@gajae-code/agent-core 0.18.10 → 0.18.12

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/CHANGELOG.md CHANGED
@@ -2,6 +2,26 @@
2
2
 
3
3
  ## [Unreleased]
4
4
 
5
+ ## [0.18.12] - 2026-10-16
6
+
7
+ ### Fixed
8
+
9
+ - Preserve a completed terminal outcome when a subscriber aborts after an accepted assistant response with no pending tool calls.
10
+
11
+ - Standalone agent loops now report `agent_end.stopReason: "cancelled"` when cancellation aborts a provider request or pending tool execution, including when a pause is requested at the same boundary.
12
+
13
+ - Preserve completed agent runs when a subscriber aborts after a final assistant message is accepted, while still reporting cancellation for interrupted tool execution and continuations.
14
+
15
+ - A project `.env` value of `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT` no longer turns GenAI message-content capture on. An operator value the project does not declare is unchanged, including a downgrade before the first read. Removing the variable before that read keeps capture off, even when a trusted agent env file still says `full`.
16
+
17
+ - The typed `incompleteArgumentsReason` an adapter recorded on a tool-call block survives the accepted attempt snapshot. The snapshot rewrote every incomplete call to `malformed`, so a call the adapter had typed as `truncated`, `conflicting`, or `ambiguous` reached dispatch with only the boolean left and was reported with the generic wording.
18
+ - An incomplete tool call with no typed reason is no longer reported as an output-token-limit truncation unless the turn actually stopped for length, or as a decode failure the provider never declared. The rejection keeps the same actionable `Re-issue the call` guidance.
19
+
20
+ - Billing callbacks now receive the provider-effective service tier after a downgrade, avoiding cost-tier mismatches.
21
+ - Lossless message snapshots preserve the provider-effective service tier when other metadata is not cloneable.
22
+
23
+ ## [0.18.11] - 2026-10-12
24
+
5
25
  ## [0.18.10] - 2026-10-12
6
26
 
7
27
  ## [0.18.9] - 2026-10-11
@@ -264,9 +264,11 @@ export interface AgentTelemetryConfig {
264
264
  * truncated to bounded snippets, not removed. `"full"` emits both summaries
265
265
  * and full OTEL message payloads.
266
266
  *
267
- * Defaults to the value of the `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`
268
- * env var (`true`/`1`/`yes` => `"summary"`, `"summary"` => `"summary"`,
269
- * `"full"` => `"full"`). Env-driven `"full"` capture emits a
267
+ * Defaults to the trusted value of
268
+ * `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`
269
+ * (`true`/`1`/`yes` => `"summary"`, `"summary"` => `"summary"`,
270
+ * `"full"` => `"full"`). A project `.env` cannot select this default.
271
+ * Env-driven `"full"` capture emits a
270
272
  * `full_content_capture_env_active` telemetry warning once per process.
271
273
  */
272
274
  readonly captureMessageContent?: TelemetryContentCapture;
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "type": "module",
3
3
  "name": "@gajae-code/agent-core",
4
- "version": "0.18.10",
4
+ "version": "0.18.12",
5
5
  "description": "General-purpose agent with transport abstraction, state management, and attachment support",
6
6
  "homepage": "https://gajae-code.com",
7
7
  "author": "Yeachan-Heo and Gajae Code Contributors",
@@ -32,9 +32,9 @@
32
32
  "fmt": "biome format --write ."
33
33
  },
34
34
  "dependencies": {
35
- "@gajae-code/ai": "0.18.10",
36
- "@gajae-code/natives": "0.18.10",
37
- "@gajae-code/utils": "0.18.10",
35
+ "@gajae-code/ai": "0.18.12",
36
+ "@gajae-code/natives": "0.18.12",
37
+ "@gajae-code/utils": "0.18.12",
38
38
  "@opentelemetry/api": "^1.9.0"
39
39
  },
40
40
  "devDependencies": {
package/src/agent-loop.ts CHANGED
@@ -405,6 +405,14 @@ export const ESCAPED_NONASCII_RECOVERY_PROMPT = escapedNonAsciiRecoveryPrompt;
405
405
 
