Add Phase 6: write tools, shell, and edit tools with reasoning-only fix

Implement 6 new agent tools — write_file, make_dir, delete_file,
str_replace, patch_apply, run_command — bringing the agent from
read-only observer to active code modifier. All write/shell operations
are gated through the existing permissions service.

Also fix a bug where qwen3.5 thinking mode produces reasoning tokens
but no content after tool results, causing the agent to silently exit.
The loop now detects reasoning-only responses, retries twice, then
injects a nudge message to break the model out of its thinking loop.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

app/agent/context.py

@@ -41,6 +41,13 @@ class SessionContext:
         self._message_count += 1
         return message
 
+    def pop_last_message(self) -> Message | None:
+        """Remove and return the last message, or None if history is empty."""
+        if self._history:
+            self._message_count -= 1
+            return self._history.pop()
+        return None
+
     def get_history(self) -> list[Message]:
         """Return a shallow copy of the conversation history."""
         return list(self._history)

app/agent/loop.py

@@ -23,6 +23,8 @@ from app.utils.logging import get_logger
 
 logger = get_logger(__name__)
 
+_MAX_REASONING_RETRIES = 2
+
 
 class AgentLoop:
     """ReAct-style agent loop that streams LLM responses and executes tool calls.
@@ -81,6 +83,7 @@ class AgentLoop:
         self._ctx.add_message("user", user_input)
 
         max_iter = self._config.agent.max_iterations
+        reasoning_only_streak = 0
         for iteration in range(1, max_iter + 1):
             # Check token budget
             if self._ctx.token_counter.is_over_budget():
@@ -102,6 +105,9 @@ class AgentLoop:
                 tool_calls=assistant_msg.tool_calls,
             )
 
+            # Detect reasoning-only response (model thought but produced nothing)
+            reasoning_only = self._handler.had_reasoning_only
+
             # Record token usage
             if self._handler.usage:
                 self._ctx.token_counter.count_usage(self._handler.usage)
@@ -114,6 +120,29 @@ class AgentLoop:
 
             self._handler.reset()
 
+            # Reasoning-only: model produced thinking tokens but no content or tool calls.
+            if reasoning_only:
+                reasoning_only_streak += 1
+                self._ctx.pop_last_message()
+
+                if reasoning_only_streak >= _MAX_REASONING_RETRIES:
+                    # Nudge the model by injecting a user hint
+                    print_warning(
+                        f"Model produced reasoning but no response {reasoning_only_streak} times. "
+                        "Nudging model to respond..."
+                    )
+                    self._ctx.add_message(
+                        "user",
+                        "Please respond with your answer. Do not just think — provide your actual response.",
+                    )
+                    reasoning_only_streak = 0
+                else:
+                    print_warning("Model produced reasoning but no response. Retrying...")
+                continue
+
+            # Successful response — reset streak
+            reasoning_only_streak = 0
+
             # No tool calls → task complete (plain text response)
             if not assistant_msg.tool_calls:
                 break