Rutger CappendijkandBen Brandt
52182e43cd
Summary This PR adds the ability to quote terminal selections into Agent threads, similar to how editor selections can be quoted. Users can now select text in the terminal and add it as context to their agent conversation. ### Features - **Context menu**: Added "Add to Agent Thread" option to the terminal's right-click menu (only shows when text is selected) - **Keyboard shortcut**: The existing `cmd->` / `ctrl->` shortcut now works in the terminal - **Collapsed display**: Terminal output appears as a collapsed "Terminal" crease in the chat, just like code references - **Works with both thread types**: Supports both ACP-style threads (Claude Code) and text threads ### Implementation - Extended `quote_selection` handler to check for terminal panel focus after checking for editor selections - Terminal content is formatted as ` ```terminal` code blocks via new `insert_terminal_crease` method - Refactored crease insertion into `insert_crease_impl` helper to support both code snippets (with `TextSnippet` icon) and terminal output (with `Terminal` icon) - Added automatic detection and folding of ` ```terminal` blocks in `set_message` so terminal content stays collapsed in sent message history - Terminal creases use a trailing space instead of newline so cursor behavior matches user expectations ### Demo 1. Select text in terminal 2. Right-click → "Add to Agent Thread" (or press `cmd->`) 3. Terminal content appears as collapsed "Terminal" block in chat 4. Send message - AI receives full terminal output, but chat history shows it collapsed ### General note I'm new to open source and the codebase in general so let me know if anything should be different! I tried to stay in line with other patterns I saw. I didn't see a way to setup an end to end test with a real terminal so I just added a unit test for the adding of the text to the chat, if it exists and I missed it then I think the test coverage could be improved with a test that uses a real terminal + running the command <img width="431" height="148" alt="image" src="https://github.com/user-attachments/assets/854732f9-2a7f-4f31-867e-d54af068c97c" /> Release Notes: - agent: Added the ability to send selections from terminals into the Agent thread --------- Co-authored-by: Ben Brandt <benjamin.j.brandt@gmail.com>
Design notes:
This crate is split into two conceptual halves:
- The terminal.rs file and the src/mappings/ folder, these contain the code for interacting with Alacritty and maintaining the pty event loop. Some behavior in this file is constrained by terminal protocols and standards. The Zed init function is also placed here.
- Everything else. These other files integrate the
Terminalstruct created in terminal.rs into the rest of GPUI. The main entry point for GPUI is the terminal_view.rs file and the modal.rs file.
ttys are created externally, and so can fail in unexpected ways. However, GPUI currently does not have an API for models than can fail to instantiate. TerminalBuilder solves this by using Rust's type system to split tty instantiation into a 2 step process: first attempt to create the file handles with TerminalBuilder::new(), check the result, then call TerminalBuilder::subscribe(cx) from within a model context.
The TerminalView struct abstracts over failed and successful terminals, passing focus through to the associated view and allowing clients to build a terminal without worrying about errors.
#Input
There are currently many distinct paths for getting keystrokes to the terminal:
-
Terminal specific characters and bindings. Things like ctrl-a mapping to ASCII control character 1, ANSI escape codes associated with the function keys, etc. These are caught with a raw key-down handler in the element and are processed immediately. This is done with the
try_keystroke()method on Terminal -
GPU Action handlers. GPUI clobbers a few vital keys by adding bindings to them in the global context. These keys are synthesized and then dispatched through the same
try_keystroke()API as the above mappings -
IME text. When the special character mappings fail, we pass the keystroke back to GPUI to hand it to the IME system. This comes back to us in the
View::replace_text_in_range()method, and we then send that to the terminal directly, bypassingtry_keystroke(). -
Pasted text has a separate pathway.
Generally, there's a distinction between 'keystrokes that need to be mapped' and 'strings which need to be written'. I've attempted to unify these under the '.try_keystroke()' API and the .input() API (which try_keystroke uses) so we have consistent input handling across the terminal