@openai/agents-core 0.2.1 → 0.3.1
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.
- package/dist/editor.d.ts +38 -0
- package/dist/editor.js +3 -0
- package/dist/editor.js.map +1 -0
- package/dist/editor.mjs +2 -0
- package/dist/editor.mjs.map +1 -0
- package/dist/extensions/handoffFilters.js +4 -0
- package/dist/extensions/handoffFilters.js.map +1 -1
- package/dist/extensions/handoffFilters.mjs +4 -0
- package/dist/extensions/handoffFilters.mjs.map +1 -1
- package/dist/index.d.ts +8 -2
- package/dist/index.js +8 -2
- package/dist/index.js.map +1 -1
- package/dist/index.mjs +3 -1
- package/dist/index.mjs.map +1 -1
- package/dist/items.d.ts +518 -4
- package/dist/items.js +22 -1
- package/dist/items.js.map +1 -1
- package/dist/items.mjs +22 -1
- package/dist/items.mjs.map +1 -1
- package/dist/memory/memorySession.d.ts +22 -0
- package/dist/memory/memorySession.js +64 -0
- package/dist/memory/memorySession.js.map +1 -0
- package/dist/memory/memorySession.mjs +60 -0
- package/dist/memory/memorySession.mjs.map +1 -0
- package/dist/memory/session.d.ts +36 -0
- package/dist/memory/session.js +3 -0
- package/dist/memory/session.js.map +1 -0
- package/dist/memory/session.mjs +2 -0
- package/dist/memory/session.mjs.map +1 -0
- package/dist/metadata.js +2 -2
- package/dist/metadata.mjs +2 -2
- package/dist/model.d.ts +10 -2
- package/dist/run.d.ts +88 -8
- package/dist/run.js +859 -347
- package/dist/run.js.map +1 -1
- package/dist/run.mjs +859 -347
- package/dist/run.mjs.map +1 -1
- package/dist/runContext.js +2 -2
- package/dist/runContext.js.map +1 -1
- package/dist/runContext.mjs +2 -2
- package/dist/runContext.mjs.map +1 -1
- package/dist/runImplementation.d.ts +37 -3
- package/dist/runImplementation.js +1116 -319
- package/dist/runImplementation.js.map +1 -1
- package/dist/runImplementation.mjs +1106 -317
- package/dist/runImplementation.mjs.map +1 -1
- package/dist/runState.d.ts +4453 -785
- package/dist/runState.js +62 -3
- package/dist/runState.js.map +1 -1
- package/dist/runState.mjs +62 -3
- package/dist/runState.mjs.map +1 -1
- package/dist/shell.d.ts +36 -0
- package/dist/shell.js +3 -0
- package/dist/shell.js.map +1 -0
- package/dist/shell.mjs +2 -0
- package/dist/shell.mjs.map +1 -0
- package/dist/tool.d.ts +62 -1
- package/dist/tool.js +30 -0
- package/dist/tool.js.map +1 -1
- package/dist/tool.mjs +28 -0
- package/dist/tool.mjs.map +1 -1
- package/dist/types/aliases.d.ts +3 -3
- package/dist/types/protocol.d.ts +5470 -1519
- package/dist/types/protocol.js +74 -1
- package/dist/types/protocol.js.map +1 -1
- package/dist/types/protocol.mjs +73 -0
- package/dist/types/protocol.mjs.map +1 -1
- package/dist/utils/applyDiff.d.ts +9 -0
- package/dist/utils/applyDiff.js +275 -0
- package/dist/utils/applyDiff.js.map +1 -0
- package/dist/utils/applyDiff.mjs +272 -0
- package/dist/utils/applyDiff.mjs.map +1 -0
- package/dist/utils/index.d.ts +1 -0
- package/dist/utils/index.js +3 -1
- package/dist/utils/index.js.map +1 -1
- package/dist/utils/index.mjs +1 -0
- package/dist/utils/index.mjs.map +1 -1
- package/dist/utils/serialize.js +12 -0
- package/dist/utils/serialize.js.map +1 -1
- package/dist/utils/serialize.mjs +12 -0
- package/dist/utils/serialize.mjs.map +1 -1
- package/dist/utils/smartString.d.ts +9 -0
- package/dist/utils/smartString.js +15 -0
- package/dist/utils/smartString.js.map +1 -1
- package/dist/utils/smartString.mjs +14 -3
- package/dist/utils/smartString.mjs.map +1 -1
- package/package.json +3 -3
|
@@ -6,15 +6,23 @@ Object.defineProperty(exports, "__esModule", { value: true });
|
|
|
6
6
|
exports.AgentToolUseTracker = exports.nextStepSchema = void 0;
|
|
7
7
|
exports.processModelResponse = processModelResponse;
|
|
8
8
|
exports.maybeResetToolChoice = maybeResetToolChoice;
|
|
9
|
-
exports.
|
|
10
|
-
exports.
|
|
9
|
+
exports.resolveInterruptedTurn = resolveInterruptedTurn;
|
|
10
|
+
exports.resolveTurnAfterModelResponse = resolveTurnAfterModelResponse;
|
|
11
11
|
exports.getToolCallOutputItem = getToolCallOutputItem;
|
|
12
12
|
exports.executeFunctionToolCalls = executeFunctionToolCalls;
|
|
13
|
+
exports.executeShellActions = executeShellActions;
|
|
14
|
+
exports.executeApplyPatchOperations = executeApplyPatchOperations;
|
|
13
15
|
exports.executeComputerActions = executeComputerActions;
|
|
14
16
|
exports.executeHandoffCalls = executeHandoffCalls;
|
|
15
17
|
exports.checkForFinalOutputFromTools = checkForFinalOutputFromTools;
|
|
16
18
|
exports.streamStepItemsToRunResult = streamStepItemsToRunResult;
|
|
17
19
|
exports.addStepToRunResult = addStepToRunResult;
|
|
20
|
+
exports.toInputItemList = toInputItemList;
|
|
21
|
+
exports.extractOutputItemsFromRunItems = extractOutputItemsFromRunItems;
|
|
22
|
+
exports.saveToSession = saveToSession;
|
|
23
|
+
exports.saveStreamInputToSession = saveStreamInputToSession;
|
|
24
|
+
exports.saveStreamResultToSession = saveStreamResultToSession;
|
|
25
|
+
exports.prepareInputItemsWithSession = prepareInputItemsWithSession;
|
|
18
26
|
const agent_1 = require("./agent.js");
|
|
19
27
|
const errors_1 = require("./errors.js");
|
|
20
28
|
const handoff_1 = require("./handoff.js");
|
|
@@ -24,25 +32,75 @@ const messages_1 = require("./utils/messages.js");
|
|
|
24
32
|
const createSpans_1 = require("./tracing/createSpans.js");
|
|
25
33
|
const tools_1 = require("./utils/tools.js");
|
|
26
34
|
const base64_1 = require("./utils/base64.js");
|
|
35
|
+
const smartString_1 = require("./utils/smartString.js");
|
|
27
36
|
const safeExecute_1 = require("./utils/safeExecute.js");
|
|
28
37
|
const context_1 = require("./tracing/context.js");
|
|
29
38
|
const events_1 = require("./events.js");
|
|
30
39
|
const zod_1 = require("zod");
|
|
31
|
-
const smartString_1 = require("./utils/smartString.js");
|
|
32
40
|
const utils_1 = require("./utils/index.js");
|
|
41
|
+
function isApprovalItemLike(value) {
|
|
42
|
+
if (!value || typeof value !== 'object') {
|
|
43
|
+
return false;
|
|
44
|
+
}
|
|
45
|
+
if (!('rawItem' in value)) {
|
|
46
|
+
return false;
|
|
47
|
+
}
|
|
48
|
+
const rawItem = value.rawItem;
|
|
49
|
+
if (!rawItem || typeof rawItem !== 'object') {
|
|
50
|
+
return false;
|
|
51
|
+
}
|
|
52
|
+
const itemType = rawItem.type;
|
|
53
|
+
return itemType === 'function_call' || itemType === 'hosted_tool_call';
|
|
54
|
+
}
|
|
55
|
+
function getApprovalIdentity(approval) {
|
|
56
|
+
const rawItem = approval.rawItem;
|
|
57
|
+
if (!rawItem) {
|
|
58
|
+
return undefined;
|
|
59
|
+
}
|
|
60
|
+
if (rawItem.type === 'function_call' && rawItem.callId) {
|
|
61
|
+
return `function_call:${rawItem.callId}`;
|
|
62
|
+
}
|
|
63
|
+
if ('callId' in rawItem && rawItem.callId) {
|
|
64
|
+
return `${rawItem.type}:${rawItem.callId}`;
|
|
65
|
+
}
|
|
66
|
+
const id = 'id' in rawItem ? rawItem.id : undefined;
|
|
67
|
+
if (id) {
|
|
68
|
+
return `${rawItem.type}:${id}`;
|
|
69
|
+
}
|
|
70
|
+
const providerData = typeof rawItem.providerData === 'object' && rawItem.providerData
|
|
71
|
+
? rawItem.providerData
|
|
72
|
+
: undefined;
|
|
73
|
+
if (providerData?.id) {
|
|
74
|
+
return `${rawItem.type}:provider:${providerData.id}`;
|
|
75
|
+
}
|
|
76
|
+
const agentName = 'agent' in approval && approval.agent ? approval.agent.name : '';
|
|
77
|
+
try {
|
|
78
|
+
return `${agentName}:${rawItem.type}:${JSON.stringify(rawItem)}`;
|
|
79
|
+
}
|
|
80
|
+
catch {
|
|
81
|
+
return `${agentName}:${rawItem.type}`;
|
|
82
|
+
}
|
|
83
|
+
}
|
|
33
84
|
/**
|
|
34
85
|
* @internal
|
|
86
|
+
* Walks a raw model response and classifies each item so the runner can schedule follow-up work.
|
|
87
|
+
* Returns both the serializable RunItems (for history/streaming) and the actionable tool metadata.
|
|
35
88
|
*/
|
|
36
89
|
function processModelResponse(modelResponse, agent, tools, handoffs) {
|
|
37
90
|
const items = [];
|
|
38
91
|
const runHandoffs = [];
|
|
39
92
|
const runFunctions = [];
|
|
40
93
|
const runComputerActions = [];
|
|
94
|
+
const runShellActions = [];
|
|
95
|
+
const runApplyPatchActions = [];
|
|
41
96
|
const runMCPApprovalRequests = [];
|
|
42
97
|
const toolsUsed = [];
|
|
43
98
|
const handoffMap = new Map(handoffs.map((h) => [h.toolName, h]));
|
|
99
|
+
// Resolve tools upfront so we can look up the concrete handler in O(1) while iterating outputs.
