Question 1

Is this an implementation I can run, or a plan on paper?

Accepted Answer

A plan. The document describes a three-layer architecture and a set of goals for the bot — explore rather than follow a fixed route, craft, fight or dodge mobs, build a house, and end at the dragon without being handed a step-by-step path — but the author's own framing is exploratory: they want to understand how Jev works and how to drive it. Treat it as a design sketch with named layers rather than something with a build script or a demo.

Question 2

Why not just let the LLM drive the bot directly?

Accepted Answer

The note separates the roles deliberately. Deterministic code observes the world, memory, inventory, crafting and player physics, so anything mathematical stays out of the model. The LLM acts as a strategist, called rarely on algorithmic triggers at one or two calls per minute and given a summary of the game, while Jev handles reflexes each tick by choosing among primitives. The LLM is explicitly forbidden from authoring Jev's questions: the question sets are fixed by the system and the LLM may only fill parameters drawn from closed value sets.

Question 3

How does the bot remember what it has already tried?

Accepted Answer

Through a structured journal rather than model memory, because the note states plainly that neither Jev nor the LLM remembers anything between two calls. The journal keeps coordinates for the base, village and caves, plus what was attempted and failed; the strategist reads it and Jev receives only a relevant extract. Because a strategist decision can arrive 5 to 30 seconds later, the code is supposed to re-check on arrival that the decision still makes sense.

Question 4

What does Jev actually decide, and what context does it get?

Accepted Answer

Jev picks among primitives — the action layer below strategy — and is handed the current intent without needing to know the end goal. The context it would receive is server-sent rather than screen-scraped: the blocks it can see, sound with the exact origin and coordinates of what emitted it, coordinates of visible mobs, craft patterns, inventory item positions and identities, block state such as furnace progress, and game events. The intent is that the bot is not given information a human player would not have.

Question 5

Why is the whole thing in French?

Accepted Answer

The README is written in French and the repository keeps it that way, so the design rationale is only available to French readers. Nothing about the architecture depends on the language, but anyone reading the note itself needs a translation.

Question 6

What is still open in the design?

Accepted Answer

Several things, and the author says so rather than glossing over them. The LLM to use is still undecided between two candidates. Character control assumes a mod server doing genuine pathfinding and authentic player behaviour, borrowed from Twitch Play and Minecraft Studio rather than implemented here. And the layered pyramid of Jev instances is only a thought about where the project could go next.

Question 7

Is anything in the repository actually implemented?

Accepted Answer

No layer described is shown running. The document is intent, open questions, and a to-do list, and the author states plainly that they are still learning how Jev works and do not know whether it has any global knowledge of Minecraft or its mechanics. Read it as a layering proposal with named responsibilities.

Jev.mods

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community