[eric] browser: per-action command timeouts (reads 12s, nav 20s) so a hung tab surfaces in seconds not 30s

backend/apps/agents/core/ws_manager.py

@@ -7,6 +7,18 @@ from backend.apps.agents.core.seq_log import TERMINAL_STATUSES, seq_log
 
 logger = logging.getLogger(__name__)
 
+# Per-action browser-command timeouts (seconds). A hung tab makes EVERY command
+# block to its timeout, so these bound how fast a freeze surfaces. Reads/clicks
+# operate on an already-loaded page and should be quick; navigation legitimately
+# loads the network so it gets a longer leash. Was a flat 30s, which let one
+# wedged page spin for ~20 minutes across retries.
+_BROWSER_CMD_TIMEOUT_DEFAULT = 12.0
+_BROWSER_CMD_TIMEOUTS = {
+    "navigate": 20.0,     # a real page load can be slow
+    "replay_route": 20.0, # an API fetch can be slow
+    "wait": 12.0,         # smart-wait already caps itself well under this
+}
+
 
 class ConnectionManager:
     """Manages WebSocket connections and HITL approval bridging; events flow through seq_log so reconnects can replay."""
@@ -242,7 +254,14 @@ class ConnectionManager:
         })
 
         try:
-            result = await asyncio.wait_for(future, timeout=30.0)
+            # Bound each command so a wedged tab can't block for 30s (the cost
+            # that turned one hung LinkedIn page into a 20-minute spin). Navigation
+            # legitimately takes longer than reads/clicks on an already-loaded page,
+            # so it gets a longer leash; everything else fails fast. A one-off slow
+            # command just times out and the next success resets the agent's streak,
+            # so only a SUSTAINED hang trips the fast-fail abort.
+            timeout = _BROWSER_CMD_TIMEOUTS.get(action, _BROWSER_CMD_TIMEOUT_DEFAULT)
+            result = await asyncio.wait_for(future, timeout=timeout)
             return result
         except asyncio.TimeoutError:
             return {"error": "Browser command timed out"}

backend/tests/test_browser_command_timeout.py

@@ -0,0 +1,78 @@
+"""Per-action browser-command timeouts.
+
+A hung tab makes every command block to its timeout; a flat 30s let one wedged
+page spin ~20 minutes across retries. These pin that the bound is now short and
+per-action, so a freeze surfaces in seconds.
+"""
+
+import asyncio
+import time
+
+import pytest
+
+from backend.apps.agents.core import ws_manager as wsm
+
+
+class _FakeSock:
+    async def send_text(self, _):
+        return None
+
+
+def _mgr():
+    m = wsm.ConnectionManager()
+    m.global_connections = [_FakeSock()]  # get past the 'no dashboard' guard
+    return m
+
+
+def test_timeout_map_reads_are_short_navigation_longer():
+    # reads/clicks act on a loaded page -> short; navigation loads network -> longer
+    assert wsm._BROWSER_CMD_TIMEOUT_DEFAULT <= 15
+    assert wsm._BROWSER_CMD_TIMEOUTS["navigate"] <= 25
+    assert wsm._BROWSER_CMD_TIMEOUTS["navigate"] > wsm._BROWSER_CMD_TIMEOUT_DEFAULT
+    # the old flat 30s is gone for the common path
+    assert wsm._BROWSER_CMD_TIMEOUT_DEFAULT < 30
+
+
+@pytest.mark.asyncio
+async def test_hung_command_returns_fast_at_the_bound(monkeypatch):
+    # shrink the bounds so the test is quick, then never resolve the future:
+    # the command must return a timeout error at ~the (default) bound, not hang.
+    monkeypatch.setattr(wsm, "_BROWSER_CMD_TIMEOUT_DEFAULT", 0.3)
+    monkeypatch.setattr(wsm, "_BROWSER_CMD_TIMEOUTS", {"navigate": 0.6})
+    m = _mgr()
+    t0 = time.monotonic()
+    res = await m.send_browser_command("rid1", "get_text", "b1", {})  # never resolved
+    elapsed = time.monotonic() - t0
+    assert res == {"error": "Browser command timed out"}
+    assert 0.25 < elapsed < 1.0, f"a read should time out near its 0.3s bound, took {elapsed:.2f}s"
+
+
+@pytest.mark.asyncio
+async def test_navigate_gets_the_longer_leash(monkeypatch):
+    monkeypatch.setattr(wsm, "_BROWSER_CMD_TIMEOUT_DEFAULT", 0.3)
+    monkeypatch.setattr(wsm, "_BROWSER_CMD_TIMEOUTS", {"navigate": 0.7})
+    m = _mgr()
+    t0 = time.monotonic()
+    await m.send_browser_command("rid2", "navigate", "b1", {"url": "x"})
+    elapsed = time.monotonic() - t0
+    assert elapsed > 0.5, "navigate should use its longer bound, not the default"
+
+
+@pytest.mark.asyncio
+async def test_a_resolved_command_returns_immediately(monkeypatch):
+    # a healthy command returns the moment the renderer resolves it, not at the bound
+    monkeypatch.setattr(wsm, "_BROWSER_CMD_TIMEOUT_DEFAULT", 5.0)
+    m = _mgr()
+
+    async def _resolve_soon():
+        await asyncio.sleep(0.05)
+        # find the pending future and resolve it like the renderer would
+        rid = next(iter(m.browser_futures))
+        m.resolve_browser_command(rid, {"text": "ok", "url": "u"})
+
+    asyncio.create_task(_resolve_soon())
+    t0 = time.monotonic()
+    res = await m.send_browser_command("rid3", "get_text", "b1", {})
+    elapsed = time.monotonic() - t0
+    assert res == {"text": "ok", "url": "u"}
+    assert elapsed < 1.0, "healthy command returns on resolve, not at the timeout"