@sema-agent/server 7.102.0 → 7.103.1

package/dist/boot/config-center.js

@@ -4,6 +4,8 @@ import { homedir } from "node:os";
 import { join } from "node:path";
 import { loadRemoteExec } from "@sema-agent/settings-schema/node";
 import { CENTER_CREDENTIAL_ENV, skillContentHash } from "@sema-agent/settings-schema";
+import { stableStringify } from "@sema-agent/settings-schema/hash";
+import { buildModelRefusalsSignal } from "../model-entry-refusal.js";
 import { CenterPromptSource, FilePromptArtifactStore, FilePromptSourceStateStore, MemoryPromptArtifactStore, MemoryPromptSourceStateStore, } from "@sema-agent/core";
 import { buildPricing } from "../budget.js";
 import { applyCenterPlugins } from "../capabilities/center-plugins.js";
@@ -19,7 +21,7 @@ import { createConfigProvider, raceBootFetch, BOOT_FETCH_DEFERRED } from "../con
 import { executionLaneCaps, executionLaneOf } from "../execution-lane-caps.js";
 import { createKeyResolver } from "../key-resolver.js";
 import { ensureSealedKeyStore, reportExecutionPublicKey } from "../sealed-key.js";
-import { applyEffective, applyCenterReadFace, mutateInPlace, logEffectiveDiff, applyCenterSkills, resolveMcpServers, resolveA2aPeers, restartReasons, modelPlaneChanged, withModelPlaneOf, buildEnvOnlyModelPlane, fetchPromptArtifact } from "../config-center/facade.js";
+import { applyEffective, applyCenterReadFace, mutateInPlace, logEffectiveDiff, applyCenterSkills, resolveMcpServers, resolveA2aPeers, restartReasons, modelPlaneChanged, withModelPlaneOf, bootSnapshotOf, fullyLandedForLkg, fetchPromptArtifact } from "../config-center/facade.js";
 import { createMcpRevocationLedger, enabledMcpEntries } from "../config-center/mcp-revocation.js";
 import { SHA256_HEX_RE } from "../digest-form.js";
 import { createConfigApplyLedger, modelPlaneCommitted, refusedGroups } from "../config-center/apply-ledger.js";
@@ -62,6 +64,7 @@ export async function createConfigCenterRuntime(ctx) {
     let planeDeferredNoHandoff;
     let bootSkillStale = false;
     let pendingRestart;
+    let modelRefusals;
     const cc = config.configCenter;
     const bootCredential = cc?.credential.read();
     const bootVerdict = bootCredential === undefined ? undefined : judgeCenterCredentialAtBoot(bootCredential);
@@ -162,6 +165,16 @@ export async function createConfigCenterRuntime(ctx) {
         if (!same)
             notifySelfReport("restart_signal");
     };
+    const publishModelRefusals = (next) => {
+        const same = stableStringify(modelRefusals ?? null) === stableStringify(next ?? null);
+        modelRefusals = next;
+        if (!same)
+            notifySelfReport("model_refusals");
+    };
+    const recordApply = (rep) => {
+        applyLedger.recordApply(rep);
+        publishModelRefusals(buildModelRefusalsSignal(rep.modelRefusals));
+    };
     let centerPrompts;
     let pendingCatalogFaces;
     const retryPendingCatalog = async (phase) => {
@@ -402,7 +415,7 @@ export async function createConfigCenterRuntime(ctx) {
                 generation: lkgGen,
                 report: (rep) => {
                     lkgReport = rep;
-                    applyLedger.recordApply(rep);
+                    recordApply(rep);
                 },
             });
             applyLedger.applied(lkgGen, lkgR.effective.version);
@@ -410,12 +423,14 @@ export async function createConfigCenterRuntime(ctx) {
             const lkgPlane = modelPlaneCommitted(lkgReport);
             if (lkgPlane)
                 markRosterLanded(lkgR.effective);
-            effective = lkgR.effective;
+            effective = bootSnapshotOf(lkgR.effective, lkgPlane ? lkgR.effective : undefined);
             if (lkgPlane)
                 appliedPlaneEff = lkgR.effective;
-            latestEffective = lkgR.effective;
-            if (lkgR.etag !== undefined)
-                ccEtag = lkgR.etag;
+            if (fullyLandedForLkg(lkgR, lkgReport)) {
+                latestEffective = lkgR.effective;
+                if (lkgR.etag !== undefined)
+                    ccEtag = lkgR.etag;
+            }
             const cfgModels = (lkgR.effective.models?.models ?? []).filter((m) => m.enabled !== false).length;
             logger.info("config_loaded", { source: "lkg", version: lkgR.effective.version, worker: cc?.worker ?? "(global)", models: cfgModels });
         };
@@ -450,15 +465,12 @@ export async function createConfigCenterRuntime(ctx) {
             const r = raced === BOOT_FETCH_DEFERRED ? null : raced;
             if (r) {
                 bootApplyPhase = true;
-                const bootPromptsRaw = r.effective.prompts;
-                const bootPromptsOk = bootPromptsRaw === undefined || validatePromptsDomain(bootPromptsRaw).ok;
-                let bootClean = (r.domainErrors ?? []).length === 0 && bootPromptsOk;
                 for (const de of r.domainErrors ?? [])
                     logger.warn("config_domain_invalid", { domain: de.domain, error: de.error, note: "tolerated at BOOT (corrupt file → schema defaults; dangling ref → reported as-is); the refresh path rejects such a candidate whole and keeps the LKG (§9.5-5)" });
                 warnDroppedEntries(r, "boot");
                 if (dryRun) {
                     logEffectiveDiff(config, r.effective, logger);
-                    if (bootClean)
+                    if (fullyLandedForLkg(r, undefined))
                         ccEtag = r.etag;
                 }
                 else {
@@ -471,12 +483,11 @@ export async function createConfigCenterRuntime(ctx) {
                         generation: bootGen,
                         report: (rep) => {
