KTCrisis

Hoshi7 — AI skill for Claude Code

AI community

A testbed for LLM agents that farm together: one watering can, partners with known policies, beliefs checked against the truth.

How to install Hoshi7

This entry records only its repository, not the path inside it, so there is no exact command to give. Open KTCrisis/hoshi7 and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Hoshi7 does

A testbed for LLM agents that farm together: one watering can, partners with known policies, beliefs checked against the truth.

Alternatives in AI

Doop — The open-source alternative to Paper.design 420 ★
Scipilot Writing Skill — SciPilot family: academic writing and polishing co-pilot, with a machine-checked writing-quality evidence chai 220 ★
Rails AI Context — 45 MCP tools that give AI coding agents ground truth about your Rails app: schema, models, routes, controllers 155 ★

README

hoshi7

A small world where AI agents farm together on a rooftop above a city that never sleeps, one in-game hour at a time. There is one watering can for everyone, so they have to cooperate. Every action is an event in a log that replays alone, every refusal of the world is recorded, and the partners can be scripted bots whose policy is known exactly, so what a model believes about its partner, and about the world's rules, can be checked against the truth.

It is a testbed for LLM agents: can they play a clear goal over a long loop, cooperate when the goal requires it, find the rules of a world by themselves, and does a persona change what they do, or only what they say?

First results

Gate 1, capacity with a fixed partner: six in-game days, two runs per cell, four cells (a bot that waters and never gives the can, a bot that gives it when asked, a bot that farms, the model with itself). Two scripted bots harvest 7.

model	harvests per run (mean)	refused actions
Claude Sonnet 5.5	7.1	3 %
Claude Haiku 4.5	3.3	43 %
Mistral Small 3.2 (local)	0.7	67 %

Sonnet plays at the bots' level; the difficulty is the partner, not the farming. A private note of beliefs, written each hour and heard by no one, separates the models as clearly as the harvest: Sonnet infers the bot's policy from the first hour, including that it cannot hear ("It may not read my messages"); Haiku turns the same bot into "a competitor" and holds rules the world does not have ("crops die if dry for one day"); Mistral does not model its partner and believes the can is needed to till and plant. Details, statistics and their limits: [docs/results.md](docs/results.md).

Run it

uv venv .venv && uv pip install --python .venv/bin/python -e '.[dev,claude]'
.venv/bin/python -m pytest -q
.venv/bin/python -m hoshi7 run worlds/rooftop.yaml --agent ada=scripted --agent hal=scripted
.venv/bin/python -m hoshi7 run worlds/rooftop.yaml --days 6 --stage 6 --w