@sema-agent/server 5.6.0 → 5.7.0

package/dist/config.js

@@ -448,7 +448,7 @@ function parseStoreDomain(ctx) {
     const needsDb = needsTidb;
     return {
         sessionBackend,
-        sessionCacheTtlSec: Number(env("SESSION_CACHE_TTL_SEC", "300")),
+        sessionCacheTtlSec: numEnv("SESSION_CACHE_TTL_SEC", "300"), // A2 同族:NaN ⇒ `> 0` 假 ⇒ 缓存静默关
         rewindSnapshotMaxMb: optFinitePositiveEnv("REWIND_SNAPSHOT_MAX_MB"), // soft knob (S20: bad value warns + default)
         dbBackend,
         dbBackendExplicit: dbBackendSet,
@@ -748,7 +748,7 @@ function parseModelDomain() {
                 baseUrl: firstSetEnv("ANTHROPIC_BASE_URL"),
                 version: process.env.ANTHROPIC_VERSION,
                 cacheBreakpoints: (process.env.ANTHROPIC_CACHE_BREAKPOINTS ?? "true") === "true",
-                ...(process.env.ANTHROPIC_MAX_RETRIES ? { maxRetries: Number(env("ANTHROPIC_MAX_RETRIES", "2")) } : {}),
+                ...(process.env.ANTHROPIC_MAX_RETRIES ? { maxRetries: numEnv("ANTHROPIC_MAX_RETRIES", "2") } : {}), // A1 同族:NaN 会把 SDK 重试面折没
             }
             : undefined,
         resilience: {
@@ -756,16 +756,19 @@ function parseModelDomain() {
             // timer only arms AFTER the first token (bumpIdle lives in sawContentToken), so with these
             // two off a stream that opens but never emits hangs unbounded — the one true blind spot the
             // fleet audit found. Values follow the [880] recommendation; =0 restores off.
-            connectTimeoutMs: Number(env("MODEL_CONNECT_TIMEOUT_MS", "30000")),
-            firstTokenTimeoutMs: Number(env("MODEL_FIRST_TOKEN_TIMEOUT_MS", "120000")),
+            // 鲁棒性批1 A1(2026-08-03):这五个是 TASK_TIMEOUT_SEC 同族的**危险方向**键——NaN 在 core 的
+            // `if (x && x > 0)` 门下判假 ⇒ 看门狗/熔断静默关(设了收紧值反而更松)。当年 gap-sweep 的棘轮
+            // 正则只抓 process.env 直读拼法,这批「裸 Number 包 env() helper」从盲区漏网;门已补第二正则。
+            connectTimeoutMs: numEnv("MODEL_CONNECT_TIMEOUT_MS", "30000"),
+            firstTokenTimeoutMs: numEnv("MODEL_FIRST_TOKEN_TIMEOUT_MS", "120000"),
             // BEHAVIOR CHANGE (CC 196 parity): mid-stream stall watchdog default ON at 5min —
             // after output starts, 5min of zero deltas cancels the stream as a retryable [network] failure
             // (retry/failover recovers; without it a hung gateway pins the turn until the user Escapes).
             // MODEL_IDLE_TIMEOUT_MS=0 restores off.
-            idleTimeoutMs: Number(env("MODEL_IDLE_TIMEOUT_MS", "300000")),
+            idleTimeoutMs: numEnv("MODEL_IDLE_TIMEOUT_MS", "300000"),
             circuitBreaker: boolEnv("MODEL_CIRCUIT_BREAKER", false),
-            failureThreshold: Number(env("MODEL_CB_FAILURE_THRESHOLD", "5")),
-            cooldownMs: Number(env("MODEL_CB_COOLDOWN_MS", "30000")),
+            failureThreshold: numEnv("MODEL_CB_FAILURE_THRESHOLD", "5"),
+            cooldownMs: numEnv("MODEL_CB_COOLDOWN_MS", "30000"),
         },
         model,
         models,
@@ -1248,7 +1251,7 @@ function parseLimitsHttpDomain() {
         costQuotaWindowSec: numEnv("COST_QUOTA_WINDOW_SEC", "86400"),
         reapIntervalSec: numEnv("REAP_INTERVAL_SEC", "60"),
         runStaleSec: numEnv("REAP_RUN_STALE_SEC", "120"),
-        toolResultTtlSec: Number(env("TOOL_RESULT_TTL_SEC", "86400")),
+        toolResultTtlSec: numEnv("TOOL_RESULT_TTL_SEC", "86400"), // A2 同族:NaN ⇒ reapOlderThan(NaN) 永不回收且被 .catch 吞
     };
 }
 /** 域:observability(可观测)—— 指标/追踪令牌、工具与思考轨迹开关、日志档位、OTLP 出口。 */

