@openai/agents-core 0.2.0 → 0.3.0

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 (49) hide show
  1. package/dist/index.d.ts +3 -0
  2. package/dist/index.js +3 -1
  3. package/dist/index.js.map +1 -1
  4. package/dist/index.mjs +1 -0
  5. package/dist/index.mjs.map +1 -1
  6. package/dist/memory/memorySession.d.ts +22 -0
  7. package/dist/memory/memorySession.js +64 -0
  8. package/dist/memory/memorySession.js.map +1 -0
  9. package/dist/memory/memorySession.mjs +60 -0
  10. package/dist/memory/memorySession.mjs.map +1 -0
  11. package/dist/memory/session.d.ts +36 -0
  12. package/dist/memory/session.js +3 -0
  13. package/dist/memory/session.js.map +1 -0
  14. package/dist/memory/session.mjs +2 -0
  15. package/dist/memory/session.mjs.map +1 -0
  16. package/dist/metadata.js +3 -3
  17. package/dist/metadata.js.map +1 -1
  18. package/dist/metadata.mjs +3 -3
  19. package/dist/metadata.mjs.map +1 -1
  20. package/dist/run.d.ts +88 -8
  21. package/dist/run.js +859 -347
  22. package/dist/run.js.map +1 -1
  23. package/dist/run.mjs +859 -347
  24. package/dist/run.mjs.map +1 -1
  25. package/dist/runImplementation.d.ts +21 -2
  26. package/dist/runImplementation.js +885 -346
  27. package/dist/runImplementation.js.map +1 -1
  28. package/dist/runImplementation.mjs +877 -344
  29. package/dist/runImplementation.mjs.map +1 -1
  30. package/dist/runState.d.ts +1360 -226
  31. package/dist/runState.js +16 -0
  32. package/dist/runState.js.map +1 -1
  33. package/dist/runState.mjs +16 -0
  34. package/dist/runState.mjs.map +1 -1
  35. package/dist/types/protocol.d.ts +1869 -193
  36. package/dist/types/protocol.js +1 -0
  37. package/dist/types/protocol.js.map +1 -1
  38. package/dist/types/protocol.mjs +1 -0
  39. package/dist/types/protocol.mjs.map +1 -1
  40. package/dist/utils/smartString.d.ts +9 -0
  41. package/dist/utils/smartString.js +15 -0
  42. package/dist/utils/smartString.js.map +1 -1
  43. package/dist/utils/smartString.mjs +14 -3
  44. package/dist/utils/smartString.mjs.map +1 -1
  45. package/dist/utils/tools.js +59 -20
  46. package/dist/utils/tools.js.map +1 -1
  47. package/dist/utils/tools.mjs +59 -20
  48. package/dist/utils/tools.mjs.map +1 -1
  49. package/package.json +2 -2
@@ -7,14 +7,59 @@ import { getLastTextFromOutputMessage } from "./utils/messages.mjs";
7
7
  import { withFunctionSpan, withHandoffSpan } from "./tracing/createSpans.mjs";
8
8
  import { getSchemaAndParserFromInputType } from "./utils/tools.mjs";
9
9
  import { encodeUint8ArrayToBase64 } from "./utils/base64.mjs";
10
+ import { isArrayBufferView, isNodeBuffer, isSerializedBufferSnapshot, toSmartString, } from "./utils/smartString.mjs";
10
11
  import { safeExecute } from "./utils/safeExecute.mjs";
11
12
  import { addErrorToCurrentSpan } from "./tracing/context.mjs";
12
13
  import { RunItemStreamEvent } from "./events.mjs";
13
14
  import { z } from 'zod';
14
- import { toSmartString } from "./utils/smartString.mjs";
15
15
  import { isZodObject } from "./utils/index.mjs";
16
+ function isApprovalItemLike(value) {
17
+ if (!value || typeof value !== 'object') {
18
+ return false;
19
+ }
20
+ if (!('rawItem' in value)) {
21
+ return false;
22
+ }
23
+ const rawItem = value.rawItem;
24
+ if (!rawItem || typeof rawItem !== 'object') {
25
+ return false;
26
+ }
27
+ const itemType = rawItem.type;
28
+ return itemType === 'function_call' || itemType === 'hosted_tool_call';
29
+ }
30
+ function getApprovalIdentity(approval) {
31
+ const rawItem = approval.rawItem;
32
+ if (!rawItem) {
33
+ return undefined;
34
+ }
35
+ if (rawItem.type === 'function_call' && rawItem.callId) {
36
+ return `function_call:${rawItem.callId}`;
37
+ }
38
+ if ('callId' in rawItem && rawItem.callId) {
39
+ return `${rawItem.type}:${rawItem.callId}`;
40
+ }
41
+ const id = 'id' in rawItem ? rawItem.id : undefined;
42
+ if (id) {
43
+ return `${rawItem.type}:${id}`;
44
+ }
45
+ const providerData = typeof rawItem.providerData === 'object' && rawItem.providerData
46
+ ? rawItem.providerData
47
+ : undefined;
48
+ if (providerData?.id) {
49
+ return `${rawItem.type}:provider:${providerData.id}`;
50
+ }
51
+ const agentName = 'agent' in approval && approval.agent ? approval.agent.name : '';
52
+ try {
53
+ return `${agentName}:${rawItem.type}:${JSON.stringify(rawItem)}`;
54
+ }
55
+ catch {
56
+ return `${agentName}:${rawItem.type}`;
57
+ }
58
+ }
16
59
  /**
17
60
  * @internal
61
+ * Walks a raw model response and classifies each item so the runner can schedule follow-up work.
62
+ * Returns both the serializable RunItems (for history/streaming) and the actionable tool metadata.
18
63
  */
19
64
  export function processModelResponse(modelResponse, agent, tools, handoffs) {
20
65
  const items = [];
@@ -24,6 +69,7 @@ export function processModelResponse(modelResponse, agent, tools, handoffs) {
24
69
  const runMCPApprovalRequests = [];
25
70
  const toolsUsed = [];
26
71
  const handoffMap = new Map(handoffs.map((h) => [h.toolName, h]));
72
+ // Resolve tools upfront so we can look up the concrete handler in O(1) while iterating outputs.
27
73
  const functionMap = new Map(tools.filter((t) => t.type === 'function').map((t) => [t.name, t]));
28
74
  const computerTool = tools.find((t) => t.type === 'computer');
29
75
  const mcpToolMap = new Map(tools
@@ -158,6 +204,10 @@ export const nextStepSchema = z.discriminatedUnion('type', [
158
204
  data: z.record(z.string(), z.any()),
159
205
  }),
160
206
  ]);
207
+ /**
208
+ * Internal convenience wrapper that groups the outcome of a single agent turn. It lets the caller
209
+ * update the RunState in one shot and decide which step to execute next.
210
+ */
161
211
  class SingleStepResult {
162
212
  originalInput;
163
213
  modelResponse;
@@ -166,7 +216,7 @@ class SingleStepResult {
166
216
  nextStep;
167
217
  constructor(
168
218
  /**
169
- * The input items i.e. the items before run() was called. May be muted by handoff input filters
219
+ * The input items (i.e., the items before run() was called). May be mutated by handoff input filters.
170
220
  */
171
221
  originalInput,
172
222
  /**
@@ -200,6 +250,8 @@ class SingleStepResult {
200
250
  }
201
251
  /**
202
252
  * @internal
253
+ * Resets the tool choice when the agent is configured to prefer a fresh tool selection after
254
+ * any tool usage. This prevents the provider from reusing stale tool hints across turns.
203
255
  */
204
256
  export function maybeResetToolChoice(agent, toolUseTracker, modelSettings) {
205
257
  if (agent.resetToolChoice && toolUseTracker.hasUsedTools(agent)) {
@@ -209,27 +261,94 @@ export function maybeResetToolChoice(agent, toolUseTracker, modelSettings) {
209
261
  }
210
262
  /**
211
263
  * @internal
264
+ * Continues a turn that was previously interrupted waiting for tool approval. Executes the now
265
+ * approved tools and returns the resulting step transition.
