feat(system): Prax: Codex+Gemini log adapters, model tags in monitor, inotify errno-specific messages

Co-Authored-By: @devpulse <devpulse@aipass>

src/aipass/prax/apps/handlers/monitoring/filesystem_handler.py

@@ -12,7 +12,7 @@ PRAX Filesystem Event Handler
 Watchdog FileSystemEventHandler that processes filesystem events and:
 1. Pushes events to the monitoring event queue for display
 2. Fires trigger events for cross-module integration
-3. Parses Claude Code JSONL sessions for agent activity tracking
+3. Parses CLI session files for agent activity tracking (Claude Code, Codex, Gemini)
 
 Extracted from monitor.py to maintain 3-layer architecture:
   module (orchestration) -> handler (implementation)
@@ -70,6 +70,8 @@ class MonitoringFileHandler(FileSystemEventHandler):
         self._jsonl_positions: Dict[str, int] = {}
         # Track last agent action per session to avoid duplicate displays
         self._last_agent_action: Dict[str, str] = {}
+        # Track model per session file for display tags
+        self._session_models: Dict[str, str] = {}
 
     # -- public property so the module can swap queues after construction ------
     @property
@@ -151,6 +153,12 @@ class MonitoringFileHandler(FileSystemEventHandler):
             return f"🚀 Agent: {inp.get('description', '')[:80]}"
         return f"🔧 {tool_name}"
 
+    @staticmethod
+    def _extract_model_from_entry(entry: dict) -> Optional[str]:
+        """Extract model name from a Claude Code JSONL entry."""
+        msg = entry.get('message', {}) if isinstance(entry.get('message'), dict) else {}
+        return msg.get('model')
+
     @staticmethod
     def _extract_action_from_entry(entry: dict) -> Optional[str]:
         """Extract a display action string from a JSONL entry."""
@@ -183,6 +191,259 @@ class MonitoringFileHandler(FileSystemEventHandler):
 
