fix: Workaround against docker exec_run not flushing stdout

docker_agent_sandbox/sandbox.py

@@ -3,8 +3,8 @@
 from __future__ import annotations
 
 import io
-import select as _select
 import tarfile
+import uuid
 from pathlib import Path
 from typing import TYPE_CHECKING
 import docker
@@ -12,9 +12,6 @@ import docker.errors
 from loguru import logger
 
 
-_original_poll = getattr(_select, "poll", None)
-
-
 if TYPE_CHECKING:
     import docker.models.containers
 
@@ -267,30 +264,41 @@ class DockerSandbox:
         if self._container is None:
             return 1, "Sandbox container is not running."
 
-        wrapped = [
-            "timeout",
-            "--kill-after",
-            f"{kill_grace}s",
-            f"{timeout}s",
-            "bash",
-            "-c",
-            command,
-        ]
+        # docker-py's exec_run reads command output off a hijacked raw socket
+        # and intermittently returns nothing because bytes that http.client
+        # already buffered while parsing the response headers are stranded and
+        # never read (https://github.com/docker/docker-py/issues/2042,
+        # https://github.com/docker/docker-py/issues/3332). To avoid that path
+        # entirely we redirect combined output to a file and read it back via
+        # get_archive (a normal HTTP body), which is reliable on every
+        # transport. The exit code from exec_run is unaffected by the bug.
+        out_path = f"/tmp/.sandbox-exec-{uuid.uuid4().hex}"
+        # The command is passed via an env var so the outer shell never has to
+        # quote-escape it; "$SANDBOX_CMD" reaches the inner bash -c verbatim.
+        runner = (
+            f"timeout --kill-after={kill_grace}s {timeout}s "
+            f'bash -c "$SANDBOX_CMD" > {out_path} 2>&1'
+        )
         kwargs = {}
         if self._working_dir is not None:
             kwargs["workdir"] = self._working_dir
 
         try:
-            exit_code, output = self._container.exec_run(
-                wrapped,
-                stdout=True,
-                stderr=True,
-                demux=False,
+            exit_code, _ = self._container.exec_run(
+                ["bash", "-c", runner],
+                environment={"SANDBOX_CMD": command},
                 **kwargs,
             )
-            text = (output or b"").decode("utf-8", errors="replace")
-            if exit_code in (124, 137):
-                return exit_code, text + f"\n[command timed out after {timeout}s]"
-            return exit_code, text
         except Exception as exc:
             return 1, f"Error running command in container: {exc}"
+
+        try:
+            text = self.read_file(out_path).decode("utf-8", errors="replace")
+        except FileNotFoundError:
+            text = ""
+        finally:
+            self._container.exec_run(["rm", "-f", out_path])
+
+        if exit_code in (124, 137):
+            return exit_code, text + f"\n[command timed out after {timeout}s]"
+        return exit_code, text