fix: prevent JSON event leaking to users and reload skills after AI creation

- OpenCode runtime: stop calling _collect_text fallback on non-text events (step_start, step_finish, etc.)
- Both runtimes: guard raw stdout fallback to only apply for non-JSON output
- main.py: reload skills after AI responses containing skill-related keywords
- main.py: return friendly message instead of empty string for tool-only responses

main.py

@@ -299,6 +299,18 @@ def ai_handler(msg: IncomingMessage) -> str:
     # Parse and execute action tags (reminders, cron, spawn)
     reply_text = _process_action_tags(response.text, msg)
 
+    # If the AI may have created/modified skills (via file tools), reload them
+    # so that `skill list` reflects the changes immediately.
+    if _skills and any(
+        kw in text_lower
+        for kw in ("skill", "create a skill", "new skill", "add a skill", "make a skill")
+    ):
+        try:
+            _skills.reload()
+            logger.info("Skills reloaded after potential skill modification")
+        except Exception as e:
+            logger.debug(f"Skills reload after AI response failed: {e}")
+
     # Log conversation to memory session log
     if _memory:
         try:
@@ -311,6 +323,11 @@ def ai_handler(msg: IncomingMessage) -> str:
         except Exception as e:
             logger.debug(f"Session logging failed: {e}")
 
+    # Guard against empty responses (e.g. when AI did tool-only work
+    # and the parser correctly stripped all non-text events)
+    if not reply_text or not reply_text.strip():
+        return "✅ Done — I processed that but had no text to show."
+
     return reply_text