@dudousxd/nestjs-agent-core 0.11.0 → 0.13.0

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.
package/dist/index.js CHANGED
@@ -22,6 +22,9 @@ var AGENT_PROMPT_CONTRIBUTORS = Symbol.for("@dudousxd/nestjs-agent:prompt-contri
22
22
  var AGENT_ACTOR_DIRECTORY = Symbol.for("@dudousxd/nestjs-agent:actor-directory");
23
23
  var AGENT_ATTACHMENT_STAGING = Symbol.for("@dudousxd/nestjs-agent:attachment-staging");
24
24
  var AGENT_APPROVAL_PORT = Symbol.for("@dudousxd/nestjs-agent:approval-port");
25
+ var AGENT_SKILLS = Symbol.for("@dudousxd/nestjs-agent:skills");
26
+ var AGENT_MEMORY = Symbol.for("@dudousxd/nestjs-agent:memory");
27
+ var AGENT_SKILL_SOURCES = Symbol.for("@dudousxd/nestjs-agent:skill-sources");
25
28
 
26
29
  // src/spi/token-stream-sink.ts
27
30
  var AgentStreamError = class extends Error {
@@ -43,6 +46,17 @@ function encodeStreamEvent(event) {
43
46
  `);
44
47
  }
45
48
  __name(encodeStreamEvent, "encodeStreamEvent");
49
+ function decodeStreamEvent(line) {
50
+ try {
51
+ const parsed = JSON.parse(line);
52
+ if (typeof parsed === "object" && parsed !== null && typeof parsed.kind === "string") {
53
+ return parsed;
54
+ }
55
+ } catch {
56
+ }
57
+ return null;
58
+ }
59
+ __name(decodeStreamEvent, "decodeStreamEvent");
46
60
 
47
61
  // src/spi/pricing-store.ts
48
62
  async function seedModelPrices(store, prices) {
@@ -52,6 +66,40 @@ async function seedModelPrices(store, prices) {
52
66
  }
53
67
  __name(seedModelPrices, "seedModelPrices");
54
68
 
69
+ // src/spi/processors.ts
70
+ var DEFAULT_INCREMENTAL_LOOKBACK_CHARS = 64;
71
+ var OutputRejectedError = class extends Error {
72
+ static {
73
+ __name(this, "OutputRejectedError");
74
+ }
75
+ /** {@link OutputProcessor.name} of the processor that refused. */
76
+ processor;
77
+ /** The reason it gave, verbatim. */
78
+ reason;
79
+ constructor(processor, reason) {
80
+ super(`Output rejected by "${processor}": ${reason}`);
81
+ this.name = "OutputRejectedError";
82
+ this.processor = processor;
83
+ this.reason = reason;
84
+ }
85
+ };
86
+ var ProcessorFailedError = class extends Error {
87
+ static {
88
+ __name(this, "ProcessorFailedError");
89
+ }
90
+ /** Which seam it was on, so a reader knows whether the prompt or the answer was in flight. */
91
+ phase;
92
+ /** The processor's `name`. */
93
+ processor;
94
+ constructor(phase, processor, cause) {
95
+ super(`${phase} processor "${processor}" failed: ${cause instanceof Error ? cause.message : String(cause)}`);
96
+ this.name = "ProcessorFailedError";
97
+ this.phase = phase;
98
+ this.processor = processor;
99
+ this.cause = cause;
100
+ }
101
+ };
102
+
55
103
  // src/spi/governance-queries.ts
56
104
  var THREAD_DETAIL_CONTENT_CHARS = 2e3;
57
105
  function truncateDetailContent(content) {
@@ -223,6 +271,34 @@ function dayBoundsUtc(range) {
223
271
  __name(dayBoundsUtc, "dayBoundsUtc");
224
272
 
225
273
  // src/tool-filters.ts
274
+ async function isToolEnabled(spec, handler) {
275
+ const declared = typeof spec.enabled === "function" ? await spec.enabled() : spec.enabled ?? true;
276
+ if (!declared) {
277
+ return false;
278
+ }
279
+ return handler?.isEnabled === void 0 ? true : await handler.isEnabled();
280
+ }
281
+ __name(isToolEnabled, "isToolEnabled");
282
+ async function filterToolsByEnabled(entries) {
283
+ const checked = await Promise.all(entries.map(async (entry) => ({
284
+ entry,
285
+ enabled: await isToolEnabled(entry.spec, entry.handler)
286
+ })));
287
+ return checked.filter((row) => row.enabled).map((row) => row.entry);
288
+ }
289
+ __name(filterToolsByEnabled, "filterToolsByEnabled");
290
+ async function canActorUseTool(actor, handler) {
291
+ return handler?.canUse === void 0 ? true : await handler.canUse(actor);
292
+ }
293
+ __name(canActorUseTool, "canActorUseTool");
294
+ async function filterToolsByCanUse(entries, actor) {
295
+ const checked = await Promise.all(entries.map(async (entry) => ({
296
+ entry,
297
+ allowed: await canActorUseTool(actor, entry.handler)
298
+ })));
299
+ return checked.filter((row) => row.allowed).map((row) => row.entry);
300
+ }
301
+ __name(filterToolsByCanUse, "filterToolsByCanUse");
226
302
  async function filterToolsByRole(tools, actor, policy) {
227
303
  const checked = await Promise.all(tools.map(async (tool) => ({
228
304
  tool,
@@ -240,433 +316,2753 @@ function filterToolsByAllowList(tools, allowedTools) {
240
316
  }
241
317
  __name(filterToolsByAllowList, "filterToolsByAllowList");
242
318
 
243
- // src/agent-registry.ts
244
- var AgentRegistry = class {
245
- static {
246
- __name(this, "AgentRegistry");
247
- }
248
- definitions = /* @__PURE__ */ new Map();
249
- register(definition) {
250
- this.definitions.set(definition.name, definition);
251
- }
252
- get(name) {
253
- return this.definitions.get(name);
254
- }
255
- has(name) {
256
- return this.definitions.has(name);
257
- }
258
- list() {
259
- return [
260
- ...this.definitions.values()
261
- ];
319
+ // src/history.ts
320
+ function estimateMessageTokens(message) {
321
+ const extras = (message.toolCalls !== void 0 ? JSON.stringify(message.toolCalls).length : 0) + (message.toolResults !== void 0 ? JSON.stringify(message.toolResults).length : 0);
322
+ return Math.ceil((message.content.length + extras) / 4) + 4;
323
+ }
324
+ __name(estimateMessageTokens, "estimateMessageTokens");
325
+ function windowHistory(options) {
326
+ const estimate = options.estimate ?? estimateMessageTokens;
327
+ const policy = {
328
+ // A suffix selection of at most `maxMessages` rows, which is exactly what `HistoryPolicy`'s row
329
+ // hint promises — so the store can bound its read to that many of the thread's newest messages.
330
+ // A token-only window declares nothing: its budget names no row count.
331
+ ...options.maxMessages !== void 0 ? {
332
+ maxMessages: options.maxMessages
333
+ } : {},
334
+ select(messages) {
335
+ let cut = 0;
336
+ if (options.maxMessages !== void 0) {
337
+ cut = Math.max(cut, messages.length - options.maxMessages);
338
+ }
339
+ if (options.maxTokens !== void 0) {
340
+ cut = Math.max(cut, tokenCut(messages, options.maxTokens, estimate));
341
+ }
342
+ cut = Math.min(cut, Math.max(0, messages.length - 1));
343
+ return {
344
+ keep: messages.slice(cut),
345
+ drop: messages.slice(0, cut)
346
+ };
347
+ }
348
+ };
349
+ return options.summarize !== void 0 ? {
350
+ ...policy,
351
+ summarize: options.summarize
352
+ } : policy;
353
+ }
354
+ __name(windowHistory, "windowHistory");
355
+ function tokenCut(messages, maxTokens, estimate) {
356
+ let index = messages.length;
357
+ let total = 0;
358
+ while (index > 0) {
359
+ const message = messages[index - 1];
360
+ if (message === void 0) {
361
+ break;
362
+ }
363
+ const cost = estimate(message);
364
+ if (index < messages.length && total + cost > maxTokens) {
365
+ break;
366
+ }
367
+ total += cost;
368
+ index -= 1;
262
369
  }
263
- };
370
+ return index;
371
+ }
372
+ __name(tokenCut, "tokenCut");
373
+ var DEFAULT_HISTORY_SUMMARY_INSTRUCTION = "Summarize this conversation for an assistant that will continue it without seeing these messages. Preserve decisions made, facts and constraints the user stated, identifiers and names mentioned, and anything left unresolved. Omit pleasantries. Reply with the summary only \u2014 no preamble, no headings.";
374
+ function summarizeWithModel(model, instruction = DEFAULT_HISTORY_SUMMARY_INSTRUCTION) {
375
+ return async (dropped) => {
376
+ const discard = {
377
+ write: /* @__PURE__ */ __name(() => {
378
+ }, "write"),
379
+ end: /* @__PURE__ */ __name(() => {
380
+ }, "end"),
381
+ fail: /* @__PURE__ */ __name(() => {
382
+ }, "fail")
383
+ };
384
+ const turn = await model.runTurn({
385
+ system: instruction,
386
+ messages: dropped,
387
+ tools: [],
388
+ sink: discard
389
+ });
390
+ return {
391
+ text: turn.text,
392
+ usage: turn.usage,
393
+ ...turn.modelId !== void 0 ? {
394
+ modelId: turn.modelId
395
+ } : {}
396
+ };
397
+ };
398
+ }
399
+ __name(summarizeWithModel, "summarizeWithModel");
264
400
 
265
- // src/tool-registry.ts
266
- var ToolForbiddenError = class extends Error {
267
- static {
268
- __name(this, "ToolForbiddenError");
401
+ // src/processors.ts
402
+ var decoder = new TextDecoder();
403
+ function createFrameBuffer() {
404
+ const frames = [];
405
+ return {
406
+ writer: {
407
+ write: /* @__PURE__ */ __name((chunk) => {
408
+ for (const line of decoder.decode(chunk).split("\n")) {
409
+ if (line.length > 0) {
410
+ frames.push(line);
411
+ }
412
+ }
413
+ }, "write"),
414
+ end: /* @__PURE__ */ __name(() => {
415
+ }, "end"),
416
+ fail: /* @__PURE__ */ __name(() => {
417
+ }, "fail")
418
+ },
419
+ frames: /* @__PURE__ */ __name(() => frames, "frames")
420
+ };
421
+ }
422
+ __name(createFrameBuffer, "createFrameBuffer");
423
+ function releaseGatedFrames(frames, text) {
424
+ const released = [];
425
+ let emitted = false;
426
+ for (const frame of frames) {
427
+ const event = decodeStreamEvent(frame);
428
+ if (event === null || event.kind === "text") {
429
+ if (!emitted && text.length > 0) {
430
+ released.push(encodeStreamEvent({
431
+ kind: "text",
432
+ text
433
+ }));
434
+ emitted = true;
435
+ }
436
+ continue;
437
+ }
438
+ released.push(encodeStreamEvent(event));
269
439
  }
270
- toolName;
271
- constructor(toolName) {
272
- super(`Tool "${toolName}" is not allowed for this role`), this.toolName = toolName;
273
- this.name = "ToolForbiddenError";
440
+ if (!emitted && text.length > 0) {
441
+ released.push(encodeStreamEvent({
442
+ kind: "text",
443
+ text
444
+ }));
274
445
  }
275
- };
276
- var ToolNotFoundError = class extends Error {
277
- static {
278
- __name(this, "ToolNotFoundError");
446
+ return released;
447
+ }
448
+ __name(releaseGatedFrames, "releaseGatedFrames");
449
+ async function runInputProcessors(processors, prompt, ctx) {
450
+ let current = prompt;
451
+ for (const processor of processors) {
452
+ try {
453
+ current = await processor.process(current, ctx);
454
+ } catch (error) {
455
+ throw new ProcessorFailedError("input", processor.name, error);
456
+ }
279
457
  }
280
- toolName;
281
- constructor(toolName) {
282
- super(`Tool "${toolName}" is not registered`), this.toolName = toolName;
283
- this.name = "ToolNotFoundError";
458
+ return current;
459
+ }
460
+ __name(runInputProcessors, "runInputProcessors");
461
+ async function runOutputProcessors(processors, answer, ctx) {
462
+ let text = answer.text;
463
+ for (const processor of processors) {
464
+ let verdict;
465
+ try {
466
+ verdict = await processor.process({
467
+ text,
468
+ toolCalls: answer.toolCalls
469
+ }, ctx);
470
+ } catch (error) {
471
+ throw new ProcessorFailedError("output", processor.name, error);
472
+ }
473
+ if (verdict.action === "reject") {
474
+ return {
475
+ text,
476
+ rejection: {
477
+ processor: processor.name,
478
+ reason: verdict.reason
479
+ }
480
+ };
481
+ }
482
+ if (verdict.action === "replace") {
483
+ text = verdict.text;
484
+ }
284
485
  }
285
- };
286
- var ToolInputInvalidError = class extends Error {
287
- static {
288
- __name(this, "ToolInputInvalidError");
486
+ return {
487
+ text
488
+ };
489
+ }
490
+ __name(runOutputProcessors, "runOutputProcessors");
491
+ async function gateFollowUps(processors, followUps, ctx) {
492
+ const kept = [];
493
+ for (const followUp of followUps) {
494
+ const settled = await runOutputProcessors(processors, {
495
+ text: followUp,
496
+ toolCalls: []
497
+ }, ctx);
498
+ if (settled.rejection === void 0 && settled.text.length > 0) {
499
+ kept.push(settled.text);
500
+ }
289
501
  }
290
- toolName;
291
- issues;
292
- constructor(toolName, issues) {
293
- super(`Invalid input for tool "${toolName}": ${issues.map((issue) => issue.message).join("; ")}`), this.toolName = toolName, this.issues = issues;
294
- this.name = "ToolInputInvalidError";
502
+ return kept;
503
+ }
504
+ __name(gateFollowUps, "gateFollowUps");
505
+ function resolveOutputGateMode(processors) {
506
+ if (processors.length === 0) {
507
+ return "off";
295
508
  }
296
- };
297
- var ToolRegistry = class {
298
- static {
299
- __name(this, "ToolRegistry");
509
+ return processors.every((processor) => processor.incremental !== void 0) ? "incremental" : "whole";
510
+ }
511
+ __name(resolveOutputGateMode, "resolveOutputGateMode");
512
+ function resolveGateLookback(processors) {
513
+ let lookback = 0;
514
+ for (const processor of processors) {
515
+ if (processor.incremental !== void 0) {
516
+ lookback = Math.max(lookback, processor.incremental.lookbackChars ?? DEFAULT_INCREMENTAL_LOOKBACK_CHARS);
517
+ }
300
518
  }
301
- entries = /* @__PURE__ */ new Map();
302
- register(spec, handler) {
303
- this.entries.set(spec.name, {
304
- spec,
305
- handler
306
- });
519
+ return lookback;
520
+ }
521
+ __name(resolveGateLookback, "resolveGateLookback");
522
+ function createIncrementalGate(options) {
523
+ const { processors, ctx, lookbackChars, writer } = options;
524
+ let raw = "";
525
+ let released = "";
526
+ let rejection;
527
+ let stalled = false;
528
+ const toolCalls = [];
529
+ let queue = Promise.resolve();
530
+ async function advance() {
531
+ if (raw.length === 0) {
532
+ return;
533
+ }
534
+ let settled;
535
+ try {
536
+ settled = await runOutputProcessors(processors, {
537
+ text: raw,
538
+ toolCalls
539
+ }, ctx);
540
+ } catch {
541
+ stalled = true;
542
+ return;
543
+ }
544
+ if (settled.rejection !== void 0) {
545
+ rejection = settled.rejection;
546
+ return;
547
+ }
548
+ const candidate = settled.text.slice(0, Math.max(0, settled.text.length - lookbackChars));
549
+ if (candidate.length <= released.length || !candidate.startsWith(released)) {
550
+ return;
551
+ }
552
+ const delta = candidate.slice(released.length);
553
+ released = candidate;
554
+ await writer.write(encodeStreamEvent({
555
+ kind: "text",
556
+ text: delta
557
+ }));
307
558
  }
308
- has(name) {
309
- return this.entries.has(name);
559
+ __name(advance, "advance");
560
+ async function handle(chunk) {
561
+ for (const line of decoder.decode(chunk).split("\n")) {
562
+ if (line.length === 0) {
563
+ continue;
564
+ }
565
+ const event = decodeStreamEvent(line);
566
+ if (event === null || rejection !== void 0) {
567
+ continue;
568
+ }
569
+ if (event.kind === "text") {
570
+ raw += event.text;
571
+ if (!stalled) {
572
+ await advance();
573
+ }
574
+ continue;
575
+ }
576
+ if (event.kind === "tool-input-available") {
577
+ toolCalls.push({
578
+ id: event.id,
579
+ name: event.name,
580
+ input: event.input
581
+ });
582
+ }
583
+ await writer.write(encodeStreamEvent(event));
584
+ }
310
585
  }
311
- spec(name) {
312
- return this.entries.get(name)?.spec;
586
+ __name(handle, "handle");
587
+ return {
588
+ writer: {
589
+ write: /* @__PURE__ */ __name((chunk) => {
590
+ queue = queue.then(() => handle(chunk));
591
+ return queue;
592
+ }, "write"),
593
+ end: /* @__PURE__ */ __name(() => {
594
+ }, "end"),
595
+ fail: /* @__PURE__ */ __name(() => {
596
+ }, "fail")
597
+ },
598
+ settled: /* @__PURE__ */ __name(() => queue, "settled"),
599
+ released: /* @__PURE__ */ __name(() => released, "released"),
600
+ rejection: /* @__PURE__ */ __name(() => rejection, "rejection")
601
+ };
602
+ }
603
+ __name(createIncrementalGate, "createIncrementalGate");
604
+ function gateTail(processors, released, text) {
605
+ if (!text.startsWith(released)) {
606
+ throw new ProcessorFailedError("output", processors.map((processor) => processor.name).join(" \u2192 "), new Error(`the gated answer is not an extension of the ${released.length} characters this chain already released \u2014 a processor that declares "incremental" promises its result on a prefix stays a prefix of its result on the whole answer, outside the last lookbackChars`));
313
607
  }
314
- allSpecs() {
315
- return [
316
- ...this.entries.values()
317
- ].map((entry) => entry.spec);
608
