@octopi-agent/engine 0.90.11 → 0.97.15

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (169) hide show
  1. package/dist/config-admin/agent-presets.d.ts +70 -0
  2. package/dist/config-admin/agent-presets.js +243 -0
  3. package/dist/config-admin/discover.d.ts +30 -0
  4. package/dist/config-admin/discover.js +48 -0
  5. package/dist/config-admin/env.d.ts +22 -0
  6. package/dist/config-admin/env.js +41 -0
  7. package/dist/config-admin/index.d.ts +20 -0
  8. package/dist/config-admin/index.js +15 -0
  9. package/dist/config-admin/model-presets.d.ts +75 -0
  10. package/dist/config-admin/model-presets.js +249 -0
  11. package/dist/config-admin/name-guard.d.ts +14 -0
  12. package/dist/config-admin/name-guard.js +24 -0
  13. package/dist/config-admin/patch.d.ts +56 -0
  14. package/dist/config-admin/patch.js +201 -0
  15. package/dist/config-admin/port.d.ts +68 -0
  16. package/dist/config-admin/port.js +21 -0
  17. package/dist/config-admin/redact.d.ts +39 -0
  18. package/dist/config-admin/redact.js +42 -0
  19. package/dist/config-schema/agent.d.ts +43 -1
  20. package/dist/config-schema/context.d.ts +6 -1
  21. package/dist/config-schema/context.js +8 -0
  22. package/dist/config-schema/engine.d.ts +106 -4
  23. package/dist/config-schema/governance.d.ts +43 -1
  24. package/dist/config-schema/governance.js +157 -0
  25. package/dist/config-schema/models.d.ts +5 -2
  26. package/dist/config-schema/models.js +4 -1
  27. package/dist/config-schema/observability.d.ts +9 -0
  28. package/dist/config-schema/observability.js +29 -0
  29. package/dist/config-schema/run.d.ts +3 -0
  30. package/dist/config-schema/run.js +14 -0
  31. package/dist/config.d.ts +29 -4
  32. package/dist/config.js +15 -4
  33. package/dist/harness/activation/dispatcher.d.ts +8 -2
  34. package/dist/harness/activation/dispatcher.js +33 -5
  35. package/dist/harness/activation/inbox.d.ts +87 -0
  36. package/dist/harness/activation/inbox.js +169 -0
  37. package/dist/harness/activation/runtime.d.ts +9 -8
  38. package/dist/harness/activation/runtime.js +100 -8
  39. package/dist/harness/activation/types.d.ts +19 -0
  40. package/dist/harness/activation/types.js +2 -0
  41. package/dist/harness/agent/builder.d.ts +15 -1
  42. package/dist/harness/agent/builder.js +31 -10
  43. package/dist/harness/agent/config-bridge.d.ts +5 -0
  44. package/dist/harness/agent/config-bridge.js +3 -0
  45. package/dist/harness/agent/config-integrity.d.ts +72 -0
  46. package/dist/harness/agent/config-integrity.js +146 -0
  47. package/dist/harness/agent/wire-agent-builder.js +2 -1
  48. package/dist/harness/collaboration/autonomous-subsystem/think/executor.js +3 -1
  49. package/dist/harness/collaboration/delegation/child-run.js +21 -5
  50. package/dist/harness/collaboration/delegation/notifier.js +4 -3
  51. package/dist/harness/collaboration/delegation/service.js +2 -3
  52. package/dist/harness/collaboration/delegation/wiring.js +6 -5
  53. package/dist/harness/collaboration/orchestration/reflector/llm-reflector.d.ts +1 -1
  54. package/dist/harness/collaboration/orchestration/reflector/llm-reflector.js +1 -1
  55. package/dist/harness/context/capabilities/compact/engine.d.ts +30 -0
  56. package/dist/harness/context/capabilities/compact/engine.js +198 -46
  57. package/dist/harness/context/capabilities/compact/handover.d.ts +78 -0
  58. package/dist/harness/context/capabilities/compact/handover.js +107 -0
  59. package/dist/harness/context/capabilities/compact/index.d.ts +3 -1
  60. package/dist/harness/context/capabilities/compact/index.js +2 -1
  61. package/dist/harness/context/capabilities/compact/types.d.ts +30 -1
  62. package/dist/harness/context/capabilities/summary/executor.js +2 -1
  63. package/dist/harness/context/capabilities/summary/policies.js +8 -5
  64. package/dist/harness/context/constitution/default-agents.md +11 -10
  65. package/dist/harness/context/default-context-engine.d.ts +34 -4
  66. package/dist/harness/context/default-context-engine.js +194 -49
  67. package/dist/harness/context/hybrid-compressor.js +6 -0
  68. package/dist/harness/context/llm-summarizer.d.ts +2 -8
  69. package/dist/harness/context/llm-summarizer.js +69 -91
  70. package/dist/harness/context/message-selector.js +4 -3
  71. package/dist/harness/context/summarize.d.ts +2 -0
  72. package/dist/harness/context/summarize.js +7 -2
  73. package/dist/harness/context/truncate-compressor.js +7 -0
  74. package/dist/harness/context/turn-envelope.d.ts +3 -2
  75. package/dist/harness/context/turn-envelope.js +3 -2
  76. package/dist/harness/context/types.d.ts +33 -1
  77. package/dist/harness/extension/execution-environment/dangerous-commands.d.ts +28 -0
  78. package/dist/harness/extension/execution-environment/dangerous-commands.js +129 -0
  79. package/dist/harness/extension/execution-environment/effect-policy.d.ts +127 -8
  80. package/dist/harness/extension/execution-environment/effect-policy.js +244 -19
  81. package/dist/harness/extension/execution-environment/index.d.ts +4 -2
  82. package/dist/harness/extension/execution-environment/index.js +2 -1
  83. package/dist/harness/extension/execution-environment/kill-switch.d.ts +34 -0
  84. package/dist/harness/extension/execution-environment/kill-switch.js +46 -0
  85. package/dist/harness/extension/execution-environment/ops-audit.d.ts +200 -0
  86. package/dist/harness/extension/execution-environment/ops-audit.js +338 -0
  87. package/dist/harness/extension/execution-environment/privileged-exec.d.ts +158 -0
  88. package/dist/harness/extension/execution-environment/privileged-exec.js +418 -0
  89. package/dist/harness/extension/execution-environment/process-info.d.ts +180 -0
  90. package/dist/harness/extension/execution-environment/process-info.js +466 -0
  91. package/dist/harness/extension/execution-environment/process-isolation.d.ts +7 -0
  92. package/dist/harness/extension/execution-environment/process-isolation.js +9 -0
  93. package/dist/harness/extension/execution-environment/sandbox.d.ts +109 -0
  94. package/dist/harness/extension/execution-environment/sandbox.js +228 -21
  95. package/dist/harness/extension/execution-environment/tool-capabilities.d.ts +1 -1
  96. package/dist/harness/extension/execution-environment/tool-capabilities.js +6 -0
  97. package/dist/harness/extension/execution-environment/types.d.ts +19 -0
  98. package/dist/harness/extension/plugin-ecosystem/channel-types.d.ts +5 -0
  99. package/dist/harness/extension/plugin-ecosystem/tools/allowlisted-run.js +33 -0
  100. package/dist/harness/extension/plugin-ecosystem/tools/ask-user.d.ts +4 -5
  101. package/dist/harness/extension/plugin-ecosystem/tools/ask-user.js +5 -15
  102. package/dist/harness/extension/plugin-ecosystem/tools/builtin.d.ts +3 -0
  103. package/dist/harness/extension/plugin-ecosystem/tools/builtin.js +53 -2
  104. package/dist/harness/extension/plugin-ecosystem/tools/index.d.ts +2 -0
  105. package/dist/harness/extension/plugin-ecosystem/tools/index.js +2 -0
  106. package/dist/harness/extension/plugin-ecosystem/tools/privileged-tool.d.ts +43 -0
  107. package/dist/harness/extension/plugin-ecosystem/tools/privileged-tool.js +350 -0
  108. package/dist/harness/extension/plugin-ecosystem/tools/process-tools.d.ts +30 -0
  109. package/dist/harness/extension/plugin-ecosystem/tools/process-tools.js +137 -0
  110. package/dist/harness/governance/human-in-the-loop/decision-cache.d.ts +54 -4
  111. package/dist/harness/governance/human-in-the-loop/decision-cache.js +42 -16
  112. package/dist/harness/governance/security/temp-path.d.ts +9 -0
  113. package/dist/harness/governance/security/temp-path.js +13 -2
  114. package/dist/harness/governance/session-acl/seed-roles.js +14 -0
  115. package/dist/harness/index.d.ts +2 -0
  116. package/dist/harness/index.js +2 -0
  117. package/dist/harness/knowledge/fts.js +1 -2
  118. package/dist/harness/knowledge/index-store.js +59 -9
  119. package/dist/harness/memory/sqlite/keyword-search.d.ts +1 -0
  120. package/dist/harness/memory/sqlite/keyword-search.js +4 -3
  121. package/dist/harness/observability/observer/hub.js +8 -0
  122. package/dist/harness/observability/observer/run-snapshot.js +1 -0
  123. package/dist/harness/observability/observer/types.d.ts +4 -0
  124. package/dist/harness/run/agent/agent.d.ts +7 -0
  125. package/dist/harness/run/agent/agent.js +23 -1
  126. package/dist/harness/run/reliability/fallback-provider.d.ts +117 -5
  127. package/dist/harness/run/reliability/fallback-provider.js +262 -22
  128. package/dist/harness/run/reliability/harness-events.d.ts +24 -1
  129. package/dist/harness/run/reliability/model-error-policy.d.ts +32 -0
  130. package/dist/harness/run/reliability/model-error-policy.js +42 -0
  131. package/dist/harness/run/reliability/provider-wrapper.js +29 -2
  132. package/dist/harness/run/reliability/refusal-guard.d.ts +20 -0
  133. package/dist/harness/run/reliability/refusal-guard.js +102 -0
  134. package/dist/harness/run/reliability/run-agent.js +231 -33
  135. package/dist/harness/run/run-guard/default-run-guard.js +1 -1
  136. package/dist/harness/run/run-scope.d.ts +31 -0
  137. package/dist/harness/run/runner.d.ts +116 -2
  138. package/dist/harness/run/runner.js +547 -41
  139. package/dist/harness/session/attachments/llm-intent.js +2 -1
  140. package/dist/harness/session/tasks/render.d.ts +4 -1
  141. package/dist/harness/session/tasks/render.js +10 -7
  142. package/dist/harness/shared/events/agent-event-map.d.ts +66 -0
  143. package/dist/harness/shared/events/agent-event-map.js +5 -0
  144. package/dist/index.d.ts +3 -3
  145. package/dist/index.js +2 -2
  146. package/dist/integration/agent-runtime/channel-message-source.js +19 -1
  147. package/dist/integration/index.d.ts +2 -0
  148. package/dist/integration/index.js +1 -0
