@sema-agent/server 7.67.0 → 7.69.0

package/USAGE.md

@@ -124,6 +124,19 @@ MODEL_DEGRADE_ON=rate_limit,breaker_open   # 可选,反应式触发器子集;词
   `med,high` 此前静默窄成只剩 `high` 一档,而这根旋钮的全部用途就是宣告高档可用,静默落空 = 高档
   永远上不去。空串 / 纯分隔(`, `)仍与不设同义(不铸键,行为不变);anthropic 腿刻意不铸 compat
   (那条腿不吃 `reasoning_effort`),`GET /v1/config/catalog` 对它诚实回 `null`。
+- 🔴 **`MODEL_THINKING_FORMAT`(7.69.0 新,同族第三根)**:这台网关按哪种拼法收「开/关思考」——
+  `openai|openrouter|deepseek|together|zai|qwen|qwen-chat-template` 七词闭集(上 core
+  `OpenAICompletionsCompat.thinkingFormat`),未知词同样 **boot 拒启并点名闭集**。缺席 ⇒ 不铸这一键
+  (core 按 baseUrl / 模型 id 自行推断,行为逐字不变)。**真需求**:vLLM / Qwen 类**缺省开思考**的网关,
+  只有 `qwen-chat-template`(`chat_template_kwargs.enable_thinking=false`)关得掉,而 `MODEL_EXTRA_BODY`
+  穿不过引擎的 OpenAI 保留键表 —— 不设它,这类部署每一轮都白烧一段思考且读数上看不出来。它与
+  `MODEL_REASONING_EFFORT_LEVELS` 铸进**同一个** `compat` 对象(两根旋钮同时设两键都在)。
+- 🔴 **逐模型声明(目录腿):`models.models[].compat`**(settings-schema ≥1.10.0;7.69.0 消费)——
+  同一张 `OpenAICompletionsCompat` 五键(`thinkingFormat` / `reasoningEffortLevels` / `maxTokensField` /
+  `requiresReasoningContentOnAssistantMessages` / `supportsReasoningEffort`)可以**按模型**写在
+  `config.d/models.json` 或配置控制面的目录条目上,逐键盖过上面两根 env 缺省(BL-8:没写的键仍继承 env)。
+  **坏形/坏词/未知键 ⇒ 丢掉这条 `compat` 声明、模型本体保留**,并打一条 `model_compat_dropped`
+  点名 warn(引擎回落到自己的推断 = 今天的行为);一条写坏的声明**不会**让这只模型或整套目录消失。
 - 反应式降级 brain 包在**最外层**;fallback brain 用自己的凭据(decorator 清掉主模型的 per-call key 防外泄给别的 provider)。**坑**:`MODEL_DEGRADE_TO` 最好别和被限流的是同一网关/账号,否则反应式切过去照样撞同一个 rate_limit。
 - **定价怎么设**:`model.cost` 来自 `MODEL_COST_INPUT/OUTPUT/CACHE_READ/CACHE_WRITE`(**USD per 1M tokens**,默认 0 = `costUsd` 读 0)。云模型(如 review-gw 的 deepseek-v4-flash)必须设,否则 spend 恒为 $0;本地自托管(qwen)留 0 即对(无 per-token 外部花费)。字段名是 `costUsd`,**非美元计价的网关要先折算**(如 DeepSeek 官方 CNY ÷ 汇率)。配置控制面管的模型走 `CenterModel.cost`(中心存价、不存 secret)。
 - 成本计量**自动开**:从 config 的 `model.cost`(per-1M 绝对 USD)注入 `pricing`,core 算出权威的整数 `costMicroUsd`(避免浮点累计误差)。`/metrics` 新增:`model_cost_micro_usd_total{model}`(覆盖所有 brain 调用=主任务+异步+council 子任务的总花费)、`brain_first_token_ms`(网关 hang 早警)、`brain_call_latency_ms`、`tool_calls_total{name,ok}`、`budget_exceeded_total{code}`、`cost_quota_rejected_total`、`degraded_total{reason}`(1.40 降级)。
@@ -638,8 +651,12 @@ READ_DENY_PATTERNS='[{"pattern":"**/secrets/**"},{"pattern":".ssh","caseSensitiv
   ① **本机 env 显式恒赢**(逐键判)——env 是这台机器的部署主权,下发值不会静默盖掉它,被忽略的键留一条
   `sema_registry_read_face_env_wins` warn;② **坏值整域拒**——任何一个成员过不了引擎自己的校验器就整域不落、
   保留 env 前值 + `sema_registry_read_face_invalid` warn(半应用一份 READ 姿态比不应用更危险);
-  ③ **restart-to-apply**——四键在 boot 期被引擎装配捕获,下发改动经 `/health` 的 `restart.reasons`
-  带 `read-face` 通告,由编排器重启兑现。当前生效档可从 `GET /v1/capabilities` 的 `readFace` 位读
+  ③ **一席热、三席重启**(server 7.69.0 起,引擎侧交付了 read-face 的活体座席):`face` 是**热换席** ——
+  下发或撤回一个档位,下一拍 refresh(≤60s,或手动 `POST /v1/admin/config/refresh`)当场换进全部 Runner,
+  **不重启、也不再签 `read-face` 重启理由**(已在飞的任务保留它准备时的档位,新任务读新档);三个 deny 键
+  (`denyPatterns`/`denyBuiltinTiers`/`denyBuiltinExclude`)**仍是 boot-baked**(core 的可换席闭集只收了
+  `readFace` 一席),下发改动仍经 `/health` 的 `restart.reasons` 带 `read-face` 通告、由编排器重启兑现。
+  当前生效档可从 `GET /v1/capabilities` 的 `readFace` 位读
   (`"open"`/`"roots"`/`null`=本部署未钉,引擎默认当家;deny 表内容属策略内容,**不上**能力面);
   **来源**(env / center / posture / engine-default)在 operator 面 `GET /v1/diagnostics/wiring` 的 `readFace.source`。
   ⚠️ 本地 `config.d/` 车道今天**接不上**这一域(registry-core 的可移植域表里没有它),只有

