fix: separate ContextWindow from MaxTokens configuration

- Add ContextWindow field to AgentDefaults config
- ContextWindow now defaults to 128K (most modern models support this)
- Previously ContextWindow was incorrectly set to MaxTokens (output limit)
- This fixes premature context compression for large-context models like GLM-5 (128K context)

Fixes issue where:
- GLM-5 with 128K context was limited to 65K (MaxTokens value)
- Compression threshold was underestimated by 50%
- Forced compression dropped 50% of history without summary

pkg/agent/instance.go

@@ -82,6 +82,13 @@ func NewAgentInstance(
 		maxTokens = 8192
 	}
 
+	// ContextWindow defaults to 128K if not specified (most modern models support this)
+	// For models with smaller context, users should configure this explicitly
+	contextWindow := defaults.ContextWindow
+	if contextWindow == 0 {
+		contextWindow = 128000
+	}
+
 	temperature := 0.7
 	if defaults.Temperature != nil {
 		temperature = *defaults.Temperature
@@ -103,7 +110,7 @@ func NewAgentInstance(
 		MaxIterations:  maxIter,
 		MaxTokens:      maxTokens,
 		Temperature:    temperature,
-		ContextWindow:  maxTokens,
+		ContextWindow:  contextWindow,
 		Provider:       provider,
 		Sessions:       sessionsManager,
 		ContextBuilder: contextBuilder,

pkg/config/config.go

@@ -177,6 +177,7 @@ type AgentDefaults struct {
 	ImageModel          string   `json:"image_model,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_IMAGE_MODEL"`
 	ImageModelFallbacks []string `json:"image_model_fallbacks,omitempty"`
 	MaxTokens           int      `json:"max_tokens"                      env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOKENS"`
+	ContextWindow       int      `json:"context_window"                  env:"PICOCLAW_AGENTS_DEFAULTS_CONTEXT_WINDOW"`
 	Temperature         *float64 `json:"temperature,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_TEMPERATURE"`
 	MaxToolIterations   int      `json:"max_tool_iterations"             env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS"`
 }