runbios-sdk 0.3.0-dev.322 → 0.3.0-dev.324

package/README.md

@@ -225,8 +225,15 @@ more header, `X-Loop-Pipeline: <tag>` (several tags: `pipeline: ['a', 'b']`).
 The gateway records the exchange -- what was asked, what the model answered and
 its tool calls -- as a sample of the pipelines those tags feed (a pipeline's
 `own_tag`), even where recording is off for that model: the header is your
-consent for that one exchange. It works for serverless models and dedicated
-deployments, including the pipeline's own `loop-<tag>` model.
+consent for that one exchange. With `chatCompletions` and `streamChatCompletions`
+it works for serverless models and dedicated deployments, including the
+pipeline's own `loop-<tag>` model. `messages` and `streamMessages` (the
+Anthropic shape) work for serverless models; a dedicated deployment, and so a
+pipeline's `loop-<tag>` model, serves the OpenAI shape and refuses
+`/v1/messages` with 400, so call it with `chatCompletions`. Only the tags of a
+pipeline that exists are recorded: a call naming a misspelt tag, or a deleted
+pipeline's, is not recorded under it, and `client.loop.stats()` lists such
+calls under `capture`.
 
 The gateway answers with the sample's id in `X-Loop-Sample-Id`. A non-streaming
 result carries it as `loop_sample_id` (non-enumerable: it is never part of the
@@ -260,10 +267,29 @@ Feedback sent before the sample is recorded is refused `409
 SAMPLE_NOT_YET_RECORDED`, and nothing is stored. `upvote`, `downvote`,
 `correct`, `gold` and `signal` send it again after the `Retry-After` seconds (2
 when none is given), at most 6 times in all, then throw that error. A stream
-that was cut short is never recorded, so its id stays unknown. If you send
+that was cut short is never recorded, so its id stays unknown (it is counted in
+`client.loop.stats()` under `capture`). If you send
 `X-Request-ID`, keep it unique per call: a repeated one is recorded once.
 `completions`, `embeddings` and `rerank` are not recorded and refuse the option.
 
+Many feedback items at once: `signalMany` sends each as `signal` does, the same
+retry included, with up to 8 in flight (`{ concurrency }`, at most 32), and
+returns every item in order with its `signal` or its `error`:
+
+```typescript
+const results = await client.loop.signalMany([
+  { trace_id: idA, verdict: 'good' },
+  { trace_id: idB, verdict: 'corrected', correction: 'Refunds take five working days.' },
+]);
+const failed = results.filter((r) => r.error);
+```
+
+Sending a sample yourself with `client.loop.capture` and already know the
+verdict? Pass `feedback: { verdict: 'good' | 'bad' | 'corrected', correction }`
+and it is written with the sample, in one transaction: both are stored or
+neither is. `good`, `bad` and `corrected` work in `signal` and `signalMany` too,
+stored as `accepted`, `rejected` and `edited`.
+
 #### Workspace serverless usage and limits (read-only)
 
 A workspace-bound platform key or hosted OAuth grant with `analytics:read` can

package/dist/client.d.ts

@@ -130,17 +130,25 @@ export type CapacityUnavailableError = GpuRejectionError;
  * (HTTP 403): TRAINING_COMING_SOON from POST /api/training/jobs,
  * DATASETS_COMING_SOON from the dataset creation endpoints (upload,
  * uploads/initiate, register-hf, integration import), and LOOP_COMING_SOON
- * from EVERY Conscious Loop route (`client.loop.*`) where the loop is not
- * offered — it is offered on the development environment only.
+ * from the Conscious Loop routes (`client.loop.*`) where the platform has the
+ * loop switched off — every one of them but the five that let a member stop
+ * paying (see {@link ComingSoonError}).
  */
 export declare const COMING_SOON_CODES: readonly ["TRAINING_COMING_SOON", "DATASETS_COMING_SOON", "LOOP_COMING_SOON"];
 /**
  * Raised when a pre-launch surface rejects a call: fine-tuning and datasets
  * launch soon, so training.create and dataset upload/import/register answer 403
  * with one of {@link COMING_SOON_CODES}, and their read and lifecycle methods
- * (list, get, status, stop, resume, delete, preview) are unaffected. The
- * Conscious Loop is offered on the development environment only; everywhere
- * else EVERY `client.loop.*` method, reads included, answers LOOP_COMING_SOON.
+ * (list, get, status, stop, resume, delete, preview) are unaffected. Where
+ * the platform has the Conscious Loop switched off, every `client.loop.*`
+ * method answers LOOP_COMING_SOON, reads included, except the five a member
+ * needs to stop paying, which answer as usual: listing pipelines
+ * (GET /api/loop/pipelines), reading one (GET /api/loop/pipelines/{id}),
+ * switching one off (PUT /api/loop/pipelines/{id} whose body is only
+ * `{"enabled": false}`, with at most `expected_revision` and `workspace_id`),
+ * deleting one, which stops its live version and removes its data
+ * (DELETE /api/loop/pipelines/{id}), and cancelling an attempt
+ * (POST /api/loop/training-runs/{id}/cancel).
  *
  * The service is the authority: the SDK never blocks client-side, so when the
  * gate lifts at launch every SDK version works again without an upgrade. This

