Skip to content

v1.0.0

Choose a tag to compare

View all tags
@aid-ninja aid-ninja released this 15 Jun 05:55
· 9 commits to main since this release
v1.0.0
acaf488

bean — a recursive convergence loop for large tasks, for Claude Code and Codex.

bean runs a task the way the Fable model worked: investigate the most decisive open front, record what you learn as typed evidence in a claim ledger, let a compiler score what's still weak or contradictory, revise the beliefs that don't hold, and loop until it converges — then deliver. It shapes the procedure a model follows, not its capability ceiling.

Highlights

  • Convergence loop, not a fixed checklist — Frame → Survey → Investigate → Record → Compile → Revise → Converged?, repeating until there's nothing decisive left to learn.
  • Runs on a runtime — a claim ledger + a compiler that scores convergence. grainulator/wheat is the primary runtime; a minimal built-in ledger (bean-stalk.md) is the fallback for Codex and bare installs.
  • Belief revision is first-class — including respectfully revising the user's stated beliefs when grounded evidence contradicts them (counter-sycophancy).
  • Ground before you assert; verify by running or rendering — read the source / run the test / render the artifact; "looks right" is not a check.
  • Persistence with proportion — drive every open front to a confirmed non-issue, a verified fix, or a true residual; but don't over-engineer trivial tasks or refuse to ask the human when asking is cheaper.
  • Cross-model Codex blindspot lane for independent verification on high-stakes work.
  • Dual packaging — Claude Code (.claude-plugin/) and Codex (.codex-plugin/), plus install.sh. /bean in both.

Install

claude plugin marketplace add https://github.com/grainulation/bean.git
claude plugin install bean

Doc-only (Markdown + JSON), no runtime dependencies. MIT. Verified by a structural smoke test (manifests, frontmatter, link resolution) and CI on Node 20 + 22.

Assets 2
Loading