Support complete large files as context in Agent panel (manual override the 16KB limit)

[vlebert]

Summary

Adding files as context in not effective for models without tools support.

Description

Since the latest update, when adding a file to the AI agent's context, the system now includes only the list of functions/symbols found in the file rather than the full file content. While this works well for models supporting tool calls (which can fetch the full file as needed), it creates limitations for models that do not support tool calling.

Actual Behavior:

• Files added to the AI agent's context are summarized as symbols/functions.
• Models without tool calling support cannot retrieve the full file content, leading to incomplete context.

You can check the actual context by clicking "open thread as markdown".

Expected Behavior:

1. Option 1: For models without tool calling support, default to including the full file content in the context (similar to the Text Thread functionality).
2. Option 2: Provide a user-friendly way to explicitly control what is included in the context (e.g., toggle between "symbols only" and "full file").

Zed Version and System Specs

Zed: v0.185.13 (Zed)
OS: macOS 15.4.1
Memory: 32 GiB
Architecture: aarch64

[notpeter]

Agent mode is built to allow the model to act as an agent -- by calling tools. I recommend using text threads if your models do not support tool calling.

[vlebert]

Hmm I think even in Agent mode, you should be able in some case to include large files without alteration. This could lead to misunderstanding.

[notpeter]

I see. Currently large files cannot be added in their entirety to the agent panel view. They are capped at 16K tokens, with the expectation that the agent can call tools to get the portions of the file desired. We made this change in response to real-world evals where the model would often get confused by very large unannotated context (e.g. zed's editor.rs is 157k tokens).

Currently the only way to explicitly include unabridged context is via text threads.
I will reopen this as a feature request to support the same (perhaps via an option) in Agent mode too.