package/dist/client.js

@@ -198,17 +198,25 @@ export const CapacityUnavailableError = GpuRejectionError;
  * (HTTP 403): TRAINING_COMING_SOON from POST /api/training/jobs,
  * DATASETS_COMING_SOON from the dataset creation endpoints (upload,
  * uploads/initiate, register-hf, integration import), and LOOP_COMING_SOON
- * from EVERY Conscious Loop route (`client.loop.*`) where the loop is not
- * offered — it is offered on the development environment only.
+ * from the Conscious Loop routes (`client.loop.*`) where the platform has the
+ * loop switched off — every one of them but the five that let a member stop
+ * paying (see {@link ComingSoonError}).
  */
 export const COMING_SOON_CODES = ['TRAINING_COMING_SOON', 'DATASETS_COMING_SOON', 'LOOP_COMING_SOON'];
 /**
  * Raised when a pre-launch surface rejects a call: fine-tuning and datasets
  * launch soon, so training.create and dataset upload/import/register answer 403
  * with one of {@link COMING_SOON_CODES}, and their read and lifecycle methods
- * (list, get, status, stop, resume, delete, preview) are unaffected. The
- * Conscious Loop is offered on the development environment only; everywhere
- * else EVERY `client.loop.*` method, reads included, answers LOOP_COMING_SOON.
+ * (list, get, status, stop, resume, delete, preview) are unaffected. Where
+ * the platform has the Conscious Loop switched off, every `client.loop.*`
+ * method answers LOOP_COMING_SOON, reads included, except the five a member
+ * needs to stop paying, which answer as usual: listing pipelines
+ * (GET /api/loop/pipelines), reading one (GET /api/loop/pipelines/{id}),
+ * switching one off (PUT /api/loop/pipelines/{id} whose body is only
+ * `{"enabled": false}`, with at most `expected_revision` and `workspace_id`),
+ * deleting one, which stops its live version and removes its data
+ * (DELETE /api/loop/pipelines/{id}), and cancelling an attempt
+ * (POST /api/loop/training-runs/{id}/cancel).
  *
  * The service is the authority: the SDK never blocks client-side, so when the
  * gate lifts at launch every SDK version works again without an upgrade. This

package/dist/index.d.ts

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export declare const VERSION = "0.3.0-dev.322";
+export declare const VERSION = "0.3.0-dev.324";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;
@@ -75,7 +75,7 @@ export { Training } from './resources/training.js';
 export { Wallet } from './resources/wallet.js';
 export { GPU, type GPURecommendation } from './resources/gpu.js';
 export { Inference, validateChatRequest, parseSSE, buildInferenceRequest, contextSizingBasis, type ContextSizingBasis, CONTEXT_DEFAULT_CEILING, CONTEXT_EDITABLE_FLOOR, type ChatMessage, type FunctionTool, type ChatCompletionParams, type ChatCompletionResponse, type ChatCompletionChunk, type InferenceVerbParams, type CompletionParams, type EmbeddingParams, type RerankParams, type AnthropicMessageParams, type AnthropicMessageResponse, type LoopCaptureOption, type LoopSampleIdOnResult, LOOP_PIPELINE_HEADER, LOOP_SAMPLE_ID_HEADER, type ServerlessUsageWindow, type ServerlessTimeseriesMetric, type ServerlessWorkspaceLimits, type ServerlessUsageEnvelope, type ServerlessUsageOverviewResponse, type ServerlessUsageRowsResponse, type ServerlessUsageTimeseriesResponse, type ServerlessUsageDailyResponse, type ServerlessSavingsTotals, type ServerlessSavingsPeriod, type ServerlessUsageSavingsResponse, } from './resources/inference.js';
-export type { BiOSConfig, PaginatedResponse, ApiErrorBody, AvailableGpuAlternative, CapacityMinimumRequirement, InferenceBookingAccepted, Model, ModelSearchParams, ModelSearchResponse, ModelDetailResponse, InferenceModel, InferenceModelListResponse, ModelConfig, Dataset, DatasetListParams, DatasetListResponse, DatasetUploadParams, DatasetPreview, DatasetPreviewParams, DatasetImportHFParams, DatasetRegisterHFParams, DatasetHubSearchParams, DatasetHubPreviewParams, DatasetValidation, DatasetFormatVariant, DatasetFormatSpec, DatasetFormatSpecs, DatasetStorageUsage, TrainingMethod, RLHFAlgorithm, AdapterType, TrainingJobStatus, TrainingStatusFilter, TrainingCreateParams, TrainingStartFromParams, TrainingStartFrom, TrainingExportStep, TrainingListParams, TrainingListResponse, TrainingJob, TrainingMetrics, TrainingResourceMetrics, TrainingDeviceMetrics, MetricPoint, MetricGraphConfig, TrainingCheckpoint, TrainingLogs, TrainingLogEntry, TrainingStopResponse, TrainingResumeResponse, CanonicalTrainingRequest, TrainingPreflightDataset, TrainingPreflightWarning, TrainingPreflightResponse, TrainingCapabilityChoice, TrainingConfigFieldCapability, TrainingCapabilities, TrainingFeature, GPUChoice, WalletBalance, Transaction, TransactionListResponse, TransactionListParams, GPUInfo, GPUPricingResponse, GPUOptionsParams, TrainingRecommendParams, TrainingAdvisorVerdict, TrainingAdvisorRecommendation, TrainingAdvisorUserShape, TrainingAdvisorJustification, TrainingAdvisorBasis, GPUOption, GPUOptionSuggestion, GPUOptionsResponse, InferenceStatus, InferenceCreateParams, InferenceUpdateParams, InferenceDeployment, InferenceDeploymentSummary, InferenceCreateResponse, InferenceListResponse, InferenceUpdateResponse, InferenceLifecycleResponse, InferenceDeleteResponse, InferenceGPUOptionMarket, InferenceGPUOption, InferenceGPUAlternative, InferenceGPUOptionsParams, InferenceGPUOptionsResponse, InferenceCanonicalRequest, InferencePreflightWarning, InferencePreflightResponse, AdapterCompatibility, AdapterCompatibilityResponse, AdapterCompatibilityParams, SupportedArchitecture, ArchitectureScope, SupportedArchitecturesResponse, ApiKey, ApiKeyScope, ApiKeyIntrospection, Organization, OrgMember, OrgInvite, Workspace, Integration, IntegrationCreateParams, StorageUsage, StorageObject, PromptTokensDetails, ChatCompletionUsage, ServingKind, TrainingRulePausedReason, TrainingTrigger, TrainingRecipe, TrainingRunState, TrainingVerdict, TrainingDecision, EvaluationStatus, EvaluationItemStatus, ItemErrorCode, EvaluationWinner, EvaluationItemWinnerFilter, ReportStatusFilter, ReportSort, EvaluationWarningCode, TrainingToolCall, TrainingMessage, TrainingRuleServing, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, TrainingRunListParams, TrainingRun, TrainingRunSummary, TrainingRunEvent, TrainingRunLinks, TrainingRunActions, EvaluationRef, EvaluationDecoding, EvaluationJudgeDimension, EvaluationDimensionScore, EvaluationWarning, EvaluationMargin, Evaluation, EvaluationItemJudge, EvaluationReferenceKind, EvaluationReference, EvaluationPairwiseOrder, EvaluationItemPairwise, EvaluationItemSide, EvaluationItem, EvaluationItemListParams, BenchmarkStatus, BenchmarkSourceKind, BenchmarkRunStatus, BenchmarkScoring, BenchmarkCheck, BenchmarkDimensionScore, Benchmark, BenchmarkItem, BenchmarkRun, BenchmarkRunItemSide, BenchmarkRunItem, BenchmarkRunItemListParams, BenchmarkHistoryPoint, BenchmarkSource, BenchmarkCreateParams, BenchmarkListParams, BenchmarkItemListParams, BenchmarkHistoryParams, BenchmarkRetireParams, InferenceAlias, InferenceAliasRequest, TrainingRuleDeleteResponse, TrainingRunListResponse, TrainingRunBenchmarkRun, TrainingRunResponse, TrainingRunActionResponse, EvaluationItemsResponse, BenchmarkRunItemsResponse, InferenceAliasListResponse, InferenceAliasResponse, InferenceAliasDeleteResponse, BenchmarkListResponse, BenchmarkItemsResponse, BenchmarkHistoryResponse, TrainingRuleRunRefusalCode, TrainingRuleSaveRefusalCode, TrainingRuleMonthlyLimitRefusal, PipelineStatus, PipelineActiveRun, PipelineVersion, PipelineAttempt, AttemptExplanation, PipelineAttemptListParams, PipelineAttemptListResponse, PipelineAttemptModelFilter, PipelineAttemptResultFilter, PipelineAttemptTrigger, Lineage, PipelineChain, PipelineChainRebuild, TrainingRunRollbackRefusalCode, LoopTraceRebuild, LoopTraceDeleteResponse, PipelineAttemptOutcome, PipelineLiveDeploymentStatus, PipelineLiveDeployment, LoopStorageWhere, LoopStorage, Pipeline, PipelineListParams, PipelineListStatus, PipelineListResponse, PipelineTrainingSettings, PipelineTrainingDefaults, PipelineTrainingRequest, PipelineResponse, PipelineTrainWhen, PipelineSchedule, PipelineWriteRequest, PipelineMutationResponse, PipelineImportParams, PipelineFloors, HeldBackFloors, LoopModel, LoopModelsResponse, LoopUnavailableModel, LoopUnavailableReasonCode, LoopTracePipeline, LoopTracePipelinesResponse, LoopMetricsHistoryParams, LoopMetricPoint, LoopSignalDay, LoopMetricsHistory, PromotionPolicyState, RuleBenchmark, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, CompareBenchmark, VersionBenchmarkScore, PromotionMeasurement, PromotionOutcome, VersionHoldout, VersionScores, VersionScoreDifference, VersionCompareResponse, VersionCompareParams, TrainingRunPromotion, VerdictCheck, VerdictCheckKey, VerdictChecksWithheld, TrainingRunCost, AttemptDirection, AttemptHoldout, AttemptCompareSide, AttemptSharedSample, AttemptSharedFigures, AttemptBenchmarkSide, AttemptCompareBenchmark, AttemptCompareItemSide, AttemptCompareItem, AttemptCompareResponse, AttemptCompareParams, } from './types.js';
+export type { BiOSConfig, PaginatedResponse, ApiErrorBody, AvailableGpuAlternative, CapacityMinimumRequirement, InferenceBookingAccepted, Model, ModelSearchParams, ModelSearchResponse, ModelDetailResponse, InferenceModel, InferenceModelListResponse, ModelConfig, Dataset, DatasetListParams, DatasetListResponse, DatasetUploadParams, DatasetPreview, DatasetPreviewParams, DatasetImportHFParams, DatasetRegisterHFParams, DatasetHubSearchParams, DatasetHubPreviewParams, DatasetValidation, DatasetFormatVariant, DatasetFormatSpec, DatasetFormatSpecs, DatasetStorageUsage, TrainingMethod, RLHFAlgorithm, AdapterType, TrainingJobStatus, TrainingStatusFilter, TrainingCreateParams, TrainingStartFromParams, TrainingStartFrom, TrainingExportStep, TrainingListParams, TrainingListResponse, TrainingJob, TrainingMetrics, TrainingResourceMetrics, TrainingDeviceMetrics, MetricPoint, MetricGraphConfig, TrainingCheckpoint, TrainingLogs, TrainingLogEntry, TrainingStopResponse, TrainingResumeResponse, CanonicalTrainingRequest, TrainingPreflightDataset, TrainingPreflightWarning, TrainingPreflightResponse, TrainingCapabilityChoice, TrainingConfigFieldCapability, TrainingCapabilities, TrainingFeature, GPUChoice, WalletBalance, Transaction, TransactionListResponse, TransactionListParams, GPUInfo, GPUPricingResponse, GPUOptionsParams, TrainingRecommendParams, TrainingAdvisorVerdict, TrainingAdvisorRecommendation, TrainingAdvisorUserShape, TrainingAdvisorJustification, TrainingAdvisorBasis, GPUOption, GPUOptionSuggestion, GPUOptionsResponse, InferenceStatus, InferenceCreateParams, InferenceUpdateParams, InferenceDeployment, InferenceDeploymentSummary, InferenceCreateResponse, InferenceListResponse, InferenceUpdateResponse, InferenceLifecycleResponse, InferenceDeleteResponse, InferenceGPUOptionMarket, InferenceGPUOption, InferenceGPUAlternative, InferenceGPUOptionsParams, InferenceGPUOptionsResponse, InferenceCanonicalRequest, InferencePreflightWarning, InferencePreflightResponse, AdapterCompatibility, AdapterCompatibilityResponse, AdapterCompatibilityParams, SupportedArchitecture, ArchitectureScope, SupportedArchitecturesResponse, ApiKey, ApiKeyScope, ApiKeyIntrospection, Organization, OrgMember, OrgInvite, Workspace, Integration, IntegrationCreateParams, StorageUsage, StorageObject, PromptTokensDetails, ChatCompletionUsage, ServingKind, TrainingRulePausedReason, TrainingTrigger, TrainingRecipe, TrainingRunState, TrainingVerdict, TrainingDecision, EvaluationStatus, EvaluationItemStatus, ItemErrorCode, EvaluationWinner, EvaluationItemWinnerFilter, ReportStatusFilter, ReportSort, EvaluationWarningCode, TrainingToolCall, TrainingMessage, TrainingRuleServing, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, TrainingRunListParams, TrainingRun, TrainingRunSummary, TrainingRunEvent, TrainingRunLinks, TrainingRunActions, EvaluationRef, EvaluationDecoding, EvaluationJudgeDimension, EvaluationDimensionScore, EvaluationWarning, EvaluationMargin, Evaluation, EvaluationItemJudge, EvaluationReferenceKind, EvaluationReference, EvaluationPairwiseOrder, EvaluationItemPairwise, EvaluationItemSide, EvaluationItem, EvaluationItemListParams, BenchmarkStatus, BenchmarkSourceKind, BenchmarkRunStatus, BenchmarkScoring, BenchmarkCheck, BenchmarkDimensionScore, Benchmark, BenchmarkItem, BenchmarkRun, BenchmarkRunItemSide, BenchmarkRunItem, BenchmarkRunItemListParams, BenchmarkHistoryPoint, BenchmarkSource, BenchmarkCreateParams, BenchmarkListParams, BenchmarkItemListParams, BenchmarkHistoryParams, BenchmarkRetireParams, InferenceAlias, InferenceAliasRequest, TrainingRuleDeleteResponse, TrainingRunListResponse, TrainingRunBenchmarkRun, TrainingRunResponse, TrainingRunActionResponse, EvaluationItemsResponse, BenchmarkRunItemsResponse, InferenceAliasListResponse, InferenceAliasResponse, InferenceAliasDeleteResponse, BenchmarkListResponse, BenchmarkItemsResponse, BenchmarkHistoryResponse, TrainingRuleRunRefusalCode, TrainingRuleSaveRefusalCode, TrainingRuleMonthlyLimitRefusal, PipelineStatus, PipelineActiveRun, PipelineVersion, PipelineAttempt, AttemptExplanation, PipelineAttemptListParams, PipelineAttemptListResponse, PipelineAttemptModelFilter, PipelineAttemptResultFilter, PipelineAttemptTrigger, Lineage, PipelineChain, PipelineChainRebuild, TrainingRunRollbackRefusalCode, LoopTraceRebuild, LoopTraceDeleteResponse, PipelineAttemptOutcome, PipelineLiveDeploymentStatus, PipelineLiveDeployment, LoopStorageWhere, LoopStorage, Pipeline, PipelineListParams, PipelineListStatus, PipelineListResponse, PipelineTrainingSettings, PipelineTrainingDefaults, PipelineTrainingRequest, PipelineResponse, PipelineTrainWhen, PipelineSchedule, PipelineWriteRequest, PipelineMutationResponse, PipelineImportParams, PipelineFloors, HeldBackFloors, LoopModel, LoopModelsResponse, LoopUnavailableModel, LoopUnavailableReasonCode, LoopTracePipeline, LoopTracePipelinesResponse, LoopMetricsHistoryParams, LoopMetricPoint, LoopSignalDay, LoopMetricsHistory, LoopCaptureFeedback, LoopFeedbackWord, LoopCaptureHealth, LoopCaptureMiss, LoopSignalManyItem, LoopSignalManyResult, PromotionPolicyState, RuleBenchmark, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, CompareBenchmark, VersionBenchmarkScore, PromotionMeasurement, PromotionOutcome, VersionHoldout, VersionScores, VersionScoreDifference, VersionCompareResponse, VersionCompareParams, TrainingRunPromotion, VerdictCheck, VerdictCheckKey, VerdictChecksWithheld, TrainingRunCost, AttemptDirection, AttemptHoldout, AttemptCompareSide, AttemptSharedSample, AttemptSharedFigures, AttemptBenchmarkSide, AttemptCompareBenchmark, AttemptCompareItemSide, AttemptCompareItem, AttemptCompareResponse, AttemptCompareParams, } from './types.js';
 /** The closed set of run states a training run never leaves. */
 export { TERMINAL_RUN_STATES } from './types.js';
 /** The most versions a pipeline may be set to make. */

