// the take
A multi-agent plugin for OpenCode (with a Claude Code MCP option) that turns a research goal into a reproducible Jupyter notebook: one agent plans, one runs Python in a persistent REPL, one adversarially checks the claims, and one writes up the results. It's aimed at data scientists and quant/ML folks who want an LLM to run exploratory analysis without just making numbers up.
The finding-gating rule is the actual idea here: no [FINDING] gets accepted without a confidence interval and effect size, and ML claims need a baseline and cross-validation, which directly targets the common failure mode of an LLM reporting a point estimate as a discovery. The persistent Python REPL with session checkpoint/resume is real engineering, not just shelling out to a fresh interpreter each time. Baksa, the adversarial reviewer agent, is a genuine architectural choice to fight LLM self-agreement instead of just chaining agents that agree with each other. There's a non-trivial TypeScript test suite plus Python bridge tests, which is more than most week-old agent plugins bother with.
Native Windows isn't supported at all — WSL2 only — which cuts out a chunk of the data science crowd that works on Windows with local venvs. The trust-score system checks for the presence of markers like [STAT:ci] and [STAT:effect_size], not whether the numbers behind them are real; nothing stops the executing agent from fabricating a plausible-looking CI to clear the gate, so the rigor is enforced syntactically, not statistically. Autonomous mode (/gyoshu-auto) hands a persistent code-execution REPL to an LLM and lets it run unattended with no stated resource, time, or cost caps, which is a real risk if the goal is underspecified or the model gets stuck in a loop. It's also entirely a layer on top of OpenCode or Claude Code — there's no standalone value — and at 242 stars it's new enough that the multi-agent handoff (professor → TA → reviewer → writer) hasn't been stress-tested on messy real-world data outside the bundled wine/titanic/iris examples.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →