bug fix missing manifest

Focus: UI stability and efficient rendering.
   * Markdown Rendering Rewrite: Removed the dependency on rich.live.Live (which caused flickering and
     high CPU usage by constantly re-rendering the entire panel).
   * New BlockMarkdownRenderer: Implemented in printer.py (alias IncrementalMarkdownParser), it
     accumulates text in a buffer and only prints to the screen when it detects a complete block (e.g.,
     line breaks or
  ` code blocks).
   * UI Optimizations (terminal_ui.py & stubs.py): Waiting spinners now stop cleanly, and the UI
     transitions smoothly to block printing. Fixed visual truncation issues in the bottom "Tab prompt"
     bar for excessively long commands.

  📦 2. Commit History (Last 4 Commits)

  9446baf - improve ai rules
   * Strict Anti-Hallucination (ai.py): Injected MANDATORY rules into the System Prompt for both
     Architect and Engineer agents. Now, if the terminal buffer is empty or only contains an idle prompt
     (e.g., iol#), the AI is strictly forced to state that it lacks data instead of inventing topologies
     or configurations.
   * Language Preference: Explicitly instructed agents to always respond in the same language the user
     used to ask the question.

  64377f7 - move context block logic to server and improvements
   * Context Precision (ai_service.py): Moved context partitioning logic to the service/server side. It
     now calculates exact start_pos and end_pos based on identified commands, preventing mixed outputs
     or residual text from bleeding into the AI's prompt.
   * Token Savings (server.py): The server now selectively strips garbage metadata and UI caches that
     add no value to the LLM before sending the payload over the wire.

  e4fd1ad - fix logclean for 6wind
   * Full ANSI/CSI Support (utils.py): Replaced the legacy rigid escape filter with a complete CSI
     (Control Sequence Introducer) parser. The client can now accurately process numeric cursor
     movements (C, D), inline dynamic erasures (K), and absolute shifts (G), ensuring connpy understands
     exactly what a 6WIND router or other VNFs render on the screen without garbage characters.

  b0a914a - updates al copilot
   * Copilot <> gRPC Integration: Implemented the async plumbing required for the Copilot to work over
     the gRPC tunnel (server.py & stubs.py).
   * Initial Streaming UI: Added the first iteration of chunk callbacks to provide real-time feedback
     while the LLM generates responses, laying the groundwork for the uncommitted block-renderer
     optimizations.

.gitignore

@@ -164,6 +164,8 @@ connpy_roadmap.md
 MULTI_USER_PLAN.md
 COPILOT_PLAN.md
 ARCHITECTURAL_DEBT_REFACTOR.md
+COPILOT_UI_FEATURES.md
+MULTI_USER_IMPLEMENTATION_STEPS.md
 
 #themes
 nord.yml

MANIFEST.in

@@ -3,6 +3,6 @@ include README.md
 include requirements.txt
 recursive-include connpy/core_plugins *
 recursive-include connpy/proto *
-recursive-include connpy/grpc *.proto
+recursive-include connpy/grpc_layer *
 recursive-exclude * __pycache__
 recursive-exclude * *.py[co]

connpy/_version.py

@@ -1 +1 @@
-__version__ = "6.0.0b8"
+__version__ = "6.0.0b11"

connpy/ai.py

@@ -205,6 +205,7 @@ class ai:
             - COMPLETE MISSIONS: Execute ALL steps of a mission before reporting back.
             - DIAGRAM: Use ASCII art or Unicode box-drawing characters directly in your responses to visualize topologies or paths when helpful.
             - EVIDENCE: Include 'Key Snippets' from tool outputs. Be token-efficient.
+            - LANGUAGE: You MUST respond in the same language used by the user in their question or instruction.
             - NO WANDERING: Do not speculate. If stuck, report attempts.
             - SAFETY: When you use 'run_commands' with configuration commands, the system automatically prompts the user for confirmation. Just execute - don't ask permission first.
 {architect_instructions}
@@ -222,6 +223,7 @@ class ai:
             - ENGINEER CAPABILITIES: Your Engineer can:
                 * Filter nodes (list_nodes), Run CLI commands (run_commands), Get metadata (get_node_info).
             - ANALYSIS: Review technical findings to identify patterns or design failures.
+            - LANGUAGE: You MUST respond in the same language used by the user in their question or instruction.
             - MEMORY: Update long-term facts ONLY when the user explicitly requests it.
             
             CRITICAL - EFFICIENT DELEGATION:
@@ -297,7 +299,6 @@ class ai:
         - response: reconstructed ModelResponse (same as non-streaming)
         - streamed: True if text was rendered to console during streaming
         """
-        from rich.live import Live
 
         stream_resp = completion(model=model, messages=messages, tools=tools, api_key=api_key, stream=True, **kwargs)
 
@@ -305,7 +306,7 @@ class ai:
         full_content = ""
         is_streaming_text = False
         has_tool_calls = False
-        live_display = None
+        header_printed = False
 
         # Determine styling based on current brain
         role_label = "Network Architect" if "architect" in label.lower() else "Network Engineer"
@@ -334,7 +335,6 @@ class ai:
 
                     if not chunk_callback:
                         if not is_streaming_text:
-                            # Stop spinner definitively
                             if status:
                                 try:
                                     status.stop()
@@ -343,35 +343,28 @@ class ai:
                             
                             # Create a stable, direct Console to bypass _ConsoleProxy recreation bugs
                             from rich.console import Console as RichConsole
-                            from .printer import connpy_theme, get_original_stdout
+                            from rich.rule import Rule
+                            from .printer import connpy_theme, get_original_stdout, IncrementalMarkdownParser
                             stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
                             
-                            live_display = Live(
-                                Panel(Markdown(full_content), title=title, border_style=border, expand=False),
-                                console=stable_console,
-                                refresh_per_second=8,
-                                transient=False
-                            )
-                            live_display.start()
+                            stable_console.print(Rule(f"[bold {border}]{title}[/bold {border}]", style=border))
+                            header_printed = True
+                            md_parser = IncrementalMarkdownParser(console=stable_console)
                             is_streaming_text = True
-                        else:
-                            live_display.update(
-                                Panel(Markdown(full_content), title=title, border_style=border, expand=False)
-                            )
+                        
+                        md_parser.feed(delta.content)
         except Exception as e:
             if not chunks:
                 raise
         finally:
-            if live_display:
-                # Render final state with complete content
+            if header_printed:
                 try:
-                    live_display.update(
-                        Panel(Markdown(full_content), title=title, border_style=border, expand=False)
-                    )
-                except Exception:
-                    pass
-                try:
-                    live_display.stop()
+                    md_parser.flush()
+                    from rich.console import Console as RichConsole
+                    from rich.rule import Rule
+                    from .printer import connpy_theme, get_original_stdout
+                    stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
+                    stable_console.print(Rule(style=border))
                 except Exception:
                     pass
         
@@ -1320,11 +1313,13 @@ class ai:
         if persona == "architect":
             system_prompt = f"""Role: NETWORK ARCHITECT. You act as a senior strategic advisor during a live SSH session.
 Rules:
-1. Answer the user's question directly based on the Terminal Context.
-2. Focus on the "why" and "how". Analyze topologies, design patterns, and validate configurations.
-3. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
-4. Keep your guide concise and authoritative.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. Answer the user's question directly and EXCLUSIVELY based on the Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like 'iol#' or 'admin@vrouter>') and no command output, it means YOU DON'T HAVE DATA. In this case, YOU MUST NOT invent any information.
+4. Focus on the "why" and "how". Analyze topologies, design patterns, and validate configurations.
+5. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
+6. Keep your guide concise and authoritative.
+7. You MUST output your response in the following strict format:
 <guide>
 Your brief tactical guide in markdown.
 </guide>
@@ -1333,7 +1328,7 @@ Your brief tactical guide in markdown.
 <risk>
 low
 </risk>
-6. Risk level is usually "low" for read-only/no commands.
+8. Risk level is usually "low" for read-only/no commands.
 
 Terminal Context:
 {terminal_buffer}
@@ -1343,11 +1338,13 @@ Node: {node_name}"""
         else:
             system_prompt = f"""Role: TERMINAL COPILOT. You assist a network engineer during a live SSH session.
 Rules:
-1. Answer the user's question directly based on the Terminal Context.
-2. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the <guide> section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
-3. If the user wants to execute an action, provide the required CLI commands inside a <commands> block, one command per line. If no commands are needed, leave it empty or omit the block.
-4. ULTRA-CONCISE. Keep your guide to the point.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. EXTREMELY IMPORTANT: Answer EXCLUSIVELY based on the provided Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like 'iol#' or 'admin@vrouter>') and no command output, it means YOU DON'T HAVE DATA. In this case, YOU MUST NOT invent any information. Instead, explicitly state that you don't see the data and offer the correct CLI commands to retrieve it.
+4. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the <guide> section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
+5. If the user wants to execute an action, provide the required CLI commands inside a <commands> block, one command per line. If no commands are needed, leave it empty or omit the block.
+6. ULTRA-CONCISE. Keep your guide to the point.
+7. You MUST output your response in the following strict format:
 <guide>
 Your brief tactical guide in markdown. 3-4 sentences max.
 </guide>
@@ -1358,7 +1355,7 @@ command 2
 <risk>
 low, high, or destructive
 </risk>
-6. Risk level: "low" for read-only/no commands, "high" for config changes, "destructive" for potentially dangerous ops.
+8. Risk level: "low" for read-only/no commands, "high" for config changes, "destructive" for potentially dangerous ops.
 
 Terminal Context:
 {terminal_buffer}
@@ -1400,6 +1397,7 @@ Node: {node_name}"""
         try:
             while iteration < max_iterations:
                 iteration += 1
+
                 response = await acompletion(
                     model=current_model,
                     messages=messages,

connpy/cli/terminal_ui.py

@@ -12,7 +12,6 @@ from textwrap import dedent
 from rich.console import Console
 from rich.panel import Panel
 from rich.markdown import Markdown
-from rich.live import Live
 from prompt_toolkit import PromptSession
 from prompt_toolkit.key_binding import KeyBindings
 from prompt_toolkit.formatted_text import HTML
@@ -57,9 +56,10 @@ class CopilotInterface:
 
     async def run_session(self, 
                           raw_bytes: bytes, 
-                          cmd_byte_positions: List[tuple], 
                           node_info: dict,
-                          on_ai_call: Callable):
+                          on_ai_call: Callable,
+                          cmd_byte_positions: List[tuple] = None, 
+                          blocks: List[tuple] = None):
         """
         Runs the interactive Copilot session.
         on_ai_call: async function(active_buffer, question) -> result_dict
@@ -69,9 +69,11 @@ class CopilotInterface:
         try:
             # Prepare UI state
             buffer = log_cleaner(raw_bytes.decode(errors='replace'))
-            blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info)
-            last_line = buffer.split('\n')[-1].strip() if buffer.strip() else "(prompt)"
-            blocks.append((len(raw_bytes), last_line[:80]))
+            
+            # Use pre-calculated blocks if provided (remote mode), otherwise calculate locally (local mode)
+            if blocks is None:
+                last_line = buffer.split('\n')[-1].strip() if buffer.strip() else "(prompt)"
+                blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info, last_line=last_line)
             
             state = {
                 'context_cmd': 1,
@@ -88,7 +90,7 @@ class CopilotInterface:
             self.console.print("") # Salto de línea real
             self.console.print(Rule(title="[bold cyan] AI TERMINAL COPILOT [/bold cyan]", style="cyan"))
             self.console.print(Panel(
-                "[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel.\n"
+                "[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel. Type / for commands.\n"
                 "Tab to change context mode. Ctrl+\u2191/\u2193 to adjust context. \u2191\u2193 for question history.[/dim]",
                 border_style="cyan"
             ))
@@ -128,9 +130,8 @@ class CopilotInterface:
                 if state['context_mode'] == self.mode_lines:
                     return '\n'.join(buffer.split('\n')[-state['context_lines']:])
                 idx = max(0, state['total_cmds'] - state['context_cmd'])
-                start, preview = blocks[idx]
-                if state['context_mode'] == self.mode_single and idx + 1 < state['total_cmds']:
-                    end = blocks[idx + 1][0]
+                start, end, preview = blocks[idx]
+                if state['context_mode'] == self.mode_single:
                     active_raw = raw_bytes[start:end]
                 else:
                     active_raw = raw_bytes[start:]
@@ -172,7 +173,39 @@ class CopilotInterface:
                     base_str = f'\u25b6 Ctrl+\u2191/\u2193 adjusts by 50 lines  [Tab: {m_label}]'
                 else:
                     idx = max(0, state['total_cmds'] - state['context_cmd'])
-                    desc = blocks[idx][1]
+                    
+                    def clean_preview(text):
+                        # Limpia saltos de línea y el prompt inicial (todo hasta #, > o $) para que quede solo el comando
+                        original = text.strip().replace('\r', '').replace('\n', ' ')
+                        cleaned = re.sub(r'^.*?[#>\$]\s*', '', original)
+                        # Si limpiar el prompt nos deja con un string vacío (ej: era solo "iol#"), devolvemos el original
+                        return cleaned if cleaned else original
+
+                    if state['context_mode'] == self.mode_range:
+                        range_blocks = blocks[idx:]
+                        # Si hay más de un bloque, el último es siempre el prompt vacío/actual. Lo omitimos visualmente.
+                        if len(range_blocks) > 1:
+                            range_blocks = range_blocks[:-1]
+                            
+                        # Limpiar y truncar comandos muy largos para que no rompan la UI
+                        previews = []
+                        for b in range_blocks:
+                            p = clean_preview(b[2])
+                            if p:
+                                # Truncar comandos individuales largos
+                                if len(p) > 25: p = p[:22] + "..."
+                                previews.append(p)
+                        
+                        if not previews:
+                            desc = clean_preview(blocks[idx][2])
+                        elif len(previews) <= 3:
+                            desc = " + ".join(previews)
+                        else:
+                            desc = f"{previews[0]} + {previews[1]} + {previews[2]} ... (+{len(previews)-3})"
+                    else:
+                        # Modo SINGLE original
+                        desc = clean_preview(blocks[idx][2])
+                        
                     base_str = f'\u25b6 {desc}  [Tab: {m_label}]'
                 
                 # Wrap base_str in a style to maintain consistency and avoid glitches
@@ -299,39 +332,66 @@ class CopilotInterface:
                 # Use persona from overrides (one-shot) or from session state
                 active_persona = merged_node_info.get('persona', self.session_state.get('persona', 'engineer'))
                 persona_color = self._get_theme_color(active_persona, fallback="cyan")
+                persona_title = "Network Architect" if active_persona == "architect" else "Network Engineer"
                 
                 active_buffer = get_active_buffer()
-                live_text = "Thinking..."
-                panel = Panel(live_text, title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color)
+                live_text = ""
+                first_chunk = True
+                
+                from rich.rule import Rule
+                from rich.status import Status
+                from connpy.printer import IncrementalMarkdownParser
+                
+                md_parser = IncrementalMarkdownParser(console=self.console)
+                
+                status_spinner = Status(
+                    f"[bold {persona_color}]{persona_title}:[/bold {persona_color}] [dim]Thinking...[/dim]",
+                    console=self.console,
+                    spinner="dots"
+                )
+                status_spinner.start()
                 
                 def on_chunk(text):
-                    nonlocal live_text
-                    if live_text == "Thinking...": live_text = ""
+                    nonlocal live_text, first_chunk
+                    if first_chunk:
+                        status_spinner.stop()
+                        # Print header rule before first chunk arrives
+                        self.console.print(Rule(
+                            f"[bold {persona_color}]{persona_title}[/bold {persona_color}]",
+                            style=persona_color
+                        ))
+                        first_chunk = False
                     live_text += text
+                    md_parser.feed(text)
                 
-                with Live(panel, console=self.console, refresh_per_second=10) as live:
-                    def update_live(t):
-                        live.update(Panel(Markdown(t), title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color))
+                # Check for interruption during AI call
+                ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, on_chunk, merged_node_info))
                 
