@cogitator-ai/core 0.22.0 → 0.24.0

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 (152) hide show
  1. package/README.md +95 -0
  2. package/dist/agent-tool.d.ts +7 -1
  3. package/dist/agent-tool.d.ts.map +1 -1
  4. package/dist/agent-tool.js +4 -1
  5. package/dist/agent-tool.js.map +1 -1
  6. package/dist/agent.d.ts +6 -25
  7. package/dist/agent.d.ts.map +1 -1
  8. package/dist/agent.js +18 -3
  9. package/dist/agent.js.map +1 -1
  10. package/dist/cogitator/handoffs.d.ts +14 -0
  11. package/dist/cogitator/handoffs.d.ts.map +1 -0
  12. package/dist/cogitator/handoffs.js +53 -0
  13. package/dist/cogitator/handoffs.js.map +1 -0
  14. package/dist/cogitator/initializers.d.ts.map +1 -1
  15. package/dist/cogitator/initializers.js +14 -2
  16. package/dist/cogitator/initializers.js.map +1 -1
  17. package/dist/cogitator/message-builder.d.ts +1 -1
  18. package/dist/cogitator/message-builder.d.ts.map +1 -1
  19. package/dist/cogitator/message-builder.js +23 -8
  20. package/dist/cogitator/message-builder.js.map +1 -1
  21. package/dist/cogitator/prompts.d.ts +69 -0
  22. package/dist/cogitator/prompts.d.ts.map +1 -0
  23. package/dist/cogitator/prompts.js +149 -0
  24. package/dist/cogitator/prompts.js.map +1 -0
  25. package/dist/cogitator/run-checkpoints.d.ts +20 -0
  26. package/dist/cogitator/run-checkpoints.d.ts.map +1 -0
  27. package/dist/cogitator/run-checkpoints.js +67 -0
  28. package/dist/cogitator/run-checkpoints.js.map +1 -0
  29. package/dist/cogitator/streaming.d.ts +9 -7
  30. package/dist/cogitator/streaming.d.ts.map +1 -1
  31. package/dist/cogitator/streaming.js +12 -1
  32. package/dist/cogitator/streaming.js.map +1 -1
  33. package/dist/cogitator/threads.d.ts +27 -0
  34. package/dist/cogitator/threads.d.ts.map +1 -0
  35. package/dist/cogitator/threads.js +54 -0
  36. package/dist/cogitator/threads.js.map +1 -0
  37. package/dist/cogitator/tool-executor.d.ts +1 -1
  38. package/dist/cogitator/tool-executor.d.ts.map +1 -1
  39. package/dist/cogitator/tool-executor.js +2 -2
  40. package/dist/cogitator/tool-executor.js.map +1 -1
  41. package/dist/constitutional/constitutional-ai.d.ts +7 -1
  42. package/dist/constitutional/constitutional-ai.d.ts.map +1 -1
  43. package/dist/constitutional/constitutional-ai.js +6 -2
  44. package/dist/constitutional/constitutional-ai.js.map +1 -1
  45. package/dist/constitutional/tool-guard.d.ts +3 -1
  46. package/dist/constitutional/tool-guard.d.ts.map +1 -1
  47. package/dist/constitutional/tool-guard.js +2 -2
  48. package/dist/constitutional/tool-guard.js.map +1 -1
  49. package/dist/index.d.ts +7 -1
  50. package/dist/index.d.ts.map +1 -1
  51. package/dist/index.js +6 -1
  52. package/dist/index.js.map +1 -1
  53. package/dist/learning/postgres-trace-store.d.ts +5 -1
  54. package/dist/learning/postgres-trace-store.d.ts.map +1 -1
  55. package/dist/learning/postgres-trace-store.js +23 -0
  56. package/dist/learning/postgres-trace-store.js.map +1 -1
  57. package/dist/learning/prompt-stores.d.ts +26 -0
  58. package/dist/learning/prompt-stores.d.ts.map +1 -0
  59. package/dist/learning/prompt-stores.js +97 -0
  60. package/dist/learning/prompt-stores.js.map +1 -0
  61. package/dist/llm/anthropic-thinking.d.ts +29 -0
  62. package/dist/llm/anthropic-thinking.d.ts.map +1 -0
  63. package/dist/llm/anthropic-thinking.js +100 -0
  64. package/dist/llm/anthropic-thinking.js.map +1 -0
  65. package/dist/llm/anthropic.d.ts +14 -0
  66. package/dist/llm/anthropic.d.ts.map +1 -1
  67. package/dist/llm/anthropic.js +139 -85
  68. package/dist/llm/anthropic.js.map +1 -1
  69. package/dist/llm/azure.d.ts +2 -0
  70. package/dist/llm/azure.d.ts.map +1 -1
  71. package/dist/llm/azure.js +1 -0
  72. package/dist/llm/azure.js.map +1 -1
  73. package/dist/llm/bedrock.d.ts +2 -0
  74. package/dist/llm/bedrock.d.ts.map +1 -1
  75. package/dist/llm/bedrock.js +139 -39
  76. package/dist/llm/bedrock.js.map +1 -1
  77. package/dist/llm/claude-models.d.ts +23 -0
  78. package/dist/llm/claude-models.d.ts.map +1 -1
  79. package/dist/llm/claude-models.js +50 -0
  80. package/dist/llm/claude-models.js.map +1 -1
  81. package/dist/llm/errors.d.ts +5 -0
  82. package/dist/llm/errors.d.ts.map +1 -1
  83. package/dist/llm/errors.js +36 -26
  84. package/dist/llm/errors.js.map +1 -1
  85. package/dist/llm/google.d.ts +12 -1
  86. package/dist/llm/google.d.ts.map +1 -1
  87. package/dist/llm/google.js +91 -15
  88. package/dist/llm/google.js.map +1 -1
  89. package/dist/llm/index.d.ts +7 -2
  90. package/dist/llm/index.d.ts.map +1 -1
  91. package/dist/llm/index.js +16 -2
  92. package/dist/llm/index.js.map +1 -1
  93. package/dist/llm/ollama.d.ts +6 -1
  94. package/dist/llm/ollama.d.ts.map +1 -1
  95. package/dist/llm/ollama.js +33 -3
  96. package/dist/llm/ollama.js.map +1 -1
  97. package/dist/llm/openai-compatible-base.d.ts.map +1 -1
  98. package/dist/llm/openai-compatible-base.js +29 -21
  99. package/dist/llm/openai-compatible-base.js.map +1 -1
  100. package/dist/llm/openai-responses.d.ts.map +1 -1
