glaip-sdk 0.6.10__py3-none-any.whl → 0.6.19__py3-none-any.whl

glaip_sdk/hitl/local.py

@@ -0,0 +1,151 @@
+"""Local HITL prompt handler with interactive console support.
+
+Author:
+    Putu Ravindra Wiguna (putu.r.wiguna@gdplabs.id)
+"""
+
+import os
+from typing import Any
+
+try:
+    from aip_agents.agent.hitl.prompt.base import BasePromptHandler
+    from aip_agents.schema.hitl import ApprovalDecision, ApprovalDecisionType, ApprovalRequest
+except ImportError as e:
+    raise ImportError("aip_agents is required for local HITL. Install with: pip install 'glaip-sdk[local]'") from e
+
+from rich.console import Console
+from rich.prompt import Prompt
+
+
+class LocalPromptHandler(BasePromptHandler):
+    """Local HITL prompt handler with interactive console prompts.
+
+    Experimental local HITL implementation with known limitations:
+    - Timeouts are not enforced (interactive prompts wait indefinitely)
+    - Relies on private renderer methods for pause/resume
+    - Only supports interactive terminal environments
+
+    The key insight from Rich documentation is that Live must be stopped before
+    using Prompt/input(), otherwise the input won't render properly.
+
+    Environment variables:
+        GLAIP_HITL_AUTO_APPROVE: Set to "true" (case-insensitive) to auto-approve
+            all requests without user interaction. Useful for integration tests and CI.
+    """
+
+    def __init__(self, *, pause_resume_callback: Any | None = None) -> None:
+        """Initialize the prompt handler.
+
+        Args:
+            pause_resume_callback: Optional callable with pause() and resume() methods
+                to control the live renderer during prompts. This is needed because
+                Rich Live interferes with Prompt/input().
+        """
+        super().__init__()
+        self._pause_resume = pause_resume_callback
+        self._console = Console()
+
+    async def prompt_for_decision(
+        self,
+        request: ApprovalRequest,
+        timeout_seconds: int,
+        context_keys: list[str] | None = None,
+    ) -> ApprovalDecision:
+        """Prompt for approval decision with live renderer pause/resume.
+
+        Supports auto-approval via GLAIP_HITL_AUTO_APPROVE environment variable
+        for integration testing and CI environments. Set to "true" (case-insensitive) to enable.
+        """
+        _ = (timeout_seconds, context_keys)  # Suppress unused parameter warnings.
+
+        # Check for auto-approve mode (for integration tests/CI)
+        auto_approve = os.getenv("GLAIP_HITL_AUTO_APPROVE", "").lower() == "true"
+
+        if auto_approve:
+            # Auto-approve without user interaction
+            return ApprovalDecision(
+                request_id=request.request_id,
+                decision=ApprovalDecisionType.APPROVED,
+                operator_input="auto-approved",
+            )
+
+        # Pause the live renderer if callback is available
+        if self._pause_resume:
+            self._pause_resume.pause()
+
+        try:
+            # POC/MVP: Show what we're approving (still auto-approve for now)
+            self._print_request_info(request)
+
+            # POC/MVP: For testing, we can do actual input here
+            # Uncomment to enable real prompting:
+            response = Prompt.ask(
+                "\n[yellow]Approve this tool call?[/yellow] [dim](y/n/s)[/dim]",
+                console=self._console,
+                default="y",
+            )
+            response = response.lower().strip()
+
+            if response in ("y", "yes"):
+                decision = ApprovalDecisionType.APPROVED
+            elif response in ("n", "no"):
+                decision = ApprovalDecisionType.REJECTED
+            else:
+                decision = ApprovalDecisionType.SKIPPED
+
+            return ApprovalDecision(
+                request_id=request.request_id,
+                decision=decision,
+                operator_input=response if decision != ApprovalDecisionType.SKIPPED else None,
+            )
+        finally:
+            # Always resume the live renderer
+            if self._pause_resume:
+                self._pause_resume.resume()
+
+    def _print_request_info(self, request: ApprovalRequest) -> None:
+        """Print the approval request information."""
+        self._console.print()
+        self._console.rule("[yellow]HITL Approval Request[/yellow]", style="yellow")
+
+        tool_name = request.tool_name or "unknown"
+        self._console.print(f"[cyan]Tool:[/cyan] {tool_name}")
+
+        if hasattr(request, "arguments_preview") and request.arguments_preview:
+            self._console.print(f"[cyan]Arguments:[/cyan] {request.arguments_preview}")
+
+        if request.context:
+            self._console.print(f"[dim]Context: {request.context}[/dim]")
+
+
+class PauseResumeCallback:
+    """Simple callback object for pausing/resuming the live renderer.
+
+    This allows the LocalPromptHandler to control the renderer without
+    directly coupling to the renderer implementation.
+    """
+
+    def __init__(self) -> None:
+        """Initialize the callback."""
+        self._renderer: Any | None = None
+
+    def set_renderer(self, renderer: Any) -> None:
+        """Set the renderer instance.
+
+        Args:
+            renderer: RichStreamRenderer instance with pause_live() and resume_live() methods.
+        """
+        self._renderer = renderer
+
+    def pause(self) -> None:
+        """Pause the live renderer before prompting."""
+        if self._renderer and hasattr(self._renderer, "_shutdown_live"):
+            self._renderer._shutdown_live()
+
+    def resume(self) -> None:
+        """Resume the live renderer after prompting."""
+        if self._renderer and hasattr(self._renderer, "_ensure_live"):
+            self._renderer._ensure_live()
+
+
+__all__ = ["LocalPromptHandler", "PauseResumeCallback"]

glaip_sdk/runner/deps.py

@@ -15,6 +15,8 @@ Example:
 
 from __future__ import annotations
 
+import importlib.util
+
 from gllm_core.utils import LoggerManager
 
 logger = LoggerManager().get_logger(__name__)
@@ -24,17 +26,12 @@ _local_runtime_available: bool | None = None
 
 
 def _probe_aip_agents_import() -> bool:
-    """Attempt to import aip_agents and return success status.
+    """Check if aip_agents is available without importing it.
 
     Returns:
-        True if aip_agents can be imported successfully, False otherwise.
+        True if aip_agents appears importable, False otherwise.
     """
-    try:
-        import aip_agents  # noqa: F401, PLC0415
-
-        return True
-    except ImportError:
-        return False
+    return importlib.util.find_spec("aip_agents") is not None
 
 
 def check_local_runtime_available() -> bool:

glaip_sdk/runner/langgraph.py

@@ -19,26 +19,56 @@ from __future__ import annotations
 
 import asyncio
 import inspect
+import logging
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any
 
+from aip_agents.agent.hitl.manager import ApprovalManager  # noqa: PLC0415
+from gllm_core.utils import LoggerManager
+
+from glaip_sdk.client.run_rendering import AgentRunRenderingManager
+from glaip_sdk.hitl import LocalPromptHandler, PauseResumeCallback
 from glaip_sdk.runner.base import BaseRunner
 from glaip_sdk.runner.deps import (
     check_local_runtime_available,
     get_local_runtime_missing_message,
 )
-from glaip_sdk.utils.a2a import A2AEventStreamProcessor
-from gllm_core.utils import LoggerManager
+from glaip_sdk.utils.tool_storage_provider import build_tool_output_manager
 
 if TYPE_CHECKING:
     from langchain_core.messages import BaseMessage
 
     from glaip_sdk.agents.base import Agent
 
-logger = LoggerManager().get_logger(__name__)
 
-# Default A2A event processor
-_event_processor = A2AEventStreamProcessor()
+_AIP_LOGS_SWALLOWED = False
+
+
+def _swallow_aip_logs(level: int = logging.ERROR) -> None:
+    """Consume noisy AIPAgents logs once (opt-in via runner flag)."""
+    global _AIP_LOGS_SWALLOWED
+    if _AIP_LOGS_SWALLOWED:
+        return
+    prefixes = ("aip_agents.",)
+
+    def _silence(name: str) -> None:
+        lg = logging.getLogger(name)
+        lg.handlers = [logging.NullHandler()]
+        lg.propagate = False
+        lg.setLevel(level)
+
+    # Silence any already-registered loggers under the given prefixes
+    for logger_name in logging.root.manager.loggerDict:
+        if any(logger_name.startswith(prefix) for prefix in prefixes):
+            _silence(logger_name)
+
+    # Also set the base prefix loggers so future children inherit silence
+    for prefix in prefixes:
+        _silence(prefix.rstrip("."))
+    _AIP_LOGS_SWALLOWED = True
+
+
+logger = LoggerManager().get_logger(__name__)
 
 
 def _convert_chat_history_to_messages(
@@ -81,9 +111,10 @@ def _convert_chat_history_to_messages(
 class LangGraphRunner(BaseRunner):
     """Runner implementation using aip-agents LangGraphReactAgent.
 
-    MVP scope:
-    - Execute via `LangGraphReactAgent.arun_a2a_stream()`
-    - Extract and return final text from the emitted `final_response` event
+    Current behavior:
+    - Execute via `LangGraphReactAgent.arun_sse_stream()` (normalized SSE-compatible stream)
+    - Route all events through `AgentRunRenderingManager.async_process_stream_events`
+      for unified rendering between local and remote agents
 
     Attributes:
         default_model: Model name to use when agent.model is not set.
@@ -99,6 +130,8 @@ class LangGraphRunner(BaseRunner):
         verbose: bool = False,
         runtime_config: dict[str, Any] | None = None,
         chat_history: list[dict[str, str]] | None = None,
+        *,
+        swallow_aip_logs: bool = True,
         **kwargs: Any,
     ) -> str:
         """Execute agent synchronously and return final response text.
@@ -113,6 +146,9 @@ class LangGraphRunner(BaseRunner):
             chat_history: Optional list of prior conversation messages.
                 Each message is a dict with "role" and "content" keys.
                 Defaults to None.
+            swallow_aip_logs: When True (default), silence noisy logs from aip-agents,
+                gllm_inference, OpenAILMInvoker, and httpx. Set to False to honor user
+                logging configuration.
             **kwargs: Additional keyword arguments passed to the backend.
 
         Returns:
@@ -141,6 +177,7 @@ class LangGraphRunner(BaseRunner):
             verbose=verbose,
             runtime_config=runtime_config,
             chat_history=chat_history,
+            swallow_aip_logs=swallow_aip_logs,
             **kwargs,
         )
 
@@ -153,6 +190,8 @@ class LangGraphRunner(BaseRunner):
         verbose: bool = False,
         runtime_config: dict[str, Any] | None = None,
         chat_history: list[dict[str, str]] | None = None,
+        *,
+        swallow_aip_logs: bool = True,
         **kwargs: Any,
     ) -> str:
         """Execute agent asynchronously and return final response text.
@@ -167,6 +206,7 @@ class LangGraphRunner(BaseRunner):
             chat_history: Optional list of prior conversation messages.
                 Each message is a dict with "role" and "content" keys.
                 Defaults to None.
+            swallow_aip_logs: When True (default), silence noisy AIPAgents logs.
             **kwargs: Additional keyword arguments passed to the backend.
 
         Returns:
@@ -181,6 +221,7 @@ class LangGraphRunner(BaseRunner):
             verbose=verbose,
             runtime_config=runtime_config,
             chat_history=chat_history,
+            swallow_aip_logs=swallow_aip_logs,
             **kwargs,
         )
 
@@ -191,6 +232,8 @@ class LangGraphRunner(BaseRunner):
         verbose: bool = False,
         runtime_config: dict[str, Any] | None = None,
         chat_history: list[dict[str, str]] | None = None,
+        *,
+        swallow_aip_logs: bool = True,
         **kwargs: Any,
     ) -> str:
         """Internal async implementation of agent execution.
@@ -201,13 +244,23 @@ class LangGraphRunner(BaseRunner):
             verbose: If True, emit debug trace output during execution.
             runtime_config: Optional runtime configuration for tools, MCPs, etc.
             chat_history: Optional list of prior conversation messages.
+            swallow_aip_logs: When True (default), silence noisy AIPAgents logs.
             **kwargs: Additional keyword arguments passed to the backend.
 
         Returns:
             The final response text from the agent.
         """
+        # Optionally swallow noisy AIPAgents logs
+        if swallow_aip_logs:
+            _swallow_aip_logs()
+
+        # POC/MVP: Create pause/resume callback for interactive HITL input
+        pause_resume_callback = PauseResumeCallback()
+
         # Build the local LangGraphReactAgent from the glaip_sdk Agent
-        local_agent = self.build_langgraph_agent(agent, runtime_config=runtime_config)
+        local_agent = self.build_langgraph_agent(
+            agent, runtime_config=runtime_config, pause_resume_callback=pause_resume_callback
+        )
 
