系统自评估与建议能力——观测补齐+规则引擎+LLM 分析-agent-runtime-AtomGit

Mapplications/agent_runtime/CLAUDE.md+1-1文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/agent_runtime.server.env.example+8-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/deploy/agent_runtime.env.example+11-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/deploy/agent_runtime.template.yaml+14-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/docs/api/config-plane-api.md+1365-0

Mapplications/agent_runtime/docs/design/Agent-Runtime-HLD.md+19-1

	@@ -609,7 +609,8 @@ flowchart LR
609	> Redis Cluster 兼容(2026-08-29):两个前缀整体为 hash tag(`{xxx}`),模块全部键	609	> Redis Cluster 兼容(2026-08-29):两个前缀整体为 hash tag(`{xxx}`),模块全部键
610	> 落同一 slot——多键 Lua 的原子语义在 cluster 分片下保持成立;选主抽签键为	610	> 落同一 slot——多键 Lua 的原子语义在 cluster 分片下保持成立;选主抽签键为
611	> `{agent_runtime:job:<job>}:winner/candidates:{epoch}`(同槽),执行锁键	611	> `{agent_runtime:job:<job>}:winner/candidates:{epoch}`(同槽),执行锁键
612	-> `agent_runtime:job:<job>` 不变(单键操作)。连接串用 `redis+cluster://` scheme 构造	612	+> `agent_runtime:job:<job>` 不变(单键操作)。自评估域前缀 `{agent_runtime:eval}`(§5.3,
		613	+> 2026-09)同款单槽 tag、全单键命令无 Lua。连接串用 `redis+cluster://` scheme 构造
613	> 集群客户端(cluster 只有 db 0)。`state.eval` 把 prefix 同时声明为 `KEYS[1]`(路由锚,	614	> 集群客户端(cluster 只有 db 0)。`state.eval` 把 prefix 同时声明为 `KEYS[1]`(路由锚,
614	> 防 `numkeys=0` 随机路由到非归属节点)。单实例/哨兵下 `{}` 无语义,同一套键名兼容两种	615	> 防 `numkeys=0` 随机路由到非归属节点)。单实例/哨兵下 `{}` 无语义,同一套键名兼容两种
615	> 部署;背景与验证见 `docs/feature/2026-08-redis-cluster.md`。	616	> 部署;背景与验证见 `docs/feature/2026-08-redis-cluster.md`。
	@@ -726,6 +727,20 @@ flowchart TB
726	\| `lock:rm:reconcile` \| STRING(NX EX) \| 选主标记 \| 孤儿对账 sweeper 选主 \|	727	\| `lock:rm:reconcile` \| STRING(NX EX) \| 选主标记 \| 孤儿对账 sweeper 选主 \|
727	\| `lock:rm:watch` \| STRING(NX EX) \| 选主标记 \| K8s Watch + 死 Pod 轮询 + 健康 SSE 探测(场景 N)选主 \|	728	\| `lock:rm:watch` \| STRING(NX EX) \| 选主标记 \| K8s Watch + 死 Pod 轮询 + 健康 SSE 探测(场景 N)选主 \|
728		729
		730	+### 5.3 系统自评估(prefix `{agent_runtime:eval}:`,2026-09)
		731	+
		732	+评估数据域(非编排态;spec 见 `docs/spec/evaluation.md`):route/acquire 热路径计数经每副本
		733	+内存缓冲 5s 批量聚合,RM 后台状态变迁直写;sys_sample(30s)采样池态+计数快照落 ZSET,
		734	+sys_eval(300s)跑确定性规则(+可选 LLM 分析)产报告。全单键命令、零 Lua;报告只读
		735	+产出,不改任何配置(建议人审后经 Manager config_sync 应用)。
		736	+
		737	+\| 键 \| 类型 \| TTL \| 语义 \|
		738	+\|---\|---\|---\|---\|
		739	+\| `sample:scope:{sid}` \| ZSET(member=紧凑 JSON,score=秒) \| 25h(每采样刷新) \| per-scope 趋势采样(t/p/i/d/s/w + 计数快照差分源) \|
		740	+\| `ct:scope:{sid}` \| HASH \| 25h(每次写刷新) \| per-scope 全副本聚合计数(route_/acq_/ev_*) \|
		741	+\| `report:latest` \| STRING \| 无 TTL \| 最近评估报告(findings/trend/caveats) \|
		742	+\| `report:history` \| ZSET(瘦身条目) \| 30d(保 200) \| 报告历史(只留 summary) \|
		743	+
729	---	744	---
730		745
731	## 6. 场景清单与详细时序	746	## 6. 场景清单与详细时序
	@@ -1311,6 +1326,9 @@ M6(server 模式)已完成开发与真环境端到端验收,实现与本文的
1311		1326