  101. package/dist/llm/openai-responses.js +20 -3
  102. package/dist/llm/openai-responses.js.map +1 -1
  103. package/dist/llm/openai.d.ts +2 -0
  104. package/dist/llm/openai.d.ts.map +1 -1
  105. package/dist/llm/openai.js +1 -0
  106. package/dist/llm/openai.js.map +1 -1
  107. package/dist/llm/retry.d.ts +32 -0
  108. package/dist/llm/retry.d.ts.map +1 -0
  109. package/dist/llm/retry.js +107 -0
  110. package/dist/llm/retry.js.map +1 -0
  111. package/dist/logger.d.ts.map +1 -1
  112. package/dist/logger.js +3 -5
  113. package/dist/logger.js.map +1 -1
  114. package/dist/runtime.d.ts +49 -1
  115. package/dist/runtime.d.ts.map +1 -1
  116. package/dist/runtime.js +571 -187
  117. package/dist/runtime.js.map +1 -1
  118. package/dist/security/pii.d.ts +44 -0
  119. package/dist/security/pii.d.ts.map +1 -0
  120. package/dist/security/pii.js +285 -0
  121. package/dist/security/pii.js.map +1 -0
  122. package/dist/skill.d.ts.map +1 -1
  123. package/dist/skill.js +16 -10
  124. package/dist/skill.js.map +1 -1
  125. package/dist/tools/audio-generate.d.ts +1 -1
  126. package/dist/tools/audio-generate.d.ts.map +1 -1
  127. package/dist/tools/audio-generate.js +2 -1
  128. package/dist/tools/audio-generate.js.map +1 -1
  129. package/dist/tools/audio-transcribe.d.ts.map +1 -1
  130. package/dist/tools/audio-transcribe.js +2 -1
  131. package/dist/tools/audio-transcribe.js.map +1 -1
  132. package/dist/tools/email.d.ts.map +1 -1
  133. package/dist/tools/email.js +10 -9
  134. package/dist/tools/email.js.map +1 -1
  135. package/dist/tools/github.d.ts.map +1 -1
  136. package/dist/tools/github.js +2 -1
  137. package/dist/tools/github.js.map +1 -1
  138. package/dist/tools/image-generate.d.ts +2 -2
  139. package/dist/tools/image-generate.d.ts.map +1 -1
  140. package/dist/tools/image-generate.js +2 -1
  141. package/dist/tools/image-generate.js.map +1 -1
  142. package/dist/tools/vector-search.d.ts.map +1 -1
  143. package/dist/tools/vector-search.js +8 -7
  144. package/dist/tools/vector-search.js.map +1 -1
  145. package/dist/tools/web-search.d.ts.map +1 -1
  146. package/dist/tools/web-search.js +7 -6
  147. package/dist/tools/web-search.js.map +1 -1
  148. package/dist/utils/env.d.ts +9 -0
  149. package/dist/utils/env.d.ts.map +1 -0
  150. package/dist/utils/env.js +26 -0
  151. package/dist/utils/env.js.map +1 -0
  152. package/package.json +6 -6
package/dist/runtime.js CHANGED
@@ -1,9 +1,11 @@
1
1
  import { nanoid } from 'nanoid';
2
- import { getPrice } from '@cogitator-ai/models';
2
+ import { calculateCost as calculateModelCost } from '@cogitator-ai/models';
3
3
  import { ToolRegistry } from './registry.js';
4
4
  import { createLLMBackend, parseModel } from './llm/index.js';
5
5
  import { isLLMProvider } from './llm/providers.js';
6
6
  import { createLLMBackendFromPlugin, hasLLMPlugin } from './llm/plugin.js';
7
+ import { withLLMRetry } from './llm/retry.js';
8
+ import { PiiMasker, withPiiMasking } from './security/pii.js';
7
9
  import { getLogger } from './logger.js';
8
10
  import { initializeMemory, initializeSandbox, initializeReflection, initializeGuardrails, initializeCostRouting, initializeSecurity, initializeContextManager, cleanupState, } from './cogitator/initializers.js';
9
11
  import { CogitatorError, ErrorCode } from '@cogitator-ai/types';
@@ -12,8 +14,12 @@ import { createSpan, getTextContent } from './cogitator/span-factory.js';
12
14
  import { executeTool, createToolMessage } from './cogitator/tool-executor.js';
13
15
  import { streamChat } from './cogitator/streaming.js';
14
16
  import { RunLimiter } from './cogitator/run-limiter.js';
17
+ import { findHandoffAgent, handoffTools } from './cogitator/handoffs.js';
18
+ import { PromptRegistry } from './cogitator/prompts.js';
19
+ import { InMemoryRunCheckpointStore, ThreadRunCheckpointStore } from './cogitator/run-checkpoints.js';
15
20
  import { parseStructuredOutput, toLLMResponseFormat } from './cogitator/response-format.js';
16
21
  import { CostEstimator } from './cost-routing/cost-estimator.js';
22
+ import { readEnv } from './utils/env.js';
17
23
  /** Run timeout when neither the run, the agent nor `limits.defaultTimeout` sets one. */
18
24
  const DEFAULT_RUN_TIMEOUT = 120_000;
19
25
  /**
@@ -67,6 +73,9 @@ const DEFAULT_RUN_TIMEOUT = 120_000;
67
73
  export class Cogitator {
68
74
  config;
69
75
  backends = new Map();
76
+ processCheckpoints;
77
+ promptRegistry;
78
+ threadCheckpoints;
70
79
  /** Global tool registry shared across all runs */
71
80
  tools = new ToolRegistry();
72
81
  state = {
@@ -130,7 +139,45 @@ export class Cogitator {
130
139
  * ```
131
140
  */
132
141
  async run(agent, options) {
133
- const runId = `run_${nanoid(12)}`;
142
+ return this.execute(agent, options);
143
+ }
144
+ /**
145
+ * Continue a run that paused for tool approvals (`status: 'paused'`) from its
146
+ * `checkpoint`. Approved calls run, declined ones answer the model with the
147
+ * reason, and the run goes on; calls without a decision pause it again.
