dsh-doublecheck 0.5.0 → 0.6.0

package/lib/guard/prose.js

@@ -5,6 +5,12 @@
  * `green gate`) stay English: they are stable session-log ids the durable
  * once-semantics fold matches, and translating them would silently break
  * `remindOnce` after a language switch. Only the model-facing text localizes.
+ *
+ * Scope of localization: every model-visible string the package injects or
+ * answers with (gate prose, the critic task, `/doublecheck` replies, the
+ * switch notice) is localized. The workspace documents (`doublecheck-spec.md`
+ * / `doublecheck-report.md`) keep their English section headings: they are
+ * stable artifacts whose structure other tooling may match.
  * @module dsh-doublecheck/guard/prose
  */
 const EN = {
@@ -55,6 +61,38 @@ const EN = {
     reviewUnavailableNoFindings: 'Adversary review returned no structured findings.',
     reviewFindingsHeader: count => `Adversary review found ${count} objection(s) the delivery must answer:`,
     reviewFindingsFooter: 'Answer each: fix what is real, and state plainly what is false.',
+    reviewHeldBack: held => `… ${held} further objection(s) held back by adversaryMaxFindings`,
+    criticTask: 'You are the delivery reviewer for this software-engineering session. The '
+        + 'conversation you inherited contains a requirements spec recorded with the '
+        + 'doublecheck_spec tool (six dimensions: goal, scope, acceptance criteria, '
+        + 'failure modes, priorities, non-goals), followed by the implementation '
+        + 'work and its test evidence. Assume the delivery FAILS its own spec. Hunt '
+        + 'for the strongest objections you can actually support from this session: '
+        + 'dimensions the work did not meet, acceptance criteria with no evidence, '
+        + 'scope or non-goal violations, failure modes left unhandled. Answer '
+        + 'through the required structured output, one entry per objection, citing '
+        + 'what in the session supports it. If — and only if — the evidence '
+        + 'genuinely satisfies every dimension, return an empty findings list. Do '
+        + 'not invent objections; the empty answer is correct when nothing is wrong.',
+    switchOnDurable: 'Double-check discipline was switched ON for this session (changed by the user): the requirements grill, the red/green gates, and the delivery review are active again.',
+    switchOnLocal: 'Double-check discipline was switched ON for this process only (changed by the user; this harness build predates the ignorable append surface, so the switch is not durable): the requirements grill, the red/green gates, and the delivery review are active again.',
+    switchOffDurable: 'Double-check discipline was switched OFF for this session (changed by the user): the discipline gates defer to the human approval chain until switched back on.',
+    switchOffLocal: 'Double-check discipline was switched OFF for this process only (changed by the user; this harness build predates the ignorable append surface, so the switch is not durable): the discipline gates defer to the human approval chain until switched back on.',
+    commandNoAgent: '/doublecheck needs an agent session to inspect.',
+    commandUnknown: input => `Unknown /doublecheck argument "${input}". Usage: /doublecheck status|report|on|off`,
+    commandAlreadyOn: 'Double-check discipline is already ON for this session.',
+    commandAlreadyOff: 'Double-check discipline is already OFF for this session.',
+    commandOnDurable: 'Double-check discipline ON for this session.',
+    commandOnLocal: 'Double-check discipline ON for this process only (not durable on this harness).',
+    commandOffDurable: 'Double-check discipline OFF for this session.',
+    commandOffLocal: 'Double-check discipline OFF for this process only (not durable on this harness).',
+    commandStatus: facts => [
+        `Double-check discipline is ${facts.enabled ? 'ON' : 'OFF'} for this session. (/doublecheck on|off)`,
+        `Modules: grill=${facts.modules.grill ? 'on' : 'off'}, tdd=${facts.modules.tdd ? 'on' : 'off'}, adversary=${facts.modules.adversary ? 'on' : 'off'} (cordis.yml).`,
+        `Intensity: ${facts.intensity}. Default switch: ${facts.defaultEnabled ? 'on' : 'off'}. remindOnce: ${facts.remindOnce ? 'on' : 'off'}.`,
+        `Stage: spec=${facts.hasSpec ? 'committed' : 'missing'}, tests=${facts.color}, review=${facts.reviewed ? 'on record' : 'not run'}, edits=${facts.editCount}.`,
+        `Usage: /doublecheck status|report|on|off`,
+    ].join('\n'),
 };
 const ZH = {
     grillReminder: '交付前三查：这个任务描述很简短，本会话还没有记录 doublecheck spec。先停下编辑，把六个需求维度确认清楚——目标、范围、验收标准、失败模式、优先级、非目标。按 grill-requirements 技能执行：用 ask_user_question 向用户提问直到达成共识，用 doublecheck_spec 记录结果，然后再继续编辑。',
@@ -75,6 +113,33 @@ const ZH = {
     reviewUnavailableNoFindings: '对抗式审查没有返回结构化发现。',
     reviewFindingsHeader: count => `对抗式审查发现 ${count} 条反对意见，交付必须回应：`,
     reviewFindingsFooter: '逐条回应：属实的修掉，不属实的明确说明。',
+    reviewHeldBack: held => `… 另有 ${held} 条反对意见因 adversaryMaxFindings 上限被保留`,
+    criticTask: '你是本次软件工程会话的交付审查者。你继承的会话里包含一份通过 '
+        + 'doublecheck_spec 工具记录的需求 spec（六个维度：目标、范围、验收标准、'
+        + '失败模式、优先级、非目标），其后是实现工作和测试证据。假设交付不满足自己的 '
+        + 'spec。寻找你能从本会话证据中真正支撑的最强反对意见：未达到的维度、没有证据的'
+        + '验收标准、范围或非目标被违反、未处理的失败模式。通过要求的结构化输出作答，'
+        + '每条反对意见占一项，并引用会话中支撑它的内容。当且仅当证据确实满足每一个维度时，'
+        + '才返回空的 findings 列表。不要编造反对意见；没有问题时空答案才是正确的。',
+    switchOnDurable: 'Double-check 纪律已对本会话开启（由用户切换）：需求盘问、红绿门和交付审查恢复生效。',
+    switchOnLocal: 'Double-check 纪律已对本进程开启（由用户切换；此 harness 版本早于 ignorable 追加接口，切换不持久）：需求盘问、红绿门和交付审查恢复生效。',
+    switchOffDurable: 'Double-check 纪律已对本会话关闭（由用户切换）：在重新开启之前，纪律门让位于人工审批链。',
+    switchOffLocal: 'Double-check 纪律已对本进程关闭（由用户切换；此 harness 版本早于 ignorable 追加接口，切换不持久）：在重新开启之前，纪律门让位于人工审批链。',
+    commandNoAgent: '/doublecheck 需要代理会话才能查看。',
+    commandUnknown: input => `未知的 /doublecheck 参数 "${input}"。用法：/doublecheck status|report|on|off`,
+    commandAlreadyOn: '本会话的 Double-check 纪律已经是开启状态。',
+    commandAlreadyOff: '本会话的 Double-check 纪律已经是关闭状态。',
+    commandOnDurable: 'Double-check 纪律已对本会话开启。',
+    commandOnLocal: 'Double-check 纪律已对本进程开启（此 harness 上不持久）。',
+    commandOffDurable: 'Double-check 纪律已对本会话关闭。',
+    commandOffLocal: 'Double-check 纪律已对本进程关闭（此 harness 上不持久）。',
+    commandStatus: facts => [
+        `本会话的 Double-check 纪律：${facts.enabled ? '已开启' : '已关闭'}。(/doublecheck on|off)`,
+        `模块：grill=${facts.modules.grill ? '开' : '关'}, tdd=${facts.modules.tdd ? '开' : '关'}, adversary=${facts.modules.adversary ? '开' : '关'}（cordis.yml）。`,
+        `强度：${facts.intensity}。默认开关：${facts.defaultEnabled ? '开' : '关'}。remindOnce：${facts.remindOnce ? '开' : '关'}。`,
+        `阶段：spec=${facts.hasSpec ? '已提交' : '缺失'}, tests=${facts.color}, review=${facts.reviewed ? '有记录' : '未运行'}, edits=${facts.editCount}。`,
+        '用法：/doublecheck status|report|on|off',
+    ].join('\n'),
 };
 /** The localized prose tables by language. */
 export const PROSE = { en: EN, zh: ZH };

package/lib/guard/review.d.ts

@@ -42,6 +42,14 @@ export interface ReviewOutcome {
 export declare const REVIEW_OUTPUT_SCHEMA: ObjectJsonSchema;
 /** Injected when the critic found nothing supportable (English contract text). */
 export declare const CLEAN_TEXT: string;
+/**
+ * Sort findings blocker-first, keeping the critic's original order within a
+ * severity. A deterministic order means the most threatening objection leads
+ * both the injected prose and the durable record.
+ * @param findings - the structured findings as produced by the critic.
+ * @returns the same findings, severity-ordered.
+ */
+export declare function sortFindings(findings: readonly ReviewFinding[]): ReviewFinding[];
 /** Render structured findings as the model-facing review text. */
 export declare function renderFindings(findings: readonly ReviewFinding[], prose?: GuardProse, totalCount?: number): string;
 /**

package/lib/guard/review.js

@@ -41,25 +41,25 @@ export const REVIEW_OUTPUT_SCHEMA = {
         },
     },
 };
-/**
- * The critic's task, delivered as the forked child's user message. The child
- * already holds the parent history; this prompt only sets the adversarial
- * stance and the answer discipline.
- */
-const CRITIC_TASK = 'You are the delivery reviewer for this software-engineering session. The '
-    + 'conversation you inherited contains a requirements spec recorded with the '
-    + 'doublecheck_spec tool (six dimensions: goal, scope, acceptance criteria, '
-    + 'failure modes, priorities, non-goals), followed by the implementation '
-    + 'work and its test evidence. Assume the delivery FAILS its own spec. Hunt '
-    + 'for the strongest objections you can actually support from this session: '
-    + 'dimensions the work did not meet, acceptance criteria with no evidence, '
-    + 'scope or non-goal violations, failure modes left unhandled. Answer '
-    + 'through the required structured output, one entry per objection, citing '
-    + 'what in the session supports it. If — and only if — the evidence '
-    + 'genuinely satisfies every dimension, return an empty findings list. Do '
-    + 'not invent objections; the empty answer is correct when nothing is wrong.';
 /** Injected when the critic found nothing supportable (English contract text). */
 export const CLEAN_TEXT = PROSE.en.reviewClean;
+/** Severity ranks for a deterministic, blocker-first findings order. */
+const SEVERITY_RANK = {
+    blocker: 0,
+    major: 1,
+    minor: 2,
+    info: 3,
+};
+/**
+ * Sort findings blocker-first, keeping the critic's original order within a
+ * severity. A deterministic order means the most threatening objection leads
+ * both the injected prose and the durable record.
+ * @param findings - the structured findings as produced by the critic.
+ * @returns the same findings, severity-ordered.
+ */
+export function sortFindings(findings) {
+    return [...findings].sort((a, b) => SEVERITY_RANK[a.severity] - SEVERITY_RANK[b.severity]);
+}
 /** Render structured findings as the model-facing review text. */
 export function renderFindings(findings, prose = PROSE.en, totalCount = findings.length) {
     const lines = [
@@ -71,7 +71,7 @@ export function renderFindings(findings, prose = PROSE.en, totalCount = findings
         lines.push(`  ${finding.detail}`);
     }
     if (totalCount > findings.length) {
-        lines.push(`… ${totalCount - findings.length} further objection(s) held back by adversaryMaxFindings`);
+        lines.push(prose.reviewHeldBack(totalCount - findings.length));
     }
     lines.push('', prose.reviewFindingsFooter);
     return lines.join('\n');
@@ -101,7 +101,7 @@ export async function runAdversaryReview(ctx, config, agent, turnSignal) {
     try {
         run = await subagents.start(config.adversaryProvider, {
             label: 'doublecheck-adversary',
-            prompt: [{ type: 'text', text: CRITIC_TASK }],
+            prompt: [{ type: 'text', text: prose.criticTask }],
             parent: agent,
             signal,
             ...config.adversaryModel !== null ? { agentOptions: { model: config.adversaryModel } } : {},
@@ -149,7 +149,7 @@ async function settleResult(run, signal, prose, maxFindings) {
     if (!Array.isArray(raw)) {
         return { verdict: 'unavailable', findings: [], text: outputText(result.output, prose.reviewUnavailableNoFindings) };
     }
-    const findings = raw.slice(0, maxFindings);
+    const findings = sortFindings(raw).slice(0, maxFindings);
     if (findings.length === 0)
         return { verdict: 'clean', findings: [], text: prose.reviewClean };
     return { verdict: 'findings', findings, text: renderFindings(findings, prose, raw.length) };

