Trimming Chain-of-Thought Leakage

清理思维链泄露内容

Chain-of-thought leakage is prose whose vantage is the authoring session rather than the repository: it cites artifacts only that session could see, narrates the change instead of the state, or argues with a reviewer who has left. The fix is never deletion alone when a passage carries factual clauses — restate each so it stands on its own, then delete the transcript around it; a passage carrying none is deleted outright. Before editing, enumerate the complete proposition: preserve each actor, action, condition, timing, modality, negative guarantee, exception, ownership, side effect, failure mode, and consequence. It is guidance, not a script.

思维链泄露内容是指视角局限于创作会话而非代码库的文本：它引用仅该会话可见的工件、叙述变更过程而非当前状态，或是提及已离开的评审人员的意见。如果段落包含事实性内容，修复方案绝不仅仅是删除——需将每个事实性内容重新表述为独立可理解的内容，再删除周围的会话记录；若段落无任何事实性内容，则直接删除。**编辑前，请列出完整的命题：**保留每个参与者、动作、条件、时间、模态、否定保证、例外情况、所有权、副作用、故障模式和后果。这是指导原则，而非固定脚本。

The one test

唯一测试标准

For every suspect passage ask: could a reader with no access to any session transcript, PR thread, or uncommitted draft resolve every reference and verify every claim? If no, restate the surviving facts and delete the rest. If yes, it is not leakage — but a resolvable change story on a current-state surface (README, docs, skill body) is still change narration and belongs in a commit, PR, or Agent Note.

对于每一段可疑文本，问自己：**没有访问任何会话记录、PR线程或未提交草稿的读者，能否解析所有引用并验证所有声明？**如果不能，重新表述留存的事实并删除其余内容；如果能，则不属于泄露内容——但在当前状态载体（README、文档、Skill主体）上的可解析变更说明仍属于变更叙述，应放在提交记录、PR或Agent Notes中。

Taxonomy

分类

Dead design-session citations —
```
(decision 7)
```
,
```
(audit C2)
```
,
```
design §4.7
```
, phase labels (
```
T4
```
,
```
W3
```
). If the decision has a committed owner, cite it by name and path; otherwise delete the citation and restate its factual clause.
Stack and PR vantage — "a later PR in this stack", "this PR adds", "the previous commit". State the shipped mechanism; deferred work becomes a
```
TODO
```
marker or an issue reference.
Change narration and version stamps — "used to", "no longer", "the old X", and indexical stamps ("v1", "this cut", "today"). State the present behavior; a fixed regression becomes a present-tense counterfactual ("without X, Y happens"), never repo history.
Review choreography — "rejected in review", "the reviewer confirmed", draft ordinals ("v5 of this note"). Keep the surviving decision and rationale as plain fact.
Reviewer-addressed justification — "the cast is safe — it simply…", "this is correct because…". State the invariant that makes the code or rule safe, or delete the comment.
Restatement and derivation transcripts — control-flow narration ("first we X, then we Y"), proofs of obvious branches. Delete; keep only a non-obvious contract.
Hedges and planning residue — "probably fine for now", "should be enough", deferrals with no marker. Promote to
```
TODO
```
/
```
FIXME
```
or restate as the actual bound.
Authoring-language slips — untranslated fragments in prose whose language is otherwise English, or the reverse in a
```
.zh.md
```
counterpart. Translate or delete.

无效设计会话引用 ——
```
(decision 7)
```
、
```
(audit C2)
```
、
```
design §4.7
```
、阶段标签（
```
T4
```
、
```
W3
```
）。如果该决策有已提交的归属文件，按名称和路径引用；否则删除引用并重新表述其事实性内容。
栈与PR视角 —— "此栈中的后续PR"、"本PR新增"、"上一次提交"。说明已发布的机制；延迟处理的工作需标记为
```
TODO
```
或关联问题引用。
变更叙述与版本标记 —— "曾经"、"不再"、"旧版X"以及索引标记（"v1"、"本次版本"、"今日"）。说明当前行为；已修复的回归问题需表述为现在时态的反事实语句（"若无X，则会发生Y"），绝不能提及代码库历史。
评审流程痕迹 —— "评审被驳回"、"评审已确认"、草稿序号（"本笔记第5版"）。将留存的决策和理由保留为平实的事实。
针对评审意见的辩解内容 —— "类型转换是安全的——它只是…"、"这是正确的，因为…"。说明确保代码或规则安全的不变量，或直接删除该注释。
重述与推导记录 —— 控制流叙述（"首先我们执行X，然后执行Y"）、对明显分支的论证。直接删除；仅保留非显而易见的约定。
模糊表述与规划残留 —— "目前可能没问题"、"应该足够"、无标记的延迟处理内容。升级为
```
TODO
```
/
```
FIXME
```
标记，或重新表述为实际的约束条件。
创作语言疏漏 —— 英文文本中夹杂未翻译的片段，或
```
.zh.md
```
对应文件中出现相反情况。进行翻译或删除。

What is not leakage

不属于泄露内容的情况

Unaided citation passes fail in both directions by deleting durable references and keeping dead ones. Apply these keep rules as written:

Issue references —
```
#1470
```
,
```
TODO(name):
```
, "issue #N owns the follow-up" resolve at HEAD; keep them on any surface, including READMEs. Do not relocate them to Agent Notes.
Merged-PR and issue citations inside Agent Notes and postmortems — sanctioned evidence per the documentation standard's change-story routing.
Suppression justifications — lint-disable
```
-- reason
```
clauses, coverage-ignore reasons, and empty-catch explanations are required prose; fix a false reason, never delete it.
Counterfactual-present regression pins — "without X, Y happens", "a naive X would…".
Measured bounds — "(measured: …)" calibrating a constant; the provenance word "measured" is load-bearing.
Runtime old/new states — "the old connection drains before the new one accepts" is runtime lifecycle, not change history.
Historical stage names inside a note's change-story sections — "the first cut shipped X" is current-state-safe there; indexical stamps ("this cut") stay banned everywhere.
External references that resolve outside the repo by design — standards sections (RFC 9110 §10.1.5), design-artifact names (a Figma frame). The §-ban covers uncommitted internal drafts, not external standards or committed docs that own their §-numbering.
Project voice and genre forms — "we" as project voice; a note's Alternatives-considered section.

未经辅助的引用检查可能会错误删除有效引用或保留无效引用，请严格遵循以下保留规则：

问题引用 ——
```
#1470
```
、
```
TODO(name):
```
、"问题#N负责后续跟进"可在HEAD版本中解析；可在任何载体上保留，包括README。无需移至Agent Notes。
Agent Notes和事后分析中的已合并PR与问题引用 —— 符合文档标准中变更说明路由规则的认可证据。
抑制理由 —— lint-disable的
```
-- reason
```
子句、覆盖率忽略理由、空catch块解释是必填文本；修正错误的理由，切勿删除。
反事实时态的回归固定说明 —— "若无X，则会发生Y"、"简单实现X会…"。
实测边界 —— "(measured: …)"用于校准常量；"measured"这个来源词是必要的。
运行时新旧状态 —— "旧连接在新连接接受请求前断开"属于运行时生命周期描述，而非变更历史。
笔记变更说明部分中的历史阶段名称 —— "第一版发布了X"在此处是安全的当前状态表述；索引标记（"本次版本"）在所有场景均禁用。
设计上需在代码库外解析的外部引用 —— 标准章节（RFC 9110 §10.1.5）、设计工件名称（Figma框架）。§符号的禁用仅针对未提交的内部草稿，不包括外部标准或已提交且自有章节编号的文档。
项目语气与体裁形式 —— 作为项目语气的"we"；笔记中的备选方案考量部分。

Workflow

工作流程

Require an explicit scope. Never touch
```
vendor/
```
, frozen archived notes, or recorded fixtures and snapshots — recorded model output and sealed history keep their original voice.
Audit read-only first: run the recall batteries, then judge every hit semantically. The batteries over-match by design and under-match by nature — also read the densest prose in scope without a pattern in hand.
Fix owner-first: for a bilingual pair, update the counterpart; for a generated or manifest-listed surface, fix the source.
Before deleting anything, enumerate the passage's propositions and check the overcorrection traps.
Verify: re-run the batteries expecting only sanctioned keeps, confirm every remaining citation resolves, and run your repository's documentation checks for touched docs.

明确范围。切勿修改
```
vendor/
```
目录、已冻结的归档笔记或已记录的测试用例与快照——已记录的模型输出和封存历史需保留原始表述。
先进行只读审核：运行召回检查工具，然后从语义上判断每个匹配结果。这些工具设计上会过度匹配，本质上也可能匹配不足——同时需手动阅读范围内最密集的文本，无需依赖固定模式。
优先修复源头：对于双语配对文件，同步更新对应版本；对于生成或清单列出的载体，修复源文件。
删除任何内容前，列出段落的所有命题并检查过度修正陷阱。
验证：重新运行检查工具，确保仅保留认可的内容；确认所有剩余引用均可解析；对修改过的文档运行代码库的文档检查。

trim-cot-leakage

Original

Translation