148
+ *
149
+ * @example
150
+ * ```ts
151
+ * const paused = await cog.run(agent, { input: 'Refund order 42' });
152
+ * if (paused.status === 'paused') {
153
+ * await store.save(paused.checkpoint);
154
+ * // ... later, once someone approved it
155
+ * const result = await cog.resume(agent, checkpoint, {
156
+ * decisions: { [paused.pendingApprovals[0].toolCallId]: { approved: true } },
157
+ * });
158
+ * }
159
+ * ```
160
+ */
161
+ async resume(agent, target, options = {}) {
162
+ if (typeof target !== 'string' && target.version !== 1) {
163
+ throw new CogitatorError({
164
+ message: `Unsupported run checkpoint version: ${String(target.version)}`,
165
+ code: ErrorCode.VALIDATION_ERROR,
166
+ });
167
+ }
168
+ const { decisions = {}, defaultDecision, userId, ...rest } = options;
169
+ const threadId = typeof target === 'string' ? target : target.threadId;
170
+ return this.execute(agent, { ...rest, input: '', threadId, ...(userId !== undefined && { userId }) }, {
171
+ ...(typeof target === 'string' ? {} : { checkpoint: target }),
172
+ decisions,
173
+ ...(defaultDecision && { defaultDecision }),
174
+ checkUser: typeof target === 'string' || userId !== undefined,
175
+ });
176
+ }
177
+ async execute(agent, options, resumeFrom) {
178
+ let checkpoint = resumeFrom?.checkpoint;
179
+ let prompt;
180
+ let runId = checkpoint?.runId ?? `run_${nanoid(12)}`;
134
181
  const threadId = options.threadId ?? `thread_${nanoid(12)}`;
135
182
  const traceId = `trace_${nanoid(16)}`;
136
183
  const startTime = Date.now();
@@ -166,69 +213,109 @@ export class Cogitator {
166
213
  try {
167
214
  releaseRunSlot = await this.acquireRunSlot(abortController.signal);
168
215
  throwIfAborted(abortController.signal);
169
- options.onRunStart?.({ runId, agentId: agent.id, input: options.input, threadId });
216
+ if (!checkpoint) {
217
+ options.onRunStart?.({ runId, agentId: agent.id, input: options.input, threadId });
218
+ }
170
219
  const agentModel = this.resolveModel(agent);
171
220
  await this.initializeAll(agentModel);
172
- const input = await buildInputWithAudio(options.input, options.audio, {
173
- apiKey: this.config.llm?.providers?.openai?.apiKey ?? process.env.OPENAI_API_KEY,
174
- signal: abortController.signal,
175
- });
176
- const runOptions = input === options.input ? options : { ...options, input };
177
- const registry = new ToolRegistry();
178
- registry.registerMany(this.tools.getAll());
179
- if (agent.tools && agent.tools.length > 0) {
180
- registry.registerMany(agent.tools);
181
- }
182
- let effectiveModel = agentModel;
183
- let backend;
184
- let model;
185
- if (this.state.costRouter && this.config.costRouting?.autoSelectModel) {
186
- const recommendation = await this.state.costRouter.recommendModel(input);
187
- effectiveModel = `${recommendation.provider}/${recommendation.modelId}`;
188
- const budgetCheck = this.state.costRouter.checkBudget(recommendation.estimatedCost);
189
- if (!budgetCheck.allowed) {
190
- throw new Error(`Budget exceeded: ${budgetCheck.reason}`);
221
+ if (resumeFrom) {
222
+ checkpoint ??= (await this.runCheckpointStore().load(threadId)) ?? undefined;
223
+ if (!checkpoint) {
224
+ throw new CogitatorError({
225
+ message: `Thread ${threadId} has no paused run`,
226
+ code: ErrorCode.RUN_NOT_PAUSED,
227
+ });
191
228
  }
192
- backend = this.getBackend(effectiveModel, recommendation.provider);
193
- model = recommendation.modelId;
194
- }
195
- else {
196
- ({ backend, model } = this.route(effectiveModel, agent.config.provider));
197
- }
198
- const messages = await buildInitialMessages(agent, runOptions, threadId, this.state.memoryAdapter, this.state.contextBuilder);
199
- if (this.state.injectionDetector) {
200
- const injectionResult = await this.state.injectionDetector.analyze(input);
201
- if (injectionResult.action === 'blocked') {
202
- const threatTypes = injectionResult.threats.map((t) => t.type).join(', ');
229
+ if (resumeFrom.checkUser && checkpoint.userId !== options.userId) {
203
230
  throw new CogitatorError({
204
- message: `Prompt injection detected: ${threatTypes}`,
205
- code: ErrorCode.PROMPT_INJECTION_DETECTED,
206
- details: { threats: injectionResult.threats },
231
+ message: `The paused run in thread ${threadId} belongs to another user`,
232
+ code: ErrorCode.THREAD_ACCESS_DENIED,
207
233
  });
208
234
  }
235
+ runId = checkpoint.runId;
236
+ options = { ...options, userId: checkpoint.userId };
209
237
  }
210
- if (this.state.constitutionalAI?.config.filterInput) {
211
- const inputResult = await this.state.constitutionalAI.filterInput(input);
212
- if (!inputResult.allowed) {
213
- throw new Error(`Input blocked: ${inputResult.blockedReason ?? 'Policy violation'}`);
238
+ prompt = checkpoint?.prompt;
239
+ if (!checkpoint && (this.config.prompts || this.promptRegistry)) {
240
+ const resolution = await this.prompts.resolve(agent, threadId);
241
+ prompt = resolution.prompt;
242
+ if (resolution.instructions !== agent.instructions) {
243
+ agent = agent.clone({ id: agent.id, instructions: resolution.instructions });
214
244
  }
215
245
  }
216
- if (options.context) {
217
- addContextToMessages(messages, options.context);
246
+ let active = agent;
247
+ if (checkpoint?.activeAgent && checkpoint.activeAgent !== agent.name) {
248
+ const found = findHandoffAgent(agent, checkpoint.activeAgent);
249
+ if (!found) {
250
+ throw new CogitatorError({
251
+ message: `The paused run is in agent "${checkpoint.activeAgent}", which ${agent.name} cannot hand over to`,
252
+ code: ErrorCode.VALIDATION_ERROR,
253
+ });
254
+ }
255
+ active = found;
256
+ }
257
+ const handoffs = [...(checkpoint?.handoffs ?? [])];
258
+ const buildRegistry = (owner) => {
259
+ const ownerRegistry = new ToolRegistry();
260
+ ownerRegistry.registerMany(this.tools.getAll());
261
+ if (owner.tools.length > 0)
262
+ ownerRegistry.registerMany(owner.tools);
263
+ const handoff = handoffTools(owner);
264
+ ownerRegistry.registerMany(handoff.tools);
265
+ return { registry: ownerRegistry, targets: handoff.targets };
266
+ };
267
+ let { registry, targets: handoffTargets } = buildRegistry(active);
268
+ let effectiveModel = agentModel;
269
+ let routeProvider = agent.config.provider;
270
+ let backend;
271
+ let model;
272
+ let input;
273
+ let messages;
274
+ if (checkpoint) {
275
+ input = checkpoint.input;
276
+ effectiveModel = checkpoint.model;
277
+ routeProvider = checkpoint.provider;
278
+ ({ backend, model } = this.route(effectiveModel, routeProvider));
279
+ messages = [...checkpoint.messages];
218
280
  }
219
- if (this.state.memoryAdapter &&
220
- options.saveHistory !== false &&
221
- options.useMemory !== false) {
222
- const currentUserMessage = messages[messages.length - 1];
223
- await saveEntry(threadId, agent.id, currentUserMessage, this.state.memoryAdapter, undefined, undefined, options.onMemoryError);
281
+ else {
282
+ input = await buildInputWithAudio(options.input, options.audio, {
283
+ apiKey: this.config.llm?.providers?.openai?.apiKey ??