package/dist/adoption/plan.js

@@ -223,6 +223,17 @@ export const REBIND_LEGS = [
         residualKey: undefined,
         why: "resume 重放日志的 scope 守卫列 —— 与 run 行**同事务**(错配会让 load 守卫拒掉整条 resume)",
     },
+    {
+        leg: "workflow_agent_session#scope",
+        table: "workflow_agent_session",
+        kind: "bulk-rebind",
+        action: "bucket-rebind",
+        columns: ["scope"],
+        matchColumn: "scope",
+        encoding: "verbatim",
+        residualKey: undefined,
+        why: "S-185 workflow 出身 park 的 join 索引行的租户列(值 = 那条 run 的 `scope`,与 workflow_run/journal 同一个字节)——与它们**同族同事务**迁:不迁 ⇒ 索引行留在旧身份名下,而 run 行已迁,一次收编后人对这条 park 的批准结构上投递不出去",
+    },
     {
         leg: "workflow_resume_claim#scope",
         table: "workflow_resume_claim",

package/dist/approval-card.d.ts

@@ -328,6 +328,26 @@ declare const DenialLimitFallbackSchema: z.ZodObject<{
     autoDenyAfterMs: z.ZodNumber;
 }, z.core.$strict>;
 export type DenialLimitFallback = z.infer<typeof DenialLimitFallbackSchema>;
+/** S-185(core 7.10.0 [ref])—— **分类器为什么答不了**。`cause` 是 core 的闭三词
+ *  (`AUTO_MODE_UNAVAILABLE_CAUSES` = error | timeout | breaker_open),本仓**刻意不枚举**它:词表的单一
+ *  属主在 core,抄一份的形会在 core 加词那天把新词静默吞成缺席 —— 而丢的正是「这次不可用是新出现的那
+ *  一类」这条信息(与 `AskRequest.origin` 顶注同一条纪律)。判据只到「非空串」这一层形门。
+ *  `.strict()` + 单成员必填:core 将来给这只对象加成员时本读面当场判假、**整键不铸**,新成员必须由人
+ *  处置过才上 wire;而一个**没有 cause** 的空壳等于说「分类器坏了但说不出为什么」,比不说更坏。 */
+declare const ClassifierUnavailableSchema: z.ZodObject<{
+    cause: z.ZodString;
+}, z.core.$strict>;
+export type ClassifierUnavailable = z.infer<typeof ClassifierUnavailableSchema>;
+/**
+ * `AskRequest.classifierUnavailable`(core 7.10.0 [ref])的**边界窄读** —— 活卡帧与 `card_json` 的唯一
+ * 铸造点(与 {@link readDenialLimitFallback} 同款分工),两面结构性同值。
+ *
+ * 🔴 **缺席不是断言**:缺席同时覆盖「分类器答上了」「这只 ask 没资格走分类器」「本部署没接分类器」
+ * 三形(core d.ts 逐字:read presence, never absence),消费端禁读成「分类器好着呢」。
+ * 🔴 **echo-only**:server 不据它做任何裁决 —— 熔断/超时的处置全在 core 的分类器站,据本键在本仓自铸
+ * 第二套回落判据就是同一语义面两个写者(源头修复纪律)。
+ */
+export declare function readClassifierUnavailable(req: unknown): ClassifierUnavailable | undefined;
 /**
  * `AskRequest.denialLimitFallback`(core 7.4.0 [ref])的**边界窄读** —— 活卡帧与 `card_json` 的唯一铸造点
  * (与 {@link readRuleOffersAbsence} / {@link readProbeCause} 同款分工),两面结构性同值。
@@ -422,6 +442,9 @@ export declare const ApprovalCardSchema: z.ZodObject<{
         }>;
         autoDenyAfterMs: z.ZodNumber;
     }, z.core.$strict>>;
+    classifierUnavailable: z.ZodOptional<z.ZodObject<{
+        cause: z.ZodString;
+    }, z.core.$strict>>;
     probeCause: z.ZodOptional<z.ZodObject<{
         code: z.ZodString;
         roots: z.ZodObject<{
@@ -532,6 +555,9 @@ export declare const ApprovalCardEnvelopeSchema: z.ZodObject<{
             }>;
             autoDenyAfterMs: z.ZodNumber;
         }, z.core.$strict>>;
+        classifierUnavailable: z.ZodOptional<z.ZodObject<{
+            cause: z.ZodString;
+        }, z.core.$strict>>;
         probeCause: z.ZodOptional<z.ZodObject<{
             code: z.ZodString;
             roots: z.ZodObject<{

package/dist/approval-card.js

@@ -237,6 +237,14 @@ const DenialLimitFallbackSchema = z
 const _denialLimitFallbackConformance = [true, true];
 void _denialLimitFallbackConformance;
 const DenialLimitFallbackEnvelopeSchema = z.object({ denialLimitFallback: DenialLimitFallbackSchema.optional() });
+const ClassifierUnavailableSchema = z.object({ cause: z.string().min(1).max(MAX_IDENT) }).strict();
+const _classifierUnavailableConformance = [true];
+void _classifierUnavailableConformance;
+const ClassifierUnavailableEnvelopeSchema = z.object({ classifierUnavailable: ClassifierUnavailableSchema.optional() });
+export function readClassifierUnavailable(req) {
+    const parsed = ClassifierUnavailableEnvelopeSchema.safeParse(req);
+    return parsed.success ? parsed.data.classifierUnavailable : undefined;
+}
 export function readDenialLimitFallback(req) {
     const parsed = DenialLimitFallbackEnvelopeSchema.safeParse(req);
     return parsed.success ? parsed.data.denialLimitFallback : undefined;
@@ -261,6 +269,7 @@ export const ApprovalCardSchema = z
     ruleOffers: z.array(RuleOfferSchema).max(MAX_RULE_OFFERS).optional(),
     ruleOffersAbsence: RuleOffersAbsenceSchema.optional(),
     denialLimitFallback: DenialLimitFallbackSchema.optional(),
+    classifierUnavailable: ClassifierUnavailableSchema.optional(),
     probeCause: ProbeCauseSchema.optional(),
     ruleEvidence: RuleEvidenceSchema.optional(),
     fromSubagent: z.literal(true).optional(),
@@ -303,6 +312,7 @@ export function buildApprovalCard(source, req, requiresRealApproval) {
     const ruleEvidence = readRuleEvidence(req);
     const ruleOffersAbsence = readRuleOffersAbsence(req);
     const denialLimitFallback = readDenialLimitFallback(req);
+    const classifierUnavailable = readClassifierUnavailable(req);
     const toolCallId = clip(source.toolCallId, MAX_IDENT);
     const sourceTaskId = clip(source.sourceTaskId, MAX_IDENT);
     const sourceAgentName = clip(source.sourceAgentName, MAX_AGENT_NAME);
@@ -325,6 +335,7 @@ export function buildApprovalCard(source, req, requiresRealApproval) {
         ...(source.ruleOffers !== undefined && source.ruleOffers.length > 0 ? { ruleOffers: source.ruleOffers.map(copyRuleOffer) } : {}),
         ...(ruleOffersAbsence !== undefined ? { ruleOffersAbsence } : {}),
         ...(denialLimitFallback !== undefined ? { denialLimitFallback } : {}),
+        ...(classifierUnavailable !== undefined ? { classifierUnavailable } : {}),
         ...(source.fromSubagent === true ? { fromSubagent: true } : {}),
         ...(sourceTaskId !== undefined ? { sourceTaskId } : {}),
         ...(sourceAgentName !== undefined ? { sourceAgentName } : {}),

package/dist/approval.js

@@ -1,4 +1,5 @@
 import { isNamespacedCoveringRuleName, namespacedRuleNameCovers } from "@sema-agent/core";
+import { operatorAsk } from "./operator-ask.js";
 import { isExplicitPosture } from "./posture-source.js";
 export function hasOperatorGateIntent(config) {
     return (config.approvalRequire.length > 0 ||
@@ -51,11 +52,11 @@ export function createDurableAskPolicy(opts) {
             if (deny(toolName))
                 return { action: "deny", message: `tool "${req.toolName}" is denied by policy` };
             if (neverAuto(toolName))
-                return { action: "ask", message: `human approval required to run "${req.toolName}"` };
+                return operatorAsk(`APPROVAL_NEVER_AUTO(${toolName})`, `human approval required to run "${req.toolName}"`);
             if (require(toolName)) {
                 const remEarly = req.budget?.resourceRemainingMicroUsd;
                 if (remEarly !== undefined && remEarly <= 0) {
-                    return { action: "ask", message: `durable resource budget exhausted — human approval required to run "${req.toolName}"` };
+                    return operatorAsk("RESOURCE_SUSPEND budget exhausted", `durable resource budget exhausted — human approval required to run "${req.toolName}"`);
                 }
                 if (opts.exempt && toolName !== "AskUserQuestion") {
                     return (async () => {
@@ -81,13 +82,13 @@ export function createDurableAskPolicy(opts) {
     function decideRequired(req, toolName) {
         const rem = req.budget?.resourceRemainingMicroUsd;
         if (rem !== undefined && rem <= 0) {
-            return { action: "ask", message: `durable resource budget exhausted — human approval required to run "${req.toolName}"` };
+            return operatorAsk("RESOURCE_SUSPEND budget exhausted", `durable resource budget exhausted — human approval required to run "${req.toolName}"`);
         }
         if (toolName !== "AskUserQuestion" && budget > 0 && autoApproved < budget) {
             autoApproved++;
             return { action: "allow" };
         }
-        return { action: "ask", message: `approval required to run "${req.toolName}"` };
+        return operatorAsk(`APPROVAL_REQUIRE(${toolName})`, `approval required to run "${req.toolName}"`);
     }
 }
 //# sourceMappingURL=approval.js.map

package/dist/boot/config-center.d.ts

@@ -1,4 +1,4 @@
-import { CenterPromptSource, type Model } from "@sema-agent/core";
+import { CenterPromptSource, type SwappableDeps } from "@sema-agent/core";
 import { buildPricing } from "../budget.js";
 import { type PromptsDomainFaces } from "../prompts-domain-validate.js";
 import { type Scenario, type ScenarioDeps, type ScenarioDetail } from "../capabilities/scenarios.js";
@@ -42,16 +42,15 @@ export interface ConfigCenterRuntime {
     /** 60s refresh cadence + boot-deferred 到货续接。
      *  `limitSync`([ref] 批1)= 恒构造的两只限额座的换代口:commit 之后**零 await 内**推一把,与 pricing /
      *  keyResolver 那两件派生重建同族(缺席 = 这条腿本部署没装配,例如 run-local)。
-     *  `swapRunnerModels`(core 5.49,[ref]①)= 模型面永久热的换装口:plane 变更候选在 commit 前喂给
-     *  全部 boot Runner 原子换代;在场 ⇒ models-tiers defer/restart 臂退役,缺席 ⇒ 旧 defer 行为逐字保留。 */
+     *  `swapRunnerDeps`(core 7.10.0 `Runner.swapDeps` 一扇门;5.49 起的 `swapModels` 口在 7.10.0 退役、
+     *  本参数同语义改名)= **可换席**的换装口:模型面 plane 变更候选在 commit 前喂给全部 boot Runner
+     *  原子换代(在场 ⇒ models-tiers defer/restart 臂退役,缺席 ⇒ 旧 defer 行为逐字保留);S-174 起
+     *  read **face** 席也走同一只函数(见 refresh 拍里的 `applyCenterReadFace` 调用点)。 */
     startRefreshLoop(a: {
         runnerTierFrozen: boolean;
         pricing: ReturnType<typeof buildPricing>;
         limitSync?: LimitSync;
-        swapRunnerModels?: (plane: {
-            models: Record<string, Model>;
-            tiers: Record<string, string>;
-        }) => void;
+        swapRunnerDeps?: (next: SwappableDeps) => void;
     }): void;
     /** [ref]-2:停刷新环(hardShutdown 收尾链;幂等,未起环时 no-op)。 */
     stopRefreshLoop(): void;