+ return text.slice(released.length);
609
+ }
610
+ __name(gateTail, "gateTail");
611
+
612
+ // src/structured-output.ts
613
+ var StructuredOutputError = class extends Error {
614
+ static {
615
+ __name(this, "StructuredOutputError");
318
616
  }
319
- /** The tools to offer the model for this actor+agent, after the two filter layers. */
320
- async definitionsFor(actor, policy, allowedTools) {
321
- const roleScoped = await filterToolsByRole(this.allSpecs(), actor, policy);
322
- const allowScoped = filterToolsByAllowList(roleScoped, allowedTools);
323
- return allowScoped.map((spec) => ({
324
- name: spec.name,
325
- kind: spec.kind,
326
- description: spec.description,
327
- inputSchema: spec.inputSchema
328
- }));
617
+ /** Why it failed the schema. Empty only when the text was not JSON at all. */
618
+ issues;
619
+ /** The last text the model produced, verbatim. */
620
+ text;
621
+ /** How many model calls were spent trying (1 = the formatting pass, no repairs). */
622
+ attempts;
623
+ constructor(issues, text, attempts) {
624
+ const detail = issues.length > 0 ? issues.map((issue4) => issue4.message).join("; ") : "the model did not reply with JSON";
625
+ super(`Structured output invalid after ${attempts} attempt(s): ${detail}`);
626
+ this.name = "StructuredOutputError";
627
+ this.issues = issues;
628
+ this.text = text;
629
+ this.attempts = attempts;
329
630
  }
330
- /** Run a tool. Re-checks the role (defense-in-depth) and re-parses the input via Zod. */
331
- async invoke(name, input, ctx, policy) {
332
- const entry = this.entries.get(name);
333
- if (entry === void 0) {
334
- throw new ToolNotFoundError(name);
335
- }
336
- if (!await policy.can(ctx.actor, entry.spec)) {
337
- throw new ToolForbiddenError(name);
631
+ };
632
+ function extractJson(text) {
633
+ const candidates = [
634
+ text,
635
+ text.match(/\{[\s\S]*\}/)?.[0],
636
+ text.match(/\[[\s\S]*\]/)?.[0]
637
+ ];
638
+ for (const candidate of candidates) {
639
+ if (candidate === void 0) {
640
+ continue;
338
641
  }
339
- const validation = await entry.spec.inputSchema["~standard"].validate(input);
340
- if (validation.issues !== void 0) {
341
- throw new ToolInputInvalidError(name, validation.issues);
642
+ try {
643
+ return JSON.parse(candidate);
644
+ } catch {
342
645
  }
343
- return entry.handler.execute(validation.value, ctx);
344
- }
345
- };
346
- var DefaultRolesPolicy = class {
347
- static {
348
- __name(this, "DefaultRolesPolicy");
349
646
  }
350
- defaultRoles;
351
- constructor(defaultRoles = [
352
- "ADMIN"
353
- ]) {
354
- this.defaultRoles = defaultRoles;
355
- }
356
- can(actor, tool) {
357
- const allowed = tool.roles ?? this.defaultRoles;
358
- return (actor.roles ?? []).some((role) => allowed.includes(role));
647
+ return void 0;
648
+ }
649
+ __name(extractJson, "extractJson");
650
+ async function validateStructured(schema, text, reported) {
651
+ const candidate = reported !== void 0 ? reported : extractJson(text);
652
+ if (candidate === void 0) {
653
+ return {
654
+ ok: false,
655
+ issues: []
656
+ };
359
657
  }
360
- };
361
-
362
- // src/agent-loop.ts
363
- import { createHash } from "node:crypto";
364
- import { trace } from "@dudousxd/nestjs-diagnostics";
658
+ const result = await schema["~standard"].validate(candidate);
659
+ return result.issues !== void 0 ? {
660
+ ok: false,
661
+ issues: result.issues
662
+ } : {
663
+ ok: true,
664
+ value: result.value
665
+ };
666
+ }
667
+ __name(validateStructured, "validateStructured");
668
+ var DEFAULT_STRUCTURED_OUTPUT_INSTRUCTION = "Restate the assistant's final answer as a single JSON value matching the required schema. Use only information already present in the conversation \u2014 add nothing, and answer nothing that was not asked. Reply with ONLY the JSON: no prose, no explanation, no code fences.";
669
+ function repairInstruction(instruction, issues) {
670
+ const detail = issues.length > 0 ? issues.map((issue4) => `${(issue4.path ?? []).join(".") || "(root)"}: ${issue4.message}`).join("\n") : "the previous reply was not valid JSON";
671
+ return `${instruction}
365
672
 
366
- // src/diagnostics.ts
367
- import { emit } from "@dudousxd/nestjs-diagnostics";
368
- function publishAgentRunStarted(payload) {
369
- emit("agent", "run.started", payload);
673
+ Your previous reply was rejected:
674
+ ${detail}`;
370
675
  }
371
- __name(publishAgentRunStarted, "publishAgentRunStarted");
372
- function publishAgentMessage(payload) {
373
- emit("agent", "message", payload);
676
+ __name(repairInstruction, "repairInstruction");
677
+
678
+ // src/elicitation.ts
679
+ function optionValues(question) {
680
+ return new Set(question.options.map((option) => option.value));
374
681
  }
375
- __name(publishAgentMessage, "publishAgentMessage");
376
- function publishAgentToolCall(payload) {
377
- emit("agent", "tool-call", payload);
682
+ __name(optionValues, "optionValues");
683
+ function normalizeElicitationReply(reply) {
684
+ if (typeof reply !== "object" || reply === null) {
685
+ return {
686
+ answers: {}
687
+ };
688
+ }
689
+ const candidate = reply;
690
+ if (typeof candidate.answers === "object" && candidate.answers !== null) {
691
+ return reply;
692
+ }
693
+ const answeredByRef = candidate.answeredByRef ?? candidate.executedByRef;
694
+ return {
695
+ answers: {},
696
+ ...candidate.approved === false || candidate.skipped === true ? {
697
+ skipped: true
698
+ } : {},
699
+ ...answeredByRef !== void 0 ? {
700
+ answeredByRef
701
+ } : {}
702
+ };
378
703
  }
379
- __name(publishAgentToolCall, "publishAgentToolCall");
380
- function publishAgentQuotaExceeded(payload) {
381
- emit("agent", "quota.exceeded", payload);
704
+ __name(normalizeElicitationReply, "normalizeElicitationReply");
705
+ function resolveElicitation(request, raw) {
706
+ const reply = normalizeElicitationReply(raw);
707
+ const answers = {};
708
+ const defaulted = [];
709
+ for (const question of request.questions) {
710
+ const submitted = reply.skipped === true ? void 0 : reply.answers[question.id];
711
+ if (submitted === void 0) {
712
+ answers[question.id] = [
713
+ ...question.defaults ?? []
714
+ ];
715
+ defaulted.push(question.id);
716
+ continue;
717
+ }
718
+ const allowed = optionValues(question);
719
+ const valid = question.allowFreeText === true ? submitted : submitted.filter((value) => allowed.has(value));
720
+ answers[question.id] = question.multiple === true ? valid : valid.slice(0, 1);
721
+ }
722
+ return {
723
+ answers,
724
+ skipped: reply.skipped === true,
725
+ defaulted
726
+ };
382
727
  }
383
- __name(publishAgentQuotaExceeded, "publishAgentQuotaExceeded");
384
- function publishAgentRunFinished(payload) {
385
- emit("agent", "run.finished", payload);
728
+ __name(resolveElicitation, "resolveElicitation");
729
+ function settleElicitation(request, reply) {
730
+ const outcome = resolveElicitation(request, reply);
731
+ return {
732
+ ...outcome,
733
+ summary: renderElicitationAnswers(request, outcome)
734
+ };
386
735
  }
387
- __name(publishAgentRunFinished, "publishAgentRunFinished");
388
- function publishAgentRunFailed(payload) {
389
- emit("agent", "run.failed", payload);
736
+ __name(settleElicitation, "settleElicitation");
737
+ function renderElicitationAnswers(request, outcome) {
738
+ const lines = request.questions.map((question) => {
739
+ const chosen = outcome.answers[question.id] ?? [];
740
+ const labels = chosen.map((value) => question.options.find((option) => option.value === value)?.label ?? value);
741
+ return `${question.prompt} \u2192 ${labels.length > 0 ? labels.join(", ") : "(no answer)"}`;
742
+ });
743
+ const preface = outcome.skipped ? "The user declined to answer and asked you to proceed on these assumptions:" : "The user answered:";
744
+ return `${preface}
745
+ ${lines.join("\n")}`;
390
746
  }
391
- __name(publishAgentRunFailed, "publishAgentRunFailed");
392
- function publishAgentDelegated(payload) {
393
- emit("agent", "delegated", payload);
747
+ __name(renderElicitationAnswers, "renderElicitationAnswers");
748
+ var ASK_TOOL_NAME = "ask";
749
+ var MAX_ASK_QUESTIONS = 5;
750
+ function issue(path, message) {
751
+ return {
752
+ message,
753
+ path
754
+ };
394
755
  }
395
- __name(publishAgentDelegated, "publishAgentDelegated");
396
- function publishAgentRetrieved(payload) {
397
- emit("agent", "retrieved", payload);
756
+ __name(issue, "issue");
757
+ function parseOption(raw, path, issues) {
758
+ if (typeof raw !== "object" || raw === null) {
759
+ issues.push(issue(path, "must be an object"));
760
+ return void 0;
761
+ }
762
+ const candidate = raw;
763
+ if (typeof candidate.value !== "string" || candidate.value.length === 0) {
764
+ issues.push(issue([
765
+ ...path,
766
+ "value"
767
+ ], "must be a non-empty string"));
768
+ return void 0;
769
+ }
770
+ if (typeof candidate.label !== "string" || candidate.label.length === 0) {
771
+ issues.push(issue([
772
+ ...path,
773
+ "label"
774
+ ], "must be a non-empty string"));
775
+ return void 0;
776
+ }
777
+ return {
778
+ value: candidate.value,
779
+ label: candidate.label,
780
+ ...typeof candidate.hotkey === "string" ? {
781
+ hotkey: candidate.hotkey
782
+ } : {}
783
+ };
398
784
  }
399
- __name(publishAgentRetrieved, "publishAgentRetrieved");
400
- function publishAgentToolRetry(payload) {
401
- emit("agent", "tool.retry", payload);
402
- }
403
- __name(publishAgentToolRetry, "publishAgentToolRetry");
404
- var AGENT_SPAN_EVENTS = [
405
- "llm.turn",
406
- "tool.execution",
407
- "retrieval",
408
- "follow-ups"
409
- ];
410
- var AGENT_DIAGNOSTIC_EVENTS = [
411
- "run.started",
412
- "message",
413
- "tool-call",
414
- "quota.exceeded",
415
- "run.finished",
416
- "run.failed",
417
- "delegated",
418
- "retrieved",
419
- "tool.retry"
420
- ];
421
- function agentDiagnosticKey(event) {
422
- return `agent:${event}`;
423
- }
424
- __name(agentDiagnosticKey, "agentDiagnosticKey");
425
-
426
- // src/tool-retry.ts
427
- function hasTransientShape(error) {
428
- if (typeof error !== "object" || error === null) {
429
- return false;
430
- }
431
- const code = "code" in error ? error.code : void 0;
432
- const errno = "errno" in error ? error.errno : void 0;
433
- const sqlState = "sqlState" in error ? error.sqlState : void 0;
434
- if (code === 1213 || code === 1205 || errno === 1213 || errno === 1205) {
435
- return true;
436
- }
437
- if (code === "ER_LOCK_DEADLOCK" || code === "ER_LOCK_WAIT_TIMEOUT") {
438
- return true;
785
+ __name(parseOption, "parseOption");
786
+ function parseQuestion(raw, path, issues) {
787
+ if (typeof raw !== "object" || raw === null) {
788
+ issues.push(issue(path, "must be an object"));
789
+ return void 0;
439
790
  }
440
- if (code === "40001" || code === "40P01" || sqlState === "40001" || sqlState === "40P01") {
441
- return true;
791
+ const candidate = raw;
792
+ if (typeof candidate.id !== "string" || candidate.id.length === 0) {
793
+ issues.push(issue([
794
+ ...path,
795
+ "id"
796
+ ], "must be a non-empty string"));
797
+ return void 0;
442
798
  }
443
- if (code === "SQLITE_BUSY") {
444
- return true;
799
+ if (typeof candidate.prompt !== "string" || candidate.prompt.length === 0) {
800
+ issues.push(issue([
801
+ ...path,
802
+ "prompt"
803
+ ], "must be a non-empty string"));
804
+ return void 0;
445
805
  }
446
- const message = "message" in error ? error.message : void 0;
447
- return typeof message === "string" && /deadlock|lock wait timeout|serialization failure/i.test(message);
448
- }
449
- __name(hasTransientShape, "hasTransientShape");
450
- function isTransientToolError(error) {
451
- if (hasTransientShape(error)) {
452
- return true;
806
+ if (!Array.isArray(candidate.options) || candidate.options.length === 0) {
807
+ issues.push(issue([
808
+ ...path,
809
+ "options"
810
+ ], "must be a non-empty array"));
811
+ return void 0;
453
812
  }
454
- if (typeof error === "object" && error !== null && "cause" in error) {
455
- const cause = error.cause;
456
- if (cause !== void 0 && cause !== error && hasTransientShape(cause)) {
457
- return true;
813
+ const options = [];
814
+ for (const [index, rawOption] of candidate.options.entries()) {
815
+ const option = parseOption(rawOption, [
816
+ ...path,
817
+ "options",
818
+ index
819
+ ], issues);
820
+ if (option !== void 0) {
821
+ options.push(option);
458
822
  }
459
823
  }
460
- return false;
461
- }
462
- __name(isTransientToolError, "isTransientToolError");
463
- var DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS = 2;
464
- var DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS = 150;
465
- function resolveToolTransientRetryNumbers(setting) {
466
- if (setting === false) {
467
- return false;
824
+ if (!Array.isArray(candidate.defaults) || candidate.defaults.length === 0) {
825
+ issues.push(issue([
826
+ ...path,
827
+ "defaults"
828
+ ], "must pre-pick at least one option \u2014 say what you would choose so the user can just confirm"));
829
+ return void 0;
830
+ }
831
+ const offered = new Set(options.map((option) => option.value));
832
+ const defaults = candidate.defaults.filter((value) => typeof value === "string" && offered.has(value));
833
+ if (defaults.length === 0) {
834
+ issues.push(issue([
835
+ ...path,
836
+ "defaults"
837
+ ], "must name values that appear in this question's options"));
838
+ return void 0;
468
839
  }
469
840
  return {
470
- attempts: setting?.attempts ?? DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS,
471
- backoffMs: setting?.backoffMs ?? DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS
841
+ id: candidate.id,
842
+ prompt: candidate.prompt,
843
+ options,
844
+ defaults,
845
+ ...candidate.multiple === true ? {
846
+ multiple: true
847
+ } : {},
848
+ ...candidate.allowFreeText === true ? {
849
+ allowFreeText: true
850
+ } : {}
472
851
  };
473
852
  }
474
- __name(resolveToolTransientRetryNumbers, "resolveToolTransientRetryNumbers");
475
- function delay(ms) {
476
- return new Promise((resolve) => {
477
- setTimeout(resolve, ms);
478
- });
479
- }
480
- __name(delay, "delay");
481
- async function invokeWithTransientRetry(fn, setting, options) {
482
- if (setting === false) {
483
- return fn();
484
- }
485
- const attempts = setting.attempts ?? DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS;
486
- const backoffMs = setting.backoffMs ?? DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS;
487
- const classify = setting.classify ?? isTransientToolError;
488
- let attempt = 1;
489
- for (; ; ) {
490
- try {
491
- return await fn();
492
- } catch (error) {
493
- if (options?.isControlFlowError?.(error) === true) {
494
- throw error;
853
+ __name(parseQuestion, "parseQuestion");
854
+ var ASK_JSON_SCHEMA = {
855
+ type: "object",
856
+ additionalProperties: false,
857
+ properties: {
858
+ preamble: {
859
+ type: "string",
860
+ description: 'One sentence shown above the form, e.g. "Three questions before I start. I have pre-picked what I would choose, so confirming is enough."'
861
+ },
862
+ questions: {
863
+ type: "array",
864
+ minItems: 1,
865
+ maxItems: MAX_ASK_QUESTIONS,
866
+ items: {
867
+ type: "object",
868
+ additionalProperties: false,
869
+ required: [
870
+ "id",
871
+ "prompt",
872
+ "options",
873
+ "defaults"
874
+ ],
875
+ properties: {
876
+ id: {
877
+ type: "string",
878
+ description: "Unique within this call; answers come back under it."
879
+ },
880
+ prompt: {
881
+ type: "string"
882
+ },
883
+ multiple: {
884
+ type: "boolean",
885
+ description: "Allow more than one option to be chosen."
886
+ },
887
+ allowFreeText: {
888
+ type: "boolean",
889
+ description: "Accept an answer that is not an option."
890
+ },
891
+ defaults: {
892
+ type: "array",
893
+ minItems: 1,
894
+ items: {
895
+ type: "string"
896
+ },
897
+ description: "REQUIRED. The option values you would pick yourself, so the user can confirm rather than decide."
