fix: llm auto continue judge for new loop

Author: codefionn

internal/config/config.go

@@ -102,11 +102,14 @@ type SandboxOutputCompactionConfig struct {
 
 // LoopConfig holds configuration for the orchestrator loop abstraction
 type LoopConfig struct {
-	Strategy                string `json:"strategy"`                   // Loop strategy: "default", "conservative", "aggressive"
-	MaxIterations           int    `json:"max_iterations"`             // Maximum number of iterations (0 = use default)
-	MaxAutoContinueAttempts int    `json:"max_auto_continue_attempts"` // Maximum auto-continue attempts (0 = use default)
-	EnableLoopDetection     bool   `json:"enable_loop_detection"`      // Enable repetitive pattern detection
-	EnableAutoContinue      bool   `json:"enable_auto_continue"`       // Enable automatic continuation on incomplete responses
+	Strategy                   string `json:"strategy"`                     // Loop strategy: "default", "conservative", "aggressive", "llm-judge"
+	MaxIterations              int    `json:"max_iterations"`               // Maximum number of iterations (0 = use default)
+	MaxAutoContinueAttempts    int    `json:"max_auto_continue_attempts"`   // Maximum auto-continue attempts (0 = use default)
+	EnableLoopDetection        bool   `json:"enable_loop_detection"`        // Enable repetitive pattern detection
+	EnableAutoContinue         bool   `json:"enable_auto_continue"`         // Enable automatic continuation on incomplete responses
+	EnableLLMAutoContinueJudge bool   `json:"enable_llm_auto_continue_judge"` // Enable LLM-based auto-continue decisions
+	LLMAutoContinueJudgeTimeout int   `json:"llm_auto_continue_judge_timeout_seconds"` // LLM judge timeout in seconds (0 = use default 15s)
+	LLMAutoContinueJudgeTokenLimit int `json:"llm_auto_continue_judge_token_limit"` // LLM judge token limit (0 = use default 1000)
 }
 
 // SandboxConfig holds configuration for shell command sandboxing

internal/orchestrator/loop/doc.go

@@ -118,9 +118,41 @@
 //   - DefaultStrategy: Balanced approach with standard limits
 //   - ConservativeStrategy: Stops earlier to prevent excessive token usage
 //   - AggressiveStrategy: Continues more aggressively for batch operations
+//   - LLMJudgeStrategy: Uses an LLM to make intelligent auto-continue decisions
 //
 // Custom strategies can be implemented by implementing the Strategy interface.
 //
+// # LLM Judge Strategy
+//
+// The LLMJudgeStrategy provides intelligent auto-continue decisions by:
+//
+// 1. Checking for repetitive patterns (loops) in recent messages
+// 2. Using an LLM to evaluate whether continuation is needed
+// 3. Applying model-specific decision logic (Qwen 3, Mistral, etc.)
+// 4. Falling back to pattern matching on any LLM failure
+//
+// Configuration options:
+//
+//   - EnableLLMAutoContinueJudge: Enable/disable the LLM judge
+//   - LLMAutoContinueJudgeTimeout: Maximum time to wait for LLM response (default: 15s)
+//   - LLMAutoContinueJudgeTokenLimit: Context limit sent to LLM judge (default: 1000 tokens)
+//
+// Example usage:
+//
+//	config := loop.DefaultConfig()
+//	config.EnableLLMAutoContinueJudge = true
+//	config.LLMAutoContinueJudgeTimeout = 15 * time.Second
+//	config.LLMAutoContinueJudgeTokenLimit = 1000
+//	factory := loop.NewStrategyFactory(config)
+//	strategy := factory.CreateWithLLMJudge("llm-judge", config, llmClient, modelID, session)
+//
+// The LLM judge automatically falls back to pattern matching if:
+//   - The LLM client is not available
+//   - The model ID is empty
+//   - The LLM call times out
+//   - The LLM returns an invalid response
+//   - Auto-continue is disabled in the config
+//
 // # Testing
 //
 // The package includes mock implementations for testing: