play-policy-insights

Play Policy Insights: data safety, login credentials, and restricted permissions

You must audit Android apps for three specific policy domains. You must check

data safety, demo login credentials, and restricted permissions.

Path Resolution

• repo_root: Absolute path to the directory containing this SKILL.md.
• app_dir:: Absolute path to the directory containing app's code.
• temp_dir: Absolute path to the scratch directory at the workspace root.

It is located at .scratch/play_policy_insights_<uuid>. **Containment

Mandate**: You must confine all file system writes, intermediate artifacts,

and logs strictly to this directory. This ensures the skill remains portable

and safe across diverse execution environments, including local harnesses

and CI/CD pipelines, by avoiding reliance on system-level temporary paths or

user home directories.

Critical mandates

• Execution Mode Awareness Before starting Phase 2, evaluate if your

execution environment provides a tool to spawn or delegate tasks to

general-purpose sub-agents (e.g., tools often named invoke_agent,

delegate_task, or spawn_worker, using generic agent profiles like

'generalist' or 'coding_agent').

• If YES, you MUST use Mode A (Delegation).

• If NO, use Mode B (Sequential Self-Execution). You must read the

prompt files intended for the subagents, follow their instructions, and

write the expected output files to disk.

• Sub-agents orchestration:

• If you use "Mode A (Delegation)", wait for "SUCCESS" confirmation from

sub-agents to know when they are done.

• Idempotency & Timeout Safeguard: If a sub-agent fails or times out,

you MUST verify the presence and integrity of its target output file

(e.g., <temp_dir>/worker_<goal_name>.json) before retrying. If the

file exists and contains valid JSON, treat the execution as SUCCESS

and proceed. Otherwise, retry up to three times.

• Fail-fast mandate: The automated audit in Phase 1 is the source of

truth. If orchestrator.py fails, you must stop immediately with an

explanation of failure. Do not use manual auditing as a fallback.

The two-phase protocol

Phase 1: Fact gathering and triage

1. Initialize and triage:

• Run python3 <repo_root>/scripts/orchestrator.py init <app_dir>.
• This will create the scratch environment, perform static analysis, map

the codebase, identify audit goals, and produce prompts for subagents

for each audit goal and prompts for designated critic and aggregator

subagents.

• You must wait (up to 5 minutes) for the script to finish.

1. Capture environment: Note values of the temp_dir, and

activated_goals from the JSON output. You will need them in Phase 2.

1. Evaluate goals: If activated_goals is empty, skip to step 3 of Phase 2

(Aggregation). Otherwise, proceed to step 1 of Phase 2 (Detailed analysis).

Phase 2: Goal-oriented audit

Determine your execution capabilities and proceed with either Mode A OR Mode B.

Mode A: Orchestrator WITH Delegation Capabilities (Parallel)

1. Detailed analysis: For each goal in activated_goals (e.g.,

permissions_and_apis, data_safety_part_1, data_safety_part_2),

delegate to a sub-agent. Concurrency Limit: You must not spawn more than

3 sub-agents simultaneously. Spawn the first batch of up to 3, wait for

their completions, and then spawn the next batch. Repeat until all goals are

complete. Pass the prompt: `"Read your instructions from

<temp_dir>/prompt_worker_<goal_name>.md and execute. MANDATORY: You must

use your file-writing capabilities to save your final JSON findings directly

to the file system at <temp_dir>/worker_<goal_name>.json. You are strictly

forbidden from outputting the JSON in your chat response. To minimize

context usage, your final response must be exactly 'SUCCESS' and nothing

else."` Validate: Confirm every

<temp_dir>/worker_<goal_name>.json exists and contains valid JSON. If a

sub-agent fails or times out, but the valid JSON output file is already

present on disk, do NOT retry; proceed normally. Only retry the

corresponding worker (up to three times) if the file is missing or invalid.

1. Aggregate Findings: Execute the python aggregation command:

python3 <repo_root>/scripts/orchestrator.py aggregate <temp_dir>. This

produces aggregated_findings.json and returns a JSON object containing

critic_chunks representing the number of chunks to verify (e.g.,

{"temp_dir": "...", "critic_chunks": 2}).

1. Parallel Critic review: For each chunk index i from 1 to

critic_chunks, delegate to a sub-agent. Concurrency Limit: You must not

spawn more than 3 critic sub-agents simultaneously. Batch them in groups of 3

as above. Pass the prompt:

"Read your instructions from <temp_dir>/prompt_critic_<i>.md and execute. MANDATORY: You must use your file-writing capabilities to save your final JSON findings directly to the file system at <temp_dir>/critic_output_<i>.json. You are strictly forbidden from outputting the JSON in your chat response. To minimize context usage, your final response must be exactly 'SUCCESS' and nothing else."

Validate: Confirm each <temp_dir>/critic_output_<i>.json exists and

contains valid JSON before proceeding. If it failed or timed out, but the

valid JSON file is present, proceed normally. Otherwise, retry that specific

critic chunk.

1. Proceed to Finalization (Step 4 below)

Mode B: Orchestrator WITHOUT Delegation Capabilities (Sequential)

1. Detailed Analysis: For each goal in activated_goals, sequentially:

• Read the contents of <temp_dir>/prompt_worker_<goal_name>.md.
• Execute the instructions contained within that file yourself.
• CRITICAL: You MUST format your findings exactly as requested in the

prompt and save them to <temp_dir>/worker_<goal_name>.json. Do not

summarize findings in your thoughts or chat; move to the next task.

• Validate: Confirm <temp_dir>/worker_<goal_name>.json exists before

moving to the next goal.

1. Aggregate Findings: Execute the python aggregation command:

python3 <repo_root>/scripts/orchestrator.py --aggregate <temp_dir>.

This produces aggregated_findings.json and returns a JSON object containing

critic_chunks representing the number of chunks to verify.

1. Sequential Critic review: For each chunk index i from 1 to

critic_chunks, sequentially:

• Read the contents of <temp_dir>/prompt_critic_<i>.md.
• Execute the steps yourself and save your findings to

<temp_dir>/critic_output_<i>.json.

• Validate: Confirm <temp_dir>/critic_output_<i>.json exists before

moving to the next chunk.

1. Proceed to Finalization (Step 4 below)

Finalization (Both Modes)

1. Present findings: Run python3 <repo_root>/scripts/generate_report.py <temp_dir>.

It will produce <temp_dir>/compliance_report.md. Present this output file to user.

1. STOP: The audit is complete. Await further instructions.