Cost-efficient model delegation playbook for Claude Code and Codex. Use BEFORE spawning subagents or workflows, when starting research/exploration/multi-step implementation, or when discussing cost or model selection. Routes Claude Fable/Opus/Sonnet/Haiku per call, and Codex GPT-5.6 Sol/Terra/Luna via predefined .codex/agents roles plus per-call overrides. Loading it for delegation work commits the main loop to plan/integrate/review only.
Evidence-grade methodology for analyzing, auditing, or retrospecting a Claude Code or Codex agent session over Junrei's MCP tools, and for closing the self-improvement loop (record a fix, check whether it helped). Use whenever asked to analyze/review/retrospect an agent session, investigate cost/behavior/looping/delegation, propose or record a learning, or produce an evaluation grounded in Junrei data. Encodes the six-tool loop order, the provenance-citation rule, the confidence-class trust order, and truncation handling. Load it before drafting any analysis or logging any learning.
Launch Junrei's isolated development environment and verify its web UI in a browser. Use when changing React UI, CSS/layout, navigation, interactive controls, or API-backed screen behavior in this repository, and whenever a dev server or port must be chosen for browser testing — with whatever browser tooling the current agent environment provides.
Prevent and diagnose "PR created but GitHub Actions never start" situations. Use on EVERY commit/push/PR-creation flow in this repo (including commit-commands:commit-push-pr) — rebase onto origin/main before pushing — and whenever `gh pr checks` shows "no checks reported", checks stay absent after a push, or CI seems stuck without any runs.
How to manage dev tools with aqua (aquaproj) in this repo — adding tools, updating versions, and keeping aqua-checksums.json in sync. Use whenever aqua.yaml or aqua-checksums.json is touched, when the user asks to add/update/pin a CLI tool (node, gh, pnpm, etc.), when tool versions drift between local and CI, or when aqua install fails with a checksum/policy error. Version bumps are NOT complete until checksums are regenerated.