212
266
  */
213
- export async function executeInterruptedToolsAndSideEffects(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
267
+ export async function resolveInterruptedTurn(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
214
268
  // call_ids for function tools
215
269
  const functionCallIds = originalPreStepItems
216
270
  .filter((item) => item instanceof RunToolApprovalItem &&
217
271
  'callId' in item.rawItem &&
218
272
  item.rawItem.type === 'function_call')
219
273
  .map((item) => item.rawItem.callId);
274
+ // We already persisted the turn once when the approval interrupt was raised, so the
275
+ // counter reflects the approval items as "flushed". When we resume the same turn we need
276
+ // to rewind it so the eventual tool output for this call is still written to the session.
277
+ const pendingApprovalItems = state
278
+ .getInterruptions()
279
+ .filter(isApprovalItemLike);
280
+ if (pendingApprovalItems.length > 0) {
281
+ const pendingApprovalIdentities = new Set();
282
+ for (const approval of pendingApprovalItems) {
283
+ const identity = getApprovalIdentity(approval);
284
+ if (identity) {
285
+ pendingApprovalIdentities.add(identity);
286
+ }
287
+ }
288
+ if (pendingApprovalIdentities.size > 0) {
289
+ let rewindCount = 0;
290
+ for (let index = originalPreStepItems.length - 1; index >= 0; index--) {
291
+ const item = originalPreStepItems[index];
292
+ if (!(item instanceof RunToolApprovalItem)) {
293
+ continue;
294
+ }
295
+ const identity = getApprovalIdentity(item);
296
+ if (!identity) {
297
+ continue;
298
+ }
299
+ if (!pendingApprovalIdentities.has(identity)) {
300
+ continue;
301
+ }
302
+ rewindCount++;
303
+ pendingApprovalIdentities.delete(identity);
304
+ if (pendingApprovalIdentities.size === 0) {
305
+ break;
306
+ }
307
+ }
308
+ // Persisting the approval request already advanced the counter once, so undo the increment
309
+ // to make sure we write the final tool output back to the session when the turn resumes.
310
+ if (rewindCount > 0) {
311
+ state._currentTurnPersistedItemCount = Math.max(0, state._currentTurnPersistedItemCount - rewindCount);
312
+ }
313
+ }
314
+ }
220
315
  // Run function tools that require approval after they get their approval results
221
316
  const functionToolRuns = processedResponse.functions.filter((run) => {
222
317
  return functionCallIds.includes(run.toolCall.callId);
223
318
  });
224
319
  const functionResults = await executeFunctionToolCalls(agent, functionToolRuns, runner, state);
225
- // Create the initial set of the output items
226
- const newItems = functionResults.map((r) => r.runItem);
320
+ // There is no built-in HITL approval surface for computer tools today, so every pending action
321
+ // is executed immediately when the turn resumes.
322
+ const computerResults = processedResponse.computerActions.length > 0
323
+ ? await executeComputerActions(agent, processedResponse.computerActions, runner, state._context)
324
+ : [];
325
+ // When resuming we receive the original RunItem references; suppress duplicates so history and streaming do not double-emit the same items.
326
+ const originalPreStepItemSet = new Set(originalPreStepItems);
327
+ const newItems = [];
328
+ const newItemsSet = new Set();
329
+ const appendIfNew = (item) => {
330
+ if (originalPreStepItemSet.has(item) || newItemsSet.has(item)) {
331
+ return;
332
+ }
333
+ newItems.push(item);
334
+ newItemsSet.add(item);
335
+ };
336
+ for (const result of functionResults) {
337
+ appendIfNew(result.runItem);
338
+ }
339
+ for (const result of computerResults) {
340
+ appendIfNew(result);
341
+ }
227
342
  // Run MCP tools that require approval after they get their approval results
228
343
  const mcpApprovalRuns = processedResponse.mcpApprovalRequests.filter((run) => {
229
344
  return (run.requestItem.type === 'tool_approval_item' &&
230
345
  run.requestItem.rawItem.type === 'hosted_tool_call' &&
231
346
  run.requestItem.rawItem.providerData?.type === 'mcp_approval_request');
232
347
  });
348
+ // Hosted MCP approvals may still be waiting on a human decision when the turn resumes.
349
+ const pendingHostedMCPApprovals = new Set();
350
+ const pendingHostedMCPApprovalIds = new Set();
351
+ // Keep track of approvals we still need to surface next turn so HITL flows can resume cleanly.
233
352
  for (const run of mcpApprovalRuns) {
234
353
  // the approval_request_id "mcpr_123..."
235
354
  const approvalRequestId = run.requestItem.rawItem.id;
@@ -245,50 +364,93 @@ export async function executeInterruptedToolsAndSideEffects(agent, originalInput
245
364
  reason: undefined,
246
365
  };
247
366
  // Tell Responses API server the approval result in the next turn
248
- newItems.push(new RunToolCallItem({
367
+ const responseItem = new RunToolCallItem({
249
368
  type: 'hosted_tool_call',
250
369
  name: 'mcp_approval_response',
251
370
  providerData,
252
- }, agent));
371
+ }, agent);
372
+ appendIfNew(responseItem);
373
+ }
374
+ else {
375
+ pendingHostedMCPApprovals.add(run.requestItem);
376
+ pendingHostedMCPApprovalIds.add(approvalRequestId);
377
+ functionResults.push({
378
+ type: 'hosted_mcp_tool_approval',
379
+ tool: run.mcpTool,
380
+ runItem: run.requestItem,
381
+ });
382
+ appendIfNew(run.requestItem);
253
383
  }
254
384
  }
255
- const checkToolOutput = await checkForFinalOutputFromTools(agent, functionResults, state);
256
- // Exclude the tool approval items, which should not be sent to Responses API,
257
- // from the SingleStepResult's preStepItems
385
+ // Server-managed conversations rely on preStepItems to re-surface pending approvals.
386
+ // Keep unresolved hosted MCP approvals in place so HITL flows still have something to approve next turn.
387
+ // Drop resolved approval placeholders so they are not replayed on the next turn, but keep
388
+ // pending approvals in place to signal the outstanding work to the UI and session store.
258
389
  const preStepItems = originalPreStepItems.filter((item) => {
259
- return !(item instanceof RunToolApprovalItem);
390
+ if (!(item instanceof RunToolApprovalItem)) {
391
+ return true;
392
+ }
393
+ if (item.rawItem.type === 'hosted_tool_call' &&
394
+ item.rawItem.providerData?.type === 'mcp_approval_request') {
395
+ if (pendingHostedMCPApprovals.has(item)) {
396
+ return true;
397
+ }
398
+ const approvalRequestId = item.rawItem.id;
399
+ if (approvalRequestId) {
400
+ return pendingHostedMCPApprovalIds.has(approvalRequestId);
401
+ }
402
+ return false;
403
+ }
404
+ return false;
260
405
  });
261
- if (checkToolOutput.isFinalOutput) {
262
- runner.emit('agent_end', state._context, agent, checkToolOutput.finalOutput);
263
- agent.emit('agent_end', state._context, checkToolOutput.finalOutput);
264
- return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
265
- type: 'next_step_final_output',
266
- output: checkToolOutput.finalOutput,
267
- });
268
- }
269
- else if (checkToolOutput.isInterrupted) {
270
- return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
271
- type: 'next_step_interruption',
272
- data: {
273
- interruptions: checkToolOutput.interruptions,
274
- },
275
- });
406
+ const completedStep = await maybeCompleteTurnFromToolResults({
407
+ agent,
408
+ runner,
409
+ state,
410
+ functionResults,
411
+ originalInput,
412
+ newResponse,
413
+ preStepItems,
414
+ newItems,
415
+ });
416
+ if (completedStep) {
417
+ return completedStep;
276
418
  }
277
419
  // we only ran new tools and side effects. We need to run the rest of the agent
278
420
  return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
279
421
  }
280
422
  /**
281
423
  * @internal
424
+ * Executes every follow-up action the model requested (function tools, computer actions, MCP flows),
425
+ * appends their outputs to the run history, and determines the next step for the agent loop.