-                    wrapped_chunk = lambda t: (on_chunk(t), update_live(live_text))
-                    
-                    # Check for interruption during AI call
-                    ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, wrapped_chunk, merged_node_info))
-                    
-                    try:
-                        while not ai_task.done():
-                            await asyncio.sleep(0.05)
-                        result = await ai_task
-                    except asyncio.CancelledError:
-                        return "cancel", None, None
-
-                if not result or result.get("error"):
-                    if result and result.get("error"): self.console.print(f"[red]Error: {result['error']}[/red]")
+                try:
+                    while not ai_task.done():
+                        await asyncio.sleep(0.05)
+                    result = await ai_task
+                except asyncio.CancelledError:
+                    status_spinner.stop()
                     return "cancel", None, None
                 
-                # 4. Handle result
-                if live_text == "Thinking..." and result.get("guide"):
-                    self.console.print(Panel(Markdown(result["guide"]), title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color))
+                # Ensure spinner is stopped if no chunks arrived
+                if first_chunk:
+                    status_spinner.stop()
+
+                # Close the streamed output with a Rule
+                if not first_chunk:
+                    md_parser.flush()
+                    self.console.print(Rule(style=persona_color))
+                
+                if not result or result.get("error"):
+                    if first_chunk and result and result.get("error"):
+                        self.console.print(f"[red]Error: {result['error']}[/red]")
+                    return "cancel", None, None
+
+                # If no chunks were streamed but we have a guide, print it as a panel
+                if first_chunk and result and result.get("guide"):
+                    self.console.print(Panel(Markdown(result["guide"]), title=f"[bold {persona_color}]{persona_title}[/bold {persona_color}]", border_style=persona_color))
 
                 commands = result.get("commands", [])
                 if not commands:
@@ -434,5 +494,4 @@ class CopilotInterface:
 
         finally:
             state['cancelled'] = True
-            self.console.print("[dim]Returning to session...[/dim]")
 

connpy/core.py

@@ -211,6 +211,7 @@ class node:
         self.output = ""
         self.status = 1
         self.result = {}
+        self.cmd_byte_positions = [(0, None)]
 
     @MethodHook
     def _passtx(self, passwords, *, keyfile=None):
@@ -385,9 +386,9 @@ class node:
             loop = asyncio.get_running_loop()
             child_reader_queue = asyncio.Queue()
             
-            # Track command byte positions for copilot context navigation
+            # Reset and track command byte positions for copilot context navigation
             # Each entry is (byte_position, command_text_or_None)
-            cmd_byte_positions = [(0, None)]
+            self.cmd_byte_positions = [(self.mylog.tell() if hasattr(self, 'mylog') else 0, None)]
             
             def _child_read_ready():
                 try:
@@ -428,7 +429,7 @@ class node:
                             node_info["prompt"] = to_str(self.tags.get("prompt", r'>$|#$|\$$|>.$|#.$|\$.$'))
                         
                         # Invoke copilot (async callback handles UI)
-                        await copilot_handler(self.mylog.getvalue(), node_info, local_stream, child_fd, cmd_byte_positions)
+                        await copilot_handler(self.mylog.getvalue(), node_info, local_stream, child_fd, self.cmd_byte_positions)
                         continue
                     
                     # Remove any stray \x00 bytes and forward normally
@@ -436,10 +437,9 @@ class node:
                     if clean_data:
                         # Track command boundaries when user hits Enter
                         if hasattr(self, 'mylog') and (b'\r' in clean_data or b'\n' in clean_data):
-                            cmd_byte_positions.append((self.mylog.tell(), None))
+                            self.cmd_byte_positions.append((self.mylog.tell(), None))
 
-                        try:
-                            os.write(child_fd, clean_data)
+                        try:                            os.write(child_fd, clean_data)
                         except OSError:
                             break
                         self.lastinput = time()
@@ -561,6 +561,45 @@ class node:
         finally:
             local_stream.teardown()
 
+    @MethodHook
+    async def inject_commands(self, commands, child_fd, on_inject=None):
+        """
+        Inject a list of commands into the node's PTY.
+        Handles screen_length_command, history tracking and delays.
+        """
+        if not commands:
+            return
+
+        # 0. Clear line
+        os.write(child_fd, b'\x15')
+        await asyncio.sleep(0.1)
+
+        # 1. Prepare list (prepend screen_length if exists)
+        slc = self.tags.get("screen_length_command") if hasattr(self, 'tags') and isinstance(self.tags, dict) else None
+        
+        to_send = list(commands)
+        if slc and slc not in to_send: # avoid duplicates if already there
+             to_send.insert(0, slc)
+
+        # 2. Inject one by one
+        for cmd in to_send:
+            # Register in node's official history (SKIP if it's the administrative screen length command)
+            if cmd != slc and hasattr(self, 'cmd_byte_positions') and self.cmd_byte_positions is not None:
+                log_pos = self.mylog.tell() if hasattr(self, 'mylog') else 0
+                self.cmd_byte_positions.append((log_pos, cmd))
+            
+            # Write physically to PTY
+            os.write(child_fd, (cmd + "\n").encode())
+            
+            # Notify (e.g., for gRPC or logs) - SKIP for administrative SLC
+            if on_inject and cmd != slc:
+                if asyncio.iscoroutinefunction(on_inject):
+                    await on_inject(cmd)
+                else:
+                    on_inject(cmd)
+            
+            # Delay to avoid overwhelming the router
+            await asyncio.sleep(0.8)
 
     @MethodHook
     def interact(self, debug=False, logger=None):
@@ -642,7 +681,7 @@ class node:
                         while True:
                             action, commands, custom_cmd = await interface.run_session(
                                 raw_bytes=raw_bytes,
-                                cmd_byte_positions=cmd_byte_positions,
+                                cmd_byte_positions=self.cmd_byte_positions,
                                 node_info=node_info,
                                 on_ai_call=on_ai_call
                             )
@@ -650,6 +689,7 @@ class node:
                                 continue
                             break
                 finally:
+                    print("\033[2m Returning to session...\033[0m", flush=True)
                     # Reiniciar el lector de la terminal para volver al modo interactivo SSH/Telnet
                     if hasattr(stream, 'start_reading'):
                         stream.start_reading()
@@ -658,20 +698,7 @@ class node:
                 
                 if action in ("send_all", "custom"):
                     cmds_to_send = commands if action == "send_all" else custom_cmd
-                    
-                    if cmds_to_send:
-                        os.write(child_fd, b'\x15') # Ctrl+U
-                        await asyncio.sleep(0.1)
-                        
-                        # Prepend screen length command to avoid pagination
-                        if "screen_length_command" in self.tags:
-                            cmds_to_send.insert(0, self.tags["screen_length_command"])
-
-                        for cmd in cmds_to_send:
-                            if cmd_byte_positions is not None:
-                                cmd_byte_positions.append((self.mylog.tell(), cmd))
-                            os.write(child_fd, (cmd + "\n").encode())
-                            await asyncio.sleep(0.8)
+                    await self.inject_commands(cmds_to_send, child_fd)
                 else:
                     os.write(child_fd, b'\x15\r')
             except Exception as e:

connpy/grpc_layer/server.py

@@ -207,15 +207,34 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                     import json
                     import asyncio
                     import os
+                    from ..services.ai_service import AIService
+                    
+                    service = AIService(self.service.config)
 
                     if node_info is None:
                         node_info = {}
                     
+                    # Calculate real command blocks from history using the central service
+                    raw_bytes = n.mylog.getvalue() if hasattr(n, 'mylog') else buffer
+                    if not isinstance(raw_bytes, bytes):
+                        raw_bytes = str(raw_bytes).encode()
+                    
+                    from connpy.utils import log_cleaner
+                    last_line = log_cleaner(raw_bytes.decode(errors='replace')).split('\n')[-1].strip()
+                    blocks = service.build_context_blocks(raw_bytes, n.cmd_byte_positions, node_info, last_line=last_line)
+                    node_info["context_blocks"] = blocks
+
                     node_info_json = json.dumps(node_info)
                     
                     # Convert buffer to string if it's bytes for the preview
                     preview_str = buffer[-200:].decode(errors='replace') if isinstance(buffer, bytes) else str(buffer)[-200:]
                     
+                    # Generate a unique session ID for this copilot interaction to prevent race conditions
+                    import uuid
+                    copilot_session_id = str(uuid.uuid4())
+                    node_info["session_id"] = copilot_session_id
+                    node_info_json = json.dumps(node_info)
+                    
                     # 1. Send prompt to client
                     response_queue.put(connpy_pb2.InteractResponse(
                         copilot_prompt=True,
@@ -224,6 +243,13 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                     ))
 
                     while True:
+                        # 0. Drain the queue of any stale messages before starting a new interaction
+                        while not remote_stream.copilot_queue.empty():
+                            try:
+                                remote_stream.copilot_queue.get_nowait()
+                            except:
+                                break
+
                         # 2. Await the question from client via the copilot_queue
                         import threading
                         def preload_ai_deps():
@@ -236,8 +262,17 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                         try:
                             req_data = await asyncio.wait_for(remote_stream.copilot_queue.get(), timeout=120)
                             if not req_data: return
-                            if "question" not in req_data or not req_data["question"] or req_data["question"] == "CANCEL" or req_data.get("action") == "cancel":
-                                os.write(child_fd, b'\x15\r')
+                            
+                            # Validate session ID if provided by client (skip validation if not provided for CLI compatibility)
+                            req_session_id = req_data.get("session_id")
+                            if req_session_id and req_session_id != copilot_session_id:
+                                continue # Ignore stale request from a previous session
+
+                            if "question" not in req_data or not req_data["question"] or req_data["question"] == "CANCEL" or req_data.get("action") in ("cancel", "web_cancel"):
+                                if req_data.get("action") == "web_cancel":
+                                    os.write(child_fd, b'\x05')
+                                else:
+                                    os.write(child_fd, b'\x15\r')
                                 return
                             question = req_data["question"]
                             
@@ -264,9 +299,6 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                             return
                             
                         # 3. Call AI Service with streaming
-                        from ..services.ai_service import AIService
-                        service = AIService(self.service.config)
-                        
                         def chunk_callback(chunk_text):
                             if chunk_text:
                                 response_queue.put(connpy_pb2.InteractResponse(
@@ -287,8 +319,11 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                         if wait_action_task in done:
                             req_data = wait_action_task.result()
                             ai_task.cancel()
-                            if req_data.get("action") == "cancel" or req_data.get("question") == "CANCEL":
-                                os.write(child_fd, b'\x15\r')
+                            if req_data.get("action") in ("cancel", "web_cancel") or req_data.get("question") == "CANCEL":
+                                if req_data.get("action") == "web_cancel":
+                                    os.write(child_fd, b'\x05')
+                                else:
+                                    os.write(child_fd, b'\x15\r')
                                 return
                             continue # Loop back instead of returning to keep session alive
                         else:
@@ -312,45 +347,27 @@ class NodeServicer(connpy_pb2_grpc.NodeServiceServicer):
                             if action == "continue":
                                 continue # Loop back for next question
                                 
-                            if action == "cancel":
-                                os.write(child_fd, b'\x15\r')
+                            if action in ("cancel", "web_cancel"):
+                                if action == "web_cancel":
+                                    os.write(child_fd, b'\x05')
+                                else:
+                                    os.write(child_fd, b'\x15\r')
                                 return
                         except asyncio.TimeoutError:
                             os.write(child_fd, b'\x15\r')
                             return
                             
+                        def on_inject(cmd):
+                            response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd))
+
                         if action == "send_all":
                             commands = result.get("commands", [])
-                            os.write(child_fd, b'\x15') # Ctrl+U to clear line
-                            await asyncio.sleep(0.1)
-
-                            # Prepend screen length command to avoid pagination
-                            if "screen_length_command" in n.tags:
-                                os.write(child_fd, (n.tags["screen_length_command"] + "\n").encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=n.tags["screen_length_command"]))
-                                await asyncio.sleep(0.8)
-
-                            for cmd in commands:
-                                os.write(child_fd, (cmd + "\n").encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd))
-                                await asyncio.sleep(0.8)
+                            await n.inject_commands(commands, child_fd, on_inject=on_inject)
                             return
                         elif action.startswith("custom:"):
-                            custom_cmds = action[7:]
-                            os.write(child_fd, b'\x15')
-                            await asyncio.sleep(0.1)
-                            
-                            # Prepend screen length command to avoid pagination
-                            if "screen_length_command" in n.tags:
-                                os.write(child_fd, (n.tags["screen_length_command"] + "\n").encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=n.tags["screen_length_command"]))
-                                await asyncio.sleep(0.8)
-
-                            for cmd in custom_cmds.split('\n'):
-                                if cmd.strip():
-                                    os.write(child_fd, (cmd.strip() + "\n").encode())
-                                    response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd.strip()))
-                                    await asyncio.sleep(0.8)
+                            custom_cmds_raw = action[7:]
+                            custom_cmds = [cmd.strip() for cmd in custom_cmds_raw.split('\n') if cmd.strip()]
+                            await n.inject_commands(custom_cmds, child_fd, on_inject=on_inject)
                             return
                         else:
                             os.write(child_fd, b'\x15\r')

connpy/grpc_layer/stubs.py

@@ -43,7 +43,7 @@ class NodeStub:
         self.remote_host = remote_host
         self.config = config
 
-    def _handle_remote_copilot(self, res, request_queue, response_queue, client_buffer_bytes, cmd_byte_positions, pause_generator, resume_generator, old_tty):
+    def _handle_remote_copilot(self, res, request_queue, response_queue, client_buffer_bytes, pause_generator, resume_generator, old_tty):
         import json, asyncio, termios, sys, tty, queue
         from ..core import copilot_terminal_mode
         from . import connpy_pb2
@@ -51,6 +51,10 @@ class NodeStub:
         pause_generator()
         
         termios.tcsetattr(sys.stdin, termios.TCSADRAIN, old_tty)
+        
+        node_info = json.loads(res.copilot_node_info_json) if res.copilot_node_info_json else {}
+        blocks = node_info.get("context_blocks", [])
+
         interface = CopilotInterface(
             self.config, 
             history=getattr(self, 'copilot_history', None),
@@ -59,8 +63,6 @@ class NodeStub:
         self.copilot_history = interface.history
         self.copilot_state = interface.session_state
 
-        node_info = json.loads(res.copilot_node_info_json) if res.copilot_node_info_json else {}
-
         async def on_ai_call_remote(active_buffer, question, chunk_callback, merged_node_info):
             # Send request to server
             request_queue.put(connpy_pb2.InteractRequest(
@@ -85,9 +87,9 @@ class NodeStub:
             while True:
                 action, commands, custom_cmd = await interface.run_session(
                     raw_bytes=bytes(client_buffer_bytes),
-                    cmd_byte_positions=cmd_byte_positions,
                     node_info=node_info,
-                    on_ai_call=on_ai_call_remote
+                    on_ai_call=on_ai_call_remote,
+                    blocks=blocks
                 )
                 
                 if action == "continue":
@@ -100,6 +102,7 @@ class NodeStub:
         with copilot_terminal_mode():
             action, commands, custom_cmd = asyncio.run(run_remote_copilot())
         
+        print("\033[2m Returning to session...\033[0m", flush=True)
         # Prepare final action for server
         action_sent = "cancel"
         if action == "send_all" and commands:
@@ -124,7 +127,6 @@ class NodeStub:
         
         request_queue = queue.Queue()
         client_buffer_bytes = bytearray()
-        cmd_byte_positions = [(0, None)]
         pause_stdin = [False]
         wake_r, wake_w = os.pipe()
 
@@ -171,8 +173,6 @@ class NodeStub:
                         data = os.read(sys.stdin.fileno(), 1024)
                         if not data:
                             break
-                        if b'\r' in data or b'\n' in data:
-                            cmd_byte_positions.append((len(client_buffer_bytes), None))
                         yield connpy_pb2.InteractRequest(stdin_data=data)
                     except OSError:
                         break
@@ -246,14 +246,11 @@ class NodeStub:
                 if res.copilot_prompt:
                     self._handle_remote_copilot(
                         res, request_queue, response_queue, 
-                        client_buffer_bytes, cmd_byte_positions, 
+                        client_buffer_bytes, 
                         pause_generator, resume_generator, old_tty
                     )
                     continue
 
-                if res.copilot_injected_command:
-                    cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
                 if res.stdout_data:
                     os.write(sys.stdout.fileno(), res.stdout_data)
                     client_buffer_bytes.extend(res.stdout_data)
@@ -275,7 +272,6 @@ class NodeStub:
         params_json = json.dumps(connection_params)
         request_queue = queue.Queue()
         client_buffer_bytes = bytearray()
-        cmd_byte_positions = [(0, None)]
         pause_stdin = [False]
         wake_r, wake_w = os.pipe()
 
@@ -323,8 +319,6 @@ class NodeStub:
                         data = os.read(sys.stdin.fileno(), 1024)
                         if not data:
                             break
-                        if b'\r' in data or b'\n' in data:
-                            cmd_byte_positions.append((len(client_buffer_bytes), None))
                         yield connpy_pb2.InteractRequest(stdin_data=data)
                     except OSError:
                         break
@@ -397,14 +391,11 @@ class NodeStub:
                 if res.copilot_prompt:
                     self._handle_remote_copilot(
                         res, request_queue, response_queue, 
-                        client_buffer_bytes, cmd_byte_positions, 
+                        client_buffer_bytes, 
                         pause_generator, resume_generator, old_tty
                     )
                     continue
 
-                if res.copilot_injected_command:
-                    cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
                 if res.stdout_data:
                     os.write(sys.stdout.fileno(), res.stdout_data)
                     client_buffer_bytes.extend(res.stdout_data)
@@ -736,7 +727,6 @@ class AIStub:
         import queue
         from rich.prompt import Prompt
         from rich.text import Text
-        from rich.live import Live
         from rich.panel import Panel
         from rich.markdown import Markdown
         
@@ -767,7 +757,7 @@ class AIStub:
         responses = self.stub.ask(request_generator())
         
         full_content = ""
-        live_display = None
+        header_printed = False
         final_result = {"response": "", "chat_history": []}
 
         # Background thread to pull responses from gRPC into a local queue
@@ -832,69 +822,53 @@ class AIStub:
                 
                 if response.debug_message:
                     if debug:
-                        if live_display:
-                            try: live_display.stop()
-                            except: pass
                         if status:
                             try: status.stop()
                             except: pass
                         printer.console.print(Text.from_ansi(response.debug_message))
-                        if live_display:
-                            try: live_display.start()
-                            except: pass
-                        elif status:
+                        if status:
                             try: status.start()
                             except: pass
                     continue
                 
                 if response.important_message:
-                    if live_display:
-                        try: live_display.stop()
-                        except: pass
                     if status:
                         try: status.stop()
                         except: pass
                     printer.console.print(Text.from_ansi(response.important_message))
-                    if live_display:
-                        try: live_display.start()
-                        except: pass
-                    elif status:
+                    if status:
                         try: status.start()
                         except: pass
                     continue
 
                 if not response.is_final:
                     if response.text_chunk:
-                        full_content += response.text_chunk
-                        
-                        if not live_display:
+                        if not header_printed:
                             if status:
                                 try: status.stop()
                                 except: pass
                             
                             from rich.console import Console as RichConsole
-                            from ..printer import connpy_theme, get_original_stdout
+                            from rich.rule import Rule
+                            from ..printer import connpy_theme, get_original_stdout, IncrementalMarkdownParser
                             stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
                             
-                            # We default to Engineer title during stream, final result will correct it if needed
-                            live_display = Live(
-                                Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False),
-                                console=stable_console,
-                                refresh_per_second=8,
-                                transient=False
-                            )
-                            live_display.start()
-                        else:
-                            live_display.update(
-                                Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False)
-                            )
+                            # Print header on first chunk
+                            stable_console.print(Rule("[bold engineer]Network Engineer[/bold engineer]", style="engineer"))
+                            header_printed = True
+                            
+                            # Initialize parser
+                            md_parser = IncrementalMarkdownParser(console=stable_console)
+                        
+                        full_content += response.text_chunk
+                        md_parser.feed(response.text_chunk)
                     continue
                 
                 if response.is_final:
-                    if live_display:
-                        try: live_display.stop()
-                        except: pass
-                    # Final stop for status to ensure it disappears before the panel
+                    if header_printed:
+                        from rich.rule import Rule
+                        md_parser.flush()
+                        
                     if status:
                         try: status.stop()
                         except: pass
@@ -905,13 +879,14 @@ class AIStub:
                     role_label = "Network Architect" if responder == "architect" else "Network Engineer"
                     title = f"[bold {alias}]{role_label}[/bold {alias}]"
                     
-                    content_to_print = full_content or final_result.get("response", "")
-                    if content_to_print:
-                        if live_display:
-                            # Re-render the final frame with correct title/colors
-                            live_display.update(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
-                        else:
-                            printer.console.print(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
+                    if header_printed:
+                        from rich.console import Console as RichConsole
+                        from ..printer import connpy_theme, get_original_stdout
+                        stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
+                        stable_console.print(Rule(style=alias))
+                    elif not full_content and final_result.get("response"):
+                        # If nothing streamed but we have response (e.g. error or direct guide)
+                        printer.console.print(Panel(Markdown(final_result["response"]), title=title, border_style=alias, expand=False))
                     break
         except Exception as e:
             # Check if it was a gRPC error that we should let handle_errors catch

connpy/printer.py

@@ -15,7 +15,17 @@ class ThreadLocalStream:
     def write(self, data):
         stream = self._get_stream()
         if stream:
-            stream.write(data)
+            import time
+            retries = 0
+            while True:
+                try:
+                    stream.write(data)
+                    break
+                except BlockingIOError:
+                    if retries > 50:
+                        raise
+                    time.sleep(0.01)
+                    retries += 1
     
     def flush(self):
         stream = self._get_stream()
@@ -496,3 +506,74 @@ class _ThemeProxy:
         return getattr(local.theme, name)
 
 connpy_theme = _ThemeProxy()
+
+class BlockMarkdownRenderer:
+    """
+    Block-buffered streaming markdown renderer.
+    Accumulates text until block boundaries are detected,
+    then renders complete blocks using Rich's Markdown.
+    """
+    def __init__(self, console=None):
+        from rich.console import Console as RichConsole
+        from .printer import connpy_theme, get_original_stdout
+        self._console = console or RichConsole(
+            theme=connpy_theme, file=get_original_stdout()
+        )
+        self._line_buf = ""        # chars waiting for \n
+        self._block_lines = []     # complete lines for current block
+        self._in_code_block = False
+
+    def feed(self, text):
+        self._line_buf += text
+        while '\n' in self._line_buf:
+            idx = self._line_buf.index('\n')
+            line = self._line_buf[:idx + 1]
+            self._line_buf = self._line_buf[idx + 1:]
+            self._process_line(line)
+
+    def flush(self):
+        if self._line_buf:
+            self._block_lines.append(self._line_buf)
+            self._line_buf = ""
+        self._flush_block()
+
+    def _process_line(self, line):
+        stripped = line.strip()
+        
+        if stripped.startswith('```'):
+            if not self._in_code_block:
+                # Flush accumulated text before code block
+                self._flush_block()
+                self._in_code_block = True
+                self._block_lines.append(line)
+            else:
+                # Include closing fence and flush code block
+                self._block_lines.append(line)
+                self._in_code_block = False
+                self._flush_block()
+            return
+
+        if self._in_code_block:
+            self._block_lines.append(line)
+            return
+
+        # Blank line = paragraph break
+        if stripped == '':
+            self._block_lines.append(line)
+            self._flush_block()
+            return
+
+        self._block_lines.append(line)
+
+    def _flush_block(self):
+        if not self._block_lines:
+            return
+        block_text = ''.join(self._block_lines).strip()
+        self._block_lines = []
+        if not block_text:
+            return
+        from rich.markdown import Markdown
+        self._console.print(Markdown(block_text))
+
+# Alias for backward compatibility
+IncrementalMarkdownParser = BlockMarkdownRenderer

connpy/services/ai_service.py

@@ -6,10 +6,10 @@ from connpy.utils import log_cleaner
 class AIService(BaseService):
     """Business logic for interacting with AI agents and LLM configurations."""
 
-    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) -> list:
+    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict, last_line: str = "") -> list:
         """Identifies command blocks in the terminal history."""
         blocks = []
-        if not (cmd_byte_positions and len(cmd_byte_positions) >= 2 and raw_bytes):
+        if not raw_bytes:
             return blocks
             
         default_prompt = r'>$|#$|\$$|>.$|#.$|\$.$'
@@ -20,29 +20,63 @@ class AIService(BaseService):
         except Exception:
             prompt_re = re.compile(re.sub(r'(?<!\\)\$', '', default_prompt))
             
-        for i in range(1, len(cmd_byte_positions)):
-            pos, known_cmd = cmd_byte_positions[i]
-            prev_pos = cmd_byte_positions[i-1][0]
+        parsed_positions = []
+        if cmd_byte_positions and len(cmd_byte_positions) >= 1:
+            for i in range(1, len(cmd_byte_positions)):
+                pos, known_cmd = cmd_byte_positions[i]
+                prev_pos = cmd_byte_positions[i-1][0]
                 
-            if known_cmd:
-                prev_chunk = raw_bytes[prev_pos:pos]
-                prev_cleaned = log_cleaner(prev_chunk.decode(errors='replace'))
-                prev_lines = [l for l in prev_cleaned.split('\n') if l.strip()]
-                prompt_text = prev_lines[-1].strip() if prev_lines else ""
-                preview = f"{prompt_text}{known_cmd}" if prompt_text else known_cmd
-                blocks.append((pos, preview[:80]))
-            else:
-                chunk = raw_bytes[prev_pos:pos]
-                cleaned = log_cleaner(chunk.decode(errors='replace'))
-                lines = [l for l in cleaned.split('\n') if l.strip()]
-                preview = lines[-1].strip() if lines else ""
+                if known_cmd:
+                    prev_chunk = raw_bytes[prev_pos:pos]
+                    prev_cleaned = log_cleaner(prev_chunk.decode(errors='replace'))
+                    prev_lines = [l for l in prev_cleaned.split('\n') if l.strip()]
+                    prompt_text = prev_lines[-1].strip() if prev_lines else ""
+                    preview = f"{prompt_text}{known_cmd}" if prompt_text else known_cmd
+                    parsed_positions.append({"pos": pos, "type": "VALID_CMD", "preview": preview[:80]})
+                else:
+                    chunk = raw_bytes[prev_pos:pos]
+                    cleaned = log_cleaner(chunk.decode(errors='replace'))
+                    lines = [l for l in cleaned.split('\n') if l.strip()]
+                    preview = lines[-1].strip() if lines else ""
+                    
+                    if preview:
+                        match = prompt_re.search(preview)
+                        if match:
+                            cmd_text = preview[match.end():].strip()
+                            if cmd_text:
+                                parsed_positions.append({"pos": pos, "type": "VALID_CMD", "preview": preview[:80]})
+                            else:
+                                parsed_positions.append({"pos": pos, "type": "EMPTY_PROMPT", "preview": ""})
+                        else:
+                            parsed_positions.append({"pos": pos, "type": "SCROLLING", "preview": ""})
+                    else:
+                        parsed_positions.append({"pos": pos, "type": "SCROLLING", "preview": ""})
+
+        last_newline = raw_bytes.rfind(b'\n')
+        current_prompt_pos = last_newline + 1 if last_newline != -1 else 0
+        current_end = len(raw_bytes)
+
+        for i, item in enumerate(parsed_positions):
+            if item["type"] == "VALID_CMD":
+                start_pos = item["pos"]
+                preview = item["preview"]
+                
+                # Find the end position: next VALID_CMD or EMPTY_PROMPT
+                end_pos = current_prompt_pos
+                for j in range(i + 1, len(parsed_positions)):
+                    next_item = parsed_positions[j]
+                    if next_item["type"] in ("VALID_CMD", "EMPTY_PROMPT"):
+                        end_pos = next_item["pos"]
+                        break
+                
+                blocks.append((start_pos, end_pos, preview))
+
+        # Always ensure there is a final block representing the current prompt
+        if not blocks:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else "CURRENT CONTEXT"))
+        elif blocks[-1][0] < current_prompt_pos:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else "CURRENT CONTEXT"))
 
-                if preview:
-                    match = prompt_re.search(preview)
-                    if match:
-                        cmd_text = preview[match.end():].strip()
-                        if cmd_text:
-                            blocks.append((pos, preview[:80]))
         return blocks
 
     def process_copilot_input(self, input_text: str, session_state: dict) -> dict:

connpy/utils.py

@@ -23,14 +23,34 @@ def log_cleaner(data: str) -> str:
             elif token in ('\b', '\x7f'):
                 if cursor > 0:
                     cursor -= 1
-            elif token == '\x1B[D': # Left Arrow
-                if cursor > 0:
-                    cursor -= 1
-            elif token == '\x1B[C': # Right Arrow
-                if cursor < len(buffer):
-                    cursor += 1
-            elif token == '\x1B[K': # Clear to end of line
-                buffer = buffer[:cursor]
+            elif token.startswith('\x1B[') and len(token) >= 3:
+                # Parse CSI: \x1B[ <params> <final_char>
+                final = token[-1]
+                param_str = token[2:-1]
+                n = int(param_str) if param_str.isdigit() else 1
+
+                if final == 'D':   # CUB – Cursor Back
+                    cursor = max(0, cursor - n)
+                elif final == 'C': # CUF – Cursor Forward
+                    cursor = min(len(buffer), cursor + n)
+                elif final == 'K': # EL  – Erase in Line
+                    if n == 0 or param_str == '':      # Clear to end
+                        buffer = buffer[:cursor]
+                    elif n == 1:                         # Clear to start
+                        buffer[:cursor] = [' '] * cursor
+                    elif n == 2:                         # Clear entire line
+                        buffer = []
+                        cursor = 0
+                elif final == 'G': # CHA – Cursor Horizontal Absolute (1-indexed)
+                    cursor = max(0, n - 1)
+                    # Pad buffer if cursor is beyond current length
+                    if cursor > len(buffer):
+                        buffer.extend([' '] * (cursor - len(buffer)))
+                elif final == 'P': # DCH – Delete Characters
+                    del buffer[cursor:cursor + n]
+                elif final == '@': # ICH – Insert Characters
+                    buffer[cursor:cursor] = [' '] * n
+                # All other CSI sequences are silently discarded
             elif token.startswith('\x1B'):
                 continue
             elif len(token) == 1 and ord(token) < 32:

docs/connpy/cli/terminal_ui.html

@@ -90,9 +90,10 @@ el.replaceWith(d);
 
     async def run_session(self, 
                           raw_bytes: bytes, 
-                          cmd_byte_positions: List[tuple], 
                           node_info: dict,
-                          on_ai_call: Callable):
+                          on_ai_call: Callable,
+                          cmd_byte_positions: List[tuple] = None, 
+                          blocks: List[tuple] = None):
         """
         Runs the interactive Copilot session.
         on_ai_call: async function(active_buffer, question) -> result_dict
@@ -102,9 +103,11 @@ el.replaceWith(d);
         try:
             # Prepare UI state
             buffer = log_cleaner(raw_bytes.decode(errors='replace'))
-            blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info)
-            last_line = buffer.split('\n')[-1].strip() if buffer.strip() else "(prompt)"
-            blocks.append((len(raw_bytes), last_line[:80]))
+            
+            # Use pre-calculated blocks if provided (remote mode), otherwise calculate locally (local mode)
+            if blocks is None:
+                last_line = buffer.split('\n')[-1].strip() if buffer.strip() else "(prompt)"
+                blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info, last_line=last_line)
             
             state = {
                 'context_cmd': 1,
@@ -121,7 +124,7 @@ el.replaceWith(d);
             self.console.print("") # Salto de línea real
             self.console.print(Rule(title="[bold cyan] AI TERMINAL COPILOT [/bold cyan]", style="cyan"))
             self.console.print(Panel(
-                "[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel.\n"
+                "[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel. Type / for commands.\n"
                 "Tab to change context mode. Ctrl+\u2191/\u2193 to adjust context. \u2191\u2193 for question history.[/dim]",
                 border_style="cyan"
             ))
@@ -161,9 +164,8 @@ el.replaceWith(d);
                 if state['context_mode'] == self.mode_lines:
                     return '\n'.join(buffer.split('\n')[-state['context_lines']:])
                 idx = max(0, state['total_cmds'] - state['context_cmd'])
-                start, preview = blocks[idx]
-                if state['context_mode'] == self.mode_single and idx + 1 < state['total_cmds']:
-                    end = blocks[idx + 1][0]
+                start, end, preview = blocks[idx]
+                if state['context_mode'] == self.mode_single:
                     active_raw = raw_bytes[start:end]
                 else:
                     active_raw = raw_bytes[start:]
@@ -205,7 +207,40 @@ el.replaceWith(d);
                     base_str = f'\u25b6 Ctrl+\u2191/\u2193 adjusts by 50 lines  [Tab: {m_label}]'
                 else:
                     idx = max(0, state['total_cmds'] - state['context_cmd'])
-                    desc = blocks[idx][1]
+                    import re
+                    
+                    def clean_preview(text):
+                        # Limpia saltos de línea y el prompt inicial (todo hasta #, > o $) para que quede solo el comando
+                        original = text.strip().replace('\r', '').replace('\n', ' ')
+                        cleaned = re.sub(r'^.*?[#>\$]\s*', '', original)
+                        # Si limpiar el prompt nos deja con un string vacío (ej: era solo "iol#"), devolvemos el original
+                        return cleaned if cleaned else original
+
+                    if state['context_mode'] == self.mode_range:
+                        range_blocks = blocks[idx:]
+                        # Si hay más de un bloque, el último es siempre el prompt vacío/actual. Lo omitimos visualmente.
+                        if len(range_blocks) > 1:
+                            range_blocks = range_blocks[:-1]
+                            
+                        # Limpiar y truncar comandos muy largos para que no rompan la UI
+                        previews = []
+                        for b in range_blocks:
+                            p = clean_preview(b[2])
+                            if p:
+                                # Truncar comandos individuales largos
+                                if len(p) > 25: p = p[:22] + "..."
+                                previews.append(p)
+                        
+                        if not previews:
+                            desc = clean_preview(blocks[idx][2])
+                        elif len(previews) <= 3:
+                            desc = " + ".join(previews)
+                        else:
+                            desc = f"{previews[0]} + {previews[1]} + {previews[2]} ... (+{len(previews)-3})"
+                    else:
+                        # Modo SINGLE original
+                        desc = clean_preview(blocks[idx][2])
+                        
                     base_str = f'\u25b6 {desc}  [Tab: {m_label}]'
                 
                 # Wrap base_str in a style to maintain consistency and avoid glitches
@@ -332,39 +367,67 @@ el.replaceWith(d);
                 # Use persona from overrides (one-shot) or from session state
                 active_persona = merged_node_info.get('persona', self.session_state.get('persona', 'engineer'))
                 persona_color = self._get_theme_color(active_persona, fallback="cyan")
+                persona_title = "Network Architect" if active_persona == "architect" else "Network Engineer"
                 
                 active_buffer = get_active_buffer()
-                live_text = "Thinking..."
-                panel = Panel(live_text, title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color)
+                live_text = ""
+                first_chunk = True
+                
+                import sys
+                from rich.rule import Rule
+                from rich.status import Status
+                from connpy.printer import IncrementalMarkdownParser
+                
+                md_parser = IncrementalMarkdownParser(console=self.console)
+                
+                status_spinner = Status(
+                    f"[bold {persona_color}]{persona_title}:[/bold {persona_color}] [dim]Thinking...[/dim]",
+                    console=self.console,
+                    spinner="dots"
+                )
+                status_spinner.start()
                 
                 def on_chunk(text):
-                    nonlocal live_text
-                    if live_text == "Thinking...": live_text = ""
+                    nonlocal live_text, first_chunk
+                    if first_chunk:
+                        status_spinner.stop()
+                        # Print header rule before first chunk arrives
+                        self.console.print(Rule(
+                            f"[bold {persona_color}]{persona_title}[/bold {persona_color}]",
+                            style=persona_color
+                        ))
+                        first_chunk = False
                     live_text += text
+                    md_parser.feed(text)
                 
-                with Live(panel, console=self.console, refresh_per_second=10) as live:
-                    def update_live(t):
-                        live.update(Panel(Markdown(t), title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color))
+                # Check for interruption during AI call
+                ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, on_chunk, merged_node_info))
                 
-                    wrapped_chunk = lambda t: (on_chunk(t), update_live(live_text))
-                    
-                    # Check for interruption during AI call
-                    ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, wrapped_chunk, merged_node_info))
-                    
-                    try:
-                        while not ai_task.done():
-                            await asyncio.sleep(0.05)
-                        result = await ai_task
-                    except asyncio.CancelledError:
-                        return "cancel", None, None
-
-                if not result or result.get("error"):
-                    if result and result.get("error"): self.console.print(f"[red]Error: {result['error']}[/red]")
+                try:
+                    while not ai_task.done():
+                        await asyncio.sleep(0.05)
+                    result = await ai_task
+                except asyncio.CancelledError:
+                    status_spinner.stop()
                     return "cancel", None, None
                 
-                # 4. Handle result
-                if live_text == "Thinking..." and result.get("guide"):
-                    self.console.print(Panel(Markdown(result["guide"]), title=f"[bold {persona_color}]Copilot Guide[/bold {persona_color}]", border_style=persona_color))
+                # Ensure spinner is stopped if no chunks arrived
+                if first_chunk:
+                    status_spinner.stop()
+
+                # Close the streamed output with a Rule
+                if not first_chunk:
+                    md_parser.flush()
+                    self.console.print(Rule(style=persona_color))
+                
+                if not result or result.get("error"):
+                    if first_chunk and result and result.get("error"):
+                        self.console.print(f"[red]Error: {result['error']}[/red]")
+                    return "cancel", None, None
+
+                # If no chunks were streamed but we have a guide, print it as a panel
+                if first_chunk and result and result.get("guide"):
+                    self.console.print(Panel(Markdown(result["guide"]), title=f"[bold {persona_color}]{persona_title}[/bold {persona_color}]", border_style=persona_color))
 
                 commands = result.get("commands", [])
                 if not commands:
@@ -466,14 +529,13 @@ el.replaceWith(d);
             return "cancel", None, None
 
         finally:
-            state['cancelled'] = True
-            self.console.print(&#34;[dim]Returning to session...[/dim]&#34;)</code></pre>
+            state[&#39;cancelled&#39;] = True</code></pre>
 </details>
 <div class="desc"></div>
 <h3>Methods</h3>
 <dl>
 <dt id="connpy.cli.terminal_ui.CopilotInterface.run_session"><code class="name flex">
-<span>async def <span class="ident">run_session</span></span>(<span>self,<br>raw_bytes: bytes,<br>cmd_byte_positions: List[tuple],<br>node_info: dict,<br>on_ai_call: Callable)</span>
+<span>async def <span class="ident">run_session</span></span>(<span>self,<br>raw_bytes: bytes,<br>node_info: dict,<br>on_ai_call: Callable,<br>cmd_byte_positions: List[tuple] = None,<br>blocks: List[tuple] = None)</span>
 </code></dt>
 <dd>
 <details class="source">
@@ -482,9 +544,10 @@ el.replaceWith(d);
 </summary>
 <pre><code class="python">async def run_session(self, 
                       raw_bytes: bytes, 
-                      cmd_byte_positions: List[tuple], 
                       node_info: dict,
-                      on_ai_call: Callable):
+                      on_ai_call: Callable,
+                      cmd_byte_positions: List[tuple] = None, 
+                      blocks: List[tuple] = None):
     &#34;&#34;&#34;
     Runs the interactive Copilot session.
     on_ai_call: async function(active_buffer, question) -&gt; result_dict
@@ -494,9 +557,11 @@ el.replaceWith(d);
     try:
         # Prepare UI state
         buffer = log_cleaner(raw_bytes.decode(errors=&#39;replace&#39;))
-        blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info)
-        last_line = buffer.split(&#39;\n&#39;)[-1].strip() if buffer.strip() else &#34;(prompt)&#34;
-        blocks.append((len(raw_bytes), last_line[:80]))
+        
+        # Use pre-calculated blocks if provided (remote mode), otherwise calculate locally (local mode)
+        if blocks is None:
+            last_line = buffer.split(&#39;\n&#39;)[-1].strip() if buffer.strip() else &#34;(prompt)&#34;
+            blocks = self.ai_service.build_context_blocks(raw_bytes, cmd_byte_positions, node_info, last_line=last_line)
         
         state = {
             &#39;context_cmd&#39;: 1,
@@ -513,7 +578,7 @@ el.replaceWith(d);
         self.console.print(&#34;&#34;) # Salto de línea real
         self.console.print(Rule(title=&#34;[bold cyan] AI TERMINAL COPILOT [/bold cyan]&#34;, style=&#34;cyan&#34;))
         self.console.print(Panel(
-            &#34;[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel.\n&#34;
+            &#34;[dim]Type your question. Enter to send, Escape/Ctrl+C to cancel. Type / for commands.\n&#34;
             &#34;Tab to change context mode. Ctrl+\u2191/\u2193 to adjust context. \u2191\u2193 for question history.[/dim]&#34;,
             border_style=&#34;cyan&#34;
         ))
@@ -553,9 +618,8 @@ el.replaceWith(d);
             if state[&#39;context_mode&#39;] == self.mode_lines:
                 return &#39;\n&#39;.join(buffer.split(&#39;\n&#39;)[-state[&#39;context_lines&#39;]:])
             idx = max(0, state[&#39;total_cmds&#39;] - state[&#39;context_cmd&#39;])
-            start, preview = blocks[idx]
-            if state[&#39;context_mode&#39;] == self.mode_single and idx + 1 &lt; state[&#39;total_cmds&#39;]:
-                end = blocks[idx + 1][0]
+            start, end, preview = blocks[idx]
+            if state[&#39;context_mode&#39;] == self.mode_single:
                 active_raw = raw_bytes[start:end]
             else:
                 active_raw = raw_bytes[start:]
@@ -597,7 +661,40 @@ el.replaceWith(d);
                 base_str = f&#39;\u25b6 Ctrl+\u2191/\u2193 adjusts by 50 lines  [Tab: {m_label}]&#39;
             else:
                 idx = max(0, state[&#39;total_cmds&#39;] - state[&#39;context_cmd&#39;])
-                desc = blocks[idx][1]
+                import re
+                
+                def clean_preview(text):
+                    # Limpia saltos de línea y el prompt inicial (todo hasta #, &gt; o $) para que quede solo el comando
+                    original = text.strip().replace(&#39;\r&#39;, &#39;&#39;).replace(&#39;\n&#39;, &#39; &#39;)
+                    cleaned = re.sub(r&#39;^.*?[#&gt;\$]\s*&#39;, &#39;&#39;, original)
+                    # Si limpiar el prompt nos deja con un string vacío (ej: era solo &#34;iol#&#34;), devolvemos el original
+                    return cleaned if cleaned else original
+
+                if state[&#39;context_mode&#39;] == self.mode_range:
+                    range_blocks = blocks[idx:]
+                    # Si hay más de un bloque, el último es siempre el prompt vacío/actual. Lo omitimos visualmente.
+                    if len(range_blocks) &gt; 1:
+                        range_blocks = range_blocks[:-1]
+                        
+                    # Limpiar y truncar comandos muy largos para que no rompan la UI
+                    previews = []
+                    for b in range_blocks:
+                        p = clean_preview(b[2])
+                        if p:
+                            # Truncar comandos individuales largos
+                            if len(p) &gt; 25: p = p[:22] + &#34;...&#34;
+                            previews.append(p)
+                    
+                    if not previews:
+                        desc = clean_preview(blocks[idx][2])
+                    elif len(previews) &lt;= 3:
+                        desc = &#34; + &#34;.join(previews)
+                    else:
+                        desc = f&#34;{previews[0]} + {previews[1]} + {previews[2]} ... (+{len(previews)-3})&#34;
+                else:
+                    # Modo SINGLE original
+                    desc = clean_preview(blocks[idx][2])
+                    
                 base_str = f&#39;\u25b6 {desc}  [Tab: {m_label}]&#39;
             
             # Wrap base_str in a style to maintain consistency and avoid glitches
@@ -724,39 +821,67 @@ el.replaceWith(d);
             # Use persona from overrides (one-shot) or from session state
             active_persona = merged_node_info.get(&#39;persona&#39;, self.session_state.get(&#39;persona&#39;, &#39;engineer&#39;))
             persona_color = self._get_theme_color(active_persona, fallback=&#34;cyan&#34;)
+            persona_title = &#34;Network Architect&#34; if active_persona == &#34;architect&#34; else &#34;Network Engineer&#34;
             
             active_buffer = get_active_buffer()
-            live_text = &#34;Thinking...&#34;
-            panel = Panel(live_text, title=f&#34;[bold {persona_color}]Copilot Guide[/bold {persona_color}]&#34;, border_style=persona_color)
+            live_text = &#34;&#34;
+            first_chunk = True
+            
+            import sys
+            from rich.rule import Rule
+            from rich.status import Status
+            from connpy.printer import IncrementalMarkdownParser
+            
+            md_parser = IncrementalMarkdownParser(console=self.console)
+            
+            status_spinner = Status(
+                f&#34;[bold {persona_color}]{persona_title}:[/bold {persona_color}] [dim]Thinking...[/dim]&#34;,
+                console=self.console,
+                spinner=&#34;dots&#34;
+            )
+            status_spinner.start()
             
             def on_chunk(text):
-                nonlocal live_text
-                if live_text == &#34;Thinking...&#34;: live_text = &#34;&#34;
+                nonlocal live_text, first_chunk
+                if first_chunk:
+                    status_spinner.stop()
+                    # Print header rule before first chunk arrives
+                    self.console.print(Rule(
+                        f&#34;[bold {persona_color}]{persona_title}[/bold {persona_color}]&#34;,
+                        style=persona_color
+                    ))
+                    first_chunk = False
                 live_text += text
+                md_parser.feed(text)
             
-            with Live(panel, console=self.console, refresh_per_second=10) as live:
-                def update_live(t):
-                    live.update(Panel(Markdown(t), title=f&#34;[bold {persona_color}]Copilot Guide[/bold {persona_color}]&#34;, border_style=persona_color))
+            # Check for interruption during AI call
+            ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, on_chunk, merged_node_info))
             
-                wrapped_chunk = lambda t: (on_chunk(t), update_live(live_text))
-                
-                # Check for interruption during AI call
-                ai_task = asyncio.create_task(on_ai_call(active_buffer, clean_question, wrapped_chunk, merged_node_info))
-                
-                try:
-                    while not ai_task.done():
-                        await asyncio.sleep(0.05)
-                    result = await ai_task
-                except asyncio.CancelledError:
-                    return &#34;cancel&#34;, None, None
-
-            if not result or result.get(&#34;error&#34;):
-                if result and result.get(&#34;error&#34;): self.console.print(f&#34;[red]Error: {result[&#39;error&#39;]}[/red]&#34;)
+            try:
+                while not ai_task.done():
+                    await asyncio.sleep(0.05)
+                result = await ai_task
+            except asyncio.CancelledError:
+                status_spinner.stop()
                 return &#34;cancel&#34;, None, None
             
