Bug Report - Chat UI blocked due to [string_above_max_length] in automatic context compaction

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Pro user chat UI completely frozen and blocked. Getting error: [StringParam] [input.arguments] [string_above_max_length]. Invalid ‘input.arguments’: string too long. Expected a string with maximum length 1048576, but got a string with length 1329691 instead. Account name: Jesus Morales Carlos (Pro).

Steps to Reproduce

  1. Open a long-running chat session (“RPAs”).
  2. Send a brief command to monitor a process.
  3. The system triggers the automatic context compaction mechanism (“Contexto compactado automáticamente”) in the background.
  4. The background process generates a text string of ~1.33 MB, which exceeds the 1 MB API limit and corrupts the local state of the chat.

Expected Behavior

The system should compress or truncate the context properly without exceeding the API token/character limits, and the chat should remain operational instead of freezing the entire UI.

Screenshots / Screen Recordings

bug_codex.png
bug_codex.png782×138 4.38 KB

Operating System

Windows 10/11

Version Information

Latest Cursor Desktop Version (Pro Account)

For AI issues: which model did you use?

Claude 3.5 Sonnet / GPT-4o

For AI issues: add Request ID with privacy disabled

Not available. The chat UI is completely frozen and does not allow opening the request logs.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

6

Hi @Jesus_Morales_Carlos Thanks for the post! Here’s what’s happening: one of the tool calls in your conversation history is carrying about 1.33 MB of arguments, and the limit for a single value is 1 MB. Auto-compaction reduces the overall size of your context, but it doesn’t split up an individual oversized item — so that one call keeps failing validation on every retry.

The quickest unblock is to start a fresh chat rather than continuing the existing one. It usually traces back to a very large file or pasted payload getting handed to a tool, so if you can spot that step, avoiding it will keep it from recurring.

One note on where to file it: this is coming from the Codex extension rather than Cursor’s own agent, so OpenAI’s channels are the right place to report the underlying behavior. Happy to keep helping troubleshoot here in the meantime, and if you hit anything similar with Cursor’s agent, just let us know and we’ll dig in.