         return None
 
+    # =========================================================================
+    # MODEL TAG HELPERS
+    # =========================================================================
+
+    @staticmethod
+    def _shorten_model(model: str) -> str:
+        """Shorten a model identifier for display."""
+        if not model:
+            return ''
+        m = model.lower()
+        # Claude models
+        if 'opus' in m:
+            return 'opus'
+        if 'sonnet' in m:
+            return 'sonnet'
+        if 'haiku' in m:
+            return 'haiku'
+        # OpenAI/Codex models
+        if m.startswith('gpt-'):
+            return m  # already short: gpt-4o, gpt-5.4
+        if m.startswith('o') and any(c.isdigit() for c in m):
+            return m  # o1, o3, o4-mini
+        # Gemini models
+        if 'gemini' in m:
+            # gemini-3-flash-preview → gemini-3-flash
+            parts = m.replace('gemini-', '').split('-')
+            return 'gemini-' + '-'.join(p for p in parts if p != 'preview')
+        # Fallback: first 15 chars
+        return model[:15]
+
+    def _tag_branch_with_model(self, path_key: str, branch: str) -> str:
+        """Append model tag to branch if known for this session."""
+        model = self._session_models.get(path_key, '')
+        if model:
+            return f"{branch}/{model}"
+        return branch
+
+    # =========================================================================
+    # CODEX AGENT ACTIVITY PARSING (JSONL sessions)
+    # =========================================================================
+
+    @staticmethod
+    def _extract_codex_action(entry: dict) -> Optional[str]:
+        """Extract a display action string from a Codex JSONL entry."""
+        entry_type = entry.get('type', '')
+        payload = entry.get('payload', {})
+
+        if entry_type == 'event_msg':
+            event_type = payload.get('type', '')
+            if event_type == 'agent_message':
+                text = payload.get('text', '') or payload.get('message', '')
+                if text:
+                    return f"💬 {str(text)[:120]}"
+                return '💬 Agent response'
+            if event_type == 'user_message':
+                return '📩 User message'
+            if event_type == 'token_count':
+                return '💭 Thinking'
+            if event_type == 'task_started':
+                return '🚀 Task started'
+            if event_type == 'task_complete':
+                return '✅ Task complete'
+            return None
+
+        if entry_type == 'response_item':
+            item = payload.get('item', payload)
+            item_type = item.get('type', '')
+            if item_type == 'function_call':
+                name = item.get('name', 'tool')
+                args = item.get('arguments', '')
+                if isinstance(args, str) and len(args) > 80:
+                    args = args[:80] + '...'
+                return f"🔧 {name}"
+            if item_type == 'function_call_output':
+                return None  # Skip output events
+            if item_type == 'message':
+                content = item.get('content', [])
+                if isinstance(content, list):
+                    for part in content:
+                        if isinstance(part, dict):
+                            text = part.get('text', '')
+                            if text:
+                                return f"💬 {text[:120]}"
+                return '💬 Agent response'
+            return None
+
+        if entry_type == 'session_meta':
+            return None  # Skip metadata
+
+        if entry_type == 'turn_context':
+            return None  # Skip context snapshots
+
+        return None
+
+    # =========================================================================
+    # GEMINI AGENT ACTIVITY PARSING (full JSON sessions)
+    # =========================================================================
+
+    @staticmethod
+    def _extract_gemini_action(message: dict) -> Optional[str]:
+        """Extract a display action string from a Gemini session message."""
+        msg_type = message.get('type', '')
+
+        if msg_type == 'user':
+            return '📩 User message'
+
+        if msg_type == 'gemini':
+            # Check for tool calls first
+            tool_calls = message.get('toolCalls', [])
+            if tool_calls:
+                last_tool = tool_calls[-1]
+                name = last_tool.get('displayName', last_tool.get('name', 'tool'))
+                return f"🔧 {name}"
+
+            # Check for thoughts
+            thoughts = message.get('thoughts', [])
+            if thoughts:
+                return '💭 Thinking'
+
+            # Text response
+            content = message.get('content', [])
+            if isinstance(content, list) and content:
+                for part in content:
+                    if isinstance(part, dict):
+                        text = part.get('text', '').strip()
+                        if text:
+                            return f"💬 {text[:120]}"
+            elif isinstance(content, str) and content.strip():
+                return f"💬 {content.strip()[:120]}"
+
+            return '💬 Agent response'
+
+        return None
+
+    def _parse_gemini_activity(self, file_path, branch):
+        """Parse Gemini session JSON to show agent actions.
+
+        Gemini rewrites the entire file on each change, so we track
+        the message count and only process new messages.
+        """
+        try:
+            path_key = str(file_path)
+            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
+                data = _json.load(f)
+
+            messages = data.get('messages', [])
+            last_count = self._jsonl_positions.get(path_key, 0)
+
+            if len(messages) <= last_count:
+                return True  # No new messages
+
+            self._jsonl_positions[path_key] = len(messages)
+
+            # Extract model from gemini messages
+            for msg in messages:
+                model = msg.get('model', '')
+                if model:
+                    self._session_models[path_key] = self._shorten_model(model)
+                    break
+
+            # Process new messages (most recent first for dedup)
+            new_messages = messages[last_count:]
+            for msg in reversed(new_messages):
+                action_text = self._extract_gemini_action(msg)
+                if not action_text:
+                    continue
+
+                if self._last_agent_action.get(path_key) == action_text:
+                    return True
+                self._last_agent_action[path_key] = action_text
+
+                tagged_branch = self._tag_branch_with_model(path_key, branch)
+                evt = MonitoringEvent(
+                    priority=1, event_type='agent', branch=tagged_branch,
+                    action='activity', message=action_text, level='info',
+                )
+                if self._event_queue:
+                    self._event_queue.enqueue(evt)
+                return True
+
+            return True
+        except (_json.JSONDecodeError, OSError) as e:
+            logger.info(f"[monitor] Gemini parse error for {file_path.name}: {e}")
+            return False
+
+    def _parse_codex_activity(self, file_path, branch):
+        """Parse Codex session JSONL to show agent actions.
+
+        Same JSONL tail-and-parse approach as Claude Code, but uses
+        _extract_codex_action() for Codex-specific event types.
+        """
+        try:
+            path_key = str(file_path)
+            current_size = file_path.stat().st_size
+            last_pos = self._jsonl_positions.get(path_key, 0)
+
+            if current_size < last_pos:
+                last_pos = 0
+            if current_size <= last_pos:
+                return True
+
+            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
+                f.seek(last_pos)
+                new_data = f.read()
+                self._jsonl_positions[path_key] = f.tell()
+
+            lines = [l for l in new_data.strip().split('\n') if l.strip()]
+            if not lines:
+                return True
+
+            for line in reversed(lines):
+                try:
+                    entry = _json.loads(line)
+                except _json.JSONDecodeError as e:
+                    logger.info(f"[monitor] Skipping malformed Codex JSONL: {e}")
+                    continue
+
+                # Extract model from session_meta or response_item
+                if entry.get('type') == 'session_meta':
+                    model = entry.get('payload', {}).get('model', '')
+                    if model:
+                        self._session_models[path_key] = self._shorten_model(model)
+                elif entry.get('type') == 'response_item':
+                    model = entry.get('payload', {}).get('item', {}).get('model', '')
+                    if model:
+                        self._session_models[path_key] = self._shorten_model(model)
+
+                action_text = self._extract_codex_action(entry)
+                if not action_text:
+                    continue
+
+                if self._last_agent_action.get(path_key) == action_text:
+                    return True
+                self._last_agent_action[path_key] = action_text
+
+                tagged_branch = self._tag_branch_with_model(path_key, branch)
+                evt = MonitoringEvent(
+                    priority=1, event_type='agent', branch=tagged_branch,
+                    action='activity', message=action_text, level='info',
+                )
+                if self._event_queue:
+                    self._event_queue.enqueue(evt)
+                return True
+
+            return True
+        except Exception as e:
+            logger.info(f"[monitor] Codex JSONL parse error for {file_path.name}: {e}")
+            return False
+
+    # =========================================================================
+    # CLAUDE CODE AGENT ACTIVITY PARSING (JSONL sessions)
+    # =========================================================================
+
     def _parse_agent_activity(self, file_path, branch):
         """Parse Claude Code session JSONL to show agent actions.
 
