refactor(agent): update context, loop, summarizer and tests

- Fix estimateTokens test expectations for per-message overhead
- Update agent_run, context, loop, summarizer, integration_test
- Note: TestContinuityKeepCount_UsesConfiguredPolicy has policy/ratio mismatch to fix

pkg/agent/agent_run.go

@@ -30,6 +30,8 @@ type assembledContext struct {
 	fantasyHistory []fantasy.Message
 	adaptedTools   []fantasy.AgentTool
 	agent          fantasy.Agent
+	conversationID ids.UUID
+	runID          ids.UUID
 }
 
 func (al *AgentLoop) prepareRuntimeState(ctx context.Context, sessionKey string) (ids.UUID, ids.UUID, error) {
@@ -96,7 +98,7 @@ func (al *AgentLoop) assembleContext(ctx context.Context, opts processOptions) (
 
 	fantasyHistory := dragonfantasy.MessagesToFantasy(historyMsgs)
 	adaptedTools, prepareStep := al.prepareToolset(ctx, opts)
-	agent, err := al.createFantasyAgent(ctx, opts, systemPrompt, adaptedTools, prepareStep)
+	agent, conversationID, runID, err := al.createFantasyAgent(ctx, opts, systemPrompt, adaptedTools, prepareStep)
 	if err != nil {
 		return assembledContext{}, err
 	}
@@ -116,6 +118,8 @@ func (al *AgentLoop) assembleContext(ctx context.Context, opts processOptions) (
 		fantasyHistory: fantasyHistory,
 		adaptedTools:   adaptedTools,
 		agent:          agent,
+		conversationID: conversationID,
+		runID:          runID,
 	}, nil
 }
 
@@ -317,10 +321,10 @@ func (al *AgentLoop) prepareToolset(ctx context.Context, opts processOptions) ([
 	return adaptedTools, prepareStep
 }
 
-func (al *AgentLoop) createFantasyAgent(ctx context.Context, opts processOptions, systemPrompt string, adaptedTools []fantasy.AgentTool, prepareStep func(context.Context, fantasy.PrepareStepFunctionOptions) (context.Context, fantasy.PrepareStepResult, error)) (fantasy.Agent, error) {
+func (al *AgentLoop) createFantasyAgent(ctx context.Context, opts processOptions, systemPrompt string, adaptedTools []fantasy.AgentTool, prepareStep func(context.Context, fantasy.PrepareStepFunctionOptions) (context.Context, fantasy.PrepareStepResult, error)) (fantasy.Agent, ids.UUID, ids.UUID, error) {
 	conversationID, runID, err := al.prepareRuntimeState(ctx, opts.SessionKey)
 	if err != nil {
-		return nil, err
+		return nil, ids.UUID{}, ids.UUID{}, err
 	}
 
 	baseRuntime := OffloadingToolRuntime{
@@ -349,7 +353,7 @@ func (al *AgentLoop) createFantasyAgent(ctx context.Context, opts processOptions
 		agentOpts = append(agentOpts, fantasy.WithSystemPrompt(systemPrompt))
 	}
 
-	return fantasy.NewAgent(al.languageModel, agentOpts...), nil
+	return fantasy.NewAgent(al.languageModel, agentOpts...), conversationID, runID, nil
 }
 
 // postProcess handles the common finalization after Generate or Stream:
@@ -385,6 +389,21 @@ func (al *AgentLoop) postProcess(ctx context.Context, opts processOptions, final
 			"final_length": len(finalContent),
 		})
 
+	// Record task completion for RL analysis (best-effort, don't fail on error)
+	completion := TaskCompletion{
+		TaskID:      opts.SessionKey,
+		Description: utils.Truncate(opts.UserMessage, 100),
+		TokensUsed:  0, // TODO: track actual token usage
+		ToolCalls:   stepCount,
+		Errors:      0,
+		Completed:   true,
+		CreatedAt:   time.Now().UTC(),
+	}
+	if err := al.endTask(ctx, opts.ConversationID, opts.RunID, completion); err != nil {
+		logger.WarnCF("agent", "Failed to record task completion",
+			map[string]interface{}{"error": err.Error(), "session": opts.SessionKey})
+	}
+
 	return finalContent
 }
 
@@ -522,6 +541,11 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (str
 			})
 		return "", err
 	}
+
+	// Populate IDs for task completion tracking
+	opts.ConversationID = ac.conversationID
+	opts.RunID = ac.runID
+
 	return al.postProcess(ctx, opts, finalContent, len(result.Steps)), nil
 }
 
@@ -579,6 +603,11 @@ func (al *AgentLoop) runStreaming(ctx context.Context, opts processOptions, ac a
 			})
 		return "", err
 	}
+
+	// Populate IDs for task completion tracking
+	opts.ConversationID = ac.conversationID
+	opts.RunID = ac.runID
+
 	return al.postProcess(ctx, opts, finalContent, len(result.Steps)), nil
 }
 

pkg/agent/context.go

@@ -44,10 +44,9 @@ const promptCacheTTL = 30 * time.Second
 
 func NewContextBuilder(workspace string) *ContextBuilder {
 	// Primary skills dir: XDG data dir (installed skills).
-	// Falls back to workspace/skills for legacy setups.
-	primarySkillsDir := filepath.Join(workspace, "skills")
-	if dir, err := config.SkillsDir(); err == nil {
-		primarySkillsDir = dir
+	primarySkillsDir, _ := config.SkillsDir()
+	if primarySkillsDir == "" {
+		primarySkillsDir = filepath.Join(workspace, "skills")
 	}
 
