@sema-agent/server 3.23.0 → 4.0.0

package/dist/bench/s1/run-firm.js

@@ -285,7 +285,7 @@ function dryRunMockFactory() {
             runWithVerification: async () => vr,
             runWithVerificationSup: async () => vr,
             resumeWithVerification: async () => vr,
-            runRepairLoop: async () => ({ ...vr, terminal: "candidate_only", bundle: { failureTrace: "", diagnostics: [], rejectedHypotheses: [], attemptCount: 1, oracleTier: "trusted_hidden" } }),
+            runRepairLoop: async () => ({ ...vr, terminal: "candidate_only", oracleCostMicroUsd: 0, bundle: { failureTrace: "", diagnostics: [], rejectedHypotheses: [], attemptCount: 1, oracleTier: "trusted_hidden" } }),
             getCheckpoint: async () => null,
             runLeaderTask: async () => ({ ok: true, reports: [], merge: undefined, workerBudgetsUsd: [0.5, 0.5] }),
             runOracle: async () => oracle,

package/dist/boot/config-center.js

@@ -780,7 +780,11 @@ export async function createConfigCenterRuntime(ctx) {
                                 // expanded admission gate while core throws "Unknown model ref" until restart. Hot-apply is safe
                                 // only when BOTH generations are tier-less.
                                 const planeDeferred = (runnerTierFrozen || planeHasActiveTiers(r.effective)) && modelPlaneChanged(appliedPlaneEff, r.effective);
-                                applyEffective(config, r.effective, logger, { teamsOnly: true, sealedKeys, ...(planeDeferred ? { deferModelPlane: true } : {}) });
+                                const committed = applyEffective(config, r.effective, logger, { teamsOnly: true, sealedKeys, ...(planeDeferred ? { deferModelPlane: true } : {}) });
+                                // [2283]③:CAS 拒绝(更旧世代)⇒ 本拍整体跳过——prompts 采用/pricing/keyResolver/restart
+                                // 对账/etag 推进/LKG 落盘都不得从旁路半应用同一个被拒世代(拒绝 warn 已在 applyEffective 留痕)。
+                                if (!committed)
+                                    return;
                                 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)" });
                                 }
@@ -954,7 +958,10 @@ export async function createConfigCenterRuntime(ctx) {
                     // this late arrival — if it froze a tier-expanded copy, the arriving center plane must not hot-apply
                     // (admission/Runner split). Tier-less env boot (the common deferred-boot shape) keeps true hot-apply.
                     const planeDeferredLate = (runnerTierFrozen || planeHasActiveTiers(r.effective)) && modelPlaneChanged(appliedPlaneEff, r.effective);
-                    applyEffective(config, r.effective, logger, { teamsOnly: true, sealedKeys, ...(planeDeferredLate ? { deferModelPlane: true } : {}) });
+                    const committedLate = applyEffective(config, r.effective, logger, { teamsOnly: true, sealedKeys, ...(planeDeferredLate ? { deferModelPlane: true } : {}) });
+                    // [2283]③(refresh 腿同款):CAS 拒绝 ⇒ 迟到 boot 拍整体跳过,旁路消费与 etag/LKG 都不动。
+                    if (!committedLate)
+                        return;
                     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" });
                     else {

package/dist/budget.d.ts

@@ -46,6 +46,12 @@ export interface ModelUsageDelta {
      *  = 运营者声明免费 ⇒ 键在场且为 0。两者可区分,这正是 #59 折零案的源头修复。
      *  与 SDK `ModelUsageDelta`(五键全可选)/ `usage-analytics.UsageModelEntry` 同轨。 */
     costMicroUsd?: number;
+    /** core 3.0.0([2296] 装车单 S10,§DESIGN-V2 V2-⑤)可选口径标记:标出这一行的 token 分量是在哪套记账
+     *  基准下写的(当前唯一值 = `"uncached-components-v1"`:`inputTokens` = cache-MISS 分量,`cacheRead`/
+     *  `cacheWrite` 分列)。**写点在 `trace/project.ts` 的 durable 行构造处**(`appendModelUsageDelta`),
+     *  不在这里 —— {@link ModelUsageTracker.record} 的累加逻辑本身零改动,永远不设置这个键(它只在 drain
+     *  之后、落盘之前被打上)。缺席 = 落盘更早、口径未知的存量行(usage-analytics 读端据此诚实标注)。 */
+    usageBasis?: string;
 }
 /**
  * E8 (shell-host contract): per-task × per-model usage accumulator, fed (synchronously) by the `brain.call`

package/dist/budget.js

@@ -206,16 +206,29 @@ export function createTracer(metrics, costQuota, modelUsage, fleetUsage, fleetLe
             // they neither leak nor enter the echo; their spend is in TaskStats.nested). Drained at turn boundaries into the
             // durable log (resume-safe); the registered entry is dropped at run end (clear). NOT cardinality-guarded by
             // model: bounded by a run's model set, and by the active-run set.
+            // core 3.0.0([2296] S1):this lane's `inputTokens` stays `e.promptTokens` — DELIBERATELY zero-changed.
+            // The E8 ledger's semantics are COMPONENT-wise (MISS + cacheRead/cacheWrite listed separately;
+            // usage-analytics' read side SUMS them back into a total), unlike the weightedTokens/fleetUsage lane just
+            // below (which wants the pre-summed normalized total and migrates to `e.totalInputTokens`). Pre-3.0.0 the
+            // gateway family's `promptTokens` already included cache, so this same read-side sum double-counted for
+            // that family — an existing accounting gap in historical rows that this version does not retroactively fix
+            // (declared, not silently carried forward): see CHANGELOG's "存量口径断层" note.
             modelUsage?.record(e.taskId, e.model, { inputTokens: e.promptTokens, outputTokens: e.completionTokens, cacheReadTokens: e.cacheRead, cacheWriteTokens: e.cacheWrite, ...(e.costMicroUsd !== undefined ? { costMicroUsd: e.costMicroUsd } : {}) }); // RB-368 透传缺席
             // weight-at-burn:加权 tokens = (prompt+completion)×模型配额倍率,在记账点固化(此后改倍率
             // 不回算历史,registry ModelEntry.quotaWeight 同句契约)。weight 源=config.modelQuotaWeights(registry
             // hot-apply;env lane 空表 ⇒ 1 兜底);非法回调值防御性归 1(tracer 契约 never-throws,不让坏表毒记账)。
+            // core 3.0.0([2296] S1):`e.promptTokens` → `e.totalInputTokens`. This axis's contract (line ~365 below)
+            // has always been "normalized total input INCLUDING cache" — pre-3.0.0 that's what `promptTokens` meant for
+            // the gateway family; 3.0.0 flips `promptTokens` to mean cache-MISS-only, so staying on it here would
+            // massively under-count the quota/lease axis for any cache-heavy load (a lease could be walked through by
+            // cache-heavy traffic — the exact same failure class as the side-query R4 finding below).
             const qwRaw = quotaWeightFor?.(e.model);
             const qw = typeof qwRaw === "number" && Number.isFinite(qwRaw) && qwRaw > 0 ? qwRaw : 1;
-            const weightedTokens = Math.round((e.promptTokens + e.completionTokens) * qw);
+            const weightedTokens = Math.round((e.totalInputTokens + e.completionTokens) * qw);
             // P2 用量批报:principal×model 窗口累计(同 ⑤b 的 ALS 归因点 — 子任务花费也归到发起 principal),
             // fleet-client 每窗口 drain→POST /api/fleet/usage。未配 fleet(accumulator 缺席)= 零行为。
-            fleetUsage?.record(currentPrincipal(), e.model, e.taskId, { inputTokens: e.promptTokens, outputTokens: e.completionTokens, ...(e.costMicroUsd !== undefined ? { costMicroUsd: e.costMicroUsd } : {}), weightedTokens, quotaWeightAtUse: qw }); // RB-368 透传缺席
+            // core 3.0.0:同上迁移理由,inputTokens=归一化总输入含 cache(totalInputTokens),与 weightedTokens 同轴。
+            fleetUsage?.record(currentPrincipal(), e.model, e.taskId, { inputTokens: e.totalInputTokens, outputTokens: e.completionTokens, ...(e.costMicroUsd !== undefined ? { costMicroUsd: e.costMicroUsd } : {}), weightedTokens, quotaWeightAtUse: qw }); // RB-368 透传缺席
             // lease 消费(D4 AP):同一 ALS 归因点做 lease 本地扣减 —— 只对当前持 lease 的 principal
             // 生效(FleetLeaseManager 内部判),sync + never throws。双轴=weightedTokens(配额轴)+costMicroUsd($ 轴)。
             fleetLease?.recordSpend(currentPrincipal(), { weightedTokens, ...(e.costMicroUsd !== undefined ? { costMicroUsd: e.costMicroUsd } : {}) }); // RB-368 透传缺席
@@ -332,6 +345,9 @@ fleetUsage, fleetLease, quotaWeightFor) {
         // (input-excludes-cached)的 cache 命中几乎全部在 input 之外,不加 cacheRead/cacheWrite 就是配额轴
         // 大幅少收(lease 可被 cache-heavy side query 绕穿)。family 未知(catalog 查不到)走**保守臂**
         // (含 cache——OpenAI 族此臂多收,方向与 cost 残差一致:宁多勿少,记档同一条)。
+        // core 3.0.0 核实(S1 提货站):此段的输入源是 `SideQueryResult.usage`(llm 层生料 `Usage`,
+        // engine/llm/types.d.ts)——不是 3.0.0 翻新过的 `brain.call`/`turn_end` trace 事件面,3.0.0 未动这个类型,
+        // 本折算继续承重(零改动)。
         const promptTokens = r.family === "input-includes-cached" ? input : input + cacheRead + cacheWrite;
         // RB-368/#59 同判据(createTracer 上方注释同款纪律):`usage.cost.total` 整键缺席 = 未知(unpriced 部署的
         // side query 真形),不是「花了 $0」——折 0 就是把「不知道」编造成「免费」。已知(哪怕显式 0)才带键。