         # Convert chat history to LangChain messages for the agent
         langchain_messages = _convert_chat_history_to_messages(chat_history)
@@ -219,20 +272,54 @@ class LangGraphRunner(BaseRunner):
                 agent.name,
             )
 
-        # Collect A2AEvents from the stream and extract final response
-        events: list[dict[str, Any]] = []
+        # Use shared render manager for unified processing
+        render_manager = AgentRunRenderingManager(logger)
+        renderer = render_manager.create_renderer(kwargs.get("renderer"), verbose=verbose)
 
-        async for event in local_agent.arun_a2a_stream(message, **kwargs):
-            if verbose:
-                self._log_event(event)
-            events.append(event)
+        # POC/MVP: Set renderer on callback so LocalPromptHandler can pause/resume Live
+        pause_resume_callback.set_renderer(renderer)
 
-        return _event_processor.extract_final_response(events)
+        meta = render_manager.build_initial_metadata(agent.name, message, kwargs)
+        render_manager.start_renderer(renderer, meta)
+
+        try:
+            # Use shared async stream processor for unified event handling
+            (
+                final_text,
+                stats_usage,
+                started_monotonic,
+                finished_monotonic,
+            ) = await render_manager.async_process_stream_events(
+                local_agent.arun_sse_stream(message, **kwargs),
+                renderer,
+                meta,
+                skip_final_render=True,
+            )
+        except KeyboardInterrupt:
+            try:
+                renderer.close()
+            finally:
+                raise
+        except Exception:
+            try:
+                renderer.close()
+            finally:
+                raise
+
+        # Use shared finalizer to avoid code duplication
+        from glaip_sdk.client.run_rendering import finalize_render_manager  # noqa: PLC0415
+
+        return finalize_render_manager(
+            render_manager, renderer, final_text, stats_usage, started_monotonic, finished_monotonic
+        )
 
     def build_langgraph_agent(
         self,
         agent: Agent,
         runtime_config: dict[str, Any] | None = None,
+        shared_tool_output_manager: Any | None = None,
+        *,
+        pause_resume_callback: Any | None = None,
     ) -> Any:
         """Build a LangGraphReactAgent from a glaip_sdk Agent definition.
 
@@ -240,6 +327,10 @@ class LangGraphRunner(BaseRunner):
             agent: The glaip_sdk Agent to convert.
             runtime_config: Optional runtime configuration with tool_configs,
                 mcp_configs, agent_config, and agent-specific overrides.
+            shared_tool_output_manager: Optional ToolOutputManager to reuse across
+                agents with tool_output_sharing enabled.
+            pause_resume_callback: Optional callback used to pause/resume the renderer
+                during interactive HITL prompts.
 
         Returns:
             A configured LangGraphReactAgent instance.
@@ -249,6 +340,7 @@ class LangGraphRunner(BaseRunner):
             ValueError: If agent has unsupported tools, MCPs, or sub-agents for local mode.
         """
         from aip_agents.agent import LangGraphReactAgent  # noqa: PLC0415
+
         from glaip_sdk.runner.tool_adapter import LangChainToolAdapter  # noqa: PLC0415
 
         # Adapt tools for local execution
@@ -260,9 +352,6 @@ class LangGraphRunner(BaseRunner):
             adapter = LangChainToolAdapter()
             langchain_tools = adapter.adapt_tools(agent.tools)
 
-        # Build sub-agents recursively
-        sub_agent_instances = self._build_sub_agents(agent.agents, runtime_config)
-
         # Normalize runtime config: merge global and agent-specific configs
         normalized_config = self._normalize_runtime_config(runtime_config, agent)
 
@@ -276,6 +365,19 @@ class LangGraphRunner(BaseRunner):
         merged_agent_config = self._merge_agent_config(agent, normalized_config)
         agent_config_params, agent_config_kwargs = self._apply_agent_config(merged_agent_config)
 
