[osolmaz]

Summary

switch asked the model about a state in two ways: a pick at each step, and the final questions after the steps.
Now every step is one set of typed questions, and the domain's transition function gets the answers.
A pick is one choice question that a helper builds, so the final questions, the answer stages and run --state-from are gone.
The plan, with every decision, is in docs/2026-10-29-questions-as-steps.md.

What Changed

The core loop has one path now. A domain gives initial, questions(state) and transition(state, answers), and the run ends when questions gives nothing.

• Core: new StateMachine type; picking, pickQuestion and pickedChoice in src/core/pick.ts (was choices.ts). switch adds no stop choice and no longer refuses a repeated pick. The limits are windowShare and an optional steps. A run ends as done, step limit or failed. The view function loses its purpose argument. Large question sets are split into batches at every step.
• Trace and replay: the entries are start, state, questions, answers, ended, note and thrown. A replay checks that each recorded answer fits the questions asked now (ids, types, and for a pick, the choice's key). A fork repeats the first n steps.
• CLI: no --state-from; result.json loses answers, run.json loses stateFrom.
• Viewer: frames show each step's questions and answers, with the pick's choices and probabilities.
• Chess: "Resign the game" is a choice; a resignation is a status of the state.
• WikiRace: "Give up and lose the race" is a choice; giving up is a field of the state.
• rubric-judge: gathers evidence by picks, then asks every item's meets and misses questions in one step. The outcome holds each item's verdict, which score and compare read.
• Docs: README "How it works" and its domain example, docs/MATH.md, AGENTS.md and the example READMEs.

Testing

All tests use the fake decider; no model was called.

• pnpm check at the root passes: format, lint, typecheck, tests with coverage, DRY, the viewer and the three examples, and slophammer.
• Tests: root 71, viewer 37, chess 75, rubric-judge 48, WikiRace 19.
• The README's domain example was typechecked and run with the fake decider outside the repository.

Risks

This is a breaking change for every domain and for old run folders.

• Old traces and run folders are not read.
• Re-judging a whole set with another judge file needs one replay --fork per case; there is no set command for it yet.
• The judge files' texts did not change, so their hashes stay the same, but the judge's version is now 0.6.0.

Follow-ups

• A reviewer that did not write the change looks for leftovers of the old interface after the merge.