[#23896] Create tools for inpection (ros2 cli)

[Danipiza]

Description

This branch adds a built-in ROS 2 default tool suite to VulcanAI and wires in the runtime, planner, validator, console, and CI changes needed to support it cleanly.

ROS 2 default tools

• Added a new ros2_default_tools entry point in pyproject.toml.
• Added a shared ROS2DefaultToolNode to support default tools when no main node is provided.

- ROS2 NODE
    - List (ros2 node list) output a List of available nodes.
    - Info (ros2 node info <node>) output Information about a node

- ROS2 TOPIC
    - List (ros2 topic list) output a List of available topics.
    - Info (ros2 topic info <topic>) print Information about a topic.
    - Type (ros2 topic type <topic>) print a topic's Type.
    - Find (ros2 topic find <type>) output a List of available topics of a given Type.
    - Bw (ros2 topic bw <topic>) display Bandwidth used by topic.
    - Delay (ros2 topic delay <topic>) display Delay of topic from timestamp in header.
    - Echo (ros2 topic echo <topic>) Output messages from a topic.
    - Hz (ros2 topic hz <topic>) print the average receiving rate (Hz)
    - Pub (ros2 topic pub <topic> <type> <data>) Publish a message to a topic.


- ROS2 SERVICE
    - Call (ros2 service call <service> <type> <args>) call a service.
    - Echo (ros2 service echo <service>) echo a service.
    - Find (ros2 service find <type>) output a List of available services of a given type.
    - Info (ros2 service info <service>) print Information about a service.
    - List (ros2 service list) output a List of available services.
    - Type (ros2 service type <service>) output a service's type.

- ROS2 ACTION
    - Info (ros2 action info <action>) print Information about an action.
    - List (ros2 action list) output a List of action names.
    - Send Goal (ros2 action send_goal <action> <type> <goal>) send an action goal.
    - Type (ros2 action type <action>) print an action's type.
    - Echo (ros2 action echo <action>) echo an action.
    - Find (ros2 action find <type>) find actions from type.

- ROS2 PARAM
    - Delete (ros2 param delete <node> <param>) delete parameter.
    - Describe (ros2 param describe <node> <param>) show descriptive Information about declared parameters.
    - Dump (ros2 param dump <node>) show all of the parameters of a node in a YAML file format.
    - Get (ros2 param get <node> <param>) get parameter.
    - List (ros2 param list) output a List of available parameters.
    - Load (ros2 param load <node> <file>) load parameter file for a node.
    - Set (ros2 param set <node> <param> <value>) set parameter.

Tool registry and tool metadata

• ToolRegistry now supports auto-loading default tools, with the ability to disable that behavior where needed.
• Added support for grouped tools via group_name.
• Added tool_description for user-facing tool listings and input_defaults for documented/defaulted inputs.
• Improved grouped registration logging and grouped tool rendering.

Planner, validator, and executor

• Added support for optional schema types like string?, int?, float?, etc.
• Executor now fills omitted arguments from input_defaults in schema order.
• Strengthened plan models so plans, plan nodes, and steps cannot be empty/missing required fields.

Console and streaming UX

• Added a dedicated subprocess/stream output panel for long-running streaming tools.
• Added support for keeping the main log anchored correctly during spinner/layout changes.
• Added /debug command and debug-only plan output logging.
• Improved /rerun handling and validation.
• Updated copy binding from F3 to F4.
• Improved grouped tool selection modal and suggestion modal behavior.

Shared tool utilities

• Moved subprocess execution and suggestion helpers into src/vulcanai/tools/utils.py.
• Improved streaming subprocess capture, cancellation, output routing, and suggestion handling.

CI and tests

• Split unit tests in CI so generic unit tests run separately from ROS 2-dependent default tool tests.
• Added a ROS 2 CI job using a Vulcanexus container for test_default_tools.py.
• Added extensive new tests for:
  • ROS 2 default tools
  • optional/default argument behavior
  • grouped tool registry behavior
  • stricter validator/model constraints

[cferreiragonz]

First review just checking the code. We can consider creating some tests to check all tools.

I am concerned about the optional input parameters. I don't know how all LLMs will treat them.

I will re-review after reviewing Textualize PR

[cferreiragonz]

The default tools also require the user to provide a ROS2 node to the console, which implies building code to use them, we should be able to provide a real default tools where the user could use them without coding

[Danipiza]

The default tools also require the user to provide a ROS2 node to the console, which implies building code to use them, we should be able to provide a real default tools where the user could use them without coding

Should we disable the publisher/subscriber tools, or add a SharedNode class in default_tool.py to create a ros node if the user have not added one?

[cferreiragonz]

Suggested change

	super().__init__(model, registry, validator, k, max(3, hist_depth), logger, _default_tools)
	super().__init__(model=model, registry=registry, validator=validator, k=k, hist_depth=max(3, hist_depth), logger=logger, default_tools=default_tools)

[Danipiza]

Done in f2549ac

[cferreiragonz]

Remember to apply this

[Danipiza]

Done in 237a622

[Danipiza]

Also managed the exception in the copy feacture when the user does not have installed a copy apt package.

• sudo apt-get install xclip,
• sudo apt-get install xselect (on X11)
• sudo apt-get install wl-clipboard (on Wayland)

[cferreiragonz]

I will test them soon to check each tool individually. In the meantime, could you rebase onto main to solve the conflicts Github complains about.