package/dist/index.js

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export const VERSION = '0.3.0-dev.322';
+export const VERSION = '0.3.0-dev.324';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/inference.d.ts

@@ -66,8 +66,12 @@ export interface FunctionTool extends Record<string, unknown> {
  * refused 409 `SAMPLE_NOT_YET_RECORDED`; send it again after the
  * `Retry-After` seconds.
  *
- * Works for serverless models and dedicated deployments, on
- * `/v1/chat/completions` and `/v1/messages` only.
+ * Works on `/v1/chat/completions` for serverless models and dedicated
+ * deployments, a pipeline's own `loop-<tag>` model included, and on
+ * `/v1/messages` for serverless models: a dedicated deployment serves the
+ * OpenAI shape and refuses an Anthropic body with 400. Only the tags of a
+ * pipeline that exists are recorded; `client.loop.stats()` lists the calls
+ * that were not, under `capture`.
  */
 export interface LoopCaptureOption {
     /** One pipeline tag, or several; each a non-empty tag with no comma. */
@@ -412,7 +416,9 @@ export declare class Inference {
     /**
      * Anthropic Messages (`/v1/messages`). With `loop: { pipeline }` the
      * exchange is recorded into those pipelines and the result carries
-     * `loop_sample_id` (see {@link LoopCaptureOption}).
+     * `loop_sample_id` (see {@link LoopCaptureOption}). For serverless models:
+     * a dedicated deployment, a pipeline's `loop-<tag>` model included, serves
+     * the OpenAI shape and refuses this with 400; call it with `chatCompletions`.
      */
     messages(params: AnthropicMessageParams): Promise<AnthropicMessageResponse>;
     streamCompletions(params: CompletionParams): AsyncGenerator<Record<string, unknown>>;

