Making generated files consumable by the Code Interpreter tool without round-tripping through model context

[Aluriak]

This follows up on #14366, which proposes generating files (e.g. from an MCP server's output) and attaching them to the conversation via a LibreChat file_id, so they show up in the chat UI like any uploaded/generated file. That proposal is already on a good track and solves the "produce the file" half of the problem.

This discussion is about the other half: making that file usable as an input to the Code Interpreter tool for further analysis, without its content ever passing through the model's context window. (that is : file being available to the code executor to run commands on it)

Context: as far as I can tell, today any data the Code Interpreter sandbox works with has to be loaded into context first (the model reads/writes the content as text/tokens). That's fine for small snippets, but it breaks down for real data analysis (large CSVs, generated datasets, intermediate artifacts from a multi-step agent pipeline, …). The cost and the context-window ceiling make it impractical past a certain size, even though the sandbox itself could perfectly well read the file directly from disk/storage.

What I'd like to figure out, concretely:

• Does the Code Interpreter tool currently accept a file_id as an input reference it resolves itself (e.g. by fetching from the configured file storage (S3, local, …)), or does every file necessarily get inlined into the request first?
• If it's the latter today, is there an existing extension point (tool input schema, file resolution step before the sandbox call) where a file_id could be passed through and fetched directly by the sandbox runner, bypassing context entirely?

Happy to help test/implement on our end ; we run our own Code Interpreter deployment (codeapi/sandbox-runner) and have a direct interest in this working end-to-end for agent-driven data analysis rather than just chat-attached files.