  149. package/dist/integration/observability/loop-watchdog.d.ts +56 -0
  150. package/dist/integration/observability/loop-watchdog.js +177 -0
  151. package/dist/integration/observability/metrics.d.ts +3 -0
  152. package/dist/integration/observability/metrics.js +18 -7
  153. package/dist/integration/providers/anthropic.d.ts +23 -0
  154. package/dist/integration/providers/anthropic.js +97 -12
  155. package/dist/integration/providers/fetch-timeout.d.ts +42 -0
  156. package/dist/integration/providers/fetch-timeout.js +51 -0
  157. package/dist/integration/providers/openai.d.ts +10 -0
  158. package/dist/integration/providers/openai.js +58 -11
  159. package/dist/integration/providers/provider-diagnostics.d.ts +80 -0
  160. package/dist/integration/providers/provider-diagnostics.js +89 -0
  161. package/dist/integration/providers/timeout-coherence.d.ts +52 -0
  162. package/dist/integration/providers/timeout-coherence.js +68 -0
  163. package/dist/integration/providers/wire-dump.d.ts +62 -0
  164. package/dist/integration/providers/wire-dump.js +111 -2
  165. package/dist/integration/storage/jsonl.js +4 -1
  166. package/dist/subsystems/cognition-steward/conceptualize/handler.js +1 -1
  167. package/dist/subsystems/memory-steward/backfill/handler.js +1 -1
  168. package/dist/subsystems/wisdom-steward/formation/handler.js +1 -1
  169. package/package.json +3 -2
@@ -7,7 +7,7 @@
7
7
  * 设计:arch/agent-runtime.md
8
8
  */
9
9
  import { CoalesceBuffer } from './coalesce.js';
10
- import { buildRunRequest, resolveSessionId } from './compiler.js';
10
+ import { buildRunRequest, compileMessages, resolveSessionId } from './compiler.js';
11
11
  import { ExplicitRouter } from './router.js';
12
12
  import { RuntimeEvents } from './types.js';
13
13
  /**
@@ -156,8 +156,9 @@ export class AgentRuntime {
156
156
  return results;
157
157
  }
158
158
  /**
159
- * 模型 A:await 至 Run 结束(或 skipped / failed)
159
+ * 模型 A:await 至 Run 结束(或 skipped / failed / accepted)
160
160
  * @param options.onEvent - 转发 Run 事件流(Gateway 流式广播);不变量 #6
161
+ * @param options.onDelivery - 活跃 Run 接收即回执(早于 Run 结束)
161
162
  */
162
163
  async dispatch(trigger, options) {
163
164
  if (!this.started) {
@@ -183,8 +184,22 @@ export class AgentRuntime {
183
184
  });
184
185
  return { status: 'skipped', reason: 'no_agent' };
185
186
  }
187
+ // 回执恒发一次:fan-out 多 target 时首个去向生效,其余静默
188
+ //(每 target 一帧 accepted 会与 channel/web 的「一次受理」契约冲突)
189
+ let deliveryFired = false;
190
+ const deliveryOptions = options
191
+ ? {
192
+ ...options,
193
+ onDelivery: (delivery) => {
194
+ if (deliveryFired)
195
+ return;
196
+ deliveryFired = true;
197
+ options.onDelivery?.(delivery);
198
+ },
199
+ }
200
+ : undefined;
186
201
  if (targets.length === 1) {
187
- const result = await this.dispatchToTarget(trigger, targets[0], options);
202
+ const result = await this.dispatchToTarget(trigger, targets[0], deliveryOptions);
188
203
  // 单目标丢弃也必须留下 TRIGGER_DROPPED(与 fan-out 对称):
189
204
  // 「trigger 未被受理必有 dropped 事件」是委派送达回执(delegation §7.1)
190
205
  // 依赖的不变量,缺事件 = 送达回执无法证伪。受理后 run 失败走 RUN_FAILED
@@ -201,11 +216,11 @@ export class AgentRuntime {
201
216
  // fan-out:串行;聚合为 FanoutDispatchResult(ran/aborted/failed 均不静默丢弃)
202
217
  const results = [];
203
218
  for (const target of targets) {
204
- const result = await this.dispatchToTarget(trigger, target, options);
219
+ const result = await this.dispatchToTarget(trigger, target, deliveryOptions);
205
220
  results.push({ agentId: target.agentId, result });
206
221
  }
207
222
  const anyFailed = results.some((r) => r.result.status === 'failed');
208
- const anyRan = results.some((r) => r.result.status === 'ran');
223
+ const anyRan = results.some((r) => r.result.status === 'ran' || r.result.status === 'accepted');
209
224
  const allSkipped = results.every((r) => r.result.status === 'skipped');
210
225
  if (anyFailed) {
211
226
  this.emitRuntime(RuntimeEvents.TRIGGER_DROPPED, {
@@ -234,6 +249,40 @@ export class AgentRuntime {
234
249
  }
235
250
  return { status: 'skipped', reason: 'no_agent' };
236
251
  }
252
+ /**
253
+ * 路由目标的活跃 inbox 接收(human 与 delegation signal 共用)。
254
+ * 仅 message / agent_signal 可接续;其余类型与带附件/换模型输入走完整新 Run。
255
+ */
256
+ offerToTarget(trigger, target, sessionId, onEvent) {
257
+ if (!target.dispatcher.inbox)
258
+ return 'dispatched';
259
+ if (trigger.type !== 'message' && trigger.type !== 'agent_signal')
260
+ return 'dispatched';
261
+ const [compiledMessage] = compileMessages([trigger]);
262
+ const human = trigger.type === 'message' && trigger.payload.kind === 'user_message';
263
+ const delegationId = trigger.metadata?.delegationId;
264
+ return target.dispatcher.inbox.offer({
265
+ agentId: target.agentId,
266
+ sessionId,
267
+ entryId: delegationId ?? trigger.id,
268
+ triggerType: trigger.type,
269
+ human,
270
+ modelOverride: trigger.metadata?.modelOverride,
271
+ modelProvider: trigger.metadata?.modelProvider,
272
+ trigger,
273
+ onEvent,
274
+ message: compiledMessage,
275
+ onAdmitted: (delivery) => {
276
+ this.emitRuntime(RuntimeEvents.INBOX_ADMITTED, {
277
+ triggerId: trigger.id,
278
+ agentId: target.agentId,
279
+ sessionId,
280
+ delegationId,
281
+ delivery,
282
+ });
283
+ },
284
+ });
285
+ }
237
286
  async dispatchToTarget(trigger, target, options) {
238
287
  const agent = this.agents.get(target.agentId);
239
288
  if (!agent) {
@@ -243,6 +292,14 @@ export class AgentRuntime {
243
292
  ...trigger,
244
293
  sessionId: target.sessionId ?? trigger.sessionId,
245
294
  });
295
+ // 单一 offer 点:活跃 Run 接收 = 不再开新 Run(原 Trigger 不 dispatch,
296
+ // Run 关闭未消费时由 inbox 按原 Trigger 重投)
297
+ const offerResult = this.offerToTarget(trigger, agent, sessionId, options?.onEvent);
298
+ if (offerResult === 'steered' || offerResult === 'duplicate') {
299
+ options?.onDelivery?.('steered');
300
+ return { status: 'accepted', agentId: agent.agentId, sessionId, delivery: 'steered' };
301
+ }
302
+ options?.onDelivery?.('run');
246
303
  if (trigger.coalesceKey) {
247
304
  this.emitRuntime(RuntimeEvents.RUN_SCHEDULED, {
248
305
  triggerId: trigger.id,
@@ -291,15 +348,46 @@ export class AgentRuntime {
291
348
  triggerType: triggers[0]?.type,
292
349
  });
293
350
  try {
351
+ // loop 终态如实镜像到 RUN_ENDED / DispatchResult:LLM 彻底失败以
352
+ // engine.error 事件收尾(不 throw),不得记成 completed——否则事故在
353
+ // 监控与 dispatch 结果里隐身(10-27 事故同族问题)
354
+ let endReason;
355
+ let endError;
294
356
  for await (const event of agent.dispatcher.execute(request, controller.signal)) {
357
+ // 所有 Trigger(含后台 signal)共用一个 run-event 出口;调用方的
358
+ // onEvent 只承担该次 dispatch 的业务收尾,不再是后台 Run 的唯一可见通道。
359
+ this.emitRuntime(RuntimeEvents.RUN_EVENT, {
360
+ requestId: request.requestId,
361
+ triggerId: triggers[0]?.id,
362
+ triggerType: triggers[0]?.type,
363
+ // 路由键必带:gateway 按 sessionId 扇出 WS(缺则整条流式/终态链路静默落空)
364
+ agentId: agent.agentId,
365
+ sessionId,
366
+ event,
367
+ });
368
+ if (event.type === 'engine.end') {
369
+ const reason = event.data?.reason;
370
+ if (endReason !== 'error') {
371
+ // error 粘滞:合批(coalesce)里前一条 entry 失败、后一条成功时
372
+ // 不得被翻回 completed——任一 entry 失败即整体失败
373
+ endReason = reason === 'should_stop' || reason === 'aborted' ? reason : 'completed';
374
+ }
375
+ }
376
+ else if (event.type === 'engine.error') {
377
+ endReason = 'error';
378
+ const rawError = event.data?.error;
379
+ // 与 runner 同口径取 message(自定义 dispatcher 直传 Error 时不留 "Error: " 前缀)
380
+ endError = rawError instanceof Error ? rawError.message : String(rawError ?? 'Unknown error');
381
+ }
295
382
  try {
296
383
  onEvent?.(event);
297
384
  }
298
385
  catch {
299
- // 消费者(WS 广播等)异常不中断 Run
386
+ // 消费者(频道最终回复等)异常不中断 Run
300
387
  }
301
388
  }
302
- if (controller.signal.aborted) {
389
+ // error 优先于 aborted:失败与中止同刻到达时,失败不得在监控里消失
390
+ if (controller.signal.aborted && endReason !== 'error') {
303
391
  this.emitRuntime(RuntimeEvents.RUN_ENDED, {
304
392
  requestId: request.requestId,
305
393
  agentId: agent.agentId,
@@ -312,8 +400,12 @@ export class AgentRuntime {
312
400
  requestId: request.requestId,
313
401
  agentId: agent.agentId,
314
402
  sessionId,
315
- reason: 'completed',
403
+ reason: endReason ?? 'completed',
404
+ ...(endError ? { error: endError } : {}),
316
405
  });
406
+ if (endReason === 'error') {
407
+ return { status: 'failed', error: endError ?? 'unknown' };
408
+ }
317
409
  return {
318
410
  status: 'ran',
319
411
  requestId: request.requestId,
@@ -95,6 +95,17 @@ export interface RunAuditRecord {
95
95
  }
96
96
  export interface RunDispatcher {
97
97
  execute(req: RunRequest, signal?: AbortSignal): AsyncIterable<AgentEvent>;
98
+ /** 可选:活跃 Run 接续;缺省保持原 Trigger 排队语义。 */
99
+ inbox?: {
100
+ offer(signal: import('./inbox.js').SessionInboxSignal): import('./inbox.js').InboxOfferResult;
101
+ };
102
+ }
103
+ /** dispatch 调用方选项 */
104
+ export interface DispatchOptions {
105
+ /** Run 事件流转发(频道回复捕获等);不变量 #6 */
106
+ onEvent?: (event: AgentEvent) => void;
107
+ /** 活跃 Run 接收即回执(早于 Run 结束);run = 走新 Run。 */
108
+ onDelivery?: (delivery: 'steered' | 'run') => void;
98
109
  }
99
110
  export interface RouteTarget {
100
111
  agentId: string;
@@ -127,6 +138,12 @@ export type DispatchResult = {
127
138
  requestId: string;
128
139
  agentId: string;
129
140
  sessionId: string;
141
+ } | {
142
+ /** 已交活跃 Run 在下一消息边界接续(不开新 Run)。 */
143
+ status: 'accepted';
144
+ agentId: string;
145
+ sessionId: string;
146
+ delivery: 'steered';
130
147
  } | {
131
148
  status: 'skipped';
132
149
  reason: 'no_agent' | 'filtered' | 'runtime_stopped' | 'aborted';
@@ -159,6 +176,8 @@ export declare const RuntimeEvents: {
159
176
  readonly RUN_FAILED: "runtime.run.failed";
160
177
  readonly AGENT_SIGNAL_EMITTED: "runtime.agent_signal.emitted";
161
178
  readonly TRIGGER_DROPPED: "runtime.trigger.dropped";
179
+ readonly RUN_EVENT: "runtime.run.event";