                             bootReport = rep;
-                            applyLedger.recordApply(rep);
+                            recordApply(rep);
                         },
                     });
                     applyLedger.applied(bootGen, r.effective.version);
-                    if (bootReport?.fullyApplied === false)
-                        bootClean = false;
+                    const bootClean = fullyLandedForLkg(r, bootReport);
                     if (bootClean)
                         ccEtag = r.etag;
                     await adoptCenterPrompts(r.effective, "boot");
@@ -494,7 +505,7 @@ export async function createConfigCenterRuntime(ctx) {
                     }
                     if (bootPlaneEff !== undefined)
                         markRosterLanded(bootPlaneEff);
-                    effective = bootPlaneEff === r.effective ? r.effective : withModelPlaneOf(r.effective, bootPlaneEff ?? buildEnvOnlyModelPlane());
+                    effective = bootSnapshotOf(r.effective, bootPlaneEff);
                     if (bootPlaneEff !== undefined)
                         appliedPlaneEff = bootPlaneEff;
                     if (bootClean)
@@ -553,8 +564,8 @@ export async function createConfigCenterRuntime(ctx) {
     let keyResolver;
     const mcpRevocations = createMcpRevocationLedger((ev) => logger.info("mcp_revocation_ledger", { ...(ev.revoked.length > 0 ? { revoked: ev.revoked } : {}), ...(ev.restored.length > 0 ? { restored: ev.restored } : {}) }));
     let mcpRevocationWired = false;
-    const restartCtx = (consolidationSeat) => ({
-        ...(config.degrade?.reactive ? { reactiveDegradeTo: config.degrade.to } : {}),
+    const restartCtx = (consolidationSeat, degradeRoute) => ({
+        degradeRoute,
         pluginsAllowHosts: config.pluginsAllowHosts,
         ...(consolidationSeat !== undefined ? { consolidationSeat } : {}),
         ...(readFaceEnvHeld.length > 0 ? { readFaceEnvHeld } : {}),
@@ -592,6 +603,7 @@ export async function createConfigCenterRuntime(ctx) {
             selfReportListeners.push(listener);
         },
         restartState: () => pendingRestart,
+        modelRefusalsState: () => modelRefusals,
         planeDeferredState: () => planeDeferredNoHandoff,
         configApplyState: () => (configProvider ? applyLedger.view() : undefined),
         centerManagedKeys: () => (configProvider ? centerManagedConfigKeys(latestEffective) : undefined),
@@ -724,7 +736,7 @@ export async function createConfigCenterRuntime(ctx) {
             }
         },
         startRefreshLoop(a) {
-            const { pricing, limitSync, swapRunnerDeps, ceilingVerdict, consolidationSeat, checkpointStorePresent } = a;
+            const { pricing, limitSync, swapRunnerDeps, ceilingVerdict, consolidationSeat, degradeRoute, checkpointStorePresent } = a;
             if (configProvider) {
                 const ccRef = config.configCenter;
                 let refreshInFlight = false;
@@ -800,7 +812,7 @@ export async function createConfigCenterRuntime(ctx) {
                                     generation: tickGen,
                                     report: (rep) => {
                                         applyReport = rep;
-                                        applyLedger.recordApply(rep);
+                                        recordApply(rep);
                                     },
                                     ...(planeChanged ? { swapPlane: swapRunnerDeps } : {}),
                                 });
@@ -833,7 +845,7 @@ export async function createConfigCenterRuntime(ctx) {
                                     return "ok";
                                 }
                                 const lkgPersisted = await persistLkgDurable(r.effective, r.etag);
-                                const reasons = restartReasons(effective, r.effective, restartCtx(consolidationSeat?.read()));
+                                const reasons = restartReasons(effective, r.effective, restartCtx(consolidationSeat?.read(), degradeRoute.read()));
                                 const skillsChangeDetected = reasons.includes("skills");
                                 if (bootSkillStale && !skillsChangeDetected)
                                     reasons.push("skills");
@@ -932,7 +944,7 @@ export async function createConfigCenterRuntime(ctx) {
                         generation: lateGen,
                         report: (rep) => {
                             applyReportLate = rep;
-                            applyLedger.recordApply(rep);
+                            recordApply(rep);
                         },
                         ...(planeChangedLate ? { swapPlane: swapRunnerDeps } : {}),
                     });
@@ -967,7 +979,7 @@ export async function createConfigCenterRuntime(ctx) {
                     latestEffective = r.effective;
                     ccEtag = r.etag;
                     const lkgPersistedLate = await persistLkgDurable(r.effective, r.etag);
-                    const reasons = restartReasons(undefined, r.effective, restartCtx(consolidationSeat?.read()));
+                    const reasons = restartReasons(undefined, r.effective, restartCtx(consolidationSeat?.read(), degradeRoute.read()));
                     if (lkgSurvivesRestart && !lkgPersistedLate) {
                         const blocked = reasons.splice(0, reasons.length);
                         planeDeferredNoHandoff = { version: r.effective.version, since: Date.now(), blocked, candidate: r.effective, ...(r.etag !== undefined ? { candidateEtag: r.etag } : {}) };