@@ -214,6 +475,11 @@ class MonitoringFileHandler(FileSystemEventHandler):
                     logger.info(f"[monitor] Skipping malformed JSONL line: {e}")
                     continue
 
+                # Extract model for display tag
+                model = self._extract_model_from_entry(entry)
+                if model:
+                    self._session_models[path_key] = self._shorten_model(model)
+
                 action_text = self._extract_action_from_entry(entry)
                 if not action_text:
                     continue
@@ -222,8 +488,9 @@ class MonitoringFileHandler(FileSystemEventHandler):
                     return True
                 self._last_agent_action[path_key] = action_text
 
+                tagged_branch = self._tag_branch_with_model(path_key, branch)
                 evt = MonitoringEvent(
-                    priority=1, event_type='agent', branch=branch,
+                    priority=1, event_type='agent', branch=tagged_branch,
                     action='activity', message=action_text, level='info',
                 )
                 if self._event_queue:
@@ -273,6 +540,18 @@ class MonitoringFileHandler(FileSystemEventHandler):
                 if self._parse_agent_activity(file_path, branch):
                     return
 
+            # Codex JSONL files: parse agent activity
+            if file_path.suffix == '.jsonl' and '.codex/sessions/' in path_str:
+                codex_branch = 'CODEX'
+                if self._parse_codex_activity(file_path, codex_branch):
+                    return
+
+            # Gemini JSON session files: parse agent activity
+            if file_path.suffix == '.json' and '.gemini/tmp/' in path_str and '/chats/' in path_str:
+                gemini_branch = 'GEMINI'
+                if self._parse_gemini_activity(file_path, gemini_branch):
+                    return
+
             self._check_command_indicator(action, file_path, branch)
 
             priority_level = get_priority(file_path, action)

src/aipass/prax/apps/handlers/monitoring/monitoring_filters.py

@@ -317,6 +317,12 @@ def should_monitor(path: Path) -> bool:
     if '.claude.json.backup' in name or '.claude.json.tmp' in name:
         return False
 
+    # CLI session files: always monitor (override ignore patterns for .codex, .gemini)
+    if '.codex/sessions/' in path_str and name.endswith('.jsonl'):
+        return True
+    if '.gemini/tmp/' in path_str and '/chats/' in path_str and name.endswith('.json'):
+        return True
+
     if _matches_always_patterns(path_str, parts, name):
         return True
 

src/aipass/prax/apps/handlers/monitoring/unified_stream.py

@@ -123,9 +123,11 @@ def print_event(event_type: str, branch: str, message: str, level: str = 'info',
         timestamp = datetime.now().strftime("%H:%M:%S")
 