282
426
  */
283
- export async function executeToolsAndSideEffects(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
427
+ export async function resolveTurnAfterModelResponse(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
428
+ // Reuse the same array reference so we can compare object identity when deciding whether to
429
+ // append new items, ensuring we never double-stream existing RunItems.
284
430
  const preStepItems = originalPreStepItems;
285
- let newItems = processedResponse.newItems;
431
+ const seenItems = new Set(originalPreStepItems);
432
+ const newItems = [];
433
+ const appendIfNew = (item) => {
434
+ if (seenItems.has(item)) {
435
+ return;
436
+ }
437
+ newItems.push(item);
438
+ seenItems.add(item);
439
+ };
440
+ for (const item of processedResponse.newItems) {
441
+ appendIfNew(item);
442
+ }
443
+ // Run function tools and computer actions in parallel; neither depends on the other's side effects.
286
444
  const [functionResults, computerResults] = await Promise.all([
287
445
  executeFunctionToolCalls(agent, processedResponse.functions, runner, state),
288
446
  executeComputerActions(agent, processedResponse.computerActions, runner, state._context),
289
447
  ]);
290
- newItems = newItems.concat(functionResults.map((r) => r.runItem));
291
- newItems = newItems.concat(computerResults);
448
+ for (const result of functionResults) {
449
+ appendIfNew(result.runItem);
450
+ }
451
+ for (const item of computerResults) {
452
+ appendIfNew(item);
453
+ }
292
454
  // run hosted MCP approval requests
293
455
  if (processedResponse.mcpApprovalRequests.length > 0) {
294
456
  for (const approvalRequest of processedResponse.mcpApprovalRequests) {
@@ -334,22 +496,18 @@ export async function executeToolsAndSideEffects(agent, originalInput, originalP
334
496
  if (processedResponse.handoffs.length > 0) {
335
497
  return await executeHandoffCalls(agent, originalInput, preStepItems, newItems, newResponse, processedResponse.handoffs, runner, state._context);
336
498
  }
337
- const checkToolOutput = await checkForFinalOutputFromTools(agent, functionResults, state);
338
- if (checkToolOutput.isFinalOutput) {
339
- runner.emit('agent_end', state._context, agent, checkToolOutput.finalOutput);
340
- agent.emit('agent_end', state._context, checkToolOutput.finalOutput);
341
- return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
342
- type: 'next_step_final_output',
343
- output: checkToolOutput.finalOutput,
344
- });
345
- }
346
- else if (checkToolOutput.isInterrupted) {
347
- return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
348
- type: 'next_step_interruption',
349
- data: {
350
- interruptions: checkToolOutput.interruptions,
351
- },
352
- });
499
+ const completedStep = await maybeCompleteTurnFromToolResults({
500
+ agent,
501
+ runner,
502
+ state,
503
+ functionResults,
504
+ originalInput,
505
+ newResponse,
506
+ preStepItems,
507
+ newItems,
508
+ });
509
+ if (completedStep) {
510
+ return completedStep;
353
511
  }
354
512
  // If the model issued any tool calls or handoffs in this turn,
355
513
  // we must NOT treat any assistant message in the same turn as the final output.
@@ -371,6 +529,7 @@ export async function executeToolsAndSideEffects(agent, originalInput, originalP
371
529
  if (typeof potentialFinalOutput === 'undefined') {
372
530
  return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
373
531
  }
532
+ // Keep looping if any tool output placeholders still require an approval follow-up.
374
533
  const hasPendingToolsOrApprovals = functionResults.some((result) => result.runItem instanceof RunToolApprovalItem);
375
534
  if (!hasPendingToolsOrApprovals) {
376
535
  if (agent.outputType === 'text') {
@@ -397,8 +556,32 @@ export async function executeToolsAndSideEffects(agent, originalInput, originalP
397
556
  }
398
557
  return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
399
558
  }
559
+ // Consolidates the logic that determines whether tool results yielded a final answer,
560
+ // triggered an interruption, or require the agent loop to continue running.
561
+ async function maybeCompleteTurnFromToolResults({ agent, runner, state, functionResults, originalInput, newResponse, preStepItems, newItems, }) {
562
+ const toolOutcome = await checkForFinalOutputFromTools(agent, functionResults, state);
563
+ if (toolOutcome.isFinalOutput) {
564
+ runner.emit('agent_end', state._context, agent, toolOutcome.finalOutput);
565
+ agent.emit('agent_end', state._context, toolOutcome.finalOutput);
566
+ return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
567
+ type: 'next_step_final_output',
568
+ output: toolOutcome.finalOutput,
569
+ });
570
+ }
571
+ if (toolOutcome.isInterrupted) {
572
+ return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
573
+ type: 'next_step_interruption',
574
+ data: {
575
+ interruptions: toolOutcome.interruptions,
576
+ },
577
+ });
578
+ }
579
+ return null;
580
+ }
400
581
  /**
401
582
  * @internal
583
+ * Normalizes tool outputs once so downstream code works with fully structured protocol items.
584
+ * Doing this here keeps API surface stable even when providers add new shapes.
402
585
  */
403
586
  export function getToolCallOutputItem(toolCall, output) {
404
587
  const maybeStructuredOutputs = normalizeStructuredToolOutputs(output);
@@ -423,318 +606,92 @@ export function getToolCallOutputItem(toolCall, output) {
423
606
  },
424
607
  };
425
608
  }
426
- /**
427
- * Accepts whatever the tool returned and attempts to coerce it into the structured protocol
428
- * shapes we expose to downstream model adapters (input_text/input_image/input_file). Tools are
429
- * allowed to return either a single structured object or an array of them; anything else falls
430
- * back to the legacy string pipeline.
431
- */
432
- function normalizeStructuredToolOutputs(output) {
433
- if (Array.isArray(output)) {
434
- const structured = [];
435
- for (const item of output) {
436
- const normalized = normalizeStructuredToolOutput(item);
437
- if (!normalized) {
438
- return null;
439
- }
440
- structured.push(normalized);
441
- }
442
- return structured;
609
+ function normalizeFileValue(value) {
610
+ const directFile = value.file;
611
+ if (typeof directFile === 'string' && directFile.length > 0) {
612
+ return directFile;
443
613
  }
444
- const normalized = normalizeStructuredToolOutput(output);
445
- return normalized ? [normalized] : null;
614
+ const normalizedObject = normalizeFileObjectCandidate(directFile);
615
+ if (normalizedObject) {
616
+ return normalizedObject;
617
+ }
618
+ const legacyValue = normalizeLegacyFileValue(value);
619
+ if (legacyValue) {
620
+ return legacyValue;
621
+ }
622
+ return null;
446
623
  }
447
- /**
448
- * Best-effort normalization of a single tool output item. If the object already matches the
449
- * protocol shape we simply cast it; otherwise we copy the recognised fields into the canonical
450
- * structure. Returning null lets the caller know we should revert to plain-string handling.
451
- */
452
- function normalizeStructuredToolOutput(value) {
624
+ function normalizeFileObjectCandidate(value) {
453
625
  if (!isRecord(value)) {
454
626
  return null;
455
627
  }
456
- const type = value.type;
457
- if (type === 'text' && typeof value.text === 'string') {
458
- const output = { type: 'text', text: value.text };
459
- if (isRecord(value.providerData)) {
460
- output.providerData = value.providerData;
461
- }
462
- return output;
463
- }
464
- if (type === 'image') {
465
- const output = { type: 'image' };
466
- let imageString;
467
- let imageFileId;
468
- const fallbackImageMediaType = isNonEmptyString(value.mediaType)
469
- ? value.mediaType
470
- : undefined;
471
- const imageField = value.image;
472
- if (typeof imageField === 'string' && imageField.length > 0) {
473
- imageString = imageField;
474
- }
475
- else if (isRecord(imageField)) {
476
- const imageObj = imageField;
477
- const inlineMediaType = isNonEmptyString(imageObj.mediaType)
478
- ? imageObj.mediaType
479
- : fallbackImageMediaType;
480
- if (isNonEmptyString(imageObj.url)) {
481
- imageString = imageObj.url;
482
- }
483
- else if (isNonEmptyString(imageObj.data)) {
484
- imageString = toInlineImageString(imageObj.data, inlineMediaType);
485
- }
486
- else if (imageObj.data instanceof Uint8Array &&
487
- imageObj.data.length > 0) {
488
- imageString = toInlineImageString(imageObj.data, inlineMediaType);
489
- }
490
- if (!imageString) {
491
- const candidateId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
492
- (isNonEmptyString(imageObj.id) && imageObj.id) ||
493
- undefined;
494
- if (candidateId) {
495
- imageFileId = candidateId;
496
- }
497
- }
498
- }
499
- if (!imageString &&
500
- typeof value.imageUrl === 'string' &&
501
- value.imageUrl.length > 0) {
502
- imageString = value.imageUrl;
503
- }
504
- if (!imageFileId &&
505
- typeof value.fileId === 'string' &&
506
- value.fileId.length > 0) {
507
- imageFileId = value.fileId;
508
- }
509
- if (!imageString &&
510
- typeof value.data === 'string' &&
511
- value.data.length > 0) {
512
- imageString = fallbackImageMediaType
513
- ? toInlineImageString(value.data, fallbackImageMediaType)
514
- : value.data;
515
- }
516
- else if (!imageString &&
517
- value.data instanceof Uint8Array &&
518
- value.data.length > 0) {
519
- imageString = toInlineImageString(value.data, fallbackImageMediaType);
628
+ if ('data' in value && value.data !== undefined) {
629
+ const dataValue = value.data;
630
+ const hasStringData = typeof dataValue === 'string' && dataValue.length > 0;
631
+ const hasBinaryData = dataValue instanceof Uint8Array && dataValue.length > 0;
632
+ if (!hasStringData && !hasBinaryData) {
633
+ return null;
520
634
  }
521
- if (typeof value.detail === 'string' && value.detail.length > 0) {
522
- output.detail = value.detail;
635
+ if (!isNonEmptyString(value.mediaType) ||
636
+ !isNonEmptyString(value.filename)) {
637
+ return null;
523
638
  }
524
- if (imageString) {
525
- output.image = imageString;
639
+ return {
640
+ data: typeof dataValue === 'string' ? dataValue : new Uint8Array(dataValue),
641
+ mediaType: value.mediaType,
642
+ filename: value.filename,
643
+ };
644
+ }
645
+ if (isNonEmptyString(value.url)) {
646
+ const result = { url: value.url };
647
+ if (isNonEmptyString(value.filename)) {
648
+ result.filename = value.filename;
526
649
  }
527
- else if (imageFileId) {
528
- output.image = { fileId: imageFileId };
650
+ return result;
651
+ }
652
+ const referencedId = (isNonEmptyString(value.id) && value.id) ||
653
+ (isNonEmptyString(value.fileId) && value.fileId);
654
+ if (referencedId) {
655
+ const result = { id: referencedId };
656
+ if (isNonEmptyString(value.filename)) {
657
+ result.filename = value.filename;
529
658
  }
530
- else {
659
+ return result;
660
+ }
661
+ return null;
662
+ }
663
+ function normalizeLegacyFileValue(value) {
664
+ const filename = typeof value.filename === 'string' && value.filename.length > 0
665
+ ? value.filename
666
+ : undefined;
667
+ const mediaType = typeof value.mediaType === 'string' && value.mediaType.length > 0
668
+ ? value.mediaType
669
+ : undefined;
670
+ if (typeof value.fileData === 'string' && value.fileData.length > 0) {
671
+ if (!mediaType || !filename) {
531
672
  return null;
532
673
  }
533
- if (isRecord(value.providerData)) {
534
- output.providerData = value.providerData;
535
- }
536
- return output;
674
+ return { data: value.fileData, mediaType, filename };
537
675
  }
538
- if (type === 'file') {
539
- const fileValue = normalizeFileValue(value);
540
- if (!fileValue) {
676
+ if (value.fileData instanceof Uint8Array && value.fileData.length > 0) {
677
+ if (!mediaType || !filename) {
541
678
  return null;
542
679
  }
543
- const output = { type: 'file', file: fileValue };
544
- if (isRecord(value.providerData)) {
545
- output.providerData = value.providerData;
680
+ return { data: new Uint8Array(value.fileData), mediaType, filename };
681
+ }
682
+ if (typeof value.fileUrl === 'string' && value.fileUrl.length > 0) {
683
+ const result = { url: value.fileUrl };
684
+ if (filename) {
685
+ result.filename = filename;
546
686
  }
547
- return output;
548
- }
549
- return null;
550
- }
551
- /**
552
- * Translates the normalized tool output into the protocol `input_*` items. This is the last hop
553
- * before we hand the data to model-specific adapters, so we generate the exact schema expected by
554
- * the protocol definitions.
555
- */
556
- function convertStructuredToolOutputToInputItem(output) {
557
- if (output.type === 'text') {
558
- const result = {
559
- type: 'input_text',
560
- text: output.text,
561
- };
562
- if (output.providerData) {
563
- result.providerData = output.providerData;
564
- }
565
- return result;
566
- }
567
- if (output.type === 'image') {
568
- const result = { type: 'input_image' };
569
- if (typeof output.detail === 'string' && output.detail.length > 0) {
570
- result.detail = output.detail;
571
- }
572
- if (typeof output.image === 'string' && output.image.length > 0) {
573
- result.image = output.image;
574
- }
575
- else if (isRecord(output.image)) {
576
- const imageObj = output.image;
577
- const inlineMediaType = isNonEmptyString(imageObj.mediaType)
578
- ? imageObj.mediaType
579
- : undefined;
580
- if (isNonEmptyString(imageObj.url)) {
581
- result.image = imageObj.url;
582
- }
583
- else if (isNonEmptyString(imageObj.data)) {
584
- result.image =
585
- inlineMediaType && !imageObj.data.startsWith('data:')
586
- ? asDataUrl(imageObj.data, inlineMediaType)
587
- : imageObj.data;
588
- }
589
- else if (imageObj.data instanceof Uint8Array &&
590
- imageObj.data.length > 0) {
591
- const base64 = encodeUint8ArrayToBase64(imageObj.data);
592
- result.image = asDataUrl(base64, inlineMediaType);
593
- }
594
- else {
595
- const referencedId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
596
- (isNonEmptyString(imageObj.id) && imageObj.id) ||
597
- undefined;
598
- if (referencedId) {
599
- result.image = { id: referencedId };
600
- }
601
- }
602
- }
603
- if (output.providerData) {
604
- result.providerData = output.providerData;
605
- }
606
- return result;
607
- }
608
- if (output.type === 'file') {
609
- const result = { type: 'input_file' };
610
- const fileValue = output.file;
611
- if (typeof fileValue === 'string') {
612
- result.file = fileValue;
613
- }
614
- else if (fileValue && typeof fileValue === 'object') {
615
- const record = fileValue;
616
- if ('data' in record && record.data) {
617
- const mediaType = record.mediaType ?? 'text/plain';
