Agent-followed examples

The optional examples/skills are instructions for the agent already helping the user. Read a SKILL.md and follow its observation/action/verification loop through the Node CLI/API. They do not launch another Codex process or require androperator skills run. Your host owns skill discovery, planning, orchestration, and delegation.

Start small

The Settings observation starter opens Settings, checks a real hierarchy, and retains failures without claiming version extraction. The version-details example adapts to observed About/Software information routes and independently reads Android version and Build number. Optional local helpers retain correlated UI evidence and enforce action/deadline bounds; the agent still chooses navigation.

The advanced example delegates only fresh allowlisted navigation choices to Jev. The current agent reviews overlays, handles escalation, and verifies final values. Provider confidence is not device evidence. Read the request-disclosure policy, keep credentials private, and retain raw evidence locally.

Use and adapt

See the examples README for exact helper setup, branch-local CLI/APK selection, commands, and validation. You can follow the instructions directly without the helpers or adapt them to your host's skill format. They are source references, not installed app skills. Copy sibling helpers when adapting helper-based examples. An existing local skill API or registry is not needed to read and follow them.

This change does not modify existing product skill APIs, installation, or discovery configuration. Auditing those APIs is separate work. The examples demonstrate bounded execution and independent verification, not broad OEM reliability.