+        tool_output_manager = self._resolve_tool_output_manager(
+            agent,
+            merged_agent_config,
+            shared_tool_output_manager,
+        )
+
+        # Build sub-agents recursively, sharing tool output manager when enabled.
+        sub_agent_instances = self._build_sub_agents(
+            agent.agents,
+            runtime_config,
+            shared_tool_output_manager=tool_output_manager,
+        )
+
         # Build the LangGraphReactAgent with tools, sub-agents, and configs
         local_agent = LangGraphReactAgent(
             name=agent.name,
@@ -285,6 +387,7 @@ class LangGraphRunner(BaseRunner):
             tools=langchain_tools,
             agents=sub_agent_instances if sub_agent_instances else None,
             tool_configs=tool_configs if tool_configs else None,
+            tool_output_manager=tool_output_manager,
             **agent_config_params,
             **agent_config_kwargs,
         )
@@ -292,6 +395,11 @@ class LangGraphRunner(BaseRunner):
         # Add MCP servers if configured
         self._add_mcp_servers(local_agent, agent, mcp_configs)
 
+        # Inject local HITL manager only if hitl_enabled is True (master switch).
+        # This matches remote behavior: hitl_enabled gates the HITL plumbing.
+        # Tool-level HITL configs are only enforced when hitl_enabled=True.
+        self._inject_hitl_manager(local_agent, merged_agent_config, agent.name, pause_resume_callback)
+
         logger.debug(
             "Built local LangGraphReactAgent for agent '%s' with %d tools, %d sub-agents, and %d MCPs",
             agent.name,
@@ -301,16 +409,60 @@ class LangGraphRunner(BaseRunner):
         )
         return local_agent
 
+    def _resolve_tool_output_manager(
+        self,
+        agent: Agent,
+        merged_agent_config: dict[str, Any],
+        shared_tool_output_manager: Any | None,
+    ) -> Any | None:
+        """Resolve tool output manager for local agent execution."""
+        tool_output_sharing_enabled = merged_agent_config.get("tool_output_sharing", False)
+        if not tool_output_sharing_enabled:
+            return None
+        if shared_tool_output_manager is not None:
+            return shared_tool_output_manager
+        return build_tool_output_manager(agent.name, merged_agent_config)
+
+    def _inject_hitl_manager(
+        self,
+        local_agent: Any,
+        merged_agent_config: dict[str, Any],
+        agent_name: str,
+        pause_resume_callback: Any | None,
+    ) -> None:
+        """Inject HITL manager when enabled, mirroring remote gating behavior."""
+        hitl_enabled = merged_agent_config.get("hitl_enabled", False)
+        if hitl_enabled:
+            try:
+                local_agent.hitl_manager = ApprovalManager(
+                    prompt_handler=LocalPromptHandler(pause_resume_callback=pause_resume_callback)
+                )
+                # Store callback reference for setting renderer later
+                if pause_resume_callback:
+                    local_agent._pause_resume_callback = pause_resume_callback
+                logger.debug("HITL manager injected for agent '%s' (hitl_enabled=True)", agent_name)
+            except ImportError as e:
+                # Missing dependencies - fail fast
+                raise ImportError("Local HITL requires aip_agents. Install with: pip install 'glaip-sdk[local]'") from e
+            except Exception as e:
+                # Other errors during HITL setup - fail fast
+                raise RuntimeError(f"Failed to initialize HITL manager for agent '{agent_name}'") from e
+        else:
+            logger.debug("HITL manager not injected for agent '%s' (hitl_enabled=False)", agent_name)
+
     def _build_sub_agents(
         self,
         sub_agents: list[Any] | None,
         runtime_config: dict[str, Any] | None,
+        shared_tool_output_manager: Any | None = None,
     ) -> list[Any]:
         """Build sub-agent instances recursively.
 
         Args:
             sub_agents: List of sub-agent definitions.
             runtime_config: Runtime config to pass to sub-agents.
+            shared_tool_output_manager: Optional ToolOutputManager to reuse across
+                agents with tool_output_sharing enabled.
 
         Returns:
             List of built sub-agent instances.
@@ -324,7 +476,13 @@ class LangGraphRunner(BaseRunner):
         sub_agent_instances = []
         for sub_agent in sub_agents:
             self._validate_sub_agent_for_local_mode(sub_agent)
-            sub_agent_instances.append(self.build_langgraph_agent(sub_agent, runtime_config))
+            sub_agent_instances.append(
+                self.build_langgraph_agent(
+                    sub_agent,
+                    runtime_config,
+                    shared_tool_output_manager=shared_tool_output_manager,
+                )
+            )
         return sub_agent_instances
 
     def _add_mcp_servers(
@@ -571,7 +729,13 @@ class LangGraphRunner(BaseRunner):
         if "planning" in agent_config:
             direct_params["planning"] = agent_config["planning"]
 