898
+ },
899
+ options: {
900
+ type: "array",
901
+ minItems: 2,
902
+ items: {
903
+ type: "object",
904
+ additionalProperties: false,
905
+ required: [
906
+ "value",
907
+ "label"
908
+ ],
909
+ properties: {
910
+ value: {
911
+ type: "string"
912
+ },
913
+ label: {
914
+ type: "string"
915
+ },
916
+ hotkey: {
917
+ type: "string",
918
+ maxLength: 1
919
+ }
920
+ }
921
+ }
922
+ }
923
+ }
495
924
  }
496
- const attemptsRemain = attempt < attempts;
497
- if (!attemptsRemain || !classify(error)) {
498
- throw error;
925
+ }
926
+ },
927
+ required: [
928
+ "questions"
929
+ ]
930
+ };
931
+ var askInputSchema = {
932
+ "~standard": {
933
+ version: 1,
934
+ vendor: "nestjs-agent",
935
+ validate: /* @__PURE__ */ __name((value) => {
936
+ const issues = [];
937
+ if (typeof value !== "object" || value === null) {
938
+ return {
939
+ issues: [
940
+ issue([], "must be an object")
941
+ ]
942
+ };
499
943
  }
500
- options?.onRetry?.(attempt, error);
501
- await delay(backoffMs * attempt);
502
- attempt += 1;
944
+ const candidate = value;
945
+ if (!Array.isArray(candidate.questions) || candidate.questions.length === 0) {
946
+ return {
947
+ issues: [
948
+ issue([
949
+ "questions"
950
+ ], "must be a non-empty array")
951
+ ]
952
+ };
953
+ }
954
+ if (candidate.questions.length > MAX_ASK_QUESTIONS) {
955
+ return {
956
+ issues: [
957
+ issue([
958
+ "questions"
959
+ ], `must hold at most ${MAX_ASK_QUESTIONS} questions`)
960
+ ]
961
+ };
962
+ }
963
+ const questions = [];
964
+ for (const [index, raw] of candidate.questions.entries()) {
965
+ const question = parseQuestion(raw, [
966
+ "questions",
967
+ index
968
+ ], issues);
969
+ if (question !== void 0) {
970
+ questions.push(question);
971
+ }
972
+ }
973
+ if (issues.length > 0) {
974
+ return {
975
+ issues
976
+ };
977
+ }
978
+ return {
979
+ value: {
980
+ questions,
981
+ ...typeof candidate.preamble === "string" ? {
982
+ preamble: candidate.preamble
983
+ } : {}
984
+ }
985
+ };
986
+ }, "validate"),
987
+ jsonSchema: {
988
+ input: /* @__PURE__ */ __name(() => ASK_JSON_SCHEMA, "input")
503
989
  }
504
990
  }
991
+ };
992
+ var ASK_TOOL_DESCRIPTION = "Ask the user to settle the scope of the work before you do it. Use it when a reasonable person would produce a materially different result depending on the answer \u2014 not to confirm something the conversation already says. Every question must pre-pick the answer you would choose, so the user can confirm instead of deciding. The user may decline, in which case you proceed on those pre-picked answers.";
993
+ function askToolDefinition() {
994
+ return {
995
+ name: ASK_TOOL_NAME,
996
+ kind: "ask",
997
+ description: ASK_TOOL_DESCRIPTION,
998
+ inputSchema: askInputSchema
999
+ };
505
1000
  }
506
- __name(invokeWithTransientRetry, "invokeWithTransientRetry");
1001
+ __name(askToolDefinition, "askToolDefinition");
1002
+ var DEFAULT_INTAKE_PREAMBLE = "A few questions before I start. I have pre-picked what I would choose, so confirming is enough.";
507
1003
 
508
- // src/agent-loop.ts
509
- function resolveCostUsd(usage, reportedCostUsd, price) {
510
- if (reportedCostUsd !== void 0) {
511
- return reportedCostUsd;
512
- }
513
- return price === void 0 ? null : estimateCost(usage, price);
1004
+ // src/skills.ts
1005
+ var GLOBAL_SCOPE = "global";
1006
+ function actorScope(actor) {
1007
+ return `actor:${actor.id}`;
514
1008
  }
515
- __name(resolveCostUsd, "resolveCostUsd");
516
- function buildContextBlock(passages) {
517
- const items = passages.map((passage, index) => {
518
- const label = passage.source !== void 0 ? ` (${passage.source})` : "";
519
- return `[${index + 1}]${label} ${passage.text}`;
520
- }).join("\n\n");
521
- return `<retrieved_context>
522
- ${items}
523
- </retrieved_context>
524
- Use the retrieved context above to answer when relevant, and cite sources by their bracket number.`;
1009
+ __name(actorScope, "actorScope");
1010
+ function tenantScope(tenantRef) {
1011
+ return `tenant:${tenantRef}`;
525
1012
  }
526
- __name(buildContextBlock, "buildContextBlock");
527
- var QuotaExceededError = class extends Error {
528
- static {
529
- __name(this, "QuotaExceededError");
530
- }
531
- constructor() {
532
- super("Daily token quota exceeded");
533
- this.name = "QuotaExceededError";
534
- }
1013
+ __name(tenantScope, "tenantScope");
1014
+ var defaultScopeResolver = {
1015
+ resolve: /* @__PURE__ */ __name(({ actor }) => [
1016
+ actorScope(actor),
1017
+ ...actor.tenantRef !== void 0 ? [
1018
+ tenantScope(actor.tenantRef)
1019
+ ] : [],
1020
+ GLOBAL_SCOPE
1021
+ ], "resolve")
535
1022
  };
536
- var MAX_DELEGATION_DEPTH = 5;
537
- async function resolvePrompt(prompt, ctx) {
538
- return typeof prompt === "function" ? prompt(ctx) : prompt;
1023
+ function staticSkillProvider(skills) {
1024
+ return {
1025
+ list: /* @__PURE__ */ __name(({ scopes }) => skills.filter((skill) => scopes.includes(skill.scope)), "list"),
1026
+ load: /* @__PURE__ */ __name(({ name, scope }) => skills.find((skill) => skill.name === name && skill.scope === scope)?.body ?? null, "load")
1027
+ };
539
1028
  }
540
- __name(resolvePrompt, "resolvePrompt");
541
- async function resolveSystemPrompt(deps, input) {
542
- const ctx = {
543
- actor: input.actor,
544
- agentName: input.agentName ?? "default",
545
- ...input.pageContext !== void 0 ? {
546
- pageContext: input.pageContext
547
- } : {}
1029
+ __name(staticSkillProvider, "staticSkillProvider");
1030
+ function compositeSkillProvider(providers) {
1031
+ return {
1032
+ list: /* @__PURE__ */ __name(async ({ scopes, ctx }) => {
1033
+ const seen = /* @__PURE__ */ new Set();
1034
+ const merged = [];
1035
+ for (const provider of providers) {
1036
+ for (const summary of await provider.list({
1037
+ scopes,
1038
+ ctx
1039
+ })) {
1040
+ const key = `${summary.scope}\0${summary.name}`;
1041
+ if (!seen.has(key)) {
1042
+ seen.add(key);
1043
+ merged.push(summary);
1044
+ }
1045
+ }
1046
+ }
1047
+ return merged;
1048
+ }, "list"),
1049
+ load: /* @__PURE__ */ __name(async ({ name, scope, ctx }) => {
1050
+ for (const provider of providers) {
1051
+ const body = await provider.load({
1052
+ name,
1053
+ scope,
1054
+ ctx
1055
+ });
1056
+ if (body !== null) {
1057
+ return body;
1058
+ }
1059
+ }
1060
+ return null;
1061
+ }, "load")
548
1062
  };
549
- const sections = [
550
- await resolvePrompt(deps.systemPrompt, ctx)
551
- ];
552
- for (const contribute of deps.promptContributors ?? []) {
553
- const section = await contribute(ctx);
554
- if (section !== null && section.length > 0) {
555
- sections.push(section);
1063
+ }
1064
+ __name(compositeSkillProvider, "compositeSkillProvider");
1065
+ var DEFAULT_MAX_SKILLS = 20;
1066
+ function resolveSkillCatalog(summaries, scopes, maxSkills = DEFAULT_MAX_SKILLS) {
1067
+ const rank = new Map(scopes.map((scope, index) => [
1068
+ scope,
1069
+ index
1070
+ ]));
1071
+ const byName = /* @__PURE__ */ new Map();
1072
+ for (const summary of summaries) {
1073
+ if (!rank.has(summary.scope)) {
1074
+ continue;
1075
+ }
1076
+ const existing = byName.get(summary.name);
1077
+ if (existing === void 0) {
1078
+ byName.set(summary.name, [
1079
+ summary
1080
+ ]);
1081
+ } else {
1082
+ existing.push(summary);
556
1083
  }
557
1084
  }
558
- return sections.join("\n\n");
559
- }
560
- __name(resolveSystemPrompt, "resolveSystemPrompt");
561
- function extractTask(input) {
562
- if (typeof input === "object" && input !== null && "task" in input) {
563
- const task = input.task;
564
- if (typeof task === "string") {
565
- return task;
1085
+ const resolved = [];
1086
+ for (const candidates of byName.values()) {
1087
+ const ordered = [
1088
+ ...candidates
1089
+ ].sort((a, b) => (rank.get(a.scope) ?? 0) - (rank.get(b.scope) ?? 0));
1090
+ const winner = ordered[0];
1091
+ if (winner === void 0) {
1092
+ continue;
566
1093
  }
1094
+ const shadows = ordered.slice(1).map((candidate) => candidate.scope);
1095
+ resolved.push({
1096
+ name: winner.name,
1097
+ description: winner.description,
1098
+ scope: winner.scope,
1099
+ ...shadows.length > 0 ? {
1100
+ shadows
1101
+ } : {}
1102
+ });
567
1103
  }
568
- return JSON.stringify(input);
1104
+ resolved.sort((a, b) => {
1105
+ const byScope = (rank.get(a.scope) ?? 0) - (rank.get(b.scope) ?? 0);
1106
+ return byScope !== 0 ? byScope : compare(a.name, b.name);
1107
+ });
1108
+ return {
1109
+ entries: resolved.slice(0, maxSkills),
1110
+ omitted: Math.max(0, resolved.length - maxSkills)
1111
+ };
569
1112
  }
570
- __name(extractTask, "extractTask");
571
- function deriveTitle(userText) {
572
- const trimmed = userText.trim().replace(/\s+/g, " ");
573
- return trimmed.length > 60 ? `${trimmed.slice(0, 57)}...` : trimmed || "New chat";
1113
+ __name(resolveSkillCatalog, "resolveSkillCatalog");
1114
+ function compare(a, b) {
1115
+ return a < b ? -1 : a > b ? 1 : 0;
574
1116
  }
575
- __name(deriveTitle, "deriveTitle");
576
- var ToolTimeoutError = class ToolTimeoutError2 extends Error {
577
- static {
578
- __name(this, "ToolTimeoutError");
579
- }
580
- constructor(toolName, ms) {
581
- super(`Tool "${toolName}" exceeded its ${ms}ms timeout`);
582
- this.name = "ToolTimeoutError";
1117
+ __name(compare, "compare");
1118
+ async function offerSkills(config, ctx) {
1119
+ const scopes = await (config.scopes ?? defaultScopeResolver).resolve(ctx);
1120
+ const summaries = await config.provider.list({
1121
+ scopes,
1122
+ ctx
1123
+ });
1124
+ return {
1125
+ scopes,
1126
+ ...resolveSkillCatalog(summaries, scopes, config.maxSkills)
1127
+ };
1128
+ }
1129
+ __name(offerSkills, "offerSkills");
1130
+ function buildSkillsBlock(entries) {
1131
+ const lines = entries.map((entry) => {
1132
+ const shadows = entry.shadows === void 0 ? "" : ` (overrides the one from ${entry.shadows.join(", ")})`;
1133
+ return `- ${entry.name} [${entry.scope}] \u2014 ${entry.description}${shadows}`;
1134
+ });
1135
+ return `<skills>
1136
+ Procedures available to you. Each line is a name, the scope it came from, and what it covers \u2014 not the procedure itself.
1137
+ Call the \`${SKILL_TOOL_NAME}\` tool with a name from this list to read one BEFORE doing the task it covers, and follow what it says; never guess its contents from the description.
1138
+ A skill from a narrower scope overrides a same-named one from a wider scope. When you follow one that overrides another, say which scope it came from, so the user knows their setting differs from the wider default.
1139
+ ${lines.join("\n")}
1140
+ </skills>`;
1141
+ }
1142
+ __name(buildSkillsBlock, "buildSkillsBlock");
1143
+ var SKILL_TOOL_NAME = "skill";
1144
+ var SKILL_JSON_SCHEMA = {
1145
+ type: "object",
1146
+ additionalProperties: false,
1147
+ required: [
1148
+ "name"
1149
+ ],
1150
+ properties: {
1151
+ name: {
1152
+ type: "string",
1153
+ description: "The name of a skill from the <skills> list, exactly as written there."
1154
+ }
583
1155
  }
584
1156
  };
585
- function withToolTimeout(work, ms, toolName) {
586
- return new Promise((resolve, reject) => {
587
- const timer = setTimeout(() => reject(new ToolTimeoutError(toolName, ms)), ms);
588
- work.then((value) => {
589
- clearTimeout(timer);
590
- resolve(value);
591
- }, (error) => {
592
- clearTimeout(timer);
593
- reject(error);
594
- });
595
- });
1157
+ function issue2(path, message) {
1158
+ return {
1159
+ message,
1160
+ path
1161
+ };
596
1162
  }
597
- __name(withToolTimeout, "withToolTimeout");
598
- function parseFollowUps(text, count) {
599
- const source = text.match(/\[[\s\S]*\]/)?.[0] ?? text;
600
- try {
601
- const parsed = JSON.parse(source);
602
- if (Array.isArray(parsed)) {
603
- return parsed.filter((item) => typeof item === "string").slice(0, count);
1163
+ __name(issue2, "issue");
1164
+ var skillInputSchema = {
1165
+ "~standard": {
1166
+ version: 1,
1167
+ vendor: "nestjs-agent",
1168
+ validate: /* @__PURE__ */ __name((value) => {
1169
+ if (typeof value !== "object" || value === null) {
1170
+ return {
1171
+ issues: [
1172
+ issue2([], "must be an object")
1173
+ ]
1174
+ };
1175
+ }
1176
+ const candidate = value;
1177
+ if (typeof candidate.name !== "string" || candidate.name.length === 0) {
1178
+ return {
1179
+ issues: [
1180
+ issue2([
1181
+ "name"
1182
+ ], "must be a non-empty string")
1183
+ ]
1184
+ };
1185
+ }
1186
+ return {
1187
+ value: {
1188
+ name: candidate.name
1189
+ }
1190
+ };
1191
+ }, "validate"),
1192
+ jsonSchema: {
1193
+ input: /* @__PURE__ */ __name(() => SKILL_JSON_SCHEMA, "input")
604
1194
  }
605
- } catch {
606
1195
  }
607
- return [];
1196
+ };
1197
+ var SKILL_TOOL_DESCRIPTION = "Read one of the procedures listed in <skills>. Call it before doing a task a listed skill covers, and follow what it returns. It performs nothing and changes nothing \u2014 it only gives you instructions you do not yet have.";
1198
+ function skillToolDefinition() {
1199
+ return {
1200
+ name: SKILL_TOOL_NAME,
1201
+ kind: "skill",
1202
+ description: SKILL_TOOL_DESCRIPTION,
1203
+ inputSchema: skillInputSchema
1204
+ };
608
1205
  }
609
- __name(parseFollowUps, "parseFollowUps");
610
- async function generateFollowUps(model, messages, count) {
611
- const discard = {
612
- write: /* @__PURE__ */ __name(() => {
613
- }, "write"),
614
- end: /* @__PURE__ */ __name(() => {
615
- }, "end"),
616
- fail: /* @__PURE__ */ __name(() => {
617
- }, "fail")
1206
+ __name(skillToolDefinition, "skillToolDefinition");
1207
+ function withSkillTool({ tools, enabled }) {
1208
+ return enabled ? [
1209
+ ...tools,
1210
+ skillToolDefinition()
1211
+ ] : tools;
1212
+ }
1213
+ __name(withSkillTool, "withSkillTool");
1214
+ async function loadSkill(config, offer, name, ctx) {
1215
+ const entry = offer.entries.find((candidate) => candidate.name === name);
1216
+ if (entry === void 0) {
1217
+ const available = offer.entries.map((candidate) => candidate.name).join(", ");
1218
+ return {
1219
+ ok: false,
1220
+ error: available.length === 0 ? `No skill named "${name}" is available to you, and no others are either.` : `No skill named "${name}" is available to you. Available: ${available}.`
1221
+ };
1222
+ }
1223
+ const body = await config.provider.load({
1224
+ name: entry.name,
1225
+ scope: entry.scope,
1226
+ ctx
1227
+ });
1228
+ if (body === null) {
1229
+ return {
1230
+ ok: false,
1231
+ error: `Skill "${name}" is listed but its body could not be read.`
1232
+ };
1233
+ }
1234
+ return {
1235
+ ok: true,
1236
+ skill: {
1237
+ name: entry.name,
1238
+ description: entry.description,
1239
+ scope: entry.scope,
1240
+ body
1241
+ },
1242
+ ...entry.shadows !== void 0 ? {
1243
+ shadows: entry.shadows
1244
+ } : {}
618
1245
  };
619
- const turn = await model.runTurn({
620
- system: `Based on the conversation so far, propose up to ${count} short, distinct follow-up questions the user is likely to ask next. Respond with ONLY a JSON array of strings \u2014 no prose, no code fences.`,
621
- messages,
622
- tools: [],
1246
+ }
1247
+ __name(loadSkill, "loadSkill");
1248
+ function skillWriteVerdict(request) {
1249
+ const { scope, scopes, author, actor } = request;
1250
+ if (!scopes.includes(scope)) {
1251
+ return {
1252
+ allowed: false,
1253
+ reason: `"${scope}" is not a scope this actor belongs to`
1254
+ };
1255
+ }
1256
+ const own = actorScope(actor);
1257
+ if (scope === own) {
1258
+ return {
1259
+ allowed: true
1260
+ };
1261
+ }
1262
+ if (author.kind !== "human") {
1263
+ return {
1264
+ allowed: false,
1265
+ reason: `only a human may author a skill at "${scope}"; an agent may write only at "${own}"`
1266
+ };
1267
+ }
1268
+ return request.elevated === true ? {
1269
+ allowed: true
1270
+ } : {
1271
+ allowed: false,
1272
+ reason: `authoring at "${scope}" requires an elevated human author`
1273
+ };
1274
+ }
1275
+ __name(skillWriteVerdict, "skillWriteVerdict");
1276
+
1277
+ // src/memory.ts
1278
+ var DEFAULT_MAX_MEMORIES = 20;
1279
+ var DEFAULT_MAX_FACT_CHARS = 240;
1280
+ function resolveMemoryDigest({ records, scopes, maxMemories = DEFAULT_MAX_MEMORIES, ranked = false }) {
1281
+ const rank = new Map(scopes.map((scope, index) => [
1282
+ scope,
1283
+ index
1284
+ ]));
1285
+ const byKey = /* @__PURE__ */ new Map();
1286
+ for (const entry of records) {
1287
+ if (!rank.has(entry.scope)) {
1288
+ continue;
1289
+ }
1290
+ const existing = byKey.get(entry.key);
1291
+ if (existing === void 0) {
1292
+ byKey.set(entry.key, [
1293
+ entry
1294
+ ]);
1295
+ } else {
1296
+ existing.push(entry);
1297
+ }
1298
+ }
1299
+ const resolved = [];
1300
+ for (const candidates of byKey.values()) {
1301
+ const ordered = [
1302
+ ...candidates
1303
+ ].sort((a, b) => (rank.get(a.scope) ?? 0) - (rank.get(b.scope) ?? 0));
1304
+ const winner = ordered[0];
1305
+ if (winner === void 0) {
1306
+ continue;
1307
+ }
1308
+ const overrides = ordered.slice(1).map((beaten) => ({
1309
+ scope: beaten.scope,
1310
+ text: beaten.text,
1311
+ author: beaten.origin.author
1312
+ }));
1313
+ const pinned = ordered.some((candidate) => candidate.pinned === true);
1314
+ resolved.push({
1315
+ ...winner,
1316
+ ...pinned ? {
1317
+ pinned: true
1318
+ } : {},
1319
+ ...overrides.length > 0 ? {
1320
+ overrides
1321
+ } : {}
1322
+ });
1323
+ }
1324
+ const byPrecedence = /* @__PURE__ */ __name((a, b) => (rank.get(a.scope) ?? 0) - (rank.get(b.scope) ?? 0) || compare2(b.updatedAt, a.updatedAt) || compare2(a.key, b.key), "byPrecedence");
1325
+ const selected = resolved.map((entry, index) => ({
1326
+ entry,
1327
+ index
1328
+ })).sort((a, b) => pinRank(a.entry) - pinRank(b.entry) || (ranked && a.entry.pinned !== true ? a.index - b.index : byPrecedence(a.entry, b.entry))).slice(0, maxMemories).map((candidate) => candidate.entry);
1329
+ const pinnedOmitted = resolved.filter((entry) => entry.pinned === true).length - selected.filter((entry) => entry.pinned === true).length;
1330
+ selected.sort((a, b) => pinRank(a) - pinRank(b) || (rank.get(a.scope) ?? 0) - (rank.get(b.scope) ?? 0) || compare2(a.key, b.key));
1331
+ return {
1332
+ entries: selected,
1333
+ omitted: Math.max(0, resolved.length - maxMemories),
1334
+ pinnedOmitted
1335
+ };
1336
+ }
1337
+ __name(resolveMemoryDigest, "resolveMemoryDigest");
1338
+ function pinRank(entry) {
1339
+ return entry.pinned === true ? 0 : 1;
1340
+ }
1341
+ __name(pinRank, "pinRank");
1342
+ function compare2(a, b) {
1343
+ return a < b ? -1 : a > b ? 1 : 0;
1344
+ }
1345
+ __name(compare2, "compare");
1346
+ async function offerMemories({ config, ctx, query }) {
1347
+ const scopes = await (config.scopes ?? defaultScopeResolver).resolve(ctx);
1348
+ const maxMemories = config.maxMemories ?? DEFAULT_MAX_MEMORIES;
1349
+ const trimmed = query?.trim() ?? "";
1350
+ const search = trimmed.length > 0 ? config.provider.search : void 0;
1351
+ const records = search !== void 0 ? await search({
1352
+ scopes,
1353
+ query: trimmed,
1354
+ limit: maxMemories,
1355
+ ctx
1356
+ }) : await config.provider.list({
1357
+ scopes,
1358
+ ctx
1359
+ });
1360
+ return {
1361
+ scopes,
1362
+ recalled: search !== void 0,
1363
+ ...resolveMemoryDigest({
1364
+ records,
1365
+ scopes,
1366
+ maxMemories,
1367
+ ranked: search !== void 0
1368
+ })
1369
+ };
1370
+ }
1371
+ __name(offerMemories, "offerMemories");
1372
+ var ASSERTED_FRAMING = "Stated by people. A person wrote each of these deliberately \u2014 the user about themselves, or someone administering their organisation. Treat them as you would any other instruction you were given. If the user says something that contradicts one, say which note it conflicts with and at what scope it is held, rather than quietly setting it aside.";
1373
+ var CONCLUDED_FRAMING = "Concluded by you. These are your own notes from earlier turns, not documents anyone wrote: they may be wrong or out of date, so prefer what the user says now, and say where a note came from if you act on it.";
1374
+ var PARTIAL_FRAMING = "You are being shown a selection, not everything on file. Other notes exist that are not in this list, so never read something\u2019s absence from it as evidence that it was never recorded.";
1375
+ function memorySection(framing, entries) {
1376
+ if (entries.length === 0) {
1377
+ return [];
1378
+ }
1379
+ return [
1380
+ framing,
1381
+ ...entries.flatMap((entry) => [
1382
+ `- [${entry.scope}] ${entry.key}: ${entry.text}`,
1383
+ ...(entry.overrides ?? []).map((beaten) => ` \u21B3 [${beaten.scope}] ${beaten.author === "human" ? "a person stated" : "instead has"}: ${beaten.text}`)
1384
+ ])
1385
+ ];
1386
+ }
1387
+ __name(memorySection, "memorySection");
1388
+ function buildMemoryBlock({ entries, writable, partial }) {
1389
+ const writeLine = writable ? [
1390
+ `When you learn something durable about this user, record it with the \`${REMEMBER_TOOL_NAME}\` tool; when one of these turns out to be wrong, record the corrected fact under the same key. What you record that way is one of your own notes, not something stated by a person.`
1391
+ ] : [];
1392
+ const body = [
1393
+ "What is on file about this user and their organisation, carried over from earlier conversations. Each line is a scope, a key and the fact. A narrower scope overrides a wider one; where the wider value is shown underneath, tell the user their setting differs from it rather than silently choosing one.",
1394
+ ...partial ? [
1395
+ PARTIAL_FRAMING
1396
+ ] : [],
1397
+ ...writeLine,
1398
+ // People before inferences: the model meets what it was told before what it worked out.
1399
+ ...memorySection(ASSERTED_FRAMING, entries.filter((entry) => entry.origin.author === "human")),
1400
+ ...memorySection(CONCLUDED_FRAMING, entries.filter((entry) => entry.origin.author !== "human"))
1401
+ ];
1402
+ return `<memory>
1403
+ ${body.join("\n")}
1404
+ </memory>`;
1405
+ }
1406
+ __name(buildMemoryBlock, "buildMemoryBlock");
1407
+ var REMEMBER_TOOL_NAME = "remember";
1408
+ var REMEMBER_JSON_SCHEMA = {
1409
+ type: "object",
1410
+ additionalProperties: false,
1411
+ required: [
1412
+ "key",
1413
+ "fact"
1414
+ ],
1415
+ properties: {
1416
+ key: {
1417
+ type: "string",
1418
+ description: 'A short, stable handle for what this fact is ABOUT (e.g. "fiscal-year", "preferred-units"). Reusing an existing key replaces that fact rather than adding a second one.'
