Frontier-model-plans, cheap-model-executes delegation loop. The expensive model writes a decision-complete spec, dispatches implementation to a cheap lane (cheaper Claude subagent or Codex CLI), enforces an implementer/reviewer split with fresh-context adversarial review, and accepts work only by verified evidence — spending frontier tokens on judgment, not typing. Use when the user says "architect this", "spec it and delegate", "plan expensive execute cheap", "have the big model plan and a cheap model build", or invokes /architect <task>.
Get an independent, separate-model (OpenAI Codex CLI, flagship GPT tier) review of code — a just-shipped feature, a risky change, or a substrate you're about to build on. Use to complement your own Claude subagent audits with a genuinely different model that has different blind spots. A separate model catches what same-model reviewers miss.
The operating method Fable (Claude Fable 5) left behind — an evidence-first working loop plus the judgment layer for calibration, scope, and reporting. Use at the start of any substantial or ambiguous task, when unsure how to proceed, when work feels like flailing, or when the user says "work like Fable", "fable method", "how would Fable do this", or invokes /fable-method.
Validate every plugin in this marketplace for structural + registration consistency — valid plugin.json, name matches its directory, registered in marketplace.json and the README, SKILL.md frontmatter present, scripts executable and compilable, references over 100 lines carrying a Contents TOC, no pre-plugin-era script paths, and SKILL.md spines under 500 lines. Use when the user says "lint the marketplace", "validate the plugins", "check the skills repo", "are all my plugins registered", or invokes /marketplace-lint. Run it before pushing a new or edited plugin.
Cross-project forgotten-work radar. Sweeps every local git repo for uncommitted, unpushed, stashed, and never-pushed work, cross-references the context packs' open threads, and ranks what is most at risk of being forgotten — with one concrete first move each. Use when the user asks "what am I forgetting", "open threads", "loose ends", "what's at risk", "what did I leave hanging", or invokes /open-threads. (Distinct from /rehydrate, which restores ONE project's context, and /session-recap, which is a timeline of past sessions — this is the radar across everything.)
Address review comments on your own PR via an interactive walkthrough. Fetches unresolved comment threads, walks through them one at a time, lets you choose Fix / Push back / Defer / Skip per comment, dispatches an implementer subagent for each Fix, batches the fixes into one commit, posts replies (with commit SHA for fixes, reasoning for push-backs), and resolves the threads. Use when the user says "respond to PR comments", "address the review", "respond to feedback on PR 42", or invokes /pr-respond.
AI-powered PR review with contract verification, frontend and backend rules, inbox mode, and incremental re-review. Auto-detects project type and applies relevant rules. Use when user says "review PR 42", "review my PRs", "check my PR inbox", "pr review", or invokes /pr-review. This is the standard single-model review lane that posts to GitHub — for a multi-model panel use /review-board; unsure how much review a change deserves, /risk-gate first.
Quality-gated push pipeline. Auto-detects package manager (pnpm/yarn/npm) and runs available scripts (check-types, lint, test, build) in parallel before pushing. Auto-fixes failures in a retry loop, creates a PR, and polls CI. Use when user says "push", "push to remote", "ship it", or invokes /push.