284
+ (options.audio?.length ? readEnv('OPENAI_API_KEY') : undefined),
285
+ signal: abortController.signal,
286
+ });
287
+ const runOptions = input === options.input ? options : { ...options, input };
288
+ if (this.state.costRouter && this.config.costRouting?.autoSelectModel) {
289
+ const recommendation = await this.state.costRouter.recommendModel(input);
290
+ effectiveModel = `${recommendation.provider}/${recommendation.modelId}`;
291
+ routeProvider = recommendation.provider;
292
+ const budgetCheck = this.state.costRouter.checkBudget(recommendation.estimatedCost);
293
+ if (!budgetCheck.allowed) {
294
+ throw new Error(`Budget exceeded: ${budgetCheck.reason}`);
295
+ }
296
+ backend = this.getBackend(effectiveModel, recommendation.provider);
297
+ model = recommendation.modelId;
298
+ }
299
+ else {
300
+ ({ backend, model } = this.route(effectiveModel, agent.config.provider));
301
+ }
302
+ await this.abandonPausedRun(agent, runOptions, threadId);
303
+ messages = await this.prepareMessages(agent, runOptions, input, threadId);
224
304
  }
225
- const allToolCalls = [];
226
- let totalInputTokens = 0;
227
- let totalOutputTokens = 0;
228
- let iterations = 0;
305
+ const allToolCalls = [...(checkpoint?.toolCalls ?? [])];
306
+ let totalInputTokens = checkpoint?.usage.inputTokens ?? 0;
307
+ let totalOutputTokens = checkpoint?.usage.outputTokens ?? 0;
308
+ let cachedInputTokens = checkpoint?.usage.cachedInputTokens ?? 0;
309
+ let cacheWriteTokens = checkpoint?.usage.cacheWriteTokens ?? 0;
310
+ let reasoningTokens = checkpoint?.usage.reasoningTokens ?? 0;
311
+ const reasoningParts = [...(checkpoint?.reasoning ?? [])];
312
+ let reasoning = options.reasoning ?? active.config.reasoning;
313
+ const promptCache = this.config.llm?.promptCache ?? {};
314
+ let iterations = checkpoint?.iterations ?? 0;
229
315
  const maxIterations = agent.config?.maxIterations ?? 10;
230
- let lastToolCallSig = '';
231
- const responseFormat = toLLMResponseFormat(agent.config.responseFormat);
316
+ let lastToolCallSig = checkpoint?.lastToolCallSignature ?? '';
317
+ let pausedTurn;
318
+ let responseFormat = toLLMResponseFormat(active.config.responseFormat);
232
319
  const allReflections = [];
233
320
  const allActions = [];
234
321
  const agentContext = {
@@ -241,10 +328,189 @@ export class Cogitator {
241
328
  previousActions: [],
242
329
  availableTools: registry.getNames(),
243
330
  };
244
- if (this.state.reflectionEngine && this.config.reflection?.enabled) {
331
+ if (!checkpoint && this.state.reflectionEngine && this.config.reflection?.enabled) {
245
332
  await enrichMessagesWithInsights(messages, this.state.reflectionEngine, agentContext);
246
333
  }
247
- while (iterations < maxIterations) {
334
+ const switchTo = (target, reason) => {
335
+ const event = {
336
+ from: active.name,
337
+ to: target.name,
338
+ ...(typeof reason === 'string' && reason && { reason }),
339
+ };
340
+ handoffs.push(event);
341
+ options.onHandoff?.(event);
342
+ const system = messages[0];
343
+ const content = system?.role === 'system' &&
344
+ typeof system.content === 'string' &&
345
+ system.content.startsWith(active.instructions)
346
+ ? target.instructions + system.content.slice(active.instructions.length)
347
+ : target.instructions;
348
+ if (system?.role === 'system')
349
+ messages[0] = { role: 'system', content };
350
+ else
351
+ messages.unshift({ role: 'system', content });
352
+ spans.push(createSpan('agent.handoff', traceId, rootSpanId, Date.now(), Date.now(), { 'handoff.from': event.from, 'handoff.to': event.to }, 'ok', 'internal', options.onSpan));
353
+ active = target;
354
+ ({ registry, targets: handoffTargets } = buildRegistry(target));
355
+ effectiveModel = this.resolveModel(target);
356
+ routeProvider = target.config.provider;
357
+ ({ backend, model } = this.route(effectiveModel, routeProvider));
358
+ reasoning = options.reasoning ?? target.config.reasoning;
359
+ responseFormat = toLLMResponseFormat(target.config.responseFormat);
360
+ lastToolCallSig = '';
361
+ };
362
+ const handleToolTurn = async (toolCalls, resumed) => {
363
+ if (!resumed) {
364
+ const currentSig = toolCalls
365
+ .map((tc) => `${tc.name}:${JSON.stringify(tc.arguments)}`)
366
+ .join('|');
367
+ if (currentSig === lastToolCallSig) {
368
+ for (const tc of toolCalls) {
369
+ const errorResult = {
370
+ callId: tc.id,
371
+ name: tc.name,
372
+ result: null,
373
+ error: 'Duplicate tool call detected. Try a different approach.',
374
+ };
375
+ const duplicateMessage = createToolMessage(tc, errorResult);
376
+ messages.push(duplicateMessage);
377
+ if (this.state.memoryAdapter &&
378
+ options.saveHistory !== false &&
379
+ options.useMemory !== false) {
380
+ await saveEntry(threadId, active.id, duplicateMessage, this.state.memoryAdapter, undefined, [errorResult], options.onMemoryError, options.userId);
381
+ }
382
+ }
383
+ lastToolCallSig = '';
384
+ return undefined;
385
+ }
386
+ lastToolCallSig = currentSig;
387
+ for (const toolCall of toolCalls) {
388
+ allToolCalls.push(toolCall);
389
+ options.onToolCall?.(toolCall);
390
+ }
391
+ }
392
+ const decisions = new Map(Object.entries(resumed?.decisions ?? {}));
393
+ const pending = [];
394
+ for (const toolCall of toolCalls) {
395
+ const tool = registry.get(toolCall.name);
396
+ if (!tool || decisions.has(toolCall.id) || !needsApproval(tool, toolCall.arguments)) {
397
+ continue;
398
+ }
399
+ const request = {
400
+ toolCallId: toolCall.id,
401
+ toolName: toolCall.name,
402
+ arguments: toolCall.arguments,
403
+ description: tool.description,
404
+ ...(tool.sideEffects && { sideEffects: [...tool.sideEffects] }),
405
+ };
406
+ const decision = resumed?.fallback ?? (await this.decideApproval(request, options));
407
+ if (decision === 'pause')
408
+ pending.push(request);
409
+ else
410
+ decisions.set(toolCall.id, decision);
411
+ }
412
+ if (pending.length > 0) {
413
+ return { toolCalls, decisions: Object.fromEntries(decisions), pending };
414
+ }
415
+ if (resumed)
416
+ await this.runCheckpointStore().delete(threadId);
417
+ const executeToolCall = async (toolCall) => {
418
+ const toolSpanStart = Date.now();
419
