dsh-perm-gate 5.4.2 → 5.4.3

package/CHANGELOG.md

@@ -2,6 +2,14 @@
 
 ## Unreleased
 
+## 5.4.3 (2026-11-03)
+
+- **#6 fail-open 修复**：宿主把 volatile 字段以**引用**交给 `apply`，此前 `...config` 把引用原样展开进运行时选项；引用对象既不等于 `deny` 也不等于 `ask`，运行时两条分支都落空，未命中任何规则的调用直接放过——即使配的是 ask / 默认审查。现在构造前先 `resolveVolatileConfig` 解引用，`coerceRuleAction` 把非三值一律 fail-closed 成 `ask`（规则文档路径同样过一遍），`defaultAction` / `caseInsensitivePaths` 改由 live 钩子读取并纳入热更新 key，改设置后 `reload()` 生效（此前只有首轮播种，改值不生效）。
+- **#7 假 SAFE 修复**：host 审查只采纳 `text-delta`，`reasoning-delta`（模型内心独白）不再拼进待解析文本；`max-tokens` / `length` / `content-filter` 等截断收尾判为失败而不是短答案；`parseRiskVerdict` 删掉裸 `safe` 关键字兜底——“not safe”、半截 JSON、纯文本 `SAFE` 一律 unresolved，留在人工确认上。
+- **#8 host-review 集成修复**：(1) `reasoningEffort` 不再硬编码 `'off'`，改由新 volatile 配置 `classifierEffort` 控制（设置卡新增「推理强度」，留空＝不发送、交给宿主模型配置），被宿主拒绝时可选 low/off；(2) `POST /api/dsh-perm-gate/health` 曾用 `{ ...result, ok: true }` 把检查自己给出的 `ok:false` 覆盖成健康，现改为以检查结果为准；(3) host 模式清空 Model 真正回落到会话模型组——`classifierModel` 不再带 schema 默认值，空值不再被当成显式覆盖（custom 端点模式仍在请求时兜底 `deepseek-chat`）。
+
+## 5.4.2 (2026-11-02)
+
 - **拦截意见不再冒充用户消息**：被拦后给主模型的 `工具审核意见` 曾以 `mode:'queue'` 投到 next-turn——那是用户的输入队列，意见投进去会在 input-traffic 等待区里像用户自己排的一条，并与待发消息抢队列，且要等整个回合结束才送达（模型多反刍一轮）。现在交由 `buildBlockNoteSender` 投递：首选 `agent.inject()`（next-step 带内、不唤醒 driver，贴着被拒的 tool 结果在下一个 pre-step 送达）并标上自有 `source.kind='perm-gate'`，因为宿主的语义里只有 `source.kind==='user'`（带 `rpcId`）才是“人说的话”，其余一律按注入上下文渲染；没有活 agent 时才回退到 steer 车道（`session.prompt mode:'steer'`），绝不静默降级成 `followup`。yolo / 机械直通 / 降级审查 / 确定性 deny 四条拦截路径统一如此；评审页「撤销此改动」仍走 next-turn（那是人想发的一句话）。
 - 配套：`dsh-input-traffic` 同步把「插话中」限定为真人插话（`source.kind === 'user'`），宿主自有通知（如 `user-approval` 的审批策略变更）与第三方意见不再占位等待区；没带 `source` 的旧行仍照常显示，不会误伤真队列。
 

package/dsh.plugin.json

@@ -1,7 +1,7 @@
 {
   "id": "dsh-perm-gate",
   "name": "dsh-perm-gate",
-  "version": "5.4.2",
+  "version": "5.4.3",
   "description": {
     "en": "Deterministic-first, fail-closed permission gate for DeepSeek Harness 0.1.2+: a P0 hard-deny -> P1 precise session grant -> P2 static allow/deny rule chain -> P3 optional LLM classifier -> P4 ask pipeline over the tools/pre-execute waterfall, with a standalone CLI dry-run evaluator and an auditable decision log. Credential handling is a seven-family ladder (ssh/cloud/package/gitcreds/env/cli/db) with deny/ask/readonly/readwrite rungs kept in the rules document plus a paired write-exemption list; a protected-branch push switch keeps the P0 floor by default. It only vetoes with deny/ask and always delegates otherwise, and unknown or malformed rules fail loud instead of silently standing the gate down.",
     "zh": "面向 DeepSeek Harness 0.1.2+ 的确定性优先、失败即拒的权限闸门：在 tools/pre-execute 瀑布上串起 P0 硬拒绝 → P1 精确会话授权 → P2 静态 allow/deny 规则链 → P3 可选 LLM 语义分类 → P4 ask，并附带独立 CLI 干跑评估器与可审计的决策日志。凭据处理为七族阶梯（ssh/cloud/package/gitcreds/env/cli/db）× deny/ask/readonly/readwrite 四档，档位存规则文档 + 成对写豁免清单；保护分支强推开关默认保留 P0 地板。只以 deny/ask 否决、其余一律放行，规则未知或非法时直接报错，绝不静默失效。"