406
406
  const MAX_ESCAPED_NONASCII_RESAMPLES = 2;
407
407
 
408
+ /**
409
+ * Recovery guidance kept byte-for-byte for every incomplete-argument rejection
410
+ * that carries a typed reason, and identical in meaning for the fallback where
411
+ * only the boolean was set.
412
+ */
413
+ const INCOMPLETE_ARGUMENTS_REISSUE_GUIDANCE =
414
+ "Re-issue the call with complete arguments, splitting the work into smaller steps if needed.";
415
+
408
416
  /** Whether any tool call in the turn carried `\uXXXX`-escaped arguments. */
409
417
  interface EscapedToolCallMetadata {
410
418
  guarded: boolean;
@@ -416,6 +424,13 @@ interface EscapedToolCallMetadata {
416
424
 
417
425
  const acceptedToolCallMetadata = new WeakMap<object, EscapedToolCallMetadata>();
418
426
 
427
+ /** The closed reason vocabulary the tool-call contract accepts. */
428
+ const INCOMPLETE_ARGUMENTS_REASONS = new Set(["truncated", "malformed", "conflicting", "ambiguous"]);
429
+
430
+ function isIncompleteArgumentsReason(value: unknown): value is "truncated" | "malformed" | "conflicting" | "ambiguous" {
431
+ return typeof value === "string" && INCOMPLETE_ARGUMENTS_REASONS.has(value);
432
+ }
433
+
419
434
  function escapedToolCallMetadata(
420
435
  block: Extract<AssistantMessage["content"][number], { type: "toolCall" }>,
421
436
  ): EscapedToolCallMetadata {
@@ -442,6 +457,15 @@ function escapedToolCallMetadata(
442
457
  (typeof incompleteReasonRead.value === "string" && SANITIZER_SENTINELS.has(incompleteReasonRead.value)) ||
443
458
  (evidenceRead.present && (evidenceRead.value === undefined || evidence === undefined || evidence.malformed));
444
459
  const incomplete = incompleteRead.ok && incompleteRead.value === true;
460
+ // A producer that typed the reason has already answered the question the
461
+ // metadata check is guessing at, so the declaration survives a failed or
462
+ // unverifiable evidence read instead of being overwritten with `malformed`.
463
+ // Re-labelling it would report a decode failure for a call the wire said was
464
+ // truncated (or conflicting, or ambiguously identified).
465
+ const declaredReason =
466
+ incompleteReasonRead.ok && typeof incompleteReasonRead.value === "string"
467
+ ? incompleteReasonRead.value
468
+ : undefined;
445
469
  return {
446
470
  guarded:
447
471
  malformedMetadata ||
@@ -451,7 +475,7 @@ function escapedToolCallMetadata(
451
475
  evidence,
452
476
  incompleteArguments: incomplete,
453
477
  incompleteArgumentsReason:
454
- !incompleteReasonRead.ok || malformedMetadata ? "malformed" : incompleteReasonRead.value,
478
+ declaredReason ?? (!incompleteReasonRead.ok || malformedMetadata ? "malformed" : incompleteReasonRead.value),
455
479
  };
456
480
  }
457
481
 
