Skip to content

[discuss] : CursorTab as the input method for the Agent era? #90

Open

Description

@jensenojs

Nowadays, users often need to write more than just code. Frequently, they are organizing the next prompt, or preparing to respond to an agent's deviations, supplement edge cases, tighten the scope, or correct misunderstandings.

This experience could work like an input method: the system combines the current context to provide a candidate expression in advance; the user can then accept, rewrite, or continue typing.

I think this could be a very interesting direction for cursortab

I'd like to hear if you're interested in this direction: could Cursor Tab expand from next-edit suggestions to more general next-intent suggestions in the future?

Since the conversations in codex/opencode are all stored locally, I can use it to conduct an A/B test. When it can handle over 90% of my daily simple interactions with the agent, I think it will be a solid conversational solution. I've noticed you seem to be focusing on building your agent lately, and perhaps this could be achieved through some well-designed hooks and workflows.

Activity

  1. jensenojs commented on Jun 27, 2026

    @jensenojs
    ContributorAuthor

    In fact, my refactoring aims to achieve this goal under simpler conditions. Both the Lua and Go components of the project hold significant value in their own right. In particular, by appropriately exposing the Go portion (which may require some more detailed design), the functionality discussed in this issue could be made fully available for use in your smelt project.

    I really hope you'll be interested in this; otherwise, I'll just have to work on the design in my fork.

  2. jensenojs commented on Jun 30, 2026

    @jensenojs
    ContributorAuthor

    I have conducted some brief competitive research (most implementations seem quite unrefined), but I may not have the full energy to dedicate to this project over the next month or two. I will sync here from time to time when there is milestone progress in the design; if you allow, please do not close this issue.

    The simple ultimate expectation is that the match rate for reply prompts under my common workflow can reach a respectable percentage; otherwise, this feature would be meaningless. Some prompts might not be predictable, as I myself might need to think for a considerable amount of time to know how to respond.


    • Imagine a scenario: When we open a diffview in nvim (and select a specific block of content), can it automatically predict—collecting relevant context when necessary—what the primary issues or points of dissatisfaction are regarding this piece of code?
      • The quality of the prediction might depend on the context I choose to feed into the workflow; perhaps there are times when the agent needs to execute a simple grep itself to gather additional context?
    • To realize the above scenario, what hooks should the Lua side provide for writing the necessary workflows?
      • Once the nvim platform is ready, I expect it to seamlessly transition to other platforms (whether it's codex or smelt, as long as the corresponding platform exposes the appropriate hooks)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions