Repository navigation
npm run format mass-reformats ~125 files: no Prettier config, unpinned major range resolves to 3.9.6 #445
Description
来源: #442 (#442) 分析时发现（PR #443 实现过程中触发）
Repro: on pristine master, npm ci && npx prettier --check . fails on ~125 files. Running npm run format rewrites package.json (4→2 space indent), regenerates package-lock.json, and wraps/reindents virtually every src/ and tests/ file.
Cause:
- No Prettier config file exists in the repo, so Prettier runs on defaults (printWidth 80) that do not match the committed style.
- devDependency is
prettier ^3.3.0, which now resolves to 3.9.6 — the committed style predates the current major's defaults. - CI (.github/workflows/ci.yml runs: npm ci, typecheck, test, build, check-pr.sh) does not run format:check, so the drift is invisible to CI and only explodes locally.
Impact: AGENTS.md lists npm run format / npm run format:check as standard build commands, but running them today produces a 125-file churn diff that buries any real change and defeats the PR diff-cleanliness gate (§7.5). Any agent following the documented commands hits this (hit during PR #443 implementation; the mass reformat was reverted and formatting was skipped instead).
Suggested fix (owner decision): either (a) add a .prettierrc matching the committed style AND pin prettier to an exact tested version, then re-run format once deliberately; or (b) remove format/format:check from AGENTS.md build commands until a config exists. Option (a) restores the documented workflow; option (b) is lower-risk but leaves the command broken-by-omission.
Activity
ranxianglei commentedon Sep 26, 2026
[bot] 🏷 Reproduced + root-caused — recommending option (a); your call before I push
Repro confirmed on pristine master (
476d440):npm ciresolvesprettier ^3.3.0→ 3.9.6;npx prettier --check .fails on 125 files. No.prettierrc,.editorconfig, or.prettierignoreexists in the repo.Key finding — the committed style was never prettier-formatted. I tested printWidth 80/100/120/140 against Prettier defaults: no config makes pristine master pass (best case = 125 files at width 80). The code mixes three habits: long single-line signatures up to ~120 cols (e.g.
src/nudge-text.ts:19), manually-wrapped ~80-col blocks, and hand-packed multi-field object lines (src/nudge-text.ts:83) that Prettier would never produce. So "match the committed style" can only be approximate — option (a)'s deliberate one-time reformat is unavoidable under any width. (Sibling check: opencode-acp has a.prettierrcbut its no-semicolons style contradicts acp-kernel's committed semicolon style — not transferable. Everything else in the committed style already equals Prettier v3 defaults: semi, double quotes, 2-space, trailing commas → printWidth is the only real knob.)Measured churn of the one-time run:
Recommendation: option (a), concretely:
.prettierrc={ "printWidth": 120 }— matches how new code is actually written (recent PRs use long single-line signatures); fewer total changed lines than 80."prettier": "^3.3.0"→"3.9.6"exact (= current lockfile resolution = the version tested here). Kills silent future re-drift..prettierignore:dist/,coverage/,tmp/,.tmp/,package-lock.json(npm-owned — Prettier would rewrite it),*.tgz,.acp-*.json. Without this,formatwould mangle build output and the lockfile.Verified on the formatted tree:
prettier --checkclean · typecheck clean · 1122/1122 tests pass · build OK. Branch is fully staged locally (2 commits: config+pin, then the labeled reformat) — push + PR happens the moment you say go.One extra decision for you: CI never runs⚠️ Behavior change: unformatted PRs will start failing CI (that's the point). Say if you want it out of scope.
format:check— that's why this drift stayed invisible for months. I recommend adding it to theci.ymltest job as a third labeled commit.Decision requested: (a) as above [my recommendation] or (b) strip the commands from AGENTS.md? And yes/no on the CI gate?
中文摘要：已复现并定位根因——仓库从未用 Prettier 格式化过(任何宽度都无法让 master 通过 check,提交风格本身混用长签名与手工换行)。建议方案 (a):加
.prettierrc(printWidth 120)+ 精确锁定 prettier 3.9.6 + 新增 .prettierignore(排除 dist/lockfile 等生成物),一次性重排 161 文件(纯空白、零语义变更,已验证 format:check/typecheck/1122 项测试/build 全绿);另建议在 CI 增加 format:check 防止再次静默漂移(行为变更:未格式化 PR 将失败 CI)。请拍板 (a)/(b) 及是否加 CI 门禁。ranxianglei commentedon Sep 26, 2026
[bot] 🏷 Correction: this is a duplicate of PR #438 — closing here, relaying one open item there
Duplicate screening (per repo policy, before implementing) turned up prior work covering this issue's full scope, opened ~9h before this issue was filed:
PR #438 —
chore: prettier baseline — config, CI format gate, tree reformat(Closes #435), branch2026-09-26_prettier-baseline@9654f8b, based directly on current master476d440:.prettierrc: explicit pin of Prettier defaults (printWidth 80 + all other defaults written out, so a future major's default change can't silently re-drift the tree).prettierignore:dist/,node_modules/,.tmp/,package-lock.jsonci.yml: new standaloneformatjob runningnpm run format:checkon PRs + pushes to masterRelated threads for completeness: #361 (same scope, older base, superseded by #438 — closure candidate) and #418 (fixes the prettier two-pass non-idempotence quirk in
src/hide-consumed.tsthat #438 documents but leaves in place — recommended small follow-up after #438 lands).Why this issue still got filed: during PR #443 implementation (on master, where #438 was still unmerged) the working agent hit the broken
npm run format, reverted the churn, and filed this without linking #438. Root cause is identical — missing Prettier config (+ caret range) → mass reformat — so per the one-issue-per-root-cause policy I'm closing #445 as a duplicate rather than starting parallel work. My own staged width-120 variant stays local, unpushed.Independent measurements from my analysis (for the record, feeding #438's open choices):
80 touches fewer files (its stated rationale holds); 120 changes fewer total lines and keeps long single-line signatures as-written. Also confirmed: no prettier config can make pristine master pass — the committed style was never prettier-formatted (it mixes long single-line signatures, manual ~80-col wraps, and hand-packed object lines), so the one-time reformat is unavoidable under option (a) at any width.
Open item relayed to #438: this issue's cause section independently supports the optional exact pin from #438's PR body —
"prettier": "3.9.6"instead of^3.3.0. The caret range is what let 3.9.6 float in silently; lockfile +npm cimake resolution deterministic today, but the pin costs one line and removes the last drift vector. Comment posted in #438's thread.Closing #445 as duplicate of #438 (labeling
重复). If you'd rather keep the (a)-vs-(b) question open as a fresh decision independent of #438, say so and I'll reopen — but note the substantive choice is exactly #438's merge gate.中文摘要：查重发现本 issue 与已开放且 CI 全绿、独立评审可合并的 PR #438 同根因(缺 Prettier 配置+caret 版本漂移→批量重排;#438 已实现方案 (a):显式 .prettierrc(printWidth 80)+ .prettierignore + CI format 门禁 + 一次性全树重排 124 文件)。按「一根因一 issue」原则以重复关闭本 issue 并转链 #438/#361/#418;本 issue 独立支持精确锁定 prettier 3.9.6(一行改动,#438 正文已提议),该建议已转达至 #438 线程。如需保留 (a)/(b) 作为独立决策可说明后重开。