618
- if (typeof record.data === 'string') {
619
- result.file = asDataUrl(record.data, mediaType);
620
- }
621
- else {
622
- const base64 = encodeUint8ArrayToBase64(record.data);
623
- result.file = asDataUrl(base64, mediaType);
624
- }
625
- }
626
- else if (typeof record.url === 'string' && record.url.length > 0) {
627
- result.file = { url: record.url };
628
- }
629
- else {
630
- const referencedId = (typeof record.id === 'string' &&
631
- record.id.length > 0 &&
632
- record.id) ||
633
- (typeof record.fileId === 'string' && record.fileId.length > 0
634
- ? record.fileId
635
- : undefined);
636
- if (referencedId) {
637
- result.file = { id: referencedId };
638
- }
639
- }
640
- if (typeof record.filename === 'string' && record.filename.length > 0) {
641
- result.filename = record.filename;
642
- }
643
- }
644
- if (output.providerData) {
645
- result.providerData = output.providerData;
646
- }
647
- return result;
648
- }
649
- const exhaustiveCheck = output;
650
- return exhaustiveCheck;
651
- }
652
- function normalizeFileValue(value) {
653
- const directFile = value.file;
654
- if (typeof directFile === 'string' && directFile.length > 0) {
655
- return directFile;
656
- }
657
- const normalizedObject = normalizeFileObjectCandidate(directFile);
658
- if (normalizedObject) {
659
- return normalizedObject;
660
- }
661
- const legacyValue = normalizeLegacyFileValue(value);
662
- if (legacyValue) {
663
- return legacyValue;
664
- }
665
- return null;
666
- }
667
- function normalizeFileObjectCandidate(value) {
668
- if (!isRecord(value)) {
669
- return null;
670
- }
671
- if ('data' in value && value.data !== undefined) {
672
- const dataValue = value.data;
673
- const hasStringData = typeof dataValue === 'string' && dataValue.length > 0;
674
- const hasBinaryData = dataValue instanceof Uint8Array && dataValue.length > 0;
675
- if (!hasStringData && !hasBinaryData) {
676
- return null;
677
- }
678
- if (!isNonEmptyString(value.mediaType) ||
679
- !isNonEmptyString(value.filename)) {
680
- return null;
681
- }
682
- return {
683
- data: typeof dataValue === 'string' ? dataValue : new Uint8Array(dataValue),
684
- mediaType: value.mediaType,
685
- filename: value.filename,
686
- };
687
- }
688
- if (isNonEmptyString(value.url)) {
689
- const result = { url: value.url };
690
- if (isNonEmptyString(value.filename)) {
691
- result.filename = value.filename;
692
- }
693
- return result;
694
- }
695
- const referencedId = (isNonEmptyString(value.id) && value.id) ||
696
- (isNonEmptyString(value.fileId) && value.fileId);
697
- if (referencedId) {
698
- const result = { id: referencedId };
699
- if (isNonEmptyString(value.filename)) {
700
- result.filename = value.filename;
701
- }
702
- return result;
703
- }
704
- return null;
705
- }
706
- function normalizeLegacyFileValue(value) {
707
- const filename = typeof value.filename === 'string' && value.filename.length > 0
708
- ? value.filename
709
- : undefined;
710
- const mediaType = typeof value.mediaType === 'string' && value.mediaType.length > 0
711
- ? value.mediaType
712
- : undefined;
713
- if (typeof value.fileData === 'string' && value.fileData.length > 0) {
714
- if (!mediaType || !filename) {
715
- return null;
716
- }
717
- return { data: value.fileData, mediaType, filename };
718
- }
719
- if (value.fileData instanceof Uint8Array && value.fileData.length > 0) {
720
- if (!mediaType || !filename) {
721
- return null;
722
- }
723
- return { data: new Uint8Array(value.fileData), mediaType, filename };
724
- }
725
- if (typeof value.fileUrl === 'string' && value.fileUrl.length > 0) {
726
- const result = { url: value.fileUrl };
727
- if (filename) {
728
- result.filename = filename;
729
- }
730
- return result;
731
- }
732
- if (typeof value.fileId === 'string' && value.fileId.length > 0) {
733
- const result = { id: value.fileId };
734
- if (filename) {
735
- result.filename = filename;
736
- }
737
- return result;
687
+ return result;
688
+ }
689
+ if (typeof value.fileId === 'string' && value.fileId.length > 0) {
690
+ const result = { id: value.fileId };
691
+ if (filename) {
692
+ result.filename = filename;
693
+ }
694
+ return result;
738
695
  }
739
696
  return null;
740
697
  }
@@ -759,6 +716,8 @@ function asDataUrl(base64, mediaType) {
759
716
  }
760
717
  /**
761
718
  * @internal
719
+ * Runs every function tool call requested by the model and returns their outputs alongside
720
+ * the `RunItem` instances that should be appended to history.
762
721
  */
763
722
  export async function executeFunctionToolCalls(agent, toolRuns, runner, state) {
764
723
  async function runSingleTool(toolRun) {
@@ -771,6 +730,7 @@ export async function executeFunctionToolCalls(agent, toolRuns, runner, state) {
771
730
  parsedArgs = JSON.parse(parsedArgs);
772
731
  }
773
732
  }
733
+ // Some tools require a human or policy check before execution; defer until approval is recorded.
774
734
  const needsApproval = await toolRun.tool.needsApproval(state._context, parsedArgs, toolRun.toolCall.callId);
775
735
  if (needsApproval) {
776
736
  const approval = state._context.isToolApproved({
@@ -924,6 +884,8 @@ async function _runComputerActionAndScreenshot(computer, toolCall) {
924
884
  }
925
885
  /**
926
886
  * @internal
887
+ * Executes any computer-use actions emitted by the model and returns the resulting items so the
888
+ * run history reflects the computer session.
927
889
  */
928
890
  export async function executeComputerActions(agent, actions, runner, runContext, customLogger = undefined) {
929
891
  const _logger = customLogger ?? logger;
@@ -956,7 +918,7 @@ export async function executeComputerActions(agent, actions, runner, runContext,
956
918
  toolCall,
957
919
  });
958
920
  }
959
- // Always return a screenshot as a base64 data URL
921
+ // Return the screenshot as a data URL when available; fall back to an empty string on failures.
960
922
  const imageUrl = output ? `data:image/png;base64,${output}` : '';
961
923
  const rawItem = {
962
924
  type: 'computer_call_result',
@@ -969,6 +931,8 @@ export async function executeComputerActions(agent, actions, runner, runContext,
969
931
  }
970
932
  /**
971
933
  * @internal
934
+ * Drives handoff calls by invoking the downstream agent and capturing any generated items so
935
+ * the current agent can continue with the new context.
972
936
  */
973
937
  export async function executeHandoffCalls(agent, originalInput, preStepItems, newStepItems, newResponse, runHandoffs, runner, runContext) {
974
938
  newStepItems = [...newStepItems];
@@ -1037,6 +1001,8 @@ const NOT_FINAL_OUTPUT = {
1037
1001
  };
1038
1002
  /**
1039
1003
  * @internal
1004
+ * Determines whether tool executions produced a final agent output, triggered an interruption,
1005
+ * or whether the agent loop should continue collecting more responses.
1040
1006
  */
1041
1007
  export async function checkForFinalOutputFromTools(agent, toolResults, state) {
1042
1008
  if (toolResults.length === 0) {
@@ -1163,4 +1129,571 @@ export class AgentToolUseTracker {
1163
1129
  }));
1164
1130
  }
1165
1131
  }
1132
+ /**
1133
+ * @internal
1134
+ * Convert a user-provided input into a list of input items.
1135
+ */
1136
+ export function toInputItemList(input) {
1137
+ if (typeof input === 'string') {
1138
+ return [
1139
+ {
1140
+ type: 'message',
1141
+ role: 'user',
1142
+ content: input,
1143
+ },
1144
+ ];
1145
+ }
1146
+ return [...input];
1147
+ }
1148
+ /**
1149
+ * @internal
1150
+ * Extract model output items from run items, excluding tool approval items.