180
+ readonly INBOX_ADMITTED: "runtime.inbox.admitted";
162
181
  };
163
182
  export type RuntimeEventType = (typeof RuntimeEvents)[keyof typeof RuntimeEvents];
164
183
  export interface RuntimeEvent {
@@ -12,4 +12,6 @@ export const RuntimeEvents = {
12
12
  RUN_FAILED: 'runtime.run.failed',
13
13
  AGENT_SIGNAL_EMITTED: 'runtime.agent_signal.emitted',
14
14
  TRIGGER_DROPPED: 'runtime.trigger.dropped',
15
+ RUN_EVENT: 'runtime.run.event',
16
+ INBOX_ADMITTED: 'runtime.inbox.admitted',
15
17
  };
@@ -18,7 +18,8 @@ import type { ModelProvider } from '@octopi-agent/core/interfaces/model-provider
18
18
  import { Agent } from '../run/agent/index.js';
19
19
  import type { ReliabilityConfig, ReliabilityHarness } from '../run/reliability/run-agent.js';
20
20
  import type { ContextEngine, SummarizeFunction } from '../context/types.js';
21
- import type { ErrorStrategy } from '@octopi-agent/core/interfaces/error-strategy.js';
21
+ import type { ErrorStrategy, ClassifiedError, ErrorAction, OverflowAction } from '@octopi-agent/core/interfaces/error-strategy.js';
22
+ import type { SecurityViolation, SecurityAction } from '@octopi-agent/core/security-guard.js';
22
23
  import type { RunGuard } from '@octopi-agent/core/interfaces/run-guard.js';
23
24
  import type { RunGuardConfig } from '../run/run-guard/default-run-guard.js';
24
25
  import type { Observer } from '@octopi-agent/core/interfaces/observer.js';
@@ -31,6 +32,13 @@ import { SessionAwareRunner } from '../run/runner.js';
31
32
  import type { SessionAwareRunnerConfig } from '../run/runner.js';
32
33
  import type { McpClientFactory } from '../extension/plugin-ecosystem/mcp/manager.js';
33
34
  import type { McpServerConfig, McpManager } from '../extension/plugin-ecosystem/mcp/types.js';
35
+ /** 默认错误策略(按原因分级重试 + 中止;上限见 MODEL_ERROR_MAX_ATTEMPTS) */
36
+ export declare class DefaultErrorStrategy implements ErrorStrategy {
37
+ onModelError(error: ClassifiedError, attempt: number): ErrorAction;
38
+ onToolError(error: ClassifiedError, _call: any): ErrorAction;
39
+ onContextOverflow(_tokenCount: number, _limit: number): OverflowAction;
40
+ onSecurityViolation(violation: SecurityViolation): SecurityAction;
41
+ }
34
42
  /** 工具运行时上下文提供者 */
35
43
  export interface ToolContextProvider {
36
44
  get(): {
@@ -46,6 +54,12 @@ export interface ToolContextProvider {
46
54
  pathContainment?: 'strict' | 'soft' | 'none';
47
55
  confirmEffect?: import('@octopi-agent/core/types/tools.js').ToolExecutionContext['confirmEffect'];
48
56
  effectChannel?: 'engine' | 'host-gate';
57
+ /** 提权策略(ops P3/P4,resolved;与 core ToolExecutionContext 同源) */
58
+ escalation?: import('@octopi-agent/core/types/tools.js').ToolExecutionContext['escalation'];
59
+ /** 危险命令减速带 resolved(§6;与 core ToolExecutionContext 同源) */
60
+ dangerousCommands?: import('@octopi-agent/core/types/tools.js').ToolExecutionContext['dangerousCommands'];
61
+ /** 参与者身份(SEC-2 / ops P5;与 core ToolExecutionContext 同源) */
62
+ sessionParticipants?: import('@octopi-agent/core/types/tools.js').ToolExecutionContext['sessionParticipants'];
49
63
  processLimits?: boolean;
50
64
  shellEscape?: 'deny' | 'confirm' | 'allow';
51
65
  shellJournal?: 'snapshot' | 'off';
@@ -37,24 +37,30 @@ import { createProviderSummarize } from '../context/summarize.js';
37
37
  import { createDefaultSystemPromptAssembler } from '../context/system-prompt-assembler.js';
38
38
  import { SessionAwareRunner } from '../run/runner.js';
39
39
  import { DefaultMcpManager } from '../extension/plugin-ecosystem/mcp/manager.js';
40
+ import { maxAttemptsFor } from '../run/reliability/model-error-policy.js';
40
41
  // ── 默认实现 ──
41
- /** 默认错误策略(简单重试 + 中止) */
42
- class DefaultErrorStrategy {
42
+ /** 默认错误策略(按原因分级重试 + 中止;上限见 MODEL_ERROR_MAX_ATTEMPTS) */
43
+ export class DefaultErrorStrategy {
43
44
  onModelError(error, attempt) {
45
+ const maxAttempts = maxAttemptsFor(error.reason);
44
46
  // ── 可重试错误:限流、超时、网络、服务端 ──
45
- if (error.reason === 'rate_limit' && attempt < 3) {
47
+ if (error.reason === 'rate_limit' && attempt < maxAttempts) {
46
48
  const delayMs = error.retryAfterMs ?? (attempt + 1) * 1000;
47
49
  return { action: 'retry', delayMs };
48
50
  }
49
- if (error.reason === 'timeout' && attempt < 3) {
51
+ if (error.reason === 'timeout' && attempt < maxAttempts) {
50
52
  return { action: 'retry', delayMs: (attempt + 1) * 1000 };
51
53
  }
52
- if (error.reason === 'network' && attempt < 3) {
54
+ if (error.reason === 'network' && attempt < maxAttempts) {
53
55
  return { action: 'retry', delayMs: (attempt + 1) * 1500 };
54
56
  }
55
- if (error.reason === 'server' && attempt < 2) {
57
+ if (error.reason === 'server' && attempt < maxAttempts) {
56
58
  return { action: 'retry', delayMs: (attempt + 1) * 2000 };
57
59
  }
60
+ // ── content_filter:内容过滤拒答多为阈值型误杀,原样重试常过;不教模型换说法绕过 ──
61
+ if (error.reason === 'content_filter' && attempt < maxAttempts) {