package/lib/client.js

@@ -78,6 +78,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "llmAssist 接收 LLM（OpenAI 兼容，可用自定义 API）",
 				"card.llmEndpoint": "Endpoint（BaseURL，如 https://api.openai.com/v1）",
 				"card.llmModel": "Model（模型 id，如 deepseek-chat）",
+				"card.llmEffort": "推理强度（仅 host 模式）",
+				"card.llmEffortHost": "跟随宿主默认",
+				"card.llmEffortHint": "留空＝不干预，由宿主模型配置决定；被宿主拒绝时改 low 或 off。",
 				"card.llmKey": "API Key（保密）",
 				"card.llmKeyHidden": "未设置 · 首次输入后以 * 显示",
 				"card.llmKeyMasked": "已设置 · 输入新值覆盖",
@@ -277,6 +280,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "llmAssist receiving LLM (OpenAI-compatible; any custom API)",
 				"card.llmEndpoint": "Endpoint (BaseURL, e.g. https://api.openai.com/v1)",
 				"card.llmModel": "Model (model id, e.g. deepseek-chat)",
+				"card.llmEffort": "Reasoning effort (host mode only)",
+				"card.llmEffortHost": "Follow the host default",
+				"card.llmEffortHint": "Empty = send no effort and let the host model config decide; switch to low or off if the host rejects it.",
 				"card.llmKey": "API Key (secret)",
 				"card.llmKeyHidden": "Unset · shown as * after first entry",
 				"card.llmKeyMasked": "Set · type a new value to overwrite",
@@ -476,6 +482,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "llmAssist 受信 LLM（OpenAI 互換、カスタム API 可）",
 				"card.llmEndpoint": "Endpoint（BaseURL、例：https://api.openai.com/v1）",
 				"card.llmModel": "Model（モデル id、例：deepseek-chat）",
+				"card.llmEffort": "推論強度（host モードのみ）",
+				"card.llmEffortHost": "ホスト既定に従う",
+				"card.llmEffortHint": "空 = 指定せずホストのモデル設定に任せる。ホストに拒否されたら low か off に変更。",
 				"card.llmKey": "API Key（機密）",
 				"card.llmKeyHidden": "未設定・初回入力後は * 表示",
 				"card.llmKeyMasked": "設定済み・新しい値を入力すると上書き",
@@ -675,6 +684,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "llmAssist 수신 LLM (OpenAI 호환, 사용자 지정 API 가능)",
 				"card.llmEndpoint": "Endpoint (BaseURL, 예: https://api.openai.com/v1)",
 				"card.llmModel": "Model (모델 id, 예: deepseek-chat)",
+				"card.llmEffort": "추론 강도 (host 모드 전용)",
+				"card.llmEffortHost": "호스트 기본값 따름",
+				"card.llmEffortHint": "비움 = 지정하지 않고 호스트 모델 설정을 따름. 호스트가 거부하면 low 나 off 로 변경.",
 				"card.llmKey": "API Key (비밀)",
 				"card.llmKeyHidden": "미설정 · 최초 입력 후 * 표시",
 				"card.llmKeyMasked": "설정됨 · 새 값을 입력하면 덮어씀",
@@ -874,6 +886,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "LLM récepteur llmAssist (compatible OpenAI ; API personnalisée possible)",
 				"card.llmEndpoint": "Endpoint (BaseURL, ex. https://api.openai.com/v1)",
 				"card.llmModel": "Model (identifiant du modèle, ex. deepseek-chat)",
+				"card.llmEffort": "Effort de raisonnement (mode host)",
+				"card.llmEffortHost": "Suivre l’hôte",
+				"card.llmEffortHint": "Vide = ne rien imposer, la config du modèle hôte décide ; passer à low ou off si l’hôte refuse.",
 				"card.llmKey": "API Key (secret)",
 				"card.llmKeyHidden": "Non défini · affiché comme * après la première saisie",
 				"card.llmKeyMasked": "Défini · saisir une nouvelle valeur pour écraser",
@@ -1073,6 +1088,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "llmAssist empfangendes LLM (OpenAI-kompatibel; jede benutzerdefinierte API)",
 				"card.llmEndpoint": "Endpoint (BaseURL, z. B. https://api.openai.com/v1)",
 				"card.llmModel": "Model (Modell-ID, z. B. deepseek-chat)",
+				"card.llmEffort": "Reasoning-Aufwand (nur Host-Modus)",
+				"card.llmEffortHost": "Host-Standard folgen",
+				"card.llmEffortHint": "Leer = nichts vorgeben, die Host-Modellkonfiguration entscheidet; auf low oder off wechseln, wenn der Host es ablehnt.",
 				"card.llmKey": "API Key (geheim)",
 				"card.llmKeyHidden": "Nicht gesetzt · nach erster Eingabe als * angezeigt",
 				"card.llmKeyMasked": "Gesetzt · neuen Wert eingeben zum Überschreiben",
