All experiments

LAB 005 / Conversation models

Threads & Subthreads

Branch from an AI response and keep the parent, its subthreads, and their origins navigable.

Conversations are linear. Ideas are not. One answer can open three directions at once, and a transcript makes you follow one while the others sink out of view. This lab looks at how a response could become a place to pursue one of those directions, and how to stay oriented when you do.

The conversation below is seeded from my own research workflow. It is a scenario, not a product plan.

Try the prototype

Branch from any assistant response, then choose one of its avenues. Those avenue labels are written into this seeded demo. No model found or summarized them. The new subthread opens with its origin and a preview of inherited context. Maximize opens an in-page fullscreen mode. Escape closes it.

Pilipinas Interface architecture

Pilipinas Interface architecture

Parent · L0

Original transcript · branches stay separate

6 messages
  1. You · 09:14

    I'm shaping the architecture for Pilipinas Interface, a public knowledge surface for Philippine civic and cultural data. It has to feel calm and quick on a phone outside a Metro Manila café and on a laptop. What should the architecture look like?

  2. Assistant · 09:14

    Start with a boring server and a thin client. The interesting decisions are not about frameworks; they are about where the edges sit.

    First edge: ingestion. Source material arrives as agency PDFs, uneven CSVs, and pages that change without announcement. Normalize that into a versioned content layer before the interface ever sees it. Second edge: retrieval. Decide what the interface can answer directly and what it should honestly hand off to a source. A confident half-answer costs more trust than a clean link out.

    The direction I'd start with is language. Tagalog and English share a sentence here, often in the same breath. If the interface treats that as an edge case, every later feature inherits the assumption.

    The avenues I'd pull apart: Data ingestion pipeline, Retrieval boundary, Tagalog-first typography, Offline-first delivery.

  3. You · 09:18

    The language question is the one I keep circling. People code-switch mid-sentence. What does that mean for the interface, not just the copy?

  4. Assistant · 09:18

    It means code-switching is an input model, not a translation task. Search has to accept a query that changes language halfway through, and results have to stay legible when a heading is in Tagalog and a metadata line is in English.

    There is a quieter cost too. Font fallback and line length decide whether that mixed sentence breaks cleanly or leaves a word stranded at the end of a line. On a narrow screen, the typography is the interaction.

    The avenues I'd pull apart: Code-switching in search, Font fallback and line length, Editorial voice across languages.

  5. You · 09:23

    Keep going on the search side. Where does the data actually come from, and who keeps it honest?

  6. Assistant · 09:23

    Both questions are provenance questions. Ingestion should record where each field came from, when it was last seen, and who reviewed the change. Without that, 'fresh' is a feeling rather than a claim.

    Normalization is where the politics hide. A barangay named three ways across three agencies is not a data-cleaning problem; it is an editorial decision that needs a paper trail. Build the review queue before the pipeline, or the pipeline will outrun the people maintaining it.

    The avenues I'd pull apart: Source provenance, Normalization rules, Freshness and review queues.

Enter to send · Shift+Enter for a new line · Simulated response

Prototype state lives in this page only. Reloading restores the seeded conversation.

01 / The problem

In a long research conversation, the useful part of an answer is often the second-order question it exposes. Follow that question in the same thread and the conversation starts carrying several topics at once.

Starting a fresh chat fixes the drift but breaks the relationship to the discussion that produced the question. I end up restating the premise in every new conversation.

02 / The observation

Looking back at my transcripts, the branches are already there. They are flattened: a sentence quoted into another chat, a link in a note, a question I meant to return to.

What is missing is a visible link from the response to the conversation that grew out of it. That is an observation about my workflow, not a claim about how everyone works.

03 / Existing approaches

Branching already ships in several tools. The differences are in what you can branch from, what the new thread receives, and how you find your way back.