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

@@ -8,11 +8,14 @@ import type { EffectiveConfig } from "./types.js";
  *  `this.deps.models/roles/pricing` LIVE per-task and `/v1/models` reads `config.models` — both share the
  *  reference captured at boot, so mutating it (vs reassigning) updates both with no split, no Runner rebuild. */
 export declare function mutateInPlace<V>(target: Record<string, V>, source: Record<string, V>): void;
+/** 返回值=是否 COMMIT(false 仅在 [2283]③ 世代序 CAS 拒绝时)。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。 */
 export declare function applyEffective(config: ServiceConfig, eff: EffectiveConfig, logger?: Logger, opts?: {
     teamsOnly?: boolean;
     sealedKeys?: SealedKeyOpener;
     deferModelPlane?: boolean;
-}): void;
+}): boolean;
 /**
  * Apply the center `runtime` governance/limit gates OVER the env-derived config (mutates `config`).
  *
@@ -42,23 +45,6 @@ export type RuntimeGateKey = (typeof RUNTIME_GATE_KEYS)[number];
  *  divides by it; center's `.positive()` already enforces, belt-and-suspenders here). */
 export declare function runtimeGatePresent(rt: NonNullable<EffectiveConfig["runtime"]>, key: RuntimeGateKey): boolean;
 export declare function applyRuntimeGates(config: ServiceConfig, rt: EffectiveConfig["runtime"], logger?: Logger): void;
-/**
- * Apply the runtime governance "second baton" (center §10): `autonomy` + `commandPolicy`. UNLIKE the 6 gates in
- * {@link applyRuntimeGates}, these are per-request HOT (read live in main.ts `resolveSpec` via
- * `applyRuntimeGovernance`), so they apply on BOTH boot and refresh (NOT via RUNTIME_GATE_KEYS / restart-to-apply).
- *
- * 🔴 NON-STICKY (differs from the restart-to-apply gates on purpose — adversarial-review HIGH): a hot field that
- * center STOPS publishing must REVERT to the env baseline, not keep the last center value. The restart-to-apply
- * gates can be sticky because a refresh never touches them (the running middleware holds boot values until a
- * restart re-reads env+center); a HOT field has no such reset, so "absent ⇒ keep" would silently freeze a stale
- * center override (e.g. center published `auto`, then un-published it — the gate would stay OFF forever). So we
- * recompute every call as `present ? center : envBaseline`. The env baseline = `AUTONOMY` (re-derived, env is
- * immutable at runtime) for autonomy; `undefined` (no env scalar source) for commandPolicy.
- *
- * 🔴 commandPolicy is VALIDATED here (adversarial-review HIGH): an invalid command (glob/path/operator — which
- * core's EXACT-name matcher would silently never match → a hole) is rejected FAIL-LOUD and the prior good policy
- * is KEPT (a broken publish never half-applies a silently-weakened gate).
- */
 export declare function applyRuntimeHot(config: ServiceConfig, rt: EffectiveConfig["runtime"], logger?: Logger): void;
 /**
  * [865]① 显式默认解析——applyEffective 与 dry-run(logEffectiveDiff.wouldDefaultModel)共用的单源。优先级