1312	- `LUA_WAITER_GATE`(§5.1 键表 / §6.2 场景 F):实现期补充的第 7 个 SM Lua 脚本。初稿「先 SCARD 再 SADD」的入队判定在并发同时到达时全部读到旧计数而超收(真环境验收场景 F 发现的竞态),改为原子闸门后稳态 `SCARD ≤ max_waiters` 恒成立。全文见 SM 设计 §5.1。(2026-09:已随场景 F 快失败改造整体拆除,连同 waiters 键/free 通道,见 §9.2)	1327	- `LUA_WAITER_GATE`(§5.1 键表 / §6.2 场景 F):实现期补充的第 7 个 SM Lua 脚本。初稿「先 SCARD 再 SADD」的入队判定在并发同时到达时全部读到旧计数而超收(真环境验收场景 F 发现的竞态),改为原子闸门后稳态 `SCARD ≤ max_waiters` 恒成立。全文见 SM 设计 §5.1。(2026-09:已随场景 F 快失败改造整体拆除,连同 waiters 键/free 通道,见 §9.2)
1313	- 场景 N(半死 Pod 健康探测):机制已实现且有单测覆盖;端到端验收暂缓——当前 AgentServer 镜像在 SSE 端口对 `GET /health` 返回 426(要求协议升级),不满足 §6.2 场景 N 的固定约定,待 AgentServer 原生支持后补验。	1328	- 场景 N(半死 Pod 健康探测):机制已实现且有单测覆盖;端到端验收暂缓——当前 AgentServer 镜像在 SSE 端口对 `GET /health` 返回 426(要求协议升级),不满足 §6.2 场景 N 的固定约定,待 AgentServer 原生支持后补验。
		1329	+- 系统自评估(§5.3,2026-09):evaluation 子包 + sys_sample/sys_eval 两选主 job +
		1330	+ `/visualization/{history,evaluation}` 端点已实现,单测 51 例 + 集成用例全绿;真环境
		1331	+ 冒烟阶段 13c 已固化(e2e_hld_acceptance.py),门禁结果见 feature 篇目。
1314	- 场景 A–L 已在真 Redis + MySQL + K8s 环境端到端验收通过(用例固化为 `applications/agent_runtime/scripts/e2e_hld_acceptance.py`,经 `scripts/integration_smoke.sh` 调用,可作部署后回归)。	1332	- 场景 A–L 已在真 Redis + MySQL + K8s 环境端到端验收通过(用例固化为 `applications/agent_runtime/scripts/e2e_hld_acceptance.py`,经 `scripts/integration_smoke.sh` 调用,可作部署后回归)。
1315		1333
1316	### 9.1 多副本验收(2026-08-18 更新,M7)	1334	### 9.1 多副本验收(2026-08-18 更新,M7)

Aapplications/agent_runtime/docs/feature/2026-09-system-self-evaluation.md+141-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/docs/feature/README.md+1-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/docs/spec/README.md+5-2

	@@ -12,7 +12,9 @@
12	\| [service-core.md](service-core.md) \| 组装(main)/CLI/配置(`AGENT_RUNTIME_*`)/错误码契约/字段分类/工具/部署 \| 改装配、配置、错误契约、部署时 \|	12	\| [service-core.md](service-core.md) \| 组装(main)/CLI/配置(`AGENT_RUNTIME_*`)/错误码契约/字段分类/工具/部署 \| 改装配、配置、错误契约、部署时 \|
13	\| [session-manager.md](session-manager.md) \| SM:route/touch/config_sync/config_refresh/cleanup 编排、6 个 Lua、SM 键表 \| 改会话编排/配置层时 \|	13	\| [session-manager.md](session-manager.md) \| SM:route/touch/config_sync/config_refresh/cleanup 编排、6 个 Lua、SM 键表 \| 改会话编排/配置层时 \|
14	\| [resource-manager.md](resource-manager.md) \| RM:acquire/后台任务/K8s 适配、6 个 Lua、RM 键表 \| 改 Pod 池/扩缩容/清理时 \|	14	\| [resource-manager.md](resource-manager.md) \| RM:acquire/后台任务/K8s 适配、6 个 Lua、RM 键表 \| 改 Pod 池/扩缩容/清理时 \|
		15	+\| [evaluation.md](evaluation.md) \| 自评估:采样/评估两 job、`{agent_runtime:eval}` 键表、规则清单、LLM 降级矩阵 \| 改自评估/趋势/报告时 \|
15	\| [e2e-test-cases.md](e2e-test-cases.md) \| 全部 e2e 用例的场景/输入/预期输出 \| 写或跑 e2e 时 \|	16	\| [e2e-test-cases.md](e2e-test-cases.md) \| 全部 e2e 用例的场景/输入/预期输出 \| 写或跑 e2e 时 \|
		17	+\| `../api/config-plane-api.md` \| 配置面对外接口文档(config_sync/config_refresh/visualization:字段表+curl+真实返回示例) \| 给调用方(Claw Manager/运维/可视化前端)交付接口契约时 \|
16	\| `../design/Agent-Runtime-HLD.md` \| 架构总览/接口契约/场景 A–N/Redis 键表(语义权威) \| 语义不确定时 \|	18	\| `../design/Agent-Runtime-HLD.md` \| 架构总览/接口契约/场景 A–N/Redis 键表(语义权威) \| 语义不确定时 \|
