My take on billion-context after a longer real session with Claude Code:

• ✅ It stayed out of the way. 61 requests and 22 test runs went through bili, and the coding work was fine.
• 🗜️ It compressed nothing. The export says blocks=0. Claude Code's own setup on my box already used about 375,000 tokens, bili only starts compressing after about 50,000 more, and my task added only about 35,000.
• 🔒 Claude did try to compress twice, but bili's tool needs permission, and I hadn't granted it.
• ⚠️ It didn't rescue the chat when it got too big. On resume Claude Code switched to a 200,000-token model, and bili refused rather than making it fit. Asking for the 1M model fixed it.
• 🧪 The big promise is still untested. To see compression, you'd need a much longer task, or a lower threshold, plus the tool permission.

"bili tracked a long session without problems, but it never actually compressed anything, so I still can't say whether it delivers endless context."