From 51dc98d314000d7b326b341bb2b95f23ac6814d4 Mon Sep 17 00:00:00 2001
From: thchen <pama0227@gmail.com>
Date: Sat, 25 Apr 2026 13:02:27 +0000
Subject: [PATCH] fix(agent): detect Qwen3/Ollama inline thinking after tool
 calls

Ollama serves Qwen3 thinking inside the content field as <think>...</think>
blocks rather than in the API-level reasoning_content field.  This means
_has_structured was False for these responses, so an empty-looking reply
after a tool call triggered the nudge instead of the prefill continuation,
causing a double-response loop.

Fix: detect <think>/<thinking>/<reasoning> in final_response and:
  1. Skip the nudge when thinking is present (model is still reasoning)
  2. Include _has_inline_thinking in _has_structured so prefill kicks in
---
 run_agent.py | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/run_agent.py b/run_agent.py
index 482878019..6d604f9b5 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -13438,9 +13438,22 @@ class AIAgent:
                             m.get("role") == "tool"
                             for m in messages[-5:]  # check recent messages
                         )
+                        # Detect Qwen3/Ollama-style in-content thinking blocks.
+                        # Ollama puts <think> in the content field (not in
+                        # reasoning_content), so _has_structured below would
+                        # miss it.  We check here so thinking-only responses
+                        # after tool calls route to prefill instead of nudge.
+                        _has_inline_thinking = bool(
+                            re.search(
+                                r'<think>|<thinking>|<reasoning>',
+                                final_response or "",
+                                re.IGNORECASE,
+                            )
+                        )
                         if (
                             _prior_was_tool
                             and not getattr(self, "_post_tool_empty_retried", False)
+                            and not _has_inline_thinking  # thinking model still working — let prefill handle
                         ):
                             self._post_tool_empty_retried = True
                             # Clear stale narration so it doesn't resurface
@@ -13480,10 +13493,13 @@ class AIAgent:
                         # continue — the model will see its own reasoning
                         # on the next turn and produce the text portion.
                         # Inspired by clawdbot's "incomplete-text" recovery.
+                        # Also covers Qwen3/Ollama in-content <think> blocks
+                        # (detected above as _has_inline_thinking).
                         _has_structured = bool(
                             getattr(assistant_message, "reasoning", None)
                             or getattr(assistant_message, "reasoning_content", None)
                             or getattr(assistant_message, "reasoning_details", None)
+                            or _has_inline_thinking
                         )
                         if _has_structured and self._thinking_prefill_retries < 2:
                             self._thinking_prefill_retries += 1