package/dist/fleet-client.d.ts

@@ -16,6 +16,10 @@ import { UsageReport } from "@sema-agent/registry-core/fleet";
 /** Bound on distinct principal×model keys per window — a runaway/multi-tenant burst collapses into
  *  `__other__` instead of growing unbounded (same posture as budget.ts cardinalityGuard). */
 export declare const FLEET_USAGE_KEY_CAP = 1024;
+/** C1(2026-08-03):announce/usage/deregister 三站点的 fetch 上界(fleet-lease `LEASE_FETCH_TIMEOUT_MS`
+ *  同款判据的普适形)。比 lease 的 2500 略宽——这三条不在提交热路径,5s 足以吃掉慢网 RTT,又远小于
+ *  undici 默认 ~300s headersTimeout(此前半开 center 把 stop()/心跳各钉住到那个量级)。 */
+export declare const FLEET_CLIENT_FETCH_TIMEOUT_MS = 5000;
 /**
  * Per-window principal×model usage accumulator, fed from the tracer's `brain.call` hook (which carries the
  * ALS principal — reaches council/team sub-task spend too, same attribution point costQuota uses). Sync,

package/dist/fleet-client.js

@@ -16,6 +16,10 @@ import { InstanceAnnouncement, UsageReport, DEFAULT_HEARTBEAT_MS } from "@sema-a
 /** Bound on distinct principal×model keys per window — a runaway/multi-tenant burst collapses into
  *  `__other__` instead of growing unbounded (same posture as budget.ts cardinalityGuard). */
 export const FLEET_USAGE_KEY_CAP = 1024;
+/** C1(2026-08-03):announce/usage/deregister 三站点的 fetch 上界(fleet-lease `LEASE_FETCH_TIMEOUT_MS`
+ *  同款判据的普适形)。比 lease 的 2500 略宽——这三条不在提交热路径,5s 足以吃掉慢网 RTT,又远小于
+ *  undici 默认 ~300s headersTimeout(此前半开 center 把 stop()/心跳各钉住到那个量级)。 */
+export const FLEET_CLIENT_FETCH_TIMEOUT_MS = 5000;
 /**
  * Per-window principal×model usage accumulator, fed from the tracer's `brain.call` hook (which carries the
  * ALS principal — reaches council/team sub-task spend too, same attribution point costQuota uses). Sync,
@@ -122,7 +126,10 @@ export function startFleetClient(opts) {
             heartbeatMs,
         };
         try {
-            const res = await f(`${base}/api/fleet/instances`, { method: "POST", headers, body: JSON.stringify(body) });
+            // 鲁棒性批1 C1(2026-08-03):三条对 center 的 fetch 全部界上 abort(fleet-lease O1 同款判据的
+            // 普适形)——半开 center 下此前各挂 undici 默认 ~300s headersTimeout,stop() 的滚动更新被单个
+            // 半开连接钉住;心跳侧每 tick 叠一个在飞请求。
+            const res = await f(`${base}/api/fleet/instances`, { method: "POST", headers, body: JSON.stringify(body), signal: AbortSignal.timeout(FLEET_CLIENT_FETCH_TIMEOUT_MS) });
             if (!res.ok)
                 throw new Error(`announce ${res.status}`);
             if (failStreak > 0)
@@ -147,7 +154,7 @@ export function startFleetClient(opts) {
             return; // nothing accrued — skip the POST entirely
         const report = { worker: opts.center.worker, instanceId: opts.instanceId, windowStartMs: start, windowEndMs, entries };
         try {
-            const res = await f(`${base}/api/fleet/usage`, { method: "POST", headers, body: JSON.stringify({ reports: [report] }) });
+            const res = await f(`${base}/api/fleet/usage`, { method: "POST", headers, body: JSON.stringify({ reports: [report] }), signal: AbortSignal.timeout(FLEET_CLIENT_FETCH_TIMEOUT_MS) });
             if (!res.ok)
                 throw new Error(`usage ${res.status}`);
         }
@@ -171,7 +178,7 @@ export function startFleetClient(opts) {
             clearInterval(uw);
         await flushUsage(); // final partial window — best-effort
         try {
-            await f(`${base}/api/fleet/instances`, { method: "DELETE", headers, body: JSON.stringify({ instanceId: opts.instanceId, worker: opts.center.worker }) });
+            await f(`${base}/api/fleet/instances`, { method: "DELETE", headers, body: JSON.stringify({ instanceId: opts.instanceId, worker: opts.center.worker }), signal: AbortSignal.timeout(FLEET_CLIENT_FETCH_TIMEOUT_MS) });
         }
         catch {
             /* deregistration is best-effort — staleness (2× heartbeatMs) reaps us anyway */

