personal memory agent
Fix generator timeouts and skip screen on audio-only segments
Screen generator now requires screen data (skips audio-only segments
instead of running a mismatched prompt that hangs). Generators get an
LLM request timeout derived from their token budget (120-480s) so
hanging API calls fail fast and trigger provider fallback instead of
blocking for the full 600s cortex kill. Cortex timeout events are now
broadcast to callosum for immediate detection, and dream's wait margin
is bumped to 610s to cover cortex cleanup.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
+23
-6
+1
-1
muse/screen.md
+1
-1
muse/screen.md
+8
think/agents.py
+8
think/agents.py
···
800
800
max_output_tokens = config.get("max_output_tokens") or 8192 * 6
801
801
is_json_output = output_format == "json"
802
802
803
+
# Derive LLM request timeout from token budget: scale with output size,
804
+
# floor at 120s, cap at 480s (well under cortex's 600s subprocess kill).
805
+
timeout_s = config.get("timeout_s") or min(
806
+
480, max(120, (max_output_tokens + thinking_budget) // 100)
807
+
)
808
+
803
809
# Build contents: transcript + instruction + prompt
804
810
contents = []
805
811
if transcript:
···
823
829
thinking_budget=thinking_budget,
824
830
system_instruction=system_instruction,
825
831
json_output=is_json_output,
832
+
timeout_s=timeout_s,
826
833
)
827
834
except Exception as exc:
828
835
if not _is_retryable_error(exc) or config.get("fallback_from"):
···
867
874
thinking_budget=thinking_budget,
868
875
system_instruction=system_instruction,
869
876
json_output=is_json_output,
877
+
timeout_s=timeout_s,
870
878
provider=backup,
871
879
model=backup_model,
872
880
)
+9
think/cortex.py
+9
think/cortex.py
···
336
336
f.write(json.dumps(error_event) + "\n")
337
337
except Exception as e:
338
338
self.logger.error(f"Failed to write timeout event: {e}")
339
+
340
+
# Broadcast to callosum so wait_for_agents detects immediately
341
+
try:
342
+
event_copy = error_event.copy()
343
+
event_type = event_copy.pop("event", "error")
344
+
self.callosum.emit("cortex", event_type, **event_copy)
345
+
except Exception:
346
+
pass
347
+
339
348
agent.stop()
340
349
341
350
def _monitor_stdout(self, agent: AgentProcess) -> None:
+5
-5
think/dream.py
+5
-5
think/dream.py
···
216
216
agent_ids = [agent_id for agent_id, _, _, _ in spawned]
217
217
logging.info(f"Waiting for {len(agent_ids)} agents...")
218
218
219
-
completed, timed_out = wait_for_agents(agent_ids, timeout=600)
219
+
completed, timed_out = wait_for_agents(agent_ids, timeout=610)
220
220
221
221
success = 0
222
222
failed = 0
···
692
692
agent_id=agent_id,
693
693
)
694
694
695
-
completed, timed_out = wait_for_agents([agent_id], timeout=600)
695
+
completed, timed_out = wait_for_agents([agent_id], timeout=610)
696
696
697
697
if timed_out:
698
698
logging.error(f"Generator {name} timed out (ID: {agent_id})")
···
837
837
838
838
logging.info(f"Waiting for {len(spawned_ids)} agent(s)...")
839
839
completed, timed_out = wait_for_agents(
840
-
[agent_id for agent_id, _ in spawned_ids], timeout=600
840
+
[agent_id for agent_id, _ in spawned_ids], timeout=610
841
841
)
842
842
843
843
if timed_out:
···
1021
1021
agent_ids = [aid for aid, _, _ in spawned]
1022
1022
logging.info(f"Waiting for {len(agent_ids)} agents...")
1023
1023
1024
-
completed, timed_out = wait_for_agents(agent_ids, timeout=600)
1024
+
completed, timed_out = wait_for_agents(agent_ids, timeout=610)
1025
1025
1026
1026
if timed_out:
1027
1027
logging.warning(f"{len(timed_out)} agents timed out")
···
1308
1308
if spawned:
1309
1309
_update_status(current_agents=[name for _, name, _ in spawned])
1310
1310
agent_ids = [aid for aid, _, _ in spawned]
1311
-
completed, timed_out = wait_for_agents(agent_ids, timeout=600)
1311
+
completed, timed_out = wait_for_agents(agent_ids, timeout=610)
1312
1312
1313
1313
if timed_out:
1314
1314
logging.warning(f"Flush: {len(timed_out)} agents timed out")