feat(subturn): support stateful iteration for evaluator-optimizer pattern

Add ActualSystemPrompt and InitialMessages fields to SubTurnConfig to enable
stateful worker context passing across multiple evaluation iterations.

Changes:
- Add ActualSystemPrompt field to separate system role from user task description
- Add InitialMessages field to preload ephemeral session history before agent loop starts
- Add Messages field to ToolResult for carrying session history (internal use, not serialized)
- Update runTurn to inject system prompt and preload history from InitialMessages
- Update AgentLoopSpawner to map new fields from tools.SubTurnConfig to agent.SubTurnConfig

This enables the evaluator-optimizer execution strategy in team tool to maintain
worker context across iterations while keeping SubTurn isolation intact.

pkg/agent/loop.go

@@ -64,6 +64,7 @@ type processOptions struct {
 	SenderID                string   // Current sender ID for dynamic context
 	SenderDisplayName       string   // Current sender display name for dynamic context
 	UserMessage             string   // User message content (may include prefix)
+	SystemPromptOverride    string   // Override the default system prompt (Used by SubTurns)
 	Media                   []string // media:// refs from inbound message
 	DefaultResponse         string   // Response when LLM returns empty
 	EnableSummary           bool     // Whether to trigger summarization
@@ -1069,6 +1070,17 @@ func (al *AgentLoop) runAgentLoop(
 	maxMediaSize := cfg.Agents.Defaults.GetMaxMediaSize()
 	messages = resolveMediaRefs(messages, al.mediaStore, maxMediaSize)
 
+	// 1.5 Override the System prompt (e.g., for Evaluator/Optimizer specific personas)
+	if opts.SystemPromptOverride != "" {
+		for i, msg := range messages {
+			if msg.Role == "system" {
+				messages[i].Content = opts.SystemPromptOverride
+				messages[i].SystemParts = []providers.ContentBlock{{Type: "text", Text: opts.SystemPromptOverride}}
+				break
+			}
+		}
+	}
+
 	// 2. Save user message to session
 	if !opts.SkipAddUserMessage && opts.UserMessage != "" {
 		agent.Sessions.AddMessage(opts.SessionKey, "user", opts.UserMessage)

pkg/agent/subturn.go

@@ -119,6 +119,14 @@ type SubTurnConfig struct {
 	// truncated while preserving system messages and recent context.
 	MaxContextRunes int
 
+	// ActualSystemPrompt is injected as the true 'system' role message for the childAgent.
+	// The legacy SystemPrompt field is actually used as the first 'user' message (task description).
+	ActualSystemPrompt string
+
+	// InitialMessages preloads the ephemeral session history before the agent loop starts.
+	// Used by evaluator-optimizer patterns to pass the full worker context across multiple iterations.
+	InitialMessages []providers.Message
+
 	// Can be extended with temperature, topP, etc.
 }
 
@@ -186,14 +194,16 @@ func (s *AgentLoopSpawner) SpawnSubTurn(ctx context.Context, cfg tools.SubTurnCo
 
 	// Convert tools.SubTurnConfig to agent.SubTurnConfig
 	agentCfg := SubTurnConfig{
-		Model:           cfg.Model,
+		Model:              cfg.Model,
-		Tools:           cfg.Tools,
+		Tools:              cfg.Tools,
-		SystemPrompt:    cfg.SystemPrompt,
+		SystemPrompt:       cfg.SystemPrompt,
-		MaxTokens:       cfg.MaxTokens,
+		ActualSystemPrompt: cfg.ActualSystemPrompt,
-		Async:           cfg.Async,
+		InitialMessages:    cfg.InitialMessages,
-		Critical:        cfg.Critical,
+		MaxTokens:          cfg.MaxTokens,
-		Timeout:         cfg.Timeout,
+		Async:              cfg.Async,
-		MaxContextRunes: cfg.MaxContextRunes,
+		Critical:           cfg.Critical,
+		Timeout:            cfg.Timeout,
+		MaxContextRunes:    cfg.MaxContextRunes,
 	}
 
 	return spawnSubTurn(ctx, s.al, parentTS, agentCfg)
@@ -481,6 +491,19 @@ func runTurn(ctx context.Context, al *AgentLoop, ts *turnState, cfg SubTurnConfi
 		childAgent.MaxTokens = parentAgent.MaxTokens
 	}
 