|
|
44
100
|
const functionMap = new Map(tools.filter((t) => t.type === 'function').map((t) => [t.name, t]));
|
|
45
101
|
const computerTool = tools.find((t) => t.type === 'computer');
|
|
102
|
+
const shellTool = tools.find((t) => t.type === 'shell');
|
|
103
|
+
const applyPatchTool = tools.find((t) => t.type === 'apply_patch');
|
|
46
104
|
const mcpToolMap = new Map(tools
|
|
47
105
|
.filter((t) => t.type === 'hosted_tool' && t.providerData?.type === 'mcp')
|
|
48
106
|
.map((t) => t)
|
|
@@ -112,6 +170,40 @@ function processModelResponse(modelResponse, agent, tools, handoffs) {
|
|
|
112
170
|
computer: computerTool,
|
|
113
171
|
});
|
|
114
172
|
}
|
|
173
|
+
else if (output.type === 'shell_call') {
|
|
174
|
+
items.push(new items_1.RunToolCallItem(output, agent));
|
|
175
|
+
toolsUsed.push('shell');
|
|
176
|
+
if (!shellTool) {
|
|
177
|
+
(0, context_1.addErrorToCurrentSpan)({
|
|
178
|
+
message: 'Model produced shell action without a shell tool.',
|
|
179
|
+
data: {
|
|
180
|
+
agent_name: agent.name,
|
|
181
|
+
},
|
|
182
|
+
});
|
|
183
|
+
throw new errors_1.ModelBehaviorError('Model produced shell action without a shell tool.');
|
|
184
|
+
}
|
|
185
|
+
runShellActions.push({
|
|
186
|
+
toolCall: output,
|
|
187
|
+
shell: shellTool,
|
|
188
|
+
});
|
|
189
|
+
}
|
|
190
|
+
else if (output.type === 'apply_patch_call') {
|
|
191
|
+
items.push(new items_1.RunToolCallItem(output, agent));
|
|
192
|
+
toolsUsed.push('apply_patch');
|
|
193
|
+
if (!applyPatchTool) {
|
|
194
|
+
(0, context_1.addErrorToCurrentSpan)({
|
|
195
|
+
message: 'Model produced apply_patch action without an apply_patch tool.',
|
|
196
|
+
data: {
|
|
197
|
+
agent_name: agent.name,
|
|
198
|
+
},
|
|
199
|
+
});
|
|
200
|
+
throw new errors_1.ModelBehaviorError('Model produced apply_patch action without an apply_patch tool.');
|
|
201
|
+
}
|
|
202
|
+
runApplyPatchActions.push({
|
|
203
|
+
toolCall: output,
|
|
204
|
+
applyPatch: applyPatchTool,
|
|
205
|
+
});
|
|
206
|
+
}
|
|
115
207
|
if (output.type !== 'function_call') {
|
|
116
208
|
continue;
|
|
117
209
|
}
|
|
@@ -148,13 +240,17 @@ function processModelResponse(modelResponse, agent, tools, handoffs) {
|
|
|
148
240
|
handoffs: runHandoffs,
|
|
149
241
|
functions: runFunctions,
|
|
150
242
|
computerActions: runComputerActions,
|
|
243
|
+
shellActions: runShellActions,
|
|
244
|
+
applyPatchActions: runApplyPatchActions,
|
|
151
245
|
mcpApprovalRequests: runMCPApprovalRequests,
|
|
152
246
|
toolsUsed: toolsUsed,
|
|
153
247
|
hasToolsOrApprovalsToRun() {
|
|
154
248
|
return (runHandoffs.length > 0 ||
|
|
155
249
|
runFunctions.length > 0 ||
|
|
156
250
|
runMCPApprovalRequests.length > 0 ||
|
|
157
|
-
runComputerActions.length > 0
|
|
251
|
+
runComputerActions.length > 0 ||
|
|
252
|
+
runShellActions.length > 0 ||
|
|
253
|
+
runApplyPatchActions.length > 0);
|
|
158
254
|
},
|
|
159
255
|
};
|
|
160
256
|
}
|
|
@@ -175,6 +271,10 @@ exports.nextStepSchema = zod_1.z.discriminatedUnion('type', [
|
|
|
175
271
|
data: zod_1.z.record(zod_1.z.string(), zod_1.z.any()),
|
|
176
272
|
}),
|
|
177
273
|
]);
|
|
274
|
+
/**
|
|
275
|
+
* Internal convenience wrapper that groups the outcome of a single agent turn. It lets the caller
|
|
276
|
+
* update the RunState in one shot and decide which step to execute next.
|
|
277
|
+
*/
|
|
178
278
|
class SingleStepResult {
|
|
179
279
|
originalInput;
|
|
180
280
|
modelResponse;
|
|
@@ -183,7 +283,7 @@ class SingleStepResult {
|
|
|
183
283
|
nextStep;
|
|
184
284
|
constructor(
|
|
185
285
|
/**
|
|
186
|
-
* The input items i.e
|
|
286
|
+
* The input items (i.e., the items before run() was called). May be mutated by handoff input filters.
|
|
187
287
|
*/
|
|
188
288
|
originalInput,
|
|
189
289
|
/**
|
|
@@ -217,6 +317,8 @@ class SingleStepResult {
|
|
|
217
317
|
}
|
|
218
318
|
/**
|
|
219
319
|
* @internal
|
|
320
|
+
* Resets the tool choice when the agent is configured to prefer a fresh tool selection after
|
|
321
|
+
* any tool usage. This prevents the provider from reusing stale tool hints across turns.
|
|
220
322
|
*/
|
|
221
323
|
function maybeResetToolChoice(agent, toolUseTracker, modelSettings) {
|
|
222
324
|
if (agent.resetToolChoice && toolUseTracker.hasUsedTools(agent)) {
|
|
@@ -226,33 +328,104 @@ function maybeResetToolChoice(agent, toolUseTracker, modelSettings) {
|
|
|
226
328
|
}
|
|
227
329
|
/**
|
|
228
330
|
* @internal
|
|
331
|
+
* Continues a turn that was previously interrupted waiting for tool approval. Executes the now
|
|
332
|
+
* approved tools and returns the resulting step transition.
|
|
229
333
|
*/
|
|
230
|
-
async function
|
|
334
|
+
async function resolveInterruptedTurn(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
|
|
231
335
|
// call_ids for function tools
|
|
232
336
|
const functionCallIds = originalPreStepItems
|
|
233
337
|
.filter((item) => item instanceof items_1.RunToolApprovalItem &&
|
|
234
338
|
'callId' in item.rawItem &&
|
|
235
339
|
item.rawItem.type === 'function_call')
|
|
236
340
|
.map((item) => item.rawItem.callId);
|
|
341
|
+
// We already persisted the turn once when the approval interrupt was raised, so the
|
|
342
|
+
// counter reflects the approval items as "flushed". When we resume the same turn we need
|
|
343
|
+
// to rewind it so the eventual tool output for this call is still written to the session.
|
|
344
|
+
const pendingApprovalItems = state
|
|
345
|
+
.getInterruptions()
|
|
346
|
+
.filter(isApprovalItemLike);
|
|
347
|
+
if (pendingApprovalItems.length > 0) {
|
|
348
|
+
const pendingApprovalIdentities = new Set();
|
|
349
|
+
for (const approval of pendingApprovalItems) {
|
|
350
|
+
const identity = getApprovalIdentity(approval);
|
|
351
|
+
if (identity) {
|
|
352
|
+
pendingApprovalIdentities.add(identity);
|
|
353
|
+
}
|
|
354
|
+
}
|
|
355
|
+
if (pendingApprovalIdentities.size > 0) {
|
|
356
|
+
let rewindCount = 0;
|
|
357
|
+
for (let index = originalPreStepItems.length - 1; index >= 0; index--) {
|
|
358
|
+
const item = originalPreStepItems[index];
|
|
359
|
+
if (!(item instanceof items_1.RunToolApprovalItem)) {
|
|
360
|
+
continue;
|
|
361
|
+
}
|
|
362
|
+
const identity = getApprovalIdentity(item);
|
|
363
|
+
if (!identity) {
|
|
364
|
+
continue;
|
|
365
|
+
}
|
|
366
|
+
if (!pendingApprovalIdentities.has(identity)) {
|
|
367
|
+
continue;
|
|
368
|
+
}
|
|
369
|
+
rewindCount++;
|
|
370
|
+
pendingApprovalIdentities.delete(identity);
|
|
371
|
+
if (pendingApprovalIdentities.size === 0) {
|
|
372
|
+
break;
|
|
373
|
+
}
|
|
374
|
+
}
|
|
375
|
+
// Persisting the approval request already advanced the counter once, so undo the increment
|
|
376
|
+
// to make sure we write the final tool output back to the session when the turn resumes.
|
|
377
|
+
if (rewindCount > 0) {
|
|
378
|
+
state._currentTurnPersistedItemCount = Math.max(0, state._currentTurnPersistedItemCount - rewindCount);
|
|
379
|
+
}
|
|
380
|
+
}
|
|
381
|
+
}
|
|
237
382
|
// Run function tools that require approval after they get their approval results
|
|
238
383
|
const functionToolRuns = processedResponse.functions.filter((run) => {
|
|
239
384
|
return functionCallIds.includes(run.toolCall.callId);
|
|
240
385
|
});
|
|
241
386
|
const functionResults = await executeFunctionToolCalls(agent, functionToolRuns, runner, state);
|
|
242
|
-
//
|
|
243
|
-
|
|
387
|
+
// There is no built-in HITL approval surface for computer tools today, so every pending action
|
|
388
|
+
// is executed immediately when the turn resumes.
|
|
389
|
+
const computerResults = processedResponse.computerActions.length > 0
|
|
390
|
+
? await executeComputerActions(agent, processedResponse.computerActions, runner, state._context)
|
|
391
|
+
: [];
|
|
392
|
+
// When resuming we receive the original RunItem references; suppress duplicates so history and streaming do not double-emit the same items.
|
|
393
|
+
const originalPreStepItemSet = new Set(originalPreStepItems);
|
|
394
|
+
const newItems = [];
|
|
395
|
+
const newItemsSet = new Set();
|
|
396
|
+
const appendIfNew = (item) => {
|
|
397
|
+
if (originalPreStepItemSet.has(item) || newItemsSet.has(item)) {
|
|
398
|
+
return;
|
|
399
|
+
}
|
|
400
|
+
newItems.push(item);
|
|
401
|
+
newItemsSet.add(item);
|
|
402
|
+
};
|
|
403
|
+
for (const result of functionResults) {
|
|
404
|
+
appendIfNew(result.runItem);
|
|
405
|
+
}
|
|
406
|
+
for (const result of computerResults) {
|
|
407
|
+
appendIfNew(result);
|
|
408
|
+
}
|
|
244
409
|
// Run MCP tools that require approval after they get their approval results
|
|
245
410
|
const mcpApprovalRuns = processedResponse.mcpApprovalRequests.filter((run) => {
|
|
246
411
|
return (run.requestItem.type === 'tool_approval_item' &&
|
|
247
412
|
run.requestItem.rawItem.type === 'hosted_tool_call' &&
|
|
248
413
|
run.requestItem.rawItem.providerData?.type === 'mcp_approval_request');
|
|
249
414
|
});
|
|
415
|
+
// Hosted MCP approvals may still be waiting on a human decision when the turn resumes.
|
|
416
|
+
const pendingHostedMCPApprovals = new Set();
|
|
417
|
+
const pendingHostedMCPApprovalIds = new Set();
|
|
418
|
+
// Keep track of approvals we still need to surface next turn so HITL flows can resume cleanly.
|
|
250
419
|
for (const run of mcpApprovalRuns) {
|
|
251
420
|
// the approval_request_id "mcpr_123..."
|
|
252
|
-
const
|
|
421
|
+
const rawItem = run.requestItem.rawItem;
|
|
422
|
+
if (rawItem.type !== 'hosted_tool_call') {
|
|
423
|
+
continue;
|
|
424
|
+
}
|
|
425
|
+
const approvalRequestId = rawItem.id;
|
|
253
426
|
const approved = state._context.isToolApproved({
|
|
254
427
|
// Since this item name must be the same with the one sent from Responses API server
|
|
255
|
-
toolName:
|
|
428
|
+
toolName: rawItem.name,
|
|
256
429
|
callId: approvalRequestId,
|
|
257
430
|
});
|
|
258
431
|
if (typeof approved !== 'undefined') {
|
|
@@ -262,50 +435,101 @@ async function executeInterruptedToolsAndSideEffects(agent, originalInput, origi
|
|
|
262
435
|
reason: undefined,
|
|
263
436
|
};
|
|
264
437
|
// Tell Responses API server the approval result in the next turn
|
|
265
|
-
|
|
438
|
+
const responseItem = new items_1.RunToolCallItem({
|
|
266
439
|
type: 'hosted_tool_call',
|
|
267
440
|
name: 'mcp_approval_response',
|
|
268
441
|
providerData,
|
|
269
|
-
}, agent)
|
|
442
|
+
}, agent);
|
|
443
|
+
appendIfNew(responseItem);
|
|
444
|
+
}
|
|
445
|
+
else {
|
|
446
|
+
pendingHostedMCPApprovals.add(run.requestItem);
|
|
447
|
+
pendingHostedMCPApprovalIds.add(approvalRequestId);
|
|
448
|
+
functionResults.push({
|
|
449
|
+
type: 'hosted_mcp_tool_approval',
|
|
450
|
+
tool: run.mcpTool,
|
|
451
|
+
runItem: run.requestItem,
|
|
452
|
+
});
|
|
453
|
+
appendIfNew(run.requestItem);
|
|
270
454
|
}
|
|
271
455
|
}
|
|
272
|
-
|
|
273
|
-
//
|
|
274
|
-
//
|
|
456
|
+
// Server-managed conversations rely on preStepItems to re-surface pending approvals.
|
|
457
|
+
// Keep unresolved hosted MCP approvals in place so HITL flows still have something to approve next turn.
|
|
458
|
+
// Drop resolved approval placeholders so they are not replayed on the next turn, but keep
|
|
459
|
+
// pending approvals in place to signal the outstanding work to the UI and session store.
|
|
275
460
|
const preStepItems = originalPreStepItems.filter((item) => {
|
|
276
|
-
|
|
461
|
+
if (!(item instanceof items_1.RunToolApprovalItem)) {
|
|
462
|
+
return true;
|
|
463
|
+
}
|
|
464
|
+
if (item.rawItem.type === 'hosted_tool_call' &&
|
|
465
|
+
item.rawItem.providerData?.type === 'mcp_approval_request') {
|
|
466
|
+
if (pendingHostedMCPApprovals.has(item)) {
|
|
467
|
+
return true;
|
|
468
|
+
}
|
|
469
|
+
const approvalRequestId = item.rawItem.id;
|
|
470
|
+
if (approvalRequestId) {
|
|
471
|
+
return pendingHostedMCPApprovalIds.has(approvalRequestId);
|
|
472
|
+
}
|
|
473
|
+
return false;
|
|
474
|
+
}
|
|
475
|
+
return false;
|
|
277
476
|
});
|
|
278
|
-
|
|
279
|
-
|
|
280
|
-
|
|
281
|
-
|
|
282
|
-
|
|
283
|
-
|
|
284
|
-
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
|
|
288
|
-
|
|
289
|
-
|
|
290
|
-
interruptions: checkToolOutput.interruptions,
|
|
291
|
-
},
|
|
292
|
-
});
|
|
477
|
+
const completedStep = await maybeCompleteTurnFromToolResults({
|
|
478
|
+
agent,
|
|
479
|
+
runner,
|
|
480
|
+
state,
|
|
481
|
+
functionResults,
|
|
482
|
+
originalInput,
|
|
483
|
+
newResponse,
|
|
484
|
+
preStepItems,
|
|
485
|
+
newItems,
|
|
486
|
+
});
|
|
487
|
+
if (completedStep) {
|
|
488
|
+
return completedStep;
|
|
293
489
|
}
|
|
294
490
|
// we only ran new tools and side effects. We need to run the rest of the agent
|
|
295
491
|
return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
|
|
296
492
|
}
|
|
297
493
|
/**
|
|
298
494
|
* @internal
|
|
495
|
+
* Executes every follow-up action the model requested (function tools, computer actions, MCP flows),
|
|
496
|
+
* appends their outputs to the run history, and determines the next step for the agent loop.
|
|
299
497
|
*/
|
|
300
|
-
async function
|
|
498
|
+
async function resolveTurnAfterModelResponse(agent, originalInput, originalPreStepItems, newResponse, processedResponse, runner, state) {
|
|
499
|
+
// Reuse the same array reference so we can compare object identity when deciding whether to
|
|
500
|
+
// append new items, ensuring we never double-stream existing RunItems.
|
|
301
501
|
const preStepItems = originalPreStepItems;
|
|
302
|
-
|
|
303
|
-
const
|
|
502
|
+
const seenItems = new Set(originalPreStepItems);
|
|
503
|
+
const newItems = [];
|
|
504
|
+
const appendIfNew = (item) => {
|
|
505
|
+
if (seenItems.has(item)) {
|
|
506
|
+
return;
|
|
507
|
+
}
|
|
508
|
+
newItems.push(item);
|
|
509
|
+
seenItems.add(item);
|
|
510
|
+
};
|
|
511
|
+
for (const item of processedResponse.newItems) {
|
|
512
|
+
appendIfNew(item);
|
|
513
|
+
}
|
|
514
|
+
// Run function tools and computer actions in parallel; neither depends on the other's side effects.
|
|
515
|
+
const [functionResults, computerResults, shellResults, applyPatchResults] = await Promise.all([
|
|
304
516
|
executeFunctionToolCalls(agent, processedResponse.functions, runner, state),
|
|
305
517
|
executeComputerActions(agent, processedResponse.computerActions, runner, state._context),
|
|
518
|
+
executeShellActions(agent, processedResponse.shellActions, runner, state._context),
|
|
519
|
+
executeApplyPatchOperations(agent, processedResponse.applyPatchActions, runner, state._context),
|
|
306
520
|
]);
|
|
307
|
-
|
|
308
|
-
|
|
521
|
+
for (const result of functionResults) {
|
|
522
|
+
appendIfNew(result.runItem);
|
|
523
|
+
}
|
|
524
|
+
for (const item of computerResults) {
|
|
525
|
+
appendIfNew(item);
|
|
526
|
+
}
|
|
527
|
+
for (const item of shellResults) {
|
|
528
|
+
appendIfNew(item);
|
|
529
|
+
}
|
|
530
|
+
for (const item of applyPatchResults) {
|
|
531
|
+
appendIfNew(item);
|
|
532
|
+
}
|
|
309
533
|
// run hosted MCP approval requests
|
|
310
534
|
if (processedResponse.mcpApprovalRequests.length > 0) {
|
|
311
535
|
for (const approvalRequest of processedResponse.mcpApprovalRequests) {
|
|
@@ -351,28 +575,26 @@ async function executeToolsAndSideEffects(agent, originalInput, originalPreStepI
|
|
|
351
575
|
if (processedResponse.handoffs.length > 0) {
|
|
352
576
|
return await executeHandoffCalls(agent, originalInput, preStepItems, newItems, newResponse, processedResponse.handoffs, runner, state._context);
|
|
353
577
|
}
|
|
354
|
-
const
|
|
355
|
-
|
|
356
|
-
runner
|
|
357
|
-
|
|
358
|
-
|
|
359
|
-
|
|
360
|
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
|
|
364
|
-
|
|
365
|
-
|
|
366
|
-
data: {
|
|
367
|
-
interruptions: checkToolOutput.interruptions,
|
|
368
|
-
},
|
|
369
|
-
});
|
|
578
|
+
const completedStep = await maybeCompleteTurnFromToolResults({
|
|
579
|
+
agent,
|
|
580
|
+
runner,
|
|
581
|
+
state,
|
|
582
|
+
functionResults,
|
|
583
|
+
originalInput,
|
|
584
|
+
newResponse,
|
|
585
|
+
preStepItems,
|
|
586
|
+
newItems,
|
|
587
|
+
});
|
|
588
|
+
if (completedStep) {
|
|
589
|
+
return completedStep;
|
|
370
590
|
}
|
|
371
591
|
// If the model issued any tool calls or handoffs in this turn,
|
|
372
592
|
// we must NOT treat any assistant message in the same turn as the final output.
|
|
373
593
|
// We should run the loop again so the model can see the tool results and respond.
|
|
374
594
|
const hadToolCallsOrActions = (processedResponse.functions?.length ?? 0) > 0 ||
|
|
375
595
|
(processedResponse.computerActions?.length ?? 0) > 0 ||
|
|
596
|
+
(processedResponse.shellActions?.length ?? 0) > 0 ||
|
|
597
|
+
(processedResponse.applyPatchActions?.length ?? 0) > 0 ||
|
|
376
598
|
(processedResponse.mcpApprovalRequests?.length ?? 0) > 0 ||
|
|
377
599
|
(processedResponse.handoffs?.length ?? 0) > 0;
|
|
378
600
|
if (hadToolCallsOrActions) {
|
|
@@ -388,6 +610,7 @@ async function executeToolsAndSideEffects(agent, originalInput, originalPreStepI
|
|
|
388
610
|
if (typeof potentialFinalOutput === 'undefined') {
|
|
389
611
|
return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
|
|
390
612
|
}
|
|
613
|
+
// Keep looping if any tool output placeholders still require an approval follow-up.
|
|
391
614
|
const hasPendingToolsOrApprovals = functionResults.some((result) => result.runItem instanceof items_1.RunToolApprovalItem);
|
|
392
615
|
if (!hasPendingToolsOrApprovals) {
|
|
393
616
|
if (agent.outputType === 'text') {
|
|
@@ -414,8 +637,32 @@ async function executeToolsAndSideEffects(agent, originalInput, originalPreStepI
|
|
|
414
637
|
}
|
|
415
638
|
return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, { type: 'next_step_run_again' });
|
|
416
639
|
}
|
|
640
|
+
// Consolidates the logic that determines whether tool results yielded a final answer,
|
|
641
|
+
// triggered an interruption, or require the agent loop to continue running.
|
|
642
|
+
async function maybeCompleteTurnFromToolResults({ agent, runner, state, functionResults, originalInput, newResponse, preStepItems, newItems, }) {
|
|
643
|
+
const toolOutcome = await checkForFinalOutputFromTools(agent, functionResults, state);
|
|
644
|
+
if (toolOutcome.isFinalOutput) {
|
|
645
|
+
runner.emit('agent_end', state._context, agent, toolOutcome.finalOutput);
|
|
646
|
+
agent.emit('agent_end', state._context, toolOutcome.finalOutput);
|
|
647
|
+
return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
|
|
648
|
+
type: 'next_step_final_output',
|
|
649
|
+
output: toolOutcome.finalOutput,
|
|
650
|
+
});
|
|
651
|
+
}
|
|
652
|
+
if (toolOutcome.isInterrupted) {
|
|
653
|
+
return new SingleStepResult(originalInput, newResponse, preStepItems, newItems, {
|
|
654
|
+
type: 'next_step_interruption',
|
|
655
|
+
data: {
|
|
656
|
+
interruptions: toolOutcome.interruptions,
|
|
657
|
+
},
|
|
658
|
+
});
|
|
659
|
+
}
|
|
660
|
+
return null;
|
|
661
|
+
}
|
|
417
662
|
/**
|
|
418
663
|
* @internal
|
|
664
|
+
* Normalizes tool outputs once so downstream code works with fully structured protocol items.
|
|
665
|
+
* Doing this here keeps API surface stable even when providers add new shapes.
|
|
419
666
|
*/
|
|
420
667
|
function getToolCallOutputItem(toolCall, output) {
|
|
421
668
|
const maybeStructuredOutputs = normalizeStructuredToolOutputs(output);
|
|
@@ -440,283 +687,57 @@ function getToolCallOutputItem(toolCall, output) {
|
|
|
440
687
|
},
|
|
441
688
|
};
|
|
442
689
|
}
|
|
443
|
-
|
|
444
|
-
|
|
445
|
-
|
|
446
|
-
|
|
447
|
-
* back to the legacy string pipeline.
|
|
448
|
-
*/
|
|
449
|
-
function normalizeStructuredToolOutputs(output) {
|
|
450
|
-
if (Array.isArray(output)) {
|
|
451
|
-
const structured = [];
|
|
452
|
-
for (const item of output) {
|
|
453
|
-
const normalized = normalizeStructuredToolOutput(item);
|
|
454
|
-
if (!normalized) {
|
|
455
|
-
return null;
|
|
456
|
-
}
|
|
457
|
-
structured.push(normalized);
|
|
458
|
-
}
|
|
459
|
-
return structured;
|
|
690
|
+
function normalizeFileValue(value) {
|
|
691
|
+
const directFile = value.file;
|
|
692
|
+
if (typeof directFile === 'string' && directFile.length > 0) {
|
|
693
|
+
return directFile;
|
|
460
694
|
}
|
|
461
|
-
const
|
|
462
|
-
|
|
695
|
+
const normalizedObject = normalizeFileObjectCandidate(directFile);
|
|
696
|
+
if (normalizedObject) {
|
|
697
|
+
return normalizedObject;
|
|
698
|
+
}
|
|
699
|
+
const legacyValue = normalizeLegacyFileValue(value);
|
|
700
|
+
if (legacyValue) {
|
|
701
|
+
return legacyValue;
|
|
702
|
+
}
|
|
703
|
+
return null;
|
|
463
704
|
}
|
|
464
|
-
|
|
465
|
-
* Best-effort normalization of a single tool output item. If the object already matches the
|
|
466
|
-
* protocol shape we simply cast it; otherwise we copy the recognised fields into the canonical
|
|
467
|
-
* structure. Returning null lets the caller know we should revert to plain-string handling.
|
|
468
|
-
*/
|
|
469
|
-
function normalizeStructuredToolOutput(value) {
|
|
705
|
+
function normalizeFileObjectCandidate(value) {
|
|
470
706
|
if (!isRecord(value)) {
|
|
471
707
|
return null;
|
|
472
708
|
}
|
|
473
|
-
|
|
474
|
-
|
|
475
|
-
const
|
|
476
|
-
|
|
477
|
-
|
|
478
|
-
}
|
|
479
|
-
return output;
|
|
480
|
-
}
|
|
481
|
-
if (type === 'image') {
|
|
482
|
-
const output = { type: 'image' };
|
|
483
|
-
let imageString;
|
|
484
|
-
let imageFileId;
|
|
485
|
-
const fallbackImageMediaType = isNonEmptyString(value.mediaType)
|
|
486
|
-
? value.mediaType
|
|
487
|
-
: undefined;
|
|
488
|
-
const imageField = value.image;
|
|
489
|
-
if (typeof imageField === 'string' && imageField.length > 0) {
|
|
490
|
-
imageString = imageField;
|
|
491
|
-
}
|
|
492
|
-
else if (isRecord(imageField)) {
|
|
493
|
-
const imageObj = imageField;
|
|
494
|
-
const inlineMediaType = isNonEmptyString(imageObj.mediaType)
|
|
495
|
-
? imageObj.mediaType
|
|
496
|
-
: fallbackImageMediaType;
|
|
497
|
-
if (isNonEmptyString(imageObj.url)) {
|
|
498
|
-
imageString = imageObj.url;
|
|
499
|
-
}
|
|
500
|
-
else if (isNonEmptyString(imageObj.data)) {
|
|
501
|
-
imageString = toInlineImageString(imageObj.data, inlineMediaType);
|
|
502
|
-
}
|
|
503
|
-
else if (imageObj.data instanceof Uint8Array &&
|
|
504
|
-
imageObj.data.length > 0) {
|
|
505
|
-
imageString = toInlineImageString(imageObj.data, inlineMediaType);
|
|
506
|
-
}
|
|
507
|
-
if (!imageString) {
|
|
508
|
-
const candidateId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
|
|
509
|
-
(isNonEmptyString(imageObj.id) && imageObj.id) ||
|
|
510
|
-
undefined;
|
|
511
|
-
if (candidateId) {
|
|
512
|
-
imageFileId = candidateId;
|
|
513
|
-
}
|
|
514
|
-
}
|
|
515
|
-
}
|
|
516
|
-
if (!imageString &&
|
|
517
|
-
typeof value.imageUrl === 'string' &&
|
|
518
|
-
value.imageUrl.length > 0) {
|
|
519
|
-
imageString = value.imageUrl;
|
|
520
|
-
}
|
|
521
|
-
if (!imageFileId &&
|
|
522
|
-
typeof value.fileId === 'string' &&
|
|
523
|
-
value.fileId.length > 0) {
|
|
524
|
-
imageFileId = value.fileId;
|
|
525
|
-
}
|
|
526
|
-
if (!imageString &&
|
|
527
|
-
typeof value.data === 'string' &&
|
|
528
|
-
value.data.length > 0) {
|
|
529
|
-
imageString = fallbackImageMediaType
|
|
530
|
-
? toInlineImageString(value.data, fallbackImageMediaType)
|
|
531
|
-
: value.data;
|
|
532
|
-
}
|
|
533
|
-
else if (!imageString &&
|
|
534
|
-
value.data instanceof Uint8Array &&
|
|
535
|
-
value.data.length > 0) {
|
|
536
|
-
imageString = toInlineImageString(value.data, fallbackImageMediaType);
|
|
537
|
-
}
|
|
538
|
-
if (typeof value.detail === 'string' && value.detail.length > 0) {
|
|
539
|
-
output.detail = value.detail;
|
|
540
|
-
}
|
|
541
|
-
if (imageString) {
|
|
542
|
-
output.image = imageString;
|
|
543
|
-
}
|
|
544
|
-
else if (imageFileId) {
|
|
545
|
-
output.image = { fileId: imageFileId };
|
|
546
|
-
}
|
|
547
|
-
else {
|
|
709
|
+
if ('data' in value && value.data !== undefined) {
|
|
710
|
+
const dataValue = value.data;
|
|
711
|
+
const hasStringData = typeof dataValue === 'string' && dataValue.length > 0;
|
|
712
|
+
const hasBinaryData = dataValue instanceof Uint8Array && dataValue.length > 0;
|
|
713
|
+
if (!hasStringData && !hasBinaryData) {
|
|
548
714
|
return null;
|
|
549
715
|
}
|
|
550
|
-
if (
|
|
551
|
-
|
|
716
|
+
if (!isNonEmptyString(value.mediaType) ||
|
|
717
|
+
!isNonEmptyString(value.filename)) {
|
|
718
|
+
return null;
|
|
552
719
|
}
|
|
553
|
-
return
|
|
720
|
+
return {
|
|
721
|
+
data: typeof dataValue === 'string' ? dataValue : new Uint8Array(dataValue),
|
|
722
|
+
mediaType: value.mediaType,
|
|
723
|
+
filename: value.filename,
|
|
724
|
+
};
|
|
554
725
|
}
|
|
555
|
-
if (
|
|
556
|
-
const
|
|
557
|
-
if (
|
|
558
|
-
|
|
726
|
+
if (isNonEmptyString(value.url)) {
|
|
727
|
+
const result = { url: value.url };
|
|
728
|
+
if (isNonEmptyString(value.filename)) {
|
|
729
|
+
result.filename = value.filename;
|
|
559
730
|
}
|
|
560
|
-
|
|
561
|
-
|
|
562
|
-
|
|
731
|
+
return result;
|
|
732
|
+
}
|
|
733
|
+
const referencedId = (isNonEmptyString(value.id) && value.id) ||
|
|
734
|
+
(isNonEmptyString(value.fileId) && value.fileId);
|
|
735
|
+
if (referencedId) {
|
|
736
|
+
const result = { id: referencedId };
|
|
737
|
+
if (isNonEmptyString(value.filename)) {
|
|
738
|
+
result.filename = value.filename;
|
|
563
739
|
}
|
|
564
|
-
return
|
|
565
|
-
}
|
|
566
|
-
return null;
|
|
567
|
-
}
|
|
568
|
-
/**
|
|
569
|
-
* Translates the normalized tool output into the protocol `input_*` items. This is the last hop
|
|
570
|
-
* before we hand the data to model-specific adapters, so we generate the exact schema expected by
|
|
571
|
-
* the protocol definitions.
|
|
572
|
-
*/
|
|
573
|
-
function convertStructuredToolOutputToInputItem(output) {
|
|
574
|
-
if (output.type === 'text') {
|
|
575
|
-
const result = {
|
|
576
|
-
type: 'input_text',
|
|
577
|
-
text: output.text,
|
|
578
|
-
};
|
|
579
|
-
if (output.providerData) {
|
|
580
|
-
result.providerData = output.providerData;
|
|
581
|
-
}
|
|
582
|
-
return result;
|
|
583
|
-
}
|
|
584
|
-
if (output.type === 'image') {
|
|
585
|
-
const result = { type: 'input_image' };
|
|
586
|
-
if (typeof output.detail === 'string' && output.detail.length > 0) {
|
|
587
|
-
result.detail = output.detail;
|
|
588
|
-
}
|
|
589
|
-
if (typeof output.image === 'string' && output.image.length > 0) {
|
|
590
|
-
result.image = output.image;
|
|
591
|
-
}
|
|
592
|
-
else if (isRecord(output.image)) {
|
|
593
|
-
const imageObj = output.image;
|
|
594
|
-
const inlineMediaType = isNonEmptyString(imageObj.mediaType)
|
|
595
|
-
? imageObj.mediaType
|
|
596
|
-
: undefined;
|
|
597
|
-
if (isNonEmptyString(imageObj.url)) {
|
|
598
|
-
result.image = imageObj.url;
|
|
599
|
-
}
|
|
600
|
-
else if (isNonEmptyString(imageObj.data)) {
|
|
601
|
-
result.image =
|
|
602
|
-
inlineMediaType && !imageObj.data.startsWith('data:')
|
|
603
|
-
? asDataUrl(imageObj.data, inlineMediaType)
|
|
604
|
-
: imageObj.data;
|
|
605
|
-
}
|
|
606
|
-
else if (imageObj.data instanceof Uint8Array &&
|
|
607
|
-
imageObj.data.length > 0) {
|
|
608
|
-
const base64 = (0, base64_1.encodeUint8ArrayToBase64)(imageObj.data);
|
|
609
|
-
result.image = asDataUrl(base64, inlineMediaType);
|
|
610
|
-
}
|
|
611
|
-
else {
|
|
612
|
-
const referencedId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
|
|
613
|
-
(isNonEmptyString(imageObj.id) && imageObj.id) ||
|
|
614
|
-
undefined;
|
|
615
|
-
if (referencedId) {
|
|
616
|
-
result.image = { id: referencedId };
|
|
617
|
-
}
|
|
618
|
-
}
|
|
619
|
-
}
|
|
620
|
-
if (output.providerData) {
|
|
621
|
-
result.providerData = output.providerData;
|
|
622
|
-
}
|
|
623
|
-
return result;
|
|
624
|
-
}
|
|
625
|
-
if (output.type === 'file') {
|
|
626
|
-
const result = { type: 'input_file' };
|
|
627
|
-
const fileValue = output.file;
|
|
628
|
-
if (typeof fileValue === 'string') {
|
|
629
|
-
result.file = fileValue;
|
|
630
|
-
}
|
|
631
|
-
else if (fileValue && typeof fileValue === 'object') {
|
|
632
|
-
const record = fileValue;
|
|
633
|
-
if ('data' in record && record.data) {
|
|
634
|
-
const mediaType = record.mediaType ?? 'text/plain';
|
|
635
|
-
if (typeof record.data === 'string') {
|
|
636
|
-
result.file = asDataUrl(record.data, mediaType);
|
|
637
|
-
}
|
|
638
|
-
else {
|
|
639
|
-
const base64 = (0, base64_1.encodeUint8ArrayToBase64)(record.data);
|
|
640
|
-
result.file = asDataUrl(base64, mediaType);
|
|
641
|
-
}
|
|
642
|
-
}
|
|
643
|
-
else if (typeof record.url === 'string' && record.url.length > 0) {
|
|
644
|
-
result.file = { url: record.url };
|
|
645
|
-
}
|
|
646
|
-
else {
|
|
647
|
-
const referencedId = (typeof record.id === 'string' &&
|
|
648
|
-
record.id.length > 0 &&
|
|
649
|
-
record.id) ||
|
|
650
|
-
(typeof record.fileId === 'string' && record.fileId.length > 0
|
|
651
|
-
? record.fileId
|
|
652
|
-
: undefined);
|
|
653
|
-
if (referencedId) {
|
|
654
|
-
result.file = { id: referencedId };
|
|
655
|
-
}
|
|
656
|
-
}
|
|
657
|
-
if (typeof record.filename === 'string' && record.filename.length > 0) {
|
|
658
|
-
result.filename = record.filename;
|
|
659
|
-
}
|
|
660
|
-
}
|
|
661
|
-
if (output.providerData) {
|
|
662
|
-
result.providerData = output.providerData;
|
|
663
|
-
}
|
|
664
|
-
return result;
|
|
665
|
-
}
|
|
666
|
-
const exhaustiveCheck = output;
|
|
667
|
-
return exhaustiveCheck;
|
|
668
|
-
}
|
|
669
|
-
function normalizeFileValue(value) {
|
|
670
|
-
const directFile = value.file;
|
|
671
|
-
if (typeof directFile === 'string' && directFile.length > 0) {
|
|
672
|
-
return directFile;
|
|
673
|
-
}
|
|
674
|
-
const normalizedObject = normalizeFileObjectCandidate(directFile);
|
|
675
|
-
if (normalizedObject) {
|
|
676
|
-
return normalizedObject;
|
|
677
|
-
}
|
|
678
|
-
const legacyValue = normalizeLegacyFileValue(value);
|
|
679
|
-
if (legacyValue) {
|
|
680
|
-
return legacyValue;
|
|
681
|
-
}
|
|
682
|
-
return null;
|
|
683
|
-
}
|
|
684
|
-
function normalizeFileObjectCandidate(value) {
|
|
685
|
-
if (!isRecord(value)) {
|
|
686
|
-
return null;
|
|
687
|
-
}
|
|
688
|
-
if ('data' in value && value.data !== undefined) {
|
|
689
|
-
const dataValue = value.data;
|
|
690
|
-
const hasStringData = typeof dataValue === 'string' && dataValue.length > 0;
|
|
691
|
-
const hasBinaryData = dataValue instanceof Uint8Array && dataValue.length > 0;
|
|
692
|
-
if (!hasStringData && !hasBinaryData) {
|
|
693
|
-
return null;
|
|
694
|
-
}
|
|
695
|
-
if (!isNonEmptyString(value.mediaType) ||
|
|
696
|
-
!isNonEmptyString(value.filename)) {
|
|
697
|
-
return null;
|
|
698
|
-
}
|
|
699
|
-
return {
|
|
700
|
-
data: typeof dataValue === 'string' ? dataValue : new Uint8Array(dataValue),
|
|
701
|
-
mediaType: value.mediaType,
|
|
702
|
-
filename: value.filename,
|
|
703
|
-
};
|
|
704
|
-
}
|
|
705
|
-
if (isNonEmptyString(value.url)) {
|
|
706
|
-
const result = { url: value.url };
|
|
707
|
-
if (isNonEmptyString(value.filename)) {
|
|
708
|
-
result.filename = value.filename;
|
|
709
|
-
}
|
|
710
|
-
return result;
|
|
711
|
-
}
|
|
712
|
-
const referencedId = (isNonEmptyString(value.id) && value.id) ||
|
|
713
|
-
(isNonEmptyString(value.fileId) && value.fileId);
|
|
714
|
-
if (referencedId) {
|
|
715
|
-
const result = { id: referencedId };
|
|
716
|
-
if (isNonEmptyString(value.filename)) {
|
|
717
|
-
result.filename = value.filename;
|
|
718
|
-
}
|
|
719
|
-
return result;
|
|
740
|
+
return result;
|
|
720
741
|
}
|
|
721
742
|
return null;
|
|
722
743
|
}
|
|
@@ -776,6 +797,8 @@ function asDataUrl(base64, mediaType) {
|
|
|
776
797
|
}
|
|
777
798
|
/**
|
|
778
799
|
* @internal
|
|
800
|
+
* Runs every function tool call requested by the model and returns their outputs alongside
|
|
801
|
+
* the `RunItem` instances that should be appended to history.
|
|
779
802
|
*/
|
|
780
803
|
async function executeFunctionToolCalls(agent, toolRuns, runner, state) {
|
|
781
804
|
async function runSingleTool(toolRun) {
|
|
@@ -788,6 +811,7 @@ async function executeFunctionToolCalls(agent, toolRuns, runner, state) {
|
|
|
788
811
|
parsedArgs = JSON.parse(parsedArgs);
|
|
789
812
|
}
|
|
790
813
|
}
|
|
814
|
+
// Some tools require a human or policy check before execution; defer until approval is recorded.
|
|
791
815
|
const needsApproval = await toolRun.tool.needsApproval(state._context, parsedArgs, toolRun.toolCall.callId);
|
|
792
816
|
if (needsApproval) {
|
|
793
817
|
const approval = state._context.isToolApproved({
|
|
@@ -939,8 +963,210 @@ async function _runComputerActionAndScreenshot(computer, toolCall) {
|
|
|
939
963
|
}
|
|
940
964
|
throw new Error('Computer does not implement screenshot()');
|
|
941
965
|
}
|
|
966
|
+
function toErrorMessage(error) {
|
|
967
|
+
if (error instanceof Error) {
|
|
968
|
+
return error.message || error.toString();
|
|
969
|
+
}
|
|
970
|
+
try {
|
|
971
|
+
return JSON.stringify(error);
|
|
972
|
+
}
|
|
973
|
+
catch {
|
|
974
|
+
return String(error);
|
|
975
|
+
}
|
|
976
|
+
}
|
|
977
|
+
async function executeShellActions(agent, actions, runner, runContext, customLogger = undefined) {
|
|
978
|
+
const _logger = customLogger ?? logger_1.default;
|
|
979
|
+
const results = [];
|
|
980
|
+
for (const action of actions) {
|
|
981
|
+
const shellTool = action.shell;
|
|
982
|
+
const toolCall = action.toolCall;
|
|
983
|
+
const approvalItem = new items_1.RunToolApprovalItem(toolCall, agent, shellTool.name);
|
|
984
|
+
const requiresApproval = await shellTool.needsApproval(runContext, toolCall.action, toolCall.callId);
|
|
985
|
+
if (requiresApproval) {
|
|
986
|
+
if (shellTool.onApproval) {
|
|
987
|
+
const decision = await shellTool.onApproval(runContext, approvalItem);
|
|
988
|
+
if (decision.approve === true) {
|
|
989
|
+
runContext.approveTool(approvalItem);
|
|
990
|
+
}
|
|
991
|
+
else if (decision.approve === false) {
|
|
992
|
+
runContext.rejectTool(approvalItem);
|
|
993
|
+
}
|
|
994
|
+
}
|
|
995
|
+
const approval = runContext.isToolApproved({
|
|
996
|
+
toolName: shellTool.name,
|
|
997
|
+
callId: toolCall.callId,
|
|
998
|
+
});
|
|
999
|
+
if (approval === false) {
|
|
1000
|
+
const response = 'Tool execution was not approved.';
|
|
1001
|
+
const rejectionOutput = {
|
|
1002
|
+
stdout: '',
|
|
1003
|
+
stderr: response,
|
|
1004
|
+
outcome: { type: 'exit', exitCode: null },
|
|
1005
|
+
};
|
|
1006
|
+
results.push(new items_1.RunToolCallOutputItem({
|
|
1007
|
+
type: 'shell_call_output',
|
|
1008
|
+
callId: toolCall.callId,
|
|
1009
|
+
output: [rejectionOutput],
|
|
1010
|
+
}, agent, response));
|
|
1011
|
+
continue;
|
|
1012
|
+
}
|
|
1013
|
+
if (approval !== true) {
|
|
1014
|
+
results.push(approvalItem);
|
|
1015
|
+
continue;
|
|
1016
|
+
}
|
|
1017
|
+
}
|
|
1018
|
+
runner.emit('agent_tool_start', runContext, agent, shellTool, {
|
|
1019
|
+
toolCall,
|
|
1020
|
+
});
|
|
1021
|
+
if (typeof agent.emit === 'function') {
|
|
1022
|
+
agent.emit('agent_tool_start', runContext, shellTool, { toolCall });
|
|
1023
|
+
}
|
|
1024
|
+
let shellOutputs;
|
|
1025
|
+
const providerMeta = {};
|
|
1026
|
+
let maxOutputLength;
|
|
1027
|
+
try {
|
|
1028
|
+
const shellResult = await shellTool.shell.run(toolCall.action);
|
|
1029
|
+
shellOutputs = shellResult.output ?? [];
|
|
1030
|
+
if (shellResult.providerData) {
|
|
1031
|
+
Object.assign(providerMeta, shellResult.providerData);
|
|
1032
|
+
}
|
|
1033
|
+
if (typeof shellResult.maxOutputLength === 'number') {
|
|
1034
|
+
maxOutputLength = shellResult.maxOutputLength;
|
|
1035
|
+
}
|
|
1036
|
+
}
|
|
1037
|
+
catch (err) {
|
|
1038
|
+
const errorText = toErrorMessage(err);
|
|
1039
|
+
shellOutputs = [
|
|
1040
|
+
{
|
|
1041
|
+
stdout: '',
|
|
1042
|
+
stderr: errorText,
|
|
1043
|
+
outcome: { type: 'exit', exitCode: null },
|
|
1044
|
+
},
|
|
1045
|
+
];
|
|
1046
|
+
_logger.error('Failed to execute shell action:', err);
|
|
1047
|
+
}
|
|
1048
|
+
shellOutputs = shellOutputs ?? [];
|
|
1049
|
+
runner.emit('agent_tool_end', runContext, agent, shellTool, JSON.stringify(shellOutputs), {
|
|
1050
|
+
toolCall,
|
|
1051
|
+
});
|
|
1052
|
+
if (typeof agent.emit === 'function') {
|
|
1053
|
+
agent.emit('agent_tool_end', runContext, shellTool, JSON.stringify(shellOutputs), {
|
|
1054
|
+
toolCall,
|
|
1055
|
+
});
|
|
1056
|
+
}
|
|
1057
|
+
const rawItem = {
|
|
1058
|
+
type: 'shell_call_output',
|
|
1059
|
+
callId: toolCall.callId,
|
|
1060
|
+
output: shellOutputs ?? [],
|
|
1061
|
+
};
|
|
1062
|
+
if (typeof maxOutputLength === 'number') {
|
|
1063
|
+
rawItem.maxOutputLength = maxOutputLength;
|
|
1064
|
+
}
|
|
1065
|
+
if (Object.keys(providerMeta).length > 0) {
|
|
1066
|
+
rawItem.providerData = providerMeta;
|
|
1067
|
+
}
|
|
1068
|
+
results.push(new items_1.RunToolCallOutputItem(rawItem, agent, rawItem.output));
|
|
1069
|
+
}
|
|
1070
|
+
return results;
|
|
1071
|
+
}
|
|
1072
|
+
async function executeApplyPatchOperations(agent, actions, runner, runContext, customLogger = undefined) {
|
|
1073
|
+
const _logger = customLogger ?? logger_1.default;
|
|
1074
|
+
const results = [];
|
|
1075
|
+
for (const action of actions) {
|
|
1076
|
+
const applyPatchTool = action.applyPatch;
|
|
1077
|
+
const toolCall = action.toolCall;
|
|
1078
|
+
const approvalItem = new items_1.RunToolApprovalItem(toolCall, agent, applyPatchTool.name);
|
|
1079
|
+
const requiresApproval = await applyPatchTool.needsApproval(runContext, toolCall.operation, toolCall.callId);
|
|
1080
|
+
if (requiresApproval) {
|
|
1081
|
+
if (applyPatchTool.onApproval) {
|
|
1082
|
+
const decision = await applyPatchTool.onApproval(runContext, approvalItem);
|
|
1083
|
+
if (decision.approve === true) {
|
|
1084
|
+
runContext.approveTool(approvalItem);
|
|
1085
|
+
}
|
|
1086
|
+
else if (decision.approve === false) {
|
|
1087
|
+
runContext.rejectTool(approvalItem);
|
|
1088
|
+
}
|
|
1089
|
+
}
|
|
1090
|
+
const approval = runContext.isToolApproved({
|
|
1091
|
+
toolName: applyPatchTool.name,
|
|
1092
|
+
callId: toolCall.callId,
|
|
1093
|
+
});
|
|
1094
|
+
if (approval === false) {
|
|
1095
|
+
const response = 'Tool execution was not approved.';
|
|
1096
|
+
results.push(new items_1.RunToolCallOutputItem({
|
|
1097
|
+
type: 'apply_patch_call_output',
|
|
1098
|
+
callId: toolCall.callId,
|
|
1099
|
+
status: 'failed',
|
|
1100
|
+
output: response,
|
|
1101
|
+
}, agent, response));
|
|
1102
|
+
continue;
|
|
1103
|
+
}
|
|
1104
|
+
if (approval !== true) {
|
|
1105
|
+
results.push(approvalItem);
|
|
1106
|
+
continue;
|
|
1107
|
+
}
|
|
1108
|
+
}
|
|
1109
|
+
runner.emit('agent_tool_start', runContext, agent, applyPatchTool, {
|
|
1110
|
+
toolCall,
|
|
1111
|
+
});
|
|
1112
|
+
if (typeof agent.emit === 'function') {
|
|
1113
|
+
agent.emit('agent_tool_start', runContext, applyPatchTool, {
|
|
1114
|
+
toolCall,
|
|
1115
|
+
});
|
|
1116
|
+
}
|
|
1117
|
+
let status = 'completed';
|
|
1118
|
+
let output = '';
|
|
1119
|
+
try {
|
|
1120
|
+
let result;
|
|
1121
|
+
switch (toolCall.operation.type) {
|
|
1122
|
+
case 'create_file':
|
|
1123
|
+
result = await applyPatchTool.editor.createFile(toolCall.operation);
|
|
1124
|
+
break;
|
|
1125
|
+
case 'update_file':
|
|
1126
|
+
result = await applyPatchTool.editor.updateFile(toolCall.operation);
|
|
1127
|
+
break;
|
|
1128
|
+
case 'delete_file':
|
|
1129
|
+
result = await applyPatchTool.editor.deleteFile(toolCall.operation);
|
|
1130
|
+
break;
|
|
1131
|
+
default:
|
|
1132
|
+
throw new Error('Unsupported apply_patch operation');
|
|
1133
|
+
}
|
|
1134
|
+
if (result && typeof result.status === 'string') {
|
|
1135
|
+
status = result.status;
|
|
1136
|
+
}
|
|
1137
|
+
if (result && typeof result.output === 'string') {
|
|
1138
|
+
output = result.output;
|
|
1139
|
+
}
|
|
1140
|
+
}
|
|
1141
|
+
catch (err) {
|
|
1142
|
+
status = 'failed';
|
|
1143
|
+
output = toErrorMessage(err);
|
|
1144
|
+
_logger.error('Failed to execute apply_patch operation:', err);
|
|
1145
|
+
}
|
|
1146
|
+
runner.emit('agent_tool_end', runContext, agent, applyPatchTool, output, {
|
|
1147
|
+
toolCall,
|
|
1148
|
+
});
|
|
1149
|
+
if (typeof agent.emit === 'function') {
|
|
1150
|
+
agent.emit('agent_tool_end', runContext, applyPatchTool, output, {
|
|
1151
|
+
toolCall,
|
|
1152
|
+
});
|
|
1153
|
+
}
|
|
1154
|
+
const rawItem = {
|
|
1155
|
+
type: 'apply_patch_call_output',
|
|
1156
|
+
callId: toolCall.callId,
|
|
1157
|
+
status,
|
|
1158
|
+
};
|
|
1159
|
+
if (output) {
|
|
1160
|
+
rawItem.output = output;
|
|
1161
|
+
}
|
|
1162
|
+
results.push(new items_1.RunToolCallOutputItem(rawItem, agent, output));
|
|
1163
|
+
}
|
|
1164
|
+
return results;
|
|
1165
|
+
}
|
|
942
1166
|
/**
|
|
943
1167
|
* @internal
|
|
1168
|
+
* Executes any computer-use actions emitted by the model and returns the resulting items so the
|
|
1169
|
+
* run history reflects the computer session.
|
|
944
1170
|
*/
|
|
945
1171
|
async function executeComputerActions(agent, actions, runner, runContext, customLogger = undefined) {
|
|
946
1172
|
const _logger = customLogger ?? logger_1.default;
|
|
@@ -973,7 +1199,7 @@ async function executeComputerActions(agent, actions, runner, runContext, custom
|
|
|
973
1199
|
toolCall,
|
|
974
1200
|
});
|
|
975
1201
|
}
|
|
976
|
-
//
|
|
1202
|
+
// Return the screenshot as a data URL when available; fall back to an empty string on failures.
|
|
977
1203
|
const imageUrl = output ? `data:image/png;base64,${output}` : '';
|
|
978
1204
|
const rawItem = {
|
|
979
1205
|
type: 'computer_call_result',
|
|
@@ -986,6 +1212,8 @@ async function executeComputerActions(agent, actions, runner, runContext, custom
|
|
|
986
1212
|
}
|
|
987
1213
|
/**
|
|
988
1214
|
* @internal
|
|
1215
|
+
* Drives handoff calls by invoking the downstream agent and capturing any generated items so
|
|
1216
|
+
* the current agent can continue with the new context.
|
|
989
1217
|
*/
|
|
990
1218
|
async function executeHandoffCalls(agent, originalInput, preStepItems, newStepItems, newResponse, runHandoffs, runner, runContext) {
|
|
991
1219
|
newStepItems = [...newStepItems];
|
|
@@ -1054,6 +1282,8 @@ const NOT_FINAL_OUTPUT = {
|
|
|
1054
1282
|
};
|
|
1055
1283
|
/**
|
|
1056
1284
|
* @internal
|
|
1285
|
+
* Determines whether tool executions produced a final agent output, triggered an interruption,
|
|
1286
|
+
* or whether the agent loop should continue collecting more responses.
|
|
1057
1287
|
*/
|
|
1058
1288
|
async function checkForFinalOutputFromTools(agent, toolResults, state) {
|
|
1059
1289
|
if (toolResults.length === 0) {
|
|
@@ -1181,4 +1411,571 @@ class AgentToolUseTracker {
|
|
|
1181
1411
|
}
|
|
1182
1412
|
}
|
|
1183
1413
|
exports.AgentToolUseTracker = AgentToolUseTracker;
|
|
1414
|
+
/**
|
|
1415
|
+
* @internal
|
|
1416
|
+
* Convert a user-provided input into a list of input items.
|
|
1417
|
+
*/
|
|
1418
|
+
function toInputItemList(input) {
|
|
1419
|
+
if (typeof input === 'string') {
|
|
1420
|
+
return [
|
|
1421
|
+
{
|
|
1422
|
+
type: 'message',
|
|
1423
|
+
role: 'user',
|
|
1424
|
+
content: input,
|
|
1425
|
+
},
|
|
1426
|
+
];
|
|
1427
|
+
}
|
|
1428
|
+
return [...input];
|
|
1429
|
+
}
|
|
1430
|
+
/**
|
|
1431
|
+
* @internal
|
|
1432
|
+
* Extract model output items from run items, excluding tool approval items.
|
|
1433
|
+
*/
|
|
1434
|
+
function extractOutputItemsFromRunItems(items) {
|
|
1435
|
+
return items
|
|
1436
|
+
.filter((item) => item.type !== 'tool_approval_item')
|
|
1437
|
+
.map((item) => item.rawItem);
|
|
1438
|
+
}
|
|
1439
|
+
function normalizeItemsForSessionPersistence(items) {
|
|
1440
|
+
// Persisted sessions must avoid raw binary so we convert every item into a JSON-safe shape before writing to storage.
|
|
1441
|
+
return items.map((item) => sanitizeValueForSession(stripTransientCallIds(item)));
|
|
1442
|
+
}
|
|
1443
|
+
function sanitizeValueForSession(value, context = {}) {
|
|
1444
|
+
if (value === null || value === undefined) {
|
|
1445
|
+
return value;
|
|
1446
|
+
}
|
|
1447
|
+
// Convert supported binary payloads into ArrayBuffer views before serialization.
|
|
1448
|
+
const binary = toUint8ArrayIfBinary(value);
|
|
1449
|
+
if (binary) {
|
|
1450
|
+
return toDataUrlFromBytes(binary, context.mediaType);
|
|
1451
|
+
}
|
|
1452
|
+
if (Array.isArray(value)) {
|
|
1453
|
+
return value.map((entry) => sanitizeValueForSession(entry, context));
|
|
1454
|
+
}
|
|
1455
|
+
if (!isPlainObject(value)) {
|
|
1456
|
+
return value;
|
|
1457
|
+
}
|
|
1458
|
+
const record = value;
|
|
1459
|
+
const result = {};
|
|
1460
|
+
const mediaType = typeof record.mediaType === 'string' && record.mediaType.length > 0
|
|
1461
|
+
? record.mediaType
|
|
1462
|
+
: context.mediaType;
|
|
1463
|
+
for (const [key, entry] of Object.entries(record)) {
|
|
1464
|
+
// Propagate explicit media type only when walking into binary payload containers.
|
|
1465
|
+
const nextContext = key === 'data' || key === 'fileData' ? { mediaType } : context;
|
|
1466
|
+
result[key] = sanitizeValueForSession(entry, nextContext);
|
|
1467
|
+
}
|
|
1468
|
+
return result;
|
|
1469
|
+
}
|
|
1470
|
+
function toUint8ArrayIfBinary(value) {
|
|
1471
|
+
// Normalize the diverse binary containers we may receive into a shared Uint8Array view.
|
|
1472
|
+
if (value instanceof ArrayBuffer) {
|
|
1473
|
+
return new Uint8Array(value);
|
|
1474
|
+
}
|
|
1475
|
+
if ((0, smartString_1.isArrayBufferView)(value)) {
|
|
1476
|
+
const view = value;
|
|
1477
|
+
return new Uint8Array(view.buffer, view.byteOffset, view.byteLength);
|
|
1478
|
+
}
|
|
1479
|
+
if ((0, smartString_1.isNodeBuffer)(value)) {
|
|
1480
|
+
const view = value;
|
|
1481
|
+
return new Uint8Array(view.buffer, view.byteOffset, view.byteLength);
|
|
1482
|
+
}
|
|
1483
|
+
if ((0, smartString_1.isSerializedBufferSnapshot)(value)) {
|
|
1484
|
+
const snapshot = value;
|
|
1485
|
+
return Uint8Array.from(snapshot.data);
|
|
1486
|
+
}
|
|
1487
|
+
return undefined;
|
|
1488
|
+
}
|
|
1489
|
+
function toDataUrlFromBytes(bytes, mediaType) {
|
|
1490
|
+
// Convert binary payloads into a durable data URL so session files remain self-contained.
|
|
1491
|
+
const base64 = (0, base64_1.encodeUint8ArrayToBase64)(bytes);
|
|
1492
|
+
// Note that OpenAI Responses API never accepts application/octet-stream as a media type,
|
|
1493
|
+
// so we fall back to text/plain; that said, tools are supposed to return a valid media type when this utility is used.
|
|
1494
|
+
const type = mediaType && !mediaType.startsWith('data:') ? mediaType : 'text/plain';
|
|
1495
|
+
return `data:${type};base64,${base64}`;
|
|
1496
|
+
}
|
|
1497
|
+
function isPlainObject(value) {
|
|
1498
|
+
if (typeof value !== 'object' || value === null) {
|
|
1499
|
+
return false;
|
|
1500
|
+
}
|
|
1501
|
+
const proto = Object.getPrototypeOf(value);
|
|
1502
|
+
return proto === Object.prototype || proto === null;
|
|
1503
|
+
}
|
|
1504
|
+
function stripTransientCallIds(value) {
|
|
1505
|
+
if (value === null || value === undefined) {
|
|
1506
|
+
return value;
|
|
1507
|
+
}
|
|
1508
|
+
if (Array.isArray(value)) {
|
|
1509
|
+
return value.map((entry) => stripTransientCallIds(entry));
|
|
1510
|
+
}
|
|
1511
|
+
if (!isPlainObject(value)) {
|
|
1512
|
+
return value;
|
|
1513
|
+
}
|
|
1514
|
+
const record = value;
|
|
1515
|
+
const result = {};
|
|
1516
|
+
const isProtocolItem = typeof record.type === 'string' && record.type.length > 0;
|
|
1517
|
+
const shouldStripId = isProtocolItem && shouldStripIdForType(record.type);
|
|
1518
|
+
for (const [key, entry] of Object.entries(record)) {
|
|
1519
|
+
if (shouldStripId && key === 'id') {
|
|
1520
|
+
continue;
|
|
1521
|
+
}
|
|
1522
|
+
result[key] = stripTransientCallIds(entry);
|
|
1523
|
+
}
|
|
1524
|
+
return result;
|
|
1525
|
+
}
|
|
1526
|
+
function shouldStripIdForType(type) {
|
|
1527
|
+
switch (type) {
|
|
1528
|
+
case 'function_call':
|
|
1529
|
+
case 'function_call_result':
|
|
1530
|
+
return true;
|
|
1531
|
+
default:
|
|
1532
|
+
return false;
|
|
1533
|
+
}
|
|
1534
|
+
}
|
|
1535
|
+
/**
|
|
1536
|
+
* @internal
|
|
1537
|
+
* Persist full turn (input + outputs) for non-streaming runs.
|
|
1538
|
+
*/
|
|
1539
|
+
// Persists the combination of user inputs (possibly filtered) and model outputs for a completed turn.
|
|
1540
|
+
async function saveToSession(session, sessionInputItems, result) {
|
|
1541
|
+
if (!session) {
|
|
1542
|
+
return;
|
|
1543
|
+
}
|
|
1544
|
+
const inputItems = sessionInputItems ?? [];
|
|
1545
|
+
const state = result.state;
|
|
1546
|
+
const alreadyPersisted = state._currentTurnPersistedItemCount ?? 0;
|
|
1547
|
+
// Persist only the portion of _generatedItems that has not yet been stored for this turn.
