[deploy] Start TERA agent on Jetson with OLLAMA_MODEL=gemma3:4b

[jdev-02]

Lane: deploy · Owners: Kyle (Jetson runtime) + Jon (env config) · Phase: P3 critical-path · Priority: P0

Why

Two things are running on the Jetson independently:

1. ollama serve with gemma3:4b pulled (verified via llm_dev_kmh/tera_gemma/chat_tera_gemma.py)
2. not yet: the TERA agent service (agent/app.py FastAPI)

agent/llm.py OllamaClient defaults to gemma2:2b (line 327). Without overriding, the agent will fail to pull a non-existent model on first request.

Acceptance criteria

• OLLAMA_MODEL=gemma3:4b TERA_PHASE=3 TERA_DEVICE_PROFILE=austere uvicorn agent.app:app --host 0.0.0.0 --port 8000 runs on the Jetson and stays up.
• curl http://<jetson>:8000/health returns {"status":"ok","phase":"3","profile":"austere"}
• First /plan POST against this endpoint produces a non-empty RouteQuery from Gemma 3 4B (check server logs for llm_emitted_route_query).
• OllamaClient's loopback enforcement passes (OLLAMA_HOST=http://127.0.0.1:11434 is the default; verify port 11434 is loopback-only).
• Latency from /plan POST to first byte response < 30s (PRD §11.2 target).

File pointers

• Agent runtime: agent/app.py, env vars in agent/llm.py (OLLAMA_MODEL)
• New if needed: deploy/systemd/tera-agent.service (post-MVP, [deploy] systemd unit for agent + bridge on Jetson #14)

Notes

• Defaults to keep open: OLLAMA_HOST left default (loopback enforcement is intentional)
• TERA_DEVICE_PROFILE=austere ensures no frontier client constructed even if API keys are present (PRD §8 OPSEC)
• For the demo, just running uvicorn in a tmux pane is fine — systemd unit ([deploy] systemd unit for agent + bridge on Jetson #14) is post-MVP

Dependency / blocker

Once this is up, the wire-format issue and polyline-render issue (sister #s) can be tested end-to-end.

Refs PRD §7.2 LLM rows, §11.2 latency target, §13.3 Sun 0700 gate.

[kylemhicks]

Status after PR #101/#103: the Jetson demo path now starts/checks Ollama, pulls/warms gemma3:4b, and activates TERA ATAK local mode from the Source Planner button. The compose refresh also does a nonblocking host-side warmup before restarting the web app.\n\nLeaving this open because the acceptance criteria here are for a separate agent.app:app service on port 8000 with /health returning phase/profile and /plan emitting RouteQuery logs. The shipped demo path is currently llm_dev_kmh on port 8080 plus ATAK plugin calls, not the standalone agent.app:8000 service.