Link to a specific response: jump to it, and hand it to an agent as context

[joshka]

Problem

I often want to point at one specific response in a thread: "use the plan from this response", or "I said this in ". Today I can quote words or link the whole thread, but I can't point at a single turn. Quoting copies text without saying where it came from. A thread link loses the position.

There are four uses:

1. Context for the agent. Pass a response by reference rather than by its words, so the agent reads that exact response instead of guessing which part of the thread I mean.
2. Navigation. Copy a link to a response, then open it later (or from notes, another thread, or a PR) and land on that message.
3. Talking about the conversation. Point back at where something was said: "I said this in ", "you agreed to this in ", or a link in a bug report about how an agent behaved. Here the link records where and when something was said, not just its content.
4. Other tools. Notes apps, task lists, issue trackers, PRs, scripts and other agents can link straight into a response in T3 Code. Tools can also take a link as the address of something to read: give a link to another agent, an MCP client or a CLI, and it fetches that message. The link becomes a handle that tools understand, not only something a person clicks.

Proposal

Every message gets a stable link that the UI can copy. Opening it scrolls to the message, and an agent can resolve it to that message's content.

• Copy: a "Copy link" action on each message, on web, desktop and mobile.
• Open: scroll to the message, loading earlier history and unfolding it if needed.
• In chat: a message form of the existing in-chat thread link format, so an agent can write a link to a message and clients render it as clickable, the way thread links work now.
• For tools: the link is a documented, stable format that a tool can parse into environment, thread and message, and T3's MCP read tools accept a link as well as raw IDs. A t3 CLI command to read a link would follow the same path, but the CLI has no thread commands today.
• For agents: a link pasted into the composer resolves to the message, and t3_thread_read can read the message the link names, including who wrote it, when, and where it sits in the thread (use 3).

Existing work

• #19203 (open) adds ?message=<messageId> to the app links from #18454. It covers copying and opening on web, desktop and mobile, so most of use 2. It's waiting on maintainer review, and this discussion is partly meant to settle its direction.
• What's left is the agent side. t3_thread_read reads one item by itemId (a turn item ID), while feat: link to a message in a thread from anywhere #19203 links by message ID. The in-chat thread links agents write can't target a message yet.

Related: #16153 (pin responses and jump between them), #19969 (thread links that resolve on another machine; message links would have the same problem), #19531 (copy a desktop deep link to a thread).

Questions for maintainers

• Is "every message has a copyable, stable link" the direction you want, with feat: link to a message in a thread from anywhere #19203 as the first part?
• Should agent-written message links extend the in-chat thread link with a message part, or should agents use the same https:// app link that users copy?

🤖 Claude Opus 5.5

[maxim-kich]

I was about to start a discussion myself and found yours, which covers a related use case.

One question about “Copy link”: would these links work only within the user’s connected environments, or could another person open them too? I tend to associate that action with sharing, so the distinction might need to be clear.

For referring back to a response within the conversation, I explored a shortcut using the existing Cite feature:

Problem or use case

I often go back and forth with models—exploring approaches, generating images, writing, and especially researching. Conversations get long, and I regularly need to point the model back to something it said earlier.

Citing a whole response currently means selecting all its text, clicking Cite, and dismissing the optional comment popup before writing my prompt. I often copy and paste the response into the composer instead, which clutters the draft.

Proposed solution

Add a quote icon beside Fork on assistant responses, with the tooltip “Cite this message”.

Clicking it would insert a citation for the full response at the current cursor position, reusing the existing citation chip. It wouldn’t open the optional comment popup, since I can write my follow-up directly in the composer.

This would save a few steps and make Cite easier to discover and use.

As a small related improvement, hovering over a citation chip could highlight the original message or selected text. This would make its source easier to spot, especially for people like me who find it hard to focus when there’s a lot of text on screen.

Demo

I prepared a local prototype showing the interaction:

t3-code-cite-message-demo.mp4

What do you think about adding a shortcut to the existing Cite feature for this use case?

To maintainers

I already have a working local prototype and would like to contribute the implementation myself. If maintainers approve the direction and scope, I’m ready to prepare a focused PR. I’m also happy to open a separate Ideas discussion if needed.

[joshka]

Add a quote icon beside Fork on assistant responses, with the tooltip “Cite this message”.

I think this is a great idea, but it's a different request than this. I would want this in addition to what I'm asking for here. How about creating a new discussion on this one - I'll slap my +1 on it.

Basically copy text, copy link, cite response, are all erasonable items to want. Often I'm using multiple threads and want to carry state between them - the turn id is the perfect value to be able to past for this. e.g. "read thread:123:response:xyz, and then ..."

[joshka]

One question about “Copy link”: would these links work only within the user’s connected environments, or could another person open them too? I tend to associate that action with sharing, so the distinction might need to be clear.

On this - this should likely be a local non shared link for now - shared links should be different to this request and bring all sorts of things that need to be solved.

[maxim-kich]

Thank you!

I created new discussion #21157
Let's see!

And sorry now I get it. Of course in case of mentioning/linking/citing a message in another thread it's different.

I only asked because of terminology, it may be confusing, and not fullfill expectations since users got used to links as links. Maybe just another name?

You know what theoretically could work and be easy to implement instead of additional button. Just an idea:

• user clicks on copy icon under message
• goes in another thread pasts it in chatbox
• gets chip that is link to a message in another session

Like this:

t3-code-copy-message-sent-chip.mp4

The only one point that it doesn't solve is "Other tools."
But since it meant to be internal linking, not for share, it may be fine.

What do you think?

[joshka]

And sorry now I get it.

No probs :)

I only asked because of terminology, it may be confusing, and not fullfill expectations since users got used to links as links. Maybe just another name?

Links are weird - I'm using he app, not the web version, so links are t3-code:// etc. These sort of things have been around since windows 95 days IIRC. The thing about them is that they become clickable in the rollout, transportable to other applications on your machine to be used to jump into t3code app directly etc. Shared links bring something else - they mean that you have to actually setup sharing of the session data somehow, which is quite a bit of a different.

You know what theoretically could work and be easy to implement instead of additional button. Just an idea:

I want both links and content to be able to be copied. I want that deterministic base on what I click, not based on where I pasted. Some use cases need one, others another. A link lets the model choose which elements of the session log it re-reads to understand, it may not read all, it may read things around it (the user prompt before it, the command output etc.). The content doesn't have this same effect - it's serialized to text once. That's useful when pasting something elsewhere (agent, email, webform) rather than asking a model to read a session file etc.

I also think we want bookmarks on responses etc. Think "this is the really important model output I care about that contains some reasoning for ..." where you treat the model response as useful information (for whatever reason - maybe it's a bug, maybe it's a really obvious thing that needs to be fixed later, ...)

[maxim-kich]

Links are weird - I'm using he app, not the web version, so links are t3-code:// etc.

I meant UX writing not actual implementation. Of course I didn't mean real url ;)

I want both links and content to be able to be copied.

Yes, you're right. It may make copy and paste as content to edit it or whatever much complicated.

I like the idea of Bookmarks! I found related discussion but about pinning messages, which also solves JTBD #16153