[Feature Request] Configurable Context Buffer for Spawned Agents (DeepSeek/External Models)

[ptmlst]

Hi GSD Team,

First of all, I want to say that GSD is an absolute powerhouse. Using it alongside Claude Code and DeepSeek-Reasoner (via ANTHROPIC_BASE_URL override) has been a game-changer for my workflow. The synergy is incredible.

The Context Limit Issue:
While mapping large codebases or running complex background tasks, I've encountered a 400 Invalid Request Error. It seems that during agent spawning or sub-task execution, the combined token count (messages + completion) slightly exceeds the model's hard limit.

Error Example:

API Error: 400 {"error":{"message":"This model's maximum context length is 131072 tokens. However, you requested 131973 tokens (99973 in the messages, 32000 in the completion).", "code":"invalid_request_error"}}

In this case, the tool requested a 32,000 token completion on top of a 99,973 token message history, pushing it just 901 tokens over the 131,072 limit.

The Proposal:
Could we have a configuration parameter (or an environment variable like GSD_MAX_CONTEXT_BUFFER) to manually set a safety cap for context?

If we could tell GSD to:

Treat the max context as 120k instead of the model's theoretical 131k.

Pass this constraint down to all spawned agents and background tasks.

This would prevent these "edge-case" crashes when using models like DeepSeek that have strict context windows. It would make GSD even more robust for those of us using custom API endpoints.

Thanks for building such a great tool!