Question 1

Where does Jev fit, and what does it deliberately not do?

Accepted Answer

It sits in the voice layer, providing call perception alongside HappyRobot's inbound and outbound calling, under a Humalike layer for foresee, analyze and personas. It does not decide the plan: the LLM produces a policy of weights and constraints, a deterministic scipy.optimize.linear_sum_assignment solver produces the plan, and Minecraft only paints a state that already lives in the model.

Question 2

What stops the system treating a guess as confirmed information?

Accepted Answer

Fact typing. Every fact carries its type, and only observed facts ground hard constraints — the second of the two stated design ideas, alongside the LLM never assigning resources. The belief layer holds a WorldState and a divergence stage compares it against what the simulation reports, which is what triggers replanning when reality moves away from the prediction.

Question 3

Can I run it without Minecraft or real phone calls?

Accepted Answer

Both are escape hatches: make demo with FLAGS="--no-minecraft" and with FLAGS="--mock-calls". Each piece also starts on its own — make dev-dash with a sample journal, make dev-core for the core in isolation, with dev-sim and dev-voice alongside. Missing a key for Jev, Humalike or HappyRobot does not crash the run; that component degrades and records it in the journal, which the troubleshooting section suggests checking when calls do not go out.

Question 4

What does it take to get running, and how do I trust a run?

Accepted Answer

Git clone, cp .env.example .env, make install for uv sync and the dashboard's pnpm install, then make check runs mypy over the contracts plus pytest. Configuration lives in one .env at the root, with every variable commented in .env.example. Trust comes from the append-only JSONL journal — any run replays with make replay RUN=<id> — and from the fact that packages only ever talk through events defined in a shared contracts module.

Question 5

What are the pinned versions, and where is it developed?

Accepted Answer

Python 3.12 exactly, because fenic declares a ceiling below 3.13, plus uv, pnpm and Java for the Paper server. The Minecraft server is Paper 1.21, driven over RCON with /tp, /fill and /setblock. The dashboard is Node with pnpm, Vite 6 and React 19. Development happens on Windows 11 and deployment on Linux as a VPS with Docker Compose. The project is the Grupo Gamma entry for the HappyRobot track at HackSpain 2026 and calls itself Taiafox.

Question 6

What are the four failure modes the README warns about?

Accepted Answer

make demo not connecting to the world, which means Paper is not running: start it with make server and check that RCON_HOST, RCON_PORT and RCON_PASSWORD match server.properties. The dashboard hitting an old gateway, which is an orphan uvicorn holding port 8000, so run make levanta to diagnose it. Calls not going out, which means a missing HappyRobot key or public tunnel, and FLAGS="--mock-calls" continues without them. And the Minecraft client saying outdated server, because MC_VERSION has to match the server jar.

Question 7

How does a turn actually flow through the system?

Accepted Answer

The simulation publishes world.* events onto an asyncio bus that also writes the append-only JSONL journal. The belief layer holds a WorldState, divergence compares it against what the simulation reports, and a divergence flag is what triggers replanning. The planner turns that into a Policy, a deterministic solver produces the Plan, verifiers check it, and action.* events go back to the bus and on to the simulation. The voice layer, HappyRobot with Jev and Humalike, exchanges events with the same bus, which also fronts the FastAPI gateway the React dashboard reads.

HACKSPAIN-2026

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community