@@ -2019,6 +2043,7 @@ const LOSSLESS_SNAPSHOT_KEYS = [
2019
2043
  "provider",
2020
2044
  "model",
2021
2045
  "responseId",
2046
+ "serviceTier",
2022
2047
  "usage",
2023
2048
  "stopReason",
2024
2049
  "errorMessage",
@@ -3031,11 +3056,25 @@ class ManagedAttemptTransaction {
3031
3056
  metadata.incompleteArgumentsReason ?? detachedMetadata.incompleteArgumentsReason,
3032
3057
  };
3033
3058
  if (combinedMetadata.malformed) {
3059
+ // The snapshot could not corroborate this call's account of its
3060
+ // arguments, so it must not execute. A reason the producer already
3061
+ // declared is still the accurate cause and is preserved: rewriting
3062
+ // every incomplete call to `malformed` here is what left a typed
3063
+ // reason unreachable at dispatch, where the block's own field is
3064
+ // the last surviving declaration of why the arguments stopped. If
3065
+ // no reason was declared, leave it unknown rather than claiming that
3066
+ // the terminal arguments failed to decode.
3067
+ const declaredReason = managedProperty(snapshotBlock, "incompleteArgumentsReason");
3068
+ const markedReason = combinedMetadata.incompleteArgumentsReason ?? declaredReason;
3034
3069
  const marked = {
3035
3070
  ...snapshotBlock,
3036
3071
  incompleteArguments: true,
3037
- incompleteArgumentsReason: "malformed" as const,
3038
3072
  };
3073
+ if (isIncompleteArgumentsReason(markedReason)) {
3074
+ marked.incompleteArgumentsReason = markedReason;
3075
+ } else {
3076
+ delete marked.incompleteArgumentsReason;
3077
+ }
3039
3078
  snapshot.content[index] = marked;
3040
3079
  acceptedToolCallMetadata.set(marked, combinedMetadata);
3041
3080
  } else {
@@ -3413,7 +3452,7 @@ function buildAgentEndEvent(
3413
3452
  messages: AgentMessage[],
3414
3453
  telemetry: AgentTelemetry | undefined,
3415
3454
  stepCount: number,
3416
- stopReason: "completed" | "paused" = "completed",
3455
+ stopReason: "completed" | "paused" | "cancelled" = "completed",
3417
3456
  scope?: AttemptScope,
3418
3457
  ): Extract<AgentEvent, { type: "agent_end" }> {
3419
3458
  const base = { type: "agent_end" as const, messages, stopReason, ...(scope ? { scope } : {}) };
@@ -3812,6 +3851,9 @@ async function runLoopBody(
3812
3851
  ): Promise<void> {
3813
3852
  let firstTurn = true;
3814
3853
  let lastAttemptScope: AttemptScope | undefined;
3854
+ // A subscriber may abort reentrantly from the final assistant message_end;
3855
+ // do not relabel that already-accepted answer as cancelled.
3856
+ let lastAttemptAcceptedFinalAssistantMessage = false;
3815
3857
  // Check for steering messages at start (user may have typed while waiting)
3816
3858
  let pendingMessages: AgentMessage[] = (await config.getSteeringMessages?.()) || [];
3817
3859
  let harmonyRetryAttempt = 0;
@@ -3857,6 +3899,23 @@ async function runLoopBody(
3857
3899
  // the loop could run forever. Any turn that produces a non-malformed batch
3858
3900
  // resets the counter, so healthy runs are unaffected.
3859
3901
  let consecutiveMalformedTurns = 0;
3902
+ const finishCancelled = (scope?: AttemptScope): void => {
3903
+ publishAgentEnd(
3904
+ stream,
3905
+ config,
3906
+ buildAgentEndEvent(newMessages, telemetry, stepCounter.count, "cancelled", scope),
3907
+ scope,
3908
+ );
3909
+ stream.end(newMessages);
3910
+ };
3911
+ const commitMessages = (messages: AgentMessage[], scope?: AttemptScope): void => {
3912
+ for (const message of messages) {
3913
+ stream.push({ type: "message_start", message, scope });
3914
+ stream.push({ type: "message_end", message, scope });
3915
+ currentContext.messages.push(message);
3916
+ newMessages.push(message);
3917
+ }
3918
+ };
3860
3919
 
3861
3920
  // Outer loop: continues when queued follow-up messages arrive after agent would stop
3862
3921
  while (true) {
@@ -3864,6 +3923,7 @@ async function runLoopBody(
3864
3923
 
3865
3924
  // Inner loop: process tool calls and steering messages
3866
3925
  while (hasMoreToolCalls || pendingMessages.length > 0) {
3926
+ lastAttemptAcceptedFinalAssistantMessage = false;
3867
3927
  const scope =
3868
3928
  initialScope ?? (firstTurn ? config.initialScope : undefined) ?? config.attemptMinter?.mint("main");
3869
3929
  initialScope = undefined;
@@ -3889,12 +3949,7 @@ async function runLoopBody(
3889
3949
  // Commit queued user input outside the provisional assistant transaction so a
3890
3950
  // discarded managed attempt cannot lose it before its retry continuation.
3891
3951
  if (pendingMessages.length > 0) {
3892
- for (const message of pendingMessages) {
3893
- stream.push({ type: "message_start", message, scope: attemptScope });
3894
- stream.push({ type: "message_end", message, scope: attemptScope });
3895
- currentContext.messages.push(message);
3896
- newMessages.push(message);
3897
- }
3952
+ commitMessages(pendingMessages, attemptScope);
3898
3953
  pendingMessages = [];
3899
3954
  }
3900
3955
 
@@ -4411,13 +4466,17 @@ async function runLoopBody(
4411
4466
  });
4412
4467
  }
4413
4468
  stream.push({ type: "turn_end", message, toolResults, scope: attemptScope });
4414
- publishAgentEnd(
4415
- stream,
4416
- config,
4417
- buildAgentEndEvent(newMessages, telemetry, stepCounter.count, "completed", attemptScope),
4418
- attemptScope,
4419
- );
4420
- stream.end(newMessages);
4469
+ if (message.stopReason === "aborted") {
4470
+ finishCancelled(attemptScope);
4471
+ } else {
4472
+ publishAgentEnd(
4473
+ stream,
4474
+ config,
4475
+ buildAgentEndEvent(newMessages, telemetry, stepCounter.count, "completed", attemptScope),
4476
+ attemptScope,
4477
+ );
4478
+ stream.end(newMessages);
4479
+ }
4421
4480
  return;
4422
4481
  }
4423
4482
 
@@ -4427,6 +4486,7 @@ async function runLoopBody(
4427
4486
  (c): c is ToolCallContent => c.type === "toolCall" && !isProviderResolvedToolCall(c),
4428
4487
  );
4429
4488
  hasMoreToolCalls = toolCalls.length > 0;
4489
+ lastAttemptAcceptedFinalAssistantMessage = !hasMoreToolCalls;
4430
4490
 
4431
4491
  const toolResults: ToolResultMessage[] = [];
4432
4492
  let repeatedMalformedToolCall = false;
@@ -4578,8 +4638,18 @@ async function runLoopBody(
4578
4638
  }
4579
4639
 
4580
4640
  // Agent would stop here. Check for follow-up messages.
4581
- await config.onBeforeYield?.();
4582
- if (config.shouldPause?.()) {
4641
+ if (loopSignal.aborted && !lastAttemptAcceptedFinalAssistantMessage) {
4642
+ finishCancelled(lastAttemptScope);
4643
+ return;
4644
+ }
4645
+ if (!loopSignal.aborted) {
4646
+ await config.onBeforeYield?.();
4647
+ }
4648
+ if (loopSignal.aborted && !lastAttemptAcceptedFinalAssistantMessage) {
4649
+ finishCancelled(lastAttemptScope);
4650
+ return;
4651
+ }
4652
+ if (!loopSignal.aborted && config.shouldPause?.()) {
4583
4653
  publishAgentEnd(
4584
4654
  stream,
4585
4655
  config,
@@ -4591,9 +4661,22 @@ async function runLoopBody(
4591
4661
  }
4592
4662
  // Poll the consume-on-read recovery candidate before follow-ups so a real
4593
4663
  // user message can supersede it without letting the stale recovery resurface
4594
- // after that follow-up turn.
4664
+ // after that follow-up turn. An abort from an accepted final message_end must
4665
+ // also see a pending recovery before the run can report successful completion.
4595
4666
  const syntheticRecoveryMessage = await config.getSyntheticRecoveryMessage?.();
4667
+ if (loopSignal.aborted) {
4668
+ if (lastAttemptAcceptedFinalAssistantMessage && !syntheticRecoveryMessage) break;
4669
+ finishCancelled(lastAttemptScope);
4670
+ return;
4671
+ }
4596
4672
  const followUpMessages = (await config.getFollowUpMessages?.()) || [];
4673
+ if (loopSignal.aborted) {
4674
+ commitMessages(followUpMessages, lastAttemptScope);
4675
+ if (lastAttemptAcceptedFinalAssistantMessage && followUpMessages.length === 0 && !syntheticRecoveryMessage)
4676
+ break;
4677
+ finishCancelled(lastAttemptScope);
4678
+ return;
4679
+ }
4597
4680
  if (followUpMessages.length > 0) {
4598
4681