1151
+ */
1152
+ export function extractOutputItemsFromRunItems(items) {
1153
+ return items
1154
+ .filter((item) => item.type !== 'tool_approval_item')
1155
+ .map((item) => item.rawItem);
1156
+ }
1157
+ function normalizeItemsForSessionPersistence(items) {
1158
+ // Persisted sessions must avoid raw binary so we convert every item into a JSON-safe shape before writing to storage.
1159
+ return items.map((item) => sanitizeValueForSession(stripTransientCallIds(item)));
1160
+ }
1161
+ function sanitizeValueForSession(value, context = {}) {
1162
+ if (value === null || value === undefined) {
1163
+ return value;
1164
+ }
1165
+ // Convert supported binary payloads into ArrayBuffer views before serialization.
1166
+ const binary = toUint8ArrayIfBinary(value);
1167
+ if (binary) {
1168
+ return toDataUrlFromBytes(binary, context.mediaType);
1169
+ }
1170
+ if (Array.isArray(value)) {
1171
+ return value.map((entry) => sanitizeValueForSession(entry, context));
1172
+ }
1173
+ if (!isPlainObject(value)) {
1174
+ return value;
1175
+ }
1176
+ const record = value;
1177
+ const result = {};
1178
+ const mediaType = typeof record.mediaType === 'string' && record.mediaType.length > 0
1179
+ ? record.mediaType
1180
+ : context.mediaType;
1181
+ for (const [key, entry] of Object.entries(record)) {
1182
+ // Propagate explicit media type only when walking into binary payload containers.
1183
+ const nextContext = key === 'data' || key === 'fileData' ? { mediaType } : context;
1184
+ result[key] = sanitizeValueForSession(entry, nextContext);
1185
+ }
1186
+ return result;
1187
+ }
1188
+ function toUint8ArrayIfBinary(value) {
1189
+ // Normalize the diverse binary containers we may receive into a shared Uint8Array view.
1190
+ if (value instanceof ArrayBuffer) {
1191
+ return new Uint8Array(value);
1192
+ }
1193
+ if (isArrayBufferView(value)) {
1194
+ const view = value;
1195
+ return new Uint8Array(view.buffer, view.byteOffset, view.byteLength);
1196
+ }
1197
+ if (isNodeBuffer(value)) {
1198
+ const view = value;
1199
+ return new Uint8Array(view.buffer, view.byteOffset, view.byteLength);
1200
+ }
1201
+ if (isSerializedBufferSnapshot(value)) {
1202
+ const snapshot = value;
1203
+ return Uint8Array.from(snapshot.data);
1204
+ }
1205
+ return undefined;
1206
+ }
1207
+ function toDataUrlFromBytes(bytes, mediaType) {
1208
+ // Convert binary payloads into a durable data URL so session files remain self-contained.
1209
+ const base64 = encodeUint8ArrayToBase64(bytes);
1210
+ // Note that OpenAI Responses API never accepts application/octet-stream as a media type,
1211
+ // so we fall back to text/plain; that said, tools are supposed to return a valid media type when this utility is used.
1212
+ const type = mediaType && !mediaType.startsWith('data:') ? mediaType : 'text/plain';
1213
+ return `data:${type};base64,${base64}`;
1214
+ }
1215
+ function isPlainObject(value) {
1216
+ if (typeof value !== 'object' || value === null) {
1217
+ return false;
1218
+ }
1219
+ const proto = Object.getPrototypeOf(value);
1220
+ return proto === Object.prototype || proto === null;
1221
+ }
1222
+ function stripTransientCallIds(value) {
1223
+ if (value === null || value === undefined) {
1224
+ return value;
1225
+ }
1226
+ if (Array.isArray(value)) {
1227
+ return value.map((entry) => stripTransientCallIds(entry));
1228
+ }
1229
+ if (!isPlainObject(value)) {
1230
+ return value;
1231
+ }
1232
+ const record = value;
1233
+ const result = {};
1234
+ const isProtocolItem = typeof record.type === 'string' && record.type.length > 0;
1235
+ const shouldStripId = isProtocolItem && shouldStripIdForType(record.type);
1236
+ for (const [key, entry] of Object.entries(record)) {
1237
+ if (shouldStripId && key === 'id') {
1238
+ continue;
1239
+ }
1240
+ result[key] = stripTransientCallIds(entry);
1241
+ }
1242
+ return result;
1243
+ }
1244
+ function shouldStripIdForType(type) {
1245
+ switch (type) {
1246
+ case 'function_call':
1247
+ case 'function_call_result':
1248
+ return true;
1249
+ default:
1250
+ return false;
1251
+ }
1252
+ }
1253
+ /**
1254
+ * @internal
1255
+ * Persist full turn (input + outputs) for non-streaming runs.
1256
+ */
1257
+ // Persists the combination of user inputs (possibly filtered) and model outputs for a completed turn.
1258
+ export async function saveToSession(session, sessionInputItems, result) {
1259
+ if (!session) {
1260
+ return;
1261
+ }
1262
+ const inputItems = sessionInputItems ?? [];
1263
+ const state = result.state;
1264
+ const alreadyPersisted = state._currentTurnPersistedItemCount ?? 0;
1265
+ // Persist only the portion of _generatedItems that has not yet been stored for this turn.
1266
+ const newRunItems = result.newItems.slice(alreadyPersisted);
1267
+ if (process.env.OPENAI_AGENTS__DEBUG_SAVE_SESSION) {
1268
+ console.debug('saveToSession:newRunItems', newRunItems.map((item) => item.type));
1269
+ }
1270
+ const outputItems = extractOutputItemsFromRunItems(newRunItems);
1271
+ const itemsToSave = [...inputItems, ...outputItems];
1272
+ if (itemsToSave.length === 0) {
1273
+ state._currentTurnPersistedItemCount =
1274
+ alreadyPersisted + newRunItems.length;
1275
+ return;
1276
+ }
1277
+ const sanitizedItems = normalizeItemsForSessionPersistence(itemsToSave);
1278
+ await session.addItems(sanitizedItems);
1279
+ state._currentTurnPersistedItemCount = alreadyPersisted + newRunItems.length;
1280
+ }
1281
+ /**
1282
+ * @internal
1283
+ * Persist only the user input for streaming runs at start.
1284
+ */
1285
+ // For streaming runs we persist user input as soon as it is sent so reconnections can resume.
1286
+ export async function saveStreamInputToSession(session, sessionInputItems) {
1287
+ if (!session) {
1288
+ return;
1289
+ }
1290
+ if (!sessionInputItems || sessionInputItems.length === 0) {
1291
+ return;
1292
+ }
1293
+ const sanitizedInput = normalizeItemsForSessionPersistence(sessionInputItems);
1294
+ await session.addItems(sanitizedInput);
1295
+ }
1296
+ /**
1297
+ * @internal
1298
+ * Persist only the model outputs for streaming runs at the end of a turn.
1299
+ */
1300
+ // Complements saveStreamInputToSession by recording the streaming outputs at the end of the turn.