package/lib/invariant.js

@@ -20,6 +20,7 @@
  * profile row.
  * @module dsh-doublecheck/invariant
  */
+import { SPEC_FIELD_NAMES } from "./domain/vocabulary.js";
 import { deriveReportVerdict, foldReportFacts } from "./domain/report.js";
 /** Full npm package name owning the reported failures. */
 export const PACKAGE_NAME = 'dsh-doublecheck';
@@ -28,9 +29,7 @@ export const name = 'dsh-doublecheck-invariant';
 /** Service required before the companion can reserve package ownership. */
 export const inject = ['invariants'];
 /** All six spec fields, in declaration order. */
-const SPEC_FIELDS = [
-    'goal', 'scope', 'acceptanceCriteria', 'failureModes', 'priorities', 'nonGoals',
-];
+const SPEC_FIELDS = SPEC_FIELD_NAMES;
 /**
  * Build the installer over a facts source. The standalone companion and the
  * main plugin share this body; only the facts differ.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-doublecheck",
-  "version": "0.5.0",
+  "version": "0.6.0",
   "description": "Double-check before you ship: grill the requirements, test the implementation, prove the delivery.",
   "packageManager": "pnpm@11.7.0",
   "repository": {
@@ -27,12 +27,21 @@
       "types": "./lib/guard/index.d.ts",
       "import": "./lib/guard/index.js"
     },
+    "./invariant": {
+      "types": "./lib/invariant.d.ts",
+      "import": "./lib/invariant.js"
+    },
     "./package.json": "./package.json"
   },
+  "publishConfig": {
+    "access": "public"
+  },
+  "sideEffects": false,
   "files": [
     "lib",
     "skills",
     "cordis.patch.yml",
+    "strict.patch.yml",
     "CHANGELOG.md",
     "README.md",
     "README.zh.md",
@@ -82,7 +91,8 @@
     "@deepseek-ai/schemastery": "^3.18.1"
   },
   "dependencies": {
-    "typescript": "^5.7.0"
+    "typescript": "^5.7.0",
+    "zod": "^4.4.3"
   },
   "devDependencies": {
     "@deepseek-ai/cordis": "^4.0.1",
@@ -103,7 +113,6 @@
     "@vitest/coverage-v8": "^3.2.0",
     "eslint": "^9.30.0",
     "typescript-eslint": "^8.35.0",
-    "vitest": "^3.2.0",
-    "zod": "^4.4.3"
+    "vitest": "^3.2.0"
   }
 }

package/strict.patch.yml

@@ -0,0 +1,33 @@
+# dsh-doublecheck strict overlay: every discipline gate ON at `block` intensity.
+#
+# Apply this patch AFTER the bundle patch (later layers win by position), e.g.:
+#   dsh --profile <name> --patch ./strict.patch.yml
+# A patch replaces the row's WHOLE config — Schema defaults are NOT re-applied
+# on override — so every key the guard row needs is restated here. The grill
+# row keeps its defaults and is not restated.
+- id: doublecheck-guard
+  config:
+    intensity: block
+    modules:
+      grill: true
+      tdd: true
+      adversary: true
+    adversaryModel: null
+    adversaryProvider: fork
+    adversaryMaxFindings: 5
+    adversaryTools: ['read', 'glob', 'grep']
+    adversaryTimeoutMs: 120000
+    guardTools: ['edit', 'write']
+    vagueTaskMaxChars: 200
+    remindOnce: true
+    language: en
+    enableByDefault: true
+    testToolNames: ['bash', 'pwsh']
+    testCommandPatterns:
+      - '(?:^|[;&|]\s*)(?:(?:pnpm|npm|npx|yarn|bun)(?:\s+run)?\s+(?:test|vitest|jest|mocha)(?:\s|$))'
+      - '(?:^|[;&|]\s*)(?:(?:pytest|go\s+test|cargo\s+test|make\s+test|ctest)(?:\s|$))'
+      - '(?:^|[;&|]\s*)(?:node\s+--test(?:\s|$))'
+      - '(?:^|[;&|]\s*)(?:deno\s+test|uv\s+run\s+pytest)(?:\s|$)'
+    testFilePatterns:
+      - '(^|[\\/])(tests?|__tests__|specs?)([\\/]|$)'
+      - '\\.(test|spec)\\.[A-Za-z0-9]+$'