62
+ return { action: 'retry', delayMs: (attempt + 1) * 1000 };
63
+ }
58
64
  // ── context_length:由引擎尝试 compact,这里返回 abort 让引擎决定 ──
59
65
  if (error.reason === 'context_length') {
60
66
  return { action: 'abort', reason: `Context length exceeded: ${error.message}` };
@@ -116,11 +122,16 @@ class RuntimeToolContextProvider {
116
122
  networkPolicy: scope.toolRuntime.networkPolicy,
117
123
  requireCredentialBindings: scope.toolRuntime.requireCredentialBindings,
118
124
  requireJail: scope.toolRuntime.requireJail,
125
+ disableJail: scope.toolRuntime.disableJail,
119
126
  homeWrite: scope.toolRuntime.homeWrite,
120
127
  noteSecurityDegraded: scope.toolRuntime.noteSecurityDegraded,
128
+ recordOpsEvent: scope.toolRuntime.recordOpsEvent,
121
129
  pathContainment: scope.toolRuntime.pathContainment,
122
130
  confirmEffect: scope.toolRuntime.confirmEffect,
123
131
  effectChannel: scope.toolRuntime.effectChannel,
132
+ escalation: scope.toolRuntime.escalation,
133
+ dangerousCommands: scope.toolRuntime.dangerousCommands,
134
+ sessionParticipants: scope.toolRuntime.sessionParticipants,
124
135
  processLimits: scope.toolRuntime.processLimits,
125
136
  shellEscape: scope.toolRuntime.shellEscape,
126
137
  shellJournal: scope.toolRuntime.shellJournal,
@@ -165,7 +176,7 @@ function convertToAgentTool(tool, contextProvider) {
165
176
  description: tool.definition.description,
166
177
  executionMode: tool.executionMode,
167
178
  parameters: buildToolLlmParameters(tool.definition.parameters),
168
- execute: async (toolCallId, args, signal) => {
179
+ execute: async (toolCallId, args, signal, onUpdate) => {
169
180
  const startTime = Date.now();
170
181
  try {
171
182
  const context = contextProvider.get();
@@ -176,7 +187,12 @@ function convertToAgentTool(tool, contextProvider) {
176
187
  const handlerArgs = prepareHandlerArgs(rawArgs, tool.definition);
177
188
  const { runWithToolLimits, resolveToolTimeoutMs } = await import('../extension/execution-environment/resource-limits.js');
178
189
  const processLimitsEnabled = context.processLimits !== false;
179
- const result = await runWithToolLimits((sig) => tool.handler(handlerArgs, { ...context, abortSignal: sig }), {
190
+ const result = await runWithToolLimits((sig) => tool.handler(handlerArgs, {
191
+ ...context,
192
+ abortSignal: sig,
193
+ // 工具部分输出通道(展示用途);无转发器时不给工具假通道
194
+ ...(onUpdate ? { onUpdate } : {}),
195
+ }), {
180
196
  timeoutMs: processLimitsEnabled
181
197
  ? resolveToolTimeoutMs(tool.definition)
182
198
  : undefined,
@@ -1334,7 +1350,11 @@ export class AgentBuilder {
1334
1350
  const snap = getResolvedModel();
1335
1351
  const provider = snap?.provider ?? this._model;
1336
1352
  const model = snap?.modelName;
1337
- return createProviderSummarize(provider, { model })(messages, opts);
1353
+ return createProviderSummarize(provider, {
1354
+ model,
1355
+ // F8R-04:源头钳制与 gateway 对齐——自动路径同样受模型输出上限约束
1356
+ maxOutputTokens: snap?.maxOutputTokens,
1357
+ })(messages, opts);
1338
1358
  }
1339
1359
  : undefined);
1340
1360
  agent.setConvertToLlm(async (messages) => {
@@ -1366,6 +1386,7 @@ export class AgentBuilder {
1366
1386
  tools,
1367
1387
  tokenBudget: contextWindow,
1368
1388
  contextWindow,
1389
+ maxOutputTokens: snapshot?.maxOutputTokens,
1369
1390
  compactTargetTokens: this._contextAssemblerConfig?.compactTargetTokens,
1370
1391
  summarize: summarizeFn,
1371
1392
  loadCompactState: () => agent.getSessionCompactState(sessionId, agentId),
@@ -1537,7 +1558,7 @@ export class AgentBuilder {
1537
1558
  ${toolList}
1538
1559
 
1539
1560
  Tool selection (strict):
1540
- 1. Prefer a dedicated tool when one covers the job (file_read / file_list / file_write / file_edit / file_delete / file_mkdir / file_rename / file_copy / file_stat / file_rmdir / file_search / env_info, and any task-specific tool).
1561
+ 1. Prefer a dedicated tool when one covers the job (file_read / file_list / file_write / file_edit / file_delete / file_mkdir / file_rename / file_copy / file_stat / file_rmdir / file_search / env_info / process_list, and any task-specific tool).
1541
1562
  2. If a dedicated-tool call fails because your arguments were wrong (bad path, bad params), fix the call and retry that tool — do not switch to shell for a call you can fix.
1542
1563
  3. Use shell only when: (a) no dedicated tool covers the operation, or (b) the dedicated tool is unavailable, or (c) it still fails after a correct retry and shell is the only remaining way to make progress.
1543
1564
  4. Do not prefer shell over a working dedicated tool for the same job.
@@ -85,4 +85,9 @@ export interface BuildFromConfigOptions {
85
85
  * 委派编排开关。缺省 = 只要有 agent 声明 `delegation` 策略就接线。
86
86
  */
87
87
  delegation?: boolean;
88
+ /**
89
+ * 宿主注入的 Runner 附加配置(浅合并进每个 agent)。
90
+ * ops P2:`securityIntegrity`(安全配置漂移守卫)与 `opsAuditPath` 走此通道。
91
+ */
92