-            # 4. Handle result
-            if live_text == &#34;Thinking...&#34; and result.get(&#34;guide&#34;):
-                self.console.print(Panel(Markdown(result[&#34;guide&#34;]), title=f&#34;[bold {persona_color}]Copilot Guide[/bold {persona_color}]&#34;, border_style=persona_color))
+            # Ensure spinner is stopped if no chunks arrived
+            if first_chunk:
+                status_spinner.stop()
+
+            # Close the streamed output with a Rule
+            if not first_chunk:
+                md_parser.flush()
+                self.console.print(Rule(style=persona_color))
+            
+            if not result or result.get(&#34;error&#34;):
+                if first_chunk and result and result.get(&#34;error&#34;):
+                    self.console.print(f&#34;[red]Error: {result[&#39;error&#39;]}[/red]&#34;)
+                return &#34;cancel&#34;, None, None
+
+            # If no chunks were streamed but we have a guide, print it as a panel
+            if first_chunk and result and result.get(&#34;guide&#34;):
+                self.console.print(Panel(Markdown(result[&#34;guide&#34;]), title=f&#34;[bold {persona_color}]{persona_title}[/bold {persona_color}]&#34;, border_style=persona_color))
 
             commands = result.get(&#34;commands&#34;, [])
             if not commands:
@@ -858,8 +983,7 @@ el.replaceWith(d);
         return &#34;cancel&#34;, None, None
 
     finally:
-        state[&#39;cancelled&#39;] = True
-        self.console.print(&#34;[dim]Returning to session...[/dim]&#34;)</code></pre>
+        state[&#39;cancelled&#39;] = True</code></pre>
 </details>
 <div class="desc"><p>Runs the interactive Copilot session.
 on_ai_call: async function(active_buffer, question) -&gt; result_dict</p></div>

docs/connpy/grpc_layer/server.html

@@ -807,15 +807,34 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                     import json
                     import asyncio
                     import os
+                    from ..services.ai_service import AIService
+                    
+                    service = AIService(self.service.config)
 
                     if node_info is None:
                         node_info = {}
                     
+                    # Calculate real command blocks from history using the central service
+                    raw_bytes = n.mylog.getvalue() if hasattr(n, &#39;mylog&#39;) else buffer
+                    if not isinstance(raw_bytes, bytes):
+                        raw_bytes = str(raw_bytes).encode()
+                    
+                    from connpy.utils import log_cleaner
+                    last_line = log_cleaner(raw_bytes.decode(errors=&#39;replace&#39;)).split(&#39;\n&#39;)[-1].strip()
+                    blocks = service.build_context_blocks(raw_bytes, n.cmd_byte_positions, node_info, last_line=last_line)
+                    node_info[&#34;context_blocks&#34;] = blocks
+
                     node_info_json = json.dumps(node_info)
                     
                     # Convert buffer to string if it&#39;s bytes for the preview
                     preview_str = buffer[-200:].decode(errors=&#39;replace&#39;) if isinstance(buffer, bytes) else str(buffer)[-200:]
                     
+                    # Generate a unique session ID for this copilot interaction to prevent race conditions
+                    import uuid
+                    copilot_session_id = str(uuid.uuid4())
+                    node_info[&#34;session_id&#34;] = copilot_session_id
+                    node_info_json = json.dumps(node_info)
+                    
                     # 1. Send prompt to client
                     response_queue.put(connpy_pb2.InteractResponse(
                         copilot_prompt=True,
@@ -824,6 +843,13 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                     ))
 
                     while True:
+                        # 0. Drain the queue of any stale messages before starting a new interaction
+                        while not remote_stream.copilot_queue.empty():
+                            try:
+                                remote_stream.copilot_queue.get_nowait()
+                            except:
+                                break
+
                         # 2. Await the question from client via the copilot_queue
                         import threading
                         def preload_ai_deps():
@@ -836,8 +862,17 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                         try:
                             req_data = await asyncio.wait_for(remote_stream.copilot_queue.get(), timeout=120)
                             if not req_data: return
-                            if &#34;question&#34; not in req_data or not req_data[&#34;question&#34;] or req_data[&#34;question&#34;] == &#34;CANCEL&#34; or req_data.get(&#34;action&#34;) == &#34;cancel&#34;:
-                                os.write(child_fd, b&#39;\x15\r&#39;)
+                            
+                            # Validate session ID if provided by client (skip validation if not provided for CLI compatibility)
+                            req_session_id = req_data.get(&#34;session_id&#34;)
+                            if req_session_id and req_session_id != copilot_session_id:
+                                continue # Ignore stale request from a previous session
+
+                            if &#34;question&#34; not in req_data or not req_data[&#34;question&#34;] or req_data[&#34;question&#34;] == &#34;CANCEL&#34; or req_data.get(&#34;action&#34;) in (&#34;cancel&#34;, &#34;web_cancel&#34;):
+                                if req_data.get(&#34;action&#34;) == &#34;web_cancel&#34;:
+                                    os.write(child_fd, b&#39;\x05&#39;)
+                                else:
+                                    os.write(child_fd, b&#39;\x15\r&#39;)
                                 return
                             question = req_data[&#34;question&#34;]
                             
@@ -864,9 +899,6 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                             return
                             
                         # 3. Call AI Service with streaming
-                        from ..services.ai_service import AIService
-                        service = AIService(self.service.config)
-                        
                         def chunk_callback(chunk_text):
                             if chunk_text:
                                 response_queue.put(connpy_pb2.InteractResponse(
@@ -887,8 +919,11 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                         if wait_action_task in done:
                             req_data = wait_action_task.result()
                             ai_task.cancel()
-                            if req_data.get(&#34;action&#34;) == &#34;cancel&#34; or req_data.get(&#34;question&#34;) == &#34;CANCEL&#34;:
-                                os.write(child_fd, b&#39;\x15\r&#39;)
+                            if req_data.get(&#34;action&#34;) in (&#34;cancel&#34;, &#34;web_cancel&#34;) or req_data.get(&#34;question&#34;) == &#34;CANCEL&#34;:
+                                if req_data.get(&#34;action&#34;) == &#34;web_cancel&#34;:
+                                    os.write(child_fd, b&#39;\x05&#39;)
+                                else:
+                                    os.write(child_fd, b&#39;\x15\r&#39;)
                                 return
                             continue # Loop back instead of returning to keep session alive
                         else:
@@ -912,45 +947,27 @@ interceptor chooses to service this RPC, or None otherwise.</p></div>
                             if action == &#34;continue&#34;:
                                 continue # Loop back for next question
                                 
-                            if action == &#34;cancel&#34;:
-                                os.write(child_fd, b&#39;\x15\r&#39;)
+                            if action in (&#34;cancel&#34;, &#34;web_cancel&#34;):
+                                if action == &#34;web_cancel&#34;:
+                                    os.write(child_fd, b&#39;\x05&#39;)
+                                else:
+                                    os.write(child_fd, b&#39;\x15\r&#39;)
                                 return
                         except asyncio.TimeoutError:
                             os.write(child_fd, b&#39;\x15\r&#39;)
                             return
                             
+                        def on_inject(cmd):
+                            response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd))
+
                         if action == &#34;send_all&#34;:
                             commands = result.get(&#34;commands&#34;, [])
-                            os.write(child_fd, b&#39;\x15&#39;) # Ctrl+U to clear line
-                            await asyncio.sleep(0.1)
-
-                            # Prepend screen length command to avoid pagination
-                            if &#34;screen_length_command&#34; in n.tags:
-                                os.write(child_fd, (n.tags[&#34;screen_length_command&#34;] + &#34;\n&#34;).encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=n.tags[&#34;screen_length_command&#34;]))
-                                await asyncio.sleep(0.8)
-
-                            for cmd in commands:
-                                os.write(child_fd, (cmd + &#34;\n&#34;).encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd))
-                                await asyncio.sleep(0.8)
+                            await n.inject_commands(commands, child_fd, on_inject=on_inject)
                             return
                         elif action.startswith(&#34;custom:&#34;):
-                            custom_cmds = action[7:]
-                            os.write(child_fd, b&#39;\x15&#39;)
-                            await asyncio.sleep(0.1)
-                            
-                            # Prepend screen length command to avoid pagination
-                            if &#34;screen_length_command&#34; in n.tags:
-                                os.write(child_fd, (n.tags[&#34;screen_length_command&#34;] + &#34;\n&#34;).encode())
-                                response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=n.tags[&#34;screen_length_command&#34;]))
-                                await asyncio.sleep(0.8)
-
-                            for cmd in custom_cmds.split(&#39;\n&#39;):
-                                if cmd.strip():
-                                    os.write(child_fd, (cmd.strip() + &#34;\n&#34;).encode())
-                                    response_queue.put(connpy_pb2.InteractResponse(copilot_injected_command=cmd.strip()))
-                                    await asyncio.sleep(0.8)
+                            custom_cmds_raw = action[7:]
+                            custom_cmds = [cmd.strip() for cmd in custom_cmds_raw.split(&#39;\n&#39;) if cmd.strip()]
+                            await n.inject_commands(custom_cmds, child_fd, on_inject=on_inject)
                             return
                         else:
                             os.write(child_fd, b&#39;\x15\r&#39;)

docs/connpy/grpc_layer/stubs.html

@@ -104,7 +104,6 @@ el.replaceWith(d);
         import queue
         from rich.prompt import Prompt
         from rich.text import Text
-        from rich.live import Live
         from rich.panel import Panel
         from rich.markdown import Markdown
         
@@ -135,7 +134,7 @@ el.replaceWith(d);
         responses = self.stub.ask(request_generator())
         
         full_content = ""
-        live_display = None
+        header_printed = False
         final_result = {"response": "", "chat_history": []}
 
         # Background thread to pull responses from gRPC into a local queue
@@ -200,69 +199,53 @@ el.replaceWith(d);
                 
                 if response.debug_message:
                     if debug:
-                        if live_display:
-                            try: live_display.stop()
-                            except: pass
                         if status:
                             try: status.stop()
                             except: pass
                         printer.console.print(Text.from_ansi(response.debug_message))
-                        if live_display:
-                            try: live_display.start()
-                            except: pass
-                        elif status:
+                        if status:
                             try: status.start()
                             except: pass
                     continue
                 
                 if response.important_message:
-                    if live_display:
-                        try: live_display.stop()
-                        except: pass
                     if status:
                         try: status.stop()
                         except: pass
                     printer.console.print(Text.from_ansi(response.important_message))
-                    if live_display:
-                        try: live_display.start()
-                        except: pass
-                    elif status:
+                    if status:
                         try: status.start()
                         except: pass
                     continue
 
                 if not response.is_final:
                     if response.text_chunk:
-                        full_content += response.text_chunk
-                        
-                        if not live_display:
+                        if not header_printed:
                             if status:
                                 try: status.stop()
                                 except: pass
                             
                             from rich.console import Console as RichConsole
-                            from ..printer import connpy_theme, get_original_stdout
+                            from rich.rule import Rule
+                            from ..printer import connpy_theme, get_original_stdout, IncrementalMarkdownParser
                             stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
                             
-                            # We default to Engineer title during stream, final result will correct it if needed
-                            live_display = Live(
-                                Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False),
-                                console=stable_console,
-                                refresh_per_second=8,
-                                transient=False
-                            )
-                            live_display.start()
-                        else:
-                            live_display.update(
-                                Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False)
-                            )
+                            # Print header on first chunk
+                            stable_console.print(Rule("[bold engineer]Network Engineer[/bold engineer]", style="engineer"))
+                            header_printed = True
+                            
+                            # Initialize parser
+                            md_parser = IncrementalMarkdownParser(console=stable_console)
+                        
+                        full_content += response.text_chunk
+                        md_parser.feed(response.text_chunk)
                     continue
                 
                 if response.is_final:
-                    if live_display:
-                        try: live_display.stop()
-                        except: pass
-                    # Final stop for status to ensure it disappears before the panel
+                    if header_printed:
+                        from rich.rule import Rule
+                        md_parser.flush()
+                        
                     if status:
                         try: status.stop()
                         except: pass
@@ -273,13 +256,14 @@ el.replaceWith(d);
                     role_label = "Network Architect" if responder == "architect" else "Network Engineer"
                     title = f"[bold {alias}]{role_label}[/bold {alias}]"
                     
-                    content_to_print = full_content or final_result.get("response", "")
-                    if content_to_print:
-                        if live_display:
-                            # Re-render the final frame with correct title/colors
-                            live_display.update(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
-                        else:
-                            printer.console.print(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
+                    if header_printed:
+                        from rich.console import Console as RichConsole
+                        from ..printer import connpy_theme, get_original_stdout
+                        stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
+                        stable_console.print(Rule(style=alias))
+                    elif not full_content and final_result.get("response"):
+                        # If nothing streamed but we have response (e.g. error or direct guide)
+                        printer.console.print(Panel(Markdown(final_result["response"]), title=title, border_style=alias, expand=False))
                     break
         except Exception as e:
             # Check if it was a gRPC error that we should let handle_errors catch
@@ -342,7 +326,6 @@ def ask(self, input_text, dryrun=False, chat_history=None, session_id=None, debu
     import queue
     from rich.prompt import Prompt
     from rich.text import Text
-    from rich.live import Live
     from rich.panel import Panel
     from rich.markdown import Markdown
     
@@ -373,7 +356,7 @@ def ask(self, input_text, dryrun=False, chat_history=None, session_id=None, debu
     responses = self.stub.ask(request_generator())
     
     full_content = ""
-    live_display = None
+    header_printed = False
     final_result = {"response": "", "chat_history": []}
 
     # Background thread to pull responses from gRPC into a local queue
@@ -438,69 +421,53 @@ def ask(self, input_text, dryrun=False, chat_history=None, session_id=None, debu
             
             if response.debug_message:
                 if debug:
-                    if live_display:
-                        try: live_display.stop()
-                        except: pass
                     if status:
                         try: status.stop()
                         except: pass
                     printer.console.print(Text.from_ansi(response.debug_message))
-                    if live_display:
-                        try: live_display.start()
-                        except: pass
-                    elif status:
+                    if status:
                         try: status.start()
                         except: pass
                 continue
             
             if response.important_message:
-                if live_display:
-                    try: live_display.stop()
-                    except: pass
                 if status:
                     try: status.stop()
                     except: pass
                 printer.console.print(Text.from_ansi(response.important_message))
-                if live_display:
-                    try: live_display.start()
-                    except: pass
-                elif status:
+                if status:
                     try: status.start()
                     except: pass
                 continue
 
             if not response.is_final:
                 if response.text_chunk:
-                    full_content += response.text_chunk
-                    
-                    if not live_display:
+                    if not header_printed:
                         if status:
                             try: status.stop()
                             except: pass
                         
                         from rich.console import Console as RichConsole
-                        from ..printer import connpy_theme, get_original_stdout
+                        from rich.rule import Rule
+                        from ..printer import connpy_theme, get_original_stdout, IncrementalMarkdownParser
                         stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
                         
-                        # We default to Engineer title during stream, final result will correct it if needed
-                        live_display = Live(
-                            Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False),
-                            console=stable_console,
-                            refresh_per_second=8,
-                            transient=False
-                        )
-                        live_display.start()
-                    else:
-                        live_display.update(
-                            Panel(Markdown(full_content), title="[bold engineer]Network Engineer[/bold engineer]", border_style="engineer", expand=False)
-                        )
+                        # Print header on first chunk
+                        stable_console.print(Rule("[bold engineer]Network Engineer[/bold engineer]", style="engineer"))
+                        header_printed = True
+                        
+                        # Initialize parser
+                        md_parser = IncrementalMarkdownParser(console=stable_console)
+                    
+                    full_content += response.text_chunk
+                    md_parser.feed(response.text_chunk)
                 continue
             
             if response.is_final:
-                if live_display:
-                    try: live_display.stop()
-                    except: pass
-                # Final stop for status to ensure it disappears before the panel
+                if header_printed:
+                    from rich.rule import Rule
+                    md_parser.flush()
+                    
                 if status:
                     try: status.stop()
                     except: pass
@@ -511,13 +478,14 @@ def ask(self, input_text, dryrun=False, chat_history=None, session_id=None, debu
                 role_label = "Network Architect" if responder == "architect" else "Network Engineer"
                 title = f"[bold {alias}]{role_label}[/bold {alias}]"
                 
