get_app_state (and thus all tools) crash on Linux: 'Accessible' object has no attribute 'is_text'

[dengjzy]

Environment

• OS: Ubuntu 24.04 (Linux 7.0.0-28-generic), GNOME on X11 (not Wayland)
• Python: 3.12.3 (system), gi GObject Introspection with Atspi 2.0 (gi.repository.Atspi)
• open-computer-use: v0.3.1 (npm, dist/linux/amd64/open-computer-use — a Go binary embedding apps/OpenComputerUseLinux/runtime.py)
• Node: v18.19.1 (system)

Problem 1 (critical): every tool crashes with 'Accessible' object has no attribute 'is_text'

On this system, all tools that read the accessibility tree crash:

open-computer-use call get_app_state --args '{"app":"qqmusic"}'
>>> 'Accessible' object has no attribute 'is_text'

type_text crashes the same way with a sibling attribute:

open-computer-use call type_text --args '{"app":"qqmusic","text":"hi"}'
>>> 'Accessible' object has no attribute 'is_editable_text'

Because every action tool is gated behind app state ("No app state is available for X. Run get_app_state before action tools."), one AttributeError makes the entire server unusable on Linux.

Root cause

The embedded runtime.py treats is_text / is_editable_text as attributes of the Atspi.Accessible object:

• runtime.py:220 — if not bool(safe(node.is_text, False)):
• runtime.py:410 — if focused is None or not bool(safe(focused.is_text, False)):
• runtime.py:693 — return bool(safe(node.is_editable_text, False)) and ...
• runtime.py:720 — if node is not None and bool(safe(node.is_editable_text, False)):

In the gi bindings (gi.repository.Atspi), Atspi.Accessible has no is_text/is_editable_text attribute. Interface presence is queried via node.get_interfaces() (returns e.g. ['Accessible', 'Action', 'Collection', 'Component', 'Hyperlink', 'Text']). Because the attribute is evaluated eagerly at call time, the AttributeError fires before safe() can catch it — so safe(node.is_text, False) never protects anything.

Suggested fix

Replace the attribute accesses with interface checks:

def is_text_node(node):
    try:
        return "Text" in node.get_interfaces()
    except Exception:
        return False

def is_editable_text_node(node):
    try:
        return "EditableText" in node.get_interfaces()
    except Exception:
        return False

and use those in the four call sites above.

Workaround used

I worked around it locally (without touching the Go binary) by monkeypatching the missing properties onto Atspi.Accessible from a .pth-loaded Python module:

# ocu_fix.py (loaded via ocu_fix.pth in a site-packages dir)
import gi
gi.require_version("Atspi", "2.0")
from gi.repository import Atspi
if not hasattr(Atspi.Accessible, "is_text"):
    Atspi.Accessible.is_text = property(lambda self: "Text" in self.get_interfaces())
if not hasattr(Atspi.Accessible, "is_editable_text"):
    Atspi.Accessible.is_editable_text = property(lambda self: "EditableText" in self.get_interfaces())

Note: a sitecustomize.py in the user site-packages does not work — Ubuntu ships /usr/lib/python3.12/sitecustomize.py, and Python only loads the first sitecustomize on sys.path. The .pth import ocu_fix mechanism is what reliably runs.

Problem 2: type_text cannot type CJK on Chromium/Electron apps

type_text first tries insert_text(), which requires the target node to expose the EditableText interface. On Chromium/Electron apps the search <input> is exposed with Text but no EditableText, so insert_text returns False and it falls back to send_text() (Atspi.generate_keyboard_event(..., KeySynthType.STRING)), which synthesizes keysyms and cannot produce CJK characters. On Chinese-language apps, type_text therefore cannot enter text; the only reliable route is pasting from the X clipboard into a focused field.

This would be worth documenting (or handling) if CJK input is expected to work.

Problem 3: Chromium/Electron apps expose an empty web-content tree to AT-SPI

Against QQ Music (an Electron app), get_app_state returns only empty document web nodes until the app is launched with --force-renderer-accessibility. Without that flag, no widgets (search box, buttons, song rows) are exposed at all, so element-indexed click/type_text cannot target anything. Worth a note in the README for Electron/Chromium targets.

---

Happy to provide more logs or test a patch if useful.