+ const decision = decisions.get(toolCall.id);
420
+ if (decision?.approved === false) {
421
+ const declined = {
422
+ callId: toolCall.id,
423
+ name: toolCall.name,
424
+ result: null,
425
+ error: `The user declined this tool call${decision.reason ? `: ${decision.reason}` : ''}`,
426
+ };
427
+ return { toolCall, result: declined, toolSpanStart, toolSpanEnd: Date.now() };
428
+ }
429
+ const result = await waitForAbortable(executeTool(registry, toolCall, runId, active.id, this.state.sandboxManager, this.state.constitutionalAI, this.state.constitutionalAI?.config.filterToolCalls ?? false, () => initializeSandbox(this.config, this.state), abortController.signal, {
430
+ threadId,
431
+ userId: options.userId,
432
+ channelType: options.channelType,
433
+ channelId: options.channelId,
434
+ }, decision?.approved === true), abortController.signal);
435
+ const toolSpanEnd = Date.now();
436
+ return { toolCall, result, toolSpanStart, toolSpanEnd };
437
+ };
438
+ const toolResults = options.parallelToolCalls
439
+ ? await Promise.all(toolCalls.map(executeToolCall))
440
+ : await (async () => {
441
+ const results = [];
442
+ for (const toolCall of toolCalls) {
443
+ results.push(await executeToolCall(toolCall));
444
+ }
445
+ return results;
446
+ })();
447
+ const reflectionMessages = [];
448
+ for (const { toolCall, result, toolSpanStart, toolSpanEnd } of toolResults) {
449
+ const toolSpan = createSpan(`tool.${toolCall.name}`, traceId, rootSpanId, toolSpanStart, toolSpanEnd, {
450
+ 'tool.name': toolCall.name,
451
+ 'tool.call_id': toolCall.id,
452
+ 'tool.arguments': JSON.stringify(toolCall.arguments),
453
+ 'tool.success': !result.error,
454
+ 'tool.error': result.error,
455
+ }, result.error ? 'error' : 'ok', 'internal', options.onSpan);
456
+ spans.push(toolSpan);
457
+ options.onToolResult?.(result);
458
+ const toolMessage = createToolMessage(toolCall, result);
459
+ messages.push(toolMessage);
460
+ if (this.state.memoryAdapter &&
461
+ options.saveHistory !== false &&
462
+ options.useMemory !== false) {
463
+ await saveEntry(threadId, active.id, toolMessage, this.state.memoryAdapter, undefined, [result], options.onMemoryError, options.userId);
464
+ }
465
+ const action = {
466
+ type: 'tool_call',
467
+ toolName: toolCall.name,
468
+ input: toolCall.arguments,
469
+ output: result.result,
470
+ error: result.error,
471
+ duration: toolSpanEnd - toolSpanStart,
472
+ };
473
+ allActions.push(action);
474
+ const reflection = this.config.reflection;
475
+ const reflectOnFailure = action.error !== undefined && !!reflection?.reflectAfterError;
476
+ if (this.state.reflectionEngine &&
477
+ reflection?.enabled &&
478
+ (reflectOnFailure || reflection.reflectAfterToolCall)) {
479
+ try {
480
+ const reflectionResult = reflectOnFailure
481
+ ? await this.state.reflectionEngine.reflectOnError(action, agentContext)
482
+ : await this.state.reflectionEngine.reflectOnToolCall(action, agentContext);
483
+ allReflections.push(reflectionResult.reflection);
484
+ if (reflectionResult.shouldAdjustStrategy && reflectionResult.suggestedAction) {
485
+ reflectionMessages.push({
486
+ role: 'system',
487
+ content: `Reflection: ${reflectionResult.reflection.analysis.reasoning}. Consider: ${reflectionResult.suggestedAction}`,
488
+ });
489
+ }
490
+ }
491
+ catch (reflectionError) {
492
+ getLogger().warn('Reflection failed', {
493
+ error: reflectionError instanceof Error
494
+ ? reflectionError.message
495
+ : String(reflectionError),
496
+ });
497
+ }
498
+ }
499
+ }
500
+ messages.push(...reflectionMessages);
501
+ const handoff = toolCalls.find((call) => handoffTargets.has(call.name));
502
+ const target = handoff && handoffTargets.get(handoff.name);
503
+ if (handoff && target)
504
+ switchTo(target, handoff.arguments.reason);
505
+ return undefined;
506
+ };
507
+ if (checkpoint && resumeFrom) {
508
+ pausedTurn = await handleToolTurn(checkpoint.turn.toolCalls, {
509
+ decisions: { ...checkpoint.turn.decisions, ...resumeFrom.decisions },
510
+ fallback: resumeFrom.defaultDecision,
511
+ });
512
+ }
513
+ while (!pausedTurn && iterations < maxIterations) {
248
514
  throwIfAborted(abortController.signal);
249
515
  this.assertTokenBudget(totalInputTokens + totalOutputTokens);
250
516
  if (this.state.contextManager?.shouldCompress(messages, effectiveModel)) {
@@ -258,18 +524,20 @@ export class Cogitator {
258
524
  const llmSpanStart = Date.now();
259
525
  let response;
260
526
  if (options.stream && options.onToken) {
261
- response = await waitForAbortable(streamChat(backend, model, messages, registry, agent, options.onToken, abortController.signal, responseFormat), abortController.signal);
527
+ response = await waitForAbortable(streamChat(backend, model, messages, registry, active, options.onToken, abortController.signal, responseFormat, { reasoning, cache: promptCache, onReasoning: options.onReasoning }), abortController.signal);
262
528
  }
263
529
  else {
264
530
  response = await waitForAbortable(backend.chat({
265
531
  model,
266
532
  messages,
267
533
  tools: registry.getSchemas(),
268
- temperature: agent.config.temperature,
269
- topP: agent.config.topP,
270
- maxTokens: agent.config.maxTokens,
271
- stop: agent.config.stopSequences,
534
+ temperature: active.config.temperature,
535
+ topP: active.config.topP,
536
+ maxTokens: active.config.maxTokens,
537
+ stop: active.config.stopSequences,
272
538
  responseFormat,
539
+ reasoning,
540
+ cache: promptCache,
273
541
  signal: abortController.signal,
274
542
  }), abortController.signal);
275
543
  }
@@ -278,11 +546,22 @@ export class Cogitator {
278
546
  'llm.iteration': iterations,
279
547
  'llm.input_tokens': response.usage.inputTokens,
280
548
  'llm.output_tokens': response.usage.outputTokens,
549
+ ...(response.usage.cachedInputTokens && {
550
+ 'llm.cached_input_tokens': response.usage.cachedInputTokens,
551
+ }),
552
+ ...(response.usage.reasoningTokens && {
553
+ 'llm.reasoning_tokens': response.usage.reasoningTokens,
554
+ }),
281
555
  'llm.finish_reason': response.finishReason,
282
556
  }, 'ok', 'client', options.onSpan);
283
557
  spans.push(llmSpan);
284
558
  totalInputTokens += response.usage.inputTokens;
