Before submitting
Area
apps/web
Problem or use case
Reviewing an agent's output often needs more precise feedback than a single follow-up message. Today, if I want to comment on a sentence in an assistant response or a specific line in a rendered Markdown/text file such as a plan, I have to quote or copy the relevant text into the composer and explain where it came from.
That becomes cumbersome for long responses and plans, especially when several separate passages need feedback. It is also easy to lose the association between a comment and its original location.
T3 Code already supports the analogous workflow for code diffs: select lines, attach a review comment, and send those comments to the agent as structured context. It would be useful to extend that review model to assistant responses and rendered files.
Proposed solution
Allow users to select a line or text range in:
- an assistant response; and
- a rendered Markdown/text file, including saved or proposed plan files.
The selection should expose an Add comment or Annotate action. A comment should remain visibly anchored to its selected passage while it is pending. Multiple annotations could be collected, edited, or removed, then submitted together with the next message.
When sent, T3 Code should include structured context that identifies the source and selection, rather than relying only on copied prose. For example:
- assistant message/turn identity plus the selected text;
- file path plus line or text range; and
- the user's comment.
The timeline should render this as a compact review-context card instead of exposing raw serialization.
This is similar to the annotation/review interaction in the ChatGPT desktop app, where feedback can be attached directly to selected content instead of manually quoting every passage.
Why this matters
This would make plan review and iterative agent collaboration much faster and less ambiguous. Users could give several precise pieces of feedback in one pass, and the agent would receive the exact source context for each comment.
It also creates a consistent review interaction across diffs, plans/files, and agent output instead of limiting line-level feedback to code changes.
Smallest useful scope
A first pass could support the web/desktop thread UI for:
- selecting text in assistant responses and rendered Markdown files;
- adding one or more pending comments;
- showing removable annotation chips/cards in the composer; and
- serializing the source selection and comment into the next prompt.
It does not need collaborative comments, persistent discussion threads, resolution state, or file editing. Reusing the existing diff-comment draft and structured-context patterns would be sufficient.
Alternatives considered
- Manually quote each passage in a follow-up message. This is slow and loses stable source identity.
- Copy the full response or plan and edit it elsewhere. This interrupts the review flow and adds unnecessary context.
- Use existing diff comments. They provide the desired interaction, but only for changed-code lines, not assistant messages or plan/file content.
Risks or tradeoffs
- Text-range anchors can become stale if a file changes after the annotation is created; file comments may need a content snapshot and line range.
- Assistant responses contain Markdown, code blocks, links, and expandable tool content, so selection should not break existing interactions.
- Large selections or many comments could add excessive prompt context, so the UI may need reasonable limits.
- Shared serialization would be preferable to separate incompatible formats for diff, file, and response comments.
Examples or references
Contribution
Before submitting
Area
apps/web
Problem or use case
Reviewing an agent's output often needs more precise feedback than a single follow-up message. Today, if I want to comment on a sentence in an assistant response or a specific line in a rendered Markdown/text file such as a plan, I have to quote or copy the relevant text into the composer and explain where it came from.
That becomes cumbersome for long responses and plans, especially when several separate passages need feedback. It is also easy to lose the association between a comment and its original location.
T3 Code already supports the analogous workflow for code diffs: select lines, attach a review comment, and send those comments to the agent as structured context. It would be useful to extend that review model to assistant responses and rendered files.
Proposed solution
Allow users to select a line or text range in:
The selection should expose an Add comment or Annotate action. A comment should remain visibly anchored to its selected passage while it is pending. Multiple annotations could be collected, edited, or removed, then submitted together with the next message.
When sent, T3 Code should include structured context that identifies the source and selection, rather than relying only on copied prose. For example:
The timeline should render this as a compact review-context card instead of exposing raw serialization.
This is similar to the annotation/review interaction in the ChatGPT desktop app, where feedback can be attached directly to selected content instead of manually quoting every passage.
Why this matters
This would make plan review and iterative agent collaboration much faster and less ambiguous. Users could give several precise pieces of feedback in one pass, and the agent would receive the exact source context for each comment.
It also creates a consistent review interaction across diffs, plans/files, and agent output instead of limiting line-level feedback to code changes.
Smallest useful scope
A first pass could support the web/desktop thread UI for:
It does not need collaborative comments, persistent discussion threads, resolution state, or file editing. Reusing the existing diff-comment draft and structured-context patterns would be sufficient.
Alternatives considered
Risks or tradeoffs
Examples or references
Contribution