checkpoint: finished test curation

gemini_cli_adapter.py

@@ -24,6 +24,7 @@ class GeminiCliAdapter:
             command += f' --resume {self.session_id}'
 
         accumulated_text = ""
+        tool_calls = []
         
         env = os.environ.copy()
         env["GEMINI_CLI_HOOK_CONTEXT"] = "manual_slop"
@@ -59,14 +60,22 @@ class GeminiCliAdapter:
 
                     elif msg_type == "result":
                         # Capture final usage and session persistence
-                        self.last_usage = data.get("usage")
+                        # Support both mock ('usage') and real ('stats') keys
+                        self.last_usage = data.get("usage") or data.get("stats")
                         self.session_id = data.get("session_id")
 
-                    elif msg_type in ("status", "tool_use"):
+                    elif msg_type == "tool_use":
+                        # Collect tool_use messages
+                        tool_calls.append(data)
                         # Log status/tool_use to stderr for debugging
                         sys.stderr.write(f"GeminiCliAdapter [{msg_type}]: {line}\n")
                         sys.stderr.flush()
 
+                    elif msg_type == "status":
+                        # Log status to stderr for debugging
+                        sys.stderr.write(f"GeminiCliAdapter [{msg_type}]: {line}\n")
+                        sys.stderr.flush()
+
                 except json.JSONDecodeError:
                     # Skip lines that are not valid JSON
                     continue
@@ -78,4 +87,7 @@ class GeminiCliAdapter:
         finally:
             self.last_latency = time.time() - start_time
 
-        return accumulated_text
+        return {
+            "text": accumulated_text,
+            "tool_calls": tool_calls
+        }