package/dist/boot/config-center.js

@@ -328,7 +328,8 @@ export async function createConfigCenterRuntime(ctx) {
             }
         }
     }
-    const readFaceEnvHeld = applyCenterReadFace(config, effective ?? { version: 0, updatedAt: "", models: { models: [], roles: {} } }, effective ? logger : undefined);
+    const readFaceBaseline = { ...(config.readFace !== undefined ? { readFace: config.readFace } : {}), readFaceSource: config.readFaceSource };
+    const readFaceEnvHeld = applyCenterReadFace(config, effective ?? { version: 0, updatedAt: "", models: { models: [], roles: {} } }, effective ? logger : undefined).envHeld;
     if (!modelReadyState.ready)
         logger.warn("model_roster_pending", {
             note: "CONFIG_REQUIRE_ROSTER=true and no center roster has landed yet — the env MODEL_ID is treated as a PLACEHOLDER, so billable submissions 503 (and /health reports ready:false) until the first effective pull lands ≥1 enabled model. Unset the knob if this worker's env model is authoritative.",
@@ -341,6 +342,26 @@ export async function createConfigCenterRuntime(ctx) {
         ...(readFaceEnvHeld.length > 0 ? { readFaceEnvHeld } : {}),
         ...(mcpRevocationWired && mcpRevocations.hasBaseline() ? { mcpRevocationLive: true } : {}),
     });
+    const applyLiveReadFace = (eff, swap, phase) => {
+        const before = config.readFace;
+        const beforeSource = config.readFaceSource;
+        const { faceChanged } = applyCenterReadFace(config, eff, logger, { phase: "live", baseline: readFaceBaseline });
+        if (!faceChanged)
+            return;
+        if (swap === undefined) {
+            logger.warn("read_face_swap_unwired", { face: config.readFace ?? null, phase, note: "the center's read FACE changed but no Runner swap door was wired — the value landed in config while the engine keeps the seat it booted with" });
+            return;
+        }
+        try {
+            swap({ readFace: config.readFace });
+            logger.info("read_face_swapped", { from: before ?? null, to: config.readFace ?? null, phase, note: "hot seat (core swapDeps) — legs that already prepared keep the face they prepared under" });
+        }
+        catch (err) {
+            config.readFace = before;
+            config.readFaceSource = beforeSource;
+            throw err;
+        }
+    };
     let ccTimer;
     return {
         providerKind: configProvider?.kind,
@@ -489,7 +510,7 @@ export async function createConfigCenterRuntime(ctx) {
             }
         },
         startRefreshLoop(a) {
-            const { runnerTierFrozen, pricing, limitSync, swapRunnerModels } = a;
+            const { runnerTierFrozen, pricing, limitSync, swapRunnerDeps } = a;
             if (configProvider) {
                 const ccRef = config.configCenter;
                 let refreshInFlight = false;
@@ -555,7 +576,7 @@ export async function createConfigCenterRuntime(ctx) {
                             }
                             else {
                                 const planeChanged = modelPlaneChanged(appliedPlaneEff, r.effective);
-                                const planeDeferred = swapRunnerModels === undefined && (runnerTierFrozen || planeHasActiveTiers(r.effective)) && planeChanged;
+                                const planeDeferred = swapRunnerDeps === undefined && (runnerTierFrozen || planeHasActiveTiers(r.effective)) && planeChanged;
                                 let applyReport;
                                 const committed = applyEffective(config, r.effective, logger, {
                                     teamsOnly: true,
@@ -566,7 +587,7 @@ export async function createConfigCenterRuntime(ctx) {
                                         applyLedger.recordApply(rep);
                                     },
                                     ...(planeDeferred ? { deferModelPlane: true } : {}),
-                                    ...(swapRunnerModels !== undefined && planeChanged ? { swapPlane: swapRunnerModels } : {}),
+                                    ...(swapRunnerDeps !== undefined && planeChanged ? { swapPlane: swapRunnerDeps } : {}),
                                 });
                                 if (!committed)
                                     return "ok";
@@ -574,6 +595,7 @@ export async function createConfigCenterRuntime(ctx) {
                                 keyResolver = createKeyResolver(config.modelApiKeyEnv, process.env, config.modelApiKeys);
                                 limitSync?.sync();
                                 mcpRevocations.applyEnabledEntries(enabledMcpEntries(r.effective.mcp));
+                                applyLiveReadFace(r.effective, swapRunnerDeps, "refresh");
                                 applyLedger.applied(tickGen, r.effective.version);
                                 if (planeDeferred) {
                                     logger.warn("models_tiers_plane_deferred", { version: r.effective.version, note: "tier-frozen Runner: the changed model plane (models/roles/tiers/default) is NOT hot-applied — admission stays on the Runner's generation; restart applies the new plane (models-tiers restart signal rides /health)" });
@@ -596,7 +618,7 @@ export async function createConfigCenterRuntime(ctx) {
                                 }
                                 const lkgPersisted = await persistLkgDurable(r.effective, r.etag);
                                 const reasons = restartReasons(effective, r.effective, restartCtx());
-                                if (swapRunnerModels !== undefined) {
+                                if (swapRunnerDeps !== undefined) {
                                     const mi = reasons.indexOf("models-tiers");
                                     if (mi >= 0)
                                         reasons.splice(mi, 1);
@@ -701,7 +723,7 @@ export async function createConfigCenterRuntime(ctx) {
                         return;
                     }
                     const planeChangedLate = modelPlaneChanged(appliedPlaneEff, r.effective);
-                    const planeDeferredLate = swapRunnerModels === undefined && (runnerTierFrozen || planeHasActiveTiers(r.effective)) && planeChangedLate;
+                    const planeDeferredLate = swapRunnerDeps === undefined && (runnerTierFrozen || planeHasActiveTiers(r.effective)) && planeChangedLate;
                     let applyReportLate;
                     const committedLate = applyEffective(config, r.effective, logger, {
                         teamsOnly: true,
@@ -712,13 +734,14 @@ export async function createConfigCenterRuntime(ctx) {
                             applyLedger.recordApply(rep);
                         },
                         ...(planeDeferredLate ? { deferModelPlane: true } : {}),
-                        ...(swapRunnerModels !== undefined && planeChangedLate ? { swapPlane: swapRunnerModels } : {}),
+                        ...(swapRunnerDeps !== undefined && planeChangedLate ? { swapPlane: swapRunnerDeps } : {}),
                     });
                     if (!committedLate)
                         return;
                     mutateInPlace(pricing, buildPricing(config.models));
                     keyResolver = createKeyResolver(config.modelApiKeyEnv, process.env, config.modelApiKeys);
                     limitSync?.sync();
+                    applyLiveReadFace(r.effective, swapRunnerDeps, "boot-deferred");
                     applyLedger.applied(lateGen, r.effective.version);
                     if (planeDeferredLate)
                         logger.warn("models_tiers_plane_deferred", { version: r.effective.version, note: "tier-frozen Runner (env tiers): the late-boot center model plane is NOT hot-applied — restart applies it" });
@@ -743,7 +766,7 @@ export async function createConfigCenterRuntime(ctx) {
                     ccEtag = r.etag;
                     const lkgPersistedLate = await persistLkgDurable(r.effective, r.etag);
                     const reasons = restartReasons(undefined, r.effective, restartCtx());
-                    if (swapRunnerModels !== undefined) {
+                    if (swapRunnerDeps !== undefined) {
                         const miLate = reasons.indexOf("models-tiers");
                         if (miLate >= 0)
                             reasons.splice(miLate, 1);

package/dist/boot/leader.js

@@ -50,7 +50,7 @@ export function createLeaderFace(ctx) {
                 lockedConfig: governanceSeams.lockedConfig,
                 retentionPolicy: governanceSeams.retentionPolicy,
             },
-            deploymentPosture: buildDeploymentPostureSeats(config),
+            deploymentPosture: () => buildDeploymentPostureSeats(config),
             ...(ctx.tracer ? { tracer: ctx.tracer } : {}),
             ...(config.usageWindows && ctx.usageWindowStore
                 ? { usageWindows: config.usageWindows, usageWindowStore: ctx.usageWindowStore }

package/dist/boot/resolve-spec.js

@@ -367,7 +367,7 @@ export function createResolveSpec(ctx) {
                 return merged && config.toolDeferLongtail ? applyLongtailDefer(merged, true) : merged;
             })(cap.tools),
             skills: mergeUserSkills(cap.skills, body.skills, logger),
-            mcp: resolveRequestMcp(mcpForScenario(config.mcpServers, scenarioName), body.mcpServers, config, logger),
+            mcp: resolveRequestMcp(mcpForScenario(config.mcpServers, scenarioName), body.mcpServers, config, logger, opts?.onNotice !== undefined ? { sessionId: auth?.sessionId, emit: opts.onNotice } : undefined),
             a2a: resolveRequestA2a(a2aForScenario(config.a2aPeers, scenarioName), body.a2aPeers, config, logger),
             promptProvider: centerDecls ? centerPromptProvider(centerDecls, scenarioName) : cap.promptProvider,
             toolPolicy: toolPolicyForHands(hands, durableEnabled

package/dist/boot/workflow-orchestration.d.ts

@@ -14,6 +14,7 @@
 import { type WorkflowRun, type WorkflowRunStore } from "@sema-agent/core";
 import type { ServiceConfig } from "../config.js";
 import { FleetEventBus } from "../fleet/fleet-bus.js";
+import { type WorkflowAgentSessionIndex } from "../orchestration/workflow-agent-session-index.js";
 import type { Logger } from "../observability/logger.js";
 import type { Metrics } from "../observability/metrics.js";
 import { type WorkflowCompletionInbox } from "../orchestration/workflow-completion-inbox.js";
@@ -47,5 +48,6 @@ export declare function createWorkflowOrchestration(ctx: WorkflowOrchestrationCt
         finalizeStartedBeforeMs?: number;
         publishTerminalFleetRow?: (id: string, run: WorkflowRun) => void;
     };
+    workflowAgentSessionIndex: WorkflowAgentSessionIndex | undefined;
 };
 //# sourceMappingURL=workflow-orchestration.d.ts.map

package/dist/boot/workflow-orchestration.js

@@ -1,6 +1,7 @@
 import { FileWorkflowRunStore, InMemoryWorkflowRunStore } from "@sema-agent/core";
 import { join } from "node:path";
 import { FleetEventBus } from "../fleet/fleet-bus.js";
+import { LocalWorkflowAgentSessionIndex } from "../orchestration/workflow-agent-session-index.js";
 import { FileWorkflowCompletionInbox, InMemoryWorkflowCompletionInbox, resolveCompletionRoute } from "../orchestration/workflow-completion-inbox.js";
 import { FileWorkflowNotifyJournalStore, JournalingWorkflowRunStore, WorkflowNotifyGate } from "../orchestration/workflow-notify-journal.js";
 import { SubagentSteerRegistry } from "../orchestration/subagent-steer.js";
@@ -54,8 +55,15 @@ export function createWorkflowOrchestration(ctx) {
         })
         : undefined;
     const fleetBus = new FleetEventBus();
+    const workflowAgentSessionIndex = config.selfOrchestrationEnabled
+        ? config.workflowRunStoreBackend === "memory"
+            ? new LocalWorkflowAgentSessionIndex()
+            : (sqlWorkflowRunStore
+                ? backend.workflowAgentSessionIndex()
+                : new LocalWorkflowAgentSessionIndex(join(config.localDataRoot ?? localRoot, "workflows"), undefined, (msg, meta) => logger.warn(msg, meta)))
+        : undefined;
     const workflowRunStore = baseWorkflowRunStore
-        ? new JournalingWorkflowRunStore(baseWorkflowRunStore, workflowNotifyGate, fleetBus)
+        ? new JournalingWorkflowRunStore(baseWorkflowRunStore, workflowNotifyGate, fleetBus, workflowAgentSessionIndex)
         : undefined;
     const replicaLocalNotifyJournal = workflowNotifyJournal instanceof FileWorkflowNotifyJournalStore;
     const workflowRecoverOpts = {
@@ -70,7 +78,7 @@ export function createWorkflowOrchestration(ctx) {
     return {
         sqlWorkflowRunStore, workflowNotifyJournal, workflowCompletionInbox, deliverWorkflowCompletion,
         workflowNotifyGate, fleetBus, workflowRunStore, workflowJournalStore, outcomeSink,
-        workflowAgentRegistry, subagentSteerRegistry, workflowRecoverOpts,
+        workflowAgentRegistry, subagentSteerRegistry, workflowRecoverOpts, workflowAgentSessionIndex,
     };
 }
 //# sourceMappingURL=workflow-orchestration.js.map

