I’m seeing a recurring error when I select either Auto mode or the Composer 2.5 (fast) model in Cursor.
Error shown:
Error: Something went wrong. Please try again.
[internal] Failed to run step, exceeded max retries
This happens specifically when choosing:
• Auto mode
• Composer 2.5 (fast)
Other models/modes may still work, but these options consistently fail for me.
Environment:
• OS: macOS Darwin 25.5.0
• Cursor workspace: swarm-forge
• Approximate time observed: Sunday, Jun 14, 2026 around 1:49 PM UTC-5
• Cursor version: 3.7.19
Could you help me understand whether this is an account/model availability issue, a backend retry failure, or
something related to my local Cursor setup?
Please let me know if you need logs or diagnostic info, and where I should collect them from.
Steps to Reproduce
Steps to reproduce:
Open Cursor.
Open any workspace or start a new chat.
In the model picker, select Auto mode.
Send a prompt.
Observe the error: Error: Something went wrong. Please try again. [internal] Failed to run step, exceeded max retries
Repeat with Composer 2.5 (fast) selected.
Observe the same error.
Expected Behavior
The selected mode/model should start the request and return a response.
Actual behavior:
The request fails with an internal max retries error.
This is a known bug on our side, not an account or local-setup issue. What’s happening: the chat’s context has grown past the token window for Auto / Composer 2.5 (fast), and our automatic summarization can’t shrink it back under the limit fast enough, so the request keeps retrying and eventually fails with Failed to run step, exceeded max retries. That’s also why it hits Auto and Composer 2.5 (fast) but not larger-context models like Sonnet.
What helps right now:
Start a new chat. The oversized thread can’t be recovered, and new messages in it will keep failing.
For heavy or long sessions, pick a specific larger-context model (e.g., Sonnet) instead of Auto / Composer 2.5 (fast).
If a brand-new chat with just a “hello” still fails, your persistent context is probably already near the limit before you type anything. The usual culprit is a large set of rules/skills (and sometimes MCP servers with big tool schemas) loading on every request. Open Cursor Settings (Cmd+Shift+J) → Rules, Skills, Subagents and trim what you don’t need. For ones you want to keep, add disable-model-invocation: true to the header so they only load when you call them with /name.Temporarily disabling large MCP servers and sending “hello” in an empty folder is a quick way to confirm whether project/global context is the cause.
You’re already in the right spot to reach us: I’m on the Cursor team and I’ll keep helping you here on the thread. If you’d rather handle it over email, [email protected] reaches us too.
To get you unblocked, two quick things would help me tell whether this is still the same context/token issue or something new:
In a brand-new chat, does it still fail if you specifically pick Sonnet (instead of Auto or Composer 2.5 fast)? Sonnet has a much larger context window, so if that also fails we’re likely looking at a different cause than before.
If it does fail, grab a fresh Request ID from that attempt (three-dot menu on the message → Copy Request ID) and drop it here. The one from your first report is already traced - a new one lets me see exactly what’s happening now.
The underlying bug is still actively being worked on. I don’t have an ETA to share yet, but it’s tracked on our side and I’ll follow up on this thread once there’s a fix to report.