Configurable max context

[iuriguilherme]

claude-code issue: anthropics/claude-code#34925

Original text for reference:

Allow users to configure the context window percentage at which auto-compaction triggers, via a setting in settings.json (user or project level).

{
  "autoCompactionThreshold": 80
}

If not specified, the current default behavior applies.

---

Specially when using alternative models and managed by other tools such as Ollama and LMStudio, the default maximum context may not be the optimal threshold for automatic compaction. And currently it's only possible to configure this via the API for the current session and not in the settings.json file.

[Vasanthdev2004]

This also seems like a reasonable configuration request.

The default auto-compaction threshold is fine for many cloud models, but once people start using:

• Ollama
• LM Studio
• OpenRouter-managed models
• providers with very different real context behavior

then a fixed built-in threshold can become either too conservative or too risky.

A setting like:

{
  "autoCompactionThreshold": 80
}

would be easy to explain and useful for advanced users.

The main thing I would want from an implementation is that the setting remains:

• optional
• bounded/safe
• clearly documented as an advanced tuning knob rather than something most users need to touch

So yes, I think this is a sensible idea, especially for non-default provider/model setups.