Built on the Google Antigravity SDK
📖 Read the Architectural Walkthrough on Blogger: From “Vibe Coding” to Closed-Loop SRE
Aegis-Chaos is a production-grade, zero-trust autonomous SRE control plane. It detects service anomalies, gates every destructive mutation behind a declarative security fabric, and drives remediation in isolated Git worktrees—without human intervention unless an approval checkpoint is required.
- Decoupled Declarative Safety Policies — L1 deny, L2 escrow approval, L6 wildcard allow for diagnostics.
- Parallel Subagent Orchestration — Spawns focused remediation agents in isolated Git worktrees to keep context windows small and token usage low.
- Non-Blocking Checkpoint State Machine — Math-based token budget (
C_turn,C_cumulative) with a hard kill-switch to prevent runaway loops. - At-Most-Once Execution — Firestore
delete-firstcheckpoint semantics guarantee exactly-once resumption even under Cloud Tasks retries. - End-to-End Visual Verification — Headless Chrome actuation captures design-token screenshots and
.webmrecordings for every remediation cycle. - Secure Webhook Ingestion — HMAC-SHA256 signature validation + ±60s replay guard before any task is enqueued.
flowchart TD
A[Anomaly Detected] --> B["Webhook HMAC + Timestamp Guard"]
B --> C["Cloud Tasks Enqueue run-{session_id}"]
D1["Worker: Firestore delete-first checkpoint"]
D2["Restore isolated Git worktree"]
D3["Resume Antigravity agent conversation"]
C --> D1
D1 --> D2
D2 --> D3
D3 --> G{Policy Layer}
G -->|L1 deny| H[Block]
G -->|L2 ask_user| I["CheckpointRequested → Firestore"]
G -->|L6 allow| J[Execute diagnostic]
I --> K[Out-of-band approval]
K -->|approved| D3
K -->|denied| H
J --> L["Token Ledger: C_turn, C_cumulative"]
L -->|budget breached| M["Kill-switch → scale down"]
L -->|in-budget| D3
D3 --> N["Browser actuation: verify UI + record WebM"]
N --> O["Report evidence + leave branch for review"]
# 1. Clone
git clone https://github.com/rabimba/aegis.git
cd aegis
# 2. Virtual environment
python3 -m venv venv
source venv/bin/activate
pip install -e .
# 3. Configure Antigravity CLI
mkdir -p ~/.gemini/antigravity-cli
cp .agents/skills/chaos-heal/scripts/render_title.sh ~/.gemini/antigravity-cli/
cp .agents/skills/chaos-heal/scripts/render_status.sh ~/.gemini/antigravity-cli/
# Edit ~/.gemini/antigravity-cli/settings.json and mcp_config.json per walkthrough.md
# 4. Launch services
uvicorn src.aegis_chaos.webhook_ingest:app --host 0.0.0.0 --port 8000
python tests/mock_services/app.py
# 5. Run tests
pytest -q
# 6. Run visual verification
python .agents/skills/chaos-heal/scripts/verify_ui.pyaegis/
├── src/aegis_chaos/
│ ├── policies.py # Declarative security fabric + CheckpointRequested
│ ├── orchestrator.py # Token ledger, kill-switch, background ingress monitor
│ ├── webhook_ingest.py # FastAPI HMAC + replay-guarded ingestion gateway
│ └── worker.py # Firestore delete-first checkpoint resumption
├── tests/
│ ├── conftest.py # SDK stubs for offline test execution
│ ├── test_pipeline.py # 8-case self-healing harness
│ └── mock_services/
│ └── app.py # Flask latency dashboard (port 8080)
├── .agents/skills/chaos-heal/
│ ├── SKILL.md # Agent playbook: plan → isolate → validate
│ └── scripts/
│ ├── render_title.sh
│ ├── render_status.sh
│ └── verify_ui.py # Playwright /browser actuation + WebM recording
├── artifacts/ # Screenshots + video evidence (gitignored binaries)
├── pyproject.toml
├── walkthrough.md # Deep-dive for judges / GDE Sprint panel
└── README.md
| Threat Vector | Mitigation |
|---|---|
| Tampered webhook payload | HMAC-SHA256 with hmac.compare_digest |
| Replay attack | Millisecond timestamp ±60s drift window |
| Runaway agent loop | Token budget kill-switch (C_cumulative > Θ_budget) |
| Destructive shell command | L1 explicit deny + L2 escrow checkpoint |
| Partial failure / retry | Firestore delete() before side effects |
| Context pollution | Git worktree isolation + scoped subagent contexts |
See walkthrough.md for the complete judging panel walkthrough, including architecture deep-dives, math derivations, and CLI setup commands.
Project Aegis-Chaos — Google Developer Expert Sprint