package/dist/boot/stage-01-config.js

@@ -12,6 +12,7 @@ import { installFileHistoryFailureRecorder } from "../file-history-disclosure.js
 import { installSecretEnvScrubRecorder } from "../observability/secret-env-scrub.js";
 import { setRunTerminalLogObserver } from "../observability/run-terminal.js";
 import { setRedactionObserver } from "../trace/redact.js";
+import { installOutboundDispatcher } from "../outbound-dispatcher.js";
 export function runStage01() {
     const config = loadConfig();
     const runStaleFailure = crossInvariant("run-stale-heartbeat").check({ live: config, next: config });
@@ -56,6 +57,7 @@ export function runStage01() {
             note: "LOCAL_DATA_ROOT moved the local backend's DATA root, but config.d / remote-exec.json are still read from the config root (LOCAL_DATA_ROOT is not consulted there). If you meant to move both, set CONFIG_LOCAL_DIR to the same directory.",
         });
     }
+    installOutboundDispatcher(logger);
     return { config, localRoot, logger, metrics };
 }
 //# sourceMappingURL=stage-01-config.js.map

package/dist/boot/stage-04-stores.d.ts

@@ -2,6 +2,7 @@ import type { runStage03 } from "./stage-03-store-backend.js";
 export declare function runStage04(ctx: Awaited<ReturnType<typeof runStage03>>): Promise<{
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-04-stores.js

@@ -27,7 +27,7 @@ export async function runStage04(ctx) {
     if (adoptionBoot.scanned > 0 || adoptionBoot.error !== undefined) {
         logger.info("adoption_boot_scan", { ...adoptionBoot });
     }
-    const { brain, pricing, counterDegradeHook, costQuota, modelUsageTracker, promptManifestTracker, fleetUsage, fleetLease, tracer, sideQueryAccounting, toolResultStore, sessionPolicyStore, fileHistoryStore, fileHistoryErasure, } = createBudgetAndTracing({ config, logger, metrics, backend, breakerState });
+    const { brain, degradeRouteWatch, pricing, counterDegradeHook, costQuota, modelUsageTracker, promptManifestTracker, fleetUsage, fleetLease, tracer, sideQueryAccounting, toolResultStore, sessionPolicyStore, fileHistoryStore, fileHistoryErasure, } = createBudgetAndTracing({ config, logger, metrics, backend, breakerState });
     const permissionRuleStores = config.permissionRulesEnabled && backend ? backend.permissionRule() : undefined;
     const permissionRuleProvider = permissionRuleStores
         ? createPermissionRuleStoreProvider({ durable: permissionRuleStores.durable })
@@ -44,6 +44,6 @@ export async function runStage04(ctx) {
         });
     }
     await auditDormantPermissionRules({ stores: permissionRuleStores, logger, explicit: config.permissionRulesEnabledExplicit });
-    return { ...ctx, brain, costQuota, counterDegradeHook, fileHistoryErasure, fileHistoryStore, fleetLease, fleetUsage, memoryBundleFaces, memoryComplianceFaces, memoryOriginFaces, modelUsageTracker, permissionRuleProvider, permissionRuleStores, pricing, promptManifestTracker, ruleConsent, sessionMemoryStatusFace, sessionPolicyStore, sideQueryAccounting, toolResultStore, tracer };
+    return { ...ctx, brain, costQuota, degradeRouteWatch, counterDegradeHook, fileHistoryErasure, fileHistoryStore, fleetLease, fleetUsage, memoryBundleFaces, memoryComplianceFaces, memoryOriginFaces, modelUsageTracker, permissionRuleProvider, permissionRuleStores, pricing, promptManifestTracker, ruleConsent, sessionMemoryStatusFace, sessionPolicyStore, sideQueryAccounting, toolResultStore, tracer };
 }
 //# sourceMappingURL=stage-04-stores.js.map

package/dist/boot/stage-05-execution-env.d.ts

@@ -15,6 +15,7 @@ export declare function runStage05(ctx: Awaited<ReturnType<typeof runStage04>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-06-runners.d.ts

@@ -76,6 +76,7 @@ export declare function runStage06(ctx: Awaited<ReturnType<typeof runStage05>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-07-capability-layer.d.ts

@@ -90,6 +90,7 @@ export declare function runStage07(ctx: Awaited<ReturnType<typeof runStage06>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-08-reapers.d.ts

@@ -103,6 +103,7 @@ export declare function runStage08(ctx: Awaited<ReturnType<typeof runStage07>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-08-reapers.js