package/dist/plugins/background-agent-store-sql.js

@@ -1,4 +1,4 @@
-import { BackgroundAgentStoreError, } from "@sema-agent/core";
+import { BackgroundAgentStoreError, STALE_RUNNING_REAP_ATTRIBUTION, } from "@sema-agent/core";
 import { pgSafeJsonStringify, pgSanitizeText } from "./pg-safe-json.js";
 export const BACKGROUND_AGENT_TABLE = "background_agents";
 const SUMMARY_COLS = "handle, owner, session_scoped, session_id, parent_session_id, root_session_id, name, agent_type, status, spawned_at_ms, updated_at_ms, settled_at_ms, usage_json";
@@ -66,11 +66,8 @@ function rowToSummary(r) {
             : {}),
     };
 }
-/** core 的共享归因句(`STALE_RUNNING_REAP_ATTRIBUTION`,dist/core/background-agent-store.js:44)——
- *  **值拷贝**:core 5.3.0 未把它放上包根(exports 只有 ./bench 与 ./fixtures 两条子路,deep import 会被
- *  exports 图拒绝),已在提货回执里请 core 上包根,到货后换 import。漂移钉=backgroundAgentStoreContract
- *  的 stale-running 条目逐字断言这句(store-contract-kit-db-integration §4.5,真双库)——core 改句必红。 */
-const STALE_RUNNING_REAP_ATTRIBUTION = "the host process was interrupted while this agent was running — its in-process state was lost. Check its worktree / output file for partial work before assuming the task landed (stale running row reaped).";
+// 共享归因句:core 5.4.0 起上包根([2436] 请求→[2437] 兑现),值拷贝退役——同一常量,双生与
+// InMemory 语义源头彻底同源;试剂盒条目仍逐字断言它(真双库),但漂移自此在编译期就不可能发生。
 /** core reap 的 staleRunning 翻转形(逐字段照抄 InMemory,background-agent-store.js:266-272):
  *  summary **和 error** 都只在缺席时补共享归因句——error 半边 5.2.0 起就该有,双生此前只写 summary
  *  且是手写旧句(core 5.3.0 新试剂盒首跑两方言一致揪红,[2425] bg-agent B 族同源),真失败文本永不覆写。 */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sema-agent/server",
-  "version": "5.6.0",
+  "version": "5.7.0",
   "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",
@@ -54,7 +54,7 @@
     "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": "^5.3.0",
+    "@sema-agent/core": "^5.4.0",
     "@sema-agent/registry-core": "^0.13.0",
     "e2b": "^2.28.0",
     "libsodium-wrappers": "^0.8.4",