universal-dev-standards 6.5.0 → 6.7.0

package/bundled/ai/standards/ai-response-navigation.ai.yaml

@@ -3,10 +3,14 @@
 
 id: ai-response-navigation
 meta:
-  version: "1.1.0"
-  updated: "2026-06-10"
+  version: "1.3.0"
+  updated: "2026-08-17"
   source: core/ai-response-navigation.md
-  description: Every substantive AI response must include contextual next-step suggestions with recommended options
+  description: >
+    Every substantive AI response must include contextual next-step suggestions with recommended
+    options (rules 1-6, required). Optional rules 7-11 govern the answer itself: lead with the
+    finding, restate state across turns, no preamble, plain language as the subject, and a
+    trade-off on every option rather than only the recommended one.
 
 rules:
   - id: navigation-footer
@@ -77,7 +81,70 @@ rules:
       Localized notation: 〔模型：Fast〕 / 〔模型：Standard〕 / 〔模型：Capable〕
     priority: optional
 
+  # ── Rules 7-9 (v1.2.0): the answer itself, not the navigation after it ──
+  # Borrowed from ayghri/i-have-adhd (MIT), 3 of its 10 rules. The other 7 were dropped:
+  # 2 duplicate navigation-footer/recommendation-marking above, and 5 conflict with this
+  # standard ("no recap/no closers" contradicts navigation-footer; "cap lists at 5" would
+  # truncate evidence tables) or duplicate estimation-standards.
+  # Optional in the same sense as model-tier-annotation: a project MAY promote them to required.
+  # Their triggers are deliberately precise — a rule too loose to fire is not a rule.
+
+  - id: lead-with-the-finding
+    trigger: a response that answers a question, reports an investigation result, or presents a decision
+    instruction: >
+      Open with what was found or what to do — not the method, not a restatement of the request,
+      not a plan for answering. Evidence (file:line, command output, tables, measurements) is
+      support and belongs after the claim it supports. This orders the evidence; it does not
+      license omitting it.
+    priority: optional
+
+  - id: restate-state-each-turn
+    trigger: work spanning 3 or more exchanges, or a task with 3 or more steps
+    instruction: >
+      Restate in one line where the work stands. The reader cannot be assumed to hold
+      "we are on step 3 of 5" across messages. Composes with the in-progress template:
+      this rule governs the opening, that template governs the footer.
+    priority: optional
+
+  - id: no-preamble
+    trigger: any substantive response
+    instruction: >
+      Start with the answer. Do not open with a summary of what you are about to do, an
+      acknowledgement of the request, or an assessment of the question. Generalizes the
+      anti-sycophancy-prompting prohibition on "opening critique with positive affirmation"
+      from critiques to every substantive response — for a different reason: not flattery,
+      but the delay it puts between the reader and the answer.
+      Does NOT apply to closers: navigation-footer still stands.
+    priority: optional
+
+  - id: plain-language-is-the-subject
+    trigger: any response explaining a situation, a defect, or a system's behaviour to a human
+    instruction: >
+      Explain what happened in the words the reader would use. File paths, symbol names, line
+      references, command output and version strings are support: they belong after the sentence
+      they support, not as the sentence itself. This does NOT license omitting them — a reader who
+      wants to verify must be able to. It governs which of the two is the subject.
+      Distinct from lead-with-the-finding: that rule orders finding before evidence, this one
+      governs register. A response can lead with its finding and still state that finding in
+      vocabulary only its author holds; both leave the reader unable to act.
+    priority: optional
+
+  - id: every-option-carries-its-trade-off
+    trigger: a response asking the reader to choose between two or more courses of action
+    instruction: >
+      recommendation-marking requires marking the recommended option and giving ITS reason. This
+      extends that to the rest: each option states what it buys and what it costs, in its own terms.
+      A list where only the recommendation is argued hands the comparison back to the reader, which
+      is the work they asked to have done; and an option shown without its downside reads as having
+      none. A trade-off is not a hedge — "this may be slightly harder" is not a cost, "this rewrites
+      110 files and needs a human to check the translations" is. If an option genuinely has no
+      downside worth stating, say so explicitly: an empty cell reads as "not analysed" and the
+      reader cannot tell those apart. Composes with adaptive-quantity — trade-offs make each option
+      costlier to read, so the 1-5 cap matters more, not less.
+    priority: optional
+
 related_standards:
   - ai-command-behavior
   - ai-instruction-standards
   - ai-agreement-standards
+  - anti-sycophancy-prompting

package/bundled/ai/standards/spec-driven-development.ai.yaml

@@ -3,7 +3,7 @@
 
 id: spec-driven-development
 meta:
-  version: "1.3.0"
+  version: "1.4.0"
   updated: "2026-08-12"
   source: methodologies/guides/sdd-guide.md
   description: Spec-Driven Development workflow where documentation precedes implementation
@@ -198,6 +198,33 @@ rules:
     instruction: Archive spec with links to commits/PRs
     priority: required
 
+  - id: SDD-AC-VERIFIED
+    name: An AC with no verification item is not an AC
+    name_zh: 沒有驗證項的 AC 不是 AC
+    severity: high
+    rule: >
+      Every acceptance criterion MUST have a verification item pointing at it —
+      a test, a check, a gate, or an explicitly recorded manual step. An AC that
+      no verification item references MUST be demoted to a design intent rather
+      than carried as an AC.
+    rule_zh: >
+      每一條 AC 必須有指向它的驗證項（測試／檢查／閘門／明確記錄的手動步驟）。
+      沒有任何驗證項引用的 AC 必須降級為設計意圖，不得繼續掛在 AC 欄。
+    rationale: >
+      An unverified AC does not fail loudly — it stops being true while the spec
+      continues to assert it. Measured instance (XSPEC-380): a spec's AC-7 required
+      a component be "fully preserved, no regression"; its Test Plan had seven items
+      and none pointed at AC-7. The component stopped running on the day the AC was
+      written and was found three months later by accident.
+    rationale_zh: >
+      未經驗證的 AC 不會大聲失敗——它安靜地停止成立，而規格繼續宣稱它為真。
+      實測案例：某規格的 AC-7 要求「完整保留、無回歸」，Test Plan 七項無一指向它；
+      它保護的東西死在該 AC 被寫下的同一天，三個月後偶然才被發現。
+    anti_pattern: >
+      Treating "the reviewer will notice" as a verification item. A reviewer reads
+      the spec, and the spec says the AC holds. An AC is a claim about the world;
+      only something that touches the world can falsify it.
+
 best_practices:
   do:
     - Keep specs focused and atomic (one change per spec)

package/bundled/core/ai-response-navigation.md

@@ -2,8 +2,8 @@
 
 > **Language**: English | [繁體中文](../locales/zh-TW/core/ai-response-navigation.md) | [简体中文](../locales/zh-CN/core/ai-response-navigation.md)
 
-**Version**: 1.1.0
-**Last Updated**: 2026-06-10
+**Version**: 1.3.0
+**Last Updated**: 2026-08-17
 **Applicability**: All projects using AI-assisted development
 **Scope**: universal
 **Industry Standards**: None (Emerging AI tool practice)
@@ -19,6 +19,13 @@ This standard defines navigation behavior for AI responses: every substantive AI
 
 **Solution**: A standard "Navigation Footer" appended to every substantive AI response, with contextual templates, recommendation marking, and adaptive option quantities.
 
+**Scope note (v1.2.0, extended in 1.3.0)**: Rules 1–6 govern what comes *after* the answer.
+Rules 7–11 — **all optional** — govern the answer itself: lead with the finding (R7), restate state
+across turns (R8), no preamble (R9), plain language as the subject (R10), and a trade-off on every
+option rather than only the recommended one (R11). They exist because a response can satisfy every
+one of Rules 1–6 while burying its conclusion, stating it in vocabulary only its author holds, or
+listing options the reader still has to compare themselves.
+
 ---
 
 ## Core Rules
@@ -97,6 +104,118 @@ Tier names are **vendor-neutral**. Each tool or platform maps these tiers to its
 
 ---
 
+## The Answer Before the Navigation (Rules 7–11, Optional)
+
+> **R7–R9 borrowed from**: [`ayghri/i-have-adhd`](https://github.com/ayghri/i-have-adhd) (MIT), 3 of its 10 rules.
+> **R10–R11 added in 1.3.0** from a different source — a user telling the author, twice in one session,
+> that a correct and complete answer was unreadable. R7–R9 had already shipped and were being followed.
+> The other 7 were dropped: 2 are already covered by Rules 1–2 above, and 5 either conflict with
+> this standard (its "no recap / no closers" contradicts Rule 1's Navigation Footer; its
+> "cap lists at 5" would truncate evidence tables and traversal denominators) or duplicate
+> [estimation-standards](estimation-standards.md).
+
+**Why this section exists**: Rules 1–6 govern what follows the answer. Nothing governed the answer
+itself — a response could bury its conclusion under a wall of evidence and still satisfy every rule
+in this standard by appending a correct Navigation Footer. A reader who cannot find the answer is
+not helped by being told what to do next.
+
+**These five rules are optional**, in the same sense as Rule 6: adopting projects are not required
+to enable them, and existing skills need no retroactive update. A project MAY promote any of them to
+required in its own configuration. What is *not* optional is that they have precise triggers — a rule
+phrased so loosely that it never fires is indistinguishable from not having the rule.
+
+### Rule 7: Lead With the Finding, Not the Process （Optional）
+
+**Trigger**: a response that answers a question, reports an investigation result, or presents a decision.
+
+The first line states **what was found or what to do**. Not the method, not a restatement of the
+request, not a plan for answering.
+
+Evidence — file:line references, command output, tables, measurements — is **support**, and belongs
+after the claim it supports. Leading with evidence forces the reader to reconstruct the conclusion
+themselves, which is the work they asked to have done.
+
+| Instead of | Write |
+|-----------|-------|
+| "I checked 44 days of data across 63 domains and found that…" | "Delete those three queries. 46% of what they return is download pages." |
+| "Let me look at how this is configured." | "It is configured in `x.yaml:12`; the value is wrong because…" |
+
+**This does not license omitting the evidence.** It orders it.
+
+### Rule 8: Restate State in Multi-Turn Work （Optional）
+
+**Trigger**: work spanning 3 or more exchanges, or a task with 3 or more steps.
+
+Each response restates where the work stands, in one line. The reader cannot be assumed to hold
+"we are on step 3 of 5" across messages, and the cost of restating it is one sentence.
+
+This composes with Template 4 (*In Progress*) below: Rule 8 governs the **opening**, Template 4
+governs the **footer**.
+
+### Rule 9: No Preamble （Optional）
+
+**Trigger**: any substantive response.
+
+Start with the answer. Do not open with a summary of what you are about to do, an acknowledgement
+of the request, or an assessment of the question.
+
+This generalizes one existing prohibition: [anti-sycophancy-prompting](anti-sycophancy-prompting.md)
+already forbids *"Opening critique with positive affirmation"* — but only for critiques. Rule 9
+extends the same prohibition to every substantive response, for a different reason: not flattery,
+but the delay it puts between the reader and the answer.
+
+**Rule 9 does not apply to closers.** Rule 1 requires a Navigation Footer, and that requirement
+stands — the end of a response is where this standard puts the reader's next move.
+
+### Rule 10: Plain Language Is the Subject; Identifiers Are Support （Optional）
+
+**Trigger**: any response explaining a situation, a defect, or a system's behaviour to a human.
+
+Explain what happened in the words the reader would use. File paths, symbol names, line
+references, command output and version strings are **support** — they belong after the sentence
+they support, not as the sentence itself.
+
+| Instead of | Write |
+|-----------|-------|
+| "`load_topics()` at `intel-scout.py:883` reads `intel-topics.yaml`, while `load_feeds()` at `:1033` reads `intel-feeds.yaml`." | "It gathers material two ways: by searching keywords, and by subscribing to a fixed set of blogs." *(then cite both call sites)* |
+| "`manifest.skillHashes` has 137 entries keyed by `claude-code/project/<n>/SKILL.md` while the lookup uses `.claude/skills/<n>`." | "The hashes are stored under one naming scheme and looked up under another, so none are ever found." *(then show both keys)* |
+
+This is **not** a licence to omit the identifiers — a reader who wants to verify must be able to.
+It governs which of the two is the subject of the sentence.
+
+**Why it is separate from R7**: R7 orders *finding before evidence*. R10 governs *register* — a
+response can lead with its finding and still state that finding in vocabulary only its author
+holds. Both failures leave the reader unable to act; they are different failures.
+
+### Rule 11: Every Option Carries Its Own Trade-off （Optional）
+
+**Trigger**: a response that asks the reader to choose between two or more courses of action.
+
+Rule 2 requires marking the recommended option and giving *its* reason. Rule 11 extends that to
+the rest: **each** option states what it buys and what it costs, in its own terms.
+
+A list of options where only the recommended one is argued hands the comparison back to the
+reader — which is the work they asked to have done. And an option presented without its downside
+reads as having none, which is rarely true and never verifiable from the list alone.
+
+```markdown
+> **Please choose:**
+> | Option | Buys you | Costs you |
+> |---|---|---|
+> | **(A) …** ⭐ **Recommended** — [why this one] | … | … |
+> | **(B) …** | … | … |
+```
+
+**A trade-off is not a hedge.** "This may be slightly harder" is not a cost; "this rewrites 110
+files and needs a human to check the translations" is. If an option genuinely has no downside
+worth stating, say so explicitly rather than leaving the column empty — an empty cell reads as
+"not analysed", and the reader cannot tell those apart.
+
+**Composes with Rule 4**: the option count stays bounded (1–5). Trade-offs make each option
+costlier to read, so this rule makes Rule 4's cap matter more, not less.
+
+---
+
 ## Contextual Templates
 
 ### Template 1: Task Completed
@@ -284,6 +403,11 @@ Each tool's integration layer is responsible for rendering the Navigation Footer
 | R4 | 1–5 options, adapt to context |
 | R5 | Use `/command` format when applicable |
 | R6 | *(Optional)* Append `〔model: Fast\|Standard\|Capable〕` when tier is clear |
+| R7 | *(Optional)* Lead with the finding; evidence follows the claim it supports |
+| R8 | *(Optional)* 3+ turns or 3+ steps → restate state in one line |
+| R9 | *(Optional)* No preamble. Closers still required — see R1 |
+| R10 | *(Optional)* Plain language is the subject; identifiers support it, after the claim |
+| R11 | *(Optional)* Every option states what it buys and costs — not only the recommended one |
 
 | Exempt | Not Exempt |
 |--------|------------|
@@ -307,6 +431,8 @@ Each tool's integration layer is responsible for rendering the Navigation Footer
 
 | Version | Date | Changes |
 |---------|------|---------|
+| 1.3.0 | 2026-08-17 | Add optional R10–R11. R10 governs register: plain language is the subject of the sentence and identifiers support it — distinct from R7, which orders finding before evidence, because a response can lead with its finding and still state it in vocabulary only its author holds. R11 extends Rule 2 from the recommended option to all of them: a list where only the recommendation is argued hands the comparison back to the reader, and an option shown without its cost reads as having none |
+| 1.2.0 | 2026-08-17 | Add optional R7–R9 governing the answer itself (lead with the finding, restate state, no preamble). Borrowed from `ayghri/i-have-adhd` (MIT), 3 of its 10 rules; the other 7 were dropped as duplicated by R1–R2, in conflict with R1, or covered by estimation-standards. Rules 1–6 could all be satisfied by a response that buries its conclusion — R7–R9 close that |
 | 1.1.0 | 2026-06-10 | Add R6 optional model tier annotation (`〔model: Fast\|Standard\|Capable〕`); vendor-neutral; no forced changes to existing skills |
 | 1.0.0 | 2026-03-25 | Initial release |
 

package/bundled/core/spec-driven-development.md

@@ -1,7 +1,7 @@
 # Spec-Driven Development (SDD) Standards
 
-**Version**: 2.3.0
-**Last Updated**: 2026-06-08
+**Version**: 2.4.0
+**Last Updated**: 2026-08-17
 **Applicability**: All projects adopting Spec-Driven Development
 **Scope**: universal
 **Industry Standards**: None (Emerging 2025+ methodology)
@@ -50,6 +50,61 @@ UDS supports two AC notations. **GWT is the default and preferred** (Forward Der
 
 Provide **GWT or EARS** per AC (`.ac.yaml`: `given/when/then` **or** `ears`). Prefer GWT for BDD-derivable behaviour; reach for EARS when GWT feels forced. Do not require both; do not remove GWT.
 
+## An AC With No Verification Item Is Not an AC | 沒有驗證項的 AC 不是 AC
+
+Every acceptance criterion must have a **verification item that points at it** — a test, a
+check, a gate, or an explicitly recorded manual step. An AC that no verification item
+references is a **promise nobody kept**, and it does not fail loudly: it simply stops being
+true while the spec continues to assert it.
+
+每一條驗收標準都必須有一個**指向它的驗證項**——測試、檢查、閘門，或一則明確記錄的
+手動步驟。沒有任何驗證項引用的 AC 是**一張沒有人兌現的支票**，而且它不會大聲失敗：
+它只是安靜地停止成立，而規格繼續宣稱它為真。
+
+**Rule**: an AC without a verification item must be **demoted to a design intent**, not
+carried as an AC. Demotion is honest; an unverified AC is not.
+**規則**：沒有驗證項的 AC 必須**降級為設計意圖**，不得繼續掛在 AC 欄。降級是誠實的，
+未經驗證的 AC 不是。
+
+### The measured instance | 實測案例
+
+A 2026-05-14 spec carried `AC-7: the legacy system-report is fully preserved, no
+regression`. Its Test Plan had seven items and **none of them pointed at AC-7**. The
+report's timer was disabled and its deploy function was never called — **from the same day
+the AC was written**. It was found three months later, by accident, while verifying an
+unrelated install.
+
+**It was never wired to a check and later came loose. It was never wired at all.**
+
+一份 2026-05-14 的規格寫著 `AC-7：舊版系統報告完整保留（無回歸）`。它的 Test Plan
+有七項，**沒有一項指向 AC-7**。該報告的 timer 是 disabled、部署函式從未被呼叫——
+**從那條 AC 被寫下的同一天起**。三個月後在驗證另一件無關的安裝時偶然發現。
+
+**它不是後來斷線的。它從來沒有被接上過。**
+
+### Why "someone will check it" is not a verification item
+
+A verification item must be **executable or recorded**, not implied. "The reviewer will
+notice" is not one, because a reviewer reads the spec — and the spec says the AC holds.
+An AC is a claim **about the world**, and only something that touches the world can
+falsify it.
+
+「有人會檢查」不是驗證項。它必須**可執行或有紀錄**，不能是隱含的。「審查者會注意到」
+不算——因為審查者讀的是規格，而規格說那條 AC 成立。**AC 是一個關於世界的宣稱，
+只有碰得到世界的東西才能證偽它。**
+
+### Related | 關聯
+
+- [verification-evidence](verification-evidence.md) — VE-011 requires evidence to come from a
+  fresh run **after the last edit**; this standard is the upstream question of whether any
+  run was ever pointed at the claim in the first place.
+- [class-level-fix](class-level-fix.md) — the same discipline applied to *scope*: traverse the
+  set rather than enumerate it.
+
+## What's New in v2.4.0
+
+- **An AC with no verification item is not an AC** (XSPEC-380 R5). Every acceptance criterion must have a verification item pointing at it; one that has none is demoted to a design intent rather than carried as an AC. Measured instance: a spec's `AC-7` had no matching Test Plan item, and the thing it protected stopped running **on the day the AC was written** — found three months later by accident. An AC is a claim about the world, and only something that touches the world can falsify it.
+
 ## What's New in v2.3.0
 
 - **EARS notation** as an optional AC format (XSPEC-263): 5 EARS templates + `.ac.yaml` `ears` field. GWT remains default & preferred; `given/when/then` relaxed from `required` (backward compatible).

package/bundled/locales/zh-CN/CHANGELOG.md

@@ -1,8 +1,8 @@
 ---
 source: ../../CHANGELOG.md
-source_version: 6.5.0
-translation_version: 6.5.0
-last_synced: 2026-08-14
+source_version: 6.7.0
+translation_version: 6.7.0
+last_synced: 2026-08-17
 status: current
 ---
 
@@ -17,6 +17,37 @@ status: current
 
 ## [Unreleased]
 
+## [6.7.0] - 2026-08-17
+
+### 修正
+
+- **skill 触发面在三层全数复原——55 个 skill × 英文、zh-TW、zh-CN。** 一个 skill 的 `description` 是模型决定要不要调用它时**唯一看得到的东西**。commit `d415937e`（2026-02-10）把其中 17 个改写成 `[UDS] <标签>`，删掉 `Use when:` 与 `Keywords:` 两行、**连中文关键字一起**；相邻 commit 的标题写着 `token optimization`。当时**只有 token 可数**——没有任何东西在量触发面，于是那个取舍看起来是单边的，而它不是。2026-08-14 测量：55 个含 `SKILL.md` 的 skill 中，**27 个有触发条件、27 个有关键字、0 个有排除条件、28 个两者皆无**——而那 28 个正是方法论核心：tdd、bdd、atdd、spec-driven-dev、code-review、commit-standards、checkin、requirement。
+  - **英文来源**：28 个复原，且**全部 55 个补上排除条件**（`Not for:`）。只加触发不加排除，换来的是过度触发；而一个不该响却响的 skill 会被整个关掉，**连带拖走还能用的那些**。15 个可从 git 历史救回者中**有 8 个被重写**，因为历史文字已不描述现行行为——其中六个宣称自己会引导某套生命周期，而那已于 XSPEC-095 移交采用层。**救回一个过期的描述，比不救更糟。**
+  - **语系层**：zh-TW 与 zh-CN **各 55/55，是翻译不是转码**。**这一半才是重点**——以 `--locale zh-tw` 安装的项目，在英文来源已修好时仍拿到被剥过的描述，这个修正本来到不了它要给的那个读者。繁体与简体各自用地道用词，另修正三份 zh-CN 描述中混入的繁体字。
+  - **翻译 drift 62 → 38**，剩下的 38 是刻意的停点：24 份漂移完全来自本次 description 编辑者更新了 hash，12 份正文早已漂移者保留过期 hash——更新它们等于宣称整份文件已同步，而那件事没有人验证过。
+  - ⚠️ **28 个中有 10 个带着 `disable-model-invocation: true`**，由同一个 `d415937e` 加入。对那 10 个而言，补描述**并不会**让它们变成可被选中——挡住的是那个标志，而要不要拿掉它是**设计决定不是缺陷**。模型真正选得到的数量是 **45，不是 55**。
+
+### 新增
+
+- **`ai-response-navigation` 1.2.0 → 1.3.0 —— 可选规则 R10 与 R11。** 它们的来源与 R7–R9 不同：用户在同一次工作会话中**两度**指出，一个正确且完整的回答读不懂，而当时 R7–R9 已经出货且正在被遵守。**先讲发现并不足够。**
+  - **R10 —— 白话是主语，标识符是佐证。** *触发*：任何向人解释一个情况、一个缺陷、或一个系统行为的响应。用读者会用的话说清楚发生了什么；路径、符号、行号、命令输出、版本字符串属于它们所支持的那句话**之后**，而不是那句话本身。它**不是**可以省略它们的许可——想验证的读者必须验得了。**与 R7 分开是刻意的**：R7 管的是「先发现后证据」的顺序，而一个响应可以先讲发现、却仍用只有作者持有的词汇讲它。两者都让读者无法行动，但它们是不同的失效。
+  - **R11 —— 每个选项都要带自己的利弊。** *触发*：要求读者在两个以上做法之间选择的响应。规则 2 已要求标示推荐项并给出**它的**理由；R11 要求**每一个**选项都说明它换到什么、代价是什么。一份只有推荐项被论证的清单，等于**把比较的工作丢回给读者**——而那正是他请你做的事；而没标代价的选项读起来像是没有代价。**利弊不是模棱两可**：「稍微难一点」不是代价，「重写 110 个文件且翻译需要人审」才是。空白的代价栏读起来是「没有分析过」，而读者分不出这两者。
+
+## [6.6.0] - 2026-08-17
+
+### 新增
+
+- **`spec-driven-development` 2.3.0 → 2.4.0 —— 没有验证项的 AC 不是 AC。** 每一条验收标准都必须有一个**指向它的验证项**——测试、检查、闸门，或一则明确记录的手动步骤。没有任何验证项引用的 AC 是一张没有人兑现的支票，而且它**不会**大声失败：它只是安静地停止成立，而规格继续宣称它为真。**规则**：这样的 AC 必须**降级为设计意图**，不得继续挂在 AC 栏——降级是诚实的，未经验证的 AC 不是。标准同时写明什么**不算**验证项：「审查者会注意到」不是验证项，因为审查者读的是规格，而规格说那条 AC 成立。新增规则 `SDD-AC-VERIFIED`。
+  - **实测案例**：一份 2026-05-14 的规格写着 `AC-7：旧版系统报告完整保留（无回归）`。它的 Test Plan 有七项，**没有一项指向 AC-7**。该报告的 timer 是 disabled、部署函数从未被调用——**从那条 AC 被写下的同一天起**。三个月后在验证一件无关的安装时偶然撞到。它不是接上了检查后来松脱，**它从来没有被接上过**。
+
+- **`ai-response-navigation` 1.1.0 → 1.2.0 —— 可选规则 R7–R9，管答案本身**（XSPEC 借鉴 B-10，来源 [`ayghri/i-have-adhd`](https://github.com/ayghri/i-have-adhd)，MIT）。规则 1–6 管的是答案**之后**要附什么：导航区块、标记过的推荐、匹配响应类型的模板。**答案本身没有任何规则在管。** 于是一个响应可以把结论埋在一整面证据底下，只要结尾附上正确的导航区块，它仍然满足**本标准的每一条**——而找不到答案的读者，不会因为被告知下一步而得到帮助。
+  - **R7 —— 先讲发现，不要先讲过程。** *触发*：回答问题、汇报调查结果、或提出决策的响应。第一行写**查到了什么**或**该做什么**——不是方法、不是把问题复述一遍、不是回答的计划。证据（`file:line`、命令输出、表格、测量数字）是**佐证**，应放在它所支持的论断之后；以证据开场会迫使读者自行重建结论，而那正是他请你做的工作。本条规范的是**顺序**，**不**代表可以省略证据。
+  - **R8 —— 每一轮重述进度。** *触发*：跨 3 轮以上的对话，或含 3 个以上步骤的任务。用一行说明工作进行到哪里。不能假设读者能在消息之间记住「我们在 5 步中的第 3 步」，而重述它的成本是一个句子。与模板 4（进行中）互补：**R8 管开头，模板管结尾。**
+  - **R9 —— 不要开场白。** *触发*：任何实质性响应。本条把一项既有禁令一般化：[`anti-sycophancy-prompting`](../../core/anti-sycophancy-prompting.md) 已经禁止「以正面肯定开场批评」，但**仅限批评情境**。R9 把同一项禁令扩及每一个实质响应，理由不同——不是为了防拍马屁，而是为了消除它在读者与答案之间制造的延迟。**R9 不适用于结语**；R1 的导航区块要求依然成立。
+  - **来源十条只取三条，其余七条的淘汰理由写进标准本文**，不是只写在待办清单里。两条与 R1–R2 重复。三条与本标准或其他标准冲突：它的「不要 recap／不要结语」**与 R1 的导航区块直接矛盾**；它的「列表上限 5 项」会截断证据表格与遍历分母；它的「具体时间估计」已由 [`estimation-standards`](../../core/estimation-standards.md) 涵盖。
+  - **可选的语义同 R6**（模型级别标注）：采用者不必启用、既有 skill 不需回头补，项目**可以**在自己的配置中把任一条提升为必须。**不可选的是每一条都带有精确的触发条件**——一条松到永远不会启动的规则，与没有这条规则无从分辨，那正是 XSPEC-378 记录的失效模式。
+  - **扩充既有标准而非新建一支**：再开一支管「AI 怎么对人类写回答」的标准，会让同一条轴出现两个实现。
+
 ## [6.5.0] - 2026-08-14
 
 > ⚠️ **本节简体译文待补。** 本次发布包含两批内容：XSPEC 借鉴 B-01 的五条标准补强

package/bundled/locales/zh-CN/README.md

@@ -15,7 +15,7 @@ status: current
 
 > **语言**: [English](../../README.md) | [繁體中文](../zh-TW/README.md) | 简体中文
 
-**版本**: 6.5.0 | **发布日期**: 2026-08-14 | **授权**: [双重授权](../../LICENSE) (CC BY 4.0 + MIT)
+**版本**: 6.7.0 | **发布日期**: 2026-08-17 | **授权**: [双重授权](../../LICENSE) (CC BY 4.0 + MIT)
 
 语言无关、框架无关的软件项目文档标准。通过 AI 原生工作流，确保不同技术栈之间的一致性、质量和可维护性。
 

package/bundled/locales/zh-CN/SECURITY.md

@@ -13,7 +13,7 @@ status: current
 <!-- UDS_SUPPORTED_VERSIONS_START -->
 | 版本 | 支持状态 |
 |------|--------|
-| 6.5.0 | ✅ 最新正式版 |
+| 6.7.0 | ✅ 最新正式版 |
 | < 6.0.0 | ❌ 已终止支持 |
 <!-- UDS_SUPPORTED_VERSIONS_END -->
 

package/bundled/locales/zh-CN/core/ai-response-navigation.md

@@ -1,8 +1,8 @@
 ---
 source: ../../../core/ai-response-navigation.md
-source_version: 1.1.0
-translation_version: 1.1.0
-last_synced: 2026-06-10
+source_version: 1.3.0
+translation_version: 1.3.0
+last_synced: 2026-08-17
 status: current
 ---
 
@@ -10,8 +10,8 @@ status: current
 
 > **语言**: [English](../../../core/ai-response-navigation.md) | [繁體中文](../../zh-TW/core/ai-response-navigation.md) | 简体中文
 
-**版本**: 1.1.0
-**最后更新**: 2026-06-10
+**版本**: 1.3.0
+**最后更新**: 2026-08-17
 **适用范围**: 所有使用 AI 辅助开发的项目
 **范围**: universal
 **行业标准**: 无（新兴 AI 工具实践）
@@ -27,6 +27,12 @@ status: current
 