  // Set as pending so inner loop processes them
4599
4682
  pendingMessages = followUpMessages;
@@ -5504,8 +5587,18 @@ async function executeToolCalls(
5504
5587
  record.argumentValidationFailed = true;
5505
5588
  // The provider flagged this call's arguments as unsafe to execute.
5506
5589
  // The typed reason selects accurate recovery guidance; callers that
5507
- // only read the boolean still get a safe, actionable rejection.
5508
- const reason = metadata.incompleteArgumentsReason;
5590
+ // only read the boolean still get a safe, actionable rejection. An
5591
+ // attempt-scoped snapshot can carry metadata captured before the
5592
+ // terminal block declared its reason, so the block's own field is
5593
+ // consulted whenever the snapshot has none.
5594
+ const reason =
5595
+ metadata.incompleteArgumentsReason ?? managedProperty(toolCall, "incompleteArgumentsReason");
5596
+ // The output-token-limit sentence is an assertion about WHY the
5597
+ // arguments stopped, and only a length stop proves it. A call that
5598
+ // carries no typed reason (an older producer that set only the
5599
+ // boolean) is rejected honestly with the same actionable guidance
5600
+ // instead of blaming a limit the turn may never have hit.
5601
+ const lengthTruncation = reason === "truncated" || assistantMessage.stopReason === "length";
5509
5602
  const detail =
5510
5603
  reason === "malformed"
5511
5604
  ? `The terminal arguments for tool call "${toolCall.name}" did not decode to a valid JSON object. The arguments cannot be executed. Re-issue the call with valid, complete arguments.`
@@ -5513,7 +5606,9 @@ async function executeToolCalls(
5513
5606
  ? `The streamed and terminal arguments for tool call "${toolCall.name}" disagree. The arguments cannot be executed safely. Re-issue the call with consistent arguments.`
5514
5607
  : reason === "ambiguous"
5515
5608
  ? `The identity of tool call "${toolCall.name}" was ambiguous on the wire (duplicate call id or id/call_id collision), so its arguments cannot be safely attributed. Re-issue the call.`
5516
- : `Tool call "${toolCall.name}" was cut off before its arguments finished streaming (the response hit its output token limit). The partial arguments cannot be executed. Re-issue the call with complete arguments, splitting the work into smaller steps if needed.`;
5609
+ : lengthTruncation
5610
+ ? `Tool call "${toolCall.name}" was cut off before its arguments finished streaming (the response hit its output token limit). The partial arguments cannot be executed. ${INCOMPLETE_ARGUMENTS_REISSUE_GUIDANCE}`
5611
+ : `Tool call "${toolCall.name}" was incomplete before its arguments finished streaming. The partial arguments cannot be executed. ${INCOMPLETE_ARGUMENTS_REISSUE_GUIDANCE}`;
5517
5612
  throw markDesignedError(new Error(detail));
5518
5613
  }
5519
5614
  const displaySafeEscapedArguments =
@@ -983,6 +983,23 @@ export function boundConversationTextForSummary(
983
983
  return fitsBudget(bareMarker) ? bareMarker : "";
984
984
  }
985
985
 