1419
+ },
1420
+ fact: {
1421
+ type: "string",
1422
+ description: "The fact itself, in one sentence, stated so it still makes sense in a conversation months from now."
1423
+ }
1424
+ }
1425
+ };
1426
+ function issue3(path, message) {
1427
+ return {
1428
+ message,
1429
+ path
1430
+ };
1431
+ }
1432
+ __name(issue3, "issue");
1433
+ var rememberInputSchema = {
1434
+ "~standard": {
1435
+ version: 1,
1436
+ vendor: "nestjs-agent",
1437
+ validate: /* @__PURE__ */ __name((value) => {
1438
+ if (typeof value !== "object" || value === null) {
1439
+ return {
1440
+ issues: [
1441
+ issue3([], "must be an object")
1442
+ ]
1443
+ };
1444
+ }
1445
+ const candidate = value;
1446
+ if (typeof candidate.key !== "string" || candidate.key.length === 0) {
1447
+ return {
1448
+ issues: [
1449
+ issue3([
1450
+ "key"
1451
+ ], "must be a non-empty string")
1452
+ ]
1453
+ };
1454
+ }
1455
+ if (typeof candidate.fact !== "string" || candidate.fact.length === 0) {
1456
+ return {
1457
+ issues: [
1458
+ issue3([
1459
+ "fact"
1460
+ ], "must be a non-empty string")
1461
+ ]
1462
+ };
1463
+ }
1464
+ return {
1465
+ value: {
1466
+ key: candidate.key,
1467
+ fact: candidate.fact
1468
+ }
1469
+ };
1470
+ }, "validate"),
1471
+ jsonSchema: {
1472
+ input: /* @__PURE__ */ __name(() => REMEMBER_JSON_SCHEMA, "input")
1473
+ }
1474
+ }
1475
+ };
1476
+ var REMEMBER_TOOL_DESCRIPTION = "Record one durable fact about this user under a key, so later conversations start knowing it. For things that will still be true another day \u2014 how they work, what their organisation requires, a correction they made. Not for what this conversation is about, and not for anything you were not told or could not reasonably infer. The user can read and delete everything you record here.";
1477
+ function rememberToolDefinition() {
1478
+ return {
1479
+ name: REMEMBER_TOOL_NAME,
1480
+ kind: "memory",
1481
+ description: REMEMBER_TOOL_DESCRIPTION,
1482
+ inputSchema: rememberInputSchema
1483
+ };
1484
+ }
1485
+ __name(rememberToolDefinition, "rememberToolDefinition");
1486
+ function withMemoryTool({ tools, enabled }) {
1487
+ return enabled ? [
1488
+ ...tools,
1489
+ rememberToolDefinition()
1490
+ ] : tools;
1491
+ }
1492
+ __name(withMemoryTool, "withMemoryTool");
1493
+ function memoryWriteVerdict(request) {
1494
+ const { scope, scopes, author, actor } = request;
1495
+ if (!scopes.includes(scope)) {
1496
+ return {
1497
+ allowed: false,
1498
+ reason: `"${scope}" is not a scope this actor belongs to`
1499
+ };
1500
+ }
1501
+ const own = actorScope(actor);
1502
+ if (scope === own) {
1503
+ return {
1504
+ allowed: true
1505
+ };
1506
+ }
1507
+ if (author.kind !== "human") {
1508
+ return {
1509
+ allowed: false,
1510
+ reason: `only a human may write a memory at "${scope}"; an agent may write only at "${own}"`
1511
+ };
1512
+ }
1513
+ return request.elevated === true ? {
1514
+ allowed: true
1515
+ } : {
1516
+ allowed: false,
1517
+ reason: `writing at "${scope}" requires an elevated human author`
1518
+ };
1519
+ }
1520
+ __name(memoryWriteVerdict, "memoryWriteVerdict");
1521
+ function memoryForgetVerdict({ record, actor }) {
1522
+ return record.scope === actorScope(actor) ? {
1523
+ allowed: true
1524
+ } : {
1525
+ allowed: false,
1526
+ reason: `"${record.scope}" is not this actor's own scope; deleting it is an administrative action`
1527
+ };
1528
+ }
1529
+ __name(memoryForgetVerdict, "memoryForgetVerdict");
1530
+ async function writeMemory({ config, digest, call, ctx, runId }) {
1531
+ const write = config.provider.write;
1532
+ if (write === void 0) {
1533
+ return {
1534
+ ok: false,
1535
+ error: "Memory is read-only in this deployment."
1536
+ };
1537
+ }
1538
+ const key = call.key.trim();
1539
+ const text = call.fact.trim();
1540
+ if (key.length === 0 || text.length === 0) {
1541
+ return {
1542
+ ok: false,
1543
+ error: "A memory needs both a key and a fact."
1544
+ };
1545
+ }
1546
+ const maxFactChars = config.maxFactChars ?? DEFAULT_MAX_FACT_CHARS;
1547
+ if (text.length > maxFactChars) {
1548
+ return {
1549
+ ok: false,
1550
+ error: `A memory must be at most ${maxFactChars} characters; that was ${text.length}. State the fact more briefly.`
1551
+ };
1552
+ }
1553
+ const scope = actorScope(ctx.actor);
1554
+ const verdict = memoryWriteVerdict({
1555
+ scope,
1556
+ actor: ctx.actor,
1557
+ scopes: digest.scopes,
1558
+ author: {
1559
+ kind: "agent",
1560
+ actorRef: ctx.actor.id
1561
+ }
1562
+ });
1563
+ if (!verdict.allowed) {
1564
+ return {
1565
+ ok: false,
1566
+ error: verdict.reason
1567
+ };
1568
+ }
1569
+ const record = await write({
1570
+ key,
1571
+ text,
1572
+ scope,
1573
+ origin: {
1574
+ author: "agent",
1575
+ threadId: ctx.threadId,
1576
+ runId,
1577
+ actorRef: ctx.actor.id
1578
+ },
1579
+ ctx
1580
+ });
1581
+ return {
1582
+ ok: true,
1583
+ record
1584
+ };
1585
+ }
1586
+ __name(writeMemory, "writeMemory");
1587
+
1588
+ // src/agent-registry.ts
1589
+ var AgentRegistry = class {
1590
+ static {
1591
+ __name(this, "AgentRegistry");
1592
+ }
1593
+ definitions = /* @__PURE__ */ new Map();
1594
+ register(definition) {
1595
+ this.definitions.set(definition.name, definition);
1596
+ }
1597
+ get(name) {
1598
+ return this.definitions.get(name);
1599
+ }
1600
+ has(name) {
1601
+ return this.definitions.has(name);
1602
+ }
1603
+ list() {
1604
+ return [
1605
+ ...this.definitions.values()
1606
+ ];
1607
+ }
1608
+ };
1609
+
1610
+ // src/replay-integrity.ts
1611
+ function isReplayIntegrityError(error) {
1612
+ return error instanceof Error && error.name.toLowerCase().endsWith("nondeterminismerror");
1613
+ }
1614
+ __name(isReplayIntegrityError, "isReplayIntegrityError");
1615
+
1616
+ // src/control-flow.ts
1617
+ var CONTROL_FLOW_SIGNAL = Symbol.for("aviary:durable:control-flow");
1618
+ function isControlFlowSignal(error) {
1619
+ return typeof error === "object" && error !== null && error[CONTROL_FLOW_SIGNAL] === true;
1620
+ }
1621
+ __name(isControlFlowSignal, "isControlFlowSignal");
1622
+
1623
+ // src/delegation.ts
1624
+ function normalizeDelegation(entry) {
1625
+ return typeof entry === "string" ? {
1626
+ agent: entry,
1627
+ detached: false
1628
+ } : {
1629
+ agent: entry.agent,
1630
+ detached: entry.detached === true
1631
+ };
1632
+ }
1633
+ __name(normalizeDelegation, "normalizeDelegation");
1634
+ function detachedStarted(args) {
1635
+ return {
1636
+ detached: true,
1637
+ status: "started",
1638
+ agent: args.agent,
1639
+ runId: args.runId,
1640
+ note: `The "${args.agent}" agent is now working on this in the background. Its answer is NOT part of this turn and will arrive in this conversation as a separate message when it is ready. Tell the user the work has started; do not state or guess what it will find.`
1641
+ };
1642
+ }
1643
+ __name(detachedStarted, "detachedStarted");
1644
+ function detachedDelivered(args) {
1645
+ return {
1646
+ detached: true,
1647
+ status: "delivered",
1648
+ agent: args.agent,
1649
+ runId: args.runId,
1650
+ text: args.text
1651
+ };
1652
+ }
1653
+ __name(detachedDelivered, "detachedDelivered");
1654
+ function detachedUnsettled(args) {
1655
+ return {
1656
+ detached: true,
1657
+ status: args.status,
1658
+ agent: args.agent,
1659
+ runId: args.runId,
1660
+ ...args.error !== void 0 ? {
1661
+ error: args.error
1662
+ } : {}
1663
+ };
1664
+ }
1665
+ __name(detachedUnsettled, "detachedUnsettled");
1666
+ async function settleUnsettledDelegation(args) {
1667
+ const { store, delivery, agent, runId, status } = args;
1668
+ if (await store.getThread(delivery.threadId) === null) {
1669
+ return;
1670
+ }
1671
+ await store.appendMessage({
1672
+ threadId: delivery.threadId,
1673
+ role: "assistant",
1674
+ content: status === "cancelled" ? `The "${agent}" agent was stopped before it could answer.` : `The "${agent}" agent stopped before it could answer: ${args.error ?? "unknown error"}`,
1675
+ agentName: agent,
1676
+ runId
1677
+ });
1678
+ await store.updateToolCall({
1679
+ toolCallId: delivery.toolCallId,
1680
+ status: "executed",
1681
+ output: detachedUnsettled({
1682
+ agent,
1683
+ runId,
1684
+ status,
1685
+ ...args.error !== void 0 ? {
1686
+ error: args.error
1687
+ } : {}
1688
+ })
1689
+ });
1690
+ }
1691
+ __name(settleUnsettledDelegation, "settleUnsettledDelegation");
1692
+
1693
+ // src/tool-registry.ts
1694
+ var ToolForbiddenError = class extends Error {
1695
+ static {
1696
+ __name(this, "ToolForbiddenError");
1697
+ }
1698
+ toolName;
1699
+ constructor(toolName) {
1700
+ super(`Tool "${toolName}" is not allowed for this role`), this.toolName = toolName;
1701
+ this.name = "ToolForbiddenError";
1702
+ }
1703
+ };
1704
+ var ToolDisabledError = class extends Error {
1705
+ static {
1706
+ __name(this, "ToolDisabledError");
1707
+ }
1708
+ toolName;
1709
+ constructor(toolName) {
1710
+ super(`Tool "${toolName}" is disabled in this deployment`), this.toolName = toolName;
1711
+ this.name = "ToolDisabledError";
1712
+ }
1713
+ };
1714
+ var ToolNotFoundError = class extends Error {
1715
+ static {
1716
+ __name(this, "ToolNotFoundError");
1717
+ }
1718
+ toolName;
1719
+ constructor(toolName) {
1720
+ super(`Tool "${toolName}" is not registered`), this.toolName = toolName;
1721
+ this.name = "ToolNotFoundError";
1722
+ }
1723
+ };
1724
+ var ToolInputInvalidError = class extends Error {
1725
+ static {
1726
+ __name(this, "ToolInputInvalidError");
1727
+ }
1728
+ toolName;
1729
+ issues;
1730
+ constructor(toolName, issues) {
1731
+ super(`Invalid input for tool "${toolName}": ${issues.map((issue4) => issue4.message).join("; ")}`), this.toolName = toolName, this.issues = issues;
1732
+ this.name = "ToolInputInvalidError";
1733
+ }
1734
+ };
1735
+ var ToolRegistry = class {
1736
+ static {
1737
+ __name(this, "ToolRegistry");
1738
+ }
1739
+ entries = /* @__PURE__ */ new Map();
1740
+ register(spec, handler) {
1741
+ this.entries.set(spec.name, {
1742
+ spec,
1743
+ handler
1744
+ });
1745
+ }
1746
+ has(name) {
1747
+ return this.entries.has(name);
1748
+ }
1749
+ spec(name) {
1750
+ return this.entries.get(name)?.spec;
1751
+ }
1752
+ allSpecs() {
1753
+ return [
1754
+ ...this.entries.values()
1755
+ ].map((entry) => entry.spec);
1756
+ }
1757
+ /**
1758
+ * The tools to offer the model for this actor+agent, after the four filter layers: what this
1759
+ * agent pinned, what this deployment has enabled, what this actor's role allows, and what each
1760
+ * tool's own `canUse` allows this actor.
1761
+ *
1762
+ * Every layer only ever removes tools, so no arrangement of them can widen what a turn reaches.
1763
+ * The agent's allow-list therefore goes FIRST, even though it is the narrowest statement: it is a
1764
+ * pure name-set match, while each of the three below it may be a round trip — an authz service, a
1765
+ * feature-flag store, an MCP server — and this runs once per model step. Asking those about a
1766
+ * tool the allow-list has already excluded is a call whose answer nothing reads.
1767
+ */
1768
+ async definitionsFor(actor, policy, allowedTools) {
1769
+ const pinnedNames = new Set(filterToolsByAllowList(this.allSpecs(), allowedTools).map((spec) => spec.name));
1770
+ const pinned = [
1771
+ ...this.entries.values()
1772
+ ].filter((entry) => pinnedNames.has(entry.spec.name));
1773
+ const live = await filterToolsByEnabled(pinned);
1774
+ const allowedByRole = new Set((await filterToolsByRole(live.map((entry) => entry.spec), actor, policy)).map((spec) => spec.name));
1775
+ const roleScoped = live.filter((entry) => allowedByRole.has(entry.spec.name));
1776
+ const actorScoped = await filterToolsByCanUse(roleScoped, actor);
1777
+ return actorScoped.map(({ spec }) => ({
1778
+ name: spec.name,
1779
+ kind: spec.kind,
1780
+ description: spec.description,
1781
+ inputSchema: spec.inputSchema
1782
+ }));
1783
+ }
1784
+ /**
1785
+ * Run a tool. Re-checks that the tool is enabled and that the role allows it (defense-in-depth —
1786
+ * a call can reach here from a replayed durable step or an approval granted before the flag
1787
+ * moved, neither of which went through `definitionsFor` again) and re-parses the input via Zod.
1788
+ */
1789
+ async invoke(name, input, ctx, policy) {
1790
+ const entry = this.entries.get(name);
1791
+ if (entry === void 0) {
1792
+ throw new ToolNotFoundError(name);
1793
+ }
1794
+ if (!await isToolEnabled(entry.spec, entry.handler)) {
1795
+ throw new ToolDisabledError(name);
1796
+ }
1797
+ if (!await policy.can(ctx.actor, entry.spec)) {
1798
+ throw new ToolForbiddenError(name);
1799
+ }
1800
+ if (!await canActorUseTool(ctx.actor, entry.handler)) {
1801
+ throw new ToolForbiddenError(name);
1802
+ }
1803
+ const validation = await entry.spec.inputSchema["~standard"].validate(input);
1804
+ if (validation.issues !== void 0) {
1805
+ throw new ToolInputInvalidError(name, validation.issues);
1806
+ }
1807
+ return entry.handler.execute(validation.value, ctx);
1808
+ }
1809
+ };
1810
+ var DefaultRolesPolicy = class {
1811
+ static {
1812
+ __name(this, "DefaultRolesPolicy");
1813
+ }
1814
+ defaultRoles;
1815
+ constructor(defaultRoles = [
1816
+ "ADMIN"
1817
+ ]) {
1818
+ this.defaultRoles = defaultRoles;
1819
+ }
1820
+ can(actor, tool) {
1821
+ const allowed = tool.roles ?? this.defaultRoles;
1822
+ return (actor.roles ?? []).some((role) => allowed.includes(role));
1823
+ }
1824
+ };
1825
+
1826
+ // src/agent-loop.ts
1827
+ import { createHash } from "node:crypto";
1828
+ import { trace } from "@dudousxd/nestjs-diagnostics";
1829
+
1830
+ // src/diagnostics.ts
1831
+ import { emit } from "@dudousxd/nestjs-diagnostics";
1832
+ function publishAgentRunStarted(payload) {
1833
+ emit("agent", "run.started", payload);
1834
+ }
1835
+ __name(publishAgentRunStarted, "publishAgentRunStarted");
1836
+ function publishAgentMessage(payload) {
1837
+ emit("agent", "message", payload);
1838
+ }
1839
+ __name(publishAgentMessage, "publishAgentMessage");
1840
+ function publishAgentToolCall(payload) {
1841
+ emit("agent", "tool-call", payload);
1842
+ }
1843
+ __name(publishAgentToolCall, "publishAgentToolCall");
1844
+ function publishAgentQuotaExceeded(payload) {
1845
+ emit("agent", "quota.exceeded", payload);
1846
+ }
1847
+ __name(publishAgentQuotaExceeded, "publishAgentQuotaExceeded");
1848
+ function publishAgentRunFinished(payload) {
1849
+ emit("agent", "run.finished", payload);
1850
+ }
1851
+ __name(publishAgentRunFinished, "publishAgentRunFinished");
1852
+ function publishAgentRunFailed(payload) {
1853
+ emit("agent", "run.failed", payload);
1854
+ }
1855
+ __name(publishAgentRunFailed, "publishAgentRunFailed");
1856
+ function publishAgentDelegated(payload) {
1857
+ emit("agent", "delegated", payload);
1858
+ }
1859
+ __name(publishAgentDelegated, "publishAgentDelegated");
1860
+ function publishAgentRetrieved(payload) {
1861
+ emit("agent", "retrieved", payload);
1862
+ }
1863
+ __name(publishAgentRetrieved, "publishAgentRetrieved");
1864
+ function publishAgentToolRetry(payload) {
1865
+ emit("agent", "tool.retry", payload);
1866
+ }
1867
+ __name(publishAgentToolRetry, "publishAgentToolRetry");
1868
+ function publishAgentSkillsResolved(payload) {
1869
+ emit("agent", "skills.resolved", payload);
1870
+ }
1871
+ __name(publishAgentSkillsResolved, "publishAgentSkillsResolved");
1872
+ function publishAgentMemoryResolved(payload) {
1873
+ emit("agent", "memory.resolved", payload);
1874
+ }
1875
+ __name(publishAgentMemoryResolved, "publishAgentMemoryResolved");
1876
+ function publishAgentMemoryWritten(payload) {
1877
+ emit("agent", "memory.written", payload);
1878
+ }
1879
+ __name(publishAgentMemoryWritten, "publishAgentMemoryWritten");
1880
+ var AGENT_SPAN_EVENTS = [
1881
+ "llm.turn",
1882
+ "tool.execution",
1883
+ "retrieval",
1884
+ "follow-ups"
1885
+ ];
1886
+ var AGENT_DIAGNOSTIC_EVENTS = [
1887
+ "run.started",
1888
+ "message",
1889
+ "tool-call",
1890
+ "quota.exceeded",
1891
+ "run.finished",
1892
+ "run.failed",
1893
+ "delegated",
1894
+ "retrieved",
1895
+ "skills.resolved",
1896
+ "memory.resolved",
1897
+ "memory.written",
1898
+ "tool.retry"
1899
+ ];
1900
+ function agentDiagnosticKey(event) {
1901
+ return `agent:${event}`;
1902
+ }
1903
+ __name(agentDiagnosticKey, "agentDiagnosticKey");
1904
+
1905
+ // src/tool-retry.ts
1906
+ function hasTransientShape(error) {
1907
+ if (typeof error !== "object" || error === null) {
1908
+ return false;
1909
+ }
1910
+ const code = "code" in error ? error.code : void 0;
1911
+ const errno = "errno" in error ? error.errno : void 0;
1912
+ const sqlState = "sqlState" in error ? error.sqlState : void 0;
1913
+ if (code === 1213 || code === 1205 || errno === 1213 || errno === 1205) {
1914
+ return true;
1915
+ }
1916
+ if (code === "ER_LOCK_DEADLOCK" || code === "ER_LOCK_WAIT_TIMEOUT") {
1917
+ return true;
1918
+ }
1919
+ if (code === "40001" || code === "40P01" || sqlState === "40001" || sqlState === "40P01") {
1920
+ return true;
1921
+ }
1922
+ if (code === "SQLITE_BUSY") {
1923
+ return true;
1924
+ }
1925
+ const message = "message" in error ? error.message : void 0;
1926
+ return typeof message === "string" && /deadlock|lock wait timeout|serialization failure/i.test(message);
1927
+ }
1928
+ __name(hasTransientShape, "hasTransientShape");
1929
+ function isTransientToolError(error) {
1930
+ if (hasTransientShape(error)) {
1931
+ return true;
1932
+ }
1933
+ if (typeof error === "object" && error !== null && "cause" in error) {
1934
+ const cause = error.cause;
1935
+ if (cause !== void 0 && cause !== error && hasTransientShape(cause)) {
1936
+ return true;
1937
+ }
1938
+ }
1939
+ return false;
1940
+ }
1941
+ __name(isTransientToolError, "isTransientToolError");
1942
+ var DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS = 2;
1943
+ var DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS = 150;
1944
+ function resolveToolTransientRetryNumbers(setting) {
1945
+ if (setting === false) {
1946
+ return false;
1947
+ }
1948
+ return {
1949
+ attempts: setting?.attempts ?? DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS,
1950
+ backoffMs: setting?.backoffMs ?? DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS
1951
+ };
1952
+ }
1953
+ __name(resolveToolTransientRetryNumbers, "resolveToolTransientRetryNumbers");
1954
+ function delay(ms) {
1955
+ return new Promise((resolve) => {
1956
+ setTimeout(resolve, ms);
1957
+ });
1958
+ }
1959
+ __name(delay, "delay");
1960
+ async function invokeWithTransientRetry(fn, setting, options) {
1961
+ if (setting === false) {
1962
+ return fn();
1963
+ }
1964
+ const attempts = setting.attempts ?? DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS;
1965
+ const backoffMs = setting.backoffMs ?? DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS;
1966
+ const classify = setting.classify ?? isTransientToolError;
1967
+ let attempt = 1;
1968
+ for (; ; ) {
1969
+ try {
1970
+ return await fn();
1971
+ } catch (error) {
1972
+ if (isControlFlowSignal(error) || options?.isControlFlowError?.(error) === true) {
1973
+ throw error;
1974
+ }
1975
