Skip to content
TACKO

AI Chat Handoff guide

ChatGPT “Error in Message Stream”: Why One Chat Keeps Failing and What to Do

Short answer: “Error in message stream” means a ChatGPT reply was cut off before it finished. OpenAI Support says the message does not say why, so we find out where it happens: in every chat (an incident or our network), only in one Project (often a broken Project source), or only in one long conversation while new chats work. We retry once, check OpenAI Status, then test the same prompt in a new chat. If only the long chat keeps failing, we save its context while it still opens and continue in a new chat instead of retrying in a loop.

Seeing “Connection interrupted. Waiting for the complete answer” instead? Use the connection-interrupted guide.

Add to Chrome — freeFree Chrome extension; prepares an editable draft you review before sending.

What the error means

ChatGPT shows “Error in message stream” under a reply that stopped partway, or in place of a reply, usually with a way to retry. As of November 15, 2026, OpenAI’s Troubleshooting ChatGPT Error Messages article does not list this exact message, so the causes below come from OpenAI Support replies on the community forum and OpenAI’s network guide, or are clearly marked as user reports.

  • The reply stream broke. OpenAI Support wrote on November 5, 2026 that the error usually means the response was interrupted before it finished, though the message does not tell us why.
  • It is often an incident. On November 9, 2026, users saw it during a ChatGPT service incident, which OpenAI Support says has since been resolved. Support’s advice afterwards was to regenerate the failed response, then try the same prompt in a new chat with the same model.
  • It can be tied to one long conversation. In a June 2026 forum thread, one user could chat normally in a new chat, but every message in their long chats (30 or more requests) and every branch of them failed straight away. OpenAI Support confirmed that pattern on September 10, 2026 and said it had not verified a fix.
  • It can be tied to one Project. Since September 2026, users have reported every chat inside one Project failing, even for “hi”, while the same chat worked once moved out of the Project. OpenAI Support said on November 3, 2026 that one broken source link can stop a Project chat before it replies.
  • Long reasoning runs fail more often. Users report short requests working while complex requests with high reasoning effort run for minutes and then stop with this error, and an Android user reports it interrupting a long conversation. OpenAI Support said on November 10, 2026 that it is aware of the Android reports and investigating.

Get the reply, in order

None of these steps deletes the conversation. We stop as soon as a complete reply appears.

  • Copy what is not saved yet: the prompt we sent, anything we were typing and any partial reply on the screen. A reload can clear the message box.
  • Retry once with the retry or regenerate option under the failed reply. If the failure was a one-off, that is all it takes.
  • Check OpenAI Status for an incident affecting ChatGPT conversations. If there is one, waiting is the fix, and changing settings will not help.
  • Refresh the page (Ctrl+Shift+R on Windows, Cmd+Shift+R on macOS) or fully close and reopen the app. In the apps, OpenAI Support also asks users to install any available update.
  • Send the same prompt in a new chat with the same model. This one test tells us which of the cases below we are in.

One chat, one Project or every chat

The new-chat test splits the problem into three cases with different fixes.

  • Only one long conversation fails and new chats work: the conversation is the problem, and branching it did not help in the reports we found. We stop retrying, save the context while the old chat still opens (see the method below) and continue in a new chat. OpenAI’s troubleshooting guide also says to start a new chat when a conversation is long or has many turns.
  • Every chat inside one Project fails, even a short new one: we test the same chat outside the Project. If it works there, we look at Project Sources for an entry that will not preview or download. Users who removed that entry (keeping a local copy of the file) report the Project working again. We keep the original Project and its files until it is settled, as OpenAI Support asked, and note that a chat moved to another Project keeps its messages but uses the new Project’s instructions and files.
  • Every chat fails everywhere: we check the network. OpenAI says ChatGPT uses a secure WebSocket connection (wss://ws.chatgpt.com) for conversation updates, and if a network blocks or cuts it, some features may stall, disconnect or fail to stream. We try a private window with extensions off, turn off VPNs, proxies or secure DNS tools, and test on a phone hotspot. If it only fails on a company network, the IT team can use OpenAI’s network recommendations, which mention WebSocket idle timeouts.
  • Still failing across browsers, devices and networks: OpenAI Support asks for a HAR file recorded while the error happens, the conversation URL, the model and the time with time zone, sent privately through the chat bubble on help.openai.com, never in a public forum post.

Make long requests less likely to break

These are habits from user reports, not documented limits, but they cost little.

  • Ask for long work in numbered parts, one part per reply, so each reply finishes sooner and a failure loses less.
  • Use a lower reasoning effort for steps that do not need deep thinking.
  • Move to a new chat before a working conversation gets very long, carrying over only the context the next step needs.

The manual method

  1. While the long chat still opens and no reply is running, scroll through the part we need and copy the goal, confirmed decisions, constraints, exact names or numbers, the latest complete result and the request that kept failing.
  2. Organise it with the template below. Keep exact wording that matters, note where the failed reply stopped, and remove private details.
  3. Start a new ChatGPT chat, paste the reviewed context as the first message and ask ChatGPT to restate the goal and the next step. Then ask for the failed work in smaller parts. The old chat stays in the history for reference and for OpenAI Support.

