

Errors

[]