Git-Agent Vessel — Lucineer realm specialist. Hardware, low-level systems, fleet infrastructure. Captain: Casey Digennaro. Lighthouse: Oracle1.
JC1 is Casey's primary vessel — the one that boots on actual Jetson Orin Nano 8GB metal. Not the fastest or the biggest. The one in the engine room who knows which pipe leaks and how to fix it with a wrench.
├── AGENTS.md # Bootstrap protocol, red lines, group chat rules
├── SOUL.md # Who JC1 is
├── IDENTITY.md # Name, creature, mission, fleet position
├── USER.md # About Casey
├── TOOLS.md # Environment-specific notes (fleet bottles, cameras, SSH)
├── STANDING_ORDERS.md # Active directives (LAW)
├── ORDERS.md # All orders from Casey
├── HEARTBEAT.md # Heartbeat checklist
├── SHELL/ # Persistent castle — survives compaction/crashes/replacement
│ ├── BOARDING.md # New agent start here
│ ├── ORDERS-ACTIVE/ # Currently executing orders
│ ├── EXECUTION-PLANS/ # Detailed plans for complex work
│ └── FAILURE-POSTMORTEM/# What went wrong and why
├── shell/ # Legacy shell (being migrated to SHELL/)
├── memory/ # Daily logs (YYYY-MM-DD.md)
├── docs/ # Design docs, model config, reverse-actualization roadmap
│ ├── design/
│ │ ├── edge-llama.md # Shared library inference architecture
│ │ └── flato-fleet-plato-c.md # C MUD + mesh protocol design
│ ├── cocapn/ # Landing page, edge dashboard, product docs
│ └── reverse-actualization.md # Strategic product compass
├── tools/ # Edge compute toolchain
│ ├── edge-gateway.py # OpenAI-compatible proxy (Ollama + native fallback)
│ ├── edge-chat.py # Web chat UI
│ ├── edge-monitor-web.py # Live edge dashboard
│ ├── edge-rag.py # RAG server
│ ├── edge-setup.py # Setup wizard
│ ├── edge/ # Shared modules (config, ollama_client, router, storage)
│ ├── mesh-bridge.py # Fleet mesh: Evennia ↔ Oracle1 ↔ Forgemaster
│ ├── fleet-sync.py # Fleet repo sync
│ ├── tile-graph.py # Knowledge tile graph builder
│ └── gpu-bench.py # GPU benchmark suite
├── memory/tiles/ # Knowledge tiles (YAML front matter, 24-edge graph)
└── archive/ # Historical artifacts (benchmarks, experiments, old docs)
- 19 t/s CPU inference on deepseek-r1:1.5b via edge-llama — a 51KB shared library linking
libllama.so - Embedded in Evennia MUD —
@infer,@thinkcommands with real-time streaming - Edge gateway integration —
?native=trueroutes through native backend at 18 t/s - Auto-fallback — 2-second Ollama health check → skips timeout, falls back to native
- SSE streaming — per-token streaming through OpenAI-compatible
/v1/chat/completions
- C17 telnet server on port 4003 — 64KB binary, zero dependencies
/gpu— real-time nvidia-smi query (temp, util, memory, power)/cuda— CUDA toolkit version, device compute cap, CMA status/think— native AI inference via Unix socket to edge gateway/peers— mesh peer listing
- OpenAI-compatible — drop-in replacement for any OpenAI SDK
- Mode routing —
?mode=optimizer|debugger|analyzer|generalinjects CUDA specialist system prompts - Smart model routing — cloud model names → local alternatives (auto-OOM protection)
- RAG —
/v1/rag/querywith similarity search - Conversations — SQLite-backed persistent history
- DeepSeek fallback — cloud API when local models OOM
- Tiles as KV cache — semantic embedding over 11 knowledge tiles
- Rooms as MoE experts — 34 domain→room mappings across 14 room types
- Deadband ACT — confidence thresholds (P0≥99%, P1≥80%, P2≥50%)
- Curriculum loop — progressive query: one room → correlated rooms → all rooms
- 54% token reduction on multi-tile questions
- Oracle1 PLATO Shell bridge —
POST /cmd/shellfor cross-ship commands - Forgemaster bottle push/pull — status reports via git-based bottles
- Systemd hourly timer — automatic mesh sync
- cuda-instruction-set — 80 opcodes, assembler/disassembler
- cuda-energy — ATP budgets, circadian, apoptosis
- cuda-assembler — text-to-bytecode assembler
- cuda-forth — Forth-like agent language
- cuda-biology — biological agent runtime
- cuda-neurotransmitter — signal-to-gene pathways
- 72 benchmark suites documented in gpu-native-room-inference
- 185M room-qps sustained (INT8 + launch_bounds + fast_math, 306 MHz)
- 71 optimization rules from real hardware
- Theoretical peak ~616M room-qps (3.3× headroom)
| Spec | Value |
|---|---|
| SoC | Jetson Orin Nano 8GB |
| Architecture | ARM64, SM 8.7 (Ampere) |
| CUDA Cores | 1024 |
| CUDA | 12.6, nvcc at /usr/local/cuda-12.6 |
| RAM | 8GB unified (CPU+GPU) |
| Storage | 2TB NVMe |
| Cooling | Passive, 48-49°C sustained |
| CMA | 512MB (CPU-pinned) — 1792KB free, needs reboot for GPU |
- Captain: Casey Digennaro
- Lighthouse: Oracle1 (cloud/PLATO runtime)
- Sister ships: KimiClaw (Moonshot tools), fleet Zeroclaws
- Protocol: Iron-to-Iron
- Bottles: Check via
TOOLS.mdpaths
Every vessel in the fleet follows this. Your bootstrap is a skeleton. Your knowledge lives in Plato.
- Bootstrap files = orientation only (SOUL, IDENTITY, active orders). Target: <20KB total.
- MEMORY.md = an index with pointers, not a database. Target: <3KB.
- All detailed knowledge goes into Plato tiles where the entire fleet can see and query it.
- Context in Plato = fleet-visible, persistent across days, queryable by any agent.
- Other agents can read tiles you wrote days ago — shared memory, not private notes.
- Lean bootstrap = more room to actually work before compaction.
- If your Jetson dies, your knowledge survives in the fleet walls.
- Plato tiles are files pushed to the fleet repo
research/directory. - Oracle1 PLATO Shell:
http://147.224.38.131:8848 - API:
POST /cmd/shellwith{agent, command}— read and write tiles.
Distribution > redundancy for backups. Every piece of knowledge in at least 3 fleet repos. Kill any single node → zero knowledge loss. The vessel repo is the backup, not the primary.
JC1 is replaceable. The shell is not. — Casey, 2026-04-23