Strip <think>...</think> blocks from model output

llm_cmd.py

@@ -1,5 +1,6 @@
 import click
 import llm
+import re
 import subprocess
 from prompt_toolkit import PromptSession
 from prompt_toolkit.lexers import PygmentsLexer
@@ -14,6 +15,12 @@
 You return only: git reset --soft HEAD~1
 """.strip()
 
+THINKING_RE = re.compile(r"<think>.*?</think>", re.DOTALL)
+
+def strip_thinking(text):
+    # Thinking models may emit <think>...</think> blocks alongside the command
+    return THINKING_RE.sub("", text).strip()
+
 @llm.hookimpl
 def register_commands(cli):
     @cli.command()
@@ -30,7 +37,7 @@ def cmd(args, model, system, key):
         if model_obj.needs_key:
             model_obj.key = llm.get_key(key, model_obj.needs_key, model_obj.key_env_var)
         result = model_obj.prompt(prompt, system=system or SYSTEM_PROMPT)
-        interactive_exec(str(result))
+        interactive_exec(strip_thinking(str(result)))
 
 def interactive_exec(command):
     session = PromptSession(lexer=PygmentsLexer(BashLexer))

tests/test_cmd.py

@@ -1,6 +1,18 @@
-from llm.plugins import pm
+from llm.plugins import load_plugins, pm
+from llm_cmd import strip_thinking
 
 
 def test_plugin_is_installed():
+    load_plugins()
     names = [mod.__name__ for mod in pm.get_plugins()]
     assert "llm_cmd" in names
+
+
+def test_strip_thinking():
+    # Response shape from https://github.com/simonw/llm-cmd/issues/27
+    text = "<think>\nOkay, the user wants to show their bash history.\n</think>\n\n  history\n"
+    assert strip_thinking(text) == "history"
+
+
+def test_strip_thinking_no_block():
+    assert strip_thinking("ls -la") == "ls -la"