-                content_to_print = full_content or final_result.get("response", "")
-                if content_to_print:
-                    if live_display:
-                        # Re-render the final frame with correct title/colors
-                        live_display.update(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
-                    else:
-                        printer.console.print(Panel(Markdown(content_to_print), title=title, border_style=alias, expand=False))
+                if header_printed:
+                    from rich.console import Console as RichConsole
+                    from ..printer import connpy_theme, get_original_stdout
+                    stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
+                    stable_console.print(Rule(style=alias))
+                elif not full_content and final_result.get("response"):
+                    # If nothing streamed but we have response (e.g. error or direct guide)
+                    printer.console.print(Panel(Markdown(final_result["response"]), title=title, border_style=alias, expand=False))
                 break
     except Exception as e:
         # Check if it was a gRPC error that we should let handle_errors catch
@@ -1024,7 +992,7 @@ def set_reserved_names(self, names):
         self.remote_host = remote_host
         self.config = config
 
-    def _handle_remote_copilot(self, res, request_queue, response_queue, client_buffer_bytes, cmd_byte_positions, pause_generator, resume_generator, old_tty):
+    def _handle_remote_copilot(self, res, request_queue, response_queue, client_buffer_bytes, pause_generator, resume_generator, old_tty):
         import json, asyncio, termios, sys, tty, queue
         from ..core import copilot_terminal_mode
         from . import connpy_pb2
@@ -1032,6 +1000,10 @@ def set_reserved_names(self, names):
         pause_generator()
         
         termios.tcsetattr(sys.stdin, termios.TCSADRAIN, old_tty)
+        
+        node_info = json.loads(res.copilot_node_info_json) if res.copilot_node_info_json else {}
+        blocks = node_info.get("context_blocks", [])
+
         interface = CopilotInterface(
             self.config, 
             history=getattr(self, 'copilot_history', None),
@@ -1040,8 +1012,6 @@ def set_reserved_names(self, names):
         self.copilot_history = interface.history
         self.copilot_state = interface.session_state
 
-        node_info = json.loads(res.copilot_node_info_json) if res.copilot_node_info_json else {}
-
         async def on_ai_call_remote(active_buffer, question, chunk_callback, merged_node_info):
             # Send request to server
             request_queue.put(connpy_pb2.InteractRequest(
@@ -1066,9 +1036,9 @@ def set_reserved_names(self, names):
             while True:
                 action, commands, custom_cmd = await interface.run_session(
                     raw_bytes=bytes(client_buffer_bytes),
-                    cmd_byte_positions=cmd_byte_positions,
                     node_info=node_info,
-                    on_ai_call=on_ai_call_remote
+                    on_ai_call=on_ai_call_remote,
+                    blocks=blocks
                 )
                 
                 if action == "continue":
@@ -1081,6 +1051,7 @@ def set_reserved_names(self, names):
         with copilot_terminal_mode():
             action, commands, custom_cmd = asyncio.run(run_remote_copilot())
         
+        print("\033[2m Returning to session...\033[0m", flush=True)
         # Prepare final action for server
         action_sent = "cancel"
         if action == "send_all" and commands:
@@ -1105,7 +1076,6 @@ def set_reserved_names(self, names):
         
         request_queue = queue.Queue()
         client_buffer_bytes = bytearray()
-        cmd_byte_positions = [(0, None)]
         pause_stdin = [False]
         wake_r, wake_w = os.pipe()
 
@@ -1152,8 +1122,6 @@ def set_reserved_names(self, names):
                         data = os.read(sys.stdin.fileno(), 1024)
                         if not data:
                             break
-                        if b'\r' in data or b'\n' in data:
-                            cmd_byte_positions.append((len(client_buffer_bytes), None))
                         yield connpy_pb2.InteractRequest(stdin_data=data)
                     except OSError:
                         break
@@ -1227,14 +1195,11 @@ def set_reserved_names(self, names):
                 if res.copilot_prompt:
                     self._handle_remote_copilot(
                         res, request_queue, response_queue, 
-                        client_buffer_bytes, cmd_byte_positions, 
+                        client_buffer_bytes, 
                         pause_generator, resume_generator, old_tty
                     )
                     continue
 
-                if res.copilot_injected_command:
-                    cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
                 if res.stdout_data:
                     os.write(sys.stdout.fileno(), res.stdout_data)
                     client_buffer_bytes.extend(res.stdout_data)
@@ -1256,7 +1221,6 @@ def set_reserved_names(self, names):
         params_json = json.dumps(connection_params)
         request_queue = queue.Queue()
         client_buffer_bytes = bytearray()
-        cmd_byte_positions = [(0, None)]
         pause_stdin = [False]
         wake_r, wake_w = os.pipe()
 
@@ -1304,8 +1268,6 @@ def set_reserved_names(self, names):
                         data = os.read(sys.stdin.fileno(), 1024)
                         if not data:
                             break
-                        if b'\r' in data or b'\n' in data:
-                            cmd_byte_positions.append((len(client_buffer_bytes), None))
                         yield connpy_pb2.InteractRequest(stdin_data=data)
                     except OSError:
                         break
@@ -1378,14 +1340,11 @@ def set_reserved_names(self, names):
                 if res.copilot_prompt:
                     self._handle_remote_copilot(
                         res, request_queue, response_queue, 
-                        client_buffer_bytes, cmd_byte_positions, 
+                        client_buffer_bytes, 
                         pause_generator, resume_generator, old_tty
                     )
                     continue
 
-                if res.copilot_injected_command:
-                    cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
                 if res.stdout_data:
                     os.write(sys.stdout.fileno(), res.stdout_data)
                     client_buffer_bytes.extend(res.stdout_data)
@@ -1553,7 +1512,6 @@ def connect_dynamic(self, connection_params, debug=False):
     params_json = json.dumps(connection_params)
     request_queue = queue.Queue()
     client_buffer_bytes = bytearray()
-    cmd_byte_positions = [(0, None)]
     pause_stdin = [False]
     wake_r, wake_w = os.pipe()
 
@@ -1601,8 +1559,6 @@ def connect_dynamic(self, connection_params, debug=False):
                     data = os.read(sys.stdin.fileno(), 1024)
                     if not data:
                         break
-                    if b'\r' in data or b'\n' in data:
-                        cmd_byte_positions.append((len(client_buffer_bytes), None))
                     yield connpy_pb2.InteractRequest(stdin_data=data)
                 except OSError:
                     break
@@ -1675,14 +1631,11 @@ def connect_dynamic(self, connection_params, debug=False):
             if res.copilot_prompt:
                 self._handle_remote_copilot(
                     res, request_queue, response_queue, 
-                    client_buffer_bytes, cmd_byte_positions, 
+                    client_buffer_bytes, 
                     pause_generator, resume_generator, old_tty
                 )
                 continue
 
-            if res.copilot_injected_command:
-                cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
             if res.stdout_data:
                 os.write(sys.stdout.fileno(), res.stdout_data)
                 client_buffer_bytes.extend(res.stdout_data)
@@ -1713,7 +1666,6 @@ def connect_node(self, unique_id, sftp=False, debug=False, logger=None):
     
     request_queue = queue.Queue()
     client_buffer_bytes = bytearray()
-    cmd_byte_positions = [(0, None)]
     pause_stdin = [False]
     wake_r, wake_w = os.pipe()
 
@@ -1760,8 +1712,6 @@ def connect_node(self, unique_id, sftp=False, debug=False, logger=None):
                     data = os.read(sys.stdin.fileno(), 1024)
                     if not data:
                         break
-                    if b'\r' in data or b'\n' in data:
-                        cmd_byte_positions.append((len(client_buffer_bytes), None))
                     yield connpy_pb2.InteractRequest(stdin_data=data)
                 except OSError:
                     break
@@ -1835,14 +1785,11 @@ def connect_node(self, unique_id, sftp=False, debug=False, logger=None):
             if res.copilot_prompt:
                 self._handle_remote_copilot(
                     res, request_queue, response_queue, 
-                    client_buffer_bytes, cmd_byte_positions, 
+                    client_buffer_bytes, 
                     pause_generator, resume_generator, old_tty
                 )
                 continue
 
-            if res.copilot_injected_command:
-                cmd_byte_positions.append((len(client_buffer_bytes), res.copilot_injected_command))
-
             if res.stdout_data:
                 os.write(sys.stdout.fileno(), res.stdout_data)
                 client_buffer_bytes.extend(res.stdout_data)

docs/connpy/index.html

@@ -84,10 +84,10 @@ cd connpy
 docker compose build
 
 # Run it like a native app (completely silent)
-docker compose --log-level ERROR run --rm --remove-orphans connpy-app [command]
+docker compose run --rm --remove-orphans connpy-app [command]
 
 # Pro Tip: Add this alias for a 100% native experience from any folder
-alias conn='docker compose -f /path/to/connpy/docker-compose.yml --log-level ERROR run --rm --remove-orphans connpy-app'
+alias conn='docker compose -f /path/to/connpy/docker-compose.yml run --rm --remove-orphans connpy-app'
 </code></pre>
 <hr>
 <h2 id="privacy-integration">🔒 Privacy &amp; Integration</h2>
@@ -737,6 +737,7 @@ class ai:
             - COMPLETE MISSIONS: Execute ALL steps of a mission before reporting back.
             - DIAGRAM: Use ASCII art or Unicode box-drawing characters directly in your responses to visualize topologies or paths when helpful.
             - EVIDENCE: Include &#39;Key Snippets&#39; from tool outputs. Be token-efficient.
+            - LANGUAGE: You MUST respond in the same language used by the user in their question or instruction.
             - NO WANDERING: Do not speculate. If stuck, report attempts.
             - SAFETY: When you use &#39;run_commands&#39; with configuration commands, the system automatically prompts the user for confirmation. Just execute - don&#39;t ask permission first.
 {architect_instructions}
@@ -754,6 +755,7 @@ class ai:
             - ENGINEER CAPABILITIES: Your Engineer can:
                 * Filter nodes (list_nodes), Run CLI commands (run_commands), Get metadata (get_node_info).
             - ANALYSIS: Review technical findings to identify patterns or design failures.
+            - LANGUAGE: You MUST respond in the same language used by the user in their question or instruction.
             - MEMORY: Update long-term facts ONLY when the user explicitly requests it.
             
             CRITICAL - EFFICIENT DELEGATION:
@@ -829,7 +831,6 @@ class ai:
         - response: reconstructed ModelResponse (same as non-streaming)
         - streamed: True if text was rendered to console during streaming
         &#34;&#34;&#34;
-        from rich.live import Live
 
         stream_resp = completion(model=model, messages=messages, tools=tools, api_key=api_key, stream=True, **kwargs)
 
@@ -837,7 +838,7 @@ class ai:
         full_content = &#34;&#34;
         is_streaming_text = False
         has_tool_calls = False
-        live_display = None
+        header_printed = False
 
         # Determine styling based on current brain
         role_label = &#34;Network Architect&#34; if &#34;architect&#34; in label.lower() else &#34;Network Engineer&#34;
@@ -866,7 +867,6 @@ class ai:
 
                     if not chunk_callback:
                         if not is_streaming_text:
-                            # Stop spinner definitively
                             if status:
                                 try:
                                     status.stop()
@@ -875,35 +875,28 @@ class ai:
                             
                             # Create a stable, direct Console to bypass _ConsoleProxy recreation bugs
                             from rich.console import Console as RichConsole
-                            from .printer import connpy_theme, get_original_stdout
+                            from rich.rule import Rule
+                            from .printer import connpy_theme, get_original_stdout, IncrementalMarkdownParser
                             stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
                             
-                            live_display = Live(
-                                Panel(Markdown(full_content), title=title, border_style=border, expand=False),
-                                console=stable_console,
-                                refresh_per_second=8,
-                                transient=False
-                            )
-                            live_display.start()
+                            stable_console.print(Rule(f&#34;[bold {border}]{title}[/bold {border}]&#34;, style=border))
+                            header_printed = True
+                            md_parser = IncrementalMarkdownParser(console=stable_console)
                             is_streaming_text = True
-                        else:
-                            live_display.update(
-                                Panel(Markdown(full_content), title=title, border_style=border, expand=False)
-                            )
+                        
+                        md_parser.feed(delta.content)
         except Exception as e:
             if not chunks:
                 raise
         finally:
-            if live_display:
-                # Render final state with complete content
+            if header_printed:
                 try:
-                    live_display.update(
-                        Panel(Markdown(full_content), title=title, border_style=border, expand=False)
-                    )
-                except Exception:
-                    pass
-                try:
-                    live_display.stop()
+                    md_parser.flush()
+                    from rich.console import Console as RichConsole
+                    from rich.rule import Rule
+                    from .printer import connpy_theme, get_original_stdout
+                    stable_console = RichConsole(theme=connpy_theme, file=get_original_stdout())
+                    stable_console.print(Rule(style=border))
                 except Exception:
                     pass
         
@@ -1852,11 +1845,13 @@ class ai:
         if persona == &#34;architect&#34;:
             system_prompt = f&#34;&#34;&#34;Role: NETWORK ARCHITECT. You act as a senior strategic advisor during a live SSH session.
 Rules:
-1. Answer the user&#39;s question directly based on the Terminal Context.
-2. Focus on the &#34;why&#34; and &#34;how&#34;. Analyze topologies, design patterns, and validate configurations.
-3. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
-4. Keep your guide concise and authoritative.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. Answer the user&#39;s question directly and EXCLUSIVELY based on the Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like &#39;iol#&#39; or &#39;admin@vrouter&gt;&#39;) and no command output, it means YOU DON&#39;T HAVE DATA. In this case, YOU MUST NOT invent any information.
+4. Focus on the &#34;why&#34; and &#34;how&#34;. Analyze topologies, design patterns, and validate configurations.
+5. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
+6. Keep your guide concise and authoritative.
+7. You MUST output your response in the following strict format:
 &lt;guide&gt;
 Your brief tactical guide in markdown.
 &lt;/guide&gt;
@@ -1865,7 +1860,7 @@ Your brief tactical guide in markdown.
 &lt;risk&gt;
 low
 &lt;/risk&gt;
-6. Risk level is usually &#34;low&#34; for read-only/no commands.
+8. Risk level is usually &#34;low&#34; for read-only/no commands.
 
 Terminal Context:
 {terminal_buffer}
@@ -1875,11 +1870,13 @@ Node: {node_name}&#34;&#34;&#34;
         else:
             system_prompt = f&#34;&#34;&#34;Role: TERMINAL COPILOT. You assist a network engineer during a live SSH session.
 Rules:
-1. Answer the user&#39;s question directly based on the Terminal Context.
-2. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the &lt;guide&gt; section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
-3. If the user wants to execute an action, provide the required CLI commands inside a &lt;commands&gt; block, one command per line. If no commands are needed, leave it empty or omit the block.
-4. ULTRA-CONCISE. Keep your guide to the point.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. EXTREMELY IMPORTANT: Answer EXCLUSIVELY based on the provided Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like &#39;iol#&#39; or &#39;admin@vrouter&gt;&#39;) and no command output, it means YOU DON&#39;T HAVE DATA. In this case, YOU MUST NOT invent any information. Instead, explicitly state that you don&#39;t see the data and offer the correct CLI commands to retrieve it.
+4. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the &lt;guide&gt; section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
+5. If the user wants to execute an action, provide the required CLI commands inside a &lt;commands&gt; block, one command per line. If no commands are needed, leave it empty or omit the block.
+6. ULTRA-CONCISE. Keep your guide to the point.
+7. You MUST output your response in the following strict format:
 &lt;guide&gt;
 Your brief tactical guide in markdown. 3-4 sentences max.
 &lt;/guide&gt;
@@ -1890,7 +1887,7 @@ command 2
 &lt;risk&gt;
 low, high, or destructive
 &lt;/risk&gt;
-6. Risk level: &#34;low&#34; for read-only/no commands, &#34;high&#34; for config changes, &#34;destructive&#34; for potentially dangerous ops.
+8. Risk level: &#34;low&#34; for read-only/no commands, &#34;high&#34; for config changes, &#34;destructive&#34; for potentially dangerous ops.
 
