Skip to content

Personal AI, On Personal Devices

Python

fix: suppress Qwen3.5 thinking mode, retry on empty content, add logging

Jon Saad-Falcon885b2caf996c
diff --git a/src/openjarvis/agents/executor.py b/src/openjarvis/agents/executor.py
index f7f1b086..565f22f1 100644
--- a/src/openjarvis/agents/executor.py
+++ b/src/openjarvis/agents/executor.py
@@ -231,10 +231,15 @@ class AgentExecutor:
pass # Fall back to configured model
# Construct agent instance (BaseAgent requires engine, model as positional args)
+ # Append /no_think to suppress Qwen3.5 extended thinking that
+ # can consume all tokens and produce empty visible output.
+ sys_prompt = config.get("system_prompt", "") or ""
+ if sys_prompt and "/no_think" not in sys_prompt:
+ sys_prompt = sys_prompt.rstrip() + "\n/no_think"
agent_instance = agent_cls(
engine,
model,
- system_prompt=config.get("system_prompt"),
+ system_prompt=sys_prompt or None,
tools=[],
)
@@ -320,6 +325,16 @@ class AgentExecutor:
)
_t0 = time.time()
result = agent_instance.run(input_text, context=agent_ctx)
+
+ # Retry once if the model returned empty content (common with
+ # Qwen3.5 thinking mode consuming all tokens).
+ if not (result.content or "").strip():
+ logger.warning(
+ "Agent %s: empty content, retrying once",
+ agent["name"],
+ )
+ result = agent_instance.run(input_text, context=agent_ctx)
+
_elapsed = time.time() - _t0
logger.info(
"Agent %s: agent.run() completed in %.1fs, "
diff --git a/src/openjarvis/server/agent_manager_routes.py b/src/openjarvis/server/agent_manager_routes.py
index 5b7b311e..919b997a 100644
--- a/src/openjarvis/server/agent_manager_routes.py
+++ b/src/openjarvis/server/agent_manager_routes.py
@@ -82,8 +82,8 @@ def _make_lightweight_system(
The server's ``app.state.engine`` is heavily wrapped
(MultiEngine → InstrumentedEngine → GuardrailsEngine) and can
- hang when reused from a background thread. Instead, create a
- fresh OllamaEngine pointed at the same backend.
+ return empty content when reused from a background thread.
+ Instead, create a fresh OllamaEngine pointed at the same backend.
"""
try:
from openjarvis.core.config import load_config
@@ -92,11 +92,23 @@ def _make_lightweight_system(
cfg = config if config is not None else load_config()
resolved = get_engine(cfg)
if resolved is not None:
- _, plain_engine = resolved
+ engine_key, plain_engine = resolved
+ logger.info(
+ "Lightweight system: fresh %s engine (key=%s), "
+ "model=%s",
+ type(plain_engine).__name__, engine_key, model,
+ )
return _LightweightSystem(plain_engine, model, cfg)
- except Exception:
- pass
+ except Exception as exc:
+ logger.warning(
+ "Failed to create fresh engine, falling back to "
+ "server engine: %s", exc,
+ )
# Fallback to the (wrapped) server engine
+ logger.warning(
+ "Using wrapped server engine %s — may return empty content",
+ type(engine).__name__,
+ )
return _LightweightSystem(engine, model, config)