1301
+ export async function saveStreamResultToSession(session, result) {
1302
+ if (!session) {
1303
+ return;
1304
+ }
1305
+ const state = result.state;
1306
+ const alreadyPersisted = state._currentTurnPersistedItemCount ?? 0;
1307
+ const newRunItems = result.newItems.slice(alreadyPersisted);
1308
+ const itemsToSave = extractOutputItemsFromRunItems(newRunItems);
1309
+ if (itemsToSave.length === 0) {
1310
+ state._currentTurnPersistedItemCount =
1311
+ alreadyPersisted + newRunItems.length;
1312
+ return;
1313
+ }
1314
+ const sanitizedItems = normalizeItemsForSessionPersistence(itemsToSave);
1315
+ await session.addItems(sanitizedItems);
1316
+ state._currentTurnPersistedItemCount = alreadyPersisted + newRunItems.length;
1317
+ }
1318
+ export async function prepareInputItemsWithSession(input, session, sessionInputCallback, options) {
1319
+ if (!session) {
1320
+ return {
1321
+ preparedInput: input,
1322
+ sessionItems: undefined,
1323
+ };
1324
+ }
1325
+ const includeHistoryInPreparedInput = options?.includeHistoryInPreparedInput ?? true;
1326
+ const preserveDroppedNewItems = options?.preserveDroppedNewItems ?? false;
1327
+ const history = await session.getItems();
1328
+ const newInputItems = Array.isArray(input)
1329
+ ? [...input]
1330
+ : toInputItemList(input);
1331
+ if (!sessionInputCallback) {
1332
+ return {
1333
+ preparedInput: includeHistoryInPreparedInput
1334
+ ? [...history, ...newInputItems]
1335
+ : newInputItems,
1336
+ sessionItems: newInputItems,
1337
+ };
1338
+ }
1339
+ // Capture snapshots before invoking the callback so we can reason about the original state even
1340
+ // if the callback mutates the history array in-place.
1341
+ const historySnapshot = history.slice();
1342
+ const newInputSnapshot = newInputItems.slice();
1343
+ // Delegate history reconciliation to the user-supplied callback. It must return a concrete list
1344
+ // to keep downstream model requests well-typed.
1345
+ const combined = await sessionInputCallback(history, newInputItems);
1346
+ if (!Array.isArray(combined)) {
1347
+ throw new UserError('Session input callback must return an array of AgentInputItem objects.');
1348
+ }
1349
+ const historyCounts = buildItemFrequencyMap(historySnapshot);
1350
+ const newInputCounts = buildItemFrequencyMap(newInputSnapshot);
1351
+ const historyRefs = buildItemReferenceMap(historySnapshot);
1352
+ const newInputRefs = buildItemReferenceMap(newInputSnapshot);
1353
+ const appended = [];
1354
+ for (const item of combined) {
1355
+ const key = sessionItemKey(item);
1356
+ if (consumeReference(newInputRefs, key, item)) {
1357
+ decrementCount(newInputCounts, key);
1358
+ appended.push(item);
1359
+ continue;
1360
+ }
1361
+ // Prioritize exact history matches before payload-based counts so callbacks that surface
1362
+ // history ahead of identical new inputs keep previously persisted items out of the new queue.
1363
+ if (consumeReference(historyRefs, key, item)) {
1364
+ decrementCount(historyCounts, key);
1365
+ continue;
1366
+ }
1367
+ const historyRemaining = historyCounts.get(key) ?? 0;
1368
+ if (historyRemaining > 0) {
1369
+ historyCounts.set(key, historyRemaining - 1);
1370
+ continue;
1371
+ }
1372
+ const newRemaining = newInputCounts.get(key) ?? 0;
1373
+ if (newRemaining > 0) {
1374
+ newInputCounts.set(key, newRemaining - 1);
1375
+ appended.push(item);
1376
+ continue;
1377
+ }
1378
+ appended.push(item);
1379
+ }
1380
+ // Preserve redacted inputs for model delivery when requested (e.g. server-managed histories).
1381
+ const preparedItems = includeHistoryInPreparedInput
1382
+ ? combined
1383
+ : appended.length > 0
1384
+ ? appended
1385
+ : preserveDroppedNewItems
1386
+ ? newInputSnapshot
1387
+ : [];
1388
+ return {
1389
+ preparedInput: preparedItems,
1390
+ // Respect callbacks that intentionally drop the latest inputs (e.g. to redact sensitive
1391
+ // values) by persisting only the items they kept in the combined array.
1392
+ sessionItems: appended,
1393
+ };
1394
+ }
1395
+ /**
1396
+ * Accepts whatever the tool returned and attempts to coerce it into the structured protocol
1397
+ * shapes we expose to downstream model adapters (input_text/input_image/input_file). Tools are
1398
+ * allowed to return either a single structured object or an array of them; anything else falls
1399
+ * back to the legacy string pipeline.
1400
+ */
1401
+ function normalizeStructuredToolOutputs(output) {
1402
+ if (Array.isArray(output)) {
1403
+ const structured = [];
1404
+ for (const item of output) {
1405
+ const normalized = normalizeStructuredToolOutput(item);
1406
+ if (!normalized) {
1407
+ return null;
1408
+ }
1409
+ structured.push(normalized);
1410
+ }
1411
+ return structured;
1412
+ }
1413
+ const normalized = normalizeStructuredToolOutput(output);
1414
+ return normalized ? [normalized] : null;
1415
+ }
1416
+ /**
1417
+ * Best-effort normalization of a single tool output item. If the object already matches the
1418
+ * protocol shape we simply cast it; otherwise we copy the recognised fields into the canonical
1419
+ * structure. Returning null lets the caller know we should revert to plain-string handling.
1420
+ */
1421
+ function normalizeStructuredToolOutput(value) {
1422
+ if (!isRecord(value)) {
1423
+ return null;
1424
+ }
1425
+ const type = value.type;
1426
+ if (type === 'text' && typeof value.text === 'string') {
1427
+ const output = { type: 'text', text: value.text };
1428
+ if (isRecord(value.providerData)) {
1429
+ output.providerData = value.providerData;
1430
+ }
1431
+ return output;
1432
+ }
1433
+ if (type === 'image') {
1434
+ const output = { type: 'image' };
1435
+ let imageString;
1436
+ let imageFileId;
1437
+ const fallbackImageMediaType = isNonEmptyString(value.mediaType)
1438
+ ? value.mediaType
1439
+ : undefined;
1440
+ const imageField = value.image;
1441
+ if (typeof imageField === 'string' && imageField.length > 0) {
1442
+ imageString = imageField;
1443
+ }
1444
+ else if (isRecord(imageField)) {
1445
+ const imageObj = imageField;
1446
+ const inlineMediaType = isNonEmptyString(imageObj.mediaType)
1447
+ ? imageObj.mediaType
1448
+ : fallbackImageMediaType;
1449
+ if (isNonEmptyString(imageObj.url)) {
1450
+ imageString = imageObj.url;
1451
+ }
1452
+ else if (isNonEmptyString(imageObj.data)) {
1453
+ imageString = toInlineImageString(imageObj.data, inlineMediaType);
1454
+ }
1455
+ else if (imageObj.data instanceof Uint8Array &&
1456
+ imageObj.data.length > 0) {
1457
+ imageString = toInlineImageString(imageObj.data, inlineMediaType);
1458
+ }
1459
+ if (!imageString) {
1460
+ const candidateId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
1461
+ (isNonEmptyString(imageObj.id) && imageObj.id) ||
1462
+ undefined;
1463
+ if (candidateId) {
1464
+ imageFileId = candidateId;
1465
+ }
1466
+ }
1467
+ }
1468
+ if (!imageString &&
1469
+ typeof value.imageUrl === 'string' &&
1470
+ value.imageUrl.length > 0) {
1471
+ imageString = value.imageUrl;
1472
+ }
1473
+ if (!imageFileId &&
1474
+ typeof value.fileId === 'string' &&
1475
+ value.fileId.length > 0) {
1476
+ imageFileId = value.fileId;
1477
+ }
1478
+ if (!imageString &&
1479
+ typeof value.data === 'string' &&
1480
+ value.data.length > 0) {
1481
+ imageString = fallbackImageMediaType
1482
+ ? toInlineImageString(value.data, fallbackImageMediaType)
1483
+ : value.data;
1484
+ }
1485
+ else if (!imageString &&
1486
+ value.data instanceof Uint8Array &&
1487
+ value.data.length > 0) {
1488
+ imageString = toInlineImageString(value.data, fallbackImageMediaType);
1489
+ }
1490
+ if (typeof value.detail === 'string' && value.detail.length > 0) {
1491
+ output.detail = value.detail;
1492
+ }
1493
+ if (imageString) {
1494
+ output.image = imageString;
1495
+ }
1496
+ else if (imageFileId) {
1497
+ output.image = { fileId: imageFileId };
1498
+ }
1499
+ else {
1500
+ return null;
1501
+ }
1502
+ if (isRecord(value.providerData)) {
1503
+ output.providerData = value.providerData;
1504
+ }
1505
+ return output;
1506
+ }
1507
+ if (type === 'file') {
1508
+ const fileValue = normalizeFileValue(value);
1509
+ if (!fileValue) {
1510
+ return null;
1511
+ }
1512
+ const output = { type: 'file', file: fileValue };
1513
+ if (isRecord(value.providerData)) {
1514
+ output.providerData = value.providerData;
1515
+ }
1516
+ return output;
1517
+ }
1518
+ return null;
1519
+ }
1520
+ /**
1521
+ * Translates the normalized tool output into the protocol `input_*` items. This is the last hop
1522
+ * before we hand the data to model-specific adapters, so we generate the exact schema expected by
1523
+ * the protocol definitions.
