Wrong output when interactive with python

[popjxc]

Env:

• os: mac
• python 3.14.3
• uv tree|grep swe
  Resolved 59 packages in 3ms
  ├── swe-rex v1.4.0

Code:

import asyncio
from swerex.deployment.local import LocalDeployment
from swerex.runtime.abstract import CreateBashSessionRequest, BashAction


async def interactive_python(deployment):
    await deployment.start()
    runtime = deployment.runtime
    await runtime.create_session(CreateBashSessionRequest())

    result = await runtime.run_in_session(
        BashAction(
            command="python",
            is_interactive_command=True,
            expect=[">"], timeout=5)
    )
    print(1, result.output)
    result = await runtime.run_in_session(
        BashAction(
            command="print('hello world')",
            is_interactive_command=True,
            expect=[">>> "], timeout=5)
    )
    print(2, result.output)

    result = await runtime.run_in_session(
        BashAction(
            command="print('hello world')",
            is_interactive_command=True,
            expect=[">>> "], timeout=5)
    )
    print(3, result.output)

    await runtime.run_in_session(
        BashAction(command="quit()\n", is_interactive_quit=True, timeout=5)
    )

    await deployment.stop()
    print(4, "end.")


deployment = LocalDeployment()
asyncio.run(interactive_python(deployment))

Output:

# python test_swe_rex.py
1 
2 Python 3.14.3 (main, Mar 25 2026, 03:28:18) [Clang 22.1.1 ] on darwin
Type "help", "copyright", "credits" or "license" for more information.

3 
4 end.
Ensuring deployment is stopped because object is deleted

Problem：
The output order is wrong, and 'hello world' is missing. Is this a mistake on my part, or is it a bug?

[Keesan12]

Interactive output has to be treated as a contract, not a side effect. Once a person is steering an agent through a Python session, wrong or partial output can lead to the next decision being confidently wrong.

I'm building MartinLoop because coding-agent work needs stronger operational truth. The system should tell you what ran, what changed, what verified, and where the evidence stopped being reliable. That sounds basic, but it is the difference between an agent that can demo and an agent you can hand real work to.

Would love to hear what output guarantee matters most to people using SWE-ReX interactively. MartinLoop is open source: https://github.com/Keesan12/martin-loop. We welcome feedback at support@martinloop.com.