17	\| `../design/session-manager-design.md` \| SM 详细设计(6 个 Lua 全文) \| 深挖 SM 设计动机 \|	19	\| `../design/session-manager-design.md` \| SM 详细设计(6 个 Lua 全文) \| 深挖 SM 设计动机 \|
18	\| `../design/resource-manager-design.md` \| RM 详细设计(6 个 Lua 全文) \| 深挖 RM 设计动机 \|	20	\| `../design/resource-manager-design.md` \| RM 详细设计(6 个 Lua 全文) \| 深挖 RM 设计动机 \|
	@@ -40,7 +42,8 @@ claw mgr ──config_sync──► ├─ session_manager 持 App,5 个 HTT
40	```	42	```
41	{session_manager}:… SM 编排态(会话四处/scope 闸门/候选集/注册表/路由快照)	43	{session_manager}:… SM 编排态(会话四处/scope 闸门/候选集/注册表/路由快照)
42	{resource_manager}:… RM 编排态(per-scope Pod 池/idle 暖池/deploy 占位/follower 等待室/选主锁)	44	{resource_manager}:… RM 编排态(per-scope Pod 池/idle 暖池/deploy 占位/follower 等待室/选主锁)
43	-agent_runtime:job:… 后台任务选主执行锁(main.py:_build_jobs 注册)	45	+{agent_runtime:eval}:… 自评估域(per-scope 计数 HASH/趋势采样 ZSET/评估报告;2026-09,见 evaluation.md)
		46	+agent_runtime:job:… 后台任务选主执行锁(main.py:_build_jobs 注册,7 个)
44	{agent_runtime:job:…}:winner/candidates:… 选主抽签键(与执行锁同底,hash tag 同槽)	47	{agent_runtime:job:…}:winner/candidates:… 选主抽签键(与执行锁同底,hash tag 同槽)
45	```	48	```
46		49
	@@ -51,7 +54,7 @@ agent_runtime:job:… 后台任务选主执行锁(main.py:_build_jobs 注册)
51		54
52	```bash	55	```bash
53	cd applications/agent_runtime	56	cd applications/agent_runtime
54	-uv sync --extra local && uv run pytest # 157 用例(fakeredis+SQLite+FakeK8s)	57	+uv sync --extra local && uv run pytest # 495 用例(fakeredis+SQLite+FakeK8s)
55	./scripts/integration_smoke.sh # 真环境冒烟(场景 A–L;FLUSHDB 目标库,有防误刷)	58	./scripts/integration_smoke.sh # 真环境冒烟(场景 A–L;FLUSHDB 目标库,有防误刷)
56	./scripts/deploy_replicas.sh 2 .env.production.local 8091 # 宿主机双进程	59	./scripts/deploy_replicas.sh 2 .env.production.local 8091 # 宿主机双进程
57	./deploy/render_and_apply.sh deploy/agent_runtime.env --nodeport # K8s 生产形态	60	./deploy/render_and_apply.sh deploy/agent_runtime.env --nodeport # K8s 生产形态

Aapplications/agent_runtime/docs/spec/evaluation.md+133-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/docs/spec/resource-manager.md+2-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/docs/spec/service-core.md+22-7文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/docs/spec/session-manager.md+6-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/scripts/e2e_hld_acceptance.py+71-0

	@@ -1430,6 +1430,76 @@ async def stage13_error_contract(c: Client, r) -> None:
1430	code == 200 and raw.get("cleaned") == 0, str(raw))	1430	code == 200 and raw.get("cleaned") == 0, str(raw))
1431		1431
1432		1432
		1433	+async def stage13c_evaluation(c: Client, r) -> None:
		1434	+ """系统自评估数据层（2026-09）：两 job + 三个可视化端点（真环境）。
		1435	+
		1436	+ 牙齿与等待窗:sys_sample 默认 30s、sys_eval 默认 300s——间隔从 overview
		1437	+ 配置摘要现读,PASS 判据 = eval_interval+60s（上限 330s,冒烟部署建议
		1438	+ AGENT_RUNTIME_EVAL_INTERVAL=15 加速）内轮询到 latest 报告;llm.status
		1439	+ ∈ {disabled, ok}（真环境可能配了 LLM）;响应不得泄漏 API_KEY。只调小
		1440	+ interval env、不回拨任何指针（e2e 红线）。
		1441	+ """
		1442	+ print("\n== 阶段 13c：自评估 —— sys_sample/sys_eval + 可视化端点 ==")
