weiyangzen

Build or repair a source-to-target learning cron for code understanding, subset learning, code-to-code transformation, API/schema/runtime/tool transformation, or human-language translation. Use when a repo needs code-to-human learning notes, strict one-to-one source file plus source-tree output coverage, explicit or fuzzy subset learning, code-to-code transformation, documentation translation, route-policy selection, dual-cursor checklists, master-only acceptance, or cleanup after all learning artifacts are validated.

Build or repair a resource-aware, reward-aware, ROI-tracked bridge controller for a DAG-driven repository. Use when blueprint nodes, bridge surfaces, bridge metrics, product validation goals, benchmark lanes, monitoring signals, repeated validation failures, nested b3ehive skill attempts, or operator-controlled feedback loops need bounded attempts, explicit resource leases, side-effect gates, compact evidence ledgers, no-reward pause rules, and re-funded resume under a master acceptance gate.

Run or embed a bounded proposal competition for coding, research, planning, coverage, repair, blueprint, SEO, audit, or execution-choice tasks. Use when a task needs n workers, m proposals, choose k, all-valid coverage union, three-way challenge output under run_a/run_b/run_c, verifier checks, peer review, revision, vote selection, repair synthesis, or execution/looper handoff.

Build or repair repository-agnostic, blueprint-driven execution crons with same-name Gantt Kanban monitoring, isolated agent workers, durable handoff, validation gates, explicit worker lifecycle, request accounting, and cleanup. When Codex is selected, each admitted worker generation uses one independent interactive Codex TUI process in its own task-local tmux server with a private CODEX_HOME and exactly one submitted `/goal`; the repository blueprint chooses bounded terminal executions or a persistent goal pool whose dead workers are replenished to an exact target. Codex app-server, `codex exec`, shared daemons, and no-tmux workers are forbidden. Use for continuous blueprint execution or for repairing execution-controller boundaries, transport, concurrency, replenishment, request storms, integration, monitoring, or cleanup.

Build or repair a design-idea-guided optimization cron for a repository. Use when the user provides a design philosophy and wants a Stage_*_AR_Blueprint.md with at most 100 checklist items, per-item SOTA optimization research docs under Docs/researches/Stage_*_AR/, parallel tmux workers, Codex, Claude Code, Cursor, Grok Build, opencode, OpenClaw, or Hermes agent-runner batches, and cleanup-on-complete.