Play Policy Insights

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.
2. 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}).
3. 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.
4. 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.
2. 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.
3. 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.
4. Proceed to Finalization (Step 4 below)

Finalization (Both Modes)

4. 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.
5. STOP: The audit is complete. Await further instructions.