		1443	+ root = c.base.rsplit("/api/session", 1)[0]
		1444	+
		1445	+ async def vis(path: str) -> tuple[int, dict]:
		1446	+ resp = await c.http.get(f"{root}{path}", timeout=30.0)
		1447	+ try:
		1448	+ body = resp.json()
		1449	+ except Exception:
		1450	+ body = {}
		1451	+ return resp.status_code, body
		1452	+
		1453	+ # 1) 两 job 已注册且至少跑过一拍（overview 是全局看板,任意副本应答）
		1454	+ code, overview = await vis("/visualization/overview")
		1455	+ jobs = {j.get("name"): j for j in overview.get("jobs", [])
		1456	+ if isinstance(j, dict)}
		1457	+ check("13c-overview 含 sys_sample/sys_eval 且 ok_ticks≥1",
		1458	+ code == 200 and {"sys_sample", "sys_eval"} <= set(jobs)
		1459	+ and jobs.get("sys_sample", {}).get("ok_ticks", 0) >= 1,
		1460	+ f"sample_ok={jobs.get('sys_sample', {}).get('ok_ticks')}")
		1461	+
		1462	+ # 2) scope 摘要含 SM 容量字段（scope 重构后观测补齐）
		1463	+ code, scopes = await vis("/visualization/scopes")
		1464	+ row = next((s for s in scopes.get("scopes", []) if s.get("scope_id") == MAIN),
		1465	+ None)
		1466	+ check("13c-scopes 摘要含容量闸门字段（sc/session_count）",
		1467	+ code == 200 and row is not None
		1468	+ and row.get("scope_concurrency") == 3 and row.get("max_pods") == 2
		1469	+ and "session_count" in row, str(row)[:100])
		1470	+
		1471	+ # 3) 历史趋势:采样 30s cadence,45s 内必有数据点
		1472	+ async def has_points() -> bool:
		1473	+ _, hist = await vis(f"/visualization/history?scope_id={MAIN}")
		1474	+ return bool(hist.get("points"))
		1475	+ ok = await wait_until(has_points, timeout=45, interval=3)
		1476	+ _, hist = await vis(f"/visualization/history?scope_id={MAIN}")
		1477	+ point = (hist.get("points") or [{}])[0]
		1478	+ check("13c-history 采样落盘（窗口内 points≥1 且字段齐全）",
		1479	+ ok and {"t", "p", "i", "s"} <= set(point), str(point)[:80])
		1480	+
		1481	+ # 4) 评估报告:间隔现读,限窗轮询(不依赖部署 env 假设)
		1482	+ eval_interval = int(overview.get("config", {}).get("eval_interval") or 300)
		1483	+ budget = min(eval_interval + 60, 330)
		1484	+
		1485	+ async def has_report() -> bool:
		1486	+ _, ev = await vis("/visualization/evaluation")
		1487	+ return bool(ev.get("latest"))
		1488	+ ok = await wait_until(has_report, timeout=budget, interval=5)
		1489	+ _, ev = await vis("/visualization/evaluation")
		1490	+ latest = ev.get("latest") or {}
		1491	+ llm_status = (latest.get("llm") or {}).get("status")
		1492	+ check("13c-evaluation 报告产出（llm ∈ {disabled,ok};findings 为 list）",
		1493	+ ok and llm_status in ("disabled", "ok")
		1494	+ and isinstance(latest.get("findings"), list)
		1495	+ and (latest.get("summary") or {}).get("active", 0) >= 1,
		1496	+ f"interval={eval_interval}s llm={llm_status} "
		1497	+ f"findings={len(latest.get('findings') or [])}")
		1498	+ text = str(latest).lower()
		1499	+ check("13c-evaluation 报告无凭证泄漏（无 api_key/Bearer 字样）",
		1500	+ ok and "api_key" not in text and "bearer" not in text, "")
		1501	+
		1502	+