@@ -5,7 +5,7 @@ import { startRetentionLane, RETENTION_LEASE_TTL_FACTOR } from "./retention-lane
 import { createLimitSync } from "./limit-sync.js";
 import { runDenySweep } from "./late-binds.js";
 export function runStage08(ctx) {
-    const { approvalNonceStore, backend, backgroundAgentStore, ceilingVerdict, checkpointStore, config, configCenter, consolidationSeatWatch, costQuota, fileHistoryErasure, handslessRunner, handslessSubRunner, hookAgentRunner, imageBakes, instanceId, localRoot, logger, mailboxStore, metrics, permissionRuleStores, pricing, rateLimiter, retentionExecutor, retentionLaneStore, rosterStore, runStore, runner, sqlWorkflowRunStore, subRunner, taskAttachmentStore, toolApproval, toolResultStore, workflowJournalStore, workflowNotifyGate, workflowNotifyJournal, workflowRecoverOpts, worktreeReap } = ctx;
+    const { approvalNonceStore, backend, backgroundAgentStore, ceilingVerdict, checkpointStore, config, configCenter, consolidationSeatWatch, costQuota, degradeRouteWatch, fileHistoryErasure, handslessRunner, handslessSubRunner, hookAgentRunner, imageBakes, instanceId, localRoot, logger, mailboxStore, metrics, permissionRuleStores, pricing, rateLimiter, retentionExecutor, retentionLaneStore, rosterStore, runStore, runner, sqlWorkflowRunStore, subRunner, taskAttachmentStore, toolApproval, toolResultStore, workflowJournalStore, workflowNotifyGate, workflowNotifyJournal, workflowRecoverOpts, worktreeReap } = ctx;
     const reaper = startReapers({
         config, logger, metrics, localRoot, backend, subRunner, runStore, checkpointStore,
         rateLimiter, costQuota, toolResultStore, fileHistoryErasure, taskAttachmentStore, imageBakes, worktreeReap,
@@ -71,7 +71,7 @@ export function runStage08(ctx) {
             }
         }
     };
-    configCenter.startRefreshLoop({ pricing, limitSync, swapRunnerDeps, ceilingVerdict, checkpointStorePresent: checkpointStore !== undefined, ...(consolidationSeatWatch !== undefined ? { consolidationSeat: consolidationSeatWatch } : {}) });
+    configCenter.startRefreshLoop({ pricing, limitSync, swapRunnerDeps, ceilingVerdict, checkpointStorePresent: checkpointStore !== undefined, degradeRoute: degradeRouteWatch, ...(consolidationSeatWatch !== undefined ? { consolidationSeat: consolidationSeatWatch } : {}) });
     configCenter.initKeyResolver();
     return { ...ctx, otelExporter, reaper, retentionLaneStarted };
 }

package/dist/boot/stage-09-leader.d.ts

@@ -137,6 +137,7 @@ export declare function runStage09(ctx: Awaited<ReturnType<typeof runStage08>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-10-http-server.d.ts

@@ -137,6 +137,7 @@ export declare function runStage10(ctx: Awaited<ReturnType<typeof runStage09>>):
     worktreeReap: (() => Promise<void>) | undefined;
     brain: import("@sema-agent/core").Brain;
     costQuota: import("../observability/cost-quota.js").CostQuota | import("../plugins/pg-cost-quota.js").PgCostQuota | import("../plugins/tidb-cost-quota.js").TiDBCostQuota;
+    degradeRouteWatch: import("../brain.js").DegradeRouteWatch;
     counterDegradeHook: (info: {
         table: string;
         kind: string;

package/dist/boot/stage-10-http-server.js

@@ -1,6 +1,6 @@
 import { autoRunOnRecommendationEffective, defaultTaskRegistry } from "@sema-agent/core";
 import { posIntEnv } from "../session-watch.js";
-import { brainSummary } from "../brain.js";
+import { brainSummary, buildFailoverFirstByteNotice } from "../brain.js";
 import { resolveBindHost } from "../config.js";
 import { buildWriteProtectionPosture } from "../write-protection.js";
 import { isPricingConfigured } from "../budget.js";
@@ -151,6 +151,7 @@ export async function runStage10(ctx) {
         planeDeferredState: () => configCenter.planeDeferredState(),
         configApplyState: () => configCenter.configApplyState(),
         centerCredentialState: () => configCenter.centerCredentialState(),
+        modelRefusalsState: () => configCenter.modelRefusalsState(),
         mcpRevocations,
         mcpRevocationState: () => configCenter.mcpRevocationState(),
         configRefreshNow: () => configCenter.refreshNow(),
@@ -241,12 +242,16 @@ export async function runStage10(ctx) {
         version: serviceVersion(),
         readiness: () => readinessOf(deployment),
         restart: () => deployment.restartState?.(),
+        modelRefusals: () => deployment.modelRefusalsState?.(),
         ...(fleetUsage ? { usage: fleetUsage } : {}),
         logger: { info: (m, d) => logger.info(m, d ?? {}), warn: (m, d) => logger.warn(m, d ?? {}) },
     });
     configCenter.onSelfReportChange(() => void fleetClient?.announceNow());
     if (fleetClient)
         logger.info("fleet_client_started", { worker: config.configCenter?.worker, address: buildDisplaySafeUrl(process.env.FLEET_ADVERTISE_ADDRESS ?? "") });
+    const failoverFirstByte = buildFailoverFirstByteNotice(config);
+    if (failoverFirstByte)
+        logger.warn("failover_first_byte_lane", failoverFirstByte);
     logger.info("listening", {
         port: config.port,
         bindHost: bindHost ?? "0.0.0.0/::(all interfaces)",