285
559
  totalOutputTokens += response.usage.outputTokens;
560
+ cachedInputTokens += response.usage.cachedInputTokens ?? 0;
561
+ cacheWriteTokens += response.usage.cacheWriteTokens ?? 0;
562
+ reasoningTokens += response.usage.reasoningTokens ?? 0;
563
+ if (response.reasoning)
564
+ reasoningParts.push(response.reasoning);
286
565
  let outputContent = response.content;
287
566
  if (this.state.constitutionalAI?.config.filterOutput) {
288
567
  const outputResult = await this.state.constitutionalAI.filterOutput(outputContent, messages);
@@ -306,122 +585,81 @@ export class Cogitator {
306
585
  if (this.state.memoryAdapter &&
307
586
  options.saveHistory !== false &&
308
587
  options.useMemory !== false) {
309
- await saveEntry(threadId, agent.id, assistantMessage, this.state.memoryAdapter, response.toolCalls, undefined, options.onMemoryError);
588
+ await saveEntry(threadId, active.id, assistantMessage, this.state.memoryAdapter, response.toolCalls, undefined, options.onMemoryError, options.userId);
310
589
  }
311
590
  if (response.finishReason === 'tool_calls' &&
312
591
  response.toolCalls &&
313
592
  response.toolCalls.length > 0) {
314
- let toolCalls = response.toolCalls;
315
- const currentSig = toolCalls
316
- .map((tc) => `${tc.name}:${JSON.stringify(tc.arguments)}`)
317
- .join('|');
318
- if (currentSig === lastToolCallSig) {
319
- for (const tc of toolCalls) {
320
- const errorResult = {
321
- callId: tc.id,
322
- name: tc.name,
323
- result: null,
324
- error: 'Duplicate tool call detected. Try a different approach.',
325
- };
326
- const duplicateMessage = createToolMessage(tc, errorResult);
327
- messages.push(duplicateMessage);
328
- if (this.state.memoryAdapter &&
329
- options.saveHistory !== false &&
330
- options.useMemory !== false) {
331
- await saveEntry(threadId, agent.id, duplicateMessage, this.state.memoryAdapter, undefined, [errorResult], options.onMemoryError);
332
- }
333
- }
334
- lastToolCallSig = '';
335
- continue;
336
- }
337
- lastToolCallSig = currentSig;
338
- for (const toolCall of toolCalls) {
339
- allToolCalls.push(toolCall);
340
- options.onToolCall?.(toolCall);
341
- }
342
- const executeToolCall = async (toolCall) => {
343
- const toolSpanStart = Date.now();
344
- const result = await waitForAbortable(executeTool(registry, toolCall, runId, agent.id, this.state.sandboxManager, this.state.constitutionalAI, this.state.constitutionalAI?.config.filterToolCalls ?? false, () => initializeSandbox(this.config, this.state), abortController.signal, {
345
- threadId,
346
- userId: options.userId,
347
- channelType: options.channelType,
348
- channelId: options.channelId,
349
- }), abortController.signal);
350
- const toolSpanEnd = Date.now();
351
- return { toolCall, result, toolSpanStart, toolSpanEnd };
352
- };
353
- const toolResults = options.parallelToolCalls
354
- ? await Promise.all(toolCalls.map(executeToolCall))
355
- : await (async () => {
356
- const results = [];
357
- for (const toolCall of toolCalls) {
358
- results.push(await executeToolCall(toolCall));
359
- }
360
- return results;
361
- })();
362
- const reflectionMessages = [];
363
- for (const { toolCall, result, toolSpanStart, toolSpanEnd } of toolResults) {
364
- const toolSpan = createSpan(`tool.${toolCall.name}`, traceId, rootSpanId, toolSpanStart, toolSpanEnd, {
365
- 'tool.name': toolCall.name,
366
- 'tool.call_id': toolCall.id,
367
- 'tool.arguments': JSON.stringify(toolCall.arguments),
368
- 'tool.success': !result.error,
369
- 'tool.error': result.error,
370
- }, result.error ? 'error' : 'ok', 'internal', options.onSpan);
371
- spans.push(toolSpan);
372
- options.onToolResult?.(result);
373
- const toolMessage = createToolMessage(toolCall, result);
374
- messages.push(toolMessage);
375
- if (this.state.memoryAdapter &&
376
- options.saveHistory !== false &&
377
- options.useMemory !== false) {
378
- await saveEntry(threadId, agent.id, toolMessage, this.state.memoryAdapter, undefined, [result], options.onMemoryError);
379
- }
380
- const action = {
381
- type: 'tool_call',
382
- toolName: toolCall.name,
383
- input: toolCall.arguments,
384
- output: result.result,
385
- error: result.error,
386
- duration: toolSpanEnd - toolSpanStart,
387
- };
388
- allActions.push(action);
389
- const reflection = this.config.reflection;
390
- const reflectOnFailure = action.error !== undefined && !!reflection?.reflectAfterError;
391
- if (this.state.reflectionEngine &&
392
- reflection?.enabled &&
393
- (reflectOnFailure || reflection.reflectAfterToolCall)) {
394
- try {
395
- const reflectionResult = reflectOnFailure
396
- ? await this.state.reflectionEngine.reflectOnError(action, agentContext)
397
- : await this.state.reflectionEngine.reflectOnToolCall(action, agentContext);
398
- allReflections.push(reflectionResult.reflection);
399
- if (reflectionResult.shouldAdjustStrategy && reflectionResult.suggestedAction) {
400
- reflectionMessages.push({
401
- role: 'system',
402
- content: `Reflection: ${reflectionResult.reflection.analysis.reasoning}. Consider: ${reflectionResult.suggestedAction}`,
403
- });
404
- }
405
- }
406
- catch (reflectionError) {
407
- getLogger().warn('Reflection failed', {
408
- error: reflectionError instanceof Error
409
- ? reflectionError.message
410
- : String(reflectionError),
411
- });
412
- }
413
- }
414
- }
415
- messages.push(...reflectionMessages);
593
+ pausedTurn = await handleToolTurn(response.toolCalls);
594
+ if (pausedTurn)
595
+ break;
416
596
  }
417
597
  else {
418
598
  break;
419
599
  }
420
600
  }
601
+ if (pausedTurn) {
602
+ const pausedCheckpoint = {
603
+ version: 1,
604
+ runId,
605
+ agentId: active.id,
606
+ threadId,
607
+ ...(options.userId !== undefined && { userId: options.userId }),
608
+ model: effectiveModel,
609
+ ...(routeProvider !== undefined && { provider: routeProvider }),
610
+ input,
611
+ messages,
612
+ toolCalls: allToolCalls,
613
+ ...(prompt && { prompt }),
614
+ ...(active !== agent && { activeAgent: active.name }),
615
+ ...(handoffs.length > 0 && { handoffs }),
616
+ turn: { toolCalls: pausedTurn.toolCalls, decisions: pausedTurn.decisions },
617
+ iterations,
618
+ lastToolCallSignature: lastToolCallSig,