package/dist/resources/inference.js

@@ -1027,7 +1027,9 @@ export class Inference {
     /**
      * Anthropic Messages (`/v1/messages`). With `loop: { pipeline }` the
      * exchange is recorded into those pipelines and the result carries
-     * `loop_sample_id` (see {@link LoopCaptureOption}).
+     * `loop_sample_id` (see {@link LoopCaptureOption}). For serverless models:
+     * a dedicated deployment, a pipeline's `loop-<tag>` model included, serves
+     * the OpenAI shape and refuses this with 400; call it with `chatCompletions`.
      */
     messages(params) {
         return this.sendVerb('/v1/messages', params, true);

package/dist/resources/loop.d.ts

@@ -1,5 +1,5 @@
 import { type HttpClient } from '../client.js';
-import type { LoopTrace, LoopTraceDeleteResponse, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopImportParams, LoopImportResult, LoopSignal, LoopSignalParams, LoopLabel, LoopLabelCount, LoopConfig, LoopConfigParams, LoopStats, LoopTracePipelinesResponse, LoopMetricsHistory, LoopMetricsHistoryParams, TrainingRuleDeleteResponse, TrainingRunListParams, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, Evaluation, EvaluationItemListParams, EvaluationItemsResponse, Benchmark, BenchmarkCreateParams, BenchmarkHistoryParams, BenchmarkHistoryResponse, BenchmarkItemListParams, BenchmarkItemsResponse, BenchmarkListParams, BenchmarkListResponse, BenchmarkRetireParams, BenchmarkRun, BenchmarkRunItemListParams, BenchmarkRunItemsResponse, Pipeline, PipelineAttemptListParams, PipelineAttemptListResponse, PipelineListParams, PipelineListResponse, PipelineWriteRequest, PipelineMutationResponse, LoopModelsResponse, PipelineImportParams, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, VersionCompareResponse, VersionCompareParams, AttemptCompareResponse, AttemptCompareParams } from '../types.js';
+import type { LoopTrace, LoopTraceDeleteResponse, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopImportParams, LoopImportResult, LoopSignal, LoopSignalParams, LoopSignalManyItem, LoopSignalManyResult, LoopLabel, LoopLabelCount, LoopConfig, LoopConfigParams, LoopStats, LoopTracePipelinesResponse, LoopMetricsHistory, LoopMetricsHistoryParams, TrainingRuleDeleteResponse, TrainingRunListParams, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, Evaluation, EvaluationItemListParams, EvaluationItemsResponse, Benchmark, BenchmarkCreateParams, BenchmarkHistoryParams, BenchmarkHistoryResponse, BenchmarkItemListParams, BenchmarkItemsResponse, BenchmarkListParams, BenchmarkListResponse, BenchmarkRetireParams, BenchmarkRun, BenchmarkRunItemListParams, BenchmarkRunItemsResponse, Pipeline, PipelineAttemptListParams, PipelineAttemptListResponse, PipelineListParams, PipelineListResponse, PipelineWriteRequest, PipelineMutationResponse, LoopModelsResponse, PipelineImportParams, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, VersionCompareResponse, VersionCompareParams, AttemptCompareResponse, AttemptCompareParams } from '../types.js';
 export declare class Loop {
     private readonly _http;
     /** @internal */
@@ -27,6 +27,10 @@ export declare class Loop {
      * When the source is not switched on this resolves with `captured: false`
      * and a reason instead of throwing -- capture must never be the thing that
      * breaks your application.
+     *
+     * Already know the answer was good, bad or wrong? Pass `feedback` and it is
+     * written with the exchange, so no second call is needed:
+     * `feedback: { verdict: 'corrected', correction: 'Thirty days from delivery.' }`.
      */
     capture(params: LoopCaptureParams): Promise<LoopCaptureResult>;
     /**
@@ -140,6 +144,27 @@ export declare class Loop {
      * that error. No other error is retried.
      */
     signal(traceId: string, params: LoopSignalParams): Promise<LoopSignal>;
+    /**
+     * Send many feedback items at once, several in flight together.
+     *
+     * Each item is `signal(item.trace_id, item)`: the same call, the same
+     * `SAMPLE_NOT_YET_RECORDED` retry. At most `concurrency` calls are in flight
+     * (8 by default, at most 32). It never throws for one item: every item
+     * comes back, in the order given, with its `signal` or the `error` it ended
+     * with, so one refused item costs no other.
+     *
+     * @example
+     * ```ts
+     * const results = await client.loop.signalMany([
+     *   { trace_id: idA, verdict: 'good' },
+     *   { trace_id: idB, verdict: 'corrected', correction: 'Refunds take five working days.' },
+     * ]);
+     * const failed = results.filter((r) => r.error);
+     * ```
+     */
+    signalMany(items: LoopSignalManyItem[], options?: {
+        concurrency?: number;
+    }): Promise<LoopSignalManyResult[]>;
     /**
      * Thumbs up: this answer was good. Shorthand for `signal(id, { verdict:
      * 'accepted' })` — there is one stored vocabulary behind both, so a thumbs-up
@@ -260,6 +285,13 @@ export declare class Loop {
      *
      * `ready.sft` is an upper bound: duplicates, the held-back set and
      * conversations too long to train on come out of it when an attempt trains.
+     *
+     * `capture` is what the one-header capture did not record over the last
+     * `capture.days` days: per tag, `no_pipeline` (no pipeline has the tag),
+     * `too_large` or `incomplete` (the answer could not be kept). Empty when
+     * every call was recorded. Rows are per tag: a call naming two tags is
+     * counted under each, so never add them up as calls. One pipeline's share
+     * is `capture` on its detail.
      */
     stats(): Promise<LoopStats>;
     /**