package/dist/brain.d.ts

@@ -1,5 +1,6 @@
-import { type Brain, type BreakerState } from "@sema-agent/core";
+import { type Brain, type BreakerState, type Model } from "@sema-agent/core";
 import type { ServiceConfig } from "./config-types.js";
+import { type RestartSliceCtx } from "./config-center/restart-signal.js";
 /**
  * 配了快切层(failover ≥2 路 / 断路器)时,**每路**的重试上限。
  *
@@ -12,12 +13,14 @@ export declare const FAST_FAIL_MAX_RETRIES = 2;
  * Build the external "Brain" (LLM) from config.
  *
  * Composition (A1 + the 1.38 resilience stack, ordered timeout-innermost → breaker → failover-outermost):
- *   - **Tiered timeouts** (1.38, `connectTimeoutMs`/`firstTokenTimeoutMs`) are passed into each
- *     openai-compatible brain at creation — they bound "gateway never accepts the connection" and
- *     "stream opened but hangs without a token" as retryable `[network]` failures. Default ON
- *     (connect 30s per the fleet-audit recommendation; first-token 600s since S-202 — a self-hosted backend's long
- *     prefill legitimately takes minutes to the first byte, and 120s turned that into a retry handing the
- *     same prompt back to the same queue. env =0 disables; env =120000 restores the old posture).
+ *   - **Tiered timeouts** (1.38; re-cut by core 7.33.0 along the segments of a call) are passed into each
+ *     brain at creation: `firstByteTimeoutMs` (fetch() → response headers; S-753: **passed only when the
+ *     operator set `MODEL_FIRST_BYTE_TIMEOUT_MS`** — absent = no key, core's own default, which carries an
+ *     upload allowance and one widened retry of its own), `firstTokenTimeoutMs` (headers → first delta,
+ *     600s since S-202 — a self-hosted backend's long prefill legitimately takes minutes, and 120s turned that
+ *     into a retry handing the same prompt back to the same queue) and `idleTimeoutMs`. The TCP + TLS connect
+ *     is the transport's own clock (core 7.33.0), not a knob here. env =0 = no engine clock on that segment
+ *     and the transport's header / body clocks off too (the connect keeps the transport's own clock).
  *   - **Circuit breaker** (`createCircuitBreakerBrain`) wraps the *primary* gateway route(s) when
  *     `MODEL_CIRCUIT_BREAKER=true`: after N transient failures it opens and fast-fails, so failover
  *     switches to the bare backup immediately instead of waiting out each timeout. The last gateway
@@ -30,14 +33,62 @@ export declare const FAST_FAIL_MAX_RETRIES = 2;
  *     ↔ vLLM is a *routing* choice, not failover.)
  *
  * When no fallback URL, no Anthropic key, and no resilience thresholds are configured, this returns the
- * bare openai brain (carrying the default-on watchdog timeouts — set the three env knobs to 0 for the
- * pre-[ref] fully-bare shape). The Model itself (id / contextWindow / maxTokens / provider) flows
+ * bare openai brain (carrying the watchdogs: core's own first-byte default unless `MODEL_FIRST_BYTE_TIMEOUT_MS`
+ * is set, plus the server's first-token / idle defaults; a knob at 0 = no engine clock on that segment and the
+ * transport's header / body clocks off too). The Model itself (id / contextWindow / maxTokens / provider) flows