+ runnerConfig?: import('../run/runner.js').SessionAwareRunnerConfig;
88
93
  }
@@ -131,6 +131,7 @@ export async function buildFromConfig(config, options) {
131
131
  modelCallConfig: config.modelCall,
132
132
  concurrencyConfig: config.concurrency,
133
133
  observabilityConfig: config.observability,
134
+ runnerConfig: options?.runnerConfig,
134
135
  });
135
136
  agents.set(agentConfig.id, built);
136
137
  }
@@ -171,6 +172,8 @@ export async function buildFromConfig(config, options) {
171
172
  */
172
173
  async function buildAgent(agentConfig, shared) {
173
174
  const builder = new AgentBuilder();
175
+ if (shared.runnerConfig)
176
+ builder.runnerConfig(shared.runnerConfig);
174
177
  // 解析主模型 provider(宿主可在此包装 circuit breaker;config-bridge 用原始 provider)
175
178
  const resolvedModel = resolveModelConfig(agentConfig.model, shared.flatModels);
176
179
  const provider = shared.providers.get(resolvedModel.provider);
@@ -0,0 +1,72 @@
1
+ /**
2
+ * 安全配置指纹与漂移守卫(ops P2,`arch/ops-privileged-mode.md` §4.3 / §8.4 拍板)
3
+ *
4
+ * 「agent 可写自身配置」是自提权面:host 模式下 agent 改写 `octopi.json` 的安全字段
5
+ * 等于自我升档。边界(§8.4):安全字段**重启才生效**,运行中检测到漂移 → **拒新 run**
6
+ * + `ops.config.security_changed`,直至宿主(重启 / WebUI 署名通道)确认。
7
+ *
8
+ * 本模块只做「指纹 + 对比」;重读配置文件由宿主注入(引擎不越界做 IO 策略)。
9
+ */
10
+ /** 安全字段指纹(锚定值 + 字段摘要) */
11
+ export interface SecurityConfigFingerprint {
12
+ /** sha256(canonicalJson(fields)) */
13
+ fingerprint: string;
14
+ /** 扁平字段摘要(`effectPolicy.jail` → 值),审计记录与 diff 用 */
15
+ fields: Record<string, unknown>;
16
+ }
17
+ /** 指纹对比结果 */
18
+ export interface SecurityConfigDrift {
19
+ ok: boolean;
20
+ /** 发生变化的字段名(指纹不同则至少一项) */
21
+ changed: string[];
22
+ /** 当前指纹(审计记录的「新摘要」) */
23
+ currentFingerprint: string;
24
+ }
25
+ /**
26
+ * 计算安全配置指纹(启动期锚定)。
27
+ *
28
+ * @param config - 已解析配置的安全相关切面(effectPolicy + security)
29
+ * @returns 指纹 + 字段摘要
30
+ */
31
+ export declare function computeSecurityConfigFingerprint(config: {
32
+ effectPolicy?: Record<string, unknown>;
33
+ security?: Record<string, unknown>;
34
+ agents?: Array<Record<string, unknown>>;
35
+ plugins?: Record<string, unknown>;
36
+ }): SecurityConfigFingerprint;
37
+ /**
38
+ * 从原始 JSON 文本提取安全字段切面(漂移复核用;不走完整 schema 解析——
39
+ * 被篡改的配置可能根本过不了 schema,指纹对比必须先于校验)。
40
+ *
41
+ * 覆盖面(§8.4 + 评审补洞):顶层 effectPolicy / security、**agents[].effectPolicy**
42
+ * (每 agent 覆盖顶层,改单个 agent 的 jail 也必须算漂移)、**agents[].tools**(工具
43
+ * allow/deny)、**plugins.loadPaths**(插件加载路径)。
44
+ *
45
+ * @param jsonText - 配置文件全文
46
+ * @returns 安全字段切面
47
+ */
48
+ export declare function extractSecurityConfigSlice(jsonText: string): {
49
+ effectPolicy: Record<string, unknown>;
50
+ security: Record<string, unknown>;
51
+ agents: Array<Record<string, unknown>>;
52
+ plugins: Record<string, unknown>;
53
+ };
54
+ /**
55
+ * 对比锚定指纹与当前指纹。
56
+ *
57
+ * @param anchor - 启动期指纹
58
+ * @param current - 当前指纹
59
+ * @returns 漂移结论(changed 为字段名列表)
60
+ */
61
+ export declare function diffSecurityConfigFingerprint(anchor: SecurityConfigFingerprint, current: SecurityConfigFingerprint): SecurityConfigDrift;
62
+ /**
63
+ * 基于配置文件的漂移复核器(宿主注入 Runner 的 `securityIntegrity.recheck`)。
64
+ *
65
+ * @param options.configPath - 配置文件绝对路径
66
+ * @param options.anchor - 启动期锚定指纹
67
+ * @returns 复核函数:ok=false 时 changed 非空
68
+ */
69
+ export declare function createFileSecurityIntegrity(options: {
70
+ configPath: string;
71
+ anchor: SecurityConfigFingerprint;
72
+ }): () => SecurityConfigDrift;
@@ -0,0 +1,146 @@
1
+ /**
2
+ * 安全配置指纹与漂移守卫(ops P2,`arch/ops-privileged-mode.md` §4.3 / §8.4 拍板)
3
+ *
4
+ * 「agent 可写自身配置」是自提权面:host 模式下 agent 改写 `octopi.json` 的安全字段
5
+ * 等于自我升档。边界(§8.4):安全字段**重启才生效**,运行中检测到漂移 → **拒新 run**
6
+ * + `ops.config.security_changed`,直至宿主(重启 / WebUI 署名通道)确认。
7
+ *
8
+ * 本模块只做「指纹 + 对比」;重读配置文件由宿主注入(引擎不越界做 IO 策略)。
9
+ */
10
+ import { readFileSync } from 'node:fs';
11
+ import { createHash } from 'node:crypto';
12
+ /**
13
+ * 参与指纹的安全字段(§8.4:jail / escalation / effects / 审计开关 + 作用域面)。
14
+ * 键序无关;值走稳定序列化。
15
+ */
16
+ const EFFECT_POLICY_SECURITY_KEYS = [
17
+ 'mode',
18
+ 'untrustedAdmission',
19
+ 'pathContainment',
20
+ 'shellEscape',
21
+ 'jail',
22
+ 'onJailUnavailable',
23
+ 'privilegedAck',
24
+ 'effects',
25
+ 'toolIsolation',
26
+ 'opsAudit',
27
+ 'opsAuditPath',
28
+ // P5 审计外送也是安全面(换收口 = 换见证人)
29
+ 'opsAuditExportPath',
30
+ 'network',
31
+ 'escalation',
32
+ // ops P4:危险命令减速带也是安全闸(改档 = 放松/收紧确认闸,须重启 + 宿主署名)
33
+ 'dangerousCommands',
34
+ ];
35
+ const SECURITY_CONFIG_SECURITY_KEYS = ['enforce', 'injectionSensitivity', 'allowedPaths'];
36
+ function pickFields(source, keys, prefix) {
37
+ const out = {};
38
+ for (const key of keys) {
39
+ if (source[key] !== undefined)
40
+ out[`${prefix}${key}`] = source[key];
41
+ }
42
+ return out;
43
+ }
44
+ function stableStringify(value) {
45
+ if (value === null || typeof value !== 'object')
46
+ return JSON.stringify(value) ?? 'null';
47
+ if (Array.isArray(value))
48