+	if cfg.ActualSystemPrompt != "" {
+		childAgent.Sessions.AddMessage(ts.turnID, "system", cfg.ActualSystemPrompt)
+	}
+
+	promptAlreadyAdded := false
+
+	// Preload ephemeral session history
+	if len(cfg.InitialMessages) > 0 {
+		existing := childAgent.Sessions.GetHistory(ts.turnID)
+		childAgent.Sessions.SetHistory(ts.turnID, append(existing, cfg.InitialMessages...))
+		promptAlreadyAdded = true // InitialMessages 中已含 user 消息，跳过再次添加
+	}
+
 	// Resolve MaxContextRunes configuration
 	maxContextRunes := utils.ResolveMaxContextRunes(cfg.MaxContextRunes, childAgent.ContextWindow)
 
@@ -501,7 +524,6 @@ func runTurn(ctx context.Context, al *AgentLoop, ts *turnState, cfg SubTurnConfi
 	truncationRetryCount := 0
 	contextRetryCount := 0
 	currentPrompt := cfg.SystemPrompt
-	promptAlreadyAdded := false
 
 	for {
 		// Soft context limit: check and truncate before LLM call
@@ -535,12 +557,13 @@ func runTurn(ctx context.Context, al *AgentLoop, ts *turnState, cfg SubTurnConfi
 
 		// Call the agent loop
 		finalContent, err := al.runAgentLoop(ctx, childAgent, processOptions{
-			SessionKey:         ts.turnID,
+			SessionKey:           ts.turnID,
-			UserMessage:        currentPrompt,
+			UserMessage:          currentPrompt,
-			DefaultResponse:    "",
+			SystemPromptOverride: cfg.ActualSystemPrompt,
-			EnableSummary:      false,
+			DefaultResponse:      "",
-			SendResponse:       false,
+			EnableSummary:        false,
-			SkipAddUserMessage: promptAlreadyAdded,
+			SendResponse:         false,
+			SkipAddUserMessage:   promptAlreadyAdded,
 		})
 
 		// Mark the prompt as added so subsequent truncation retries
@@ -600,8 +623,11 @@ func runTurn(ctx context.Context, al *AgentLoop, ts *turnState, cfg SubTurnConfi
 			continue // Retry with recovery prompt
 		}
 
-		// 3. Success - return result
+		// 3. Success - return result with session history
-		return &tools.ToolResult{ForLLM: finalContent}, nil
+		return &tools.ToolResult{
+			ForLLM:   finalContent,
+			Messages: childAgent.Sessions.GetHistory(ts.turnID),
+		}, nil
 	}
 }
 

pkg/tools/result.go

@@ -1,6 +1,10 @@
 package tools
 
-import "encoding/json"
+import (
+	"encoding/json"
+
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
 
 // ToolResult represents the structured return value from tool execution.
 // It provides clear semantics for different types of results and supports
@@ -34,6 +38,11 @@ type ToolResult struct {
 	// Media contains media store refs produced by this tool.
 	// When non-empty, the agent will publish these as OutboundMediaMessage.
 	Media []string `json:"media,omitempty"`
+
+	// Messages holds the ephemeral session history after execution.
+	// Only populated by SubTurn executions; used by evaluator_optimizer
+	// to carry stateful worker context across evaluation iterations.
+	Messages []providers.Message `json:"-"`
 }
 
 // NewToolResult creates a basic ToolResult with content for the LLM.

pkg/tools/subagent.go

@@ -17,15 +17,17 @@ type SubTurnSpawner interface {
 
 // SubTurnConfig holds configuration for spawning a sub-turn.
 type SubTurnConfig struct {
-	Model        string
+	Model              string
-	Tools        []Tool
+	Tools              []Tool
-	SystemPrompt string
+	SystemPrompt       string
-	MaxTokens    int
+	MaxTokens          int
-	Temperature  float64
+	Temperature        float64
-	Async        bool          // true for async (spawn), false for sync (subagent)
+	Async              bool          // true for async (spawn), false for sync (subagent)
-	Critical     bool          // continue running after parent finishes gracefully
+	Critical           bool          // continue running after parent finishes gracefully
-	Timeout      time.Duration // 0 = use default (5 minutes)
+	Timeout            time.Duration // 0 = use default (5 minutes)
-	MaxContextRunes int        // 0 = auto, -1 = no limit, >0 = explicit limit
+	MaxContextRunes    int           // 0 = auto, -1 = no limit, >0 = explicit limit
+	ActualSystemPrompt string
+	InitialMessages    []providers.Message
 }
 
 type SubagentTask struct {
@@ -203,7 +205,7 @@ After completing the task, provide a clear summary of what was done.`
 			MaxIterations: maxIter,
 			LLMOptions:    llmOptions,
 		}, messages, task.OriginChannel, task.OriginChatID)
-		
+
 		if err == nil {
 			result = &ToolResult{
 				ForLLM: fmt.Sprintf(