Unhandled Non-Atomic File Operations & Permission Crashes in A2A Context State Persistence (a2a_context.py)

[Adityakk9031]

Location:

a2a_context.py:37-43

Code snippet:

def _atomic_write(path: Path, value: Any) -> None:
    tmp = path.with_suffix(f"{path.suffix}.tmp")
    tmp.write_text(json.dumps(value, sort_keys=True) + "\n")
    tmp.chmod(0o600)
    os.replace(tmp, path)
    path.chmod(0o600)

Root Cause & Mechanism:

On Windows systems (or POSIX filesystems with strict ACL permissions), invoking path.chmod(0o600) or tmp.chmod(0o600) raises an OSError / PermissionError or is ignored with unexpected ACL side effects.

More critically, if _atomic_write fails mid-execution (e.g. after writing tmp but before os.replace), tmp remains left behind. On subsequent calls, os.replace(tmp, path) on Windows will throw PermissionError: [WinError 5] Access is denied or FileExistsError if path is locked or opened by another thread or process reading context asynchronously via read_a2a_turn_context().

Impact:

• Inbound Agent-to-Agent (A2A) turns throw unhandled file system exceptions.
• The thread context queue files lock up, causing loss of multi-agent turn state and crashing the gateway runner.

[Adityakk9031]

opened the pr #85