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

@@ -78,7 +78,74 @@ export function mutateInPlace(target, source) {
         delete target[k];
     Object.assign(target, source);
 }
+/** 每个 config 对象上一次 COMMIT 的世代号([2283]③ CAS 的 live 端)。WeakMap——config 回收即回收;
+ *  version 0(未发布/本地一次性 lane)不参与登记。 */
+const appliedGeneration = new WeakMap();
+/** 返回值=是否 COMMIT(false 仅在 [2283]③ 世代序 CAS 拒绝时)。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。 */
 export function applyEffective(config, eff, logger, opts = {}) {
+    const staged = stageEffective(config, eff, logger, opts);
+    const live = appliedGeneration.get(config);
+    if (staged.version > 0 && live !== undefined && staged.version < live) {
+        // [2283]③:两次拉取的 staging 并发/乱序完成时,慢的旧世代不得整体覆盖快的新世代——每次都是
+        // 「整世代」,但方向反了。拒绝必须留痕(C6);caller 的 etag 纪律本就只在 apply 成功后推进,
+        // 下一拍拉到的自然是更新的世代。等版本重放(etag 未推进的幂等重放)照常放行。
+        logger?.warn("sema_registry_stale_generation_refused", { staged: staged.version, live, note: "an older staged generation must not overwrite a newer committed one — refused whole; next poll replays" });
+        return false;
+    }
+    commitStaged(config, staged);
+    // ── post-commit notifications(全部世代描述性通知在最后一笔赋值之后,[2283]②)──
+    if (staged.modelPlane)
+        logger?.info("sema_registry_models", staged.modelPlane.infoLine);
+    if (staged.gatesInfo)
+        logger?.info("sema_registry_runtime", staged.gatesInfo);
+    if (Object.keys(staged.hot.infoLine).length > 0)
+        logger?.info("sema_registry_runtime_hot", staged.hot.infoLine);
+    for (const n of staged.teamsPlane?.notices ?? []) {
+        if (n.level === "warn")
+            logger?.warn(n.msg, n.fields);
+        else
+            logger?.info(n.msg, n.fields);
+    }
+    if (Object.keys(staged.projects).length > 0)
+        logger?.info("sema_registry_projects", { projects: Object.keys(staged.projects) });
+    return true;
+}
+/** COMMIT 段:纯赋值,零 await/零回调/零发射(源码钉守着——[2283]①②)。mutateInPlace/整对象重赋值/
+ *  registry 表整体换装(registerTeams/registerCollabWorkflows 皆为过滤+swap 的赋值形,无回调面)。 */
+function commitStaged(config, staged) {
+    const mp = staged.modelPlane;
+    if (mp) {
+        config.modelApiKeyEnv = mp.modelApiKeyEnv; // reassign ok — keyResolver is rebuilt from it on refresh (main.ts)
+        config.modelApiKeys = mp.modelApiKeys; // plaintext values — memory-only; poison markers ride the same map
+        mutateInPlace(config.modelQuotaWeights, mp.quotaWeights);
+        mutateInPlace(config.models, mp.models); // IN PLACE: keep the ref the Runner + /v1/models share (hot-apply)
+        config.model = config.models.default;
+        if (Object.keys(mp.roles).length > 0)
+            mutateInPlace(config.roles, mp.roles); // IN PLACE: Runner reads this.deps.roles live
+        mutateInPlace(config.tiers, mp.activeTiers);
+    }
+    if (staged.gateAssignments) {
+        for (const [k, v] of staged.gateAssignments)
+            config[k] = v;
+    }
+    if (staged.hot.setAutonomy)
+        config.autonomy = staged.hot.autonomyNext;
+    if (staged.hot.setCommandPolicy)
+        config.commandPolicy = staged.hot.commandPolicyNext;
+    const tp = staged.teamsPlane;
+    if (tp) {
+        registerTeams(tp.teams);
+        registerCollabWorkflows(tp.workflows);
+    }
+    mutateInPlace(config.projects, staged.projects); // IN PLACE: keep the ref per-request consumers captured at boot
+    if (staged.version > 0)
+        appliedGeneration.set(config, staged.version);
+}
+/** STAGE 段:一切计算/校验/解封/投影(可抛;抛=候选整体拒绝,活配置零触碰)。内容判定性 warn/error
+ *  (描述 eff 真伪,与是否 commit 无关)在此发;世代描述性 info 只装载荷,post-commit 发。 */
+function stageEffective(config, eff, logger, opts) {
     // version 0 / empty effective = the config-center has nothing for us yet — typically CONFIG_PUBLISH_MODE
     // is ON but nothing has been published. We do NOT wipe: models/roles fall back to env (the enabled>0
     // guard below), teams to BUILTIN_TEAMS (registerTeams resets to built-ins). Warn once at boot so the
@@ -96,6 +163,7 @@ export function applyEffective(config, eff, logger, opts = {}) {
     // split, no Runner rebuild, no core change. Caller (main.ts refresh) additionally refreshes `pricing` (same
     // ref) + rebuilds `keyResolver`. (Was startup-only when reassignment split the Runner's ref from config.)
     const enabled = (eff.models?.models ?? []).filter((m) => m.enabled !== false);
+    let modelPlane;
     // codex R10: `deferModelPlane` skips the WHOLE model plane (models/roles/tiers + per-model keys) — main.ts
     // sets it when the Runner is tier-frozen (private expanded copy) and the plane changed: hot-applying would
     // split admission (/v1/models, catalog gates) from the Runner's frozen generation, letting a same-key retarget
@@ -189,11 +257,6 @@ export function applyEffective(config, eff, logger, opts = {}) {
                 logger?.warn("sema_registry_model_no_brain", { model: m.name, provider: m.provider, hint: "set ANTHROPIC_API_KEY in this service's env, else it mis-routes to the gateway brain" });
             }
         }
-        config.modelApiKeyEnv = modelApiKeyEnv; // reassign ok — keyResolver is rebuilt from it on refresh (main.ts)
-        // same rebuild contract; plaintext values — memory-only, never logged. Poison markers ride the same
-        // map (they ARE per-model key state: "configured but broken"); the cast bridges config.ts's
-        // plaintext-only field type until it is widened to Record<string, string | SealedKeyPoison>.
-        config.modelApiKeys = modelApiKeys; // 类型已放真(string | SealedKeyPoison),桥接 cast 退役
         // weight-at-burn 源:quotaWeight 按 name+id 双键索引(tracer 的 brain.call e.model=模型 id,
         // 目录键=name——双键免猜);非法值(≤0/NaN)按缺省 1 丢弃。IN PLACE 与 models 同批 hot-apply。
         const quotaWeights = {};
@@ -205,7 +268,6 @@ export function applyEffective(config, eff, logger, opts = {}) {
                     quotaWeights[m.id] = qw;
             }
         }
-        mutateInPlace(config.modelQuotaWeights, quotaWeights);
         const roles = {};
         for (const [role, tgt] of Object.entries(eff.models.roles ?? {})) {
             if ("model" in tgt)
@@ -225,15 +287,16 @@ export function applyEffective(config, eff, logger, opts = {}) {
         // 静默翻转,default 必须消费 wire 里已有的显式意图。解析单源 = resolveDefaultModelName(dry-run 的
         // wouldDefaultModel 共用同一只,报数与真 apply 永不撕裂——codex M2)。
         const picked = resolveDefaultModelName(eff, (n) => models[n] !== undefined, enabled[0].name, (source, name) => logger?.warn("sema_registry_default_dangling", { source, name, hint: "explicit default names a model that is not in the enabled catalog — falling to the next source" }));
-        const defaultName = picked.name;
-        const defaultSource = picked.source;
-        models.default = models[defaultName];
-        mutateInPlace(config.models, models); // IN PLACE: keep the ref the Runner + /v1/models share (hot-apply)
-        config.model = config.models.default;
-        if (Object.keys(roles).length > 0)
-            mutateInPlace(config.roles, roles); // IN PLACE: Runner reads this.deps.roles live
-        mutateInPlace(config.tiers, activeTiers);
-        logger?.info("sema_registry_models", { count: enabled.length, default: config.model.id, defaultSource, roles: Object.keys(roles), tiers: Object.keys(config.tiers), sealedKeys: Object.values(modelApiKeys).filter((v) => typeof v === "string").length, sealedPoisoned: Object.values(modelApiKeys).filter((v) => typeof v !== "string").length });
+        models.default = models[picked.name];
+        modelPlane = {
+            models,
+            modelApiKeyEnv,
+            modelApiKeys,
+            quotaWeights,
+            roles,
+            activeTiers,
+            infoLine: { count: enabled.length, default: models.default.id, defaultSource: picked.source, roles: Object.keys(roles), tiers: Object.keys(activeTiers), sealedKeys: Object.values(modelApiKeys).filter((v) => typeof v === "string").length, sealedPoisoned: Object.values(modelApiKeys).filter((v) => typeof v !== "string").length },
+        };
     }
     // Runtime governance/limit gates (phase-2): STARTUP only (restart-to-apply) — the live RateLimiter / CostQuota
     // / approval gate are built from `config` AFTER this in main.ts, so mutating it here before they're constructed
@@ -241,18 +304,18 @@ export function applyEffective(config, eff, logger, opts = {}) {
     // governance 切新位:治理三件优先读 governance 域(真值),runtime 旧位(双写镜像)fallback——
     // 双写期两处逐键相等语义不变;center 撤双写后 governance 即唯一来源。限额残余(rateLimit/cost 五件)仍在 runtime。
     const gatesView = eff.governance ? { ...eff.runtime, ...eff.governance } : eff.runtime;
-    if (!opts.teamsOnly)
-        applyRuntimeGates(config, gatesView, logger);
+    const gates = opts.teamsOnly ? undefined : stageRuntimeGates(gatesView);
     // Runtime governance "second baton" (center §10): autonomy + commandPolicy are per-request HOT (read live in
     // resolveSpec), NOT baked into boot middleware → apply on BOTH boot and refresh (outside the teamsOnly guard) so
     // they hot-reload. No restart-to-apply signal (they take effect on the next task without a restart).
-    applyRuntimeHot(config, gatesView, logger);
+    const hot = stageRuntimeHot(config, gatesView, logger);