  * per-request through `Runner({ models })`.
  */
 export declare function createBrain(config: ServiceConfig, deps?: {
     fetchImpl?: typeof fetch;
     breakerState?: BreakerState;
 }): Brain;
+/** {@link degradeRouteFingerprint} 读的那几张表(全是活引用:refresh 就地换代,读数每次现取)。 */
+export type DegradeRoutePlane = Pick<ServiceConfig, "degrade" | "models" | "gatewayBaseUrl" | "modelApiKeyEnv" | "modelApiKeys">;
+/**
+ * S-794:反应式降级壳焊的是哪一只目标 —— **唯一判据**(createBrain 组壳、启动摘要、`degrade-route` 片的两侧读数同读这一只)。
+ * 车道没开(非 reactive)或表里没有叫 `degrade.to` 的条目 ⇒ `undefined`(createBrain 不装降级壳)。
+ */
+export declare function degradeTargetOf(config: Pick<ServiceConfig, "degrade" | "models">): Model | undefined;
+/**
+ * S-794:降级壳的**路由指纹**(纯数据)—— `degrade-route` 重启片 `baked` / `current` 两侧读数的唯一读法。
+ *
+ * 壳在 createBrain 那一刻焊死目标那只 `Model` 对象(env 派生或目录派生):请求打 `model.baseUrl`(缺席回网关根)、body 带
+ * `model.id` / `maxTokens` / `compat` / …,走哪条腿看 `provider`;refresh 就地换的是表里的**值**,够不到那只对象。于是指纹 =
+ *  · 那只 Model 经 `catalogEntryFingerprint`(与 swap 门同一张热字段减法:单价**数值**热、在场性不热;其余字段一律算 —— deny-list,
+ *    漂移的失败形是多签一次,绝不漏签);
+ *  · 有效端点根(S-623 同一只解析器 `openaiRouteBaseUrl(baseUrl, gatewayBaseUrl)`);
+ *  · 它配的凭据**身份**:`apiKeyEnv`(变量名,不是值)与托管密钥**在场位**(不取值:值是热的,且是秘密)。两次查表的键与
+ *    `fbOwnApiKey` 同(目标的真名,再 `degrade.to`)。
+ * 没有焊壳(车道没开 / 没有目标)⇒ `null`。在场 ↔ 缺席的翻转因此自然是一次变化。
+ */
+export declare function degradeRouteFingerprint(config: DegradeRoutePlane): unknown;
+/** `degrade-route` 片的席:boot 焊进去的那一只 vs 活配置此刻的同一只目标(`restartReasons` 的 `degradeRoute` 两侧读数)。 */
+export interface DegradeRouteWatch {
+    read(): NonNullable<RestartSliceCtx["degradeRoute"]>;
+}
+/**
+ * `create*`:返回带捕获状态(组壳那一刻的指纹)的活对象。**须在 {@link createBrain} 的同一时刻、同一份 config 上调**
+ * (生产:boot/budget-tracing.ts 紧随 createBrain)—— 否则 `baked` 就不是壳真焊进去的那一只。与 boot/memory-consolidation.ts
+ * `createConsolidationSeatWatch` 同形。
+ */
+export declare function createDegradeRouteWatch(config: DegradeRoutePlane): DegradeRouteWatch;
+/**
+ * S-753 复审(opus F2 + codex C1)—— **failover 拓扑 × 首字节段没有部署窗**的启动期披露(纯数据,`build*`;告警由 boot 出)。
+ *
+ * 缺省下首字节段不铸键(core 自缺省当家,[ref])、`0` = 该段不设引擎钟;而切备只发生在一次调用**失败之后**。于是主网关接了连接
+ * 却不回响应头时:缺省 ⇒ 要等 core 的首字节 lane 走完(缺省首窗 + 一次加宽重试;本仓不铸 `firstByteRetries`,S-753 车PA)才切备,
+ * 时间预算短于那一段的调用在备网关被尝试之前就被取消 —— 以 core 7.33.x 缺省计 lane ≈ 15 min,code-review council 每个镜头的 180 s 与 team
+ * discuss 的 600 s 都在其内(车PA codex r1:600 s 在主网关的加宽重试窗里到点,不是「只剩余量」);
+ * `0` ⇒ 首字节 lane 永不切备。启动日志里同行的 `gatewayRetryCap:"fast-fail:…"` 只管共享重试梯子,读起来却像「会快切」—— 这条披露把它补齐。
+ * 显式设了正数窗 ⇒ `undefined`(部署已表态);单路由 ⇒ `undefined`(没有备可切)。**不替运维铸窗**:出路是旋钮,不是本仓的数。
+ */
+export declare function buildFailoverFirstByteNotice(config: ServiceConfig): {
+    readonly knob: "MODEL_FIRST_BYTE_TIMEOUT_MS";
+    readonly fallbacks: number;
+    readonly firstByteTimeoutMs: number | null;
+    readonly effect: string;
+    readonly fix: string;
+} | undefined;
 /** One-line summary of the active brain composition, for the startup log. */
 export declare function brainSummary(config: ServiceConfig): Record<string, unknown>;
 //# sourceMappingURL=brain.d.ts.map