package/dist/resources/loop.js

@@ -64,6 +64,9 @@ const FEEDBACK_TRIES = 6;
 /** The wait when the 409 names none, and the longest wait honoured. */
 const FEEDBACK_RETRY_DEFAULT_SECONDS = 2;
 const FEEDBACK_RETRY_MAX_SECONDS = 30;
+/** How many feedback calls `signalMany` keeps in flight, by default and at most. */
+const SIGNAL_MANY_CONCURRENCY = 8;
+const SIGNAL_MANY_CONCURRENCY_MAX = 32;
 export class Loop {
     _http;
     /** @internal */
@@ -94,6 +97,10 @@ export class Loop {
      * When the source is not switched on this resolves with `captured: false`
      * and a reason instead of throwing -- capture must never be the thing that
      * breaks your application.
+     *
+     * Already know the answer was good, bad or wrong? Pass `feedback` and it is
+     * written with the exchange, so no second call is needed:
+     * `feedback: { verdict: 'corrected', correction: 'Thirty days from delivery.' }`.
      */
     async capture(params) {
         return this._http.fetchPost('/api/loop/traces', params);
@@ -273,6 +280,48 @@ export class Loop {
             }
         }
     }
+    /**
+     * Send many feedback items at once, several in flight together.
+     *
+     * Each item is `signal(item.trace_id, item)`: the same call, the same
+     * `SAMPLE_NOT_YET_RECORDED` retry. At most `concurrency` calls are in flight
+     * (8 by default, at most 32). It never throws for one item: every item
+     * comes back, in the order given, with its `signal` or the `error` it ended
+     * with, so one refused item costs no other.
+     *
+     * @example
+     * ```ts
+     * const results = await client.loop.signalMany([
+     *   { trace_id: idA, verdict: 'good' },
+     *   { trace_id: idB, verdict: 'corrected', correction: 'Refunds take five working days.' },
+     * ]);
+     * const failed = results.filter((r) => r.error);
+     * ```
+     */
+    async signalMany(items, options = {}) {
+        const asked = Math.trunc(options.concurrency ?? SIGNAL_MANY_CONCURRENCY);
+        const limit = Math.max(1, Math.min(Number.isFinite(asked) ? asked : SIGNAL_MANY_CONCURRENCY, SIGNAL_MANY_CONCURRENCY_MAX));
+        const out = new Array(items.length);
+        let next = 0;
+        const worker = async () => {
+            for (;;) {
+                const i = next++;
+                if (i >= items.length)
+                    return;
+                const { trace_id, ...params } = items[i];
+                try {
+                    if (!trace_id)
+                        throw new Error('each feedback item needs its trace_id');
+                    out[i] = { trace_id, signal: await this.signal(trace_id, params) };
+                }
+                catch (error) {
+                    out[i] = { trace_id, error };
+                }
+            }
+        };
+        await Promise.all(Array.from({ length: Math.min(limit, items.length) }, worker));
+        return out;
+    }
     /**
      * Thumbs up: this answer was good. Shorthand for `signal(id, { verdict:
      * 'accepted' })` — there is one stored vocabulary behind both, so a thumbs-up
@@ -425,6 +474,13 @@ export class Loop {
      *
      * `ready.sft` is an upper bound: duplicates, the held-back set and
      * conversations too long to train on come out of it when an attempt trains.
+     *
+     * `capture` is what the one-header capture did not record over the last
+     * `capture.days` days: per tag, `no_pipeline` (no pipeline has the tag),
+     * `too_large` or `incomplete` (the answer could not be kept). Empty when
+     * every call was recorded. Rows are per tag: a call naming two tags is
+     * counted under each, so never add them up as calls. One pipeline's share
+     * is `capture` on its detail.
      */
     async stats() {
         const res = await this._http.fetchGet('/api/loop/stats');

