---

name: "exploration" description: "Guides AppAgent's autonomous exploration and human demonstration flows that generate UI documentation." metadata: disco-role: operating disable-model-invocation: true license: MIT

Exploration

Use this sub-skill when the task is to create or refresh the documentation base that AppAgent uses later during deployment.

What this sub-skill covers

• Autonomous exploration of an Android app.
• Human demonstration capture of a similar task.
• Document generation for UI elements after a demo.
• Output structure under apps/<app>/demos/, apps/<app>/auto_docs/, and apps/<app>/demo_docs/.

When to route here

• The user wants to "explore" an app.
• The user wants to "teach" AppAgent by demonstration.
• The user needs to regenerate docs for a specific app.
• The user is asking why AppAgent has no documentation base for a target app.

Read first

• references/workflows.md for the autonomous versus human-demo flow.
• references/troubleshooting.md for capture, prompt, and doc-generation failures.
• ../../references/configuration.md when the model backend or output directories need adjustment.
• ../../references/api-reference.md when you need parser or controller details.

Bundled script

Use the launcher in this sub-skill instead of trying to reconstruct the source launch sequence yourself:

• scripts/start_exploration.py

It accepts an explicit --mode, uses subprocess calls rather than shell-string execution, and takes --repo-root <AppAgent checkout> when you are not already inside the repository.

Typical workflow

1. Validate the environment with ../../scripts/check_setup.py --repo-root <AppAgent checkout> if the host/device status is uncertain.
2. Choose an app name and a writable root_dir.
3. Run the bundled exploration launcher in autonomous mode or human mode.
4. Review the generated docs in auto_docs/ or demo_docs/.
5. If the UI meaning is still wrong, rerun with DOC_REFINE=true or manually edit the docs.

What to keep in mind

• Exploration is interactive and requires a live device and adb.
• The model backend must understand images; text-only backends are not enough.
• step_recorder.py and document_generation.py are internal steps of the human-demo path.
• self_explorer.py is the internal autonomous loop.
• The exact model response format matters because the action parser is strict.

Common failure signals

• No device found.
• Invalid device size.
• Screenshot/XML capture failure.
• Malformed exploration or reflection response.
• Existing doc entry skipped because the UID already has documentation.

Useful outputs

• apps/<app>/demos/<demo_name>/record.txt
• apps/<app>/demos/<demo_name>/task_desc.txt
• apps/<app>/auto_docs/*.txt
• apps/<app>/demo_docs/*.txt

Cross-links

• Use ../deployment/SKILL.md after docs are ready.
• Use ../../references/troubleshooting.md when the failure is shared across both phases.