fix: approval-aware stall threshold (#99) (#101)

nathanschram · claude · web-flow · commit 415c5f19f3ce · 2026-03-08T11:14:53.000+11:00
* fix: suppress stall warnings while waiting for user approval (#99) Use 30-minute threshold instead of 5-minute when the most recent action is a pending approval (has inline_keyboard buttons). Prevents false stall warnings when user is away from Telegram while a permission request waits. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: update changelog and test counts for approval-aware stall threshold Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -9,6 +9,7 @@
   - early PID threading: `last_pid` set at subprocess spawn, polled by `run_runner_with_cancel` before `StartedEvent`
   - standalone `/cancel` fallback: cancels single active run without requiring reply; prompts when multiple runs active
   - `queued_for_chat()` method on `ThreadScheduler` for standalone cancel of queued jobs
+  - approval-aware stall threshold: 30 min when waiting for user approval (inline keyboard detected), 5 min otherwise
 
 ## v0.34.1 (2026-03-07)
 
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -138,11 +138,11 @@ Rules in `.claude/rules/` auto-load when editing matching files:
 
 ## Tests
 
-1470 tests, 80% coverage threshold. Key test files:
+1472 tests, 80% coverage threshold. Key test files:
 
 - `test_claude_control.py` — 56 tests: control requests, response routing, registry lifecycle, auto-approve/auto-deny, tool auto-approve, custom deny messages, discuss action, early toast, progressive cooldown, auto permission mode
 - `test_callback_dispatch.py` — 28 tests: callback parsing, dispatch toast/ephemeral behaviour, early answering
-- `test_exec_bridge.py` — 77 tests: ephemeral notification cleanup, approval push notifications, progressive stall warnings, stall diagnostics, stall auto-cancel, session summary, PID/stream threading
+- `test_exec_bridge.py` — 80 tests: ephemeral notification cleanup, approval push notifications, progressive stall warnings, stall diagnostics, stall auto-cancel, approval-aware stall threshold, session summary, PID/stream threading
 - `test_ask_user_question.py` — 27 tests: AskUserQuestion control request handling, question extraction, pending request registry, answer routing, option button rendering, multi-question flows, structured answer responses, ask mode toggle auto-deny
 - `test_diff_preview.py` — 10 tests: Edit diff display, Write content preview, Bash command display, line/char truncation
 - `test_cost_tracker.py` — 56 tests: cost accumulation, per-run/daily budget thresholds, warning levels, daily reset, auto-cancel flag
diff --git a/src/untether/runner_bridge.py b/src/untether/runner_bridge.py
@@ -574,7 +574,13 @@ async def _stall_monitor(self) -> None:
             elapsed = self.clock() - self._last_event_at
             self._peak_idle = max(self._peak_idle, elapsed)
 
-            if elapsed < self._STALL_THRESHOLD_SECONDS:
+            # Use longer threshold when waiting for user approval
+            threshold = (
+                self._STALL_THRESHOLD_APPROVAL
+                if self._has_pending_approval()
+                else self._STALL_THRESHOLD_SECONDS
+            )
+            if elapsed < threshold:
                 continue
             now = self.clock()
             if (
@@ -687,6 +693,14 @@ async def _stall_monitor(self) -> None:
                     exc_info=True,
                 )
 
+    def _has_pending_approval(self) -> bool:
+        """Check if the most recent non-completed action is waiting for user approval."""
+        for action_state in reversed(list(self.tracker._actions.values())):
+            if not action_state.completed:
+                return bool(action_state.action.detail.get("inline_keyboard"))
+            break  # only check the most recent
+        return False
+
     def _last_action_summary(self) -> str | None:
         """Return a short description of the most recent action."""
         for action_state in reversed(list(self.tracker._actions.values())):
@@ -799,6 +813,7 @@ async def _delete_notify(ref: MessageRef) -> None:
             self.rendered_seq = seq_at_render
 
