Early compression of the 1M Subagent context window (broken subagent resume, yep))

[Artemonim]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

A subagent with a context size of one million compressed the context to approximately 300k window size.

Steps to Reproduce

Try running an Agent that launches a Subagent, which reads a lot of information, stops, and then resumes via resume to read even more information, until the summarization happen.

Screenshots / Screen Recordings

Operating System

Windows 10/11

Version Information

Version: 3.16.29 (system setup)
VS Code Extension API: 1.128.0
Commit: 6246455961129c32969845e00aa25d87ae926ec0
Date: 2026-08-18T01:26:26.285Z
Layout: IDE
Build Type: Stable
Release Track: Nightly
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.291
OS: Windows_NT x64 10.0.22631

For AI issues: which model did you use?

Agent Kimi K3 Max
Subagent GPT-5.6-Terra-Max-1M

For AI issues: add Request ID with privacy disabled

agent: 3795497f-48bb-46b5-939b-205dba9c43e1
subagent: e8ee4abb-b059-4b03-999f-db158a1678e4

Additional Information

The screenshot of the continued Subagent shown above was taken not at the moment the chat was compressed, but afterward; however, it is clear that he has accumulated less than half a million tokens in the cache as a result of his work.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[Colin]

Hey @Artemonim, thanks for the report!

When a subagent is resumed, it can silently fall back to the model’s default variant instead of the one it was spawned with. For GPT-5.6-Terra the default variant has a 272k context window, so your resumed subagent started summarizing around 300k even though it was launched on the 1M variant. This is a known issue we’re already tracking, and I’ve added your report to it.

In the meantime, select the 1M variant of the model in the main model picker, not just in the subagent configuration. Resumed subagents fall back to your per-model preference, so pinning the 1M variant there keeps resumes on the full context window.

[Artemonim]

@Colin

In the meantime, select the 1M variant of the model in the main model picker, not just in the subagent configuration.

This is very inconvenient. Especially since the values sometimes get reset during restarts or updates (I think this has happened to me at least once).

I hope this issue will be resolved soon.

[Artemonim]

@Colin

Resumed subagents fall back to your per-model preference

Version: 3.18.3 (system setup)
VS Code Extension API: 1.128.0
Commit: ea04a76369d62063da9a0af1c67c437bb885ee50
Date: 2026-08-25T21:42:34.353Z
Layout: IDE
Build Type: Stable
Release Track: Nightly
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.291
OS: Windows_NT x64 10.0.22631