Fable is gone but the loop isn't

A new open-source tool called Bean replaces fixed-step task execution with a recursive convergence loop for large tasks in Claude Code and Codex. It uses a claim ledger and compiler to investigate, record evidence, revise beliefs, and loop until the answer converges, then delivers the result.

A recursive convergence loop for large tasks. A skill for Claude Code and Codex that runs a task the way the Fable model worked: investigate, record what you learn as evidence, let a compiler tell you what's still weak or contradictory, revise the beliefs that don't hold, and loop until the answer converges — then deliver. bean replaces "do these N phases" with "keep going until there's nothing decisive left to learn." It runs on a runtime — a claim ledger plus a compiler that scores convergence — and uses that signal to decide its own next move. Investigate the most decisive open front — the compiler's signal unresolved conflict, weakest-evidence claim, coverage gap picks the target, not a fixed plan. Record evidence in a ledger — typed claims at honest evidence tiers stated → production . The ledger is bean's memory across rounds — its "own notes." Compile — a check that can fail at the whole-task level: conflicts, gaps, single-source topics, weak evidence, and undischarged risks. Revise beliefs — when new evidence overturns an earlier claim, supersede it "kill the incorrect belief" rather than letting the contradiction stand. Loop until converged — stop when there are no unresolved conflicts, the evidence bar is met, and a full round adds nothing new. High-stakes work gets an independent cross-model Codex blindspot check before declaring done. Continuous orchestration — every round re-surveys the available skills/connectors and spins up subagents for whatever the compiler flagged next. Verbose by default — each round surfaces the ledger, the compiler signal, the next move, and any belief it revised, so you can watch the answer converge. Step 1 — Add the marketplace one-time : claude plugin marketplace add https://github.com/grainulation/bean.git Step 2 — Install: claude plugin install bean Inside Claude Code, use /plugin instead of claude plugin . Requirements: Claude Code with Node.js = 20. Pinning & updates.bean installs into your agent's context, so treat it like any third-party plugin: prefer a tagged release over a moving branch pin a --ref , e.g. --ref v1.1.2 and re-review on update. bean runs locally with no network surface by default; the only optional outbound path is the grainulator/wheat remote dashboard, which stays off unless you wire it in. Codex codex plugin marketplace add grainulation/bean --ref main codex plugin add bean@bean On the Codex side bean is invoked explicitly as /bean allow implicit invocation: false . Alternative: clone directly git clone https://github.com/grainulation/bean.git cd bean && ./install.sh /bean <your task /bean quiet <your task run every step, report tersely failures still surfaced bean also triggers on phrases like "do this thoroughly", "be systematic", or "deep work mode", and when a task objectively spans multiple files, sources, or sessions. For a trivial one-pass task it stays out of the way. See EXAMPLE.md /grainulation/bean/blob/main/EXAMPLE.md for a worked before/after. Frame once, then iterate until the compiler signal goes green: | Step | What happens | |---|---| Frame once | State the goal; seed the ledger with known constraints. Does the task earn the loop? | 1. Survey | Re-assess available skills/subagents/connectors this round; read real data. | 2. Investigate | Attack the most decisive open front the compiler flagged; fan out subagents. | 3. Record | Write findings as typed claims at honest evidence tiers into the ledger. | 4. Compile | Score convergence: conflicts, gaps, single-source, weak evidence, undischarged risks. | 5. Revise beliefs | Supersede claims that new evidence overturns; resolve conflicts. | 6. Converged? | No unresolved conflicts + evidence bar met + a dry round → deliver. Else loop to 1. | bean runs on a claim ledger + a compiler. It ships its own compiler — bean-check — a zero-dependency Node CLI built from the Bran https://github.com/grainulation/grainulator core that scores convergence and exits nonzero until it is reached conflicts, undischarged risks, load-bearing claims below the evidence bar, dry-round, budget . That is the default control plane and runs anywhere Node does. is an optional richer backend; where neither can run, bean falls back to a hand-checked grainulator/wheat https://github.com/grainulation/grainulator bean-stalk.md ledger. With grainulator connected, bean also taps the wider stack, degrading gracefully when a piece is absent:— the ledger + compiler; record claims and gate convergence on wheat https://github.com/grainulation/grainulator wheat compile . grainulator — dispatch autoresearch subagents for investigation rounds. deepwiki — corroborate external claims against real repositories. silo + connectors — the canonical read surface for grounding claims in real data. farmer — route the Codex blindspot review through the dashboard. Details: skills/bean/references/grainulation.md /grainulation/bean/blob/main/skills/bean/references/grainulation.md . Lean. Markdown, JSON, and one zero-dependency CLI bean-check . No servers, no network, no telemetry, no runtime dependencies TypeScript is a CI-only devDependency . A real gate. bean-check makes convergence falsifiable — it exits nonzero until the loop honestly converges; grainulator/wheat is an optional richer backend. Verbose by default. Show the ledger, the compiler signal, and the belief revisions — watch the answer converge. See CONTRIBUTING.md /grainulation/bean/blob/main/CONTRIBUTING.md . Run npm run format and npm test before a PR. MIT /grainulation/bean/blob/main/LICENSE © grainulation contributors.