fix: separate ContextWindow from MaxTokens configuration

[QuietyAwe]

Problem

The current implementation incorrectly sets ContextWindow to MaxTokens value in pkg/agent/instance.go:

ContextWindow:  maxTokens,  // Wrong: maxTokens is output limit, not context window

This causes several issues:

1. GLM-5 with 128K context window was incorrectly limited to 65K
2. Compression threshold was underestimated by ~50%
3. Forced compression dropped 50% of history without preserving summary

Solution

• Add new ContextWindow field to AgentDefaults config
• Default to 128K (most modern models support this)
• Users can configure smaller values for models with limited context

Changes

1. pkg/config/config.go: Add ContextWindow field to AgentDefaults
2. pkg/agent/instance.go: Use separate ContextWindow value instead of MaxTokens

Configuration Example

{
  "agents": {
    "defaults": {
      "max_tokens": 8192,
      "context_window": 128000
    }
  }
}

Testing

• Code compiles successfully
• Backward compatible: existing configs without context_window will use 128K default

Related

This fixes the context window handling for large-context models like GLM-5, Claude, GPT-4-turbo, etc.

[nikolasdehor]

This is a correct and important fix. Setting ContextWindow = maxTokens meant that for models like GLM-5 (128K context, 65K output max), the compression threshold was anchored to 65K instead of 128K, triggering premature forced compression and discarding half the conversation history unnecessarily.

The fix properly separates the two concerns:

• MaxTokens = output token limit (how many tokens the model can generate)
• ContextWindow = input context limit (how many tokens can be sent)

Defaulting to 128K is reasonable for modern models. Users with smaller-context models can override via config.

The config field is properly tagged with both json and env attributes, so it can be set via JSON config or environment variable.

LGTM.

[QuietyAwe]

Thanks for the thorough review @nikolasdehor! 🙏 Glad the fix addresses the issue correctly. Looking forward to getting this merged.