Existing conversation branching patterns and what this lab investigates in relation to each
Existing patternWhat it does wellWhat this lab investigates
ChatGPT Branch in new chatBranches from a message into a new chat and keeps the original thread intact.Whether the branch point can be a specific idea inside the response, and how the new chat stays grouped with its parent.
Codex fork and resumeForks a saved session or the previous user message, preserves the source transcript, and can isolate work in a Git worktree.A visible parent-child relationship in the conversation list, beyond a picker of separate sessions.
Claude Code /branch and --fork-sessionCopies history up to a point, leaves the original in the session picker, and returns through resume.Orientation while working inside the branch, and where that branch sits relative to its origin.
Conduit fork on editEditing an earlier message forks a new conversation before that message and keeps the original.Choosing a passage or topic explicitly instead of editing a whole message to fork.
Chatvas canvas nodesShows each ChatGPT branch as a connected node on an infinite canvas.A lighter relationship view that stays inside the conversation workspace.

These are shipped features described in their own documentation. Where the docs do not say how branches are grouped or how a user returns to the source, I have left that out instead of assuming it.

04 / The idea

Branch from a passage, phrase, or topic inside an answer. Continue it as its own thread. Leave the original conversation untouched, and keep the link back to where it started.

Selective branching
The branch point should be the idea worth pursuing, not the whole message. The current build stops one step short. It branches from a response and offers a short list of named avenues drawn from that response. Selecting a phrase is the direction, not a working interaction here.
Parent-child organization
A flat list becomes hard to scan once there are many branches. The sidebar keeps one parent row with its own navigation target and a separate disclosure for its subthreads. Branches nest directly underneath and carry their own color.
Conversational orientation
A breadcrumb names the parent conversation and the active subthread. The origin banner and the Branches panel show where the thread began and offer a way back.

05 / What this explores differently

None of these pieces is new on its own. Message branching, session forks, resumes, and canvas views already exist. The experiment is about their composition: a selective branch point, a parent-child sidebar, and a breadcrumb that keeps the relationship visible while you work inside the branch.

Each product in the comparison solves part of this. I am interested in what changes when the three are treated as one interaction rather than separate features.

06 / Limitations and unresolved questions

  • The prototype branches from a whole response. It offers named avenues, but there is no branching from selected text or a phrase.
  • Inherited context is fixed at one preceding turn plus the anchor response. There is no summarization, token budget, or sync if the parent changes.
  • Only one level of branching is modeled. Nested branches and larger trees are not represented.
  • Grouped sidebar rows may not scale. I do not know how they behave after dozens of branches.
  • Names come from preset avenues. Naming, search, and reorganizing branches are open questions.
  • Breadcrumbs help in a short tree. Long conversations and branches that drift from their origin are harder to read.
  • The Review, Terminal, and Browser tabs are placeholders. Only Branches is functional. Replies are canned and state lives in the page.

07 / Reflection

Does branching let me pursue an idea without disturbing the conversation it came from? Does grouping subthreads under a parent make them easier to find, or does it only move the clutter?

I still do not know whether selecting a phrase is worth the extra interaction. That is the question I want the next version to answer. The prototype is a way to ask it, not evidence that the answer is yes.

Footnotes

  1. Pilipinas Interface is a seeded scenario. The architecture discussion is written for this prototype and is not a published project plan.
  2. OpenAI, ChatGPT release notes, 4 September 2025, "Branch conversations on web": help.openai.com.
  3. OpenAI Codex developer commands for fork, /fork, and resume: developers.openai.com.
  4. Claude Code sessions and checkpointing. /rewind restores a checkpoint, while /branch and --fork-session create separate sessions: sessions and checkpointing.
  5. Conduit desktop documents fork on editing an earlier message and fork from a reply version: features and docs. Several unrelated projects share the name Conduit.
  6. Chatvas is a small open-source desktop canvas that connects each ChatGPT branch to its parent node: github.com/kaleab-ayenew/chatvas.

If you enjoyed reading this or found it insightful, feel free to share it with others or send feedback. It helps me keep on researching and writing.

Say hi

Published on Nov. 22, 2026