     _STALL_THRESHOLD_SECONDS: float = 300.0  # 5 minutes
+    _STALL_THRESHOLD_APPROVAL: float = 1800.0  # 30 minutes when waiting for approval
     _STALL_MAX_WARNINGS: int = 10  # absolute cap
     _STALL_MAX_WARNINGS_NO_PID: int = 3  # aggressive cap when pid=None + no events
 
diff --git a/tests/test_exec_bridge.py b/tests/test_exec_bridge.py
@@ -2106,3 +2106,126 @@ async def drive() -> None:
         c for c in transport.send_calls if "Auto-cancelled" in c["message"].text
     ]
     assert len(auto_cancel_msgs) == 1
+
+
+@pytest.mark.anyio
+async def test_stall_suppressed_while_waiting_for_approval() -> None:
+    """Stall monitor uses longer threshold when pending approval action exists."""
+    transport = FakeTransport()
+    presenter = _KeyboardPresenter()
+    clock = _FakeClock(start=100.0)
+    edits = _make_edits(transport, presenter, clock=clock)
+    edits._stall_check_interval = 0.01
+    edits._STALL_THRESHOLD_SECONDS = 0.05
+    edits._STALL_THRESHOLD_APPROVAL = 0.5  # 500ms for test
+
+    # Simulate a pending approval action (has inline_keyboard in detail)
+    from untether.model import Action, ActionEvent
+
+    evt = ActionEvent(
+        engine="codex",
+        action=Action(
+            id="ctrl.1",
+            kind="warning",
+            title="Permission Request [CanUseTool] - tool: ExitPlanMode",
+            detail={"inline_keyboard": {"buttons": [[{"text": "Approve"}]]}},
+        ),
+        phase="started",
+    )
+    await edits.on_event(evt)
+    clock.set(100.0)  # reset after on_event
+
+    async with anyio.create_task_group() as tg:
+
+        async def drive() -> None:
+            # Advance past normal threshold (0.05) but NOT past approval threshold (0.5)
+            clock.set(100.2)
+            await anyio.sleep(0.05)
+            # Should NOT have warned yet
+            edits.signal_send.close()
+
+        tg.start_soon(edits.run)
+        tg.start_soon(drive)
+
+    # No stall warning should have fired
+    assert edits._stall_warn_count == 0
+    stall_msgs = [c for c in transport.send_calls if "No progress" in c["message"].text]
+    assert len(stall_msgs) == 0
+
+
+@pytest.mark.anyio
+async def test_stall_fires_after_approval_threshold() -> None:
+    """Stall monitor fires after the longer approval threshold is exceeded."""
+    transport = FakeTransport()
+    presenter = _KeyboardPresenter()
+    clock = _FakeClock(start=100.0)
+    edits = _make_edits(transport, presenter, clock=clock)
+    edits._stall_check_interval = 0.01
+    edits._STALL_THRESHOLD_SECONDS = 0.05
+    edits._STALL_THRESHOLD_APPROVAL = 0.1  # short for test
+
+    from untether.model import Action, ActionEvent
+
+    evt = ActionEvent(
+        engine="codex",
+        action=Action(
+            id="ctrl.1",
+            kind="warning",
+            title="Permission Request [CanUseTool] - tool: Bash",
+            detail={"inline_keyboard": {"buttons": [[{"text": "Approve"}]]}},
+        ),
+        phase="started",
+    )
+    await edits.on_event(evt)
+    clock.set(100.0)
+
+    async with anyio.create_task_group() as tg:
+
+        async def drive() -> None:
+            # Advance past the approval threshold (0.1)
+            clock.set(100.2)
+            await anyio.sleep(0.05)
+            edits.signal_send.close()
+
+        tg.start_soon(edits.run)
+        tg.start_soon(drive)
+
+    assert edits._stall_warn_count >= 1
+    stall_msgs = [c for c in transport.send_calls if "No progress" in c["message"].text]
+    assert len(stall_msgs) >= 1
+
+
+@pytest.mark.anyio
+async def test_stall_normal_threshold_without_approval() -> None:
+    """Stall monitor uses normal threshold when no pending approval."""
+    transport = FakeTransport()
+    presenter = _KeyboardPresenter()
+    clock = _FakeClock(start=100.0)
+    edits = _make_edits(transport, presenter, clock=clock)
+    edits._stall_check_interval = 0.01
+    edits._STALL_THRESHOLD_SECONDS = 0.05
+    edits._STALL_THRESHOLD_APPROVAL = 10.0  # very long, shouldn't matter
+
+    # No pending approval — normal action without inline_keyboard
+    from untether.model import Action, ActionEvent
+
+    evt = ActionEvent(
+        engine="codex",
+        action=Action(id="a1", kind="tool", title="Bash"),
+        phase="started",
+    )
+    await edits.on_event(evt)
+    clock.set(100.0)
+
+    async with anyio.create_task_group() as tg:
+
+        async def drive() -> None:
+            clock.set(100.1)  # past normal threshold (0.05)
+            await anyio.sleep(0.05)
+            edits.signal_send.close()
+
+        tg.start_soon(edits.run)
+        tg.start_soon(drive)
+
+    # Should have warned using the normal threshold
+    assert edits._stall_warn_count >= 1
