celEngine 的 classifyError 把「括号/方括号不配对」这类真语法错答成 kind: 'runtime',而该 kind 会原样出现在作者可见的拒写消息与 REST 响应体里

[baozhoutao]

范围外发现,出自 #4812 / PR #6130 的对拍测试(一条 fixture 断言 kind === 'parse' 意外翻红)。按 Prime Directive #10 记录,不指派。

事实(origin/main bc67c28e2 基线,实测)

packages/formula/src/cel-engine.ts 的 classifyError 靠错误文案关键词分类:

function classifyError(err: unknown): EvalResult<never> {
  const message = err instanceof Error ? err.message : String(err);
  let kind: 'parse' | 'type' | 'runtime' | 'bounds' = 'runtime';
  if (/Exceeded max/i.test(message)) kind = 'bounds';
  else if (/parse|unexpected|syntax/i.test(message)) kind = 'parse';
  else if (/type|unknown variable|undeclared/i.test(message)) kind = 'type';
  return { ok: false, error: { kind, message } };
}

而 cel-js 8.0.0 对语法错并不只有一种措辞。实测五条:

源码	cel-js 首行错误	命中分支	得到的 kind
record.budget >	Unexpected token: EOF	unexpected	parse ✅
record.a $$ 1	Unexpected character: $	unexpected	parse ✅
record.a ?? 3	Unexpected token: QUESTION	unexpected	parse ✅
((record.a)	Expected RPAREN, got EOF	无	runtime ❌
[1,2	Expected RBRACKET, got EOF	无	runtime ❌

Expected RPAREN, got EOF 里既没有 parse / unexpected / syntax,也没有 type / unknown variable / undeclared,于是落到默认值 runtime。括号、方括号、花括号不配对是最常见的手写语法错之一,恰好整类都落在这个洞里。

为什么这条是用户可见的(不是纯内部字段)

kind 不止用于内部分支,它被原样拼进作者读到的文案和 API 响应体:

• packages/objectql/src/validation/rule-validator.ts:1262
`Validation rule '${rule.name}' predicate failed to evaluate (${result.error.kind}: ${result.error.message}) — write rejected (#4649)`
• 同文件 :1417,when-predicate 版本同构
• packages/objectql/src/cel-fault.ts:75
return `${error.kind}: ${first || 'unknown error'}`;
• packages/rest/src/rest-server.ts:571
...(error?.kind ? { reason: error.kind } : {}) —— 进 HTTP 错误响应的 reason 字段

所以一条少写了一个右括号的校验规则,作者拿到的是 (runtime: Expected RPAREN, got EOF),REST 消费方拿到的是 reason: "runtime"。message 本身是对的,分类是错的 —— 而分类正是用来告诉作者"这该去哪儿修"的:runtime 指向数据/求值期,parse 指向"你的表达式写错了"。ADR-0032 D1d 要求消息面向自纠,这一格与之相悖。

影响面

classifyError 同时服务 celEngine.compile 与 celEngine.evaluate,所以 build 期(os build / os validate / os lint)与运行期(写入拒绝、REST)两侧都受影响。

未量化 / 未主张

• 没有统计真实 metadata 里不配对分隔符的出现频率。
• 没有排查 cel-js 是否还有别的语法错措辞同样漏网(上表只穷举了我构造的五条);这个洞的成因是"靠文案关键词分类"这个做法本身,补关键词只是补当前已知的洞。

建议方向(供 triage,不代裁)

关键词匹配是脆的:cel-js 换一次措辞就再破一次。可考虑改读 cel-js 抛出的错误对象上的结构化信息(ErrorOptions 带 code,见 lib/index.d.ts:82-86),把分类建立在契约上而不是文案上;若结构化信息不足以区分,则至少把关键词表补全并加一组 fixture 钉住每一类措辞 —— 后者是止血,前者是根治。

关联

#4812 / PR #6130(发现出处;该 PR 的对拍测试里刻意没有把这条错误分类断言进去,并在注释里写明了原因 —— 断言它等于把这个 bug 钉成契约)。ADR-0032 D1d(自纠消息)。

[claude]

Triage: pm:queue + domain:engine-core.