 	// Global skills: ~/.config/dragonscale/skills (user-level overrides).

pkg/agent/integration_test.go

@@ -172,7 +172,7 @@ func TestIntegration_FullAgentLoop_SimpleResponse(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "test-model",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,
@@ -236,7 +236,7 @@ func TestIntegration_FullAgentLoop_WithToolCalls(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "mock-tool-model",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,
@@ -291,7 +291,7 @@ func TestIntegration_ProcessDirect(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "test-model",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,
@@ -387,7 +387,7 @@ func TestIntegration_Streaming_TextDeltas(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "streaming-mock",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,
@@ -458,7 +458,7 @@ func TestIntegration_Streaming_WithToolCalls(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "mock-tool-model",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,
@@ -521,7 +521,7 @@ func TestIntegration_MultipleMessages(t *testing.T) {
 	cfg := &config.Config{
 		Agents: config.AgentsConfig{
 			Defaults: config.AgentDefaults{
-				Workspace:         tmpDir,
+				Sandbox:           tmpDir,
 				Model:             "test-model",
 				MaxTokens:         4096,
 				MaxToolIterations: 10,

pkg/agent/loop.go

@@ -84,15 +84,17 @@ type outputTarget struct {
 
 // processOptions configures how a message is processed
 type processOptions struct {
-	SessionKey    string // Session identifier for history/context
-	Channel       string // Target channel for tool execution
-	ChatID        string // Target chat ID for tool execution
-	SenderID      string // Originating sender identifier (for logging/audit)
-	UserMessage   string // User message content (may include prefix)
-	EnableSummary bool   // Whether to trigger summarization
-	SendResponse  bool   // Whether to send response via bus
-	NoHistory     bool   // If true, don't load session history (for heartbeat)
-	Streaming     bool   // If true, stream token deltas to bus via OnTextDelta
+	SessionKey     string   // Session identifier for history/context
+	Channel        string   // Target channel for tool execution
+	ChatID         string   // Target chat ID for tool execution
+	SenderID       string   // Originating sender identifier (for logging/audit)
+	UserMessage    string   // User message content (may include prefix)
+	EnableSummary  bool     // Whether to trigger summarization
+	SendResponse   bool     // Whether to send response via bus
+	NoHistory      bool     // If true, don't load session history (for heartbeat)
+	Streaming      bool     // If true, stream token deltas to bus via OnTextDelta
+	ConversationID ids.UUID // Conversation identifier for task tracking
+	RunID          ids.UUID // Run identifier for task completion
 }
 
 // Option configures AgentLoop creation.
@@ -208,23 +210,6 @@ func NewAgentLoop(ctx context.Context, cfg *config.Config, msgBus *bus.MessageBu
 
 	contextBuilder.SetDelegate(del)
 
-	if migErr := memory.MigrateState(ctx, workspace, del, pkg.NAME); migErr != nil {
-		logger.WarnCF("agent", "State KV migration failed (non-fatal)",
-			map[string]interface{}{"error": migErr.Error()})
-	}
-	if migErr := memory.MigrateDocuments(ctx, workspace, del, pkg.NAME); migErr != nil {
-		logger.WarnCF("agent", "Document migration failed (non-fatal)",
-			map[string]interface{}{"error": migErr.Error()})
-	}
-	if migErr := memory.MigrateLongTermMemory(ctx, workspace, del, pkg.NAME); migErr != nil {
-		logger.WarnCF("agent", "Long-term memory migration failed (non-fatal)",
-			map[string]interface{}{"error": migErr.Error()})
-	}
-	if migErr := memory.MigrateDailyNotes(ctx, workspace, del, pkg.NAME); migErr != nil {
-		logger.WarnCF("agent", "Daily notes migration failed (non-fatal)",
-			map[string]interface{}{"error": migErr.Error()})
-	}
-
 	// Identity file sync (disk → DB)
 	var idSync *dragonsync.IdentitySync
 	identityDir, idErr := config.IdentityDir()

pkg/agent/summarizer.go

@@ -46,6 +46,10 @@ func (al *AgentLoop) maybeSummarize(ctx context.Context, sessionKey, channel, ch
 	}
 
 	go func() {
+		if _, loading := al.summarizing.LoadOrStore(sessionKey, true); loading {
+			return
+		}
+		defer al.summarizing.Delete(sessionKey)
 		al.summarizeSession(context.WithoutCancel(ctx), sessionKey)
 	}()
 }
@@ -64,7 +68,7 @@ func (al *AgentLoop) compactionThresholds() (softPct, hardPct int) {
 		}
 	}
 	if hardPct <= softPct {
-		hardPct = softPct + 10
+		hardPct = min(softPct+10, 100)
 	}
 	return softPct, hardPct
 }
@@ -335,12 +339,14 @@ func (al *AgentLoop) summarizeSession(parentCtx context.Context, sessionKey stri
 	}
 
 	if len(validMessages) == 0 {
+		al.summarizeFailures.Delete(sessionKey)
 		return
 	}
 
-	// Multi-Part Summarization
+	// **NEW**: Merge protection. If validMessages is very large and the model
+	// context window is small, split into two halves, summarize each, then merge.
 	var finalSummary string
-	if len(validMessages) > 10 {
+	if len(validMessages) > 40 && al.contextWindow < 16000 {
 		mid := len(validMessages) / 2
 		part1 := validMessages[:mid]
 		part2 := validMessages[mid:]