dominus-cli 0.5.6 → 0.5.7

package/dist/index.js

@@ -680,8 +680,20 @@ These are built-in lessons that ship with Dominus. Follow them exactly.
 - Full UI tree \u2192 \`create_ui\` (one call, never loop \`insert_instance\` for UI)
 - **Convert/export existing UI \u2192 \`serialize_ui\`** (NOT get_descendants_properties)
 - Adjusting an existing instance \u2192 \`set_properties\` (not \`run_code\`)
+- **Inspecting user's selection \u2192 \`get_selection\`, then \`get_properties\` or \`serialize_ui\`**
+- **Reading properties \u2192 \`get_properties\` (NEVER \`run_code\` with dump scripts)**
+- **Browsing the tree \u2192 \`get_explorer\`**
 - Anything exotic or procedural \u2192 \`run_code\` as last resort
 
+### NEVER use run_code for:
+- Getting properties of an instance \u2192 use \`get_properties\` or \`serialize_ui\`
+- Dumping selection info \u2192 use \`get_selection\` + \`get_properties\`
+- Creating instances \u2192 use \`create_part\`, \`create_ui\`, \`insert_instance\`, \`build_multiple\`
+- Modifying properties \u2192 use \`set_properties\` or \`bulk_set_properties\`
+- Reading scripts \u2192 use \`read_script\`
+- Browsing the explorer \u2192 use \`get_explorer\`
+\`run_code\` is ONLY for procedural logic, terrain generation, raycasts, physics queries, or custom algorithms with no tool equivalent.
+
 ### Verification habits
 - After a build, call \`get_explorer\` or \`get_selection\` to confirm the tree looks right.
 - After editing a script, read it back.
@@ -2624,7 +2636,7 @@ async function startMcpServer() {
   );
   mcp.tool(
     "run_code",
-    "Execute Luau code in Roblox Studio and return output",
+    "Execute arbitrary Luau code in Roblox Studio. LAST RESORT \u2014 only use when no dedicated tool exists. Do NOT use for: reading properties (use get_properties/get_selection/serialize_ui), creating instances (use create_part/create_ui/insert_instance/build_multiple), modifying properties (use set_properties/bulk_set_properties), or inspecting the tree (use get_explorer). Valid uses: complex procedural generation, terrain algorithms, raycasting, custom logic with no tool equivalent.",
     { code: z.string().describe("Luau code to execute") },
     async ({ code }) => {
       assertConnected();
@@ -2720,7 +2732,7 @@ async function startMcpServer() {
   );
   mcp.tool(
     "get_selection",
-    "Get the currently selected instances in Roblox Studio",
+    'Get the currently selected instances in Roblox Studio. Returns paths and class names of all selected objects. Use this FIRST when the user says "my selection" or "what I have selected". Then use get_properties or serialize_ui on the returned paths to inspect details.',
     {},
     async () => {
       assertConnected();
@@ -4259,10 +4271,22 @@ You are deeply knowledgeable about:
 
 You have direct access to Roblox Studio through tools. You can read/edit scripts, run code, browse the explorer tree, manage instances, and run tests -- all in real-time via WebSocket.
 
+## CRITICAL: NEVER use run_code for tasks that have a dedicated tool
+run_code is a LAST RESORT. You have 30+ specialized tools \u2014 use them.
+- Reading properties \u2192 get_properties, get_selection, get_descendants_properties
+- Inspecting selection \u2192 get_selection (returns paths), then get_properties or serialize_ui
+- Creating parts \u2192 create_part, build_multiple
+- Creating UI \u2192 create_ui
+- Modifying properties \u2192 set_properties, bulk_set_properties
+- Browsing the tree \u2192 get_explorer
+- Reading/editing scripts \u2192 read_script, edit_script
+- Converting UI \u2192 serialize_ui
+Only use run_code for: procedural generation, terrain algorithms, raycasting, physics queries, or other logic that no tool covers.
+
 When the user asks you to do something, you should:
 1. Use your tools to gather context (read scripts, check explorer, recall memory)
 2. Plan your approach for complex tasks
-3. Execute changes using your tools (edit scripts, run code, insert instances)
+3. Execute changes using the appropriate dedicated tools (NOT run_code unless truly necessary)
 4. Verify your changes worked (read back, check output for errors)
 5. Remember important facts for future sessions`);
   if (opts.connected) {