Change max context size - local llm

[marcorobiati]

Is it possible to set a different max context size? Useful when using LLMs locally. In my case I'm currently using a local model with 262144 max context size, so single agent could work with higher than default context but this also means if 4 agents work in parallel they would have 65536 each and not 200k

[trek-e]

Context window size is controlled by the AI runtime (Claude Code, OpenCode, etc.), not by GSD directly. GSD's architecture assumes 200K tokens per agent.

For local LLMs with different context limits, /gsd:set-profile and model_overrides in config let you choose which model handles each role — but the context window itself depends on your runtime's configuration.

---

Generated by Claude Code

[marcorobiati]

'GSD's architecture assumes 200K tokens per agent.' this is a concern to me and what I'd actually like to be able to change. I don't have 200k tokens per agent. Is it possible to change it?

[Aashish-po]

No, Claude Code does not currently let you configure a custom “max context size” in settings.
The effective context size is determined by the model and the agent orchestration, not by a user‑changeable parameter.
But there’s an important nuance about parallel agents that’s easy to misunderstand.

[trek-e]

@Aashish-po — correct that Claude Code does not expose a "max context size" setting, but to add precision to your response: the relevant config in GSD is config.json context_window key (set via /gsd-config-set context_window 500000), which tells GSD workflows how large of a context window to assume when constructing agent prompts. For local LLMs, the appropriate value depends on the model loaded. The agent execution itself happens in the host runtime.

[marcorobiati]

tried "/gsd-config-set context_window 262144" nothing changed, config file wasn't update and while typing command, commands suggestion doesn't show anything

[txhno]

Yes, to a point.

GSD has a context_window setting in .planning/config.json (default 200000). That tells GSD how much context to assume when building prompts.

But the real hard limit still comes from your runtime/model. So for a local 262k model, you can raise GSD’s assumed window, but GSD cannot make the model exceed what the runtime actually supports.

[marcorobiati]

Ok after re-reading my initial post I think I wan't clear enough, my bad. What I meant is that, 200k is the assumed max context size for both Claude Code and GSD, but if running local modes, the actual max context size could be different either shorter or longer, so I wanted to know if it was possible to change it in GSD (and maybe in CC too if you know how to) to align it to the actual max content size of the underlying local model used.

I'll try setting it in config.json, thanks

[trek-e]

To directly answer your follow-up @marcorobiati: yes — set context_window in .planning/config.json to your local model's actual limit.

Fastest path:

/gsd-settings

Then set context_window to your model's real maximum (e.g. 131072 for 128K, 262144 for 256K, 1000000 for 1M). GSD uses this value when sizing prompts and deciding when to compact — raising it for larger windows lets GSD keep more context in each agent call.

This does not change what the model can actually hold — it just tells GSD how much room to assume when constructing prompts. The runtime hard limit is still set by your model and Claude Code configuration.