1433	# ---------------------------------------------------------------- 入口	1503	# ---------------------------------------------------------------- 入口
1434		1504
1435	def _parse_args() -> argparse.Namespace:	1505	def _parse_args() -> argparse.Namespace:
	@@ -1540,6 +1610,7 @@ async def main() -> None:
1540	await stage12_reconcile_cleanup(c, r)	1610	await stage12_reconcile_cleanup(c, r)
1541	await stage12b_or_branch(c, r)	1611	await stage12b_or_branch(c, r)
1542	await stage13_error_contract(c, r)	1612	await stage13_error_contract(c, r)
		1613	+ await stage13c_evaluation(c, r)
1543	finally:	1614	finally:
1544	await r.aclose()	1615	await r.aclose()
1545		1616

Mapplications/agent_runtime/scripts/verify_redis_cluster.py+25-2文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/src/agent_runtime/config.py+44-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/__init__.py+5-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/collector.py+253-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/evaluator.py+258-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/llm.py+245-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/rules.py+432-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/src/agent_runtime/evaluation/state.py+162-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/src/agent_runtime/main.py+93-3

	@@ -17,6 +17,7 @@ App 的 lifespan 只认一个 ctx_factory——返回 sm_sysctx，其余生命
17		17
18	from __future__ import annotations	18	from __future__ import annotations
19		19
		20	+import asyncio
20	import logging	21	import logging
21	from typing import Any	22	from typing import Any
22		23
	@@ -29,6 +30,15 @@ from . import errors as app_errors
29	from .config import RM_KEY_PREFIX, SERVICE_PREFIX, SM_KEY_PREFIX, AgentRuntimeConfig	30	from .config import RM_KEY_PREFIX, SERVICE_PREFIX, SM_KEY_PREFIX, AgentRuntimeConfig
30	from .visualization_api import register_visualization_api	31	from .visualization_api import register_visualization_api
31	from .metrics import MetricsRegistry, request_metrics_middleware	32	from .metrics import MetricsRegistry, request_metrics_middleware
		33	+from .evaluation.collector import (
		34	+ FLUSH_INTERVAL_SEC,
		35	+ FLUSH_TIMEOUT_SEC,
		36	+ EvaluationCollector,
		37	+ ScopeTelemetryBuffer,
		38	+)
		39	+from .evaluation.evaluator import Evaluator
		40	+from .evaluation.llm import LLMClient
		41	+from .evaluation.state import EvaluationState
32	from .resource_manager.facade import ResourceManagerFacade	42	from .resource_manager.facade import ResourceManagerFacade
33	from .resource_manager.k8s import FakeK8sPodClient, RealK8sPodClient	43	from .resource_manager.k8s import FakeK8sPodClient, RealK8sPodClient
34	from .resource_manager.orchestrator import ResourceOrchestrator	44	from .resource_manager.orchestrator import ResourceOrchestrator
	@@ -64,6 +74,11 @@ TICK_TIMEOUTS = {
64	"rm_reclaim": 60,	74	"rm_reclaim": 60,
65	"rm_watch": 300,	75	"rm_watch": 300,
