Avoid emitting thinking tokens

aider_gitea/__init__.py

@@ -158,16 +158,19 @@ AIDER_LINT = bash_cmd(
 )
 
 
-LLM_MESSAGE_FORMAT = (
-    """{issue}\nGo ahead with the changes you deem appropriate without waiting for explicit approval."""
-)
+LLM_MESSAGE_FORMAT = """{issue}
+
+Go ahead with the changes you deem appropriate without waiting for explicit approval.
+
+Do not draft changes beforehand; produce changes only once prompted for a specific file.
+"""
 
 CODE_MODEL = None
 EVALUATOR_MODEL = 'ollama/gemma3:27b'
 
 MODEL_EDIT_MODES = {
         'ollama/qwen3:32b': 'diff',
-        'ollama/hf.co/unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M': 'whole',
+        'ollama/hf.co/unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M': 'diff-fenced',
 }
 
 
@@ -186,6 +189,7 @@ def create_aider_command(issue: str) -> list[str]:
         '--auto-test',
         '--no-auto-lint',
         '--yes',
+        '--disable-playwright',
         '--timeout',
         str(10_000),
     ]
@@ -360,6 +364,9 @@ def issue_solution_round(repository_path, issue_content):
 
     return True
 
+def remove_thinking_tokens(text: str) -> str:
+    text = re.sub('^<think>.*?</think>', '', text, flags=re.MULTILINE)
+    return text.strip()
 
 def run_ollama(cwd: Path, texts: list[str]) -> str:
     cmd = ['ollama', 'run', EVALUATOR_MODEL.removeprefix('ollama/')]
@@ -373,6 +380,7 @@ def run_ollama(cwd: Path, texts: list[str]) -> str:
         text=True,
     )
     stdout, stderr = process.communicate('\n'.join(texts))
+    stdout = remove_thinking_tokens(stdout)
     print(stdout)
     return stdout