package/dist/budget.js

@@ -119,7 +119,7 @@ export function createTracer(metrics, costQuota, modelUsage, fleetUsage, fleetLe
         else if (e.kind.startsWith("compaction.")) {
             metrics.inc("compaction_events_total", { outcome: e.kind.slice("compaction.".length), trigger: e.trigger ?? "none" });
         }
-        else if (e.kind === "permission.persisted_rule_allowed" || e.kind === "permission.rule_store_unreadable") {
+        else if (e.kind === "permission.persisted_rule_allowed" || e.kind === "permission.rule_store_unreadable" || e.kind === "permission.read_only_allowed") {
             metrics.inc("permission_rule_events_total", { event: e.kind.slice("permission.".length) });
         }
     };

package/dist/config-catalog.js

@@ -1,4 +1,6 @@
 import { CROSS_SESSION_INBOUND_SETTINGS, CROSS_SESSION_DIALOG_EXPIRY_VALUES } from "@sema-agent/core";
+import { THINKING_FORMATS } from "./model-compat.js";
+import { CENTER_READ_FACE_LIVE_KEYS } from "./config-center/read-face.js";
 export const CATALOG_VALUE_TYPES = ["boolean", "number", "string", "enum", "csv", "json", "path", "url", "secret", "opaque"];
 export const CATALOG_TIMINGS = ["hot", "restart", "deploy", "inert"];
 export const CATALOG_DANGER_AXES = ["security", "billing", "data"];
@@ -24,7 +26,14 @@ const cLimits = (path, over = {}) => ({
 });
 const cGov = (path, over = {}) => ({ domain: "governance", path, centerKey: path, precedence: "center-wins", timing: "hot", ...over });
 const cModels = (path) => ({ domain: "models", path, centerKey: "models", precedence: "center-wins", timing: "models-plane" });
-const cRead = (path, centerKey) => ({ domain: "readFace", path, centerKey, precedence: "env-wins", timing: "restart", restartSlice: "read-face" });
+const cRead = (path, centerKey, over = {}) => ({
+    domain: "readFace",
+    path,
+    centerKey,
+    precedence: "env-wins",
+    ...(CENTER_READ_FACE_LIVE_KEYS.has(path) ? { timing: "hot" } : { timing: "restart", restartSlice: "read-face" }),
+    ...over,
+});
 const SEC = ["security"];
 const BILL = ["billing"];
 const DATA = ["data"];
@@ -284,6 +293,16 @@ export const CONFIG_CATALOG = [
         },
     }),
     r("MODEL_ROUTER_ID", "modelPlane", "string", "leader 路由分类模型覆写", { derivedDefaultNote: "缺席 ⇒ MODEL_CHEAP_ID,再退 default", danger: BILL }),
+    r("MODEL_THINKING_FORMAT", "modelPlane", "enum", "本网关收「开/关思考」的拼法(上 core OpenAICompletionsCompat.thinkingFormat;缺省开思考的 vLLM/Qwen 网关只有 qwen-chat-template 关得掉;anthropic 腿不铸此键;未知词拒启)", {
+        enumValues: [...THINKING_FORMATS],
+        derivedDefaultNote: "缺席 ⇒ 不铸此键,core 按 baseUrl/模型 id 自行推断(默认 openai 的 reasoning_effort)",
+        resolve: (c) => {
+            const compat = c.model.compat;
+            if (compat === undefined || !("thinkingFormat" in compat))
+                return null;
+            return compat.thinkingFormat ?? null;
+        },
+    }),
     r("MODEL_VERIFIER", "modelPlane", "string", "verify 门的异构 verifier(须命中目录;typo 回 default 并 boot warn=自评自批)", { derivedDefaultNote: "缺席/不在目录 ⇒ default(=implementer,去相关性失效警告)" }),
     r("MODEL_VISION", "modelPlane", "boolean", "主模型 vision 声明(false 让 model_no_vision 预检真拦)", { staticDefault: "true", resolve: (c) => (Array.isArray(c.model.input) ? c.model.input.includes("image") : true), center: cModels("models[].vision") }),
     r("MYSQL_DATABASE", "store", "string", "MySQL 协议库名(选中该引擎时必填)", { danger: DATA }),

package/dist/config-center/apply-effective.d.ts

@@ -1,3 +1,10 @@
+/**
+ * EffectiveConfig application — mutate the boot `ServiceConfig` OVER with a pulled `EffectiveConfig`
+ * (models/roles/tiers/projects `applyEffective`, the boot-only approval gate + runtime limit residue
+ * `applyRuntimeGates`, the per-request-hot governance baton `applyRuntimeHot`) plus their shared helpers
+ * (`toModel`, `mutateInPlace`, `resolveDefaultModelName`, `logEffectiveDiff`). Split out of `facade.ts`
+ * ([ref] A13, internal-lossless — the facade re-exports every symbol below unchanged).
+ */
 import type { Model } from "@sema-agent/core";
 import type { ServiceConfig } from "../config-types.js";
 import { type SealedKeyOpener } from "../sealed-key.js";
@@ -23,7 +30,7 @@ export declare function mutateInPlace<V>(target: Record<string, V>, source: Reco
  * 🔴 只活在内存判定面:不进 wire(`/v1/models` 只发布尔)、不进日志(warn 打的是原始坏值的类型标记)。
  */
 export declare const UNPARSABLE_AT_MODEL_ALLOWLIST_ENTRY = "\0<unparsable-at-model-allowlist-entry>";
-/** 返回值=是否 COMMIT(false 两臂:[ref]③ 世代序 CAS 拒绝 / 5.49 swapPlane 换装拒绝——两臂都在
+/** 返回值=是否 COMMIT(false 两臂:[ref]③ 世代序 CAS 拒绝 / 7.10.0 swapPlane 换装拒绝——两臂都在
  *  commit 之前,活配置零触碰)。caller 收到 false 必须把**同一世代的旁路消费**(prompts 采用/pricing/
  *  keyResolver/etag 推进/LKG 落盘)一并跳过——否则 applyEffective 拒了主面、旁路却半应用同一个被拒
  *  世代,混合世代从侧门回来。首次 apply(live 未登记)恒 true。 */

