Error: 400 Input length ____ exceeds the maximum allowed input length of ____ tokens

Summary

When using the ~recently released "Laguna-s-2.1" model, which appears to do a considerable amount of 'thinking', I have regularly encountered the message "Error: 400 Input length ____ exceeds the maximum allowed input length of ____ tokens" (in many cases, with a very small minority :: <~100/100000'th overflow) after interacting with a project's code for awhile. Shortly thereafter, it will refuse to compact (yet still emit too large an input, thus rendering the 'task' effectively unable to complete). The first few such can usually be overcome by a manual /compact (which seems to compact more than the 'auto-compact' sort), followed by 'continue'. After which command, it fails to 'continue' even when 'compact' is claimed to have 'saved' hundreds of thousands of tokens. At which point I have to switch to DeepSeek-V4-Flash-0731 or similar to ~complete. Commands such as "/usage" do not appear to show useful information about the issue.

Expected Behavior

It would seem to me that if context is indeed irretrievably ~overstuffed, then that should be reflected by the /compact command

Actual Behavior

manual /compact claims to recover many tokens, yet overflow errors continue

Steps to reproduce the issue

As shown in report overview

Command Code Version

1.14.1

Operating System

macOS

Terminal/IDE

OSX 'Terminal' w / cmd

Shell

zsh

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

No response

首先，在 macOS Terminal 上使用 zsh，並在經過長時間互動後使用 Laguna-s-2.1 模型，重現 Command Code 1.14.1 中的溢位。比較手動 /compact、auto-compact、/usage 和 continue 的行為及回報的 token 數；當壓縮準確反映剩餘的 context 並防止輸入長度錯誤反覆發生時，即視為完成。