 **解决方案**：在每个实质性 AI 响应结尾附加标准化的「导航区块」，包含情境模板、推荐标记和弹性选项数量。
 
+**范围注记（v1.2.0，v1.3.0 扩充）**：规则 1–6 管的是答案**之后**要附什么。
+规则 7–11 **全部属可选**，管的是答案本身：先讲发现（R7）、每轮重述进度（R8）、不要开场白（R9）、
+白话是主语（R10）、每个选项都要带自己的利弊而非只有推荐项有（R11）。新增的理由是——
+一个响应可以满足规则 1–6 的每一条，同时把结论埋起来；
+**而找不到答案的读者，不会因为结尾有一个正确的导航区块被告知下一步而得到帮助。**
+
 ---
 
 ## 核心规则
@@ -103,6 +109,103 @@ status: current
 
 ---
 
+## 导航之前的那个答案（规则 7–11，可选）
+
+> **规则 7–9 借鉴自**：[`ayghri/i-have-adhd`](https://github.com/ayghri/i-have-adhd)（MIT），十条中取三条。
+> **规则 10–11 于 1.3.0 新增**，来源不同——用户在同一次工作会话中两度指出，
+> 一个正确且完整的回答读不懂。当时规则 7–9 已经出货且正在被遵守。
+> 其余七条删去：两条已被上方规则 1–2 涵盖，五条与本标准冲突
+> （它的「不要 recap／不要结语」与规则 1 的导航区块直接矛盾；它的「列表上限 5 项」
+> 会截断证据表格与遍历分母）或与 [estimation-standards](estimation-standards.md) 重复。
+
+**这一节为何存在**：规则 1–6 管的是答案**之后**要附什么，而答案本身没有任何规则在管——
+一个响应可以把结论埋在一整面证据底下，只要结尾附上正确的导航区块，
+它仍然满足本标准的每一条。**找不到答案的读者，不会因为被告知下一步而得到帮助。**
+
+**这五条是可选的**，语义同规则 6：采用项目不必启用，既有 skill 也不需回头补。
+项目**可以**在自己的配置中把任一条提升为必须。**不可选的是它们必须有精确的触发条件**——
+一条松到永远不会启动的规则，与没有这条规则无从分辨。
+
+### 规则 7：先讲发现，不要先讲过程（可选）
+
+**触发条件**：回答问题、汇报调查结果、或提出决策的响应。
+
+第一行写**查到了什么**或**该做什么**。不是方法、不是把问题复述一遍、不是回答的计划。
+
+证据——`file:line`、命令输出、表格、测量数字——是**佐证**，应放在它所支持的论断**之后**。
+以证据开场会迫使读者自行重建结论，而那正是他请你做的工作。
+
+| 不要写 | 改写成 |
+|---|---|
+| 「我检查了 44 天、63 个域名的数据，发现……」 | 「那三组查询删掉。它返回的东西有 46% 是下载页。」 |
+| 「让我看看这是怎么配置的。」 | 「配置在 `x.yaml:12`，值是错的，因为……」 |
+
+**这不代表可以省略证据**，它规范的是顺序。
+
+### 规则 8：多轮工作中每一轮重述进度（可选）
+
+**触发条件**：跨 3 轮以上的对话，或含 3 个以上步骤的任务。
+
+每条响应用一行说明工作进行到哪里。不能假设读者能在消息之间记住
+「我们在 5 步中的第 3 步」，而重述它的成本是一个句子。
+
+本条与下方模板 4（进行中）互补：**规则 8 管开头，模板 4 管结尾。**
+
+### 规则 9：不要开场白（可选）
+
+**触发条件**：任何实质性响应。
+
+从答案开始。不要以「我接下来要做什么」的预告、对请求的确认、或对问题本身的评价开场。
+
+本条把一项既有禁令一般化：[anti-sycophancy-prompting](anti-sycophancy-prompting.md)
+已经禁止「以正面肯定开场批评」，但**仅限批评情境**。规则 9 把同一项禁令扩及每一个实质响应，
+理由不同：不是为了防拍马屁，而是为了消除它在读者与答案之间制造的延迟。
+
+**规则 9 不适用于结语。** 规则 1 要求的导航区块依然成立——
+响应的结尾正是本标准安放「读者下一步」的位置。
+
+### 规则 10：白话是主语，标识符是佐证（可选）
+
+**触发条件**：任何向人解释一个情况、一个缺陷、或一个系统行为的响应。
+
+用读者会用的话说清楚发生了什么。文件路径、符号名称、行号、命令输出、版本字符串是**佐证**——
+它们属于它们所支持的那句话**之后**，而不是那句话本身。
+
+| 不要写 | 改写成 |
+|---|---|
+| 「`intel-scout.py:883` 的 `load_topics()` 读 `intel-topics.yaml`，`:1033` 的 `load_feeds()` 读 `intel-feeds.yaml`。」 | 「它有两种取材方式：搜关键字，和订阅固定几个博客。」*（然后再附两个调用点）* |
+| 「`manifest.skillHashes` 有 137 条，key 是 `claude-code/project/<n>/SKILL.md`，而查询用的是 `.claude/skills/<n>`。」 | 「哈希值存在一套命名下，查询用的是另一套，所以一个都找不到。」*（然后再贴两边的 key）* |
+
+这**不是**可以省略标识符的许可——想验证的读者必须验得了。它规范的是**两者之中哪一个当主语**。
+
+**为什么它与 R7 是两条**：R7 规范的是**先讲发现再给证据**的顺序。R10 规范的是**语域**——
+一个响应可以先讲发现，却仍然用只有作者持有的词汇讲那个发现。两者都让读者无法行动，但它们是不同的失效。
+
+### 规则 11：每个选项都要带自己的利弊（可选）
+
+**触发条件**：要求读者在两个以上做法之间选择的响应。
+
+规则 2 要求标示推荐选项并给出**它的**理由。规则 11 把这件事扩及其余：
+**每一个**选项都要用它自己的话说明它换到什么、代价是什么。
+
+一份只有推荐项被论证的选项清单，等于**把比较的工作丢回给读者**——而那正是他请你做的事。
+而一个没有标出代价的选项，读起来像是没有代价，这很少为真，且从清单本身无从验证。
+
+```markdown
+> **请选择：**
+> | 选项 | 换到什么 | 代价是什么 |
+> |---|---|---|
+> | **(A) …** ⭐ **推荐** — [为什么是这个] | … | … |
+> | **(B) …** | … | … |
+```
+
+**利弊不是模棱两可。**「这可能会稍微难一点」不是代价；「这要重写 110 个文件，而且翻译需要人审」才是。
+若某个选项真的没有值得一提的代价，**明说**，不要留空——空白读起来是「没有分析过」，而读者分不出这两者。
+
+**与规则 4 相辅**：选项数维持在 1–5。利弊让每个选项读起来更花力气，所以这条规则让规则 4 的上限**更**要紧，不是更不要紧。
+
+---
+
 ## 情境模板
 