|
|
1548
|
+
const newRunItems = result.newItems.slice(alreadyPersisted);
|
|
1549
|
+
if (process.env.OPENAI_AGENTS__DEBUG_SAVE_SESSION) {
|
|
1550
|
+
console.debug('saveToSession:newRunItems', newRunItems.map((item) => item.type));
|
|
1551
|
+
}
|
|
1552
|
+
const outputItems = extractOutputItemsFromRunItems(newRunItems);
|
|
1553
|
+
const itemsToSave = [...inputItems, ...outputItems];
|
|
1554
|
+
if (itemsToSave.length === 0) {
|
|
1555
|
+
state._currentTurnPersistedItemCount =
|
|
1556
|
+
alreadyPersisted + newRunItems.length;
|
|
1557
|
+
return;
|
|
1558
|
+
}
|
|
1559
|
+
const sanitizedItems = normalizeItemsForSessionPersistence(itemsToSave);
|
|
1560
|
+
await session.addItems(sanitizedItems);
|
|
1561
|
+
state._currentTurnPersistedItemCount = alreadyPersisted + newRunItems.length;
|
|
1562
|
+
}
|
|
1563
|
+
/**
|
|
1564
|
+
* @internal
|
|
1565
|
+
* Persist only the user input for streaming runs at start.
|
|
1566
|
+
*/
|
|
1567
|
+
// For streaming runs we persist user input as soon as it is sent so reconnections can resume.
|
|
1568
|
+
async function saveStreamInputToSession(session, sessionInputItems) {
|
|
1569
|
+
if (!session) {
|
|
1570
|
+
return;
|
|
1571
|
+
}
|
|
1572
|
+
if (!sessionInputItems || sessionInputItems.length === 0) {
|
|
1573
|
+
return;
|
|
1574
|
+
}
|
|
1575
|
+
const sanitizedInput = normalizeItemsForSessionPersistence(sessionInputItems);
|
|
1576
|
+
await session.addItems(sanitizedInput);
|
|
1577
|
+
}
|
|
1578
|
+
/**
|
|
1579
|
+
* @internal
|
|
1580
|
+
* Persist only the model outputs for streaming runs at the end of a turn.
|
|
1581
|
+
*/
|
|
1582
|
+
// Complements saveStreamInputToSession by recording the streaming outputs at the end of the turn.
|
|
1583
|
+
async function saveStreamResultToSession(session, result) {
|
|
1584
|
+
if (!session) {
|
|
1585
|
+
return;
|
|
1586
|
+
}
|
|
1587
|
+
const state = result.state;
|
|
1588
|
+
const alreadyPersisted = state._currentTurnPersistedItemCount ?? 0;
|
|
1589
|
+
const newRunItems = result.newItems.slice(alreadyPersisted);
|
|
1590
|
+
const itemsToSave = extractOutputItemsFromRunItems(newRunItems);
|
|
1591
|
+
if (itemsToSave.length === 0) {
|
|
1592
|
+
state._currentTurnPersistedItemCount =
|
|
1593
|
+
alreadyPersisted + newRunItems.length;
|
|
1594
|
+
return;
|
|
1595
|
+
}
|
|
1596
|
+
const sanitizedItems = normalizeItemsForSessionPersistence(itemsToSave);
|
|
1597
|
+
await session.addItems(sanitizedItems);
|
|
1598
|
+
state._currentTurnPersistedItemCount = alreadyPersisted + newRunItems.length;
|
|
1599
|
+
}
|
|
1600
|
+
async function prepareInputItemsWithSession(input, session, sessionInputCallback, options) {
|
|
1601
|
+
if (!session) {
|
|
1602
|
+
return {
|
|
1603
|
+
preparedInput: input,
|
|
1604
|
+
sessionItems: undefined,
|
|
1605
|
+
};
|
|
1606
|
+
}
|
|
1607
|
+
const includeHistoryInPreparedInput = options?.includeHistoryInPreparedInput ?? true;
|
|
1608
|
+
const preserveDroppedNewItems = options?.preserveDroppedNewItems ?? false;
|
|
1609
|
+
const history = await session.getItems();
|
|
1610
|
+
const newInputItems = Array.isArray(input)
|
|
1611
|
+
? [...input]
|
|
1612
|
+
: toInputItemList(input);
|
|
1613
|
+
if (!sessionInputCallback) {
|
|
1614
|
+
return {
|
|
1615
|
+
preparedInput: includeHistoryInPreparedInput
|
|
1616
|
+
? [...history, ...newInputItems]
|
|
1617
|
+
: newInputItems,
|
|
1618
|
+
sessionItems: newInputItems,
|
|
1619
|
+
};
|
|
1620
|
+
}
|
|
1621
|
+
// Capture snapshots before invoking the callback so we can reason about the original state even
|
|
1622
|
+
// if the callback mutates the history array in-place.
|
|
1623
|
+
const historySnapshot = history.slice();
|
|
1624
|
+
const newInputSnapshot = newInputItems.slice();
|
|
1625
|
+
// Delegate history reconciliation to the user-supplied callback. It must return a concrete list
|
|
1626
|
+
// to keep downstream model requests well-typed.
|
|
1627
|
+
const combined = await sessionInputCallback(history, newInputItems);
|
|
1628
|
+
if (!Array.isArray(combined)) {
|
|
1629
|
+
throw new errors_1.UserError('Session input callback must return an array of AgentInputItem objects.');
|
|
1630
|
+
}
|
|
1631
|
+
const historyCounts = buildItemFrequencyMap(historySnapshot);
|
|
1632
|
+
const newInputCounts = buildItemFrequencyMap(newInputSnapshot);
|
|
1633
|
+
const historyRefs = buildItemReferenceMap(historySnapshot);
|
|
1634
|
+
const newInputRefs = buildItemReferenceMap(newInputSnapshot);
|
|
1635
|
+
const appended = [];
|
|
1636
|
+
for (const item of combined) {
|
|
1637
|
+
const key = sessionItemKey(item);
|
|
1638
|
+
if (consumeReference(newInputRefs, key, item)) {
|
|
1639
|
+
decrementCount(newInputCounts, key);
|
|
1640
|
+
appended.push(item);
|
|
1641
|
+
continue;
|
|
1642
|
+
}
|
|
1643
|
+
// Prioritize exact history matches before payload-based counts so callbacks that surface
|
|
1644
|
+
// history ahead of identical new inputs keep previously persisted items out of the new queue.
|
|
1645
|
+
if (consumeReference(historyRefs, key, item)) {
|
|
1646
|
+
decrementCount(historyCounts, key);
|
|
1647
|
+
continue;
|
|
1648
|
+
}
|
|
1649
|
+
const historyRemaining = historyCounts.get(key) ?? 0;
|
|
1650
|
+
if (historyRemaining > 0) {
|
|
1651
|
+
historyCounts.set(key, historyRemaining - 1);
|
|
1652
|
+
continue;
|
|
1653
|
+
}
|
|
1654
|
+
const newRemaining = newInputCounts.get(key) ?? 0;
|
|
1655
|
+
if (newRemaining > 0) {
|
|
1656
|
+
newInputCounts.set(key, newRemaining - 1);
|
|
1657
|
+
appended.push(item);
|
|
1658
|
+
continue;
|
|
1659
|
+
}
|
|
1660
|
+
appended.push(item);
|
|
1661
|
+
}
|
|
1662
|
+
// Preserve redacted inputs for model delivery when requested (e.g. server-managed histories).
|
|
1663
|
+
const preparedItems = includeHistoryInPreparedInput
|
|
1664
|
+
? combined
|
|
1665
|
+
: appended.length > 0
|
|
1666
|
+
? appended
|
|
1667
|
+
: preserveDroppedNewItems
|
|
1668
|
+
? newInputSnapshot
|
|
1669
|
+
: [];
|
|
1670
|
+
return {
|
|
1671
|
+
preparedInput: preparedItems,
|
|
1672
|
+
// Respect callbacks that intentionally drop the latest inputs (e.g. to redact sensitive
|
|
1673
|
+
// values) by persisting only the items they kept in the combined array.
|
|
1674
|
+
sessionItems: appended,
|
|
1675
|
+
};
|
|
1676
|
+
}
|
|
1677
|
+
/**
|
|
1678
|
+
* Accepts whatever the tool returned and attempts to coerce it into the structured protocol
|
|
1679
|
+
* shapes we expose to downstream model adapters (input_text/input_image/input_file). Tools are
|
|
1680
|
+
* allowed to return either a single structured object or an array of them; anything else falls
|
|
1681
|
+
* back to the legacy string pipeline.
|
|
1682
|
+
*/
|
|
1683
|
+
function normalizeStructuredToolOutputs(output) {
|
|
1684
|
+
if (Array.isArray(output)) {
|
|
1685
|
+
const structured = [];
|
|
1686
|
+
for (const item of output) {
|
|
1687
|
+
const normalized = normalizeStructuredToolOutput(item);
|
|
1688
|
+
if (!normalized) {
|
|
1689
|
+
return null;
|
|
1690
|
+
}
|
|
1691
|
+
structured.push(normalized);
|
|
1692
|
+
}
|
|
1693
|
+
return structured;
|
|
1694
|
+
}
|
|
1695
|
+
const normalized = normalizeStructuredToolOutput(output);
|
|
1696
|
+
return normalized ? [normalized] : null;
|
|
1697
|
+
}
|
|
1698
|
+
/**
|
|
1699
|
+
* Best-effort normalization of a single tool output item. If the object already matches the
|
|
1700
|
+
* protocol shape we simply cast it; otherwise we copy the recognised fields into the canonical
|
|
1701
|
+
* structure. Returning null lets the caller know we should revert to plain-string handling.
|
|
1702
|
+
*/
|
|
1703
|
+
function normalizeStructuredToolOutput(value) {
|
|
1704
|
+
if (!isRecord(value)) {
|
|
1705
|
+
return null;
|
|
1706
|
+
}
|
|
1707
|
+
const type = value.type;
|
|
1708
|
+
if (type === 'text' && typeof value.text === 'string') {
|
|
1709
|
+
const output = { type: 'text', text: value.text };
|
|
1710
|
+
if (isRecord(value.providerData)) {
|
|
1711
|
+
output.providerData = value.providerData;
|
|
1712
|
+
}
|
|
1713
|
+
return output;
|
|
1714
|
+
}
|
|
1715
|
+
if (type === 'image') {
|
|
1716
|
+
const output = { type: 'image' };
|
|
1717
|
+
let imageString;
|
|
1718
|
+
let imageFileId;
|
|
1719
|
+
const fallbackImageMediaType = isNonEmptyString(value.mediaType)
|
|
1720
|
+
? value.mediaType
|
|
1721
|
+
: undefined;
|
|
1722
|
+
const imageField = value.image;
|
|
1723
|
+
if (typeof imageField === 'string' && imageField.length > 0) {
|
|
1724
|
+
imageString = imageField;
|
|
1725
|
+
}
|
|
1726
|
+
else if (isRecord(imageField)) {
|
|
1727
|
+
const imageObj = imageField;
|
|
1728
|
+
const inlineMediaType = isNonEmptyString(imageObj.mediaType)
|
|
1729
|
+
? imageObj.mediaType
|
|
1730
|
+
: fallbackImageMediaType;
|
|
1731
|
+
if (isNonEmptyString(imageObj.url)) {
|
|
1732
|
+
imageString = imageObj.url;
|
|
1733
|
+
}
|
|
1734
|
+
else if (isNonEmptyString(imageObj.data)) {
|
|
1735
|
+
imageString = toInlineImageString(imageObj.data, inlineMediaType);
|
|
1736
|
+
}
|
|
1737
|
+
else if (imageObj.data instanceof Uint8Array &&
|
|
1738
|
+
imageObj.data.length > 0) {
|
|
1739
|
+
imageString = toInlineImageString(imageObj.data, inlineMediaType);
|
|
1740
|
+
}
|
|
1741
|
+
if (!imageString) {
|
|
1742
|
+
const candidateId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
|
|
1743
|
+
(isNonEmptyString(imageObj.id) && imageObj.id) ||
|
|
1744
|
+
undefined;
|
|
1745
|
+
if (candidateId) {
|
|
1746
|
+
imageFileId = candidateId;
|
|
1747
|
+
}
|
|
1748
|
+
}
|
|
1749
|
+
}
|
|
1750
|
+
if (!imageString &&
|
|
1751
|
+
typeof value.imageUrl === 'string' &&
|
|
1752
|
+
value.imageUrl.length > 0) {
|
|
1753
|
+
imageString = value.imageUrl;
|
|
1754
|
+
}
|
|
1755
|
+
if (!imageFileId &&
|
|
1756
|
+
typeof value.fileId === 'string' &&
|
|
1757
|
+
value.fileId.length > 0) {
|
|
1758
|
+
imageFileId = value.fileId;
|
|
1759
|
+
}
|
|
1760
|
+
if (!imageString &&
|
|
1761
|
+
typeof value.data === 'string' &&
|
|
1762
|
+
value.data.length > 0) {
|
|
1763
|
+
imageString = fallbackImageMediaType
|
|
1764
|
+
? toInlineImageString(value.data, fallbackImageMediaType)
|
|
1765
|
+
: value.data;
|
|
1766
|
+
}
|
|
1767
|
+
else if (!imageString &&
|
|
1768
|
+
value.data instanceof Uint8Array &&
|
|
1769
|
+
value.data.length > 0) {
|
|
1770
|
+
imageString = toInlineImageString(value.data, fallbackImageMediaType);
|
|
1771
|
+
}
|
|
1772
|
+
if (typeof value.detail === 'string' && value.detail.length > 0) {
|
|
1773
|
+
output.detail = value.detail;
|
|
1774
|
+
}
|
|
1775
|
+
if (imageString) {
|
|
1776
|
+
output.image = imageString;
|
|
1777
|
+
}
|
|
1778
|
+
else if (imageFileId) {
|
|
1779
|
+
output.image = { fileId: imageFileId };
|
|
1780
|
+
}
|
|
1781
|
+
else {
|
|
1782
|
+
return null;
|
|
1783
|
+
}
|
|
1784
|
+
if (isRecord(value.providerData)) {
|
|
1785
|
+
output.providerData = value.providerData;
|
|
1786
|
+
}
|
|
1787
|
+
return output;
|
|
1788
|
+
}
|
|
1789
|
+
if (type === 'file') {
|
|
1790
|
+
const fileValue = normalizeFileValue(value);
|
|
1791
|
+
if (!fileValue) {
|
|
1792
|
+
return null;
|
|
1793
|
+
}
|
|
1794
|
+
const output = { type: 'file', file: fileValue };
|
|
1795
|
+
if (isRecord(value.providerData)) {
|
|
1796
|
+
output.providerData = value.providerData;
|
|
1797
|
+
}
|
|
1798
|
+
return output;
|
|
1799
|
+
}
|
|
1800
|
+
return null;
|
|
1801
|
+
}
|
|
1802
|
+
/**
|
|
1803
|
+
* Translates the normalized tool output into the protocol `input_*` items. This is the last hop
|
|
1804
|
+
* before we hand the data to model-specific adapters, so we generate the exact schema expected by
|
|
1805
|
+
* the protocol definitions.
