It could still be a harness or prompting issue, which is part of what I’m testing.
Homeby Jayson2026-09-17
It could still be a harness or prompting issue, which is part of what I’m testing.
Each model chooses one relative action: `forward`, `left`, or `right`.
Jev receives a TypeSafe Choice:
```json
{
"type": "choice",
"criteria": {
"forward": "Continue in the current heading.",
"left": "Turn left.",
"right": "Turn right."
}
}
```
Luna receives the same options through a strict JSON enum:
```json
{
"action": {
"type": "string",
"enum": ["forward", "left", "right"]
}
}
```
Both receive the same prompt and game state. A shortened state example:
```json
{
"board": {
"width": 10,
"height": 10,
"map": ["..........", "...TBH.A..", ".........."]
},
"heading": "right",
"apple": {"x": 7, "y": 1},
"body_head_to_tail": [
{"role": "head", "x": 5, "y": 1},
{"role": "body", "x": 4, "y": 1},
{"role": "tail", "x": 3, "y": 1}
],
"moves": {
"forward": {
"next": {"x": 6, "y": 1},
"collision": false,
"apple_distance": 1,
"safe_cells": 98
},
"left": {
"next": {"x": 5, "y": 0},
"collision"