v3/docs/releases/v3.32.31.md
Ruflo's self-improvement loop should explore useful candidates without relaxing its promotion guardrails, and Autopilot should never search a broader task surface than the operator requested. v3.32.31 fixes both contracts.
The Flywheel safety envelope now describes the complete retrieval policy that the built-in proposer actually emits. Valid candidates can reach evaluation again, while every tunable value remains bounded and promotion remains subject to signed receipts, policy authorization, and the atomic transaction gate.
Autopilot now treats explicitly configured task sources strictly. Misspelled, empty, or unsupported sources fail with a useful error instead of silently falling back to every default source. Recovery from absent or damaged legacy state remains backward compatible.
alpha, subjectWeight, mmrLambda, bodyWeight, and
typePenaltyFactor inside explicit finite bounds.npm install --global [email protected]
ruflo doctor
Or run without a global installation:
npx [email protected] --version
The supported sources are team-tasks, swarm-tasks, and file-checklist.
ruflo autopilot config \
--task-sources swarm-tasks,file-checklist
Unsupported input now exits unsuccessfully and leaves the stored configuration unchanged:
ruflo autopilot config --task-sources issues
Inspect the current transaction and ledger state:
ruflo metaharness flywheel status
Run one governed evaluation using the built-in proposer:
ruflo metaharness flywheel run --proposer local
Evaluation still does not grant promotion authority. Promotion requires an accepted receipt, trusted signing key, current baseline, matching ledger head, policy authorization, and explicit confirmation.
@claude-flow/cli, claude-flow, and ruflo.Release acceptance requires:
latest, alpha, and v3alpha npm tags.