To Android Probe

Turn one natural-language Android runtime question into a generated probe, one controlled execution, and an evidence-backed answer. Keep code generation and execution details hidden unless the user asks for them.

Workflow

1. Read the complete sibling ../to-android-probe-script/SKILL.md and follow its generation contract.
2. Read the complete sibling ../to-android-run/SKILL.md and follow its connection and execution contract.
3. Resolve the requested target, starting scene, allowed interactions, and required chain, state, or UI evidence. Inspect the target App source and existing presets when symbols or navigation steps are unknown; never guess them. Ask only when missing information makes execution unsafe or materially ambiguous.
4. Use to-android-probe-script to generate the smallest probe that answers the original question. Use a task-specific directory under the workspace artifacts/ for Module output unless the user supplies another directory.
5. Use to-android-run to execute exactly one generated Inline operation or one Module export. Preserve its single-operation and no-unsafe-retry rules.
6. After execution returns, inspect its result and only the evidence files it returned. Read only the references matching the returned evidence types: chain, state, or UI.
7. Relate observed runtime events to original App source, distinguish observed facts from source-based inference, and decide whether the question is answered.
8. If it is not answered, revise and execute another probe only when doing so is safe. Do not automatically retry a failed or timed-out operation that may have mutated App state; first restore or confirm a known starting scene.

Result

Answer the user's original question directly, then support the conclusion with one coherent account of the call outcome, runtime evidence, and original App source. Use only the evidence types relevant to the question; do not force chain, state, and UI sections when they were not requested or returned.