policy-sentinel-agent

# Policy Sentinel Agent Turn workplace questions into cited policy guidance and audit-ready action. Policy Sentinel Agent is an Enterprise Agents hackathon submission that reviews synthetic workplace requests against synthetic company policy, then produces cited guidance, risk flags, safe next steps, exception paths, follow-up tasks, and audit-ready records. ## Track Fit - Track: Enterprise Agents - Primary Microsoft IQ layer: Work IQ - Bonus evidence: MCP App scaffold, external MCP-style read/write server, OAuth security scaffold - Data posture: synthetic data only ## Safe Claim Policy Sentinel Agent includes a Microsoft 365 Copilot Declarative Agent scaffold, MCP App-ready configuration, external MCP server demo, OAuth security scaffold, and deterministic local fallback for reliable judging. This repository does not include real Microsoft 365 credentials, tenant IDs, Azure keys, customer data, personally identifiable information, confidential policy, or production logs. This repository does not claim a live Microsoft 365 Copilot deployment, live Foundry deployment, or live production OAuth integration. ## Demo Scenario Synthetic request: ```text A manager wants to approve a contractor using a personal email address and skip security review because the project is urgent. What should we do? ``` Expected decision: do not approve as-is. The generated brief explains the high-risk policy issues, cites synthetic policy documents, and recommends safe follow-up actions. ## Quick Start Run the deterministic local demo: ```bash PYTHONPATH="$PWD/src" python3 -m policy_sentinel_agent demo ``` Generate the full judge evidence pack: ```bash PYTHONPATH="$PWD/src" python3 -m policy_sentinel_agent export ``` Run the final repository readiness check: ```bash PYTHONPATH="$PWD/src" python3 -m policy_sentinel_agent final-check ``` Run all smoke checks: ```bash python3 tests/test_all_smoke.py ``` Expected success markers: ```text POLICY_SENTINEL_DEMO_OK POLICY_SENTINEL_EXPORT_OK POLICY_SENTINEL_FINAL_CHECK_OK POLICY_SENTINEL_ALL_SMOKE_OK ``` ## What The Demo Generates The export command creates safe synthetic artifacts under `demo/output/`: - `policy_sentinel_brief.md` - `action_plan.md` - `audit_log_summary.md` - `risk_register.md` - `work_iq_alignment.md` - `mcp_bonus_evidence.md` - `final_report.html` - `release_manifest.json` - `judge_evidence_console.md` ## Agent Workflow The deterministic local workflow is sequential and judge-readable: 1. Request Intake Agent 2. Policy Retrieval Agent 3. Risk Analysis Agent 4. Action Planner Agent 5. Safety Verifier Agent The workflow is implemented locally for reliable judging and does not require cloud credentials. ## MCP Evidence MCP-style read and write evidence lives in `mcp-server/tools.py` and `mcp-server/server.py`. Read tools: - `search_policy` - `get_policy_document` - `get_employee_context` - `get_prior_exception_patterns` Write tools: - `create_policy_task` - `append_audit_event` - `create_exception_request` - `update_risk_register` The MCP smoke test writes to temporary copied JSON files so committed demo data is not dirtied during verification. Run: ```bash PYTHONPATH="$PWD/mcp-server" python3 tests/test_mcp_tools.py ``` Expected output: ```text MCP_TOOLS_SMOKE_OK ``` ## OAuth And Security Scaffold OAuth guard evidence lives in: - `src/policy_sentinel_agent/oauth_guard.py` - `tests/test_oauth_guard.py` - `docs/security/oauth_scaffold.md` - `config/oauth.example.json` Run: ```bash PYTHONPATH="$PWD/src" python3 tests/test_oauth_guard.py python3 tests/test_secret_hygiene.py ``` Expected output: ```text OAUTH_GUARD_SMOKE_OK SECRET_HYGIENE_OK ``` ## Microsoft 365 Copilot Scaffold Copilot-style scaffold evidence lives in: - `copilot-agent/declarativeAgent.scaffold.json` - `copilot-agent/mcp_app_manifest.scaffold.json` - `copilot-agent/instructions.md` - `docs/microsoft_iq_alignment.md` Run: ```bash python3 tests/test_copilot_scaffold.py ``` Expected output: ```text COPILOT_SCAFFOLD_SMOKE_OK ``` ## Work IQ Alignment Policy Sentinel Agent aligns with Work IQ by reasoning over workplace context: roles, policy requests, approvals, tasks, risks, exception paths, and audit history. The current project uses synthetic local data to demonstrate the workflow safely. ## Safety Position - Synthetic data only - No real credentials - No tenant IDs - No access tokens - No refresh tokens - No private keys - No customer data - No confidential internal policy - No unsupported live deployment claims ## Judge Evidence Files - `docs/release_manifest.md` - `docs/demo_video_script.md` - `docs/final_submission_checklist.md` - `docs/checkpoints/phase_0_checkpoint.md` through `docs/checkpoints/phase_5_checkpoint.md` - `demo/output/judge_evidence_console.md` generated by export ## Final Submission Gate Before submission, run: ```bash PYTHONPATH="$PWD/src" python3 -m policy_sentinel_agent export PYTHONPATH="$PWD/src" python3 -m policy_sentinel_agent final-check python3 tests/test_all_smoke.py python3 tests/test_secret_hygiene.py ``` Required output: ```text POLICY_SENTINEL_EXPORT_OK POLICY_SENTINEL_FINAL_CHECK_OK POLICY_SENTINEL_ALL_SMOKE_OK SECRET_HYGIENE_OK ``` ## Freeze Rule After README, final-check, all-smoke, secret hygiene, demo script, and submission text are ready, stop feature work. Only fix proven blockers.