1524
+ */
1525
+ function convertStructuredToolOutputToInputItem(output) {
1526
+ if (output.type === 'text') {
1527
+ const result = {
1528
+ type: 'input_text',
1529
+ text: output.text,
1530
+ };
1531
+ if (output.providerData) {
1532
+ result.providerData = output.providerData;
1533
+ }
1534
+ return result;
1535
+ }
1536
+ if (output.type === 'image') {
1537
+ const result = { type: 'input_image' };
1538
+ if (typeof output.detail === 'string' && output.detail.length > 0) {
1539
+ result.detail = output.detail;
1540
+ }
1541
+ if (typeof output.image === 'string' && output.image.length > 0) {
1542
+ result.image = output.image;
1543
+ }
1544
+ else if (isRecord(output.image)) {
1545
+ const imageObj = output.image;
1546
+ const inlineMediaType = isNonEmptyString(imageObj.mediaType)
1547
+ ? imageObj.mediaType
1548
+ : undefined;
1549
+ if (isNonEmptyString(imageObj.url)) {
1550
+ result.image = imageObj.url;
1551
+ }
1552
+ else if (isNonEmptyString(imageObj.data)) {
1553
+ result.image =
1554
+ inlineMediaType && !imageObj.data.startsWith('data:')
1555
+ ? asDataUrl(imageObj.data, inlineMediaType)
1556
+ : imageObj.data;
1557
+ }
1558
+ else if (imageObj.data instanceof Uint8Array &&
1559
+ imageObj.data.length > 0) {
1560
+ const base64 = encodeUint8ArrayToBase64(imageObj.data);
1561
+ result.image = asDataUrl(base64, inlineMediaType);
1562
+ }
1563
+ else {
1564
+ const referencedId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
1565
+ (isNonEmptyString(imageObj.id) && imageObj.id) ||
1566
+ undefined;
1567
+ if (referencedId) {
1568
+ result.image = { id: referencedId };
1569
+ }
1570
+ }
1571
+ }
1572
+ if (output.providerData) {
1573
+ result.providerData = output.providerData;
1574
+ }
1575
+ return result;
1576
+ }
1577
+ if (output.type === 'file') {
1578
+ const result = { type: 'input_file' };
1579
+ const fileValue = output.file;
1580
+ if (typeof fileValue === 'string') {
1581
+ result.file = fileValue;
1582
+ }
1583
+ else if (fileValue && typeof fileValue === 'object') {
1584
+ const record = fileValue;
1585
+ if ('data' in record && record.data) {
1586
+ const mediaType = record.mediaType ?? 'text/plain';
1587
+ if (typeof record.data === 'string') {
1588
+ result.file = asDataUrl(record.data, mediaType);
1589
+ }
1590
+ else {
1591
+ const base64 = encodeUint8ArrayToBase64(record.data);
1592
+ result.file = asDataUrl(base64, mediaType);
1593
+ }
1594
+ }
1595
+ else if (typeof record.url === 'string' && record.url.length > 0) {
1596
+ result.file = { url: record.url };
1597
+ }
1598
+ else {
1599
+ const referencedId = (typeof record.id === 'string' &&
1600
+ record.id.length > 0 &&
1601
+ record.id) ||
1602
+ (typeof record.fileId === 'string' && record.fileId.length > 0
1603
+ ? record.fileId
1604
+ : undefined);
1605
+ if (referencedId) {
1606
+ result.file = { id: referencedId };
1607
+ }
1608
+ }
1609
+ if (typeof record.filename === 'string' && record.filename.length > 0) {
1610
+ result.filename = record.filename;
1611
+ }
1612
+ }
1613
+ if (output.providerData) {
1614
+ result.providerData = output.providerData;
1615
+ }
1616
+ return result;
1617
+ }
1618
+ const exhaustiveCheck = output;
1619
+ return exhaustiveCheck;
1620
+ }
1621
+ function buildItemFrequencyMap(items) {
1622
+ const counts = new Map();
1623
+ for (const item of items) {
1624
+ const key = sessionItemKey(item);
1625
+ counts.set(key, (counts.get(key) ?? 0) + 1);
1626
+ }
1627
+ return counts;
1628
+ }
1629
+ function buildItemReferenceMap(items) {
1630
+ const refs = new Map();
1631
+ for (const item of items) {
1632
+ const key = sessionItemKey(item);
1633
+ const list = refs.get(key);
1634
+ if (list) {
1635
+ list.push(item);
1636
+ }
1637
+ else {
1638
+ refs.set(key, [item]);
1639
+ }
1640
+ }
1641
+ return refs;
1642
+ }
1643
+ function consumeReference(refs, key, target) {
1644
+ const candidates = refs.get(key);
1645
+ if (!candidates || candidates.length === 0) {
1646
+ return false;
1647
+ }
1648
+ const index = candidates.findIndex((candidate) => candidate === target);
1649
+ if (index === -1) {
1650
+ return false;
1651
+ }
1652
+ candidates.splice(index, 1);
1653
+ if (candidates.length === 0) {
1654
+ refs.delete(key);
1655
+ }
1656
+ return true;
1657
+ }
1658
+ function decrementCount(map, key) {
1659
+ const remaining = (map.get(key) ?? 0) - 1;
1660
+ if (remaining <= 0) {
1661
+ map.delete(key);
1662
+ }
1663
+ else {
1664
+ map.set(key, remaining);
1665
+ }
1666
+ }
1667
+ function sessionItemKey(item) {
1668
+ return JSON.stringify(item, sessionSerializationReplacer);
1669
+ }
1670
+ function sessionSerializationReplacer(_key, value) {
1671
+ if (value instanceof ArrayBuffer) {
1672
+ return {
1673
+ __type: 'ArrayBuffer',
1674
+ data: encodeUint8ArrayToBase64(new Uint8Array(value)),
1675
+ };
1676
+ }
1677
+ if (isArrayBufferView(value)) {
1678
+ const view = value;
1679
+ return {
1680
+ __type: view.constructor.name,
1681
+ data: encodeUint8ArrayToBase64(new Uint8Array(view.buffer, view.byteOffset, view.byteLength)),
1682
+ };
1683
+ }
1684
+ if (isNodeBuffer(value)) {
1685
+ const view = value;
1686
+ return {
1687
+ __type: 'Buffer',
1688
+ data: encodeUint8ArrayToBase64(new Uint8Array(view.buffer, view.byteOffset, view.byteLength)),
1689
+ };
1690
+ }
1691
+ if (isSerializedBufferSnapshot(value)) {
1692
+ return {
1693
+ __type: 'Buffer',
1694
+ data: encodeUint8ArrayToBase64(Uint8Array.from(value.data)),
1695
+ };
1696
+ }
1697
+ return value;
1698
+ }
1166
1699
  //# sourceMappingURL=runImplementation.mjs.map