[Feature]: Reference an existing thread as context in another conversation

[Aneaire]

Before submitting

• I searched existing issues and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/desktop

Problem or use case

Context

I use T3 Code to run agents that sometimes execute with custom skills or configuration. When an agent does something unexpected — for example, a custom skill steers the AI the wrong way during a test run — I want to have another AI review that conversation to figure out what went wrong and propose a fix to the skill.

The problem

Today there is no way to point one conversation at another:

• I have to manually copy and paste the transcript into a new thread.
• Copy/paste drops tool calls, reasoning output, diffs, and other structured context.
• Large threads are impractical to reproduce manually, and retyping loses fidelity.

What I'm trying to do

• Run an analysis thread that references an earlier conversation.

• Ask the new agent questions about that conversation, such as:

  "Analyze the transcript of thread X and tell me why my skill misbehaved, then propose a fix to the skill file."

• Keep the referenced thread read-only; the analysis happens in the new conversation.

Why it's hard today

There is no built-in way to attach a thread as context, so this workflow is slow, lossy, and error-prone.

Proposed solution

Add a way to reference an existing thread from within another conversation:

• "Reference thread" action on a thread (sidebar / thread menu) that lets you pick it as context, or an @reference-style mention in the composer that resolves to a saved thread.
• When a thread is referenced, the new conversation receives that thread's transcript as read-only context (user + assistant messages, and where available tool calls and reasoning).
• The referenced thread stays untouched — analysis and follow-up happen only in the referencing conversation.
• The UI shows a clear "Referenced: " chip in the composer so you can see what context will be sent before you send.

A pragmatic v1: attach the transcript as structured context on thread.turn.start (extending ChatAttachment or adding a referencedThread field), letting the provider see it as additional context for that turn.

Why this matters

• Lets a different model/agent review what happened in a past run — e.g. diagnosing why a custom skill misbehaved and proposing a fix to the skill — without manual copy/paste.
• Copy/paste today loses tool calls, reasoning, diffs, and timestamps, so the reviewing AI works from incomplete information.
• Unlocks retrospective workflows: audit a conversation, review agent behavior after the fact, and iterate on skills/prompts based on full evidence rather than a pasted summary.
• The workflow is impossible today, so this is net-new capability rather than a polish improvement.

Smallest useful scope

A first pass only needs to:

• Allow referencing one existing thread from a new conversation.
• Inject that thread's transcript (user + assistant text) into the referencing conversation's first turn as context.
• Render a small "Referenced thread" indicator so the user knows context was attached.

No cross-thread linking UI, no live updates from the referenced thread, no graph/visualization, and no multi-thread references.

Alternatives considered

No response

Risks or tradeoffs

No response

Examples or references

No response

Contribution

• I would be open to helping implement this.

[lastobelus]

chatgpt 5.6 has no problem when I say things like

• "what is 8d05a5ee-6eec-47b5-9db6-df9544c3dea0 up to?"
• "analyse t3 session 8d05a5ee-6eec-47b5-9db6-df9544c3dea0 for churn"
• "which thread did I do XXXXX in?"
• "spawn a subagent to figure out what 8d05a5ee-6eec-47b5-9db6-df9544c3dea0 had to do to do XXXXX and summarize it for you"

whether I'm running the thread in t3code or codex.

I remember I had to help it a bit a few months ago with those kinds of requests, but I don't anymore. I don't know what changed (it's not codex memories, I just enabled them recently, and there's nothing about finding t3code threads in my global AGENTS.md) but it Just Works.

Claude is less adept at it out of the box, and tends to just make something up instead of admitting it doesn't know how to do it. Here's an example of how to prompt it:

Look up T3 Code thread `<T3CODE_THREAD_ID>` in T3’s local conversation database at `~/.t3/userdata/state.sqlite`.

Read the thread’s messages from `projection_thread_messages`, including archived threads, in chronological order. Search for discussion of `<XXXX>`.

Tell me:

1. What we ultimately decided.
2. The reasoning and constraints behind it.
3. Any unresolved questions or later reversals.
4. The most relevant message dates or short excerpts supporting your answer.

Distinguish an accepted decision from suggestions, plans, and alternatives. If the thread does not establish a clear decision, say so rather than inferring one. Do not modify the database.

I don't use Claude a lot; I'm sure if you did a few requests like that it'd store some memories and you'd be able to prompt it as tersely a codex.

[lastobelus]

by the way, when you're analyzing big threads it's best to tell your main thread to use subagents to do it in chunks and summarize for the main thread, or you'll use a lot of quota.

[juliusmarminge]

Closing as a duplicate of #6701, which already tracks referencing another thread so an agent can use its conversation as context.