Land Phase

You are landing one approved candidate. One candidate, one branch, one PR. Merges here are squash-only, so a two-candidate branch collapses into a squash message that cannot honestly describe either.

Preconditions

Stop and say why if any fails:

• A maintainer has assented to this id — on the ledger, or in this session; record an in-session assent on the ledger before opening the PR. No assent, no PR.
• The id is not tombstoned.
• The candidate is in landable territory (SKILL.md § Territory).
• You are on a fresh branch off current origin/main — cut only after a successful git fetch origin || exit 1 — named simplify/<id>, and no other simplify/* PR is open. Without the fetch, git checkout -b succeeds against the stale cached ref exactly as git worktree add does in references/survey.md § 0, and §1 below then re-certifies the survey's own base as if it were current. Assent commonly arrives days or weeks after the survey, so that drift is the design norm, not an edge case.

1 — Re-verify before touching anything

The evidence was proven against the survey's base, which has since moved. Re-run the proof protocol's steps 2 through 8 (references/survey.md § 3) against the branch base — which is current origin/main only because the precondition above fetched first; against a stale ref every step re-passes vacuously. Run every step the candidate's shape triggers and record the ones you skip and why. Steps 6 through 8 exist exactly for the assent gap: a test that now pins the surface for a user, a commit that deliberately unwired it, or a design document arguing for it can all land between survey and assent, and steps 2 through 5 see none of them. If a consumer now exists, if a design doc now argues for the surface, or if the surface has changed shape, the candidate is stale: record that on the ledger and stop. Do not adapt the deletion to fit new code.

2 — Delete everything that served it

A deletion is incomplete until nothing left in the tree exists only for the thing you removed:

• its test file and any __snapshots__ entry;
• its entry in eslint.legacy-filenames.mjs, if the deleted file had one — otherwise you have created a stale allowlist row while removing another; but first check the stem exempts nothing else ("$RG" --files packages/core/src packages/cli/src -g '**/<name>.ts' -g '**/<name>.*.ts'): each entry expands to both globs, and the rule they serve reaches only packages/core/src and packages/cli/src, so if a surviving file there shares the stem, keep the row and note that on the ledger;
• its locale keys across all 9 files in packages/cli/src/i18n/locales/;
• its rows in docs/** (users and developers alike) when the surface was documented;
• the doc comment attached to it. That is the one comment deletion this skill allows, and only in the same commit as the symbol.

When cutting mechanically, do not trust brace matching to find a function's end: a multi-line return type such as Record<string, { key: string }> opens and closes a brace before the body starts, and a counter will truncate the function there. In this prettier-formatted repo everything nested is indented, so a top-level declaration ends at the next column-0 line that is exactly }, };, });, ];, or ); — a const holding a call or array literal closes with its paren or bracket, not a bare };. Whatever rule you use, prove it after the cut. For a declaration or whole-file deletion, the diff must show zero added lines in git diff --numstat, and the deleted line count must match the declaration's extent as you established it by reading the declaration before cutting: re-measuring with the cut rule itself is circular, and zero added lines alone cannot catch over-deletion. The one exception is survey.md § 1 class 6 — removing only the export keyword from a declaration that stays. Git records export const x → const x as one deletion and one addition (numstat 1 1), so the zero-added-lines invariant wrongly rejects it; there, verify instead that the diff touches exactly that declaration line and the only change on it is the leading export .

Nothing else. No neighbouring cleanup, no rename, no reformat, no "while I was here". The diff must contain exactly one idea.

3 — Verify

Always, in this order:

npm run build && npm run bundle && npm run typecheck # integration tests spawn dist/cli.js
(cd packages/<pkg> && npx vitest run src/path/to/file.test.ts)   # per AGENTS.md; see below
npm run lint:ci

The vitest line runs only against test files that survive the deletion. A clean-kill candidate (survey.md § 6, worked example 1) removes the symbol's only test as part of the deletion, and npx vitest run against a path that no longer exists exits 1 with No test files found — so an unconditional run makes a valid approved deletion impossible to verify. When the deletion carries every test that covered the surface, or the surface had none, skip the line and record "no applicable targeted unit test" in the § 6 details block; the test-corpus re-grep below is what catches any surviving consumer.

Then re-grep the test corpus on this checkout for every removed symbol and run every file it hits — a deletion's characteristic failure is a distant test that imports the symbol, which a targeted run never executes. Branch on the symbol's shape: identifiers keep the word boundaries; sentence-shaped symbols (locale keys) must search fixed-string with no anchors — \b cannot form a boundary after punctuation, and keys like {{name}} are regex parse errors:

# identifier symbols:
"$RG" -n '\b<Symbol>\b' packages integrations integration-tests scripts \
  .github docs-site \
  -g '*.test.*' -g '*.spec.*' -g '**/__snapshots__/**'
# sentence-shaped symbols (locale keys); -e keeps a leading `--` a pattern:
"$RG" -nF -e '<exact key>' packages integrations integration-tests scripts \
  .github docs-site \
  -g '*.test.*' -g '*.spec.*' -g '**/__snapshots__/**'

Then the gate matching what you touched:

4 — Blind spots to plan around

Green CI is weaker evidence than it looks for a deletion:

• npm run typecheck does not run in ci.yml at all. A type-only break from a removed export reaches main unless you run it locally. This is the single most important local command in this phase.
• Test (macos-latest), Test (windows-latest), and the CLI integration job are merge_group-gated: they give no PR signal, so a deletion that breaks a platform-specific path stays green until the merge queue.
• ci.yml classifies each PR into a CI profile; a diff that looks docs-only skips most steps. Deleting docs alongside code can quietly buy you less CI than you think.
• The Prettier CI step (the Run Prettier step in ci.yml → runPrettier() in scripts/lint.js) runs prettier --write ., not --check, so formatting never fails a build — and a formatting-only diff carries zero signal.
• The pre-commit hook runs formatters over staged files. Re-read the diff after committing; if the hook reformatted lines you did not otherwise touch, restore them in a follow-up commit (history is additive here — never amend).

5 — Self-audit

Per AGENTS.md § General workflow: read the full diff you are about to ship presuming it wrong, twice. Write down what each pass raised and how it resolved; a pass that raised something is not a clean pass, and a fix resets the count. For a single-candidate deletion, two clean passes is the whole budget — if the third pass is still finding things, the candidate is bigger than it looked and belongs back on the ledger.

6 — Commit and PR

• Commit subject: plain Conventional Commits — e.g. refactor(cli): remove unused EnumSelector component. This repo squashes with COMMIT_OR_PR_TITLE: a single-commit branch lands the commit subject on main, and any follow-up commit (§4's restore) flips the squash source to the PR title. Keep the id out of both titles so main always looks like every other commit there; ledger correlation rides on the simplify/<id> branch name, the body marker below, and §7's ledger line.
• Body: call /prepare-pr. The repo template verbatim, prose-first; AGENTS.md says explain motivation and changes in prose and avoid naming files and functions there. Do not hard-wrap the body.
• Put the machine detail — id, class, every consumer found and its kind, the proof steps run, lines removed — in ONE collapsed <details> block after the prose, and include the marker line <!-- find-simplifications:id=<id> --> so a later run can find this PR.
• End with the complete collapsed 中文说明 translation, per the repo's PR convention.

7 — After

Append <id> — landed — <date> to the ledger, or hand that line to whoever has the credentials. Then STOP. Do not start the next candidate: the review round on this one is evidence about whether this skill should keep running at all.