 Terminal Context:
 {terminal_buffer}
@@ -1932,6 +1929,7 @@ Node: {node_name}&#34;&#34;&#34;
         try:
             while iteration &lt; max_iterations:
                 iteration += 1
+
                 response = await acompletion(
                     model=current_model,
                     messages=messages,
@@ -2177,11 +2175,13 @@ def engineer_system_prompt(self):
         if persona == &#34;architect&#34;:
             system_prompt = f&#34;&#34;&#34;Role: NETWORK ARCHITECT. You act as a senior strategic advisor during a live SSH session.
 Rules:
-1. Answer the user&#39;s question directly based on the Terminal Context.
-2. Focus on the &#34;why&#34; and &#34;how&#34;. Analyze topologies, design patterns, and validate configurations.
-3. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
-4. Keep your guide concise and authoritative.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. Answer the user&#39;s question directly and EXCLUSIVELY based on the Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like &#39;iol#&#39; or &#39;admin@vrouter&gt;&#39;) and no command output, it means YOU DON&#39;T HAVE DATA. In this case, YOU MUST NOT invent any information.
+4. Focus on the &#34;why&#34; and &#34;how&#34;. Analyze topologies, design patterns, and validate configurations.
+5. Do NOT provide commands to execute unless specifically requested. Instead, explain the consequences and best practices.
+6. Keep your guide concise and authoritative.
+7. You MUST output your response in the following strict format:
 &lt;guide&gt;
 Your brief tactical guide in markdown.
 &lt;/guide&gt;
@@ -2190,7 +2190,7 @@ Your brief tactical guide in markdown.
 &lt;risk&gt;
 low
 &lt;/risk&gt;
-6. Risk level is usually &#34;low&#34; for read-only/no commands.
+8. Risk level is usually &#34;low&#34; for read-only/no commands.
 
 Terminal Context:
 {terminal_buffer}
@@ -2200,11 +2200,13 @@ Node: {node_name}&#34;&#34;&#34;
         else:
             system_prompt = f&#34;&#34;&#34;Role: TERMINAL COPILOT. You assist a network engineer during a live SSH session.
 Rules:
-1. Answer the user&#39;s question directly based on the Terminal Context.
-2. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the &lt;guide&gt; section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
-3. If the user wants to execute an action, provide the required CLI commands inside a &lt;commands&gt; block, one command per line. If no commands are needed, leave it empty or omit the block.
-4. ULTRA-CONCISE. Keep your guide to the point.
-5. You MUST output your response in the following strict format:
+1. MANDATORY: You MUST respond in the same language used by the user in their question.
+2. EXTREMELY IMPORTANT: Answer EXCLUSIVELY based on the provided Terminal Context. 
+3. NO HALLUCINATIONS. The Terminal Context is a live buffer. If it contains only a shell prompt (like &#39;iol#&#39; or &#39;admin@vrouter&gt;&#39;) and no command output, it means YOU DON&#39;T HAVE DATA. In this case, YOU MUST NOT invent any information. Instead, explicitly state that you don&#39;t see the data and offer the correct CLI commands to retrieve it.
+4. If the user asks you to analyze, parse, or extract data from the Terminal Context, DO IT directly in the &lt;guide&gt; section (you can use markdown tables or lists). Do NOT just give them a command to do it themselves.
+5. If the user wants to execute an action, provide the required CLI commands inside a &lt;commands&gt; block, one command per line. If no commands are needed, leave it empty or omit the block.
+6. ULTRA-CONCISE. Keep your guide to the point.
+7. You MUST output your response in the following strict format:
 &lt;guide&gt;
 Your brief tactical guide in markdown. 3-4 sentences max.
 &lt;/guide&gt;
@@ -2215,7 +2217,7 @@ command 2
 &lt;risk&gt;
 low, high, or destructive
 &lt;/risk&gt;
-6. Risk level: &#34;low&#34; for read-only/no commands, &#34;high&#34; for config changes, &#34;destructive&#34; for potentially dangerous ops.
+8. Risk level: &#34;low&#34; for read-only/no commands, &#34;high&#34; for config changes, &#34;destructive&#34; for potentially dangerous ops.
 
 Terminal Context:
 {terminal_buffer}
@@ -2257,6 +2259,7 @@ Node: {node_name}&#34;&#34;&#34;
         try:
             while iteration &lt; max_iterations:
                 iteration += 1
+
                 response = await acompletion(
                     model=current_model,
                     messages=messages,
@@ -4119,6 +4122,7 @@ class node:
         self.output = &#34;&#34;
         self.status = 1
         self.result = {}
+        self.cmd_byte_positions = [(0, None)]
 
     @MethodHook
     def _passtx(self, passwords, *, keyfile=None):
@@ -4293,9 +4297,9 @@ class node:
             loop = asyncio.get_running_loop()
             child_reader_queue = asyncio.Queue()
             
-            # Track command byte positions for copilot context navigation
+            # Reset and track command byte positions for copilot context navigation
             # Each entry is (byte_position, command_text_or_None)
-            cmd_byte_positions = [(0, None)]
+            self.cmd_byte_positions = [(self.mylog.tell() if hasattr(self, &#39;mylog&#39;) else 0, None)]
             
             def _child_read_ready():
                 try:
@@ -4336,7 +4340,7 @@ class node:
                             node_info[&#34;prompt&#34;] = to_str(self.tags.get(&#34;prompt&#34;, r&#39;&gt;$|#$|\$$|&gt;.$|#.$|\$.$&#39;))
                         
                         # Invoke copilot (async callback handles UI)
-                        await copilot_handler(self.mylog.getvalue(), node_info, local_stream, child_fd, cmd_byte_positions)
+                        await copilot_handler(self.mylog.getvalue(), node_info, local_stream, child_fd, self.cmd_byte_positions)
                         continue
                     
                     # Remove any stray \x00 bytes and forward normally
@@ -4344,10 +4348,9 @@ class node:
                     if clean_data:
                         # Track command boundaries when user hits Enter
                         if hasattr(self, &#39;mylog&#39;) and (b&#39;\r&#39; in clean_data or b&#39;\n&#39; in clean_data):
-                            cmd_byte_positions.append((self.mylog.tell(), None))
+                            self.cmd_byte_positions.append((self.mylog.tell(), None))
 
-                        try:
-                            os.write(child_fd, clean_data)
+                        try:                            os.write(child_fd, clean_data)
                         except OSError:
                             break
                         self.lastinput = time()
@@ -4469,6 +4472,45 @@ class node:
         finally:
             local_stream.teardown()
 
+    @MethodHook
+    async def inject_commands(self, commands, child_fd, on_inject=None):
+        &#34;&#34;&#34;
+        Inject a list of commands into the node&#39;s PTY.
+        Handles screen_length_command, history tracking and delays.
+        &#34;&#34;&#34;
+        if not commands:
+            return
+
+        # 0. Clear line
+        os.write(child_fd, b&#39;\x15&#39;)
+        await asyncio.sleep(0.1)
+
+        # 1. Prepare list (prepend screen_length if exists)
+        slc = self.tags.get(&#34;screen_length_command&#34;) if hasattr(self, &#39;tags&#39;) and isinstance(self.tags, dict) else None
+        
+        to_send = list(commands)
+        if slc and slc not in to_send: # avoid duplicates if already there
+             to_send.insert(0, slc)
+
+        # 2. Inject one by one
+        for cmd in to_send:
+            # Register in node&#39;s official history (SKIP if it&#39;s the administrative screen length command)
+            if cmd != slc and hasattr(self, &#39;cmd_byte_positions&#39;) and self.cmd_byte_positions is not None:
+                log_pos = self.mylog.tell() if hasattr(self, &#39;mylog&#39;) else 0
+                self.cmd_byte_positions.append((log_pos, cmd))
+            
+            # Write physically to PTY
+            os.write(child_fd, (cmd + &#34;\n&#34;).encode())
+            
+            # Notify (e.g., for gRPC or logs) - SKIP for administrative SLC
+            if on_inject and cmd != slc:
+                if asyncio.iscoroutinefunction(on_inject):
+                    await on_inject(cmd)
+                else:
+                    on_inject(cmd)
+            
+            # Delay to avoid overwhelming the router
+            await asyncio.sleep(0.8)
 
     @MethodHook
     def interact(self, debug=False, logger=None):
@@ -4550,7 +4592,7 @@ class node:
                         while True:
                             action, commands, custom_cmd = await interface.run_session(
                                 raw_bytes=raw_bytes,
-                                cmd_byte_positions=cmd_byte_positions,
+                                cmd_byte_positions=self.cmd_byte_positions,
                                 node_info=node_info,
                                 on_ai_call=on_ai_call
                             )
@@ -4558,6 +4600,7 @@ class node:
                                 continue
                             break
                 finally:
+                    print(&#34;\033[2m Returning to session...\033[0m&#34;, flush=True)
                     # Reiniciar el lector de la terminal para volver al modo interactivo SSH/Telnet
                     if hasattr(stream, &#39;start_reading&#39;):
                         stream.start_reading()
@@ -4566,20 +4609,7 @@ class node:
                 
                 if action in (&#34;send_all&#34;, &#34;custom&#34;):
                     cmds_to_send = commands if action == &#34;send_all&#34; else custom_cmd
-                    
-                    if cmds_to_send:
-                        os.write(child_fd, b&#39;\x15&#39;) # Ctrl+U
-                        await asyncio.sleep(0.1)
-                        
-                        # Prepend screen length command to avoid pagination
-                        if &#34;screen_length_command&#34; in self.tags:
-                            cmds_to_send.insert(0, self.tags[&#34;screen_length_command&#34;])
-
-                        for cmd in cmds_to_send:
-                            if cmd_byte_positions is not None:
-                                cmd_byte_positions.append((self.mylog.tell(), cmd))
-                            os.write(child_fd, (cmd + &#34;\n&#34;).encode())
-                            await asyncio.sleep(0.8)
+                    await self.inject_commands(cmds_to_send, child_fd)
                 else:
                     os.write(child_fd, b&#39;\x15\r&#39;)
             except Exception as e:
@@ -5073,6 +5103,57 @@ class node:
 </code></pre></div>
 <h3>Methods</h3>
 <dl>
+<dt id="connpy.node.inject_commands"><code class="name flex">
+<span>async def <span class="ident">inject_commands</span></span>(<span>self, commands, child_fd, on_inject=None)</span>
+</code></dt>
+<dd>
+<details class="source">
+<summary>
+<span>Expand source code</span>
+</summary>
+<pre><code class="python">@MethodHook
+async def inject_commands(self, commands, child_fd, on_inject=None):
+    &#34;&#34;&#34;
+    Inject a list of commands into the node&#39;s PTY.
+    Handles screen_length_command, history tracking and delays.
+    &#34;&#34;&#34;
+    if not commands:
+        return
+
+    # 0. Clear line
+    os.write(child_fd, b&#39;\x15&#39;)
+    await asyncio.sleep(0.1)
+
+    # 1. Prepare list (prepend screen_length if exists)
+    slc = self.tags.get(&#34;screen_length_command&#34;) if hasattr(self, &#39;tags&#39;) and isinstance(self.tags, dict) else None
+    
+    to_send = list(commands)
+    if slc and slc not in to_send: # avoid duplicates if already there
+         to_send.insert(0, slc)
+
+    # 2. Inject one by one
+    for cmd in to_send:
+        # Register in node&#39;s official history (SKIP if it&#39;s the administrative screen length command)
+        if cmd != slc and hasattr(self, &#39;cmd_byte_positions&#39;) and self.cmd_byte_positions is not None:
+            log_pos = self.mylog.tell() if hasattr(self, &#39;mylog&#39;) else 0
+            self.cmd_byte_positions.append((log_pos, cmd))
+        
+        # Write physically to PTY
+        os.write(child_fd, (cmd + &#34;\n&#34;).encode())
+        
+        # Notify (e.g., for gRPC or logs) - SKIP for administrative SLC
+        if on_inject and cmd != slc:
+            if asyncio.iscoroutinefunction(on_inject):
+                await on_inject(cmd)
+            else:
+                on_inject(cmd)
+        
+        # Delay to avoid overwhelming the router
+        await asyncio.sleep(0.8)</code></pre>
+</details>
+<div class="desc"><p>Inject a list of commands into the node's PTY.
+Handles screen_length_command, history tracking and delays.</p></div>
+</dd>
 <dt id="connpy.node.interact"><code class="name flex">
 <span>def <span class="ident">interact</span></span>(<span>self, debug=False, logger=None)</span>
 </code></dt>
@@ -6189,6 +6270,7 @@ def test(self, commands, expected, vars = None,*, folder = None, prompt = None,
 <li>
 <h4><code><a title="connpy.node" href="#connpy.node">node</a></code></h4>
 <ul class="">
+<li><code><a title="connpy.node.inject_commands" href="#connpy.node.inject_commands">inject_commands</a></code></li>
 <li><code><a title="connpy.node.interact" href="#connpy.node.interact">interact</a></code></li>
 <li><code><a title="connpy.node.run" href="#connpy.node.run">run</a></code></li>
 <li><code><a title="connpy.node.test" href="#connpy.node.test">test</a></code></li>

docs/connpy/services/ai_service.html

@@ -58,10 +58,10 @@ el.replaceWith(d);
 <pre><code class="python">class AIService(BaseService):
     &#34;&#34;&#34;Business logic for interacting with AI agents and LLM configurations.&#34;&#34;&#34;
 
-    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) -&gt; list:
+    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict, last_line: str = &#34;&#34;) -&gt; list:
         &#34;&#34;&#34;Identifies command blocks in the terminal history.&#34;&#34;&#34;
         blocks = []
-        if not (cmd_byte_positions and len(cmd_byte_positions) &gt;= 2 and raw_bytes):
+        if not raw_bytes:
             return blocks
             
         default_prompt = r&#39;&gt;$|#$|\$$|&gt;.$|#.$|\$.$&#39;
@@ -72,29 +72,63 @@ el.replaceWith(d);
         except Exception:
             prompt_re = re.compile(re.sub(r&#39;(?&lt;!\\)\$&#39;, &#39;&#39;, default_prompt))
             
-        for i in range(1, len(cmd_byte_positions)):
-            pos, known_cmd = cmd_byte_positions[i]
-            prev_pos = cmd_byte_positions[i-1][0]
+        parsed_positions = []
+        if cmd_byte_positions and len(cmd_byte_positions) &gt;= 1:
+            for i in range(1, len(cmd_byte_positions)):
+                pos, known_cmd = cmd_byte_positions[i]
+                prev_pos = cmd_byte_positions[i-1][0]
                 
-            if known_cmd:
-                prev_chunk = raw_bytes[prev_pos:pos]
-                prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
-                prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
-                prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
-                preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
-                blocks.append((pos, preview[:80]))
-            else:
-                chunk = raw_bytes[prev_pos:pos]
-                cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
-                lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
-                preview = lines[-1].strip() if lines else &#34;&#34;
+                if known_cmd:
+                    prev_chunk = raw_bytes[prev_pos:pos]
+                    prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
+                    prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
+                    prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
+                    preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
+                    parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                else:
+                    chunk = raw_bytes[prev_pos:pos]
+                    cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
+                    lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
+                    preview = lines[-1].strip() if lines else &#34;&#34;
+                    
+                    if preview:
+                        match = prompt_re.search(preview)
+                        if match:
+                            cmd_text = preview[match.end():].strip()
+                            if cmd_text:
+                                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                            else:
+                                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;EMPTY_PROMPT&#34;, &#34;preview&#34;: &#34;&#34;})
+                        else:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+                    else:
+                        parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+
+        last_newline = raw_bytes.rfind(b&#39;\n&#39;)
+        current_prompt_pos = last_newline + 1 if last_newline != -1 else 0
+        current_end = len(raw_bytes)
+
+        for i, item in enumerate(parsed_positions):
+            if item[&#34;type&#34;] == &#34;VALID_CMD&#34;:
+                start_pos = item[&#34;pos&#34;]
+                preview = item[&#34;preview&#34;]
+                
+                # Find the end position: next VALID_CMD or EMPTY_PROMPT
+                end_pos = current_prompt_pos
+                for j in range(i + 1, len(parsed_positions)):
+                    next_item = parsed_positions[j]
+                    if next_item[&#34;type&#34;] in (&#34;VALID_CMD&#34;, &#34;EMPTY_PROMPT&#34;):
+                        end_pos = next_item[&#34;pos&#34;]
+                        break
+                
+                blocks.append((start_pos, end_pos, preview))
+
+        # Always ensure there is a final block representing the current prompt
+        if not blocks:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
+        elif blocks[-1][0] &lt; current_prompt_pos:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
 