package/dist/brain.js

@@ -4,6 +4,7 @@ import { buildDefaultResilience } from "./config-types.js";
 import { resolveModelApiKey } from "./key-resolver.js";
 import { dispatchBrainLane } from "./model-provider.js";
 import { openaiRouteBaseUrl } from "./model-route-endpoint.js";
+import { catalogEntryFingerprint } from "./config-center/restart-signal.js";
 import { buildDisplaySafeRefusal, buildDisplaySafeUrl } from "./trace/redact.js";
 function isOffDeploymentRoot(baseUrl, config) {
     if (!baseUrl)
@@ -38,7 +39,7 @@ export function createBrain(config, deps = {}) {
     const { fetchImpl, breakerState } = deps;
     const r = config.resilience ?? DEFAULT_RESILIENCE;
     const timeouts = {
-        connectTimeoutMs: r.connectTimeoutMs,
+        ...(r.firstByteTimeoutMs !== undefined ? { firstByteTimeoutMs: r.firstByteTimeoutMs } : {}),
         firstTokenTimeoutMs: r.firstTokenTimeoutMs,
         idleTimeoutMs: r.idleTimeoutMs,
     };
@@ -93,7 +94,7 @@ export function createBrain(config, deps = {}) {
         : gatewayBrain;
     if (config.degrade?.reactive) {
         const degradeTo = config.degrade.to;
-        const fallbackModel = config.models[degradeTo];
+        const fallbackModel = degradeTargetOf(config);
         if (fallbackModel) {
             const fbOwnApiKey = () => resolveModelApiKey(fallbackModel.name, config.modelApiKeyEnv ?? {}, process.env, config.modelApiKeys ?? {}) ??
                 resolveModelApiKey(degradeTo, config.modelApiKeyEnv ?? {}, process.env, config.modelApiKeys ?? {});
@@ -135,11 +136,46 @@ export function createBrain(config, deps = {}) {
     }
     return brain;
 }
+export function degradeTargetOf(config) {
+    return config.degrade?.reactive ? config.models?.[config.degrade.to] : undefined;
+}
+export function degradeRouteFingerprint(config) {
+    const target = degradeTargetOf(config);
+    if (target === undefined || config.degrade === undefined)
+        return null;
+    const to = config.degrade.to;
+    return {
+        model: catalogEntryFingerprint(target),
+        endpoint: openaiRouteBaseUrl(target.baseUrl, config.gatewayBaseUrl) ?? null,
+        apiKeyEnv: config.modelApiKeyEnv?.[target.name] ?? config.modelApiKeyEnv?.[to] ?? null,
+        sealedKey: config.modelApiKeys?.[target.name] !== undefined || config.modelApiKeys?.[to] !== undefined,
+    };
+}
+export function createDegradeRouteWatch(config) {
+    const baked = degradeRouteFingerprint(config);
+    return { read: () => ({ baked, current: degradeRouteFingerprint(config) }) };
+}
+export function buildFailoverFirstByteNotice(config) {
+    const r = config.resilience ?? DEFAULT_RESILIENCE;
+    const fallbacks = config.gatewayFallbackUrls.length;
+    if (fallbacks === 0 || (r.firstByteTimeoutMs !== undefined && r.firstByteTimeoutMs > 0))
+        return undefined;
+    const zero = r.firstByteTimeoutMs === 0;
+    return {
+        knob: "MODEL_FIRST_BYTE_TIMEOUT_MS",
+        fallbacks,
+        firstByteTimeoutMs: r.firstByteTimeoutMs ?? null,
+        effect: zero
+            ? "the first-byte lane never fails over a primary that accepts the connection but sends no response headers (0 = no engine clock on that segment); only a run-level limit ends that wait"
+            : "a primary that accepts the connection but sends no response headers is failed over only after core's default first-byte lane runs out (the first window, then one widened retry); a call with a time budget shorter than that — the code-review council's 180 s lenses and team discuss's 600 s, with core 7.33.x defaults — is cancelled before the backup is ever tried",
+        fix: "set MODEL_FIRST_BYTE_TIMEOUT_MS to the longest response-header wait a healthy primary needs (queue + prefill), so a hung primary is cut over after that window and its one widened retry",
+    };
+}
 export function brainSummary(config) {
     const r = config.resilience ?? DEFAULT_RESILIENCE;
     const hasFailover = config.gatewayFallbackUrls.length > 0;
     const breakerActive = r.circuitBreaker && hasFailover;
-    const degradeTarget = config.degrade?.reactive ? config.models?.[config.degrade.to] : undefined;
+    const degradeTarget = degradeTargetOf(config);
     const degradeEndpoint = degradeTarget ? openaiRouteBaseUrl(degradeTarget.baseUrl, config.gatewayBaseUrl) : undefined;
     return {
         gateway: buildDisplaySafeUrl(config.gatewayBaseUrl ?? "") || null,
@@ -147,9 +183,9 @@ export function brainSummary(config) {
         failover: hasFailover,
         anthropicRoute: Boolean(config.anthropic),
         anthropicCacheBreakpoints: config.anthropic?.cacheBreakpoints ?? null,
-        connectTimeoutMs: r.connectTimeoutMs || null,
-        firstTokenTimeoutMs: r.firstTokenTimeoutMs || null,
-        idleTimeoutMs: r.idleTimeoutMs || null,
+        firstByteTimeoutMs: r.firstByteTimeoutMs ?? null,
+        firstTokenTimeoutMs: r.firstTokenTimeoutMs,
+        idleTimeoutMs: r.idleTimeoutMs,
         circuitBreaker: breakerActive,
         circuitBreakerRequested: r.circuitBreaker,
         gatewayRetryCap: config.gatewayMaxRetries !== undefined
@@ -158,10 +194,10 @@ export function brainSummary(config) {
                 ? `fast-fail:${FAST_FAIL_MAX_RETRIES}`
                 : "engine-default",
         circuitBreakerNoop: r.circuitBreaker && !hasFailover ? "needs ≥2 gateway routes" : null,
-        reactiveDegradeTo: config.degrade?.reactive && config.models?.[config.degrade.to] ? config.degrade.to : null,
+        reactiveDegradeTo: degradeTarget !== undefined ? (config.degrade?.to ?? null) : null,
         reactiveDegradeBaseUrl: buildDisplaySafeUrl(degradeEndpoint ?? "") || null,
         ...((() => {
-            const target = config.degrade?.reactive ? config.models?.[config.degrade.to] : undefined;
+            const target = degradeTarget;
             if (!target)
                 return { reactiveDegradeRoutePairing: null, reactiveDegradeRouteRefusal: null };
             if (dispatchBrainLane(target.provider, config.anthropic !== undefined) === "anthropic") {

