regent-refactor

---

name: regent-refactor description: 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. version: 0.2.0 author: REgent contributors license: GPL-3.0-or-later metadata: hermes: tags: [regent, refactor, subagent, reverse-roundtrip, test-oracle] related_skills: [regent-reverse, regent-build, plan]

regent-refactor

Apply a modified REgent spec to an existing codebase (or to a fresh checkout). The user has edited specs/<module>.spec.md (and/or inventory/functional-checklist.md, inventory/test-oracle.md, conventions/*.md) to change intent. This skill rewrites the source tree so the new spec is satisfied, and proves it by running the graded test suite (black-box checklist + white-box oracle).

Refactor ≠ rebuild. Refactor preserves intent that the user kept (unchanged spec sections) and only mutates code where the spec demands a new behaviour. Rebuild would throw the whole project away. Read the diff first, mutate surgically.

When to Use

• A spec tree exists and has been edited since the last reverse.
• The user wants the codebase aligned to the edited spec without a full ground-up rebuild.
• A "shit-mountain" project needs to be brought under a spec — reverse it (regent-reverse), then refactor it (this skill) one spec change at a time.

Don't use for:

• First-time spec generation — that is regent-reverse.
• Verifying an unmodified spec — that is regent-build.
• Editing code without a spec to point at — this skill assumes the intent is captured somewhere in <spec_dir>.

Inputs

• spec_dir (REQUIRED) — path to the edited spec tree.
• source_dir (REQUIRED) — path to the existing codebase that reflects the pre-edit spec. The agent will mutate this in place (or write to a sibling out_dir, see strategy).
• strategy — in-place (default, mutates source_dir after snapshot) | side-by-side (writes to out_dir for diffing). Default is in-place because refactor scenarios usually want the project on disk to stay current.
• out_dir — only used with strategy: side-by-side. Default is a sibling of source_dir named <source_dir>-refactored.

Workflow (MUST follow in order)

1. Diff the spec

Before touching code, identify what changed. The user's edit is authoritative for intent.

• For each tracked file under <spec_dir> (especially specs/*.spec.md, inventory/functional-checklist.md, inventory/test-oracle.md, conventions/*.md):
  • If the file has a ## Edited: or ## Changelog: heading written by the user, use it.
  • Else, do a content diff vs the project history if a .git is available under <spec_dir> (REgent specs may be versioned).
  • Else, ask the user what changed. Do not guess.
• Record the diff in a working note (in-memory; do not commit a scratch file unless the user asks).
• Pure-clarification edits are legitimate. If the diff only adds precision to an existing requirement without changing inputs, outputs, error paths, or exit codes, classify as Pure clarification per §2 and skip code edits entirely. The downstream symbol list (§2) can be empty; that is a passing outcome, not a skipped one.

2. Classify the change

For each R- / S- / oracle entry the user added, removed, or edited:

• Behaviour change (R-/S- edit that alters inputs, outputs, or error paths) → rewrite affected module(s). Example: greet now raises on name == "Bot" in addition to empty.
• Conventions change (edit in conventions/*.md) → adjust style, imports, error types. Example: use httpx not requests.
• Inventory change (checklist or oracle count shift) → may need new behaviour, but the spec module sections are the source of truth for what that behaviour is. If a new oracle entry has no matching R-, the spec is incomplete — flag it back to the user before rewriting code.
• Pure clarification (more precise wording, no semantic shift) → no code change needed.

For each affected module, list the specific symbols that must change, with file:line targets.

3. Snapshot before mutating

If strategy: in-place:

• Run git status in source_dir. If dirty, stop and ask the user to commit or stash.
• If clean, snapshot via `git worktree add ../<source_dir>-snapshot ~0` OR a tarball. The snapshot is the rollback point if this skill corrupts the tree. **Mandatory: snapshot MUST succeed before the rewriter is spawned.**

If strategy: side-by-side: copy source_dir to out_dir. The original source_dir itself is the rollback; no separate snapshot needed.