     // Teams → registry (hot-reloadable; center overrides/extends the built-ins).
     // codex R11: teams + collab workflows carry MODEL REFERENCES (member.model / workflow model args) — when the
     // model plane is deferred (tier-frozen Runner, see the deferModelPlane guard above) these faces must defer
     // WITH it, or a candidate that atomically adds/retargets a model AND updates a team/workflow to reference it
     // publishes a mixed generation: the new template goes live while config.models/the Runner stay old (a new ref
     // fails as unknown; a same-name retarget silently executes stale). One catalog generation = one visibility.
+    let teamsPlane;
     if (opts.deferModelPlane !== true) {
         const teams = {};
         for (const t of (eff.teams?.teams ?? []).filter((t) => t.enabled !== false)) {
@@ -265,22 +328,22 @@ export function applyEffective(config, eff, logger, opts = {}) {
                 synthesizer: t.synthesizer ? { role: t.synthesizer.role, modelRole: t.synthesizer.modelRole, systemPrompt: t.synthesizer.systemPrompt } : undefined,
             };
         }
-        registerTeams(teams);
-        if (Object.keys(teams).length > 0)
-            logger?.info("sema_registry_teams", { teams: Object.keys(teams) });
         // collab → named-workflow projection(切片 1.5,design/140 统一解;切片① 的
         // TeamTemplate 投影已整体替换——纪律「别留双路径降级」)。可投影子集翻译成命名 workflow 注册条目
         // (执行体=core 内置 team-discussion 脚本,center 模板=defaultArgs 合并链第二级;键=collab id,shell
         // `/team`/LLM 经 Workflow({name}) 调用);子集外整条跳过+结构化上报,绝不静默降级。HOT:boot+refresh
-        // 双腿整表替换(非粘——center 停发即空表,内置 workflow 不受影响)。
+        // 双腿整表替换(非粘——center 停发即空表,内置 workflow 不受影响)。register* 在 commit 段成对换装。
         const projected = projectCollabToWorkflows(eff.collab?.templates);
-        registerCollabWorkflows(projected.workflows);
+        const notices = [];
+        if (Object.keys(teams).length > 0)
+            notices.push({ level: "info", msg: "sema_registry_teams", fields: { teams: Object.keys(teams) } });
         if (Object.keys(projected.workflows).length > 0)
-            logger?.info("sema_registry_collab_workflows", { workflows: Object.keys(projected.workflows) });
+            notices.push({ level: "info", msg: "sema_registry_collab_workflows", fields: { workflows: Object.keys(projected.workflows) } });
         if (projected.skipped.length > 0)
-            logger?.warn("sema_registry_collab_skipped", { skipped: projected.skipped });
+            notices.push({ level: "warn", msg: "sema_registry_collab_skipped", fields: { skipped: projected.skipped } });
         if (projected.notes.length > 0)
-            logger?.info("sema_registry_collab_notes", { notes: projected.notes });
+            notices.push({ level: "info", msg: "sema_registry_collab_notes", fields: { notes: projected.notes } });
+        teamsPlane = { teams, workflows: projected.workflows, notices };
     }
     // 142-S4 projects 域(registry-core 0.10.0):center 项目登记簿 → config.projects(键=
     // projectId)。消费是 per-request 查表(memoryScope 派生 + defaultScopes 种子,security.ts/main.ts),
@@ -304,9 +367,14 @@ export function applyEffective(config, eff, logger, opts = {}) {
             ...(Array.isArray(r.defaultScopes) ? { defaultScopes: r.defaultScopes.filter((x) => typeof x === "string") } : {}),
         };
     }
-    mutateInPlace(config.projects, projects); // IN PLACE: keep the ref per-request consumers captured at boot
-    if (Object.keys(projects).length > 0)
-        logger?.info("sema_registry_projects", { projects: Object.keys(projects) });
+    return {
+        version: typeof eff.version === "number" && Number.isFinite(eff.version) ? eff.version : 0,
+        ...(modelPlane !== undefined ? { modelPlane } : {}),
+        ...(gates !== undefined ? { gateAssignments: gates.assignments, ...(gates.info !== undefined ? { gatesInfo: gates.info } : {}) } : {}),
+        hot,
+        ...(teamsPlane !== undefined ? { teamsPlane } : {}),
+        projects,
+    };
 }
 /**
  * Apply the center `runtime` governance/limit gates OVER the env-derived config (mutates `config`).
@@ -342,18 +410,28 @@ export function runtimeGatePresent(rt, key) {
         return typeof v === "number" && v > 0;
     return typeof v === "number";
 }
-export function applyRuntimeGates(config, rt, logger) {
+/** [2283] stage 半场:六闸的赋值清单(纯计算,零变异)。commit 段照单赋值,通知载荷 post-commit 发。 */
+function stageRuntimeGates(rt) {
     if (!rt)
-        return;
+        return { assignments: [] };
+    const assignments = [];
     const applied = {};
     for (const key of RUNTIME_GATE_KEYS) {
         if (!runtimeGatePresent(rt, key))
             continue; // undefined / sentinel → keep env
-        config[key] = rt[key]; // present (incl. explicit 0/[]) → override
+        assignments.push([key, rt[key]]); // present (incl. explicit 0/[]) → override
         applied[key] = rt[key];
     }
-    if (Object.keys(applied).length > 0)
-        logger?.info("sema_registry_runtime", applied);
+    return { assignments, ...(Object.keys(applied).length > 0 ? { info: applied } : {}) };
+}
+export function applyRuntimeGates(config, rt, logger) {
+    // 独立调用面的兼容壳(测试/外部):stage → 就地赋值 → 通知。applyEffective 不走这里——它把
+    // assignments 并进自己的 commit 段以保住整世代原子性([2283]②)。
+    const s = stageRuntimeGates(rt);
+    for (const [k, v] of s.assignments)
+        config[k] = v;
+    if (s.info)
+        logger?.info("sema_registry_runtime", s.info);
 }
 /**
  * Apply the runtime governance "second baton" (center §10): `autonomy` + `commandPolicy`. UNLIKE the 6 gates in
@@ -372,34 +450,49 @@ export function applyRuntimeGates(config, rt, logger) {
  * core's EXACT-name matcher would silently never match → a hole) is rejected FAIL-LOUD and the prior good policy
  * is KEPT (a broken publish never half-applies a silently-weakened gate).
  */
-export function applyRuntimeHot(config, rt, logger) {
-    const applied = {};
+/** [2283] stage 半场:hot 二件的赋值决定(计算+校验;`sema_registry_commandpolicy_invalid` 是内容判定,
+ *  stage 期发——它描述 eff 的真伪,与是否 commit 无关)。对比基准=stage 时刻的 config 现值:stage 与
+ *  commit 同一同步 tick,期间无人能改 config(单线程),对比不失效。 */
+function stageRuntimeHot(config, rt, logger) {
+    const infoLine = {};
     // autonomy: present ⇒ center value; absent ⇒ revert to the env baseline (never a stale center override).
     const envAutonomy = parseAutonomy(process.env.AUTONOMY);
-    const nextAutonomy = rt?.autonomy !== undefined ? rt.autonomy : envAutonomy;
-    if (config.autonomy !== nextAutonomy) {
-        config.autonomy = nextAutonomy;
-        applied.autonomy = nextAutonomy ?? "(env-baseline)";
-    }
+    const autonomyNext = rt?.autonomy !== undefined ? rt.autonomy : envAutonomy;
+    const setAutonomy = config.autonomy !== autonomyNext;
+    if (setAutonomy)
+        infoLine.autonomy = autonomyNext ?? "(env-baseline)";
     // commandPolicy: present+valid ⇒ apply; present+invalid ⇒ fail-loud + keep prior; absent ⇒ revert to baseline
     // (undefined — no env scalar source for structured command rules). Only re-set + log on an ACTUAL change
     // (deep-equal compare) — refresh runs every ~60s and the policy is usually unchanged; logging every tick = noise.
+    let setCommandPolicy = false;
+    let commandPolicyNext;
     if (rt?.commandPolicy !== undefined) {
         const errors = validateCommandRules(rt.commandPolicy);
         if (errors.length > 0) {
             logger?.error("sema_registry_commandpolicy_invalid", { errors, kept: config.commandPolicy?.length ?? 0 });
         }
         else if (JSON.stringify(config.commandPolicy) !== JSON.stringify(rt.commandPolicy)) {
-            config.commandPolicy = rt.commandPolicy;
-            applied.commandPolicy = rt.commandPolicy.length; // log the COUNT, not the rules (avoid leaking on every refresh)
+            setCommandPolicy = true;
+            commandPolicyNext = rt.commandPolicy;
+            infoLine.commandPolicy = rt.commandPolicy.length; // log the COUNT, not the rules (avoid leaking on every refresh)
         }
     }
     else if (config.commandPolicy !== undefined) {
-        config.commandPolicy = undefined; // center stopped managing → revert to baseline (no env source)
-        applied.commandPolicy = "(env-baseline)";
+        setCommandPolicy = true;
+        commandPolicyNext = undefined; // center stopped managing → revert to baseline (no env source)
+        infoLine.commandPolicy = "(env-baseline)";
     }
-    if (Object.keys(applied).length > 0)
-        logger?.info("sema_registry_runtime_hot", applied);
+    return { setAutonomy, autonomyNext, setCommandPolicy, commandPolicyNext, infoLine };
+}
+export function applyRuntimeHot(config, rt, logger) {
+    // 独立调用面的兼容壳(测试/外部)——applyEffective 不走这里(同 applyRuntimeGates 的注)。
+    const s = stageRuntimeHot(config, rt, logger);
+    if (s.setAutonomy)
+        config.autonomy = s.autonomyNext;
+    if (s.setCommandPolicy)
+        config.commandPolicy = s.commandPolicyNext;
+    if (Object.keys(s.infoLine).length > 0)
+        logger?.info("sema_registry_runtime_hot", s.infoLine);
 }
 /** [865]①/H3:activeTierGroup 档绑定当默认时的档梯,与 core 单一语义源逐字对齐(core roles.js:
  *  `ROLE_TIER_DEFAULTS.default = "pro"` + `resolveTier` 从本档位置**只向低档**扫 DEFAULT_TIER_ORDER

