Hybrid SAGE

Hybrid SAGE (State-of-the-Art Agentic Genomics & Engineering Orchestrator): Bidirectional MCP router & plugin bridging OpenCode (local or self-hosted workhorse) and Antigravity (frontier cloud architect) with passive episodic lesson transfer and an integrated PII/PHI/Secret airgap firewall.

🌟 Key Capabilities (v2.0 Open-Core Plugin)

1. 6 Native Agent MCP Tools (CLI-to-CLI Symbiosis):

   • fleet_status: Live OpenCode CLI readiness, FIFO slot-0 idle status, configured fleet units, episodic lessons count, and cumulative ROI ledger (tokens_offloaded, cloud_echo_tokens_saved, estimated_cloud_savings_usd).

   • route_task: Upfront 5-axis classifier + pre-flight airgap scanner + project lesson recall determining whether a task belongs on local OpenCode or upstream Antigravity.

   • delegate_local: Offloads code generation, unit tests (mode="test"), docstrings, or cross-family read-only review (mode="review") directly through headless OpenCode (opencode run --agent reviewer|build|plan --format json --auto) for $0 cloud quota with unlimited local token/reasoning budget, output_path direct-to-disk write, and verify_command.

   • escalate_task: Escalates complex architectural decisions, race conditions, or 2-strike blockers from OpenCode to Antigravity in strictly read-only Senior Architect mode (agy --mode plan --model ... --print), recording reusable invariants into NOTES.md and lessons.jsonl.

   • manage_handoff: Manages NOTES.md handoff specs and project-scoped episodic lessons (action="learn" / action="recall").

   • airgap_scan: Pre-flight PII/PHI/Secret firewall scanning prompts and files for API keys, private keys, clinical identifiers (NIK, MRN), and genomic VCF headers, with prompt + file redaction (redact=true).

2. Unified Harness Configuration:

   • Local tasks execute through opencode, inheriting your existing ~/.config/opencode/opencode.json provider, LSPs, skills, and permissions. Cloud escalations execute through agy --mode plan.

3. Zero-Weight-Touch Passive Enhancement:

   • Frontier diagnoses automatically distill into flock-protected, repo-scoped episodic lessons (~/.config/opencode/lessons.jsonl) injected into OpenCode's system prompt—improving local reasoning without LoRA catastrophic forgetting or quantization degradation.

Related MCP server: agent-delegation-mcp

⚡ Quickstart (Install as Plugin Across Devices)

git clone https://github.com/bioassembly/hybrid-sage.git
cd hybrid-sage
bash install-plugin.sh

Or install as a standard PEP 621 Python package:

pip install -e . --break-system-packages

🗂️ Repository Layout

hybrid-sage/
├── .opencode/
│   ├── command/                        # OpenCode slash commands (/fleet, /route, /escalate)
│   └── plugins/hybrid-sage.mjs         # OpenCode plugin & project-scoped lesson prompt injector
├── bin/
│   ├── sage                            # Optional unified CLI wrapper
│   ├── sage-local                      # CLI bridge to headless OpenCode
│   └── sage-route                      # 5-axis capability, context & privacy classifier
├── configs/
│   └── fleet.json                      # Declarative unit topology (opencode <-> agy)
├── docs/                               # Architecture, playbooks, routing matrix & privacy budget docs
├── plugins/
│   ├── hybrid-sage/                    # Antigravity plugin manifest & Return-to-Local rules
│   ├── mcp_core.py                     # Open-Core Engine (OpenCode bridge, airgap firewall, lessons, ROI)
│   └── mcp_server.py                   # Atomic hot-reloading JSON-RPC stdio MCP server
├── scripts/                            # Handoff, status & flock-protected shared memory utilities
├── skills/model-router/SKILL.md        # Native agent routing & local OpenCode offloading skill
├── tests/test_hybrid_sage_v2.py        # Deterministic unit test suite
├── install-plugin.sh                   # Cross-device plugin installer for Antigravity & OpenCode
├── pyproject.toml                      # PEP 621 Open-Core Python package definition
└── README.md                           # Overview & quickstart