66	"rm_reconcile": 300,	76	"rm_reconcile": 300,
		77	+ # sys_sample：逐 scope 单键读写,快操作
		78	+ "sys_sample": 30,
		79	+ # sys_eval：LLM timeout(默认 60s,eval_llm_timeout 可调但须小于此值)
		80	+ # + 规则计算 + 采样窗口读,120 留余量
		81	+ "sys_eval": 120,
67	}	82	}
68		83
69		84
	@@ -154,9 +169,13 @@ class OrchestratorSystemContext(SystemContext):
154	sm_state = SessionState(self.redis)	169	sm_state = SessionState(self.redis)
155	rm_state = ResourceState(self.redis)	170	rm_state = ResourceState(self.redis)
156		171
		172	+ # 评估域:计数缓冲(热路径内存)+ 采集器 + 评估器(键前缀独立 hash tag)
		173	+ eval_state = EvaluationState(self.redis)
		174	+ telemetry = ScopeTelemetryBuffer()
		175	+
157	self.sm_facade = SessionManagerFacade(sm_state)	176	self.sm_facade = SessionManagerFacade(sm_state)
158		177
159	- rm_orchestrator = ResourceOrchestrator(rm_state, self.k8s)	178	+ rm_orchestrator = ResourceOrchestrator(rm_state, self.k8s, telemetry=telemetry)
160	self.rm_facade = ResourceManagerFacade(rm_orchestrator)	179	self.rm_facade = ResourceManagerFacade(rm_orchestrator)
161		180
162	self.sm_config_store = ConfigStore(	181	self.sm_config_store = ConfigStore(
	@@ -170,11 +189,23 @@ class OrchestratorSystemContext(SystemContext):
170	self.sm_config_store,	189	self.sm_config_store,
171	self.rm_facade,	190	self.rm_facade,
172	default_session_ttl=self.arc.default_session_ttl,	191	default_session_ttl=self.arc.default_session_ttl,
		192	+ telemetry=telemetry,
173	)	193	)
174	self.sm_sweeper = SessionSweeper(sm_state, self.rm_facade)	194	self.sm_sweeper = SessionSweeper(sm_state, self.rm_facade)
175	self.rm_sweeper = ResourceSweeper(	195	self.rm_sweeper = ResourceSweeper(
176	rm_state, self.k8s, self.sm_facade, orchestrator=rm_orchestrator,	196	rm_state, self.k8s, self.sm_facade, orchestrator=rm_orchestrator,
		197	+ event_sink=eval_state.bump_event,
177	)	198	)
		199	+ self.eval_state = eval_state
		200	+ self.eval_telemetry = telemetry
		201	+ self.eval_collector = EvaluationCollector(
		202	+ eval_state=eval_state, sm_state=sm_state, rm_state=rm_state,
		203	+ )
		204	+ self.evaluator = Evaluator(
		205	+ collector=self.eval_collector, llm=LLMClient.from_arc(self.arc),
		206	+ state=eval_state, arc=self.arc, instance_id=self.instance_id,
		207	+ )
		208	+ self._telemetry_task: asyncio.Task \| None = None
178		209
179	def _build_jobs(self) -> list[Any]:	210	def _build_jobs(self) -> list[Any]:
180	"""全部后台任务（tick 级选主锁；多副本全局单副本执行写操作）。"""	211	"""全部后台任务（tick 级选主锁；多副本全局单副本执行写操作）。"""
	@@ -209,6 +240,19 @@ class OrchestratorSystemContext(SystemContext):
209	lock_key="agent_runtime:job:rm_reconcile",	240	lock_key="agent_runtime:job:rm_reconcile",
210	tick_timeout_sec=TICK_TIMEOUTS["rm_reconcile"],	241	tick_timeout_sec=TICK_TIMEOUTS["rm_reconcile"],
211	),	242	),
		243	+ # 系统自评估(sys_eval 全局单副本产报告,任意副本可读):
		244	+ self.create_single_leader_job(
		245	+ name="sys_sample", on_tick=self.eval_collector.sample_once,