Landing site. packages/formula/src/cel-engine.ts — classifyError verified verbatim on origin/main at :803-809, matching the quoted body character for character (default 'runtime', three keyword branches). packages/formula is on the domain:engine-core row of the SKILL domain table (CEL / matches-filter / RLS predicate evaluation). Routed by landing site, not by the title's "REST response" vocabulary — the REST and rule-validator sites the body lists are consumers that re-emit kind, not the fix site.

The freeze does not apply. The 2026-08-05 maintainer freeze on the driver-memory / driver-mongodb families (anchor #5499) explicitly exempts formula and driver-sql. This is formula, so it queues normally — recording it so a later sweep does not re-derive the question.

Why a defect rather than an observation. Named file, five-row repro table, and a user-visible blast radius the filer did not overstate: kind is string-interpolated into the author-facing rejection text (rule-validator.ts:1262 / :1417, cel-fault.ts:75) and into the HTTP reason field (rest-server.ts:571). A missing closing paren — one of the most common hand-written syntax errors there is — is reported to the author as runtime, i.e. pointed at their data instead of at their expression. ADR-0032 D1d asks these messages to be self-correcting; this one misdirects.

Not target:v17. The message is correct and complete; only the classification is wrong, so the author still sees Expected RPAREN, got EOF and can act. Misdirection, not data loss, silent drop, security, or an unusable surface ⇒ improvement-class for release purposes. Flagging the judgement explicitly so it can be overridden cheaply if the maintainer reads the misclassification as a first-hour-experience item.

Direction — the issue's own framing survives triage, and is worth preserving into dispatch. Topping up the keyword table is stop-the-bleeding; classifying off cel-js's structured error object (ErrorOptions.code, lib/index.d.ts:82-86) is the actual fix, because the root cause is classifying on prose — the table breaks again on the next cel-js wording change. Whoever claims it should establish whether the structured field discriminates all cases before falling back, and pin one fixture per wording class either way.

Stale-premise check. git log origin/main -- packages/formula/src/cel-engine.ts — most recent is 58f3220 (#4812 / PR #6130, the PR this was discovered from). That PR deliberately did not assert the classification, so nothing has pinned the bug as contract.

Dedup (full enumeration; search API unavailable to this seat). No open issue or PR in any of the three repos covers CEL error classification. #6071 (ExecutionContext drift) matched only on the substring "kind" and is unrelated.

本评论来自分诊座位 Routine(#5474 试点),不构成认领。

---

Generated by Claude Code

[baozhoutao]

认领(执行席 PM,engine-core 车道,第 10 轮)

• 会话:session_019Q7oc7ASjh8yxyS3Yz78We
• 分支:claude/issue-6133-classify-parse-errors
• 工作树:../objectstack-issue-6133
• 文件面:packages/formula(classifyError 及其测试)+ changeset;⛔ 不触消费方(rule-validator / cel-fault / rest-server 只作验证读)、不触 formula 内部还剩第三个 CEL 解析入口:cel-to-filter.ts 自建 limitless env,与 celEngine 对「什么能解析」仍不一致 #6132 的 cel-to-filter 面

---

Generated by Claude Code

[baozhoutao]

ACCEPT(执行席 PM 验收,第 10 轮)

PR #6202 验收通过,已翻 ready + auto-merge(CI 23 项全绿)。要点记档:① 两条前提诚实证伪 —— 洞的规模是 20 类 parse code 漏 16 类(非单据列举的 5 条),且派发令里的目标 kind 'syntax' 词表无此格,按词表现实取 'parse'(裁定:采纳,派发令措辞不精确,dev 处置正确);② 实现路线偏离(按错误类判定取代补关键词)由决定性实测背书:cel-js 把作者源码嵌进 message,字段名能决定分类('((record.type_id)' 被判 type)—— 补词表原理上补不对,而错误类是 public 契约(index.d.ts 导出);关键词表逐字保留作非 ParseError 兜底,严格非回归;③ type/runtime 两支刻意不动的理由成立(cel-js 按阶段选错误类,整体结构化会迁移今天正确的判定),反向残留已立 #6223 交分诊;④ 消费面 pin 落在字符串成形处(objectql 只读验证),AGENTS §9 dist 陈旧陷阱的方法学记录有价值;⑤ #6132 定价参考(错误类可共用、缺 limits 的行为变更点)已入档。

---

Generated by Claude Code