         # Get branch color (unique per branch)
-        # For subagent labels like 'DEVPULSE AGENT', use base branch color
+        # Strip suffixes for color lookup: 'DEVPULSE AGENT' → DEVPULSE, 'DEVPULSE/opus' → DEVPULSE
         branch_upper = branch.upper()
-        base_branch = branch_upper[:-6] if branch_upper.endswith(' AGENT') else branch_upper
+        base_branch = branch_upper.split('/')[0]  # strip model tag
+        if base_branch.endswith(' AGENT'):
+            base_branch = base_branch[:-6]
         branch_color = BRANCH_COLORS.get(base_branch, 'white')
 
         # Format branch label with optional PID

src/aipass/prax/apps/modules/monitor.py

@@ -332,11 +332,19 @@ def _get_watch_directories(repo_root: Path) -> list[tuple[Path, bool]]:
         except (ValueError, OSError) as e:
             logger.warning(f"[monitor] Failed to read registry: {e}")
 
-    # Watch Claude Code project sessions for agent activity tracking
+    # Watch CLI session directories for agent activity tracking
     claude_projects = Path.home() / ".claude" / "projects"
     if claude_projects.exists():
         dirs.append((claude_projects, True))
 
+    codex_sessions = Path.home() / ".codex" / "sessions"
+    if codex_sessions.exists():
+        dirs.append((codex_sessions, True))
+
+    gemini_tmp = Path.home() / ".gemini" / "tmp"
+    if gemini_tmp.exists():
+        dirs.append((gemini_tmp, True))
+
     return dirs
 
 
@@ -352,6 +360,19 @@ def _emit_watcher_event(level: str, message: str) -> None:
     ))
 
 
+def _inotify_fix_message(err: OSError) -> str:
+    """Return the correct sysctl fix for the specific inotify limit hit."""
+    import errno as _errno
+    if err.errno == _errno.ENOSPC:  # Errno 28 — max_user_watches
+        return ("inotify watch limit reached (max_user_watches). "
+                "Fix: sudo sysctl -w fs.inotify.max_user_watches=524288")
+    elif err.errno == _errno.EMFILE:  # Errno 24 — max_user_instances
+        return ("inotify instance limit reached (max_user_instances). "
+                "Fix: sudo sysctl -w fs.inotify.max_user_instances=1024")
+    else:
+        return f"inotify error ({err}). Check system inotify limits."
+
+
 def _start_observer_with_fallback(handler, watch_dirs):
     """Start watchdog observer, falling back to polling on inotify failure.
 
@@ -367,10 +388,9 @@ def _start_observer_with_fallback(handler, watch_dirs):
         observer.start()
         return observer
     except OSError as e:
+        fix_msg = _inotify_fix_message(e)
         logger.warning(f"[monitor] inotify unavailable: {e} — switching to polling")
-        _emit_watcher_event('warning',
-            "File watcher: inotify watch limit reached (VSCode/editors consume most of the 65K default). "
-            "Using polling fallback (slower). Fix: sudo sysctl -w fs.inotify.max_user_watches=524288")
+        _emit_watcher_event('warning', f"File watcher: {fix_msg} Using polling fallback (slower).")
 
     try:
         from watchdog.observers.polling import PollingObserver
@@ -435,10 +455,9 @@ def _start_log_watcher_with_fallback(event_queue) -> bool:
         start_log_watcher(event_queue)
         return True
     except OSError as e:
+        fix_msg = _inotify_fix_message(e)
         logger.warning(f"[monitor] Log watcher inotify failed: {e} — switching to polling")
-        _emit_watcher_event('warning',
-            "Log watcher: inotify watch limit reached (VSCode/editors consume most of the 65K default). "
-            "Using polling fallback (slower). Fix: sudo sysctl -w fs.inotify.max_user_watches=524288")
+        _emit_watcher_event('warning', f"Log watcher: {fix_msg} Using polling fallback (slower).")
 
     try:
         start_log_watcher(event_queue, use_polling=True)