REgent

Use when the user asks to refactor an existing project against a modified spec, regenerate code from a changed spec, or refactor a "shit-mountain" codebase (屎山重构) back to documented intent. Triggers on phrases like "refactor to spec", "按新 spec 改", "regenerate from <spec_dir> with new requirements", or "重写这个屎山". Pairs with regent-reverse + regent-build — reverse produces a spec, the user edits specs/ to change intent, this skill rewrites the codebase to match the edited spec. Grade key is the same as regent-build: both functional-checklist.md and inventory/test-oracle.md must stay green.

Use when the user asks to rebuild a project from a REgent spec tree, run "spec → code + tests" roundtrip, or verify that a reverse-engineered spec is sufficient to reconstruct the original. Triggers on phrases like "按 spec 重建", "build from spec", "reconstruct from <spec_dir>", "verify spec sufficiency", or just "build". Pairs with `regent-reverse` — reverse produces the spec, this skill consumes it and produces runnable code graded against both `functional-checklist.md` (black-box) and `inventory/test-oracle.md` (white-box, per-function invariants). Works across languages via hermes subagent delegation.

Use when the user asks to reverse-engineer a Git repository into a structured natural-language spec, write a structured spec for any codebase, or generate "repo → spec/" output. Triggers on phrases like "还原这个仓库", "把仓库转成 spec", "reverse engineer <url>", "生成结构化说明", or when the user supplies a repo URL/path and asks REgent to describe it. Drives the agent to deeply explore the entire codebase (every file, every test, every config) before writing the spec tree. Works across Python, Rust, Go, and other languages — the rules are language-aware only where they have to be (e.g. `trait` vs `Protocol` vs interface).