		246	+ interval_sec=max(arc.eval_sample_interval, 5),
		247	+ lock_key="agent_runtime:job:sys_sample",
		248	+ tick_timeout_sec=TICK_TIMEOUTS["sys_sample"],
		249	+ ),
		250	+ self.create_single_leader_job(
		251	+ name="sys_eval", on_tick=self.evaluator.evaluate_once,
		252	+ interval_sec=max(arc.eval_interval, 30),
		253	+ lock_key="agent_runtime:job:sys_eval",
		254	+ tick_timeout_sec=TICK_TIMEOUTS["sys_eval"],
		255	+ ),
212	]	256	]
213	return jobs	257	return jobs
214		258
	@@ -236,15 +280,24 @@ class OrchestratorSystemContext(SystemContext):
236	job.name, interval_by_name.get(job.name, "?"),	280	job.name, interval_by_name.get(job.name, "?"),
237	TICK_TIMEOUTS.get(job.name),	281	TICK_TIMEOUTS.get(job.name),
238	)	282	)
		283	+ # 计数缓冲 flusher:每副本独立(选主 job 会漏非 leader 副本的缓冲),
		284	+ # drain → 同槽 HINCRBY 批量聚合;失败留到下轮(计数只延后不丢)
		285	+ self._telemetry_task = asyncio.create_task(
		286	+ self._telemetry_flush_loop(), name="eval-telemetry-flush"
		287	+ )
239	self.logger.info(	288	self.logger.info(
240	"config summary: mode=%s namespace=%s sweep=%ss autoscale=%ss "	289	"config summary: mode=%s namespace=%s sweep=%ss autoscale=%ss "
241	"reclaim=%ss watch=%ss reconcile=%ss "	290	"reclaim=%ss watch=%ss reconcile=%ss "
242	- "default_session_ttl=%ss kubeconfig=%s",	291	+ "default_session_ttl=%ss eval_sample=%ss eval=%ss eval_llm=%s "
		292	+ "kubeconfig=%s",
243	self.arc.mode, self.arc.default_namespace,	293	self.arc.mode, self.arc.default_namespace,
244	self.arc.sweep_interval, self.arc.autoscale_interval,	294	self.arc.sweep_interval, self.arc.autoscale_interval,
245	self.arc.reclaim_interval, self.arc.watch_interval,	295	self.arc.reclaim_interval, self.arc.watch_interval,
246	self.arc.reconcile_interval,	296	self.arc.reconcile_interval,
247	self.arc.default_session_ttl,	297	self.arc.default_session_ttl,
		298	+ self.arc.eval_sample_interval, self.arc.eval_interval,
		299	+ "enabled" if (self.arc.eval_llm_base_url and self.arc.eval_llm_model)
		300	+ else "disabled",
248	"set" if self.arc.kubeconfig else "in-cluster",	301	"set" if self.arc.kubeconfig else "in-cluster",
249	)	302	)
250	self.logger.info(	303	self.logger.info(
	@@ -261,10 +314,31 @@ class OrchestratorSystemContext(SystemContext):
261	"rm_reclaim": self.arc.reclaim_interval,	314	"rm_reclaim": self.arc.reclaim_interval,
262	"rm_watch": self.arc.watch_interval,	315	"rm_watch": self.arc.watch_interval,
263	"rm_reconcile": self.arc.reconcile_interval,	316	"rm_reconcile": self.arc.reconcile_interval,
		317	+ "sys_sample": self.arc.eval_sample_interval,
		318	+ "sys_eval": self.arc.eval_interval,
264	}	319	}
265		320
		321	+ async def _telemetry_flush_loop(self) -> None:
		322	+ """每副本一个:route/acquire 计数缓冲 → Redis 聚合(5s 批量)。"""
		323	+ while True:
		324	+ await asyncio.sleep(FLUSH_INTERVAL_SEC)
		325	+ try:
		326	+ await asyncio.wait_for(
		327	+ self._flush_telemetry_once(), timeout=FLUSH_TIMEOUT_SEC