package/dist/config-center/apply-effective.js

@@ -1,16 +1,29 @@
-import { isThinkingLevel } from "@sema-agent/core";
 import { resolveActiveTiers, SEALED_BOX_ALG } from "@sema-agent/settings-schema";
 import { sealedKeyPoison } from "../sealed-key.js";
 import { applyAutoCompactWindow, findUnmatchableToolNames, formatUnmatchableToolNames, parseAutonomy } from "../config.js";
+import { buildModelCompat, readModelCompat } from "../model-compat.js";
 import { validateCommandRules } from "../runtime-governance.js";
 import { projectCollabToWorkflows, registerCollabWorkflows } from "../capabilities/collab-workflows.js";
 import { stageLimits } from "./stage-limits.js";
 import { BATCH1_LIMIT_KEYS } from "./hot-keys-registry.js";
-const asModelThinking = (v) => (v && isThinkingLevel(v) && v !== "off" ? v : undefined);
 const charsPerTokenOf = (v) => (typeof v === "number" && Number.isFinite(v) && v > 0 ? v : undefined);
-function toModel(m, envDefaults = {}) {
-    const thinkingDefault = asModelThinking(m.defaultThinking) ?? envDefaults.defaultThinking;
-    const effortLevels = (m.reasoningEffortLevels ? m.reasoningEffortLevels.map(asModelThinking).filter((x) => x !== undefined) : undefined) ?? envDefaults.reasoningEffortLevels ?? [];
+function envDerivedModel(config) {
+    const m = config.envModel;
+    if (m === undefined) {
+        throw new Error("config.envModel is missing — it is the BL-8 inheritance baseline (the boot env-derived model) and is REQUIRED on ServiceConfig. " +
+            "A hand-rolled `as ServiceConfig` fixture must set it to the same object as `model`; there is deliberately no fallback to `config.model` (that fallback IS the bug this seat removes).");
+    }
+    return m;
+}
+function toModel(m, envDefaults = {}, logger) {
+    const thinkingDefault = envDefaults.defaultThinking;
+    const declared = readModelCompat(m.compat);
+    if (!declared.ok)
+        logger?.warn("model_compat_dropped", { model: m.name, issue: declared.issue, hint: "the per-model wire-compat declaration was dropped whole — the engine falls back to its own inference (today's behaviour); fix the entry to make it live" });
+    else if (declared.compat !== undefined && m.provider === "anthropic") {
+        logger?.warn("model_compat_dropped", { model: m.name, issue: "declared on an anthropic model", hint: "`compat` mirrors OpenAICompletionsCompat — the Anthropic brain reads AnthropicMessagesCompat and never sees these keys; the declaration is dropped" });
+    }
+    const compat = m.provider === "anthropic" ? undefined : buildModelCompat(declared.ok ? declared.compat : undefined, envDefaults.compat);
     const model = {
         id: m.id,
         name: m.name,
@@ -29,7 +42,7 @@ function toModel(m, envDefaults = {}) {
             : { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
         params: { ...(m.tier ? { tier: m.tier } : {}), promptCacheFamily: m.provider === "anthropic" ? "input-excludes-cached" : "input-includes-cached" },
         ...(thinkingDefault ? { defaultThinking: thinkingDefault } : {}),
-        ...(effortLevels.length > 0 && m.provider !== "anthropic" ? { compat: { reasoningEffortLevels: effortLevels } } : {}),
+        ...(compat ? { compat } : {}),
         ...(m.extraBody ? { extraBody: m.extraBody } : {}),
         ...(Array.isArray(m.promptGuidance) && m.promptGuidance.length > 0 ? { promptGuidance: m.promptGuidance.map(String) } : {}),
     };
@@ -97,7 +110,7 @@ export function applyEffective(config, eff, logger, opts = {}) {
             logger?.warn("sema_registry_model_plane_swap_refused", {
                 version: staged.version,
                 error: err instanceof Error ? err.message : String(err),
-                note: "the Runner refused the candidate model plane (swapModels validates atomically — the current generation stays in force untouched); candidate refused whole, re-judged next poll (/health shows target≠applied while it persists)",
+                note: "the Runner refused the candidate model plane (swapDeps screens the whole candidate atomically — the current generation stays in force untouched); candidate refused whole, re-judged next poll (/health shows target≠applied while it persists)",
             });
             return false;
         }
@@ -174,7 +187,8 @@ function stageEffective(config, eff, logger, opts) {
         const models = {};
         const modelApiKeyEnv = {};
         const modelApiKeys = {};
-        const envDefaults = { maxTokens: config.model.maxTokens, contextWindow: config.model.contextWindow, reasoning: config.model.reasoning, charsPerToken: config.model.charsPerToken, defaultThinking: config.model.defaultThinking, reasoningEffortLevels: config.model.compat?.reasoningEffortLevels };
+        const envModel = envDerivedModel(config);
+        const envDefaults = { maxTokens: envModel.maxTokens, contextWindow: envModel.contextWindow, reasoning: envModel.reasoning, charsPerToken: envModel.charsPerToken, defaultThinking: envModel.defaultThinking, compat: envModel.compat };
         for (const m of enabled) {
             const rawBaseUrl = m.baseUrl;
             let safeUrl = rawBaseUrl;
@@ -194,7 +208,7 @@ function stageEffective(config, eff, logger, opts) {
                     safeUrl = undefined;
                 }
             }
-            models[m.name] = toModel(safeUrl === rawBaseUrl ? m : { ...m, baseUrl: safeUrl }, envDefaults);
+            models[m.name] = toModel(safeUrl === rawBaseUrl ? m : { ...m, baseUrl: safeUrl }, envDefaults, logger);
             if (m.sealedApiKey) {
                 const s = m.sealedApiKey;
                 if (s.alg !== SEALED_BOX_ALG) {

