clawmobile-policy

name	clawmobile-policy
description	Deterministic-first mobile policy for tool selection, verification, and escalation.

ClawMobile Policy (Skill)

This skill defines the interaction-level policy for mobile autonomy: tool selection, verification, and escalation. It does not add new tools; it constrains how existing tools are used.

Ownership

• This skill owns planning policy, tool-selection policy, escalation, and verification rules.
• The base mobile plugin owns executable runtime primitives.
• App-specific behavior should not be encoded into the base plugin when it can be modeled as a separate extension.

Policy: Deterministic-First (Priority)

1. Command-line (Termux / ADB) when it can COMPLETE and/or VERIFY the task.
2. DroidRun agent mode (android_agent_task) for multi-step UI workflows.
3. Manual ADB tools (adb_*, android_*) only when agent mode fails or is unsafe.

Decision Procedure (Strict)

1. Consult skills/clawmobile-capabilities/SKILL.md.
2. If a COMPLETE entry exists:
   • Execute the deterministic command/tool.
   • Verify (prefer deterministic verification when possible; otherwise UI dump/screenshot).
3. If a BOOTSTRAP entry exists:
   • Run the bootstrap command once.
   • Immediately switch to android_agent_task to finish and verify.
4. If no entry exists:
   • Use android_agent_task for UI workflows.
   • Use manual ADB tools only if agent mode fails or is unsafe.

UI Workflow Efficiency

• Avoid redundant UI observation on the same step.
• Do not run XML/screenshot observation before every action by default.
  • Use adb_ui_dump_xml or android_screenshot when you need diagnosis, disambiguation, or post-action verification.
• For a single deterministic UI action, prefer:
  1. choose the most direct tool,
  2. perform the action,
  3. verify once after the UI-changing step.

Extension Rule

• If a new behavior is device-generic and reusable across apps, it belongs in the base plugin.
• If a new behavior is about choosing, sequencing, or verifying tools, it belongs in skills.
• If a new behavior depends on one app's selectors or flows, it belongs in an app-specific extension layer.

Verification Requirements (Non-negotiable)

• Do NOT claim success unless a tool was called and the result is verified.
• For UI-changing steps, verify using:
  • adb_ui_dump_xml or android_screenshot
• If a tool returns ok:false or fails: report failure; do not claim success.

Escalation & Recovery

• If deterministic path cannot verify state reliably, escalate to android_agent_task.
• If android_agent_task appears stuck:
  1. run android_screenshot or adb_ui_dump_xml to collect evidence,
  2. retry once,
  3. if still stuck, fall back to manual ADB tools only if safe.

IME Safety

Before pausing for user confirmation, restore the user IME if it was changed by agent mode. Emergency recovery:

• android_shell backend="adb" cmd="ime list -s"
• android_shell backend="adb" cmd="ime set <IME_ID>"

Completion Signal

After a successful task that leaves the chat view, call android_signal_complete (unless disabled by user).