Autonomous unified B″ Worker for the backlog.md board. Drives BOTH lanes from one Monitor session: basic-ready events execute kind:basic tasks in isolated git worktrees that merge back on success; epic-ready events auto-decompose a kind:epic task (Epic: Ready → Decomposing → children at Basic: Backlog → Awaiting Children); child-done events reconcile the parent epic and, when all children are Basic: Done, run evaluate and write a FINISH/ITERATE recommendation for human confirmation. Invoke /loop-backlog once; it runs until the backlog/.loop-stop sentinel is written.
Interactive epic authoring for the B″ epic lane. Drives a brief multi-child epic idea through propose then plan, each gated by human-reviewed cycles, and parks it at Epic: Backlog. The plan enumerates intended child sub-tasks but creates none — the human promotes Epic: Backlog to Epic: Ready to authorize autonomous decomposition by the epic worker.
Drives the Proposal-to-Backlog pipeline for a software feature: two architect-reviewed cycles (one for Proposal, one for Plan), each converging on APPROVED within 8 rounds. Outputs a planSet entry ready for implementation; does not create branches or open PRs.
Schedules a non-coding work item — research inquiry, audit, survey, experiment, or written report — as a backlog entry. Single draft-and-critique cycle produces a phased execution checklist with shell-command gates. No compilation or test runner required.
One-time initializer for the backlog task board. Checks that the backlog CLI is installed, initializes a backlog project if none exists, and verifies that all columns required by loop-backlog and feature-to-backlog are present. Idempotent — safe to run multiple times.
Creates a Ready-status backlog task from a pre-approved template, bypassing the full review cycle. Performs a single LLM freshness check against recent git changes; if FRESH creates the task immediately, if STALE explains why and prompts the user to regenerate via task-to-backlog.
Run a pre-registered quantitative experiment with held-out fixtures, multi-model k=5 execution, statistical verdict, and evidence write-back. Use when starting a new domain experiment that requires hard verdicts (CONFIRMED/NULL/REJECTED) rather than soft self-assessment.
Track and optimize agent specialization during methodology development. Use when agent specialization emerges (generic agents show >5x performance gap), multi-experiment comparison needed, or methodology transferability analysis required. Captures agent set evolution (Aₙ tracking), meta-agent evolution (Mₙ tracking), specialization decisions (when/why to create specialized agents), and reusability assessment (universal vs domain-specific vs task-specific). Enables systematic cross-experiment learning and optimized M₀ evolution. 2-3 hours overhead per experiment.