Retry w/ different model #222

Open
opened 2026-02-16 17:25:49 -05:00 by yindo · 2 comments
Owner

Originally created by @aarvinroshin on GitHub (Jun 22, 2025).

Originally assigned to: @adamdotdevin on GitHub.

Ability to retry a prompt with the same or different model would be nice.

Originally created by @aarvinroshin on GitHub (Jun 22, 2025). Originally assigned to: @adamdotdevin on GitHub. Ability to retry a prompt with the same or different model would be nice.
Author
Owner

@khalido commented on GitHub (Jun 23, 2025):

Seconding this request - a number of chat tools I've used have an option to go to a user input, edit it, change the model and resend.

There are two options here:

  1. Easy - Edit and rerun the prompt at the current state of edits.
    .
  2. Harder: Rewind back to state before the prompt you are editing run, then run the edited prompt (with same/diff model).

This is linked to #338 - ability to rewind. Really, its basically the same request, as to edit a prompt we:

  • first rewind to the last good one
  • then enter a new prompt by
    • replacing the one we want to edit altogether
    • OR start with that and edit it (first nuking any changes made by the prompt we are editing)

Git makes it easy to roll back changes, but the way all these agentic tools are currently setup is to make a bunch of changes then commit to git.

It is a interesting issue how best to track these changes - commit after every prompt - which makes it easy but makes the git log very messy and likely unusable for humans or projects. We want to see commits after successful work done, which might invole many prompts and human edits.

Gemini proposes using git patches:

  • Capture Changes as Patches: After each agent action, run $git diff$ to capture the file changes and save them as a unique patch file.

  • Create a Session History: This sequence of saved patches acts as a temporary, step-by-step history for the session, completely separate from the Git log.

  • Rewind by Reversing Patches: To "undo" or "rewind," apply the saved patches in reverse to revert the files in your working directory to a previous state.

  • Commit Manually: The user creates a single, clean $git commit$ only when the work is truly done, bundling all the desired changes.

@khalido commented on GitHub (Jun 23, 2025): Seconding this request - a number of chat tools I've used have an option to go to a user input, edit it, change the model and resend. There are two options here: 1. Easy - Edit and rerun the prompt at the current state of edits. . 2. Harder: Rewind back to state before the prompt you are editing run, then run the edited prompt (with same/diff model). This is linked to #338 - ability to rewind. Really, its basically the same request, as to edit a prompt we: - first rewind to the last good one - then enter a new prompt by - replacing the one we want to edit altogether - OR start with that and edit it (first nuking any changes made by the prompt we are editing) Git makes it easy to roll back changes, but the way all these agentic tools are currently setup is to make a bunch of changes then commit to git. It is a interesting issue how best to track these changes - commit after every prompt - which makes it easy but makes the git log very messy and likely unusable for humans or projects. We want to see commits after successful work done, which might invole many prompts and human edits. ### Gemini proposes using git patches: * **Capture Changes as Patches:** After each agent action, run `$git diff$` to capture the file changes and save them as a unique patch file. * **Create a Session History:** This sequence of saved patches acts as a temporary, step-by-step history for the session, completely separate from the Git log. * **Rewind by Reversing Patches:** To "undo" or "rewind," apply the saved patches in reverse to revert the files in your working directory to a previous state. * **Commit Manually:** The user creates a single, clean `$git commit$` only when the work is truly done, bundling all the desired changes.
Author
Owner

@kierr commented on GitHub (Oct 8, 2025):

You can make model specific subagents, junior dev vs senior dev, and provide prompting/instruction when to hand over between them.

@kierr commented on GitHub (Oct 8, 2025): You can make model specific subagents, junior dev vs senior dev, and provide prompting/instruction when to hand over between them.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: anomalyco/opencode#222