#Canvas Session Desync & Corrupted Output

1 messages · Page 1 of 1 (latest)

surreal chasmBOT
#

Reported by @crimson hatch

Bug Report: Canvas Session Desync & Corrupted Output
`Steps to Reproduce`
  1. Assistant initiates canvas edits without opening UI.
  2. User manually opens canvas.
  3. Assistant continues thinking canvas is active, but all further output appears in chat instead of canvas.
  4. Canvas stops responding to assistant commands, throws InvalidRecipient: canvas errors, and breaks interactivity.
`Expected Result`

Canvas session is maintained correctly, even if the user manually opens the canvas after assistant begins the session.

`Actual Result`

Canvas becomes desynced. The assistant no longer controls the document and begins replying into chat with markdown-formatted code. Assistant-side canvas commands begin failing.

`Environment`

GPT-4o Web Chat Firefox on Linux (user-specific) Assistant attempted code editing over several messages before desync was noticed Canvas session ID still appears to persist but is no longer valid

#
Additional Information

Please provide relevant details to help resolve the issue, such as:

  • ChatGPT Shared Link (if applicable).
  • Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

crimson hatch
sweet girder
#

I've noted these and more anomalies a number of times here. Copy/paste the following into the Search here and you'll see similar weirdness. My personal bottom line is that I never use this well-intentioned but sadly broken Canvas feature.

canvas from: @captainstarbuck in: bug-reports

crimson hatch
#

i didn't think of it originally but is the code that leaks into chat still sand boxed? don't answer that i don't need to know but its something that i could see happening

#

and also something im not willing to test

sweet girder
#

Oh, I think I know exactly what you're describing in this specific case...

The model's markdown output processing is not tuned well enough to understand when to escape markdown syntax. So it will occasionally break out of the ``` triple backtick mode required for code, and absolutely not know that it's done so. From there everything gets formatted as plain text, which looks horrible with asterisks and other code-specific syntax that gets rendered by markdown.

#

To handle that specific case, I used the following in my code project instructions:

# Explicit Documentation Requests

If the user explicitly asks you to "generate documentation" or uses similar terminology (e.g., "create a manual," "generate a guide," "write up instructions"), it is absolutely mandatory to use this special Explicit Documentation protocol: 
1) Recognize the difference between your response, the "Explicit Documentation" within your response, and the content within that Explicit Documentation.
2) Begin and end your response normally.
3) Begin the "Explicit Documentation" as a standard Markdown codeblock using triple backticks.
4) Within the Explicit Documentation block, preface/escape Every backtick with a backslash: "\`". The content is not to be formatted. All formatting characters must be escaped.
5) Always use this format for these special Explicit Documentation responses, never for other responses.