 ### 模板 1：任务完成
@@ -290,6 +393,11 @@ AI 需要用户做出选择或提供信息时使用。
 | R4 | 1–5 个选项，依情境调整 |
 | R5 | 适用时使用 `/command` 格式 |
 | R6 | *（可选）* 级别明确时附加 `〔模型：Fast｜Standard｜Capable〕` |
+| R7 | *（可选）* 先讲发现；证据放在它所支持的论断之后 |
+| R8 | *（可选）* 跨 3 轮或 3 步以上 → 用一行重述进度 |
+| R9 | *（可选）* 不要开场白。结语仍为必须——见 R1 |
+| R10 | *（可选）* 白话是主语；标识符放在论断之后当佐证 |
+| R11 | *（可选）* 每个选项都要说明换到什么、代价是什么——不只推荐那一个 |
 
 | 豁免 | 不豁免 |
 |------|--------|
@@ -313,6 +421,8 @@ AI 需要用户做出选择或提供信息时使用。
 
 | 版本 | 日期 | 变更 |
 |------|------|------|
+| 1.3.0 | 2026-08-17 | 新增可选规则 R10–R11。R10 管语域:白话是句子的主语、标识符当佐证——与 R7 不同,R7 管的是「先发现后证据」的顺序,而一个响应可以先讲发现却仍用只有作者持有的词汇讲它。R11 把规则 2 从推荐选项扩及全部:只论证推荐项的清单等于把比较丢回给读者,而没标代价的选项读起来像没有代价 |
+| 1.2.0 | 2026-08-17 | 新增可选规则 R7–R9，管答案本身（先讲发现、重述进度、不要开场白）。借鉴自 `ayghri/i-have-adhd`（MIT），十条取三；其余七条因已被 R1–R2 涵盖、与 R1 冲突、或与 estimation-standards 重复而删去。规则 1–6 全部可以被一个把结论埋起来的响应满足——R7–R9 补上这个缺口 |
 | 1.1.0 | 2026-06-10 | 新增规则 R6 可选模型级别标注（`〔模型：Fast｜Standard｜Capable〕`）；与厂商无关；不强制既有技能回改 |
 | 1.0.0 | 2026-03-25 | 初始版本 |
 

package/bundled/locales/zh-CN/core/spec-driven-development.md

@@ -4,7 +4,7 @@ source_version: 2.3.0
 translation_version: 2.3.0
 last_synced: 2026-06-10
 source_hash: 08dd8c2bee20
-status: current
+status: stale
 ---
 
 # 规格驱动开发 (SDD) 标准

package/bundled/locales/zh-CN/skills/ac-coverage/SKILL.md

@@ -2,12 +2,16 @@
 name: ac-coverage
 source: ../../../../skills/ac-coverage/SKILL.md
 source_version: 1.1.0
-translation_version: 1.1.0
-last_synced: 2026-07-09
-source_hash: eac0755fd844
+translation_version: 1.2.0
+last_synced: 2026-08-17
+source_hash: 168298f3a994
 scope: universal
 status: current
-description: "[UDS] 分析验收条件（AC）与测试之间的追踪关系并生成覆盖率报告"
+description: |
+  [UDS] 分析验收条件（AC）与测试之间的追踪关系，并生成需求层级的覆盖率报告。
+  Use when: 审计哪些验收条件已有测试、从 SPEC 文件建立追踪矩阵、发布前找出尚未被覆盖的 AC。
+  Not for: 代码层级的行／分支／函数覆盖率——请用 /coverage；补写缺少的测试——请用 /tdd 或 /spec-derive。
+  Keywords: AC coverage, traceability, acceptance criteria, SPEC, traceability matrix, 验收条件, 需求追踪, 覆盖率矩阵, 追踪矩阵.
 allowed-tools: Read, Grep, Glob
 argument-hint: "[规格文件路径]"
 ---