+ const attemptsRemain = attempt < attempts;
1976
+ if (!attemptsRemain || !classify(error)) {
1977
+ throw error;
1978
+ }
1979
+ options?.onRetry?.(attempt, error);
1980
+ await delay(backoffMs * attempt);
1981
+ attempt += 1;
1982
+ }
1983
+ }
1984
+ }
1985
+ __name(invokeWithTransientRetry, "invokeWithTransientRetry");
1986
+
1987
+ // src/agent-loop.ts
1988
+ function resolveCostUsd(usage, reportedCostUsd, price) {
1989
+ if (reportedCostUsd !== void 0) {
1990
+ return reportedCostUsd;
1991
+ }
1992
+ return price === void 0 ? null : estimateCost(usage, price);
1993
+ }
1994
+ __name(resolveCostUsd, "resolveCostUsd");
1995
+ function buildContextBlock(passages) {
1996
+ const items = passages.map((passage, index) => {
1997
+ const label = passage.source !== void 0 ? ` (${passage.source})` : "";
1998
+ return `[${index + 1}]${label} ${passage.text}`;
1999
+ }).join("\n\n");
2000
+ return `<retrieved_context>
2001
+ ${items}
2002
+ </retrieved_context>
2003
+ Use the retrieved context above to answer when relevant, and cite sources by their bracket number.`;
2004
+ }
2005
+ __name(buildContextBlock, "buildContextBlock");
2006
+ function buildSummaryBlock(summary) {
2007
+ return `<conversation_summary>
2008
+ ${summary}
2009
+ </conversation_summary>
2010
+ Earlier messages in this thread are no longer included verbatim. Treat the summary above as an accurate record of them.`;
2011
+ }
2012
+ __name(buildSummaryBlock, "buildSummaryBlock");
2013
+ function settleAll(tasks) {
2014
+ const started = tasks.map((task) => task());
2015
+ return Promise.all(started.map((work) => work.then((value) => ({
2016
+ ok: true,
2017
+ value
2018
+ }), (error) => ({
2019
+ ok: false,
2020
+ error
2021
+ }))));
2022
+ }
2023
+ __name(settleAll, "settleAll");
2024
+ var QuotaExceededError = class extends Error {
2025
+ static {
2026
+ __name(this, "QuotaExceededError");
2027
+ }
2028
+ constructor() {
2029
+ super("Daily token quota exceeded");
2030
+ this.name = "QuotaExceededError";
2031
+ }
2032
+ };
2033
+ var RunCancelledError = class extends Error {
2034
+ static {
2035
+ __name(this, "RunCancelledError");
2036
+ }
2037
+ constructor() {
2038
+ super("Run cancelled");
2039
+ this.name = "RunCancelledError";
2040
+ }
2041
+ };
2042
+ function agentFailureCode(error) {
2043
+ if (error instanceof RunCancelledError) {
2044
+ return "cancelled";
2045
+ }
2046
+ if (error instanceof QuotaExceededError) {
2047
+ return "quota_exceeded";
2048
+ }
2049
+ if (error instanceof OutputRejectedError) {
2050
+ return "output_rejected";
2051
+ }
2052
+ if (error instanceof StructuredOutputError) {
2053
+ return "structured_output_invalid";
2054
+ }
2055
+ return "run_failed";
2056
+ }
2057
+ __name(agentFailureCode, "agentFailureCode");
2058
+ var MAX_DELEGATION_DEPTH = 5;
2059
+ var DEFAULT_MAX_AGENT_APPEARANCES = 1;
2060
+ function delegationRefusal(args) {
2061
+ const { deps, input, targetAgent } = args;
2062
+ const ancestry = input.delegationPath ?? [];
2063
+ const appearances = ancestry.filter((name) => name === targetAgent).length;
2064
+ const maxAppearances = deps.maxAgentAppearances ?? DEFAULT_MAX_AGENT_APPEARANCES;
2065
+ if (appearances >= maxAppearances) {
2066
+ const chain = [
2067
+ ...ancestry,
2068
+ targetAgent
2069
+ ].join(" \u2192 ");
2070
+ const times = appearances + 1;
2071
+ return `(delegation cycle: ${chain} \u2014 ${targetAgent} ${times} times on one chain)`;
2072
+ }
2073
+ const maxDepth = deps.maxDelegationDepth ?? MAX_DELEGATION_DEPTH;
2074
+ if ((input.delegationDepth ?? 0) >= maxDepth) {
2075
+ return `(delegation depth limit of ${maxDepth} reached)`;
2076
+ }
2077
+ return null;
2078
+ }
2079
+ __name(delegationRefusal, "delegationRefusal");
2080
+ async function resolvePrompt(prompt, ctx) {
2081
+ return typeof prompt === "function" ? prompt(ctx) : prompt;
2082
+ }
2083
+ __name(resolvePrompt, "resolvePrompt");
2084
+ async function resolveSystemPrompt(deps, input) {
2085
+ const ctx = {
2086
+ actor: input.actor,
2087
+ agentName: input.agentName ?? "default",
2088
+ ...input.pageContext !== void 0 ? {
2089
+ pageContext: input.pageContext
2090
+ } : {}
2091
+ };
2092
+ const sections = [
2093
+ await resolvePrompt(deps.systemPrompt, ctx)
2094
+ ];
2095
+ for (const contribute of deps.promptContributors ?? []) {
2096
+ const section = await contribute(ctx);
2097
+ if (section !== null && section.length > 0) {
2098
+ sections.push(section);
2099
+ }
2100
+ }
2101
+ return sections.join("\n\n");
2102
+ }
2103
+ __name(resolveSystemPrompt, "resolveSystemPrompt");
2104
+ function extractTask(input) {
2105
+ if (typeof input === "object" && input !== null && "task" in input) {
2106
+ const task = input.task;
2107
+ if (typeof task === "string") {
2108
+ return task;
2109
+ }
2110
+ }
2111
+ return JSON.stringify(input);
2112
+ }
2113
+ __name(extractTask, "extractTask");
2114
+ function deriveTitle(userText) {
2115
+ const trimmed = userText.trim().replace(/\s+/g, " ");
2116
+ return trimmed.length > 60 ? `${trimmed.slice(0, 57)}...` : trimmed || "New chat";
2117
+ }
2118
+ __name(deriveTitle, "deriveTitle");
2119
+ var ToolTimeoutError = class ToolTimeoutError2 extends Error {
2120
+ static {
2121
+ __name(this, "ToolTimeoutError");
2122
+ }
2123
+ constructor(toolName, ms) {
2124
+ super(`Tool "${toolName}" exceeded its ${ms}ms timeout`);
2125
+ this.name = "ToolTimeoutError";
2126
+ }
2127
+ };
2128
+ function withToolTimeout(work, ms, toolName) {
2129
+ return new Promise((resolve, reject) => {
2130
+ const timer = setTimeout(() => reject(new ToolTimeoutError(toolName, ms)), ms);
2131
+ work.then((value) => {
2132
+ clearTimeout(timer);
2133
+ resolve(value);
2134
+ }, (error) => {
2135
+ clearTimeout(timer);
2136
+ reject(error);
2137
+ });
2138
+ });
2139
+ }
2140
+ __name(withToolTimeout, "withToolTimeout");
2141
+ function parseFollowUps(text, count) {
2142
+ const source = text.match(/\[[\s\S]*\]/)?.[0] ?? text;
2143
+ try {
2144
+ const parsed = JSON.parse(source);
2145
+ if (Array.isArray(parsed)) {
2146
+ return parsed.filter((item) => typeof item === "string").slice(0, count);
2147
+ }
2148
+ } catch {
2149
+ }
2150
+ return [];
2151
+ }
2152
+ __name(parseFollowUps, "parseFollowUps");
2153
+ async function generateFollowUps(model, messages, count) {
2154
+ const discard = {
2155
+ write: /* @__PURE__ */ __name(() => {
2156
+ }, "write"),
2157
+ end: /* @__PURE__ */ __name(() => {
2158
+ }, "end"),
2159
+ fail: /* @__PURE__ */ __name(() => {
2160
+ }, "fail")
2161
+ };
2162
+ const turn = await model.runTurn({
2163
+ system: `Based on the conversation so far, propose up to ${count} short, distinct follow-up questions the user is likely to ask next. Respond with ONLY a JSON array of strings \u2014 no prose, no code fences.`,
2164
+ messages,
2165
+ tools: [],
623
2166
  sink: discard
624
2167
  });
625
2168
  return {
626
- followUps: parseFollowUps(turn.text, count),
627
- usage: turn.usage,
628
- ...turn.modelId !== void 0 ? {
629
- modelId: turn.modelId
630
- } : {}
2169
+ followUps: parseFollowUps(turn.text, count),
2170
+ usage: turn.usage,
2171
+ ...turn.modelId !== void 0 ? {
2172
+ modelId: turn.modelId
2173
+ } : {}
2174
+ };
2175
+ }
2176
+ __name(generateFollowUps, "generateFollowUps");
2177
+ async function spanned(event, runId, payload, run, summarize) {
2178
+ let value;
2179
+ await trace("agent", event, async () => {
2180
+ value = await run();
2181
+ return summarize(value);
2182
+ }, payload, {
2183
+ traceId: runId
2184
+ });
2185
+ return value;
2186
+ }
2187
+ __name(spanned, "spanned");
2188
+ function traceLlmTurn(runId, step, run) {
2189
+ return spanned("llm.turn", runId, {
2190
+ runId,
2191
+ step
2192
+ }, run, (turn) => ({
2193
+ ...turn.modelId !== void 0 ? {
2194
+ modelId: turn.modelId
2195
+ } : {},
2196
+ inputTokens: turn.usage.inputTokens,
2197
+ outputTokens: turn.usage.outputTokens,
2198
+ textLength: turn.text.length,
2199
+ toolCalls: turn.toolCalls.length
2200
+ }));
2201
+ }
2202
+ __name(traceLlmTurn, "traceLlmTurn");
2203
+ function traceToolExecution(runId, call, run) {
2204
+ return spanned("tool.execution", runId, {
2205
+ runId,
2206
+ ...call
2207
+ }, run, () => ({}));
2208
+ }
2209
+ __name(traceToolExecution, "traceToolExecution");
2210
+ function toModelMessages(messages) {
2211
+ return messages.map((message) => ({
2212
+ role: message.role,
2213
+ content: message.content,
2214
+ ...message.toolCalls !== void 0 ? {
2215
+ toolCalls: message.toolCalls
2216
+ } : {},
2217
+ ...message.toolResults !== void 0 ? {
2218
+ toolResults: message.toolResults
2219
+ } : {},
2220
+ ...message.attachments !== void 0 ? {
2221
+ attachments: message.attachments
2222
+ } : {}
2223
+ }));
2224
+ }
2225
+ __name(toModelMessages, "toModelMessages");
2226
+ function historyContext(input) {
2227
+ return {
2228
+ threadId: input.threadId,
2229
+ actor: input.actor,
2230
+ ...input.agentName !== void 0 ? {
2231
+ agentName: input.agentName
2232
+ } : {}
2233
+ };
2234
+ }
2235
+ __name(historyContext, "historyContext");
2236
+ function splitHistory(policy, input, messages) {
2237
+ return policy === void 0 ? {
2238
+ keep: messages,
2239
+ drop: []
2240
+ } : policy.select(messages, historyContext(input));
2241
+ }
2242
+ __name(splitHistory, "splitHistory");
2243
+ async function loadSelectedHistory(deps, input, hooks) {
2244
+ return hooks.step("load:thread", async () => {
2245
+ const thread = await readThreadForTurn(deps, input.threadId);
2246
+ const { keep, drop } = splitHistory(deps.historyPolicy, input, toModelMessages(thread.messages));
2247
+ return {
2248
+ messages: keep,
2249
+ dropped: deps.historyPolicy?.summarize === void 0 ? [] : drop,
2250
+ title: thread.title,
2251
+ hasAssistantMessage: thread.hasAssistantMessage
2252
+ };
2253
+ });
2254
+ }
2255
+ __name(loadSelectedHistory, "loadSelectedHistory");
2256
+ async function readThreadForTurn(deps, threadId) {
2257
+ const windowing = deps.store;
2258
+ if (typeof windowing.loadThreadForTurn === "function") {
2259
+ const messageLimit = turnMessageLimit(deps.historyPolicy);
2260
+ const page = await windowing.loadThreadForTurn({
2261
+ threadId,
2262
+ ...messageLimit !== void 0 ? {
2263
+ messageLimit
2264
+ } : {}
2265
+ });
2266
+ return page === null ? {
2267
+ messages: [],
2268
+ title: null,
2269
+ hasAssistantMessage: false
2270
+ } : {
2271
+ messages: page.messages,
2272
+ title: page.title,
2273
+ hasAssistantMessage: page.hasAssistantMessage
2274
+ };
2275
+ }
2276
+ const thread = await deps.store.getThread(threadId);
2277
+ const stored = thread?.messages ?? [];
2278
+ return {
2279
+ messages: stored,
2280
+ title: thread?.title ?? null,
2281
+ hasAssistantMessage: stored.some((message) => message.role === "assistant")
2282
+ };
2283
+ }
2284
+ __name(readThreadForTurn, "readThreadForTurn");
2285
+ function turnMessageLimit(policy) {
2286
+ if (policy === void 0 || policy.summarize !== void 0) {
2287
+ return void 0;
2288
+ }
2289
+ return policy.maxMessages;
2290
+ }
2291
+ __name(turnMessageLimit, "turnMessageLimit");
2292
+ async function loadWholeThread(deps, input, hooks) {
2293
+ const thread = await hooks.step("load:thread", () => deps.store.getThread(input.threadId));
2294
+ const stored = thread?.messages ?? [];
2295
+ const { keep, drop } = splitHistory(deps.historyPolicy, input, toModelMessages(stored));
2296
+ return {
2297
+ messages: keep,
2298
+ dropped: drop,
2299
+ title: thread?.title ?? null,
2300
+ hasAssistantMessage: stored.some((message) => message.role === "assistant")
2301
+ };
2302
+ }
2303
+ __name(loadWholeThread, "loadWholeThread");
2304
+ async function foldDroppedHistory(summarize, deps, input, hooks, history) {
2305
+ const summary = await hooks.step("history:summarize", () => summarize(history.dropped, historyContext(input)));
2306
+ if (summary.usage !== void 0) {
2307
+ const usage = summary.usage;
2308
+ await hooks.step("persist:usage:history", () => deps.store.recordUsage({
2309
+ threadId: input.threadId,
2310
+ actorRef: input.actor.id,
2311
+ modelId: summary.modelId ?? deps.modelId ?? "unknown",
2312
+ purpose: "history_summary",
2313
+ usage
2314
+ }));
2315
+ }
2316
+ return [
2317
+ {
2318
+ role: "system",
2319
+ content: buildSummaryBlock(summary.text)
2320
+ },
2321
+ ...history.messages
2322
+ ];
2323
+ }
2324
+ __name(foldDroppedHistory, "foldDroppedHistory");
2325
+ function processorContext(input, step) {
2326
+ return {
2327
+ threadId: input.threadId,
2328
+ actor: input.actor,
2329
+ step,
2330
+ ...input.agentName !== void 0 ? {
2331
+ agentName: input.agentName
2332
+ } : {}
2333
+ };
2334
+ }
2335
+ __name(processorContext, "processorContext");
2336
+ function restatementPrompt(messages, answer, fromTranscript) {
2337
+ const restated = {
2338
+ role: "assistant",
2339
+ content: answer
2340
+ };
2341
+ if (fromTranscript) {
2342
+ return [
2343
+ ...messages,
2344
+ restated
2345
+ ];
2346
+ }
2347
+ for (let index = messages.length - 1; index >= 0; index -= 1) {
2348
+ const message = messages[index];
2349
+ if (message?.role === "user") {
2350
+ return [
2351
+ message,
2352
+ restated
2353
+ ];
2354
+ }
2355
+ }
2356
+ return [
2357
+ restated
2358
+ ];
2359
+ }
2360
+ __name(restatementPrompt, "restatementPrompt");
2361
+ async function structureAnswer(schema, deps, input, hooks, messages, step) {
2362
+ const instruction = deps.outputInstruction ?? DEFAULT_STRUCTURED_OUTPUT_INSTRUCTION;
2363
+ const maxAttempts = 1 + (deps.outputRepairAttempts ?? 1);
2364
+ const outputProcessors = deps.outputProcessors ?? [];
2365
+ let issues = [];
2366
+ let text = "";
2367
+ for (let attempt = 0; attempt < maxAttempts; attempt += 1) {
2368
+ const system = attempt === 0 ? instruction : repairInstruction(instruction, issues);
2369
+ const reply = await hooks.step(`structured:${step}:${attempt}`, () => spanned("structured-output", hooks.runId, {
2370
+ runId: hooks.runId,
2371
+ step,
2372
+ attempt
2373
+ }, async () => {
2374
+ const discard = {
2375
+ write: /* @__PURE__ */ __name(() => {
2376
+ }, "write"),
2377
+ end: /* @__PURE__ */ __name(() => {
2378
+ }, "end"),
2379
+ fail: /* @__PURE__ */ __name(() => {
2380
+ }, "fail")
2381
+ };
2382
+ const turn = await deps.model.runTurn({
2383
+ system,
2384
+ messages,
2385
+ tools: [],
2386
+ sink: discard,
2387
+ outputSchema: schema
2388
+ });
2389
+ return {
2390
+ text: turn.text,
2391
+ usage: turn.usage,
2392
+ ...turn.object !== void 0 ? {
2393
+ object: turn.object
2394
+ } : {},
2395
+ ...turn.modelId !== void 0 ? {
2396
+ modelId: turn.modelId
2397
+ } : {}
2398
+ };
2399
+ }, (result) => ({
2400
+ inputTokens: result.usage.inputTokens,
2401
+ outputTokens: result.usage.outputTokens,
2402
+ textLength: result.text.length
2403
+ })));
2404
+ await hooks.step(`persist:usage:structured:${step}:${attempt}`, () => deps.store.recordUsage({
2405
+ threadId: input.threadId,
2406
+ actorRef: input.actor.id,
2407
+ modelId: reply.modelId ?? deps.modelId ?? "unknown",
2408
+ purpose: "structured_output",
2409
+ usage: reply.usage
2410
+ }));
2411
+ text = reply.text;
2412
+ if (outputProcessors.length > 0) {
2413
+ const gate = await hooks.step(`process:output:structured:${step}:${attempt}`, () => runOutputProcessors(outputProcessors, {
2414
+ text: reply.text,
2415
+ toolCalls: []
2416
+ }, processorContext(input, step)));
2417
+ if (gate.rejection !== void 0) {
2418
+ throw new OutputRejectedError(gate.rejection.processor, gate.rejection.reason);
2419
+ }
2420
+ text = gate.text;
2421
+ }
2422
+ const reported = text === reply.text ? reply.object : void 0;
2423
+ const outcome = await validateStructured(schema, text, reported);
2424
+ if (outcome.ok) {
2425
+ return outcome.value;
2426
+ }
2427
+ issues = outcome.issues;
2428
+ }
2429
+ throw new StructuredOutputError(issues, text, maxAttempts);
2430
+ }
2431
+ __name(structureAnswer, "structureAnswer");
2432
+ function declaredKind(deps, name) {
2433
+ if (deps.ask === true && name === ASK_TOOL_NAME) {
2434
+ return "ask";
2435
+ }
2436
+ if (deps.skills !== void 0 && name === SKILL_TOOL_NAME) {
2437
+ return "skill";
2438
+ }
2439
+ if (memoryIsWritable(deps) && name === REMEMBER_TOOL_NAME) {
2440
+ return "memory";
2441
+ }
2442
+ return deps.registry.spec(name)?.kind ?? "read";
2443
+ }
2444
+ __name(declaredKind, "declaredKind");
2445
+ function memoryIsWritable(deps) {
2446
+ return deps.memory?.provider.write !== void 0;
2447
+ }
2448
+ __name(memoryIsWritable, "memoryIsWritable");
2449
+ function withAskTool({ tools, ask }) {
2450
+ return ask === true ? [
2451
+ ...tools,
2452
+ askToolDefinition()
2453
+ ] : tools;
2454
+ }
2455
+ __name(withAskTool, "withAskTool");
2456
+ async function awaitElicitation(hooks, request, ctx) {
2457
+ if (hooks.awaitAnswers !== void 0) {
2458
+ return normalizeElicitationReply(await hooks.awaitAnswers(request, ctx));
2459
+ }
2460
+ return normalizeElicitationReply(await hooks.awaitApproval({
2461
+ id: request.id,
2462
+ name: ASK_TOOL_NAME,
2463
+ input: request,
2464
+ kind: "ask"
2465
+ }, ctx));
2466
+ }
2467
+ __name(awaitElicitation, "awaitElicitation");
2468
+ function toolContext(deps, input, hooks) {
2469
+ return {
2470
+ actor: input.actor,
2471
+ threadId: input.threadId,
2472
+ runId: hooks.runId,
2473
+ requestId: hooks.runId,
2474
+ ...input.agentName !== void 0 ? {
2475
+ agentName: input.agentName
2476
+ } : {},
2477
+ ...input.pageContext !== void 0 ? {
2478
+ pageContext: input.pageContext
2479
+ } : {},
2480
+ ...deps.host !== void 0 ? {
2481
+ host: deps.host
2482
+ } : {}
2483
+ };
2484
+ }
2485
+ __name(toolContext, "toolContext");
2486
+ function intakeApplies(intake, threadHasAssistant) {
2487
+ return intake.when === "every-turn" || !threadHasAssistant;
2488
+ }
2489
+ __name(intakeApplies, "intakeApplies");
2490
+ async function runIntake(intake, deps, input, hooks, writer, threadHasAssistant) {
2491
+ const preamble = intake.preamble ?? DEFAULT_INTAKE_PREAMBLE;
2492
+ const request = {
2493
+ id: `intake-${hooks.runId}`,
2494
+ source: "intake",
2495
+ preamble,
2496
+ questions: intake.questions
2497
+ };
2498
+ const call = {
2499
+ id: request.id,
2500
+ name: ASK_TOOL_NAME,
2501
+ input: {
2502
+ preamble,
2503
+ questions: intake.questions
2504
+ },
2505
+ kind: "ask"
2506
+ };
2507
+ const asked = await hooks.step("intake:ask", async () => {
2508
+ if (!intakeApplies(intake, threadHasAssistant)) {
2509
+ return null;
2510
+ }
2511
+ const message = await deps.store.appendMessage({
2512
+ threadId: input.threadId,
2513
+ role: "assistant",
2514
+ content: preamble,
2515
+ runId: hooks.runId,
2516
+ toolCalls: [
2517
+ call
2518
+ ],
2519
+ ...input.agentName !== void 0 ? {
2520
+ agentName: input.agentName
2521
+ } : {}
2522
+ });
2523
+ await deps.store.recordToolCall({
2524
+ toolCallId: request.id,
2525
+ messageId: message.id,
2526
+ toolName: ASK_TOOL_NAME,
2527
+ // The store knows read/action only. An unanswered question is work waiting on a human, which
2528
+ // is what `action` + `pending_approval` already mean — so it surfaces in an approvals inbox
2529
+ // rather than needing one of its own.