619
+ usage: {
620
+ inputTokens: totalInputTokens,
621
+ outputTokens: totalOutputTokens,
622
+ cachedInputTokens,
623
+ cacheWriteTokens,
624
+ reasoningTokens,
625
+ },
626
+ reasoning: reasoningParts,
627
+ startedAt: checkpoint?.startedAt ?? startTime,
628
+ };
629
+ const paused = {
630
+ output: getTextContent(messages[messages.length - 1]?.content ?? ''),
631
+ runId,
632
+ agentId: agent.id,
633
+ threadId,
634
+ status: 'paused',
635
+ pendingApprovals: pausedTurn.pending,
636
+ ...(handoffs.length > 0 && { handoffs, finalAgent: active.name }),
637
+ checkpoint: pausedCheckpoint,
638
+ usage: {
639
+ inputTokens: totalInputTokens,
640
+ outputTokens: totalOutputTokens,
641
+ totalTokens: totalInputTokens + totalOutputTokens,
642
+ cost: this.calculateCost(effectiveModel, {
643
+ inputTokens: totalInputTokens,
644
+ outputTokens: totalOutputTokens,
645
+ cachedInputTokens,
646
+ cacheWriteTokens,
647
+ }),
648
+ duration: Date.now() - startTime,
649
+ },
650
+ ...(reasoningParts.length > 0 && { reasoning: reasoningParts.join('\n\n') }),
651
+ toolCalls: allToolCalls,
652
+ messages,
653
+ trace: { traceId, spans },
654
+ };
655
+ await this.runCheckpointStore().save(pausedCheckpoint);
656
+ options.onRunComplete?.(paused);
657
+ return paused;
658
+ }
421
659
  const endTime = Date.now();
422
660
  const lastAssistantMessage = messages.filter((m) => m.role === 'assistant').pop();
423
661
  const finalOutput = lastAssistantMessage ? getTextContent(lastAssistantMessage.content) : '';
424
- const structured = parseStructuredOutput(agent.config.responseFormat, finalOutput);
662
+ const structured = parseStructuredOutput(active.config.responseFormat, finalOutput);
425
663
  if (this.state.reflectionEngine &&
426
664
  this.config.reflection?.enabled &&
427
665
  this.config.reflection.reflectAtEnd) {
@@ -447,7 +685,12 @@ export class Cogitator {
447
685
  'run.output_tokens': totalOutputTokens,
448
686
  }, 'ok', 'server', options.onSpan, rootSpanId);
449
687
  spans.unshift(rootSpan);
450
- const runCost = this.calculateCost(effectiveModel, totalInputTokens, totalOutputTokens);
688
+ const runCost = this.calculateCost(effectiveModel, {
689
+ inputTokens: totalInputTokens,
690
+ outputTokens: totalOutputTokens,
691
+ cachedInputTokens,
692
+ cacheWriteTokens,
693
+ });
451
694
  if (this.state.costRouter) {
452
695
  this.state.costRouter.recordCost({
453
696
  runId,
@@ -462,6 +705,9 @@ export class Cogitator {
462
705
  const result = {
463
706
  output: finalOutput,
464
707
  ...(structured !== undefined && { structured }),
708
+ status: 'completed',
709
+ ...(prompt && { prompt }),
710
+ ...(handoffs.length > 0 && { handoffs, finalAgent: active.name }),
465
711
  runId,
466
712
  agentId: agent.id,
467
713
  threadId,
@@ -472,7 +718,11 @@ export class Cogitator {
472
718
  totalTokens: totalInputTokens + totalOutputTokens,
473
719
  cost: runCost,
474
720
  duration: endTime - startTime,
721
+ ...(reasoningTokens > 0 && { reasoningTokens }),
722
+ ...(cachedInputTokens > 0 && { cachedInputTokens }),
723
+ ...(cacheWriteTokens > 0 && { cacheWriteTokens }),
475
724
  },
725
+ ...(reasoningParts.length > 0 && { reasoning: reasoningParts.join('\n\n') }),
476
726
  toolCalls: allToolCalls,
477
727
  messages,
478
728
  trace: {
@@ -484,10 +734,14 @@ export class Cogitator {
484
734
  ? await this.state.reflectionEngine.getSummary(agent.id)
485
735
  : undefined,
486
736
  };
737
+ if (prompt)
738
+ await this.recordPrompt(prompt, result, Date.now() - startTime);
487
739
  options.onRunComplete?.(result);
488
740
  return result;
489
741
  }
490
742
  catch (error) {
743
+ if (prompt)
744
+ await this.recordPrompt(prompt, undefined, Date.now() - startTime);
491
745
  const endTime = Date.now();
492
746
  const errorSpan = createSpan('agent.run', traceId, undefined, startTime, endTime, {
493
747
  'agent.id': agent.id,
@@ -508,6 +762,129 @@ export class Cogitator {
508
762
  removeParentAbortListener?.();
509
763
  }
510
764
  }
765
+ /**
766
+ * The messages a new run starts with: the thread's history and the input,
767
+ * once the input passed the injection and constitutional checks, with the
768
+ * input saved to memory.
769
+ */
770
+ async prepareMessages(agent, options, input, threadId) {
771
+ const messages = await buildInitialMessages(agent, options, threadId, this.state.memoryAdapter, this.state.contextBuilder);
772
+ const pii = this.config.security?.pii;
773
+ if (pii?.mode === 'block') {
774
+ const found = new PiiMasker(pii).find(input);
775
+ if (found.length > 0) {
776
+ throw new CogitatorError({
777
+ message: `The input contains personal data: ${[...new Set(found.map((f) => f.type))].join(', ')}`,
778
+ code: ErrorCode.PII_DETECTED,
779
+ details: { types: [...new Set(found.map((f) => f.type))] },
780
+ });
781
+ }
782
+ }
783
+ if (this.state.injectionDetector) {
784
+ const injectionResult = await this.state.injectionDetector.analyze(input);
785
+ if (injectionResult.action === 'blocked') {
786
+ const threatTypes = injectionResult.threats.map((t) => t.type).join(', ');
787
+ throw new CogitatorError({
788
+ message: `Prompt injection detected: ${threatTypes}`,
789
+ code: ErrorCode.PROMPT_INJECTION_DETECTED,
790
+ details: { threats: injectionResult.threats },
791
+ });
792
+ }
793
+ }
794
+ if (this.state.constitutionalAI?.config.filterInput) {
795
+ const inputResult = await this.state.constitutionalAI.filterInput(input);
796
+ if (!inputResult.allowed) {
797
+ throw new Error(`Input blocked: ${inputResult.blockedReason ?? 'Policy violation'}`);
798
+ }
799
+ }
800
+ if (options.context) {
801
+ addContextToMessages(messages, options.context);
802
+ }
803
+ if (this.state.memoryAdapter && options.saveHistory !== false && options.useMemory !== false) {
804
+ const currentUserMessage = messages[messages.length - 1];
805
+ await saveEntry(threadId, agent.id, currentUserMessage, this.state.memoryAdapter, undefined, undefined, options.onMemoryError, options.userId);
806
+ }
807
+ return messages;
808
+ }
809
+ /**
810
+ * Versioned instructions and A/B tests of this Cogitator's agents
811
+ * (`prompts` config, in process memory by default).
