skills

Non-interactive interface to the other coding backends. Runs one prompt against a chosen backend through the ask-claude, ask-codex, ask-glm, and ask-kimi wrappers and returns its output as stream-json or text. Use when delegating a self-contained build or investigation task to another backend, collecting an independent review of a design, plan, or diff, asking the same question of several backends to compare answers, or resuming an earlier backend session by ID. Triggers on "ask codex", "ask kimi", "ask glm", "ask claude", "get a second opinion from another model", or any request to hand work to a different coding backend without an interactive session.

Iteratively drives the current branch's pull request to a clean Codex review. Reads Codex's PR-body reaction signals (thumbs-up meaning approved, eyes meaning still reviewing, new review comments meaning findings), verifies any findings locally, fixes real bugs with one or more commits, pushes everything in a single push, and posts a comment that re-triggers a full Codex review. After the first pass it schedules a recurring /loop 15m tick that re-runs the same check-fix-trigger cycle, and deletes that cron once Codex signals approval. Use when you want to babysit a PR against Codex feedback until it has no more findings. Triggers on "codex pr loop", "drive PR to green with codex", "iterate codex review", or requests to auto-iterate with Codex until approval.

Read-only third-party overseer for a separate Codex CLI session that is autonomously running /goal. A dedicated Claude Code session audits the target's progress on a recurring self-managed cron tick using read-only Explore subagents, derives the project's drift criteria from the target's own transcript, and -- only through a hardened tmux injection script behind a human-authorization gate -- steers the running Codex. Use when you want to supervise an autonomous Codex /goal run for direction drift, fabricated data, fake/stub implementations, and broken project invariants, with phone alerts. Triggers on "monitor codex goal", "oversee codex /goal", "watch the codex session", "babysit the codex goal run", or any request to supervise or steer an autonomous Codex build loop from a separate session.

Research paper development workflow for academic projects. Orchestrates the iterative loop of paper writing, concept figure design, experiment planning, code development, data collection, and chart generation. Use when entering "Develop for Paper Mode" or when the project goal is producing an academic publication (conference/journal paper). Triggers on "paper mode", "develop for paper", "write paper", or any request to set up a research paper development workflow.

Superman development flow for large-scale feature development, combining superpowers brainstorming/planning with humanize RLCR execution loops and parallel sub-agent execution. Use this skill when starting major features, significant refactors, or complex multi-component implementations. Triggers on "superman flow", "superman mode", or requests for structured development with brainstorming, planning, sub-agent delegation, and code review loops.