-                if preview:
-                    match = prompt_re.search(preview)
-                    if match:
-                        cmd_text = preview[match.end():].strip()
-                        if cmd_text:
-                            blocks.append((pos, preview[:80]))
         return blocks
 
     def process_copilot_input(self, input_text: str, session_state: dict) -&gt; dict:
@@ -317,17 +351,17 @@ el.replaceWith(d);
 <div class="desc"><p>Ask the AI copilot for terminal assistance.</p></div>
 </dd>
 <dt id="connpy.services.ai_service.AIService.build_context_blocks"><code class="name flex">
-<span>def <span class="ident">build_context_blocks</span></span>(<span>self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) ‑> list</span>
+<span>def <span class="ident">build_context_blocks</span></span>(<span>self,<br>raw_bytes: bytes,<br>cmd_byte_positions: list,<br>node_info: dict,<br>last_line: str = '') ‑> list</span>
 </code></dt>
 <dd>
 <details class="source">
 <summary>
 <span>Expand source code</span>
 </summary>
-<pre><code class="python">def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) -&gt; list:
+<pre><code class="python">def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict, last_line: str = &#34;&#34;) -&gt; list:
     &#34;&#34;&#34;Identifies command blocks in the terminal history.&#34;&#34;&#34;
     blocks = []
-    if not (cmd_byte_positions and len(cmd_byte_positions) &gt;= 2 and raw_bytes):
+    if not raw_bytes:
         return blocks
         
     default_prompt = r&#39;&gt;$|#$|\$$|&gt;.$|#.$|\$.$&#39;
@@ -338,29 +372,63 @@ el.replaceWith(d);
     except Exception:
         prompt_re = re.compile(re.sub(r&#39;(?&lt;!\\)\$&#39;, &#39;&#39;, default_prompt))
         
-    for i in range(1, len(cmd_byte_positions)):
-        pos, known_cmd = cmd_byte_positions[i]
-        prev_pos = cmd_byte_positions[i-1][0]
+    parsed_positions = []
+    if cmd_byte_positions and len(cmd_byte_positions) &gt;= 1:
+        for i in range(1, len(cmd_byte_positions)):
+            pos, known_cmd = cmd_byte_positions[i]
+            prev_pos = cmd_byte_positions[i-1][0]
             
-        if known_cmd:
-            prev_chunk = raw_bytes[prev_pos:pos]
-            prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
-            prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
-            prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
-            preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
-            blocks.append((pos, preview[:80]))
-        else:
-            chunk = raw_bytes[prev_pos:pos]
-            cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
-            lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
-            preview = lines[-1].strip() if lines else &#34;&#34;
+            if known_cmd:
+                prev_chunk = raw_bytes[prev_pos:pos]
+                prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
+                prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
+                prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
+                preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
+                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+            else:
+                chunk = raw_bytes[prev_pos:pos]
+                cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
+                lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
+                preview = lines[-1].strip() if lines else &#34;&#34;
+                
+                if preview:
+                    match = prompt_re.search(preview)
+                    if match:
+                        cmd_text = preview[match.end():].strip()
+                        if cmd_text:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                        else:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;EMPTY_PROMPT&#34;, &#34;preview&#34;: &#34;&#34;})
+                    else:
+                        parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+                else:
+                    parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+
+    last_newline = raw_bytes.rfind(b&#39;\n&#39;)
+    current_prompt_pos = last_newline + 1 if last_newline != -1 else 0
+    current_end = len(raw_bytes)
+
+    for i, item in enumerate(parsed_positions):
+        if item[&#34;type&#34;] == &#34;VALID_CMD&#34;:
+            start_pos = item[&#34;pos&#34;]
+            preview = item[&#34;preview&#34;]
+            
+            # Find the end position: next VALID_CMD or EMPTY_PROMPT
+            end_pos = current_prompt_pos
+            for j in range(i + 1, len(parsed_positions)):
+                next_item = parsed_positions[j]
+                if next_item[&#34;type&#34;] in (&#34;VALID_CMD&#34;, &#34;EMPTY_PROMPT&#34;):
+                    end_pos = next_item[&#34;pos&#34;]
+                    break
+            
+            blocks.append((start_pos, end_pos, preview))
+
+    # Always ensure there is a final block representing the current prompt
+    if not blocks:
+        blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
+    elif blocks[-1][0] &lt; current_prompt_pos:
+        blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
 
-            if preview:
-                match = prompt_re.search(preview)
-                if match:
-                    cmd_text = preview[match.end():].strip()
-                    if cmd_text:
-                        blocks.append((pos, preview[:80]))
     return blocks</code></pre>
 </details>
 <div class="desc"><p>Identifies command blocks in the terminal history.</p></div>

docs/connpy/services/index.html

@@ -113,10 +113,10 @@ el.replaceWith(d);
 <pre><code class="python">class AIService(BaseService):
     &#34;&#34;&#34;Business logic for interacting with AI agents and LLM configurations.&#34;&#34;&#34;
 
-    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) -&gt; list:
+    def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict, last_line: str = &#34;&#34;) -&gt; list:
         &#34;&#34;&#34;Identifies command blocks in the terminal history.&#34;&#34;&#34;
         blocks = []
-        if not (cmd_byte_positions and len(cmd_byte_positions) &gt;= 2 and raw_bytes):
+        if not raw_bytes:
             return blocks
             
         default_prompt = r&#39;&gt;$|#$|\$$|&gt;.$|#.$|\$.$&#39;
@@ -127,29 +127,63 @@ el.replaceWith(d);
         except Exception:
             prompt_re = re.compile(re.sub(r&#39;(?&lt;!\\)\$&#39;, &#39;&#39;, default_prompt))
             
-        for i in range(1, len(cmd_byte_positions)):
-            pos, known_cmd = cmd_byte_positions[i]
-            prev_pos = cmd_byte_positions[i-1][0]
+        parsed_positions = []
+        if cmd_byte_positions and len(cmd_byte_positions) &gt;= 1:
+            for i in range(1, len(cmd_byte_positions)):
+                pos, known_cmd = cmd_byte_positions[i]
+                prev_pos = cmd_byte_positions[i-1][0]
                 
-            if known_cmd:
-                prev_chunk = raw_bytes[prev_pos:pos]
-                prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
-                prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
-                prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
-                preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
-                blocks.append((pos, preview[:80]))
-            else:
-                chunk = raw_bytes[prev_pos:pos]
-                cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
-                lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
-                preview = lines[-1].strip() if lines else &#34;&#34;
+                if known_cmd:
+                    prev_chunk = raw_bytes[prev_pos:pos]
+                    prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
+                    prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
+                    prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
+                    preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
+                    parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                else:
+                    chunk = raw_bytes[prev_pos:pos]
+                    cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
+                    lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
+                    preview = lines[-1].strip() if lines else &#34;&#34;
+                    
+                    if preview:
+                        match = prompt_re.search(preview)
+                        if match:
+                            cmd_text = preview[match.end():].strip()
+                            if cmd_text:
+                                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                            else:
+                                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;EMPTY_PROMPT&#34;, &#34;preview&#34;: &#34;&#34;})
+                        else:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+                    else:
+                        parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+
+        last_newline = raw_bytes.rfind(b&#39;\n&#39;)
+        current_prompt_pos = last_newline + 1 if last_newline != -1 else 0
+        current_end = len(raw_bytes)
+
+        for i, item in enumerate(parsed_positions):
+            if item[&#34;type&#34;] == &#34;VALID_CMD&#34;:
+                start_pos = item[&#34;pos&#34;]
+                preview = item[&#34;preview&#34;]
+                
+                # Find the end position: next VALID_CMD or EMPTY_PROMPT
+                end_pos = current_prompt_pos
+                for j in range(i + 1, len(parsed_positions)):
+                    next_item = parsed_positions[j]
+                    if next_item[&#34;type&#34;] in (&#34;VALID_CMD&#34;, &#34;EMPTY_PROMPT&#34;):
+                        end_pos = next_item[&#34;pos&#34;]
+                        break
+                
+                blocks.append((start_pos, end_pos, preview))
+
+        # Always ensure there is a final block representing the current prompt
+        if not blocks:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
+        elif blocks[-1][0] &lt; current_prompt_pos:
+            blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
 
-                if preview:
-                    match = prompt_re.search(preview)
-                    if match:
-                        cmd_text = preview[match.end():].strip()
-                        if cmd_text:
-                            blocks.append((pos, preview[:80]))
         return blocks
 
     def process_copilot_input(self, input_text: str, session_state: dict) -&gt; dict:
@@ -372,17 +406,17 @@ el.replaceWith(d);
 <div class="desc"><p>Ask the AI copilot for terminal assistance.</p></div>
 </dd>
 <dt id="connpy.services.AIService.build_context_blocks"><code class="name flex">
-<span>def <span class="ident">build_context_blocks</span></span>(<span>self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) ‑> list</span>
+<span>def <span class="ident">build_context_blocks</span></span>(<span>self,<br>raw_bytes: bytes,<br>cmd_byte_positions: list,<br>node_info: dict,<br>last_line: str = '') ‑> list</span>
 </code></dt>
 <dd>
 <details class="source">
 <summary>
 <span>Expand source code</span>
 </summary>
-<pre><code class="python">def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict) -&gt; list:
+<pre><code class="python">def build_context_blocks(self, raw_bytes: bytes, cmd_byte_positions: list, node_info: dict, last_line: str = &#34;&#34;) -&gt; list:
     &#34;&#34;&#34;Identifies command blocks in the terminal history.&#34;&#34;&#34;
     blocks = []
-    if not (cmd_byte_positions and len(cmd_byte_positions) &gt;= 2 and raw_bytes):
+    if not raw_bytes:
         return blocks
         
     default_prompt = r&#39;&gt;$|#$|\$$|&gt;.$|#.$|\$.$&#39;
@@ -393,29 +427,63 @@ el.replaceWith(d);
     except Exception:
         prompt_re = re.compile(re.sub(r&#39;(?&lt;!\\)\$&#39;, &#39;&#39;, default_prompt))
         
-    for i in range(1, len(cmd_byte_positions)):
-        pos, known_cmd = cmd_byte_positions[i]
-        prev_pos = cmd_byte_positions[i-1][0]
+    parsed_positions = []
+    if cmd_byte_positions and len(cmd_byte_positions) &gt;= 1:
+        for i in range(1, len(cmd_byte_positions)):
+            pos, known_cmd = cmd_byte_positions[i]
+            prev_pos = cmd_byte_positions[i-1][0]
             
-        if known_cmd:
-            prev_chunk = raw_bytes[prev_pos:pos]
-            prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
-            prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
-            prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
-            preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
-            blocks.append((pos, preview[:80]))
-        else:
-            chunk = raw_bytes[prev_pos:pos]
-            cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
-            lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
-            preview = lines[-1].strip() if lines else &#34;&#34;
+            if known_cmd:
+                prev_chunk = raw_bytes[prev_pos:pos]
+                prev_cleaned = log_cleaner(prev_chunk.decode(errors=&#39;replace&#39;))
+                prev_lines = [l for l in prev_cleaned.split(&#39;\n&#39;) if l.strip()]
+                prompt_text = prev_lines[-1].strip() if prev_lines else &#34;&#34;
+                preview = f&#34;{prompt_text}{known_cmd}&#34; if prompt_text else known_cmd
+                parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+            else:
+                chunk = raw_bytes[prev_pos:pos]
+                cleaned = log_cleaner(chunk.decode(errors=&#39;replace&#39;))
+                lines = [l for l in cleaned.split(&#39;\n&#39;) if l.strip()]
+                preview = lines[-1].strip() if lines else &#34;&#34;
+                
+                if preview:
+                    match = prompt_re.search(preview)
+                    if match:
+                        cmd_text = preview[match.end():].strip()
+                        if cmd_text:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;VALID_CMD&#34;, &#34;preview&#34;: preview[:80]})
+                        else:
+                            parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;EMPTY_PROMPT&#34;, &#34;preview&#34;: &#34;&#34;})
+                    else:
+                        parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+                else:
+                    parsed_positions.append({&#34;pos&#34;: pos, &#34;type&#34;: &#34;SCROLLING&#34;, &#34;preview&#34;: &#34;&#34;})
+
+    last_newline = raw_bytes.rfind(b&#39;\n&#39;)
+    current_prompt_pos = last_newline + 1 if last_newline != -1 else 0
+    current_end = len(raw_bytes)
+
+    for i, item in enumerate(parsed_positions):
+        if item[&#34;type&#34;] == &#34;VALID_CMD&#34;:
+            start_pos = item[&#34;pos&#34;]
+            preview = item[&#34;preview&#34;]
+            
+            # Find the end position: next VALID_CMD or EMPTY_PROMPT
+            end_pos = current_prompt_pos
+            for j in range(i + 1, len(parsed_positions)):
+                next_item = parsed_positions[j]
+                if next_item[&#34;type&#34;] in (&#34;VALID_CMD&#34;, &#34;EMPTY_PROMPT&#34;):
+                    end_pos = next_item[&#34;pos&#34;]
+                    break
+            
+            blocks.append((start_pos, end_pos, preview))
+
+    # Always ensure there is a final block representing the current prompt
+    if not blocks:
+        blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
+    elif blocks[-1][0] &lt; current_prompt_pos:
+        blocks.append((current_prompt_pos, current_end, last_line[:80] if last_line else &#34;CURRENT CONTEXT&#34;))
 
-            if preview:
-                match = prompt_re.search(preview)
-                if match:
-                    cmd_text = preview[match.end():].strip()
-                    if cmd_text:
-                        blocks.append((pos, preview[:80]))
     return blocks</code></pre>
 </details>
 <div class="desc"><p>Identifies command blocks in the terminal history.</p></div>

docs/connpy/utils.html

@@ -75,14 +75,34 @@ el.replaceWith(d);
             elif token in ('\b', '\x7f'):
                 if cursor > 0:
                     cursor -= 1
-            elif token == '\x1B[D': # Left Arrow
-                if cursor > 0:
-                    cursor -= 1
-            elif token == '\x1B[C': # Right Arrow
-                if cursor < len(buffer):
-                    cursor += 1
-            elif token == '\x1B[K': # Clear to end of line
-                buffer = buffer[:cursor]
+            elif token.startswith('\x1B[') and len(token) >= 3:
+                # Parse CSI: \x1B[ <params> <final_char>
+                final = token[-1]
+                param_str = token[2:-1]
+                n = int(param_str) if param_str.isdigit() else 1
+
+                if final == 'D':   # CUB – Cursor Back
+                    cursor = max(0, cursor - n)
+                elif final == 'C': # CUF – Cursor Forward
+                    cursor = min(len(buffer), cursor + n)
+                elif final == 'K': # EL  – Erase in Line
+                    if n == 0 or param_str == '':      # Clear to end
+                        buffer = buffer[:cursor]
+                    elif n == 1:                         # Clear to start
+                        buffer[:cursor] = [' '] * cursor
+                    elif n == 2:                         # Clear entire line
+                        buffer = []
+                        cursor = 0
+                elif final == 'G': # CHA – Cursor Horizontal Absolute (1-indexed)
+                    cursor = max(0, n - 1)
+                    # Pad buffer if cursor is beyond current length
+                    if cursor > len(buffer):
+                        buffer.extend([' '] * (cursor - len(buffer)))
+                elif final == 'P': # DCH – Delete Characters
+                    del buffer[cursor:cursor + n]
+                elif final == '@': # ICH – Insert Characters
+                    buffer[cursor:cursor] = [' '] * n
+                # All other CSI sequences are silently discarded
             elif token.startswith('\x1B'):
                 continue
             elif len(token) == 1 and ord(token) < 32:

setup.cfg

@@ -60,4 +60,4 @@ console_scripts =
 connpy = 
     core_plugins/*
     proto/*
-    grpc/*.proto
+    grpc_layer/*