812
+ */
813
+ get prompts() {
814
+ this.promptRegistry ??= new PromptRegistry(this.config.prompts);
815
+ return this.promptRegistry;
816
+ }
817
+ async recordPrompt(prompt, result, durationMs) {
818
+ try {
819
+ await this.prompts.record(prompt, { result, durationMs });
820
+ }
821
+ catch (error) {
822
+ getLogger().warn('Could not record the outcome of a run against its instructions', {
823
+ prompt: prompt.key,
824
+ error: error instanceof Error ? error.message : String(error),
825
+ });
826
+ }
827
+ }
828
+ runCheckpointStore() {
829
+ if (this.config.runCheckpoints)
830
+ return this.config.runCheckpoints;
831
+ const memory = this.state.memoryAdapter;
832
+ if (memory) {
833
+ if (this.threadCheckpoints?.memory !== memory) {
834
+ this.threadCheckpoints = { memory, store: new ThreadRunCheckpointStore(memory) };
835
+ }
836
+ return this.threadCheckpoints.store;
837
+ }
838
+ this.processCheckpoints ??= new InMemoryRunCheckpointStore();
839
+ return this.processCheckpoints;
840
+ }
841
+ /**
842
+ * A new run on a thread whose run is waiting for approvals means the user
843
+ * moved on: the waiting calls are answered as declined, so the thread's
844
+ * history stays whole, and the pause is dropped.
845
+ */
846
+ async abandonPausedRun(agent, options, threadId) {
847
+ if (options.threadId === undefined)
848
+ return;
849
+ const store = this.runCheckpointStore();
850
+ const paused = await store.load(threadId).catch((error) => {
851
+ getLogger().warn('Could not read the paused run of a thread', {
852
+ threadId,
853
+ error: error instanceof Error ? error.message : String(error),
854
+ });
855
+ return null;
856
+ });
857
+ if (!paused)
858
+ return;
859
+ if (options.threadAccess !== 'shared' && paused.userId !== options.userId)
860
+ return;
861
+ if (this.state.memoryAdapter && options.useMemory !== false && options.saveHistory !== false) {
862
+ for (const toolCall of paused.turn.toolCalls) {
863
+ const result = {
864
+ callId: toolCall.id,
865
+ name: toolCall.name,
866
+ result: null,
867
+ error: 'The user moved on without approving this tool call',
868
+ };
869
+ await saveEntry(threadId, agent.id, createToolMessage(toolCall, result), this.state.memoryAdapter, undefined, [result], options.onMemoryError, options.userId);
870
+ }
871
+ }
872
+ await store.delete(threadId);
873
+ }
874
+ /**
875
+ * The decision for a call that needs approval: the run's `onApproval`, else
876
+ * `guardrails.onToolApproval`, else a pause.
877
+ */
878
+ async decideApproval(request, options) {
879
+ if (options.onApproval)
880
+ return options.onApproval(request);
881
+ const legacy = this.config.guardrails?.onToolApproval;
882
+ if (legacy) {
883
+ const approved = await legacy(request.toolName, request.arguments, request.sideEffects ?? []);
884
+ return approved ? { approved: true } : { approved: false };
885
+ }
886
+ return 'pause';
887
+ }
511
888
  async acquireRunSlot(signal) {
512
889
  const max = this.config.limits?.maxConcurrentRuns;
513
890
  if (max === undefined)
@@ -602,35 +979,30 @@ export class Cogitator {
602
979
  hasLLMPlugin(name));
603
980
  }
604
981
  backendFor(name) {
605
- const custom = this.config.llm?.backends;
606
- if (custom && Object.hasOwn(custom, name))
607
- return custom[name];
608
982
  const cached = this.backends.get(name);
609
983
  if (cached)
610
984
  return cached;
611
- let backend;
612
- if (isLLMProvider(name)) {
613
- backend = createLLMBackend(name, this.config.llm);
614
- }
615
- else if (hasLLMPlugin(name)) {
616
- backend = createLLMBackendFromPlugin(name, this.config.llm?.plugins?.[name]);
617
- }
618
- else {
619
- throw new CogitatorError({
620
- message: `Unknown LLM provider "${name}": not a built-in provider, a backend in llm.backends or a registered plugin`,
621
- code: ErrorCode.CONFIGURATION_ERROR,
622
- });
623
- }
985
+ const backend = withPiiMasking(withLLMRetry(this.createBackend(name), this.config.llm?.retry), this.config.security?.pii);
624
986
  this.backends.set(name, backend);
625
987
  return backend;
626
988
  }
627
- calculateCost(model, inputTokens, outputTokens) {
628
- const { model: modelName } = parseModel(model);
629
- const price = getPrice(modelName);
630
- if (!price) {
631
- return 0;
989
+ createBackend(name) {
990
+ const custom = this.config.llm?.backends;
991
+ if (custom && Object.hasOwn(custom, name))
992
+ return custom[name];
993
+ if (isLLMProvider(name))
994
+ return createLLMBackend(name, this.config.llm);
995
+ if (hasLLMPlugin(name)) {
996
+ return createLLMBackendFromPlugin(name, this.config.llm?.plugins?.[name]);
632
997
  }
633
- return (inputTokens * price.input + outputTokens * price.output) / 1_000_000;
998
+ throw new CogitatorError({
999
+ message: `Unknown LLM provider "${name}": not a built-in provider, a backend in llm.backends or a registered plugin`,
1000
+ code: ErrorCode.CONFIGURATION_ERROR,
1001
+ });
1002
+ }
1003
+ calculateCost(model, usage) {
1004
+ const { model: modelName } = parseModel(model);
1005
+ return calculateModelCost(modelName, usage) ?? 0;
634
1006
  }
635
1007
  /**
636
1008
  * Get accumulated insights from reflection for an agent.
@@ -803,4 +1175,16 @@ function waitForAbortable(promise, signal) {
803
1175
  });
804
1176
  });
805
1177
  }
1178
+ /** Whether a call to `tool` with `args` needs approval; a check that throws counts as yes. */
1179
+ function needsApproval(tool, args) {
1180
+ const check = tool.requiresApproval;
1181
+ if (typeof check !== 'function')
1182
+ return check === true;
1183
+ try {
1184
+ return check(args);
1185
+ }
1186
+ catch {
1187
+ return true;
1188
+ }
1189
+ }
806
1190
  //# sourceMappingURL=runtime.js.map