package/dist/config-center/http-client.d.ts

@@ -1,10 +1,17 @@
-import type { EntitlementRuntimeCaps } from "@sema-agent/registry-core";
+import { type EntitlementRuntimeCaps } from "@sema-agent/registry-core";
 import type { ScenarioRuling } from "../capabilities/scenarios.js";
 import type { EffectiveConfig, ExecutionRuling } from "./types.js";
-/** GET the effective config (Bearer + ETag). null = 304 (unchanged). Throws on transport/HTTP error. */
+/** GET the effective config (Bearer + ETag). null = 304 (unchanged). Throws on transport/HTTP error, and on a
+ *  payload that is not an effective config at all(非对象 / version 非有限数 / gate 域坏形——[2281] 裁B)。
+ *  `domainErrors`:坏 catalog 域(schema default 已落)逐域单列——与本地腿 `FetchEffectiveResult` 同键同义,
+ *  boot/refresh 的候选门直接消费。 */
 export declare function fetchEffective(baseUrl: string, token: string, etag: string | undefined, fetchImpl?: typeof fetch, worker?: string): Promise<{
     effective: EffectiveConfig;
     etag?: string;
+    domainErrors?: Array<{
+        domain: string;
+        error: string;
+    }>;
 } | null>;
 /**
  * design/99 §K (core 1.157 `RunnerDeps.runtimeCapsResolver`) — GET the PER-PRINCIPAL runtime

package/dist/config-center/http-client.js

@@ -5,7 +5,11 @@
  * facade re-exports every symbol below unchanged).
  */
 import { createHash } from "node:crypto";
-/** GET the effective config (Bearer + ETag). null = 304 (unchanged). Throws on transport/HTTP error. */
+import { readEffectiveWire } from "@sema-agent/registry-core";
+/** GET the effective config (Bearer + ETag). null = 304 (unchanged). Throws on transport/HTTP error, and on a
+ *  payload that is not an effective config at all(非对象 / version 非有限数 / gate 域坏形——[2281] 裁B)。
+ *  `domainErrors`:坏 catalog 域(schema default 已落)逐域单列——与本地腿 `FetchEffectiveResult` 同键同义,
+ *  boot/refresh 的候选门直接消费。 */
 export async function fetchEffective(baseUrl, token, etag, fetchImpl = fetch, worker) {
     // Defensive scheme guard: Node's fetch supports file:// — a mis-set
     // SEMA_REGISTRY_URL must not turn into a local-file read. Reject anything but http(s).
@@ -23,7 +27,21 @@ export async function fetchEffective(baseUrl, token, etag, fetchImpl = fetch, wo
         return null;
     if (!res.ok)
         throw new Error(`config-center HTTP ${res.status}`);
-    return { effective: (await res.json()), etag: res.headers.get("etag") ?? undefined };
+    // [2281] 裁B(§M1):wire 载荷在**这个消费端边界**过 registry-core `readEffectiveWire`(0.12.0)真校验,
+    // 不再裸断言——「两个契约共用一个拼写不是一个被检查的契约」。判据与本地腿同源(parseDomain +
+    // DOMAIN_READ_FALLBACK):catalog 域坏形 → 该域 schema default + 下面折进 domainErrors(候选门在
+    // refresh 拒候选、boot 逐域 warn);gate 域坏形/垃圾载荷 → throw(caller 整包 catch 回落 env/LKG)。
+    // 未知顶层键 verbatim 透传(open-world:新 center 新域、legacy teams 键都不丢)。grandfather 类
+    // 警告(值已被收编接受)不算 error,不进 domainErrors——与本地店同口径。
+    const warnings = [];
+    const wire = readEffectiveWire(await res.json(), (w) => warnings.push(w));
+    const domainErrors = warnings
+        .filter((w) => w.kind === "domain-defaulted")
+        .map((w) => ({ domain: w.domain, error: (w.error instanceof Error ? w.error.message : String(w.error)).slice(0, 600) }));
+    // 经真校验后的 wire 值到 service 拼写副本的换装:两型同一契约面(EffectiveWire 是 service 型的同源超集,
+    // service 型只声明自己消费的域且全 optional)——此断言的前提正是上面那次校验,不再是裸信任。
+    const effective = wire;
+    return { effective, etag: res.headers.get("etag") ?? undefined, ...(domainErrors.length > 0 ? { domainErrors } : {}) };
 }
 /**
  * design/99 §K (core 1.157 `RunnerDeps.runtimeCapsResolver`) — GET the PER-PRINCIPAL runtime

package/dist/config-provider.js

@@ -91,10 +91,16 @@ export class RemoteConfigProvider {
         this.cc = cc;
         this.deps = deps;
     }
-    fetchEffective(etag) {
+    async fetchEffective(etag) {
         const fn = this.deps.fetchEffective ?? remoteFetchEffective;
         // 5th arg = worker scope; transport (fetchImpl) stays the config-center's default.
-        return fn(this.cc.baseUrl, this.cc.token, etag, undefined, this.cc.worker);
+        const r = await fn(this.cc.baseUrl, this.cc.token, etag, undefined, this.cc.worker);
+        if (r === null || r.domainErrors === undefined)
+            return r;
+        // [2281] 裁B:远程腿从此也产 domainErrors(http-client 消费端校验)。与本地腿同一条产出边界纪律:
+        // error 文本先过 redactConfigError(zod 错误会回声违规值——operand 指纹化,防受控值经
+        // config_candidate_rejected/config_domain_invalid 进结构化日志流)。
+        return { ...r, domainErrors: r.domainErrors.map((de) => ({ domain: de.domain, error: redactConfigError(de.error) })) };
     }
     async fetchSkillContent(contentHash) {
         const fn = this.deps.fetchSkillContent ?? remoteFetchSkillContent;