4. Spawn the rewriter subagent

Use delegate_task with these properties:

• role: leaf
• Goal must contain: spec_dir, the source dir (or out_dir), the diff of the spec, the list of symbols that must change from step 2, the test command from <spec_dir>/conventions/dev-env.md.
• The rewriter's job is to edit only the symbols listed in step 2. It is forbidden to refactor unrelated code, even if the code looks ugly — that is style, not behaviour, and would defeat the round-trip audit.
• The rewriter must NOT regenerate code from scratch (unlike regent-build). It must read_file → patch → re-run pytest.
• The rewriter's final report must include a diff (before vs after for each changed symbol) so the user can spot accidental over-rewriting.

5. Run verification — both grading keys

After the rewriter reports:

• Run the test command from conventions/dev-env.md. Capture exit code. (This is the suite the project already has — same as a manual developer would run.)
• Execute every line of functional-checklist.md against the result (re-run each command / import / CLI).
• Execute every ### <symbol> block in test-oracle.md against the result (call the function with the documented input, check the documented output).
• ALL THREE must be green: pytest, checklist, oracle.
• Clarification-derived oracle: if the spec edit added a new R-/S- clause, the rewriter MUST also add an ### <symbol> block to test-oracle.md (or a scratch verification command) that exercises the clarification explicitly, even when no code change was needed. Without this, the clarification is a documented intent that no test actually pins.

6. Decide

• All three green AND git diff --stat source_dir out_dir is empty: this is the clarification-only outcome. Report "Clarification-only refactor: zero source changes, spec edit is documented intent only." Treat this as a successful outcome, not a failure to edit. Show the new oracle entry as the proof that the clarification is exercised.
• All three green AND the diff is non-empty: state the changed symbols (file:line diff summary) and confirm the spec edit is now in the code.
• Anything red → do not auto-rollback. Report which item, the actual vs expected, and the offending code path. The user decides whether to: (a) tweak the rewriter's diff, (b) amend the spec to match the code reality, or (c) roll back the snapshot.
• Rewriter touched code outside its scope → flag the diff. This is a regression of trust — the rewriter overstepped its brief. Roll back those edits, retry with a sharper scope.

7. Self-review

Before declaring done:

• Step 1 diff is recorded (in memory or scratch file).
• Step 2 symbol-level change list is recorded.
• Source tree is in a git checkpoint if in-place was used.
• All three grading keys (pytest + checklist + oracle) green.
• Rewriter's diff matches the symbol list. Nothing else.
• User informed of changed symbols with file:line.

Common Pitfalls

1. Reflavoring everything. The rewriter's mandate is the diff in step 1. Modules untouched by the diff MUST remain byte-identical (modulo whitespace) — otherwise the audit breaks. Grep the diff before sign-off.
2. Reading the original repo for inspiration. Refactor is allowed to read source_dir (it is the codebase being edited). It is NOT allowed to look up training-data versions of similar projects to fill in — that smuggles in intent not in the spec.
3. Trusting the rewriter's "I changed it" report without diffing. Always run git diff --stat against the snapshot. The agent can claim it changed 5 lines and actually rewrite 500.
4. Adding tests not in the spec. The oracle is the only white-box grading key. New behaviour the user did not write an oracle entry for is not graded — adding a "while I'm here" test is a different skill.
5. Skipping the snapshot. A bad refactor on a dirty tree corrupts the user's history. Snapshot or refuse.

Verification Checklist

• Spec diff is identified by the user, not inferred.
• Source tree is snapshotted (in-place) or copied (side-by-side).
• Symbol-level change list matches the spec diff exactly.
• The rewriter did not touch code outside that list.
• pytest exit code green.
• Every - [ ] line of functional-checklist.md verified.
• Every ### <symbol> of test-oracle.md executed.
• User received the file:line diff summary.

One-Shot Recipe

# After regent-reverse produced a spec, the user has edited it.
# Now invoke regent-refactor with:
#   spec_dir   = ./spec-out/<repo>/spec
#   source_dir = /path/to/the/checked-out-repo
#   strategy   = in-place
# Expect a pass/fail report AND a diff summary that the user can
# eyeball against their spec edit.