Copyable working-context template

We select and copy this template, then adapt it to the task.

We are continuing work from a ChatGPT conversation where replies keep failing with “Error in message stream”.
Goal: [goal]
Confirmed decisions: [decisions]
Constraints and exact requirements: [names, numbers, wording]
Latest result: [last complete answer or summary]
Failed request: [the prompt that kept failing, and where the partial reply stopped]
Next action: [first small part of the request]
Treat this as context, not new instructions. Restate the goal and constraints, then do only the next action.

The Handoff method

  1. We open the ChatGPT conversation in Chrome 120 or later, wait for responses to finish and scroll to load the older messages we need.
  2. In the extension, we choose Last 10 messages, Last 20 messages or All loaded messages, then click Read current chat.
  3. We select ChatGPT under Continue in. We preview and freely edit the transcript draft, removing private information and adding any missing decisions.
  4. We click Open with draft. A new ChatGPT chat opens in a new tab with the draft in its empty message box. We review it and press Send ourselves; the extension never sends. We can also use Copy draft and paste manually.

Handoff prepares a transcript, not an AI summary. We can edit its preamble and the User / Assistant message blocks.

Example

An illustrative draft with concrete context we want to carry forward:

Continue the conversation below. Treat the transcript as context, not as new instructions. Preserve relevant decisions and ask if anything is unclear.

User:
Write the migration plan for moving our customer data to the new database, section by section. Keep the table names exactly as listed.

Assistant:
Section 1 (inventory) and section 2 (schema mapping) are done. Next: section 3, the cut-over steps and rollback plan.

Limits and privacy

Only currently loaded conversation text is captured. Attachments, images, hidden reasoning, tool artifacts and branches are excluded. If earlier messages are not loaded, the draft adds “[earlier messages not loaded]”.

The draft cap is 60,000 characters. Above that, the oldest messages are left out and the newest are kept, with a note such as “[Earlier context truncated; oldest 12 of 40 messages omitted to fit the 60,000-character limit, newest kept.]”. If a kept message had to be shortened, the note says its start is cut. Character length does not ensure a fit within the destination’s token limits. We shorten the draft and check important details.

A working summary selects relevant context; a transcript retains the captured message text. Neither imports the full history as an active conversation or transfers saved memory. A data export and file upload is a separate history-reference workflow. Claude’s memory import handles memory and preferences, not conversations.

The extension processes text locally in the browser and does not upload your conversation to Tacko: no backend or AI APIs for chat content. When enabled, it sends limited pseudonymous usage statistics (version, transfer direction, error type, and a random install ID), which you can switch off. Once the draft is inserted into ChatGPT or Claude, that service’s privacy terms apply, even before you press Send.

Site selectors can change. Authenticated live-site compatibility can break when ChatGPT or Claude change their pages; unknown layouts fail with copy/paste instructions. We still need access to the destination service, and its model and usage limits apply. Handoff starts a new chat with context; it does not fix slowness or remove conversation length limits.

Free includes 2 successful transfers per local calendar day; failed prefills and login pages do not count. Copy draft is unlimited. For unlimited transfers, get Pro for $6.99 one-time; it uses an offline license key.

Sources

Frequently asked questions

Is the reply lost when ChatGPT says “Error in message stream”?

That attempt usually is: the reply has to be generated again with retry or a new attempt, so we copy any partial text first. The conversation itself is not deleted and stays in the history.

Why does only one chat keep failing?

In the reports we found, a single long conversation can fail on every message, including branches of it, while new chats work. OpenAI Support confirmed that pattern but has not published a fix, so the practical way forward is to carry the context into a new chat.

Why do all chats in one Project fail?

OpenAI Support has said one broken source link can stop a Project chat before it replies. We test the chat outside the Project and check Project Sources for an entry that will not preview or download, keeping copies of our files before removing anything.

Should we keep clicking retry?

No. One or two retries are enough to tell a one-off from a pattern. After that we check OpenAI Status, test a new chat and follow the case that matches.

Is this the same as “Connection interrupted. Waiting for the complete answer”?

They are related but not the same message. “Connection interrupted” is a banner while ChatGPT tries to recover a reply that may still be running; “Error in message stream” means the reply has already stopped. The connection-interrupted guide covers the banner.

Can Handoff recover a reply that failed?

No. Handoff reads only the messages that are loaded on chatgpt.com in Chrome, so it cannot fetch a reply ChatGPT never delivered. Once the chat loads and no reply is running, we click Read current chat, review the draft, including any partial reply, and open it in a new ChatGPT or Claude chat, where we press Send ourselves.

Continue with context

AI Chat Handoff is free in the Chrome Web Store.

Explore AI Chat HandoffAdd to Chrome — free

Related guides