package/dist/config.js

@@ -514,8 +514,10 @@ function parseModelDomain() {
     }
     // Model.api must reflect the brain that actually serves it (routing is by provider): an "anthropic"
     // provider runs the Anthropic brain, whose usage reports `input` EXCLUDING cached tokens. core ≥1.22
-    // normalizes the cache-hit-rate denominator per api family — a wrong api makes promptTokens too small
-    // and cacheHitRate exceed 100%. (Brains select by provider, not api, so this only fixes accounting.)
+    // normalizes the cache-hit-rate denominator per api family (`cacheFamilyOf` picks the fold formula from
+    // `Model.api`) — a wrong api picks the wrong family, so the normalized total (core 3.0.0: `totalInputTokens`,
+    // core's own cacheHitRate = cachedTokens/totalInputTokens) comes out too small and cacheHitRate exceeds
+    // 100%. (Brains select by provider, not api, so this only fixes accounting.)
     //
     // Default inference ([792]④, core ruling 2026-07-14): an EXPLICIT MODEL_PROVIDER always wins (unchanged).
     // When it is unset but the operator explicitly configured an Anthropic-protocol base URL AND a matching

package/dist/fleet/fleet-bus.js

@@ -222,13 +222,16 @@ export function fleetRunPublisher(bus, run) {
             return 0;
         if (typeof u.totalTokens === "number")
             return u.totalTokens;
-        // #9 (workflow 2nd-pass re-review): a turn_end.usage has NO totalTokens → this field-sum. core's turn_end.inputTokens
-        // is the NORMALIZED TOTAL INPUT *INCLUDING CACHE* (usage-accounting promptTokensOf — input + cacheRead + cacheWrite
-        // for the Anthropic family; OpenAI already folds cacheRead into input). So inputTokens + outputTokens ALREADY is the
-        // turn total — the old code ALSO added cacheRead/cacheWrite, double-counting them (~2x with prompt caching, the
-        // normal case). Display-only (the leader fleet-row token number); billing (costMicroUsd) + context/compaction use
-        // totalTokens and were never affected; child rows use task_progress.usage.totalTokens (the early-return) — correct.
-        return (u.inputTokens ?? 0) + (u.outputTokens ?? 0);
+        // core 3.0.0([2296] 装车单 S2, [2298] R#9 二次复审并入):`turn_end.usage.inputTokens` 的语义在 3.0.0
+        // **翻转**——此前(core ≤2.x)它是 NORMALIZED TOTAL INPUT *INCLUDING CACHE*(usage-accounting
+        // promptTokensOf:Anthropic 族 = input+cacheRead+cacheWrite,OpenAI 族已把 cache 折进 input),本函数当时
+        // 直接 `inputTokens + outputTokens` 就是整轮总量;3.0.0 起 `inputTokens` 恒为 cache-MISS 分量,新必填的
+        // `totalInputTokens` 才是那个「含 cache 的归一化总量」(旧 inputTokens 语义的正统继承者)。继续只加
+        // `inputTokens` 会在有 cache 命中时**漏计**(leader 行的展示 token 数偏低)——故总量优先取
+        // `totalInputTokens`,旧核帧(尚未升级到 3.0.0、该键缺席)回落 `inputTokens`(防御性:本引擎恒 ≥3.0.0
+        // 时两键并存;异常帧宁少算不 NaN)。Display-only(leader fleet 行的 token 数);billing(costMicroUsd)+
+        // context/compaction 走 totalTokens,不受影响;child 行走 task_progress.usage.totalTokens(上面的早退臂)。
+        return (u.totalInputTokens ?? u.inputTokens ?? 0) + (u.outputTokens ?? 0);
     };
     return {
         onStart() {

package/dist/fleet-client.d.ts

@@ -31,6 +31,10 @@ export declare class FleetUsageAccumulator {
      * 同款=**传染**(同为单 producer 累加器:已知+未知=未知);center 聚合面的「下界 + costUnknown
      * 标记」语义在 registry-core aggregateUsage(跨 worker 审计总账层,两层判据见 [2103])。
      * 配额轴(weightedTokens)不受影响,仍逐笔精确。
+     *
+     * core 3.0.0([2296] 装车单 S4c)语义注:`d.inputTokens` 在唯一调用点(budget.ts 的
+     * `fleetUsage.record(...)`,S1 已迁)传的是 `e.totalInputTokens`(归一化总输入含 cache),不是
+     * `e.promptTokens`(3.0.0 起恒 cache-MISS)——上游已切换口径,本函数只是照收不重算,值域/签名零改动。
      */
     record(principal: string | undefined, model: string, taskId: string, d: {
         inputTokens: number;

package/dist/fleet-client.js

@@ -31,6 +31,10 @@ export class FleetUsageAccumulator {
      * 同款=**传染**(同为单 producer 累加器:已知+未知=未知);center 聚合面的「下界 + costUnknown
      * 标记」语义在 registry-core aggregateUsage(跨 worker 审计总账层,两层判据见 [2103])。
      * 配额轴(weightedTokens)不受影响,仍逐笔精确。
+     *
+     * core 3.0.0([2296] 装车单 S4c)语义注:`d.inputTokens` 在唯一调用点(budget.ts 的
+     * `fleetUsage.record(...)`,S1 已迁)传的是 `e.totalInputTokens`(归一化总输入含 cache),不是
+     * `e.promptTokens`(3.0.0 起恒 cache-MISS)——上游已切换口径,本函数只是照收不重算,值域/签名零改动。
      */
     record(principal, model, taskId, d) {
         const p = principal ?? "__anonymous__"; // single-user turnkey: no principal — still meter, under a stable bucket

package/dist/observability/metrics.js

@@ -213,8 +213,13 @@ export function createMetrics() {
     m.counter("council_runs_total", "Code-review council runs (L1+L3)");
     m.histogram("council_tokens", "Total model tokens per council review (lenses + arbiter)");
     // Prefix-cache effectiveness (core 1.22): per-task hit rate (cost-critical) + a count of tasks the
-    // Runner flagged as low-hit (unstable/poisoned prefix) so it can be alerted on.
-    m.histogram("task_cache_hit_rate", "Prefix-cache hit rate per task (cachedTokens/promptTokens, 0..1)", [
+    // Runner flagged as low-hit (unstable/poisoned prefix) so it can be alerted on. The observed value is
+    // `result.stats.cacheHitRate` — core-computed (core 3.0.0: cachedTokens/totalInputTokens, the normalized
+    // total-including-cache denominator; runs.ts / http/server.ts just forward it), not derived here — this
+    // description names the field, not a formula this file owns (E3: don't restate a formula whose owner
+    // can change it without telling us; core 3.0.0's denominator switched from `promptTokens` to
+    // `totalInputTokens`, which the old "(cachedTokens/promptTokens)" text would have silently mis-described).
+    m.histogram("task_cache_hit_rate", "Prefix-cache hit rate per task (core-computed stats.cacheHitRate, 0..1)", [
         0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95, 1,
     ]);
     m.counter("prompt_cache_low_hit_total", "Tasks the Runner flagged with a low prefix-cache hit rate");

package/dist/plugins/remote-env-host.js

@@ -867,7 +867,8 @@ export class RemoteHostExecutionEnv {
             }
             catch (e) {
                 // 同步 spawn throw(如 cwd=文件的 ENOTDIR)——与管道旧形字节级对齐:executor 内 re-throw=Promise
-                // reject(消费方契约:core pruneWorktrees 等 await 抛才算失败;吞成 Result err 会让 reap 静默)。
+                // reject(消费方按 await 抛感知传输层失败;吞成 Result err 会让上层静默。⚠️ 引例更新:core 3.0.0
+                // 起 pruneWorktrees 改回执形 {ok},其消费点已双臂消费——worktree-isolation.ts;本处 spawn 层契约不变)。
                 closeSpoolFds();
                 cleanupSpool();
                 throw e;