+ return `[${value.map((v) => stableStringify(v)).join(',')}]`;
49
+ const entries = Object.entries(value)
50
+ .filter(([, v]) => v !== undefined)
51
+ .sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0));
52
+ return `{${entries.map(([k, v]) => `${JSON.stringify(k)}:${stableStringify(v)}`).join(',')}}`;
53
+ }
54
+ /**
55
+ * 计算安全配置指纹(启动期锚定)。
56
+ *
57
+ * @param config - 已解析配置的安全相关切面(effectPolicy + security)
58
+ * @returns 指纹 + 字段摘要
59
+ */
60
+ export function computeSecurityConfigFingerprint(config) {
61
+ const fields = {
62
+ ...pickFields(config.effectPolicy ?? {}, EFFECT_POLICY_SECURITY_KEYS, 'effectPolicy.'),
63
+ ...pickFields(config.security ?? {}, SECURITY_CONFIG_SECURITY_KEYS, 'security.'),
64
+ };
65
+ // per-agent 覆盖与工具面/插件加载路径都是安全面(§8.4 补洞)
66
+ if (config.agents !== undefined)
67
+ fields['agents'] = config.agents;
68
+ if (config.plugins !== undefined && Object.keys(config.plugins).length > 0) {
69
+ fields['plugins'] = pickFields(config.plugins, ['loadPaths'], 'plugins.');
70
+ }
71
+ return {
72
+ fingerprint: createHash('sha256').update(stableStringify(fields)).digest('hex'),
73
+ fields,
74
+ };
75
+ }
76
+ /**
77
+ * 从原始 JSON 文本提取安全字段切面(漂移复核用;不走完整 schema 解析——
78
+ * 被篡改的配置可能根本过不了 schema,指纹对比必须先于校验)。
79
+ *
80
+ * 覆盖面(§8.4 + 评审补洞):顶层 effectPolicy / security、**agents[].effectPolicy**
81
+ * (每 agent 覆盖顶层,改单个 agent 的 jail 也必须算漂移)、**agents[].tools**(工具
82
+ * allow/deny)、**plugins.loadPaths**(插件加载路径)。
83
+ *
84
+ * @param jsonText - 配置文件全文
85
+ * @returns 安全字段切面
86
+ */
87
+ export function extractSecurityConfigSlice(jsonText) {
88
+ const parsed = JSON.parse(jsonText);
89
+ const rawAgents = Array.isArray(parsed.agents) ? parsed.agents : [];
90
+ const agents = rawAgents.map((a) => {
91
+ const agent = (a ?? {});
92
+ return {
93
+ id: agent.id,
94
+ tools: agent.tools,
95
+ ...pickFields((agent.effectPolicy ?? {}), EFFECT_POLICY_SECURITY_KEYS, 'effectPolicy.'),
96
+ };
97
+ });
98
+ return {
99
+ effectPolicy: (parsed.effectPolicy ?? {}),
100
+ security: (parsed.security ?? {}),
101
+ agents,
102
+ plugins: (parsed.plugins ?? {}),
103
+ };
104
+ }
105
+ /**
106
+ * 对比锚定指纹与当前指纹。
107
+ *
108
+ * @param anchor - 启动期指纹
109
+ * @param current - 当前指纹
110
+ * @returns 漂移结论(changed 为字段名列表)
111
+ */
112
+ export function diffSecurityConfigFingerprint(anchor, current) {
113
+ if (anchor.fingerprint === current.fingerprint) {
114
+ return { ok: true, changed: [], currentFingerprint: current.fingerprint };
115
+ }
116
+ const changed = new Set();
117
+ for (const key of new Set([...Object.keys(anchor.fields), ...Object.keys(current.fields)])) {
118
+ if (stableStringify(anchor.fields[key]) !== stableStringify(current.fields[key]))
119
+ changed.add(key);
120
+ }
121
+ return { ok: false, changed: [...changed].sort(), currentFingerprint: current.fingerprint };
122
+ }
123
+ /**
124
+ * 基于配置文件的漂移复核器(宿主注入 Runner 的 `securityIntegrity.recheck`)。
125
+ *
126
+ * @param options.configPath - 配置文件绝对路径
127
+ * @param options.anchor - 启动期锚定指纹
128
+ * @returns 复核函数:ok=false 时 changed 非空
129
+ */
130
+ export function createFileSecurityIntegrity(options) {
131
+ return () => {
132
+ let slice;
133
+ try {
134
+ slice = extractSecurityConfigSlice(readFileSync(options.configPath, 'utf8'));
135
+ }
136
+ catch (err) {
137
+ // 读不到/解析不了 = 无法证明未被篡改,按漂移对待(fail-closed)
138
+ return {
139
+ ok: false,
140
+ changed: [`<config unreadable: ${err instanceof Error ? err.message : String(err)}>`],
141
+ currentFingerprint: '<unreadable>',
142
+ };
143
+ }
144
+ return diffSecurityConfigFingerprint(options.anchor, computeSecurityConfigFingerprint(slice));
145
+ };
146
+ }
@@ -61,8 +61,9 @@ export function resolveContextEngine(config) {
61
61
  // 缺省 type = default:只给旋钮(如 proactiveCompactRatio)也必须落到默认引擎
62
62
  if (!config || config.type == null || config.type === 'default') {
63
63
  return new DefaultContextEngine({
64
- protectFirstN: config?.protectFirstN ?? 3,
64
+ protectFirstN: config?.protectFirstN ?? 5,
65
65
  protectLastN: config?.protectLastN ?? 20,
66
+ summaryBudget: config?.summaryBudget,
66
67
  proactiveCompactRatio: config?.proactiveCompactRatio ?? 0.6,
67
68
  proactiveCooldownMs: config?.proactiveCooldownMs ?? 30_000,
68
69
  historyAging: config?.historyAging,
@@ -220,7 +220,7 @@ export class ThinkExecutor {
220
220
  .map(([key]) => key),
221
221
  },
222
222
  execute: t.handler
223
- ? async (toolCallId, args) => {
223
+ ? async (toolCallId, args, _signal, onUpdate) => {
224
224
  try {
225
225
  const result = await t.handler(args, {
226
226
  sessionId: 'subsystem',
@@ -228,6 +228,8 @@ export class ThinkExecutor {
228
228
  messages: [],
229
229
  fileEffects: createUnconfiguredFileEffects(),
230
230
  memoryEffects: createUnconfiguredMemoryEffects(),
231
+ // 与 convertToAgentTool 同构:工具部分输出通道(think 子系统 loop 未挂 onToolUpdate 时为空)
232
+ ...(onUpdate ? { onUpdate } : {}),
231
233
  });
232
234
  return {
233
235
  toolCallId,