package/dist/capabilities/center-plugins.d.ts

@@ -39,14 +39,19 @@ export type PluginValidation = {
 };
 /** 条款 1/2/4:https-only、host 全等白名单、sha 40hex。纯函数,worker 侧纵深复验(不信任 wire 已验过)。 */
 export declare function validatePluginDecl(decl: PluginDeclWire, allowHosts: readonly string[]): PluginValidation;
-/** 条款 3/5:clone 计划(argv 序列 + 隔离 env)。纯数据,让"姿势"可被锚定。 */
+/**
+ * 条款 3:clone 计划(argv 序列)。纯数据,让"步骤"可被锚定。
+ *
+ * 7.103.0(S-684 残余):计划**只**产 argv。隔离 env(收紧③:GIT_CONFIG_GLOBAL/SYSTEM=/dev/null、GIT_TERMINAL_PROMPT=0)的唯一属主是
+ * `gitSubprocessEnv({ isolateConfig: true })` —— 修前这三键写在这里、由 `runGit` 叠在 `{ ...process.env }` 上,一件姿势两个写者,
+ * 而剥密恰好落在两者之外。锚点随之移到真 spawn 的 env 上(`test/secret-env-scrub-host-legs.test.ts` R2:假 git 读到的整份 env)。
+ */
 export declare function gitClonePlan(o: {
     url: string;
     dest: string;
     sha?: string;
 }): {
     steps: string[][];
-    env: Record<string, string>;
 };
 /** manifest 本体尺寸上限(#97 redesign⑥ V2-⑦:**新的本地输入上限**,非既有格式事实——manifest 只被
  *  消费一个 `skillsPath` 字段,几行 JSON;超界视为畸形输入整条拒装)。 */
@@ -101,6 +106,7 @@ export declare function applyCenterPlugins(baseline: LoadedSkill[], eff: {
     /** 测试口:跳过 validate(materialize 只接受已验声明;真 git 腿用本地 bare repo,validate 层的
      *  https-only 单独测——两层分开,谁也不掩护谁)。生产恒 false/缺席。 */
     skipValidate?: boolean;
-    runGit?: (args: string[], env: Record<string, string>) => Promise<void>;
+    /** 测试口:替换真 git(生产恒缺席)。env 不在这一格上 —— 缺省腿自己经唯一构造取(替身腿本来就不起 git)。 */
+    runGit?: (args: string[]) => Promise<void>;
 }): Promise<ApplyPluginsResult>;
 //# sourceMappingURL=center-plugins.d.ts.map

package/dist/capabilities/center-plugins.js

@@ -4,6 +4,7 @@ import { promises as fsp, realpathSync, existsSync, readFileSync, lstatSync, sta
 import { join, resolve, isAbsolute, sep } from "node:path";
 import { assertConfined, loadSkills } from "./skills.js";
 import { buildDisplaySafeUrl, redactSecrets } from "../trace/redact.js";
+import { gitSubprocessEnv } from "../server-secret-env.js";
 const execFileP = promisify(execFile);
 export const MAX_PLUGIN_SOURCE_URL_CHARS = 2048;
 export const DEFAULT_PLUGIN_ALLOW_HOSTS = ["github.com"];
@@ -53,13 +54,8 @@ export function validatePluginDecl(decl, allowHosts) {
     return { ok: true, name, url };
 }
 export function gitClonePlan(o) {
-    const env = {
-        GIT_CONFIG_GLOBAL: "/dev/null",
-        GIT_CONFIG_SYSTEM: "/dev/null",
-        GIT_TERMINAL_PROMPT: "0",
-    };
     if (o.sha === undefined) {
-        return { steps: [["clone", "--depth", "1", "--quiet", o.url, o.dest]], env };
+        return { steps: [["clone", "--depth", "1", "--quiet", o.url, o.dest]] };
     }
     return {
         steps: [
@@ -67,7 +63,6 @@ export function gitClonePlan(o) {
             ["-C", o.dest, "fetch", "--depth", "1", "--quiet", o.url, o.sha],
             ["-C", o.dest, "checkout", "--quiet", o.sha],
         ],
-        env,
     };
 }
 export const PLUGIN_MANIFEST_MAX_BYTES = 64 * 1024;
@@ -130,8 +125,8 @@ export async function applyCenterPlugins(baseline, eff, opts) {
     const out = { skills: [...baseline], failures: [], letGo: [], lkgUsed: [] };
     if (declared.length === 0)
         return out;
-    const runGit = opts.runGit ?? (async (args, env) => {
-        await execFileP("git", args, { env: { ...process.env, ...env }, timeout: 120_000, maxBuffer: 4 * 1024 * 1024 });
+    const runGit = opts.runGit ?? (async (args) => {
+        await execFileP("git", args, { env: gitSubprocessEnv({ site: "server.center-plugins.git-clone", isolateConfig: true }), timeout: 120_000, maxBuffer: 4 * 1024 * 1024 });
     });
     const taken = new Set(baseline.map((s) => s.spec.name));
     for (const decl of declared) {
@@ -155,7 +150,7 @@ export async function applyCenterPlugins(baseline, eff, opts) {
             await fsp.mkdir(staging, { recursive: true });
             const plan = gitClonePlan({ url: v.url, dest: staging, sha: v.sha });
             for (const step of plan.steps)
-                await runGit(step, plan.env);
+                await runGit(step);
             await fsp.rm(liveDir, { recursive: true, force: true });
             await fsp.mkdir(join(opts.cacheRoot, v.name), { recursive: true });
             await fsp.rename(staging, liveDir);