package/dist/plugins/worktree-isolation.js

@@ -227,7 +227,12 @@ export async function reapOrphanWorktrees(baseEnvForGit, repoRoot, logger) {
     try {
         // #62:prune 动的是与 add/remove 同一张 `.git/worktrees` 注册表 ⇒ 必须过同一把 per-repoRoot 锁
         // (曾绕锁直跑;恶劣臂见 serializeGitAdmin 注)。
-        await serializeGitAdmin(repoRoot, () => pruneWorktrees(baseEnvForGit, norm(repoRoot)));
+        // core 3.0.0(#94 提货):pruneWorktrees 从「失败 throw」改为回执 `{ok:true}|{ok:false,detail}`——
+        // 只靠下面的 catch 发警在 3.0.0 下= prune 失败**静默化**(CI svc3 钉抓出)。回执臂与 throw 臂
+        // (serializeGitAdmin/env 传输层仍可抛)两路都进同一条 warn,判据不变:失败必须留痕。
+        const r = await serializeGitAdmin(repoRoot, () => pruneWorktrees(baseEnvForGit, norm(repoRoot)));
+        if (r.ok === false)
+            logger?.warn?.("worktree_prune_failed", { repoRoot: norm(repoRoot), error: r.detail });
     }
     catch (e) {
         logger?.warn?.("worktree_prune_failed", { repoRoot: norm(repoRoot), error: e instanceof Error ? e.message : String(e) });

package/dist/trace/project.d.ts

@@ -19,7 +19,13 @@ type AppendFn = (type: string, data: unknown) => Promise<void>;
 type GetEventsFn = (taskId: string, afterSeq: number) => Promise<RunEvent[]>;
 /** E8 (shell-host): drain the per-task model-usage accumulator into an append-only `model_usage` DELTA event.
  *  Shared by BOTH durable-write legs (runInBackground + the resume leg) so the persisted shape never drifts; the
- *  append-only deltas make a suspended/resumed run's total resume-safe (each leg appends its own, the echo sums). */
+ *  append-only deltas make a suspended/resumed run's total resume-safe (each leg appends its own, the echo sums).
+ *
+ *  core 3.0.0([2296] 装车单 S10,§DESIGN-V2 V2-⑤): every drained row is tagged with `usageBasis:
+ *  "uncached-components-v1"` — the write point for this label (the tracker itself, {@link ModelUsageTracker},
+ *  never sets it; it stays a pure token/cost accumulator). This is what lets {@link aggregateModelUsage} tell
+ *  apart rows written under the current component semantics from older rows (pre-dating this label) when it
+ *  sums across a run's whole event log. */
 export declare function appendModelUsageDelta(append: AppendFn, modelUsage: ModelUsageTracker | undefined, taskId: string): Promise<void>;
 /** [998]② drain the pending prompt-manifest records (whitelisted at RECORD time, budget.ts — ids/digests/
  *  counts only, no prompt bodies by core's manifest contract) into durable `prompt_assembled` events, one per
@@ -287,9 +293,15 @@ export interface TraceTurn {
     ts: string;
     role: "system" | "user" | "assistant" | "tool";
     blocks: TraceBlock[];
+    /** core 3.0.0([2296] 装车单 S3,§DESIGN-V2 V2-②):这是**本服务自有的投影契约**,不是 core wire 帧的
+     *  verbatim 转发——`in` 刻意保持它历史上的「总量」语义不变(旧帧/新帧读端零伤),不随 core 的
+     *  `promptTokens`/`turn_end.usage.inputTokens` 翻转为 cache-MISS。新增 `uncachedIn` 承接 core 3.0.0 才有的
+     *  cache-MISS 分量(命名取自 core `uncachedInputTokensOf`),只在能确认自己踩在新契约帧上时才发
+     *  (见 {@link toContractTokens})。 */
     tokens?: {
         in: number;
         out: number;
+        uncachedIn?: number;
         cacheRead?: number;
         cacheWrite?: number;
     };
@@ -326,7 +338,14 @@ export declare function turnEndEventData(ev: {
     stopReason?: string;
 }): Record<string, unknown>;
 /** Map core's `turn_end.usage` (1.76 / core [R59]: inputTokens/outputTokens/cacheReadTokens/cacheWriteTokens/
- *  costMicroUsd) → the contract's `TraceTurn.tokens {in,out,cacheRead?,cacheWrite?}`. Absent/empty → undefined. */
+ *  costMicroUsd; core 3.0.0 adds required `totalInputTokens`) → the contract's `TraceTurn.tokens`.
+ *  Absent/empty → undefined.
+ *
+ *  core 3.0.0([2296] 装车单 S3,§DESIGN-V2 V2-②,codex 复审「异议」采纳):`TraceTurn.tokens` 是本服务自有
+ *  的投影契约,不是 wire verbatim——`in` 保持它一直以来的**总量**语义,3.0.0 起改取 `totalInputTokens`
+ *  (归一化总输入含 cache)优先,`inputTokens` 回落(旧帧/尚未升级的 core 下 `totalInputTokens` 缺席时,
+ *  `in` 落回旧值,读端零伤)。`uncachedIn` 只在**能确认自己踩在新契约帧上**(`totalInputTokens` 在场)时
+ *  才发——旧帧的 `inputTokens` 是族依赖的旧语义,标成「uncachedIn(cache-MISS)」会是谎言。 */
 export declare function toContractTokens(usage: unknown): TraceTurn["tokens"] | undefined;
 export interface ProjectOpts {
     /** Prepend the user turn (the objective is not in the event log; the endpoint supplies it). */