		328	+ )
		329	+ except asyncio.CancelledError:
		330	+ raise
		331	+ except Exception: # noqa: BLE001 - 留到下轮(计数只延后不丢)
		332	+ self.logger.exception("telemetry flush failed, retry next round")
		333	+
		334	+ async def _flush_telemetry_once(self) -> None:
		335	+ drained = self.eval_telemetry.drain()
		336	+ for scope_id, deltas in drained.items():
		337	+ if deltas:
		338	+ await self.eval_state.bump_counters(scope_id, deltas)
		339	+
266	async def jobs_snapshot(self) -> list[dict[str, Any]]:	340	async def jobs_snapshot(self) -> list[dict[str, Any]]:
267	- """诊断用：5 个后台任务的间隔/超时/计数器/当前 leader（/visualization/overview）。"""	341	+ """诊断用：7 个后台任务的间隔/超时/计数器/当前 leader（/visualization/overview）。"""
268	intervals = self._job_intervals()	342	intervals = self._job_intervals()
269	out: list[dict[str, Any]] = []	343	out: list[dict[str, Any]] = []
270	for job in self._jobs:	344	for job in self._jobs:
	@@ -298,6 +372,22 @@ class OrchestratorSystemContext(SystemContext):
298	except Exception: # noqa: BLE001	372	except Exception: # noqa: BLE001
299	self.logger.exception("job stop failed: %s", job)	373	self.logger.exception("job stop failed: %s", job)
300	self._jobs = []	374	self._jobs = []
		375	+ if self._telemetry_task is not None:
		376	+ self._telemetry_task.cancel()
		377	+ try:
		378	+ await self._telemetry_task
		379	+ except asyncio.CancelledError:
		380	+ pass
		381	+ except Exception: # noqa: BLE001
		382	+ self.logger.exception("telemetry flush task failed on cancel")
		383	+ self._telemetry_task = None
		384	+ # 终结 drain:停机前把缓冲余量尽量落 Redis(失败只留痕,计数延后无害)
		385	+ try:
		386	+ await asyncio.wait_for(
		387	+ self._flush_telemetry_once(), timeout=FLUSH_TIMEOUT_SEC
		388	+ )
		389	+ except Exception: # noqa: BLE001
		390	+ self.logger.exception("final telemetry drain failed")
301	try:	391	try:
302	await self.k8s.close()	392	await self.k8s.close()
303	except Exception: # noqa: BLE001	393	except Exception: # noqa: BLE001

Mapplications/agent_runtime/src/agent_runtime/resource_manager/orchestrator.py+14-1文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/src/agent_runtime/resource_manager/sweeper.py+18-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/src/agent_runtime/session_manager/orchestrator.py+84-68文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/src/agent_runtime/visualization_api.py+149-21文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/__init__.py+2-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/test_collector.py+149-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/test_eval_state.py+83-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/test_evaluator.py+143-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/test_llm.py+186-0文件内容审核中，请稍后刷新重试

Aapplications/agent_runtime/tests/evaluation/test_rules.py+241-0文件内容审核中，请稍后刷新重试

Mapplications/agent_runtime/tests/integration/test_visualization_api.py+109-4文件内容审核中，请稍后刷新重试