Summary

Adds experiments: fan one prompt across multiple agent/model candidates, each in its own git worktree, then compare diffs side-by-side and merge the winner — with an optional LLM-judge crown that pre-selects the best diff and a one-line rationale (user-overridable).

This is mostly orchestration glue on top of existing primitives (threads, worktrees, diffs, the one-shot prompt runner). An experiment is a thread group (candidates share groupId === experiment.id) plus an experiments registry holding the orchestration metadata.

How to use

1. In a project draft composer, type a prompt and click the 🧪 Run as experiment button.
2. The New Experiment modal seeds the prompt + your current agent/model, lets you add more candidates and pick a fork-from branch.
3. Run experiment creates one worktree + thread per candidate and opens the compact experiment board — status, diff stats (+/−, files), and per-card Open / Crown / Merge winner.
4. Crown with AI runs an LLM judge over the candidate diffs and highlights a winner with a rationale (you can override by crowning any card).
5. Merge winner merges that branch into the base branch and archives the losers (worktrees + branches removed).

What's included

• Contracts / IPC: experiment.ts contract, judgeExperiment IPC procedure, supervisor judge (experimentJudge.ts) built on the existing runOneShotPromptWithFallback, with robust response parsing.
• State: experimentSlice (persisted), experimentLauncherStore, createThreadsBatch, and openExperiment/openThreadStandalone/closeExperiment view actions; new AppView kind "experiment".
• Orchestration (experimentActions.ts): per-candidate worktree creation (sequential — avoids index.lock races), headless startThread launch, mergeExperimentWinner (via gitMergeToSource + performWorktreeRemoval), discardExperiment, crownExperiment.
• UI: ExperimentView board + ExperimentCandidateCard, NewExperimentModal, composer trigger, sidebar experiment-group affordance.
• Plumbing: hydration keeps live candidates attached on reload; project/view selectors and dbStorage handle the new view kind.

Design notes (informed by prior art: cmux, Vibe Kanban, uzi, Crystal)

• Advisory crown, never auto-merge — the judge pre-selects; merge is a separate, confirmed step.
• Anonymized candidate labels sent to the judge ("Solution A/B/C") to avoid provider/self-enhancement bias.
• Diff noise filtering (lockfiles, minified, sourcemaps) before judging.

Testing

• pnpm run typecheck ✅
• pnpm run lint (touched files) ✅
• pnpm exec vitest run for experimentJudge, appStore, dbStorage, threadActions, and the app integration test — 99 passing ✅

Follow-ups (not in this PR)

• Feed test/CI pass-fail signals to the judge and show them as badges.
• Auto-approve permission prompts during fan-out so candidates don't stall.
• i18n: user-facing strings in the new UI still need Lingui wrapping + catalog translations.

🤖 Generated with Claude Code