feat: rebuild system prompt in-loop after tool execution

Tool execution can change workDir, MEMORY.md, and plan status, but
messages[0] was never refreshed within the LLM iteration loop. This
caused the next LLM call to use a stale system prompt, and the Mini App
to display outdated state.

- Rebuild messages[0] at the end of each tool-execution cycle in
  runLLMIteration, updating workDir from session TouchDir first
- Add promptDirty atomic flag so GetSystemPrompt() rebuilds on demand
  when external state changes (notifyStateChange sets dirty=true)
- Push prompt updates via SSE "prompt" event for live Mini App refresh
- Add SessionTracker.GetTouchDir() for O(1) session directory lookup

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

pkg/agent/loop.go

@@ -91,6 +91,7 @@ type AgentLoop struct {
 	activeTasks      sync.Map // sessionKey → *activeTask
 	sessions         *SessionTracker
 	lastSystemPrompt atomic.Value // string — last system prompt sent to LLM
+	promptDirty      atomic.Bool   // true = rebuild needed on next GetSystemPrompt read
 	OnStateChange    func() // called on plan/session/skills mutations
 	OnUserMessage    func() // called when a real user message is processed
 }
@@ -149,6 +150,7 @@ func NewAgentLoop(cfg *config.Config, msgBus *bus.MessageBus, provider providers
 }
 
 func (al *AgentLoop) notifyStateChange() {
+	al.promptDirty.Store(true)
 	if al.OnStateChange != nil {
 		al.OnStateChange()
 	}
@@ -918,6 +920,7 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, agent *AgentInstance, opt
 	// Capture the finalized system prompt for Mini App inspection
 	if len(messages) > 0 {
 		al.lastSystemPrompt.Store(messages[0].Content)
+		al.promptDirty.Store(false)
 	}
 
 	// 5. Run LLM iteration loop
@@ -2247,6 +2250,18 @@ func (al *AgentLoop) runLLMIteration(
 			}
 		}
 
+		// Refresh system prompt: tool execution may have changed workDir,
+		// memory, plan status, etc.  Update messages[0] so the next LLM
+		// call sees the current state.
+		if touchDir := al.sessions.GetTouchDir(opts.SessionKey); touchDir != "" {
+			agent.ContextBuilder.SetWorkDir(filepath.Join(agent.Workspace, touchDir))
+		}
+		if newPrompt := agent.ContextBuilder.BuildSystemPrompt(); len(messages) > 0 && messages[0].Content != newPrompt {
+			messages[0].Content = newPrompt
+			al.lastSystemPrompt.Store(newPrompt)
+			al.promptDirty.Store(false)
+		}
+
 	}
 
 	// If max iterations exhausted with tool calls still pending,
@@ -2483,17 +2498,23 @@ func (al *AgentLoop) GetContextInfo() (workDir, planWorkDir, workspace string, b
 }
 
 // GetSystemPrompt returns the system prompt last sent to the LLM.
-// Falls back to building from current state if no LLM call has occurred yet.
+// If the prompt is dirty (state changed since last capture), it rebuilds
+// from current state. Falls back to building if no LLM call has occurred yet.
 func (al *AgentLoop) GetSystemPrompt() string {
-	if v := al.lastSystemPrompt.Load(); v != nil {
+	if !al.promptDirty.Load() {
-		return v.(string)
+		if v := al.lastSystemPrompt.Load(); v != nil {
+			return v.(string)
+		}
 	}
-	// Fallback: no LLM call yet — build from current state
+	// Rebuild from current state
 	agent := al.registry.GetDefaultAgent()
 	if agent == nil {
 		return ""
 	}
-	return agent.ContextBuilder.BuildSystemPrompt()
+	prompt := agent.ContextBuilder.BuildSystemPrompt()
+	al.lastSystemPrompt.Store(prompt)
+	al.promptDirty.Store(false)
+	return prompt
 }
 
 // formatMessagesForLog formats messages for logging

pkg/agent/session_tracker.go

@@ -132,6 +132,15 @@ func (st *SessionTracker) ListActive() []SessionEntry {
 	return result
 }
 
