Support session-scoped runtime hints and remote artifact retrieval in native remote daemon mode

[thymikee]

Problem

agent-device remote daemon mode is close to a transparent remote UX, but there is still a gap for React Native dev/simulator/emulator flows when the CLI runs in one environment and the daemon runs in another.

Current setup:

• CLI runs in a Linux sandbox
• daemon runs on a remote macOS host
• local mobile artifacts can be uploaded remotely
• normal remote commands like devices work
• install/open plumbing can work remotely

However, runtime connectivity information (for example Metro host/port or other platform-specific debug runtime settings) is not transparently carried through native remote daemon mode in a way that makes later UI commands just work.

This means the following transparent workflow is still incomplete:

agent-device install MyApp ./artifact --platform ios --session my-session
agent-device open MyApp --platform ios --session my-session
agent-device snapshot --platform ios --session my-session
agent-device click @e7 --platform ios --session my-session

and similarly for Android:

agent-device install com.example ./app-debug.apk --platform android --session my-session
agent-device open com.example --platform android --session my-session
agent-device snapshot --platform android --session my-session
agent-device click @e7 --platform android --session my-session

In remote daemon mode, the CLI needs a way to associate runtime/network information with the session so later UI commands work without custom proxy-specific helpers.

Requested capability

Add native support for session-scoped runtime hints in remote daemon mode for both iOS and Android.

Examples of the kind of information needed:

• Metro host
• Metro port
• optionally resolved bundle URL / launch URL if applicable
• any Android-specific debug-runtime connectivity settings required for dev builds

Possible API shapes

Option A: session runtime command

agent-device runtime set \
  --session my-session \
  --platform ios \
  --metro-host 10.0.0.10 \
  --metro-port 8081

Then subsequent commands automatically use those runtime hints:

agent-device open MyApp --platform ios --session my-session
agent-device snapshot --platform ios --session my-session
agent-device click @e7 --platform ios --session my-session

A similar flow should exist for Android if platform-specific runtime data is needed.

Option B: environment-based runtime hints

AGENT_DEVICE_IOS_METRO_HOST=10.0.0.10
AGENT_DEVICE_IOS_METRO_PORT=8081
agent-device open MyApp --platform ios --session my-session

Option C: flags on each command

agent-device open MyApp \
  --platform ios \
  --session my-session \
  --metro-host 10.0.0.10 \
  --metro-port 8081

This is less ergonomic than session-scoped state, but still better than requiring a separate proxy-specific execution API.

Also required: remote artifact retrieval back to the caller

In the same remote setup, screenshots, recordings, and other generated artifacts must be returned in a way the remote caller can actually consume.

Today, a remote command can end up returning host-local paths like:

• /Users/.../screenshot.png
• ~/Library/...
• host-local log paths

Those are not useful to a caller running in a sandbox or other remote environment.

Requested behavior

When the daemon is remote, artifact-producing commands such as:

• screenshot
• snapshot (if it emits files)
• record
• any future file-producing output

should return results that are retrievable by the client.

Possible shapes:

• stream bytes back directly
• upload artifacts through the daemon transport and return:
  • artifact ID
  • downloadable URL
  • or client-materialized local temp path
• provide a follow-up download endpoint/command in native remote mode

The important part is:
remote callers must not receive unusable host-local paths as the primary artifact output.

Why this matters

This would complete the transparent remote UX for AI agent / sandbox workflows:

• install app remotely
• start Metro / provide runtime hints once
• run iOS and Android UI commands normally through agent-device
• receive screenshots/recordings back in the caller environment

without requiring:

• proxy-specific exec helpers
• custom transport wrappers
• special-case code paths outside agent-device

Concrete use case

A QA workflow running in a sandbox:

1. downloads an Android APK or iOS simulator .app artifact from CI
2. installs it remotely via agent-device
3. starts Metro separately if needed
4. needs to launch the app and interact with it using normal agent-device commands
5. needs screenshots/recordings to come back to the sandbox as usable artifacts

Today steps 4-5 still require custom proxy-specific handling.

This issue is about making native agent-device remote daemon mode fully transparent for that setup on both iOS and Android.

[thymikee]

closed by #199