2530
+ toolType: "action",
2531
+ input: call.input,
2532
+ status: "pending_approval",
2533
+ runId: hooks.runId
2534
+ });
2535
+ await writer.write(encodeStreamEvent({
2536
+ kind: "elicitation",
2537
+ id: request.id,
2538
+ request
2539
+ }));
2540
+ return {
2541
+ messageId: message.id
2542
+ };
2543
+ });
2544
+ if (asked === null) {
2545
+ return null;
2546
+ }
2547
+ const messageId = asked.messageId;
2548
+ const reply = await awaitElicitation(hooks, request, toolContext(deps, input, hooks));
2549
+ const result = settleElicitation(request, reply);
2550
+ await hooks.step("intake:answers", async () => {
2551
+ await deps.store.updateToolCall({
2552
+ toolCallId: request.id,
2553
+ // A skip is not an answer. Both leave the agent holding the same values, but only one of them
2554
+ // is evidence the user chose them, and a reader auditing what the agent was told has to be
2555
+ // able to tell those apart.
2556
+ status: result.skipped ? "rejected" : "executed",
2557
+ output: result,
2558
+ ...result.skipped ? {
2559
+ error: "skipped by the user"
2560
+ } : {},
2561
+ ...reply.answeredByRef !== void 0 ? {
2562
+ executedByRef: reply.answeredByRef
2563
+ } : {}
2564
+ });
2565
+ await deps.store.setMessageToolResults(messageId, [
2566
+ {
2567
+ id: request.id,
2568
+ name: ASK_TOOL_NAME,
2569
+ output: result
2570
+ }
2571
+ ]);
2572
+ await writer.write(encodeStreamEvent({
2573
+ kind: "tool-output",
2574
+ id: request.id,
2575
+ output: result
2576
+ }));
2577
+ });
2578
+ return {
2579
+ role: "assistant",
2580
+ content: preamble,
2581
+ toolCalls: [
2582
+ call
2583
+ ],
2584
+ toolResults: [
2585
+ {
2586
+ id: request.id,
2587
+ name: ASK_TOOL_NAME,
2588
+ output: result
2589
+ }
2590
+ ]
2591
+ };
2592
+ }
2593
+ __name(runIntake, "runIntake");
2594
+ var PARALLEL_TOOLS_PATCH = "agent:parallel-tools";
2595
+ var CANCELLATION_PATCH = "agent:cancellation";
2596
+ var SELECTED_HISTORY_PATCH = "agent:selected-history";
2597
+ async function haltIfCancelled(hooks, cancellable, name) {
2598
+ const observe = hooks.cancelled;
2599
+ if (!cancellable || observe === void 0) {
2600
+ return;
2601
+ }
2602
+ if (await hooks.step(name, () => observe())) {
2603
+ throw new RunCancelledError();
2604
+ }
2605
+ }
2606
+ __name(haltIfCancelled, "haltIfCancelled");
2607
+ async function claimToolCall(turn, call) {
2608
+ const { deps, input, hooks, messageId } = turn;
2609
+ const persisted = await hooks.step(`persist:toolcall:${call.id}`, async () => {
2610
+ const spec = deps.registry.spec(call.name);
2611
+ const kind = declaredKind(deps, call.name);
2612
+ const awaitsHuman = kind === "action" || kind === "ask";
2613
+ await deps.store.recordToolCall({
2614
+ toolCallId: call.id,
2615
+ messageId,
2616
+ toolName: call.name,
2617
+ toolType: awaitsHuman ? "action" : "read",
2618
+ input: call.input,
2619
+ status: awaitsHuman ? "pending_approval" : "auto_executed",
2620
+ runId: hooks.runId
2621
+ });
2622
+ return {
2623
+ kind,
2624
+ ...spec?.targetAgent !== void 0 ? {
2625
+ targetAgent: spec.targetAgent
2626
+ } : {},
2627
+ // Only an `agent` spec ever carries this, and only when the author declared it — so a
2628
+ // deployment with no detached edge writes the same bytes here it always has.
2629
+ ...spec?.detached === true ? {
2630
+ detached: true
2631
+ } : {}
2632
+ };
2633
+ });
2634
+ const toolType = persisted?.kind ?? call.kind ?? "read";
2635
+ return {
2636
+ call,
2637
+ // What the rest of the turn — the dispatched tool envelope, the stream frames — sees as this
2638
+ // call's kind, so it agrees with the branch actually taken instead of with a second lookup.
2639
+ resolvedCall: call.kind === toolType ? call : {
2640
+ ...call,
2641
+ kind: toolType
2642
+ },
2643
+ toolType,
2644
+ ...persisted?.targetAgent !== void 0 ? {
2645
+ targetAgent: persisted.targetAgent
2646
+ } : {},
2647
+ ...persisted?.detached === true ? {
2648
+ detached: true
2649
+ } : {},
2650
+ ctx: toolContext(deps, input, hooks)
2651
+ };
2652
+ }
2653
+ __name(claimToolCall, "claimToolCall");
2654
+ async function loadSkillIntoTurn(turn, claimed, startedAt) {
2655
+ const { deps, input, hooks } = turn;
2656
+ const { call } = claimed;
2657
+ const config = deps.skills;
2658
+ const offer = turn.skills;
2659
+ const outcome = await hooks.step(`tool:${call.id}`, async () => {
2660
+ if (config === void 0 || offer === void 0) {
2661
+ return {
2662
+ ok: false,
2663
+ error: "Skills are not available in this deployment."
2664
+ };
2665
+ }
2666
+ const parsed = await skillInputSchema["~standard"].validate(call.input);
2667
+ if (parsed.issues !== void 0) {
2668
+ return {
2669
+ ok: false,
2670
+ error: `invalid skill input: ${parsed.issues.map((each) => `${(each.path ?? []).join(".") || "(root)"}: ${each.message}`).join("; ")}`
2671
+ };
2672
+ }
2673
+ const loaded = await loadSkill(config, offer, parsed.value.name, skillContext(input));
2674
+ if (!loaded.ok) {
2675
+ return {
2676
+ ok: false,
2677
+ error: loaded.error
2678
+ };
2679
+ }
2680
+ return {
2681
+ ok: true,
2682
+ output: {
2683
+ name: loaded.skill.name,
2684
+ scope: loaded.skill.scope,
2685
+ body: loaded.skill.body,
2686
+ ...loaded.shadows !== void 0 ? {
2687
+ shadows: loaded.shadows
2688
+ } : {}
2689
+ }
2690
+ };
2691
+ });
2692
+ return outcome.ok ? {
2693
+ status: "executed",
2694
+ output: outcome.output,
2695
+ executionMs: Date.now() - startedAt
2696
+ } : {
2697
+ status: "failed",
2698
+ error: outcome.error,
2699
+ executionMs: Date.now() - startedAt
2700
+ };
2701
+ }
2702
+ __name(loadSkillIntoTurn, "loadSkillIntoTurn");
2703
+ async function rememberIntoTurn(turn, claimed, startedAt) {
2704
+ const { deps, input, hooks } = turn;
2705
+ const { call } = claimed;
2706
+ const config = deps.memory;
2707
+ const digest = turn.memory;
2708
+ const outcome = await hooks.step(`tool:${call.id}`, async () => {
2709
+ if (config === void 0 || digest === void 0) {
2710
+ return {
2711
+ ok: false,
2712
+ error: "Memory is not available in this deployment."
2713
+ };
2714
+ }
2715
+ const parsed = await rememberInputSchema["~standard"].validate(call.input);
2716
+ if (parsed.issues !== void 0) {
2717
+ return {
2718
+ ok: false,
2719
+ error: `invalid remember input: ${parsed.issues.map((each) => `${(each.path ?? []).join(".") || "(root)"}: ${each.message}`).join("; ")}`
2720
+ };
2721
+ }
2722
+ return await writeMemory({
2723
+ config,
2724
+ digest,
2725
+ call: parsed.value,
2726
+ ctx: skillContext(input),
2727
+ runId: hooks.runId
2728
+ });
2729
+ });
2730
+ if (!outcome.ok) {
2731
+ return {
2732
+ status: "failed",
2733
+ error: outcome.error,
2734
+ executionMs: Date.now() - startedAt
2735
+ };
2736
+ }
2737
+ publishAgentMemoryWritten({
2738
+ runId: hooks.runId,
2739
+ scope: outcome.record.scope,
2740
+ chars: outcome.record.text.length
2741
+ });
2742
+ return {
2743
+ status: "executed",
2744
+ output: outcome.record,
2745
+ executionMs: Date.now() - startedAt
2746
+ };
2747
+ }
2748
+ __name(rememberIntoTurn, "rememberIntoTurn");
2749
+ function skillContext(input) {
2750
+ return {
2751
+ actor: input.actor,
2752
+ threadId: input.threadId,
2753
+ ...input.agentName !== void 0 ? {
2754
+ agentName: input.agentName
2755
+ } : {},
2756
+ ...input.pageContext !== void 0 ? {
2757
+ pageContext: input.pageContext
2758
+ } : {}
2759
+ };
2760
+ }
2761
+ __name(skillContext, "skillContext");
2762
+ async function invokeClaimedTool(turn, claimed) {
2763
+ const { deps, input, hooks } = turn;
2764
+ const { call, resolvedCall, ctx } = claimed;
2765
+ const toolType = claimed.toolType === "action" ? "action" : "read";
2766
+ const startedAt = Date.now();
2767
+ try {
2768
+ if (claimed.toolType === "skill") {
2769
+ return await loadSkillIntoTurn(turn, claimed, startedAt);
2770
+ }
2771
+ if (claimed.toolType === "memory") {
2772
+ return await rememberIntoTurn(turn, claimed, startedAt);
2773
+ }
2774
+ let output;
2775
+ if (hooks.dispatchTool) {
2776
+ const stepCtx = {
2777
+ actor: input.actor,
2778
+ threadId: input.threadId,
2779
+ runId: hooks.runId,
2780
+ requestId: hooks.runId,
2781
+ ...input.agentName !== void 0 ? {
2782
+ agentName: input.agentName
2783
+ } : {},
2784
+ ...input.pageContext !== void 0 ? {
2785
+ pageContext: input.pageContext
2786
+ } : {}
2787
+ };
2788
+ const envelope = {
2789
+ toolName: call.name,
2790
+ input: call.input,
2791
+ ctx: stepCtx,
2792
+ ...deps.toolTimeoutMs !== void 0 ? {
2793
+ timeoutMs: deps.toolTimeoutMs
2794
+ } : {},
2795
+ // Numeric-only: the handler applies withToolTimeout AND its own local `classify` — see
2796
+ // ToolStepEnvelope.transientRetry.