[cferreiragonz]

If no tests are introduced, we should document or list a series of commands to which VulcanAI should correctly answer.

For example:

• Run a publisher in background and input: "Subscribe to topic /chatter with max message 1".

Note that this could also be done in a test-suite running in a docker container with Vulcanexus and checking the blackboard to verify the output of each call.

Lastly, I am still missing providing a default ROS 2 Node if default tools are enabled and no ROS 2 node is given. Otherwise tools won't work

[Danipiza]

• Ensured variables types in publish and subscribe tools to avoid this errors:

Execution failed for 'ros_subscribe': '<=' not supported between instances of 'str' and 'int'  

• Removed lock in the publish tool.
• Added custom_node.py with node example. Also added in the __init__.pyi

Done in 3d25f8d

[cferreiragonz]

I added the missing default ROS 2 Node and moved it to the same file as the default tools. I also corrected a bug in the ROS2TopicTool and created tests for its main arguments. I still think we have to much variability by allowing optional input arguments which only work with specific commands (for example the topic_name is required for the info subcommand, but it is not required for the find subcommand), this makes the LLM success rate very low. I also improved some logs.

• We would need to extend the test suits with a specific test for each tool sub-command to avoid having failing tests with every update.
• Calling /tools returns: Error: ColorParseError("'topic_name' is not a valid color")
• It would be great to parse the message: WARNING: topic \[/chatter] does not appear to be published yet shown when subscribing or publishing messages to avoid the trailing forward slash before the bracket.
• I could not find any informative message stating how to stop a subscriber or publisher. Something like "press Ctrl+C to stop".
• We need to review how things are displayed because right now "ros2 topic list" does not show any result. The user needs to check the Output dictionary, which right now does not use a pretty format. Additionally, subscribing to any message with max_lines argument automatically closes the new pop-up windows, so no output can be read.

[EXECUTOR] Invoking 'ros_subscribe' with args:'{'topic': '/chatter', 'max_lines': '15'}'
[TOOL ros_subscribe] Subprocess created!
[TOOL ros_subscribe] Returning only the last 10 lines in result['output'].
[EXECUTOR] Executed 'ros_subscribe' in 229.9 ms with result: //<-- This means I had 229.ms to visualize the data
[EXECUTOR] PlanNode PARALLEL succeeded on attempt 1/1```

[cferreiragonz]

Suggested change

	if max_duration is not None and not isinstance(max_duration, (int, float)):
	max_duration = None
	if max_duration is not None and max_duration <= 0:
	max_duration = None
	if not isinstance(max_duration, (int, float)) or max_duration <= 0:
	max_duration = None

[Danipiza]

I will answer the previous review comment here.

I still think we have to much variability by allowing optional input arguments which only work with specific commands (for example the topic_name is required for the info subcommand, but it is not required for the find subcommand), this makes the LLM success rate very low.

I have refactored the tools in multiple commands. And added a new optional variable in the tools called "group_name". This variable groups tools, making it look like subtools. When registering, editing the active tools or executing "/tools" it is displayed with group style.

The tools have better description and tags that makes the LLM choose rapidly the tool to execute with the input of the user. The time execution of the queries have decreased significantly.

Calling /tools returns: Error: ColorParseError("'topic_name' is not a valid color")

Added a new variable "tool_description" to print a brief description of the tools, instead of the entire description of the tool (the one that uses the LLM)

It would be great to parse the message: WARNING: topic \[/chatter] does not appear to be published yet shown when subscribing or publishing messages to avoid the trailing forward slash before the bracket.

I could not find any informative message stating how to stop a subscriber or publisher. Something like "press Ctrl+C to stop".

Already resolved.

The informative message was not being displayed because the streaming panel was created at the top of the terminal and pushed the text making it not visible unless the user scroll down. Now the streaming panel is located at the bottom and this issue does not happen.

We need to review how things are displayed because right now "ros2 topic list" does not show any result. The user needs to check the Output dictionary, which right now does not use a pretty format. Additionally, subscribing to any message with max_lines argument automatically closes the new pop-up windows, so no output can be read.

This also happened to me when I used OLlama with CPU (instead of GPU) and with the previous tools. Right now is solved.

We would need to extend the test suits with a specific test for each tool sub-command to avoid having failing tests with every update.

I am currently working with this and also:

• The implementation of the publisher.
• Parse "<xxx>" characters the user writes in the terminal

# USER QUERY print me the information of the ros2 topic </chatter>
[USER] >>> print me the information of the ros2 topic 
[MANAGER] Ollama response time: 2.726 seconds
[MANAGER] Prompt tokens: 1219, Completion tokens: 110
[MANAGER] Plan received:
- Plan Summary: Prints the information of the ROS2 topic 

- PlanNode 1: kind=SEQUENCE
[MANAGER] [ERROR] Error handling user request: 'chatter' is not a valid colo

[Danipiza]

Added keyboard support in modal screens.

RadioList (suggestions)

• up, bottom keys to navigate throw the options
• space to change the selected option
• enter to submit the option

CheckList (/edit_tools)

• up, bottom keys to navigate throw the options
• space to toggle the selected option
• enter to submit the checklist

Added two new buttons in the /edit_tools modal screen

• Toggle all tools
• Toggle default tools