Repository navigation
Replies: 1 comment 2 replies
|
Hey @Domeless! Nothing like this is planned or in progress for v4, and no open issue covers it. #8371 was closed by flattening the history for gemini-web (#8607), not by continuing the remote chat. The closest thing today is per adapter. Claude Web keeps I'm open to exploring it. The order I'd take:
One thing I couldn't confirm. I didn't find the 24,000-character cap in our ChatGPT Web executor. Where did you see it? If you want to work on this, open a design issue for the binding shape first, so we settle it before any PR. |
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi! Are there plans to take web-provider conversation support further by adding shared, persistent storage for remote chat state?
The shared session reuse discussed in #3142 and the client-to-upstream conversation mapping proposed in #8371 seem like a good foundation for this.
Some adapters already handle continuation locally. For example, Claude Web and Copilot M365 keep conversation state in memory. A possible next step would be shared, persistent storage that adapters can use to find and continue the right remote chat, even after a gateway restart.
The basic flow would be:
The history hash would check that the request continues the same client conversation on the same account. Separate chats would stay separate even if their messages match. Each adapter would still handle its provider’s chat protocol.
Potential benefits include:
22,656 cached_tokensout of22,929 prompt_tokens. This suggests a possible benefit, though it does not yet show how much quota would be saved.Is anything like this planned or already in progress for v4? Would it be worth exploring this direction further, or is there existing work that already covers it?
All reactions