+        if "enable_a2a_token_streaming" in agent_config:
+            direct_params["enable_a2a_token_streaming"] = agent_config["enable_a2a_token_streaming"]
+
         # Kwargs parameters (passed through **kwargs to BaseAgent)
+        if "enable_pii" in agent_config:
+            kwargs_params["enable_pii"] = agent_config["enable_pii"]
+
         if "memory" in agent_config:
             # Map "memory" to "memory_backend" for aip-agents compatibility
             kwargs_params["memory_backend"] = agent_config["memory"]

glaip_sdk/utils/rendering/renderer/base.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 
 import json
 import logging
+import sys
 from datetime import datetime, timezone
 from time import monotonic
 from typing import Any
@@ -349,6 +350,9 @@ class RichStreamRenderer(TranscriptModeMixin):
             self._handle_status_event(ev)
         elif kind == "content":
             self._handle_content_event(content)
+        elif kind == "token":
+            # Token events should stream content incrementally with immediate console output
+            self._handle_token_event(content)
         elif kind == "final_response":
             self._handle_final_response_event(content, metadata)
         elif kind in {"agent_step", "agent_thinking_step"}:
@@ -368,6 +372,31 @@ class RichStreamRenderer(TranscriptModeMixin):
             self.state.append_transcript_text(content)
             self._ensure_live()
 
+    def _handle_token_event(self, content: str) -> None:
+        """Handle token streaming events - print immediately for real-time streaming."""
+        if content:
+            self.state.append_transcript_text(content)
+            # Print token content directly to stdout for immediate visibility when not verbose
+            # This bypasses Rich's Live display which has refresh rate limitations
+            if not self.verbose:
+                try:
+                    # Mark that we're streaming tokens directly to prevent Live display from starting
+                    self._streaming_tokens_directly = True
+                    # Stop Live display if active to prevent it from intercepting stdout
+                    # and causing each token to appear on a new line
+                    if self.live is not None:
+                        self._stop_live_display()
+                    # Write directly to stdout - tokens will stream on the same line
+                    # since we're bypassing Rich's console which adds newlines
+                    sys.stdout.write(content)
+                    sys.stdout.flush()
+                except Exception:
+                    # Fallback to live display if direct write fails
+                    self._ensure_live()
+            else:
+                # In verbose mode, use normal live display (debug panels handle the output)
+                self._ensure_live()
+
     def _handle_final_response_event(self, content: str, metadata: dict[str, Any]) -> None:
         """Handle final response events."""
         if content:
@@ -521,6 +550,18 @@ class RichStreamRenderer(TranscriptModeMixin):
         if getattr(self, "_transcript_mode_enabled", False):
             return
 
+        # When verbose=False and tokens were streamed directly, skip final panel
+        # The user's script will print the final result, avoiding duplication
+        if not self.verbose and getattr(self, "_streaming_tokens_directly", False):
+            # Add a newline after streaming tokens for clean separation
+            try:
+                sys.stdout.write("\n")
+                sys.stdout.flush()
+            except Exception:
+                pass
+            self.state.printed_final_output = True
+            return
+
         if self.verbose:
             panel = build_final_panel(
                 self.state,
@@ -597,6 +638,19 @@ class RichStreamRenderer(TranscriptModeMixin):
 
     def _finalize_display(self) -> None:
         """Finalize live display and render final output."""
+        # When verbose=False and tokens were streamed directly, skip live display updates
+        # to avoid showing duplicate final result
+        if not self.verbose and getattr(self, "_streaming_tokens_directly", False):
+            # Just add a newline after streaming tokens for clean separation
+            try:
+                sys.stdout.write("\n")
+                sys.stdout.flush()
+            except Exception:
+                pass
+            self._stop_live_display()
+            self.state.printed_final_output = True
+            return
+
         # Final refresh
         self._ensure_live()
 
@@ -629,6 +683,10 @@ class RichStreamRenderer(TranscriptModeMixin):
         """Ensure live display is updated."""
         if getattr(self, "_transcript_mode_enabled", False):
             return
+        # When verbose=False, don't start Live display if we're streaming tokens directly
+        # This prevents Live from intercepting stdout and causing tokens to appear on separate lines
+        if not self.verbose and getattr(self, "_streaming_tokens_directly", False):
+            return
         if not self._ensure_live_stack():
             return