@@ -1272,6 +1290,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "LLM ricevente llmAssist (compatibile OpenAI; qualsiasi API personalizzata)",
 				"card.llmEndpoint": "Endpoint (BaseURL, es. https://api.openai.com/v1)",
 				"card.llmModel": "Model (ID modello, es. deepseek-chat)",
+				"card.llmEffort": "Sforzo di ragionamento (solo modalità host)",
+				"card.llmEffortHost": "Segui l’host",
+				"card.llmEffortHint": "Vuoto = non imporre nulla, decide la configurazione del modello host; passa a low o off se l’host lo rifiuta.",
 				"card.llmKey": "API Key (segreta)",
 				"card.llmKeyHidden": "Non impostata · mostrata come * dopo il primo inserimento",
 				"card.llmKeyMasked": "Impostata · inserisci un nuovo valore per sovrascrivere",
@@ -1471,6 +1492,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "Принимающая LLM для llmAssist (совместимая с OpenAI; любой пользовательский API)",
 				"card.llmEndpoint": "Endpoint (BaseURL, напр. https://api.openai.com/v1)",
 				"card.llmModel": "Model (идентификатор модели, напр. deepseek-chat)",
+				"card.llmEffort": "Усилие рассуждения (только режим host)",
+				"card.llmEffortHost": "Как у хоста",
+				"card.llmEffortHint": "Пусто = не навязывать, решает конфигурация модели хоста; выберите low или off, если хост отклоняет.",
 				"card.llmKey": "API Key (секрет)",
 				"card.llmKeyHidden": "Не задан · после первого ввода отображается как *",
 				"card.llmKeyMasked": "Задан · введите новое значение для перезаписи",
@@ -1670,6 +1694,9 @@ window.__ModuleLoader__.load({
 				"card.llmReceiver": "LLM receptor de llmAssist (compatible con OpenAI; cualquier API personalizada)",
 				"card.llmEndpoint": "Endpoint (BaseURL, ej. https://api.openai.com/v1)",
 				"card.llmModel": "Model (id del modelo, ej. deepseek-chat)",
+				"card.llmEffort": "Esfuerzo de razonamiento (solo modo host)",
+				"card.llmEffortHost": "Seguir el host",
+				"card.llmEffortHint": "Vacío = no imponer nada, decide la configuración del modelo del host; cambia a low u off si el host lo rechaza.",
 				"card.llmKey": "API Key (secreta)",
 				"card.llmKeyHidden": "No establecida · se muestra como * tras la primera entrada",
 				"card.llmKeyMasked": "Establecida · introduce un nuevo valor para sobrescribir",
@@ -3165,6 +3192,43 @@ window.__ModuleLoader__.load({
 												})
 											]
 										}),
+										/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", {
+											style: fieldStyle,
+											children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+												style: fieldLabelStyle,
+												children: t("card.llmEffort")
+											}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("select", {
+												value: value.classifierEffort ?? "",
+												disabled: readonly || receiverSource !== "host",
+												style: controlStyle,
+												title: t("card.llmEffortHint"),
+												onChange: (event) => {
+													scope.set("classifierEffort", event.currentTarget.value);
+												},
+												children: [
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+														value: "",
+														children: t("card.llmEffortHost")
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+														value: "off",
+														children: "off"
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+														value: "low",
+														children: "low"
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+														value: "medium",
+														children: "medium"
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+														value: "high",
+														children: "high"
+													})
+												]
+											})]
+										}),
 										/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", {
 											style: fieldStyle,
 											children: [