package/dist/trace/project.js

@@ -1,11 +1,21 @@
 import { redactSecrets, redactDeep } from "./redact.js";
 /** E8 (shell-host): drain the per-task model-usage accumulator into an append-only `model_usage` DELTA event.
  *  Shared by BOTH durable-write legs (runInBackground + the resume leg) so the persisted shape never drifts; the
- *  append-only deltas make a suspended/resumed run's total resume-safe (each leg appends its own, the echo sums). */
+ *  append-only deltas make a suspended/resumed run's total resume-safe (each leg appends its own, the echo sums).
+ *
+ *  core 3.0.0([2296] 装车单 S10,§DESIGN-V2 V2-⑤): every drained row is tagged with `usageBasis:
+ *  "uncached-components-v1"` — the write point for this label (the tracker itself, {@link ModelUsageTracker},
+ *  never sets it; it stays a pure token/cost accumulator). This is what lets {@link aggregateModelUsage} tell
+ *  apart rows written under the current component semantics from older rows (pre-dating this label) when it
+ *  sums across a run's whole event log. */
 export async function appendModelUsageDelta(append, modelUsage, taskId) {
     const delta = modelUsage?.drain(taskId);
-    if (delta)
-        await append("model_usage", { usage: delta });
+    if (delta) {
+        const tagged = {};
+        for (const [model, d] of Object.entries(delta))
+            tagged[model] = { ...d, usageBasis: "uncached-components-v1" };
+        await append("model_usage", { usage: tagged });
+    }
 }
 /** [998]② drain the pending prompt-manifest records (whitelisted at RECORD time, budget.ts — ids/digests/
  *  counts only, no prompt bodies by core's manifest contract) into durable `prompt_assembled` events, one per
@@ -61,6 +71,11 @@ export function aggregateModelUsage(events) {
     const num = (v) => (typeof v === "number" && Number.isFinite(v) ? v : 0);
     const total = {};
     const costUnknown = new Set(); // RB-368:该模型至少有一行没带 costMicroUsd 键 ⇒ 总量不可知
+    // S10([2296] 装车单,§DESIGN-V2 V2-⑤):usageBasis 只在**这个模型贡献的每一行**都带同一个值时才透传到
+    // 回声行——混窗(升级前落盘的旧行 / 坏行 / 未来换了别的 basis 字符串)一律诚实省略,不猜、不传染出一个
+    // 假的一致标记。Map 记「目前为止这个模型见过的一致值」:第一行定值;之后任何一行不同(含缺席,记作
+    // undefined)⇒ 永久锁定为 undefined(一旦分歧就回不去)。
+    const basisSeen = new Map();
     let any = false;
     for (const ev of events) {
         if (ev.type !== "model_usage")
@@ -74,6 +89,7 @@ export function aggregateModelUsage(events) {
             // 之后,对唯一能判别的形(null)先抛,守卫不可达;坏行按整行 0 折(#59 同口径),cost 记未知。
             if (d === null || typeof d !== "object") {
                 costUnknown.add(model);
+                basisSeen.set(model, undefined); // 坏行 = 不可知,视同分歧,不透传 basis
                 continue;
             }
             const cur = total[model] ?? { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0, costMicroUsd: 0 };
@@ -87,11 +103,19 @@ export function aggregateModelUsage(events) {
                 cur.costMicroUsd = num(cur.costMicroUsd) + num(d.costMicroUsd);
             else
                 costUnknown.add(model);
+            const rowBasis = typeof d.usageBasis === "string" ? d.usageBasis : undefined;
+            if (!basisSeen.has(model))
+                basisSeen.set(model, rowBasis);
+            else if (basisSeen.get(model) !== rowBasis)
+                basisSeen.set(model, undefined);
             total[model] = cur;
         }
     }
     for (const model of costUnknown)
         delete total[model]?.costMicroUsd;
+    for (const [model, basis] of basisSeen)
+        if (basis !== undefined && total[model])
+            total[model].usageBasis = basis;
     return any ? total : undefined;
 }
 /** Identity fields core stamps on each content event (1.121.0 `TaskEventIdentity`, shell-host contract E2):
@@ -396,16 +420,24 @@ export function turnEndEventData(ev) {
     };
 }
 /** Map core's `turn_end.usage` (1.76 / core [R59]: inputTokens/outputTokens/cacheReadTokens/cacheWriteTokens/
- *  costMicroUsd) → the contract's `TraceTurn.tokens {in,out,cacheRead?,cacheWrite?}`. Absent/empty → undefined. */
+ *  costMicroUsd; core 3.0.0 adds required `totalInputTokens`) → the contract's `TraceTurn.tokens`.
+ *  Absent/empty → undefined.
+ *
+ *  core 3.0.0([2296] 装车单 S3,§DESIGN-V2 V2-②,codex 复审「异议」采纳):`TraceTurn.tokens` 是本服务自有
+ *  的投影契约,不是 wire verbatim——`in` 保持它一直以来的**总量**语义,3.0.0 起改取 `totalInputTokens`
+ *  (归一化总输入含 cache)优先,`inputTokens` 回落(旧帧/尚未升级的 core 下 `totalInputTokens` 缺席时,
+ *  `in` 落回旧值,读端零伤)。`uncachedIn` 只在**能确认自己踩在新契约帧上**(`totalInputTokens` 在场)时
+ *  才发——旧帧的 `inputTokens` 是族依赖的旧语义,标成「uncachedIn(cache-MISS)」会是谎言。 */
 export function toContractTokens(usage) {
     if (!usage || typeof usage !== "object")
         return undefined;
     const u = usage;
-    if (u.inputTokens == null && u.outputTokens == null)
+    if (u.inputTokens == null && u.totalInputTokens == null && u.outputTokens == null)
         return undefined;
     return {
-        in: u.inputTokens ?? 0,
+        in: u.totalInputTokens ?? u.inputTokens ?? 0,
         out: u.outputTokens ?? 0,
+        ...(u.totalInputTokens != null && u.inputTokens != null ? { uncachedIn: u.inputTokens } : {}),
         ...(u.cacheReadTokens != null ? { cacheRead: u.cacheReadTokens } : {}),
         ...(u.cacheWriteTokens != null ? { cacheWrite: u.cacheWriteTokens } : {}),
     };

package/dist/usage-analytics.d.ts

@@ -40,6 +40,10 @@ export interface UsageModelEntry {
     outputTokens?: number;
     cacheReadTokens?: number;
     cacheWriteTokens?: number;
+    /** S10(core 3.0 计量语义统一,§DESIGN-V2 V2-⑤)口径标记:`"uncached-components-v1"` = 三个输入键是
+     *  分量制(互不重叠,本文件的 tokensIn 求和公式正确)。缺席/别的值 = 口径不可分辨(存量行可能把
+     *  cache 计入 inputTokens,求和会双算)⇒ fold 侧标 {@link UsageTotals.estimated}。 */
+    usageBasis?: string;
     /** 整数 micro-USD。**缺席 = 未知**,不是 0(SDK `ModelUsageDelta` 同款单轨契约)。注意在场的 `0` 仍有
      *  歧义(引擎无价目表时恒发 0),那层消歧在部署面 `capabilities.pricingConfigured`——裁定见 budget.ts。 */
     costMicroUsd?: number;
@@ -49,9 +53,11 @@ export interface UsageTotals {
     tokensIn: number;
     tokensOut: number;
     costUsd: number;
-    /** 本聚合的口径**不完整**——两种来源:①有行走了 stats.tokens fallback(无 per-model echo,token 数为估算);
-     *  ②有 per-model echo 行**缺数值键**(该键按 0 计入,总量偏低;成本键缺席时尤其:少算 ≠ 免费)。
-     *  两者都只表达「别把这个总数当精确值」,消费端一视同仁。 */
+    /** 本聚合的口径**不完整**——三种来源:①有行走了 stats.tokens fallback(无 per-model echo,token 数为估算);
+     *  ②有 per-model echo 行**缺数值键**(该键按 0 计入,总量偏低;成本键缺席时尤其:少算 ≠ 免费);
+     *  ③(S10/V2-⑤)有行的 {@link UsageModelEntry.usageBasis} 缺席或非本读端认识的分量制——存量行某族
+     *  历史上把 cache 计入 inputTokens,tokensIn 求和会**双算**,口径不可分辨。
+     *  三者都只表达「别把这个总数当精确值」,消费端一视同仁。 */
     estimated: boolean;
 }
 /** GET /v1/usage/summary — 窗内总量。 */

package/dist/usage-analytics.js

@@ -36,6 +36,10 @@ function foldModelEntry(t, d) {
     t.costUsd += (cost ?? 0) / 1e6;
     if (input === undefined || output === undefined || cacheRead === undefined || cacheWrite === undefined || cost === undefined)
         t.estimated = true;
+    // S10/V2-⑤:上面这条 tokensIn 求和公式只在「分量制」行上正确。缺标记(存量行,某族历史上 inputTokens
+    // 含 cache ⇒ 双算)或本读端不认识的标记值(将来别的口径)⇒ 走既有不精确通道,不显示为精确值。
+    if (d.usageBasis !== "uncached-components-v1")
+        t.estimated = true;
 }
 function foldRow(t, row) {
     t.tasks += 1;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sema-agent/server",
-  "version": "3.23.0",
+  "version": "4.0.0",
   "description": "Sema Server — the server/API implementation layer for Sema, wiring core, registry, model providers, and cloud agent execution. Built on @sema-agent/core.",
   "type": "module",
   "license": "BUSL-1.1",
@@ -54,8 +54,8 @@
     "build:binary:run-local:darwin-arm64": "bun build --compile --target=bun-darwin-arm64 src/run-local.ts --outfile dist/run-local-darwin-arm64"
   },
   "dependencies": {
-    "@sema-agent/core": "^2.13.0",
-    "@sema-agent/registry-core": "^0.11.0",
+    "@sema-agent/core": "^3.0.0",
+    "@sema-agent/registry-core": "^0.12.0",
     "e2b": "^2.28.0",
     "libsodium-wrappers": "^0.8.4",
     "mysql2": "^3.22.4",