First debugging session

Phone is paired with the robot. Time to actually use Scry. Five things to try, in order — each one demos a different capability.

1. "What's my robot's health?"

In the chat composer at the bottom of the Scry tab, type:

What's my robot's health?

Tap Send. The agent will:

1. Call the robot_health MCP tool on the connect
2. Get back a structured snapshot: nodes alive, topics publishing, battery if available, ROS_DOMAIN_ID, transient errors
3. Render an inline health card (the green/yellow boxes)
4. Add a short text conclusion above the card

Total time: 1–3 seconds. The data is live — every call hits your robot, no caching.

2. "What topics are publishing?"

What topics are publishing right now?

The agent calls ros_topic_list and renders a table with each topic's name, type, and publish rate (where measurable). Tap a topic to drill into details.

3. Inspect a single topic

What's on /odom?

If /odom exists on your robot, the agent calls ros_topic_echo for one message and renders the JSON inline. For high-frequency topics (IMU, camera, lidar) it instead calls ros_topic_hz and gives you the publish rate + std-dev.

4. Voice input

Tap the microphone icon in the composer (between the + and the send button). Say:

What was the last warning in /rosout?

Release. The transcript appears in the composer; tap send. Voice uses the Android system SpeechRecognizer — Google's on-device model on most phones, no audio leaves the device.

5. Image attachment

Take a screenshot of an RViz scene, a terminal showing an error, or a photo of the robot itself. In Scry:

1. Tap the + icon in the composer
2. Choose Gallery (or Take photo to capture fresh)
3. Pick the image — it appears as a chip above the input
4. Type a question, e.g. what's wrong in this RViz scene?
5. Send

The agent uses your provider's vision model (Claude Sonnet, GPT-4o, Gemini Pro, etc.) to actually look at the image. Useful for "the robot is doing weird stuff and I don't even know how to describe it."

What just happened under the hood

Every message you send goes through the same loop:

You type → ChatViewModel
  → AiProxyLoop builds: system prompt + tier-0 tools (~10 core MCP tools)
  → AI provider streams response
  → If AI calls a tool:
      - Read tool? → run immediately, send result back
      - Write tool? → show approval dialog → wait for your tap → run
  → AI continues until it sends final text
  → save to local Room DB

The AI doesn't have a persistent connection to your robot. Each tool call is a fresh HTTP request to scry-connect, scoped to that turn. Nothing crosses the cloud except the AI inference itself (and that's zero-retention with most providers).

When something breaks

Three places to look, in order:

1. Look at the message in chat. If the AI says "I tried ros_topic_list and got an error: …" — that's the connect's response. Usually means the robot's state changed (node crashed, topic stopped).
2. Check the connect logs on the robot: journalctl --user -u scry-connect -f. Per-call audit logs are in /var/log/scry/audit.log.
3. File feedback via Settings → Feedback in the app, or on the specific reply. Both go to the operator's dashboard.

What's next

Now that you've kicked the tires:

• Use Scry — the full feature tour (monitors, fleet view, multi-robot, etc.)
• Architecture — how the pieces fit together if you want to extend or contribute
• MCP tools reference — every tool the agent can call, with what it returns