Merge branch 'stream_virt' into 'reset-v0.4.0'

Virtualization now supports stream

See merge request aliasrobotics/alias_research/cai!156

src/cai/repl/commands/virtualization.py

@@ -147,17 +147,30 @@ class DockerManager:
             Tuple[bool, str]: Success status and output message
         """
         try:
-            process = subprocess.run(
+            # Use Popen to stream pull progress line by line.
+            pull_proc = subprocess.Popen(
                 ["docker", "pull", image_name],
-                capture_output=True,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
                 text=True,
-                check=False
+                bufsize=1,
+            )
+
+            # Stream output in real time to the console.
+            if pull_proc.stdout is not None:
+                for line in pull_proc.stdout:
+                    console.print(line.rstrip())
+
+            pull_proc.wait()
+
+            if pull_proc.returncode == 0:
+                return True, (
+                    f"Successfully pulled image: {image_name}"
+                )
+            return False, (
+                f"Failed to pull image: {image_name} "
+                f"(code {pull_proc.returncode})"
             )
-            
-            if process.returncode == 0:
-                return True, f"Successfully pulled image: {image_name}"
-            else:
-                return False, f"Failed to pull image: {process.stderr}"
         except (subprocess.SubprocessError, FileNotFoundError) as e:
             return False, f"Error pulling image: {str(e)}"
 

src/cai/tools/common.py

@@ -11,6 +11,7 @@ import signal
 import time
 import uuid
 import sys
+import shlex
 from wasabi import color  # pylint: disable=import-error
 from cai.util  import format_time
 
@@ -774,18 +775,37 @@ def run_command(command, ctf=None, stdout=False,  # pylint: disable=too-many-arg
                 print(f"\033[32m(Started Session {new_session_id} in {context_msg})\n{output}\033[0m") # noqa E501
             return f"Started async session {new_session_id} in container {container_id[:12]}. Use this ID to interact." # noqa E501
 
-        # Handle Streaming Container Execution - not yet implemented for containers
+        # Handle Streaming Container Execution
         if stream:
-            # For now, display that streaming isn't supported for containers
-            from cai.util import cli_print_tool_output
-            if call_id and tool_name:
-                tool_args = {"command": command, "container": container_id[:12]}
-                cli_print_tool_output(
-                    tool_name, 
-                    tool_args, 
-                    "Streaming not yet supported for container execution. Running normally...",
-                    call_id=call_id
-                )
+            # Ensure workspace directory exists inside the container first
+            mkdir_cmd = [
+                "docker", "exec", container_id,
+                "mkdir", "-p", container_workspace
+            ]
+            subprocess.run(
+                mkdir_cmd,
+                capture_output=True,
+                text=True,
+                check=False,
+                timeout=10
+            )
+
+            # Build docker exec command as a single shell string for streaming
+            docker_exec_cmd = (
+                "docker exec -w "
+                f"{shlex.quote(container_workspace)} "
+                f"{shlex.quote(container_id)} sh -c "
+                f"{shlex.quote(command)}"
+            )
+
+            # Re-use the local streaming helper to provide real-time output
+            return _run_local_streamed(
+                docker_exec_cmd,
+                call_id,
+                timeout,
+                tool_name,
+                workspace_dir=_get_workspace_dir()
+            )
 
         # Handle Synchronous Execution in Container
         try:

tests/others/test_pretty_print.py

@@ -47,7 +47,7 @@ async def test_pretty_run_result_streaming():
 RunResultStreaming:
 - Current agent: Agent(name="test_agent", ...)
 - Current turn: 1
-- Max turns: 10
+- Max turns: inf
 - Is complete: True
 - Final output (str):
     Hi there
@@ -111,7 +111,7 @@ async def test_pretty_run_result_streaming_structured_output():
 RunResultStreaming:
 - Current agent: Agent(name="test_agent", ...)
 - Current turn: 1
-- Max turns: 10
+- Max turns: inf
 - Is complete: True
 - Final output (Foo):
     {
@@ -189,7 +189,7 @@ async def test_pretty_run_result_streaming_list_structured_output():
 RunResultStreaming:
 - Current agent: Agent(name="test_agent", ...)
 - Current turn: 1
-- Max turns: 10
+- Max turns: inf
 - Is complete: True
 - Final output (list):
     [Foo(bar='Test'), Foo(bar='Test 2')]