2797
+ transientRetry: resolveToolTransientRetryNumbers(deps.toolTransientRetry)
2798
+ };
2799
+ output = await hooks.dispatchTool(resolvedCall, envelope);
2800
+ } else {
2801
+ const invocation = hooks.step(`tool:${call.id}`, () => traceToolExecution(hooks.runId, {
2802
+ toolCallId: call.id,
2803
+ toolName: call.name,
2804
+ toolType
2805
+ }, () => invokeWithTransientRetry(() => deps.registry.invoke(call.name, call.input, ctx, deps.rolesPolicy), deps.toolTransientRetry ?? {}, {
2806
+ ...hooks.isControlFlowError !== void 0 ? {
2807
+ isControlFlowError: hooks.isControlFlowError
2808
+ } : {},
2809
+ onRetry: /* @__PURE__ */ __name((attempt, retryError) => {
2810
+ publishAgentToolRetry({
2811
+ toolName: call.name,
2812
+ toolCallId: call.id,
2813
+ attempt,
2814
+ message: retryError instanceof Error ? retryError.message : String(retryError)
2815
+ });
2816
+ }, "onRetry")
2817
+ })));
2818
+ output = deps.toolTimeoutMs !== void 0 ? await withToolTimeout(invocation, deps.toolTimeoutMs, call.name) : await invocation;
2819
+ }
2820
+ return {
2821
+ status: "executed",
2822
+ output,
2823
+ executionMs: Date.now() - startedAt
2824
+ };
2825
+ } catch (error) {
2826
+ if (hooks.isControlFlowError?.(error) === true || isControlFlowSignal(error) || isReplayIntegrityError(error)) {
2827
+ throw error;
2828
+ }
2829
+ return {
2830
+ status: "failed",
2831
+ error: error instanceof Error ? error.message : String(error),
2832
+ executionMs: Date.now() - startedAt
2833
+ };
2834
+ }
2835
+ }
2836
+ __name(invokeClaimedTool, "invokeClaimedTool");
2837
+ async function recordToolOutcome(turn, claimed, outcome, deciderRef) {
2838
+ const { deps, hooks } = turn;
2839
+ const { call } = claimed;
2840
+ const toolType = claimed.toolType === "action" ? "action" : "read";
2841
+ if (outcome.status === "failed") {
2842
+ await hooks.step(`persist:toolfail:${call.id}`, () => deps.store.updateToolCall({
2843
+ toolCallId: call.id,
2844
+ status: "failed",
2845
+ error: outcome.error,
2846
+ executionMs: outcome.executionMs
2847
+ }));
2848
+ publishAgentToolCall({
2849
+ runId: hooks.runId,
2850
+ toolName: call.name,
2851
+ toolType,
2852
+ status: "failed",
2853
+ durationMs: outcome.executionMs
2854
+ });
2855
+ return {
2856
+ id: call.id,
2857
+ name: call.name,
2858
+ output: null,
2859
+ error: outcome.error
2860
+ };
2861
+ }
2862
+ await hooks.step(`persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
2863
+ toolCallId: call.id,
2864
+ status: "executed",
2865
+ output: outcome.output,
2866
+ executionMs: outcome.executionMs,
2867
+ ...toolType === "action" ? {
2868
+ executedByRef: deciderRef
2869
+ } : {}
2870
+ }));
2871
+ publishAgentToolCall({
2872
+ runId: hooks.runId,
2873
+ toolName: call.name,
2874
+ toolType,
2875
+ status: "executed",
2876
+ durationMs: outcome.executionMs
2877
+ });
2878
+ return {
2879
+ id: call.id,
2880
+ name: call.name,
2881
+ output: outcome.output
2882
+ };
2883
+ }
2884
+ __name(recordToolOutcome, "recordToolOutcome");
2885
+ async function delegateToolCall(turn, claimed) {
2886
+ const { deps, input, hooks } = turn;
2887
+ const { call, targetAgent = call.name } = claimed;
2888
+ const task = extractTask(call.input);
2889
+ const refusal = delegationRefusal({
2890
+ deps,
2891
+ input,
2892
+ targetAgent
2893
+ });
2894
+ const start = claimed.detached === true && refusal === null ? hooks.startAgent : void 0;
2895
+ publishAgentDelegated({
2896
+ runId: hooks.runId,
2897
+ toAgent: targetAgent,
2898
+ ...input.agentName !== void 0 ? {
2899
+ fromAgent: input.agentName
2900
+ } : {},
2901
+ ...start !== void 0 ? {
2902
+ detached: true
2903
+ } : {}
2904
+ });
2905
+ let sub;
2906
+ if (refusal !== null) {
2907
+ sub = {
2908
+ text: refusal
2909
+ };
2910
+ } else if (start !== void 0) {
2911
+ const started = await start({
2912
+ agentName: targetAgent,
2913
+ task,
2914
+ toolCallId: call.id
2915
+ });
2916
+ sub = detachedStarted({
2917
+ agent: targetAgent,
2918
+ runId: started.runId
2919
+ });
2920
+ } else if (hooks.runAgent) {
2921
+ sub = await hooks.runAgent(targetAgent, task);
2922
+ } else {
2923
+ sub = {
2924
+ text: `(no multi-agent support wired; cannot reach "${targetAgent}")`
2925
+ };
2926
+ }
2927
+ await hooks.step(`persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
2928
+ toolCallId: call.id,
2929
+ status: "executed",
2930
+ output: sub
2931
+ }));
2932
+ return {
2933
+ id: call.id,
2934
+ name: call.name,
2935
+ output: sub
631
2936
  };
632
2937
  }
633
- __name(generateFollowUps, "generateFollowUps");
634
- async function spanned(event, runId, payload, run, summarize) {
635
- let value;
636
- await trace("agent", event, async () => {
637
- value = await run();
638
- return summarize(value);
639
- }, payload, {
640
- traceId: runId
2938
+ __name(delegateToolCall, "delegateToolCall");
2939
+ async function elicitToolCall(turn, claimed) {
2940
+ const { deps, hooks } = turn;
2941
+ const { call, ctx } = claimed;
2942
+ const parsed = await askInputSchema["~standard"].validate(call.input);
2943
+ if (parsed.issues !== void 0) {
2944
+ const error = `invalid ask input: ${parsed.issues.map((each) => `${(each.path ?? []).join(".") || "(root)"}: ${each.message}`).join("; ")}`;
2945
+ await hooks.step(`persist:toolfail:${call.id}`, () => deps.store.updateToolCall({
2946
+ toolCallId: call.id,
2947
+ status: "failed",
2948
+ error
2949
+ }));
2950
+ return {
2951
+ id: call.id,
2952
+ name: call.name,
2953
+ output: null,
2954
+ error
2955
+ };
2956
+ }
2957
+ const request = {
2958
+ id: call.id,
2959
+ source: "ask",
2960
+ questions: parsed.value.questions,
2961
+ ...parsed.value.preamble !== void 0 ? {
2962
+ preamble: parsed.value.preamble
2963
+ } : {}
2964
+ };
2965
+ await hooks.step(`stream:elicitation:${call.id}`, async () => {
2966
+ await turn.writer.write(encodeStreamEvent({
2967
+ kind: "elicitation",
2968
+ id: call.id,
2969
+ request
2970
+ }));
641
2971
  });
642
- return value;
643
- }
644
- __name(spanned, "spanned");
645
- function traceLlmTurn(runId, step, run) {
646
- return spanned("llm.turn", runId, {
647
- runId,
648
- step
649
- }, run, (turn) => ({
650
- ...turn.modelId !== void 0 ? {
651
- modelId: turn.modelId
2972
+ const reply = await awaitElicitation(hooks, request, ctx);
2973
+ const result = settleElicitation(request, reply);
2974
+ await hooks.step(result.skipped ? `persist:toolreject:${call.id}` : `persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
2975
+ toolCallId: call.id,
2976
+ status: result.skipped ? "rejected" : "executed",
2977
+ output: result,
2978
+ ...result.skipped ? {
2979
+ error: "skipped by the user"
652
2980
  } : {},
653
- inputTokens: turn.usage.inputTokens,
654
- outputTokens: turn.usage.outputTokens,
655
- textLength: turn.text.length,
656
- toolCalls: turn.toolCalls.length
2981
+ ...reply.answeredByRef !== void 0 ? {
2982
+ executedByRef: reply.answeredByRef
2983
+ } : {}
657
2984
  }));
2985
+ publishAgentToolCall({
2986
+ runId: hooks.runId,
2987
+ toolName: call.name,
2988
+ toolType: "action",
2989
+ status: result.skipped ? "rejected" : "executed"
2990
+ });
2991
+ return {
2992
+ id: call.id,
2993
+ name: call.name,
2994
+ output: result
2995
+ };
658
2996
  }
659
- __name(traceLlmTurn, "traceLlmTurn");
660
- function traceToolExecution(runId, call, run) {
661
- return spanned("tool.execution", runId, {
662
- runId,
663
- ...call
664
- }, run, () => ({}));
2997
+ __name(elicitToolCall, "elicitToolCall");
2998
+ async function runClaimedToolCall(turn, claimed) {
2999
+ const { deps, input, hooks } = turn;
3000
+ const { call, toolType, ctx } = claimed;
3001
+ if (toolType === "agent") {
3002
+ return delegateToolCall(turn, claimed);
3003
+ }
3004
+ if (toolType === "ask") {
3005
+ return elicitToolCall(turn, claimed);
3006
+ }
3007
+ let deciderRef = input.actor.id;
3008
+ if (toolType === "action") {
3009
+ const decision = await hooks.awaitApproval(call, ctx);
3010
+ deciderRef = decision.executedByRef ?? input.actor.id;
3011
+ if (!decision.approved) {
3012
+ await hooks.step(`persist:toolreject:${call.id}`, () => deps.store.updateToolCall({
3013
+ toolCallId: call.id,
3014
+ status: "rejected",
3015
+ executedByRef: deciderRef,
3016
+ ...decision.reason !== void 0 ? {
3017
+ error: decision.reason
3018
+ } : {}
3019
+ }));
3020
+ publishAgentToolCall({
3021
+ runId: hooks.runId,
3022
+ toolName: call.name,
3023
+ toolType,
3024
+ status: "rejected"
3025
+ });
3026
+ return {
3027
+ id: call.id,
3028
+ name: call.name,
3029
+ output: {
3030
+ rejected: true,
3031
+ reason: decision.reason ?? "rejected by user"
3032
+ },
3033
+ error: "rejected"
3034
+ };
3035
+ }
3036
+ }
3037
+ return recordToolOutcome(turn, claimed, await invokeClaimedTool(turn, claimed), deciderRef);
665
3038
  }
666
- __name(traceToolExecution, "traceToolExecution");
3039
+ __name(runClaimedToolCall, "runClaimedToolCall");
3040
+ async function invokeClaimedToolsTogether(turn, parallel, claimed) {
3041
+ const settled = await parallel(claimed.map((entry) => () => invokeClaimedTool(turn, entry)));
3042
+ for (const outcome of settled) {
3043
+ if (!outcome.ok) {
3044
+ throw outcome.error;
3045
+ }
3046
+ }
3047
+ const results = [];
3048
+ for (const [index, entry] of claimed.entries()) {
3049
+ const outcome = settled[index];
3050
+ if (outcome?.ok === true) {
3051
+ results.push(await recordToolOutcome(turn, entry, outcome.value, turn.input.actor.id));
3052
+ }
3053
+ }
3054
+ return results;
3055
+ }
3056
+ __name(invokeClaimedToolsTogether, "invokeClaimedToolsTogether");
667
3057
  async function runAgentLoop(deps, input, hooks) {
668
3058
  const maxSteps = deps.maxSteps ?? 8;
669
3059
  let system = await resolveSystemPrompt(deps, input);
3060
+ const inputProcessors = deps.inputProcessors ?? [];
3061
+ const outputProcessors = deps.outputProcessors ?? [];
3062
+ const gateMode = resolveOutputGateMode(outputProcessors);
3063
+ const gated = gateMode !== "off";
3064
+ const gateLookback = resolveGateLookback(outputProcessors);
3065
+ let structured;
670
3066
  if (deps.quota !== void 0) {
671
3067
  const quota = deps.quota;
672
3068
  const state = await hooks.step("quota:check", () => quota.check(input.actor.id, deps.day));
@@ -700,25 +3096,18 @@ async function runAgentLoop(deps, input, hooks) {
700
3096
  threadId: input.threadId,
701
3097
  role: "user",
702
3098
  content: input.userText,
3099
+ runId: hooks.runId,
703
3100
  ...input.attachments !== void 0 ? {
704
3101
  attachments: input.attachments
705
3102
  } : {}
706
3103
  }));
707
3104
  }
708
- const thread = await hooks.step("load:thread", () => deps.store.getThread(input.threadId));
709
- const modelMessages = (thread?.messages ?? []).map((message) => ({
710
- role: message.role,
711
- content: message.content,
712
- ...message.toolCalls !== void 0 ? {
713
- toolCalls: message.toolCalls
714
- } : {},
715
- ...message.toolResults !== void 0 ? {
716
- toolResults: message.toolResults
717
- } : {},
718
- ...message.attachments !== void 0 ? {
719
- attachments: message.attachments
720
- } : {}
721
- }));
3105
+ const history = await (hooks.patched?.(SELECTED_HISTORY_PATCH) ?? Promise.resolve(true)) ? await loadSelectedHistory(deps, input, hooks) : await loadWholeThread(deps, input, hooks);
3106
+ let modelMessages = history.messages;
3107
+ const summarize = deps.historyPolicy?.summarize?.bind(deps.historyPolicy);
3108
+ if (summarize !== void 0 && history.dropped.length > 0) {
3109
+ modelMessages = await foldDroppedHistory(summarize, deps, input, hooks, history);
3110
+ }
722
3111
  const writer = await hooks.openSink();
723
3112
  let lastText = "";
724
3113
  let steps = 0;
@@ -742,9 +3131,41 @@ async function runAgentLoop(deps, input, hooks) {
742
3131
  ...input.agentName !== void 0 ? {
743
3132
  agentName: input.agentName
744
3133
  } : {},
3134
+ ...input.parentRunId !== void 0 ? {
3135
+ parentRunId: input.parentRunId
3136
+ } : {},
745
3137
  promptHash
746
3138
  });
747
3139
  });
3140
+ let memoryDigest;
3141
+ if (deps.memory !== void 0) {
3142
+ const config = deps.memory;
3143
+ memoryDigest = await hooks.step("memory:digest", () => offerMemories({
3144
+ config,
3145
+ ctx: skillContext(input),
3146
+ query: input.userText
3147
+ }));
3148
+ const digest = memoryDigest;
3149
+ const block = digest.entries.length > 0 ? buildMemoryBlock({
3150
+ entries: digest.entries,
3151
+ writable: memoryIsWritable(deps),
3152
+ partial: digest.omitted > 0 || digest.recalled === true
3153
+ }) : "";
3154
+ if (block.length > 0) {
3155
+ system = `${system}
3156
+
3157
+ ${block}`;
3158
+ }
3159
+ publishAgentMemoryResolved({
3160
+ runId: hooks.runId,
3161
+ scopes: digest.scopes.length,
3162
+ offered: digest.entries.length,
3163
+ omitted: digest.omitted,
3164
+ pinnedOmitted: digest.pinnedOmitted,
3165
+ recalled: digest.recalled === true,
3166
+ promptChars: block.length
3167
+ });
3168
+ }
748
3169
  let injectedPassages;
749
3170
  if (deps.retriever !== void 0) {
750
3171
  const retriever = deps.retriever;
@@ -770,6 +3191,24 @@ ${buildContextBlock(passages)}`;
770
3191
  count: passages.length
771
3192
  });
772
3193
  }
3194
+ let skillOffer;
3195
+ if (deps.skills !== void 0) {
3196
+ const config = deps.skills;
3197
+ skillOffer = await hooks.step("skills:catalog", () => offerSkills(config, skillContext(input)));
3198
+ const block = skillOffer.entries.length > 0 ? buildSkillsBlock(skillOffer.entries) : "";
3199
+ if (block.length > 0) {
3200
+ system = `${system}
3201
+
3202
+ ${block}`;
3203
+ }
3204
+ publishAgentSkillsResolved({
3205
+ runId: hooks.runId,
3206
+ scopes: skillOffer.scopes.length,
3207
+ offered: skillOffer.entries.length,
3208
+ omitted: skillOffer.omitted,
3209
+ promptChars: block.length
3210
+ });
3211
+ }
773
3212
  let prices = [];
774
3213
  if (deps.pricingStore !== void 0) {
775
3214
  const pricingStore = deps.pricingStore;
@@ -779,36 +3218,124 @@ ${buildContextBlock(passages)}`;
779
3218
  price.modelId,
780
3219
  price
781
3220
  ]));
3221
+ if (deps.intake !== void 0) {
3222
+ const asked = await runIntake(deps.intake, deps, input, hooks, writer, history.hasAssistantMessage);
3223
+ if (asked !== null) {
3224
+ modelMessages.push(asked);
3225
+ }
3226
+ }
3227
+ const cancellable = hooks.cancelled !== void 0 && await (hooks.patched?.(CANCELLATION_PATCH) ?? Promise.resolve(true));
782
3228
  for (let i = 0; i < maxSteps; i += 1) {
3229
+ await haltIfCancelled(hooks, cancellable, `cancel:check:${i}`);
783
3230
  await hooks.step(`stream:step-start:${i}`, async () => {
784
3231
  await writer.write(encodeStreamEvent({
785
3232
  kind: "step-start"
786
3233
  }));
787
3234
  });
3235
+ let prompt = {
3236
+ system,
3237
+ messages: modelMessages
3238
+ };
3239
+ if (inputProcessors.length > 0) {
3240
+ prompt = await hooks.step(`process:input:${i}`, () => runInputProcessors(inputProcessors, prompt, processorContext(input, i)));
3241
+ }
788
3242
  let turn;
789
3243
  if (hooks.dispatchLlm) {
790
3244
  turn = await hooks.dispatchLlm(i, {
791
3245
  ...input.agentName !== void 0 ? {
792
3246
  agentName: input.agentName
793
3247
  } : {},
794
- system,
795
- messages: modelMessages,
796
- actor: input.actor
3248
+ system: prompt.system,
3249
+ messages: prompt.messages,
3250
+ actor: input.actor,
3251
+ ...gated ? {
3252
+ bufferOutput: true
3253
+ } : {}
797
3254
  });
798
3255
  } else {
799
- const tools = await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList);
800
- turn = await hooks.step(`llm:${i}`, () => traceLlmTurn(hooks.runId, i, () => deps.model.runTurn({
801
- system,
802
- messages: modelMessages,
803
- tools,
804
- sink: writer
805
- })));
3256
+ const tools = withMemoryTool({
3257
+ tools: withSkillTool({
3258
+ tools: withAskTool({
3259
+ tools: await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList),
3260
+ ask: deps.ask
3261
+ }),
3262
+ enabled: deps.skills !== void 0
3263
+ }),
3264
+ enabled: memoryIsWritable(deps)
3265
+ });
3266
+ turn = await hooks.step(`llm:${i}`, async () => {
3267
+ const incremental = gateMode === "incremental" ? createIncrementalGate({
3268
+ processors: outputProcessors,
3269
+ ctx: processorContext(input, i),
3270
+ lookbackChars: gateLookback,
3271
+ writer
3272
+ }) : void 0;
3273
+ const buffer = gateMode === "whole" ? createFrameBuffer() : void 0;
3274
+ const result = await traceLlmTurn(hooks.runId, i, () => deps.model.runTurn({
3275
+ system: prompt.system,
3276
+ messages: prompt.messages,
3277
+ tools,
3278
+ sink: incremental?.writer ?? buffer?.writer ?? writer
3279
+ }));
3280
+ if (incremental !== void 0) {
3281
+ await incremental.settled();
3282
+ const refusal = incremental.rejection();
3283
+ return {
3284
+ ...result,
3285
+ releasedText: incremental.released(),
3286
+ ...refusal !== void 0 ? {
3287
+ gateRejection: refusal
3288
+ } : {}
3289
+ };
3290
+ }
3291
+ return buffer === void 0 ? result : {
3292
+ ...result,
3293
+ bufferedFrames: buffer.frames()
3294
+ };
3295
+ });
3296
+ }
3297
+ let rejection;
3298
+ if (gated) {
3299
+ const gate = await hooks.step(`process:output:${i}`, async () => {
3300
+ const released = turn.releasedText;
3301
+ if (turn.gateRejection !== void 0) {
3302
+ return {
3303
+ text: turn.text,
3304
+ rejection: turn.gateRejection
3305
+ };
3306
+ }
3307
+ const settled = await runOutputProcessors(outputProcessors, {
3308
+ text: turn.text,
3309
+ toolCalls: turn.toolCalls
3310
+ }, processorContext(input, i));
3311
+ if (settled.rejection === void 0) {
3312
+ if (released === void 0) {
3313
+ for (const chunk of releaseGatedFrames(turn.bufferedFrames ?? [], settled.text)) {
3314
+ await writer.write(chunk);
3315
+ }
3316
+ } else {
3317
+ const tail = gateTail(outputProcessors, released, settled.text);
3318
+ if (tail.length > 0) {
3319
+ await writer.write(encodeStreamEvent({
3320
+ kind: "text",
3321
+ text: tail
3322
+ }));
3323
+ }
3324
+ }
3325
+ }
3326
+ return settled;
3327
+ });
3328
+ rejection = gate.rejection;
3329
+ turn = {
3330
+ ...turn,
3331
+ text: gate.text
3332
+ };
806
3333
  }
807
3334
  const resolvedModelId = turn.modelId ?? deps.modelId ?? "unknown";
808
3335
  const costUsd = resolveCostUsd(turn.usage, turn.costUsd, priceByModel.get(resolvedModelId));
809
3336
  const toolCallsWithKind = turn.toolCalls.map((call) => ({
810
3337
  ...call,
811
- kind: deps.registry.spec(call.name)?.kind ?? "read"
3338
+ kind: declaredKind(deps, call.name)
812
3339
  }));
813
3340
  await hooks.step(`persist:usage:${i}`, () => deps.store.recordUsage({
814
3341
  threadId: input.threadId,
@@ -825,6 +3352,9 @@ ${buildContextBlock(passages)}`;
825
3352
  const quota = deps.quota;
826
3353
  await hooks.step(`quota:bump:${i}`, () => quota.bump(input.actor.id, deps.day, turn.usage.inputTokens + turn.usage.outputTokens));
827
3354
  }
3355
+ if (rejection !== void 0) {
3356
+ throw new OutputRejectedError(rejection.processor, rejection.reason);
3357
+ }
828
3358
  steps += 1;
829
3359
  totalInput += turn.usage.inputTokens;
830
3360
  totalOutput += turn.usage.outputTokens;
@@ -836,6 +3366,9 @@ ${buildContextBlock(passages)}`;
836
3366
  textLength: turn.text.length
837
3367
  });
838
3368
  const isFinalTurn = turn.toolCalls.length === 0;
3369
+ if (isFinalTurn && deps.outputSchema !== void 0) {
3370
+ structured = await structureAnswer(deps.outputSchema, deps, input, hooks, restatementPrompt(prompt.messages, turn.text, deps.outputFromTranscript === true), i);
3371
+ }
839
3372
  let followUps;
840
3373
  if (isFinalTurn && deps.followUpsCount !== void 0 && deps.followUpsCount > 0) {
841
3374
  const count = deps.followUpsCount;
@@ -857,8 +3390,12 @@ ${buildContextBlock(passages)}`;
857
3390
  modelId: result.modelId
858
3391
  } : {}
859
3392
  })));
860
- if (generated.followUps.length > 0) {
861
- followUps = generated.followUps;
3393
+ let suggestions = generated.followUps;
3394
+ if (outputProcessors.length > 0) {
3395
+ suggestions = await hooks.step(`process:output:followups:${i}`, () => gateFollowUps(outputProcessors, suggestions, processorContext(input, i)));
3396
+ }
3397
+ if (suggestions.length > 0) {
3398
+ followUps = suggestions;
862
3399
  }
863
3400
  await hooks.step(`persist:usage:followups:${i}`, () => deps.store.recordUsage({
864
3401
  threadId: input.threadId,
@@ -868,10 +3405,58 @@ ${buildContextBlock(passages)}`;
868
3405
  usage: generated.usage
869
3406
  }));
870
3407
  }
3408
+ const synthetic = [];
3409
+ if (i === 0 && injectedPassages !== void 0) {
3410
+ const call = {
3411
+ id: `retrieve-${hooks.runId}`,
3412
+ name: "retrieve",
3413
+ input: {
3414
+ query: input.userText
3415
+ },
3416
+ kind: "read"
3417
+ };
3418
+ const output = {
3419
+ passages: injectedPassages
3420
+ };
3421
+ synthetic.push({
3422
+ step: "persist:retrieval",
3423
+ call,
3424
+ result: {
3425
+ id: call.id,
3426
+ name: call.name,
3427
+ output
3428
+ }
3429
+ });
3430
+ }
3431
+ if (structured !== void 0) {
3432
+ const call = {
3433
+ id: `structured-${hooks.runId}`,
3434
+ name: "structured_output",
3435
+ input: {},
3436
+ kind: "read"
3437
+ };
3438
+ synthetic.push({
3439
+ step: "persist:structured",
3440
+ call,
3441
+ result: {
3442
+ id: call.id,
3443
+ name: call.name,
3444
+ output: structured
3445
+ }
3446
+ });
3447
+ }
3448
+ const messageCalls = [
3449
+ ...toolCallsWithKind,
3450
+ ...synthetic.map((entry) => entry.call)
3451
+ ];
3452
+ const syntheticResults = synthetic.map((entry) => entry.result);
871
3453
  const assistant = await hooks.step(`persist:assistant:${i}`, () => deps.store.appendMessage({
872
3454
  threadId: input.threadId,
873
3455
  role: "assistant",
874
3456
  content: turn.text,
3457
+ // Stamps which turn wrote this message, so a reader does not have to infer it from
3458
+ // timestamps — an inference a regenerate breaks, since it re-answers a surviving prompt.
