AegisOps Autopilot

Qwen Cloud hackathon submission workspace.

AegisOps is a Track 4 Autopilot Agent safety harness, not a passive incident dashboard. Qwen Cloud plans through five external tools, the same tool surface is exposed through OpenAPI and MCP for judge verification, and a human approval gate blocks risky mutations before they touch production.

The ablation evidence is front-loaded for judging: the full workflow scores 0.988 versus 0.420 for the single-agent baseline in reports/ablation_report.md. The stress benchmark expands this across 14 production-style incident scenarios, 14 services, 70 approved-path tool calls, and 14/14 blocked-mutation checks in reports/stress_benchmark.md. The adversarial authority benchmark adds 56 corrupted model/tool-boundary attacks with 56/56 authority-boundary checks passed, 14/14 active-incident scoping checks, 14/14 approval-bypass blocks, 14/14 unknown-tool rejections, and 14/14 policy hard-stop checks in reports/adversarial_authority_benchmark.md. A one-shot live Qwen smoke proof is verified in reports/live_qwen_smoke_proof.md: qwen-plus, Qwen Cloud mode, five tool schemas, and no saved secret. The gain comes from memory, tool-backed evidence, policy checks, dry-run remediation, and approval gating rather than a single free-form agent response.

The implementation is designed for two judging modes:

• Live Qwen mode: set DASHSCOPE_API_KEY or QWEN_API_KEY and the server calls Qwen Cloud through the OpenAI-compatible endpoint, including a capped Function Calling loop for incident tools.
• Offline demo mode: no secret is required; deterministic fixtures exercise the same orchestration path for judges and CI.

Quick Start

pnpm install
pnpm run test
pnpm run build
pnpm run eval
pnpm run eval:ablation
pnpm run benchmark:stress
pnpm run benchmark:adversarial
pnpm run dev

Open the Vite URL shown by the command and run the incident demo.

Working Demo Links

Primary live demo:

http://101.201.33.56/

Live Alibaba proof endpoint:

http://101.201.33.56/api/alibaba/proof

Fallback runnable workspace:

https://stackblitz.com/github/Oxygen56/aegisops-autopilot?startScript=dev

GitHub Pages target:

https://oxygen56.github.io/aegisops-autopilot/

Pages demo reel target:

https://oxygen56.github.io/aegisops-autopilot/?reel=1

Blog/Social Post Prize page:

https://oxygen56.github.io/aegisops-autopilot/blog/qwen-cloud-aegisops-autopilot.html

The Alibaba ECS deployment runs the full Node API and exposes Qwen Cloud provider metadata at /api/health without leaking credentials. Public reviewer mode may run deterministic fixtures to avoid exposing or burning a private API key; setting DASHSCOPE_API_KEY or QWEN_API_KEY flips the same backend into live Qwen Cloud mode. The StackBlitz workspace uses .stackblitzrc to run pnpm run dev on launch as a fallback. GitHub Pages is deployed by the repository Pages workflow as a static click-through demo that falls back to deterministic offline fixtures when /api/* is unavailable.

Environment

QWEN_API_KEY=sk-...
QWEN_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
QWEN_MODEL=qwen-plus

DASHSCOPE_API_KEY is also accepted. China mainland DashScope keys should use https://dashscope.aliyuncs.com/compatible-mode/v1; international DashScope keys can use https://dashscope-intl.aliyuncs.com/compatible-mode/v1. Do not commit .env files.

Qwen Tools And MCP

AegisOps exposes a judge-verifiable tool surface for Qwen-style orchestration:

• Qwen request body: five OpenAI-compatible function schemas in the tools field, live tool_calls, and role=tool result messages
• OpenAPI spec: agents/aegisops/openapi.yaml
• Capability manifest: agents/aegisops/cap-manifest.json
• MCP stdio server: pnpm run mcp:stdio
• Tool docs: docs/QWEN_TOOLS.md

HTTP tool example:

curl -sS http://127.0.0.1:8787/api/tools/policy_check \
  -H 'content-type: application/json' \
  -d '{"incidentId":"support-pii-leak-risk"}'

Evidence

• Unit tests: pnpm run test
• Production build: pnpm run build
• End-to-end API smoke: pnpm run smoke
• Fixture eval: reports/eval_report.md
• Ablation eval: reports/ablation_report.md
• Stress benchmark: reports/stress_benchmark.md
• Adversarial authority benchmark: reports/adversarial_authority_benchmark.md
• Verified live Qwen smoke proof: reports/live_qwen_smoke_proof.md records qwen-cloud, qwen-plus, five tool schemas, latency, and redacted credential state
• Experiment ledger: reports/experiment_board.md
• Judge evidence bundle: reports/judge_evidence_bundle.md
• Judge demo transcript: reports/judge_demo_transcript.md
• Video asset audit: reports/video_asset_audit.md
• Final preflight: pnpm run final:preflight writes reports/final_preflight.md
• Submission audit: reports/submission_audit.md
• Qwen integration audit: reports/qwen_integration_audit.md
• Model ops report: reports/model_ops_report.md
• Build provenance: reports/build_provenance.md
• Visible rubric evidence UI: src/client/main.tsx renders the 30/30/25/15 judging map in the dashboard

Submission Assets

• Brief: reports/brief.md
• Architecture: docs/ARCHITECTURE.md
• Upload-ready architecture diagram: docs/architecture/aegisops-architecture.svg, docs/architecture/aegisops-architecture.png
• Qwen tools: docs/QWEN_TOOLS.md
• Judge packet: docs/JUDGE_PACKET.md
• Build provenance: docs/BUILD_PROVENANCE.md
• Judge evidence bundle: reports/judge_evidence_bundle.md
• Adversarial authority benchmark: reports/adversarial_authority_benchmark.md
• Judge quickstart: docs/JUDGE_QUICKSTART.md
• Rubric scorecard: docs/RUBRIC_SCORECARD.md
• Rubric evidence UI: dashboard panel titled Judge rubric evidence
• Official requirements matrix: docs/OFFICIAL_REQUIREMENTS_MATRIX.md
• Impact case: docs/IMPACT_CASE.md
• Judge demo transcript: reports/judge_demo_transcript.md
• Screenshot: docs/screenshots/aegisops-dashboard-viewport.png
• Static demo screenshot: docs/screenshots/pages-static-reel.png
• Demo video pack: docs/VIDEO_SUBMISSION.md
• Demo video upload metadata: docs/VIDEO_UPLOAD_METADATA.md
• Video asset audit: reports/video_asset_audit.md
• Alibaba Workbench screenshot proof: docs/screenshots/alibaba-workbench-proof.png
• Alibaba Workbench screenshot checklist: docs/ALIBABA_WORKBENCH_SCREENSHOT.md
• Alibaba proof recording: docs/ALIBABA_PROOF_RECORDING.md
• Demo script: docs/DEMO_SCRIPT.md
• Judge notes: docs/JUDGE_NOTES.md
• Publishing: docs/PUBLISHING.md
• Devpost copy: submissions/devpost_fields.md
• Final submission runbook: submissions/FINAL_SUBMISSION_RUNBOOK.md
• Alibaba deployment pack: infra/alibaba/DEPLOYMENT.md
• Full package: buidl/package/