+// GetTouchDir returns the TouchDir for a given session key, or "" if not found.
+func (st *SessionTracker) GetTouchDir(sessionKey string) string {
+	val, ok := st.entries.Load(sessionKey)
+	if !ok {
+		return ""
+	}
+	return val.(*SessionEntry).TouchDir
+}
+
 // GetPeerPurposes returns purposes of other active sessions targeting the same project.
 // Used for lightweight coordination without context pollution.
 func (st *SessionTracker) GetPeerPurposes(sessionKey, projectPath string) []PeerInfo {

pkg/miniapp/miniapp.go

@@ -786,7 +786,7 @@ func (h *Handler) apiEvents(w http.ResponseWriter, r *http.Request) {
 	ch := h.notifier.Subscribe()
 	defer h.notifier.Unsubscribe(ch)
 
-	var lastPlan, lastSession, lastSkills, lastDev, lastContext []byte
+	var lastPlan, lastSession, lastSkills, lastDev, lastContext, lastPrompt []byte
 
 	// Send initial state immediately
 	sendSSEIfChanged(w, flusher, "plan", h.provider.GetPlanInfo(), &lastPlan)
@@ -796,6 +796,7 @@ func (h *Handler) apiEvents(w http.ResponseWriter, r *http.Request) {
 	sendSSEIfChanged(w, flusher, "skills", h.provider.ListSkills(), &lastSkills)
 	sendSSEIfChanged(w, flusher, "dev", h.devStatus(), &lastDev)
 	sendSSEIfChanged(w, flusher, "context", h.provider.GetContextInfo(), &lastContext)
+	sendSSEIfChanged(w, flusher, "prompt", map[string]string{"prompt": h.provider.GetSystemPrompt()}, &lastPrompt)
 
 	for {
 		select {
@@ -811,6 +812,7 @@ func (h *Handler) apiEvents(w http.ResponseWriter, r *http.Request) {
 			sendSSEIfChanged(w, flusher, "skills", h.provider.ListSkills(), &lastSkills)
 			sendSSEIfChanged(w, flusher, "dev", h.devStatus(), &lastDev)
 			sendSSEIfChanged(w, flusher, "context", h.provider.GetContextInfo(), &lastContext)
+			sendSSEIfChanged(w, flusher, "prompt", map[string]string{"prompt": h.provider.GetSystemPrompt()}, &lastPrompt)
 		}
 	}
 }

pkg/miniapp/miniapp_test.go

@@ -397,8 +397,8 @@ func TestSSE_NotifyDrivesSubsequentEvents(t *testing.T) {
 	defer resp.Body.Close()
 
 	scanner := bufio.NewScanner(resp.Body)
-	// Drain initial 5 events (plan, session, skills, dev, context)
+	// Drain initial 6 events (plan, session, skills, dev, context, prompt)
-	drainEvents(t, scanner, 5, 2*time.Second)
+	drainEvents(t, scanner, 6, 2*time.Second)
 
 	// Mutate state and notify — diff dedup should detect the change and send a new event
 	provider.mutated.Store(true)
@@ -426,8 +426,8 @@ func TestSSE_DiffDedupSuppressesDuplicate(t *testing.T) {
 	defer resp.Body.Close()
 
 	scanner := bufio.NewScanner(resp.Body)
-	// Drain initial events (plan, session, skills, dev, context)
+	// Drain initial events (plan, session, skills, dev, context, prompt)
-	drainEvents(t, scanner, 5, 2*time.Second)
+	drainEvents(t, scanner, 6, 2*time.Second)
 
 	// Notify with unchanged data — should produce zero new event lines
 	notifier.Notify()
@@ -1690,8 +1690,8 @@ func TestSSE_DevEventUpdatesOnActivateDeactivate(t *testing.T) {
 	defer resp.Body.Close()
 
 	scanner := bufio.NewScanner(resp.Body)
-	// Drain initial 5 events
+	// Drain initial 6 events (plan, session, skills, dev, context, prompt)
-	drainEvents(t, scanner, 5, 2*time.Second)
+	drainEvents(t, scanner, 6, 2*time.Second)
 
 	// Activate — should trigger a dev event with active=true
 	h.ActivateDevTarget(id)

pkg/miniapp/static/index.html

@@ -1605,6 +1605,15 @@ function connectSSE() {
   eventSource.addEventListener('context', function(e) {
     try { renderContextFromData(JSON.parse(e.data)); } catch(err) {}
   });
+  eventSource.addEventListener('prompt', function(e) {
+    try {
+      var d = JSON.parse(e.data);
+      var view = document.getElementById('system-prompt-view');
+      if (view && view.style.display !== 'none') {
+        view.textContent = d.prompt || '(empty)';
+      }
+    } catch(err) {}
+  });
   eventSource.onerror = function() {
     // Browser will auto-reconnect EventSource
   };