986
+ /** Convert a stored summary the same way as conversation text, for the outbound prompt only. */
987
+ function outboundSummaryText(text: string, convert: ConvertToLlm): string {
988
+ const [message] = convert([
989
+ {
990
+ role: "user",
991
+ content: [{ type: "text", text }],
992
+ timestamp: 0,
993
+ },
994
+ ]);
995
+ if (!message) return text;
996
+ if (typeof message.content === "string") return message.content;
997
+ for (const block of message.content) {
998
+ if (block.type === "text") return block.text;
999
+ }
1000
+ return text;
1001
+ }
1002
+
986
1003
  export async function generateSummary(
987
1004
  currentMessages: AgentMessage[],
988
1005
  model: Model,
@@ -1009,10 +1026,12 @@ export async function generateSummary(
1009
1026
  const llmMessages = (options?.convertToLlm ?? convertToLlm)(currentMessages);
1010
1027
  const conversationText = boundConversationTextForSummary(serializeConversation(llmMessages), model, maxTokens);
1011
1028
 
1012
- // Build the prompt with conversation wrapped in tags
1029
+ // Build the prompt with conversation wrapped in tags.
1030
+ // The stored summary stays unchanged; only this outbound copy is converted.
1013
1031
  let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
1014
1032
  if (previousSummary) {
1015
- promptText += `<previous-summary>\n${previousSummary}\n</previous-summary>\n\n`;
1033
+ const outboundPreviousSummary = outboundSummaryText(previousSummary, options?.convertToLlm ?? convertToLlm);
1034
+ promptText += `<previous-summary>\n${outboundPreviousSummary}\n</previous-summary>\n\n`;
1016
1035
  }
1017
1036
  promptText += formatAdditionalContext(options?.extraContext);
1018
1037
  promptText += basePrompt;
package/src/telemetry.ts CHANGED
@@ -30,6 +30,8 @@ import {
30
30
  completeSimple,
31
31
  type Message,
32
32
  type Model,
33
+ modelSupportsServiceTier,
34
+ modelSupportsUltrafastTier,
33
35
  type ServiceTier,
34
36
  type SimpleStreamOptions,
35
37
  type StopReason,
@@ -37,6 +39,7 @@ import {
37
39
  type ToolChoice,
38
40
  type Usage,
39
41
  } from "@gajae-code/ai";
42
+ import { $credentialEnv } from "@gajae-code/utils";
40
43
  import { isDesignedError } from "@gajae-code/utils/error-classification";
41
44
  import { recordHandledError } from "@gajae-code/utils/postmortem";
42
45
  import {
@@ -345,9 +348,11 @@ export interface AgentTelemetryConfig {
345
348
  * truncated to bounded snippets, not removed. `"full"` emits both summaries
346
349
  * and full OTEL message payloads.
347
350
  *
348
- * Defaults to the value of the `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`
349
- * env var (`true`/`1`/`yes` => `"summary"`, `"summary"` => `"summary"`,
350
- * `"full"` => `"full"`). Env-driven `"full"` capture emits a
351
+ * Defaults to the trusted value of
352
+ * `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT`
353
+ * (`true`/`1`/`yes` => `"summary"`, `"summary"` => `"summary"`,
354
+ * `"full"` => `"full"`). A project `.env` cannot select this default.
355
+ * Env-driven `"full"` capture emits a
351
356
  * `full_content_capture_env_active` telemetry warning once per process.
352
357
  */
353
358
  readonly captureMessageContent?: TelemetryContentCapture;
@@ -457,9 +462,28 @@ export function resolveTelemetry(
457
462
 
458
463
  let contentCaptureEnvCache: ResolvedTelemetryContentCapture | undefined;
459
464
  let hasWarnedFullContentCaptureEnv = false;
465
+
466
+ /** `full` > `summary` > anything else. Matches the parser below. */
467
+ function contentCaptureRank(raw: string | undefined): 0 | 1 | 2 {
468
+ if (!raw) return 0;
469
+ const normalized = raw.trim().toLowerCase();
470
+ if (normalized === "full") return 2;
471
+ if (normalized === "summary" || normalized === "true" || normalized === "1" || normalized === "yes") return 1;
472
+ return 0;
473
+ }
474
+
460
475
  function readContentCaptureEnv(): ResolvedTelemetryContentCapture {
461
476
  if (contentCaptureEnvCache !== undefined) return contentCaptureEnvCache;
462
- const raw = process.env[CONTENT_CAPTURE_ENV];
477
+ // `$credentialEnv` ignores a project `.env`, pins an inherited value, and
478
+ // falls back to trusted files. A missing current key is off, so that
479
+ // fallback cannot restore a deleted value. A lower live rank is a downgrade.
480
+ // A present higher rank cannot raise the pin.
481
+ const trusted = $credentialEnv(CONTENT_CAPTURE_ENV);
482
+ const live = Bun.env[CONTENT_CAPTURE_ENV];
483
+ let raw: string | undefined;
484
+ if (live === undefined) raw = undefined;
485
+ else if (contentCaptureRank(live) < contentCaptureRank(trusted)) raw = live;
486
+ else raw = trusted;
463
487
  if (!raw) {
464
488
  contentCaptureEnvCache = "none";
465
489
  return "none";
@@ -753,7 +777,7 @@ export function startChatSpan(
753
777
  model,
754
778
  parent: options.parent,
755
779
  stepNumber: options.stepNumber,
756
- attributes: buildChatRequestAttributes(options.stepNumber, options.request, model.provider),
780
+ attributes: buildChatRequestAttributes(options.stepNumber, options.request, model),
757
781
  });
758
782
  if (span) {
759
783
  telemetry.collector.beginChat(span, {
@@ -787,7 +811,7 @@ export interface ChatRequestSnapshot {
787
811
  readonly messages?: readonly Message[];
788
812
  }
789
813
 
790
- function buildChatRequestAttributes(stepNumber: number, request: ChatRequestSnapshot, provider: string): Attributes {
814
+ function buildChatRequestAttributes(stepNumber: number, request: ChatRequestSnapshot, model: Model): Attributes {
791
815
  const attrs: Attributes = {
792
816
  [PiGenAIAttr.AgentStepNumber]: stepNumber,
793
817
  [GenAIAttr.OutputType]: "text",
@@ -803,7 +827,16 @@ function buildChatRequestAttributes(stepNumber: number, request: ChatRequestSnap
803
827
  if (request.stopSequences && request.stopSequences.length > 0) {
804
828
  attrs[GenAIAttr.RequestStopSequences] = [...request.stopSequences];
805
829
  }
806
- if (request.serviceTier && shouldSendServiceTier(request.serviceTier, provider)) {
830
+ if (
831
+ request.serviceTier &&
832
+ shouldSendServiceTier(
833
+ request.serviceTier,
834
+ model.provider,
835
+ modelSupportsServiceTier(model),
836
+ modelSupportsUltrafastTier(model),
837
+ ) &&
838
+ (request.serviceTier !== "ultrafast" || modelSupportsUltrafastTier(model))
839
+ ) {
807
840
  attrs[OpenAIAttr.RequestServiceTier] = request.serviceTier;
808
841
  }
809
842
  if (request.reasoningEffort) attrs[PiGenAIAttr.RequestReasoningEffort] = request.reasoningEffort;
@@ -1167,17 +1200,18 @@ export async function finishChatSpan(
1167
1200
  readonly baseUrl?: string;
1168
1201
  },
1169
1202
  ): Promise<void> {
1203
+ const serviceTier = resolveCompletedServiceTier(message, options.serviceTier);
1170
1204
  // Usage-only mode (spans disabled): the chat span is absent but usage/cost
1171
1205
  // hooks must still fire per contract C3. Cost estimation and the usage
1172
1206
  // event run against a non-recording placeholder span.
1173
1207
  if (!span) {
1174
1208
  if (!telemetry || telemetry.spansEnabled) return;
1175
1209
  const placeholder = trace.wrapSpanContext(INVALID_SPAN_CONTEXT);
1176
- const usageCost = applyCostEstimate(telemetry, placeholder, message, options.serviceTier, options.stepNumber);
1210
+ const usageCost = applyCostEstimate(telemetry, placeholder, message, serviceTier, options.stepNumber);
1177
1211
  await emitChatUsage(telemetry, placeholder, {
1178
1212
  model: message.model,
1179
1213
  provider: message.provider,
1180
- serviceTier: options.serviceTier,
1214
+ serviceTier,
1181
1215
  stepNumber: options.stepNumber,
1182
1216
  usage: message.usage,
1183
1217
  applied: usageCost,
@@ -1195,12 +1229,12 @@ export async function finishChatSpan(
1195
1229
  applyChatResponseAttributes(span, message);
1196
1230
  applyUsageAttributes(span, message.usage);
1197
1231
  applyGatewayAttributes(span, options.responseHeaders, options.baseUrl);
1198
- const cost = applyCostEstimate(telemetry, span, message, options.serviceTier, options.stepNumber);
1232
+ const cost = applyCostEstimate(telemetry, span, message, serviceTier, options.stepNumber);
1199
1233
  if (telemetry) {
1200
1234
  await emitChatUsage(telemetry, span, {
1201
1235
  model: message.model,
1202
1236
  provider: message.provider,
1203
- serviceTier: options.serviceTier,
1237
+ serviceTier,
1204
1238
  stepNumber: options.stepNumber,
1205
1239
  usage: message.usage,
1206
1240
  applied: cost,
@@ -1267,6 +1301,15 @@ export function failChatSpan(
1267
1301
  span.end();
1268
1302
  }
1269
1303
 
1304
+ function resolveCompletedServiceTier(
1305
+ message: AssistantMessage,
1306
+ requestedServiceTier: ServiceTier | undefined,
1307
+ ): ServiceTier | undefined {
1308
+ if (message.serviceTier !== undefined) return message.serviceTier;
1309
+ // Ultrafast is sent only for opted-in models, so an absent completion tier means it was not applied.
1310
+ return requestedServiceTier === "ultrafast" ? undefined : requestedServiceTier;
1311
+ }
1312
+
1270
1313
  function applyChatResponseAttributes(span: Span, message: AssistantMessage): void {
1271
1314
  span.setAttribute(GenAIAttr.ResponseModel, message.model);
1272
1315
  if (message.responseId) span.setAttribute(GenAIAttr.ResponseId, message.responseId);