|
|
1806
|
+
*/
|
|
1807
|
+
function convertStructuredToolOutputToInputItem(output) {
|
|
1808
|
+
if (output.type === 'text') {
|
|
1809
|
+
const result = {
|
|
1810
|
+
type: 'input_text',
|
|
1811
|
+
text: output.text,
|
|
1812
|
+
};
|
|
1813
|
+
if (output.providerData) {
|
|
1814
|
+
result.providerData = output.providerData;
|
|
1815
|
+
}
|
|
1816
|
+
return result;
|
|
1817
|
+
}
|
|
1818
|
+
if (output.type === 'image') {
|
|
1819
|
+
const result = { type: 'input_image' };
|
|
1820
|
+
if (typeof output.detail === 'string' && output.detail.length > 0) {
|
|
1821
|
+
result.detail = output.detail;
|
|
1822
|
+
}
|
|
1823
|
+
if (typeof output.image === 'string' && output.image.length > 0) {
|
|
1824
|
+
result.image = output.image;
|
|
1825
|
+
}
|
|
1826
|
+
else if (isRecord(output.image)) {
|
|
1827
|
+
const imageObj = output.image;
|
|
1828
|
+
const inlineMediaType = isNonEmptyString(imageObj.mediaType)
|
|
1829
|
+
? imageObj.mediaType
|
|
1830
|
+
: undefined;
|
|
1831
|
+
if (isNonEmptyString(imageObj.url)) {
|
|
1832
|
+
result.image = imageObj.url;
|
|
1833
|
+
}
|
|
1834
|
+
else if (isNonEmptyString(imageObj.data)) {
|
|
1835
|
+
result.image =
|
|
1836
|
+
inlineMediaType && !imageObj.data.startsWith('data:')
|
|
1837
|
+
? asDataUrl(imageObj.data, inlineMediaType)
|
|
1838
|
+
: imageObj.data;
|
|
1839
|
+
}
|
|
1840
|
+
else if (imageObj.data instanceof Uint8Array &&
|
|
1841
|
+
imageObj.data.length > 0) {
|
|
1842
|
+
const base64 = (0, base64_1.encodeUint8ArrayToBase64)(imageObj.data);
|
|
1843
|
+
result.image = asDataUrl(base64, inlineMediaType);
|
|
1844
|
+
}
|
|
1845
|
+
else {
|
|
1846
|
+
const referencedId = (isNonEmptyString(imageObj.fileId) && imageObj.fileId) ||
|
|
1847
|
+
(isNonEmptyString(imageObj.id) && imageObj.id) ||
|
|
1848
|
+
undefined;
|
|
1849
|
+
if (referencedId) {
|
|
1850
|
+
result.image = { id: referencedId };
|
|
1851
|
+
}
|
|
1852
|
+
}
|
|
1853
|
+
}
|
|
1854
|
+
if (output.providerData) {
|
|
1855
|
+
result.providerData = output.providerData;
|
|
1856
|
+
}
|
|
1857
|
+
return result;
|
|
1858
|
+
}
|
|
1859
|
+
if (output.type === 'file') {
|
|
1860
|
+
const result = { type: 'input_file' };
|
|
1861
|
+
const fileValue = output.file;
|
|
1862
|
+
if (typeof fileValue === 'string') {
|
|
1863
|
+
result.file = fileValue;
|
|
1864
|
+
}
|
|
1865
|
+
else if (fileValue && typeof fileValue === 'object') {
|
|
1866
|
+
const record = fileValue;
|
|
1867
|
+
if ('data' in record && record.data) {
|
|
1868
|
+
const mediaType = record.mediaType ?? 'text/plain';
|
|
1869
|
+
if (typeof record.data === 'string') {
|
|
1870
|
+
result.file = asDataUrl(record.data, mediaType);
|
|
1871
|
+
}
|
|
1872
|
+
else {
|
|
1873
|
+
const base64 = (0, base64_1.encodeUint8ArrayToBase64)(record.data);
|
|
1874
|
+
result.file = asDataUrl(base64, mediaType);
|
|
1875
|
+
}
|
|
1876
|
+
}
|
|
1877
|
+
else if (typeof record.url === 'string' && record.url.length > 0) {
|
|
1878
|
+
result.file = { url: record.url };
|
|
1879
|
+
}
|
|
1880
|
+
else {
|
|
1881
|
+
const referencedId = (typeof record.id === 'string' &&
|
|
1882
|
+
record.id.length > 0 &&
|
|
1883
|
+
record.id) ||
|
|
1884
|
+
(typeof record.fileId === 'string' && record.fileId.length > 0
|
|
1885
|
+
? record.fileId
|
|
1886
|
+
: undefined);
|
|
1887
|
+
if (referencedId) {
|
|
1888
|
+
result.file = { id: referencedId };
|
|
1889
|
+
}
|
|
1890
|
+
}
|
|
1891
|
+
if (typeof record.filename === 'string' && record.filename.length > 0) {
|
|
1892
|
+
result.filename = record.filename;
|
|
1893
|
+
}
|
|
1894
|
+
}
|
|
1895
|
+
if (output.providerData) {
|
|
1896
|
+
result.providerData = output.providerData;
|
|
1897
|
+
}
|
|
1898
|
+
return result;
|
|
1899
|
+
}
|
|
1900
|
+
const exhaustiveCheck = output;
|
|
1901
|
+
return exhaustiveCheck;
|
|
1902
|
+
}
|
|
1903
|
+
function buildItemFrequencyMap(items) {
|
|
1904
|
+
const counts = new Map();
|
|
1905
|
+
for (const item of items) {
|
|
1906
|
+
const key = sessionItemKey(item);
|
|
1907
|
+
counts.set(key, (counts.get(key) ?? 0) + 1);
|
|
1908
|
+
}
|
|
1909
|
+
return counts;
|
|
1910
|
+
}
|
|
1911
|
+
function buildItemReferenceMap(items) {
|
|
1912
|
+
const refs = new Map();
|
|
1913
|
+
for (const item of items) {
|
|
1914
|
+
const key = sessionItemKey(item);
|
|
1915
|
+
const list = refs.get(key);
|
|
1916
|
+
if (list) {
|
|
1917
|
+
list.push(item);
|
|
1918
|
+
}
|
|
1919
|
+
else {
|
|
1920
|
+
refs.set(key, [item]);
|
|
1921
|
+
}
|
|
1922
|
+
}
|
|
1923
|
+
return refs;
|
|
1924
|
+
}
|
|
1925
|
+
function consumeReference(refs, key, target) {
|
|
1926
|
+
const candidates = refs.get(key);
|
|
1927
|
+
if (!candidates || candidates.length === 0) {
|
|
1928
|
+
return false;
|
|
1929
|
+
}
|
|
1930
|
+
const index = candidates.findIndex((candidate) => candidate === target);
|
|
1931
|
+
if (index === -1) {
|
|
1932
|
+
return false;
|
|
1933
|
+
}
|
|
1934
|
+
candidates.splice(index, 1);
|
|
1935
|
+
if (candidates.length === 0) {
|
|
1936
|
+
refs.delete(key);
|
|
1937
|
+
}
|
|
1938
|
+
return true;
|
|
1939
|
+
}
|
|
1940
|
+
function decrementCount(map, key) {
|
|
1941
|
+
const remaining = (map.get(key) ?? 0) - 1;
|
|
1942
|
+
if (remaining <= 0) {
|
|
1943
|
+
map.delete(key);
|
|
1944
|
+
}
|
|
1945
|
+
else {
|
|
1946
|
+
map.set(key, remaining);
|
|
1947
|
+
}
|
|
1948
|
+
}
|
|
1949
|
+
function sessionItemKey(item) {
|
|
1950
|
+
return JSON.stringify(item, sessionSerializationReplacer);
|
|
1951
|
+
}
|
|
1952
|
+
function sessionSerializationReplacer(_key, value) {
|
|
1953
|
+
if (value instanceof ArrayBuffer) {
|
|
1954
|
+
return {
|
|
1955
|
+
__type: 'ArrayBuffer',
|
|
1956
|
+
data: (0, base64_1.encodeUint8ArrayToBase64)(new Uint8Array(value)),
|
|
1957
|
+
};
|
|
1958
|
+
}
|
|
1959
|
+
if ((0, smartString_1.isArrayBufferView)(value)) {
|
|
1960
|
+
const view = value;
|
|
1961
|
+
return {
|
|
1962
|
+
__type: view.constructor.name,
|
|
1963
|
+
data: (0, base64_1.encodeUint8ArrayToBase64)(new Uint8Array(view.buffer, view.byteOffset, view.byteLength)),
|
|
1964
|
+
};
|
|
1965
|
+
}
|
|
1966
|
+
if ((0, smartString_1.isNodeBuffer)(value)) {
|
|
1967
|
+
const view = value;
|
|
1968
|
+
return {
|
|
1969
|
+
__type: 'Buffer',
|
|
1970
|
+
data: (0, base64_1.encodeUint8ArrayToBase64)(new Uint8Array(view.buffer, view.byteOffset, view.byteLength)),
|
|
1971
|
+
};
|
|
1972
|
+
}
|
|
1973
|
+
if ((0, smartString_1.isSerializedBufferSnapshot)(value)) {
|
|
1974
|
+
return {
|
|
1975
|
+
__type: 'Buffer',
|
|
1976
|
+
data: (0, base64_1.encodeUint8ArrayToBase64)(Uint8Array.from(value.data)),
|
|
1977
|
+
};
|
|
1978
|
+
}
|
|
1979
|
+
return value;
|
|
1980
|
+
}
|
|
1184
1981
|
//# sourceMappingURL=runImplementation.js.map
|