3459
+ runId: hooks.runId,
875
3460
  usage: {
876
3461
  ...turn.usage,
877
3462
  costUsd
@@ -879,8 +3464,11 @@ ${buildContextBlock(passages)}`;
879
3464
  ...input.agentName !== void 0 ? {
880
3465
  agentName: input.agentName
881
3466
  } : {},
882
- ...toolCallsWithKind.length > 0 ? {
883
- toolCalls: toolCallsWithKind
3467
+ ...messageCalls.length > 0 ? {
3468
+ toolCalls: messageCalls
3469
+ } : {},
3470
+ ...syntheticResults.length > 0 ? {
3471
+ toolResults: syntheticResults
884
3472
  } : {},
885
3473
  ...followUps !== void 0 ? {
886
3474
  followUps
@@ -889,33 +3477,43 @@ ${buildContextBlock(passages)}`;
889
3477
  const assistantMessage = {
890
3478
  role: "assistant",
891
3479
  content: turn.text,
892
- ...toolCallsWithKind.length > 0 ? {
893
- toolCalls: toolCallsWithKind
3480
+ ...messageCalls.length > 0 ? {
3481
+ toolCalls: messageCalls
3482
+ } : {},
3483
+ ...syntheticResults.length > 0 ? {
3484
+ toolResults: syntheticResults
894
3485
  } : {}
895
3486
  };
896
3487
  modelMessages.push(assistantMessage);
897
- if (i === 0 && injectedPassages !== void 0) {
898
- const passages = injectedPassages;
899
- const toolCallId = `retrieve-${assistant.id}`;
900
- await hooks.step(`persist:retrieval:${assistant.id}`, async () => {
3488
+ for (const entry of synthetic) {
3489
+ const { call, result } = entry;
3490
+ await hooks.step(`${entry.step}:${assistant.id}`, async () => {
901
3491
  await deps.store.recordToolCall({
902
- toolCallId,
3492
+ toolCallId: call.id,
903
3493
  messageId: assistant.id,
904
- toolName: "retrieve",
3494
+ toolName: call.name,
905
3495
  toolType: "read",
906
- input: {
907
- query: input.userText
908
- },
3496
+ input: call.input,
909
3497
  status: "auto_executed",
910
3498
  runId: hooks.runId
911
3499
  });
912
3500
  await deps.store.updateToolCall({
913
- toolCallId,
3501
+ toolCallId: call.id,
914
3502
  status: "executed",
915
- output: {
916
- passages
917
- }
3503
+ output: result.output
918
3504
  });
3505
+ await writer.write(encodeStreamEvent({
3506
+ kind: "tool-input-available",
3507
+ id: call.id,
3508
+ name: call.name,
3509
+ input: call.input,
3510
+ toolKind: "read"
3511
+ }));
3512
+ await writer.write(encodeStreamEvent({
3513
+ kind: "tool-output",
3514
+ id: call.id,
3515
+ output: result.output
3516
+ }));
919
3517
  });
920
3518
  }
921
3519
  if (isFinalTurn) {
@@ -928,218 +3526,46 @@ ${buildContextBlock(passages)}`;
928
3526
  });
929
3527
  break;
930
3528
  }
3529
+ await haltIfCancelled(hooks, cancellable, `cancel:tools:${i}`);
931
3530
  const results = [];
932
- for (const call of toolCallsWithKind) {
933
- const spec = deps.registry.spec(call.name);
934
- const toolType = call.kind ?? "read";
935
- const ctx = {
936
- actor: input.actor,
937
- threadId: input.threadId,
938
- runId: hooks.runId,
939
- requestId: hooks.runId,
940
- ...input.agentName !== void 0 ? {
941
- agentName: input.agentName
942
- } : {},
943
- ...input.pageContext !== void 0 ? {
944
- pageContext: input.pageContext
945
- } : {},
946
- ...deps.host !== void 0 ? {
947
- host: deps.host
948
- } : {}
949
- };
950
- if (toolType === "agent") {
951
- const targetAgent = spec?.targetAgent ?? call.name;
952
- const task = extractTask(call.input);
953
- await hooks.step(`persist:toolcall:${call.id}`, () => deps.store.recordToolCall({
954
- toolCallId: call.id,
955
- messageId: assistant.id,
956
- toolName: call.name,
957
- toolType: "read",
958
- input: call.input,
959
- status: "auto_executed",
960
- runId: hooks.runId
961
- }));
962
- const overDepth = (input.delegationDepth ?? 0) >= MAX_DELEGATION_DEPTH;
963
- publishAgentDelegated({
964
- runId: hooks.runId,
965
- toAgent: targetAgent,
966
- ...input.agentName !== void 0 ? {
967
- fromAgent: input.agentName
968
- } : {}
969
- });
970
- let sub;
971
- if (overDepth) {
972
- sub = {
973
- text: `(delegation depth limit of ${MAX_DELEGATION_DEPTH} reached)`
974
- };
975
- } else if (hooks.runAgent) {
976
- sub = await hooks.runAgent(targetAgent, task);
977
- } else {
978
- sub = {
979
- text: `(no multi-agent support wired; cannot reach "${targetAgent}")`
980
- };
981
- }
982
- await hooks.step(`persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
983
- toolCallId: call.id,
984
- status: "executed",
985
- output: sub
986
- }));
987
- results.push({
988
- id: call.id,
989
- name: call.name,
990
- output: sub
991
- });
992
- continue;
3531
+ const turnCalls = {
3532
+ deps,
3533
+ input,
3534
+ hooks,
3535
+ messageId: assistant.id,
3536
+ writer,
3537
+ ...skillOffer !== void 0 ? {
3538
+ skills: skillOffer
3539
+ } : {},
3540
+ ...memoryDigest !== void 0 ? {
3541
+ memory: memoryDigest
3542
+ } : {}
3543
+ };
3544
+ const parallel = hooks.parallel;
3545
+ if (parallel !== void 0 && toolCallsWithKind.length > 1 && await (hooks.patched?.(PARALLEL_TOOLS_PATCH) ?? Promise.resolve(true))) {
3546
+ const claimed = [];
3547
+ for (const call of toolCallsWithKind) {
3548
+ claimed.push(await claimToolCall(turnCalls, call));
993
3549
  }
994
- let deciderRef = input.actor.id;
995
- if (toolType === "action") {
996
- await hooks.step(`persist:toolcall:${call.id}`, () => deps.store.recordToolCall({
997
- toolCallId: call.id,
998
- messageId: assistant.id,
999
- toolName: call.name,
1000
- toolType: "action",
1001
- input: call.input,
1002
- status: "pending_approval",
1003
- runId: hooks.runId
1004
- }));
1005
- const decision = await hooks.awaitApproval(call, ctx);
1006
- deciderRef = decision.executedByRef ?? input.actor.id;
1007
- if (!decision.approved) {
1008
- await hooks.step(`persist:toolreject:${call.id}`, () => deps.store.updateToolCall({
1009
- toolCallId: call.id,
1010
- status: "rejected",
1011
- executedByRef: deciderRef,
1012
- ...decision.reason !== void 0 ? {
1013
- error: decision.reason
1014
- } : {}
1015
- }));
1016
- results.push({
1017
- id: call.id,
1018
- name: call.name,
1019
- output: {
1020
- rejected: true,
1021
- reason: decision.reason ?? "rejected by user"
1022
- },
1023
- error: "rejected"
1024
- });
1025
- publishAgentToolCall({
1026
- runId: hooks.runId,
1027
- toolName: call.name,
1028
- toolType,
1029
- status: "rejected"
1030
- });
1031
- continue;
1032
- }
3550
+ if (claimed.every((entry) => entry.toolType === "read" || entry.toolType === "skill" || entry.toolType === "memory")) {
3551
+ results.push(...await invokeClaimedToolsTogether(turnCalls, parallel, claimed));
1033
3552
  } else {
1034
- await hooks.step(`persist:toolcall:${call.id}`, () => deps.store.recordToolCall({
1035
- toolCallId: call.id,
1036
- messageId: assistant.id,
1037
- toolName: call.name,
1038
- toolType: "read",
1039
- input: call.input,
1040
- status: "auto_executed",
1041
- runId: hooks.runId
1042
- }));
1043
- }
1044
- const startedAt2 = Date.now();
1045
- try {
1046
- let output;
1047
- if (hooks.dispatchTool) {
1048
- const stepCtx = {
1049
- actor: input.actor,
1050
- threadId: input.threadId,
1051
- runId: hooks.runId,
1052
- requestId: hooks.runId,
1053
- ...input.agentName !== void 0 ? {
1054
- agentName: input.agentName
1055
- } : {},
1056
- ...input.pageContext !== void 0 ? {
1057
- pageContext: input.pageContext
1058
- } : {}
1059
- };
1060
- const envelope = {
1061
- toolName: call.name,
1062
- input: call.input,
1063
- ctx: stepCtx,
1064
- ...deps.toolTimeoutMs !== void 0 ? {
1065
- timeoutMs: deps.toolTimeoutMs
1066
- } : {},
1067
- // Numeric-only: the handler applies withToolTimeout AND its own local `classify` — see
1068
- // ToolStepEnvelope.transientRetry.
1069
- transientRetry: resolveToolTransientRetryNumbers(deps.toolTransientRetry)
1070
- };
1071
- output = await hooks.dispatchTool(call, envelope);
1072
- } else {
1073
- const invocation = hooks.step(`tool:${call.id}`, () => traceToolExecution(hooks.runId, {
1074
- toolCallId: call.id,
1075
- toolName: call.name,
1076
- toolType
1077
- }, () => invokeWithTransientRetry(() => deps.registry.invoke(call.name, call.input, ctx, deps.rolesPolicy), deps.toolTransientRetry ?? {}, {
1078
- ...hooks.isControlFlowError !== void 0 ? {
1079
- isControlFlowError: hooks.isControlFlowError
1080
- } : {},
1081
- onRetry: /* @__PURE__ */ __name((attempt, retryError) => {
1082
- publishAgentToolRetry({
1083
- toolName: call.name,
1084
- toolCallId: call.id,
1085
- attempt,
1086
- message: retryError instanceof Error ? retryError.message : String(retryError)
1087
- });
1088
- }, "onRetry")
1089
- })));
1090
- output = deps.toolTimeoutMs !== void 0 ? await withToolTimeout(invocation, deps.toolTimeoutMs, call.name) : await invocation;
1091
- }
1092
- const executionMs = Date.now() - startedAt2;
1093
- await hooks.step(`persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
1094
- toolCallId: call.id,
1095
- status: "executed",
1096
- output,
1097
- executionMs,
1098
- ...toolType === "action" ? {
1099
- executedByRef: deciderRef
1100
- } : {}
1101
- }));
1102
- results.push({
1103
- id: call.id,
1104
- name: call.name,
1105
- output
1106
- });
1107
- publishAgentToolCall({
1108
- runId: hooks.runId,
1109
- toolName: call.name,
1110
- toolType,
1111
- status: "executed",
1112
- durationMs: executionMs
1113
- });
1114
- } catch (error) {
1115
- if (hooks.isControlFlowError?.(error) === true) {
1116
- throw error;
3553
+ for (const entry of claimed) {
3554
+ results.push(await runClaimedToolCall(turnCalls, entry));
1117
3555
  }
1118
- const executionMs = Date.now() - startedAt2;
1119
- const message = error instanceof Error ? error.message : String(error);
1120
- await hooks.step(`persist:toolfail:${call.id}`, () => deps.store.updateToolCall({
1121
- toolCallId: call.id,
1122
- status: "failed",
1123
- error: message,
1124
- executionMs
1125
- }));
1126
- results.push({
1127
- id: call.id,
1128
- name: call.name,
1129
- output: null,
1130
- error: message
1131
- });
1132
- publishAgentToolCall({
1133
- runId: hooks.runId,
1134
- toolName: call.name,
1135
- toolType,
1136
- status: "failed",
1137
- durationMs: executionMs
1138
- });
3556
+ }
3557
+ } else {
3558
+ for (const call of toolCallsWithKind) {
3559
+ results.push(await runClaimedToolCall(turnCalls, await claimToolCall(turnCalls, call)));
1139
3560
  }
1140
3561
  }
1141
- assistantMessage.toolResults = results;
3562
+ const settledResults = [
3563
+ ...results,
3564
+ ...syntheticResults
3565
+ ];
3566
+ assistantMessage.toolResults = settledResults;
1142
3567
  await hooks.step(`stream:tool-outputs:${i}`, async () => {
3568
+ await deps.store.setMessageToolResults(assistant.id, settledResults);
1143
3569
  for (const result of results) {
1144
3570
  await writer.write(encodeStreamEvent(result.error !== void 0 ? {
1145
3571
  kind: "tool-output-error",
@@ -1160,9 +3586,34 @@ ${buildContextBlock(passages)}`;
1160
3586
  }));
1161
3587
  });
1162
3588
  }
1163
- if (thread !== null && (thread.title === "" || thread.title === "New chat")) {
3589
+ if (history.title === "" || history.title === "New chat") {
1164
3590
  await hooks.step("persist:title", () => deps.store.setTitle(input.threadId, deriveTitle(input.userText)));
1165
3591
  }
3592
+ const delivery = input.deliverTo;
3593
+ if (delivery !== void 0) {
3594
+ await hooks.step("deliver:detached", async () => {
3595
+ if (await deps.store.getThread(delivery.threadId) === null) {
3596
+ return;
3597
+ }
3598
+ const agent = input.agentName ?? "default";
3599
+ await deps.store.appendMessage({
3600
+ threadId: delivery.threadId,
3601
+ role: "assistant",
3602
+ content: lastText,
3603
+ agentName: agent,
3604
+ runId: hooks.runId
3605
+ });
3606
+ await deps.store.updateToolCall({
3607
+ toolCallId: delivery.toolCallId,
3608
+ status: "executed",
3609
+ output: detachedDelivered({
3610
+ agent,
3611
+ runId: hooks.runId,
3612
+ text: lastText
3613
+ })
3614
+ });
3615
+ });
3616
+ }
1166
3617
  await hooks.step("persist:run:end", async () => {
1167
3618
  await deps.store.recordRunEnd?.({
1168
3619
  runId: hooks.runId,
@@ -1179,7 +3630,10 @@ ${buildContextBlock(passages)}`;
1179
3630
  outputTokens: totalOutput
1180
3631
  });
1181
3632
  return {
1182
- text: lastText
3633
+ text: lastText,
3634
+ ...structured !== void 0 ? {
3635
+ object: structured
3636
+ } : {}
1183
3637
  };
1184
3638
  }
1185
3639
  __name(runAgentLoop, "runAgentLoop");
@@ -1193,6 +3647,7 @@ export {
1193
3647
  AGENT_DURABLE_RUNNER,
1194
3648
  AGENT_EMBEDDING_PROVIDER,
1195
3649
  AGENT_GOVERNANCE_QUERIES,
3650
+ AGENT_MEMORY,
1196
3651
  AGENT_MODEL,
1197
3652
  AGENT_OPTIONS,
1198
3653
  AGENT_PRICING_STORE,
@@ -1203,48 +3658,131 @@ export {
1203
3658
  AGENT_ROLES_POLICY,
1204
3659
  AGENT_RUNNER,
1205
3660
  AGENT_SINK,
3661
+ AGENT_SKILLS,
3662
+ AGENT_SKILL_SOURCES,
1206
3663
  AGENT_SPAN_EVENTS,
1207
3664
  AGENT_STORE,
1208
3665
  AGENT_TOOL_REGISTRY,
3666
+ ASK_TOOL_DESCRIPTION,
3667
+ ASK_TOOL_NAME,
1209
3668
  AgentRegistry,
1210
3669
  AgentStreamError,
3670
+ DEFAULT_HISTORY_SUMMARY_INSTRUCTION,
3671
+ DEFAULT_INCREMENTAL_LOOKBACK_CHARS,
3672
+ DEFAULT_INTAKE_PREAMBLE,
3673
+ DEFAULT_MAX_FACT_CHARS,
3674
+ DEFAULT_MAX_MEMORIES,
3675
+ DEFAULT_MAX_SKILLS,
3676
+ DEFAULT_STRUCTURED_OUTPUT_INSTRUCTION,
1211
3677
  DEFAULT_TOOL_TRANSIENT_RETRY_ATTEMPTS,
1212
3678
  DEFAULT_TOOL_TRANSIENT_RETRY_BACKOFF_MS,
1213
3679
  DefaultRolesPolicy,
3680
+ GLOBAL_SCOPE,
3681
+ MAX_ASK_QUESTIONS,
3682
+ OutputRejectedError,
3683
+ ProcessorFailedError,
1214
3684
  QuotaExceededError,
3685
+ REMEMBER_TOOL_DESCRIPTION,
3686
+ REMEMBER_TOOL_NAME,
3687
+ RunCancelledError,
3688
+ SKILL_TOOL_DESCRIPTION,
3689
+ SKILL_TOOL_NAME,
3690
+ StructuredOutputError,
1215
3691
  THREAD_DETAIL_CONTENT_CHARS,
3692
+ ToolDisabledError,
1216
3693
  ToolForbiddenError,
1217
3694
  ToolInputInvalidError,
1218
3695
  ToolNotFoundError,
1219
3696
  ToolRegistry,
3697
+ actorScope,
1220
3698
  agentDiagnosticKey,
3699
+ agentFailureCode,
3700
+ askInputSchema,
3701
+ askToolDefinition,
1221
3702
  bucketByActor,
1222
3703
  bucketByModel,
1223
3704
  bucketByThread,
1224
3705
  bucketUsageTrend,
3706
+ buildMemoryBlock,
3707
+ buildSkillsBlock,
3708
+ canActorUseTool,
3709
+ compositeSkillProvider,
3710
+ createFrameBuffer,
3711
+ createIncrementalGate,
1225
3712
  dayBoundsUtc,
3713
+ decodeStreamEvent,
3714
+ defaultScopeResolver,
3715
+ detachedDelivered,
3716
+ detachedStarted,
3717
+ detachedUnsettled,
1226
3718
  encodeStreamEvent,
1227
3719
  estimateCost,
3720
+ estimateMessageTokens,
3721
+ extractJson,
1228
3722
  filterToolsByAllowList,
3723
+ filterToolsByCanUse,
3724
+ filterToolsByEnabled,
1229
3725
  filterToolsByRole,
3726
+ gateFollowUps,
3727
+ gateTail,
1230
3728
  invokeWithTransientRetry,
3729
+ isControlFlowSignal,
3730
+ isReplayIntegrityError,
3731
+ isToolEnabled,
1231
3732
  isTransientToolError,
3733
+ loadSkill,
3734
+ memoryForgetVerdict,
3735
+ memoryWriteVerdict,
3736
+ normalizeDelegation,
3737
+ normalizeElicitationReply,
3738
+ offerMemories,
3739
+ offerSkills,
1232
3740
  publishAgentDelegated,
3741
+ publishAgentMemoryResolved,
3742
+ publishAgentMemoryWritten,
1233
3743
  publishAgentMessage,
1234
3744
  publishAgentQuotaExceeded,
1235
3745
  publishAgentRetrieved,
1236
3746
  publishAgentRunFailed,
1237
3747
  publishAgentRunFinished,
1238
3748
  publishAgentRunStarted,
3749
+ publishAgentSkillsResolved,
1239
3750
  publishAgentToolCall,
1240
3751
  publishAgentToolRetry,
3752
+ releaseGatedFrames,
3753
+ rememberInputSchema,
3754
+ rememberToolDefinition,
3755
+ renderElicitationAnswers,
3756
+ repairInstruction,
3757
+ resolveElicitation,
3758
+ resolveGateLookback,
3759
+ resolveMemoryDigest,
3760
+ resolveOutputGateMode,
3761
+ resolveSkillCatalog,
1241
3762
  resolveToolTransientRetryNumbers,
1242
3763
  rollupThreadUsage,
1243
3764
  runAgentLoop,
3765
+ runInputProcessors,
3766
+ runOutputProcessors,
1244
3767
  seedModelPrices,
3768
+ settleAll,
3769
+ settleElicitation,
3770
+ settleUnsettledDelegation,
3771
+ skillInputSchema,
3772
+ skillToolDefinition,
3773
+ skillWriteVerdict,
3774
+ staticSkillProvider,
3775
+ summarizeWithModel,
3776
+ tenantScope,
1245
3777
  traceLlmTurn,
1246
3778
  traceToolExecution,
1247
3779
  truncateDetailContent,
1248
- withToolTimeout
3780
+ validateStructured,
3781
+ windowHistory,
3782
+ withAskTool,
3783
+ withMemoryTool,
3784
+ withSkillTool,
3785
+ withToolTimeout,
3786
+ writeMemory
1249
3787
  };
1250
3788
  //# sourceMappingURL=index.js.map