package/dist/config-center/facade.d.ts

@@ -44,6 +44,6 @@ export { fetchEffective, fetchPrincipalCaps, fetchPrincipalOrgMemory, ConfigCent
 export { mutateInPlace, applyEffective, type ApplyReport, applyRuntimeGates, applyRuntimeHot, resolveDefaultModelName, logEffectiveDiff, } from "./apply-effective.js";
 export { restartReasons, planeHasActiveTiers, modelPlaneChanged, type RestartSlice, type RestartSliceCtx, type RestartSignal, } from "./restart-signal.js";
 export { applyCenterSkills, resolveMcpServers, mcpForScenario, resolveA2aPeers, a2aForScenario } from "./skills-mcp.js";
-export { applyCenterReadFace } from "./read-face.js";
+export { applyCenterReadFace, CENTER_READ_FACE_LIVE_KEYS, type CenterReadFaceApply, type CenterReadFaceReport, type ReadFaceBaseline } from "./read-face.js";
 export type { CenterSkillManifest, CenterMcpServer, CenterA2aPeer, CenterReadFace, EffectiveConfig, ExecutionRuling, SessionMirrorRuling, } from "./types.js";
 //# sourceMappingURL=facade.d.ts.map

package/dist/config-center/facade.js

@@ -2,5 +2,5 @@ export { fetchEffective, fetchPrincipalCaps, fetchPrincipalOrgMemory, ConfigCent
 export { mutateInPlace, applyEffective, applyRuntimeGates, applyRuntimeHot, resolveDefaultModelName, logEffectiveDiff, } from "./apply-effective.js";
 export { restartReasons, planeHasActiveTiers, modelPlaneChanged, } from "./restart-signal.js";
 export { applyCenterSkills, resolveMcpServers, mcpForScenario, resolveA2aPeers, a2aForScenario } from "./skills-mcp.js";
-export { applyCenterReadFace } from "./read-face.js";
+export { applyCenterReadFace, CENTER_READ_FACE_LIVE_KEYS } from "./read-face.js";
 //# sourceMappingURL=facade.js.map

package/dist/config-center/read-face.d.ts

@@ -1,6 +1,35 @@
 import type { ServiceConfig } from "../config-types.js";
+import type { ReadFaceSource } from "../read-face-posture.js";
 import type { Logger } from "../observability/logger.js";
 import type { CenterReadFace, EffectiveConfig } from "./types.js";
+/**
+ * S-174 —— 哪些 center 键是**热换席**。判据**不是本仓的意见**:`KEY_MAP` 的 `cfgKey` 落在 core 的闭集
+ * `SWAPPABLE_DEP_SEATS` 里(经 `isSwappableDepSeat` 亲问)⇒ 这个键 refresh 期可换;否则它是 boot-baked
+ * 座,只有重启能换。core 加/删可换席 ⇒ 本集合与 `read-face` restart 片的指纹**同时**跟着走,不留第二份
+ * 手抄表(restart-signal.ts 读的就是这一只)。
+ */
+export declare const CENTER_READ_FACE_LIVE_KEYS: ReadonlySet<keyof CenterReadFace>;
+/** 本进程的**前 center** READ face 姿态(env / posture / 引擎默认)。`live` 相必带 —— 一次发布**撤回**
+ *  `face` 键时,活席要还原到它,而不是把上一份发布的档位无限期latch 住(那是这条轴上的 fail-OPEN 方向:
+ *  组织把 `open` 撤回之后,worker 还在 open)。boot 相不需要:那一刻 `config` 上的值就是基线。 */
+export interface ReadFaceBaseline {
+    readonly readFace?: "open" | "roots";
+    readonly readFaceSource: ReadFaceSource;
+}
+/** 本次调用落在哪一相(见文件头注「一席热、三席重启」)。判别联合而不是可选字段:`baseline` 只在
+ *  `live` 相有意义,写成可选就等于允许一次「没有基线的 live 调用」——那正是撤回臂会静默失效的形。 */
+export type CenterReadFaceApply = {
+    readonly phase: "boot";
+} | {
+    readonly phase: "live";
+    readonly baseline: ReadFaceBaseline;
+};
+/** 调用回执。`envHeld` 同旧返回值(喂 `RestartSliceCtx.readFaceEnvHeld`);`faceChanged` 是 live 相的
+ *  换席信号 —— 真 ⇒ 调用方必须 `swapDeps({ readFace: config.readFace })` 一次,否则 config 与 Runner 分家。 */
+export interface CenterReadFaceReport {
+    readonly envHeld: readonly (keyof CenterReadFace)[];
+    readonly faceChanged: boolean;
+}
 /**
  * 应用 center 的 readFace 域(**boot-only**:调用点在 boot 拉取之后、`createRunnerDeps` 之前)。
  * 缺席 ⇒ 零触碰。规则见文件头注 ①②③。
@@ -10,5 +39,5 @@ import type { CenterReadFace, EffectiveConfig } from "./types.js";
  * 否则 center 改一个本机永远用不上的键,整片 fleet 会为一次对自己毫无影响的发布滚动重启一遍
  * (codex 交叉复审 finding 2,验真后采纳的那一半)。
  */
-export declare function applyCenterReadFace(config: ServiceConfig, eff: EffectiveConfig, logger?: Logger): readonly (keyof CenterReadFace)[];
+export declare function applyCenterReadFace(config: ServiceConfig, eff: EffectiveConfig, logger?: Logger, apply?: CenterReadFaceApply): CenterReadFaceReport;
 //# sourceMappingURL=read-face.d.ts.map