package/dist/types.d.ts

@@ -1740,6 +1740,14 @@ export interface InferenceDeploymentSummary {
      * unreachable or slow (a money read never blocks the list).
      */
     charged_cents?: number | null;
+    /**
+     * A Conscious Loop comparison machine that no model name points at: the
+     * loop's own machine for comparing an attempt, listed and billed like any
+     * other, but never a model to offer. Leave it out of a model picker; the
+     * pipeline's own name (`loop-<tag>`) is the one to call. Absent from a
+     * service from before it was sent.
+     */
+    loop_comparison?: boolean;
     /** History view (`include_terminated`) only: how the lifecycle closed, and when. */
     disposition?: string;
     deleted_at?: string | null;
@@ -2436,6 +2444,13 @@ export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
 export type LoopVerdict = 'accepted' | 'rejected' | 'edited' | 'scored' | 'gold';
 /** The verdicts feedback can send: a person's good, bad, correction or reference answer. */
 export type LoopFeedbackVerdict = 'accepted' | 'rejected' | 'edited' | 'gold';
+/**
+ * The plain words for the three common verdicts: `good` is `accepted`, `bad`
+ * is `rejected` and `corrected` is `edited` (with its `correction`). A capture's
+ * `feedback` takes them, and so does `signal`; the stored verdict, and the one
+ * a signal answers with, is the `LoopVerdict` word.
+ */
+export type LoopFeedbackWord = 'good' | 'bad' | 'corrected';
 /** A function an assistant turn asked to invoke. `arguments` is JSON-encoded. */
 export interface LoopToolCall {
     id?: string;
@@ -2483,6 +2498,19 @@ export interface LoopCaptureParams {
      * one conversation reaches another pipeline too. At most 32 in all.
      */
     labels?: string[];
+    /**
+     * The feedback you already have for this exchange, written with it, so no
+     * second call is needed. Refused (400) before anything is stored when it is
+     * unusable. Send `request_id` too: a retry under it records the feedback
+     * once.
+     */
+    feedback?: LoopCaptureFeedback;
+}
+/** Feedback sent with a capture: good, bad, or corrected with the right answer. */
+export interface LoopCaptureFeedback {
+    verdict: LoopFeedbackWord;
+    /** Required with `corrected`, and read only then: the answer that should have been given. */
+    correction?: string;
 }
 export interface LoopCaptureResult {
     captured: boolean;
@@ -2661,11 +2689,22 @@ export interface LoopSignal {
     metadata: Record<string, unknown> | null;
     created_at: string;
 }
+/** One feedback item for `loop.signalMany`: the sample's id and its feedback. */
+export interface LoopSignalManyItem extends LoopSignalParams {
+    trace_id: string;
+}
+/** What became of one `signalMany` item: its signal, or the error it ended with. */
+export interface LoopSignalManyResult {
+    trace_id: string;
+    signal?: LoopSignal;
+    error?: unknown;
+}
 export interface LoopSignalParams {
-    verdict: LoopFeedbackVerdict;
+    /** A `LoopFeedbackVerdict`, or its plain word (`good`, `bad`, `corrected`: see `LoopFeedbackWord`). */
+    verdict: LoopFeedbackVerdict | LoopFeedbackWord;
     /** Always `human` when sent: feedback is a person's. Other sources are refused 400. */
     source?: 'human';
-    /** Required when verdict is `edited` or `gold`. The answer the model should have given. */
+    /** Required when verdict is `edited`, `corrected` or `gold`. The answer the model should have given. */
     correction?: string;
     /** A value or fact the answer can be checked against. Kept with the verdict; a pipeline trains on `correction`, not on this. */
     ground_truth?: string;
@@ -2847,6 +2886,32 @@ export interface LoopStats {
     ready: {
         sft: number;
     };
+    /**
+     * The calls sent with `X-Loop-Pipeline` in the last `days` days (UTC) that
+     * were not recorded under a tag, most first. Empty when every one was.
+     */
+    capture: LoopCaptureHealth;
+}
+/** `LoopStats.capture` and `Pipeline.capture`: what the one-header capture did not record. */
+export interface LoopCaptureHealth {
+    /** The window, in days, today included. */
+    days: number;
+    not_recorded: LoopCaptureMiss[];
+}
+/**
+ * Calls under ONE tag that were not recorded, and why. A call naming two tags
+ * is counted under each, so rows are not added up as calls.
+ */
+export interface LoopCaptureMiss {
+    tag: string;
+    /**
+     * `no_pipeline`: no pipeline has the tag (misspelt, never made, or deleted).
+     * `too_large`: the answer was larger than the platform keeps for a sample.
+     * `incomplete`: the answer was cut short, or had no text and no tool call.
+     */
+    reason: 'no_pipeline' | 'too_large' | 'incomplete';
+    calls: number;
+    last_at: string;
 }
 /** One pipeline a conversation falls inside (GET /api/loop/traces/{id}/pipelines). */
 export interface LoopTracePipeline {
@@ -2977,9 +3042,10 @@ export type TrainingRulePausedReason = 'consent_invalid' | 'insufficient_funds'
 /** A peer was unreachable when the rule was saved, so it has not been validated yet. */
  | 'validation_pending'
 /**
- * Training is in beta and the rule's organisation does not hold the grant: a
- * run was refused with the platform's pre-launch gate. Nothing on the rule is
- * wrong; the platform team enables training, then a member saves the rule.
+ * The platform's training service refused an attempt with its pre-launch
+ * gate while the platform's services were being updated (the Conscious Loop
+ * trains for every organisation, so nothing else gives this). Nothing on the
+ * pipeline is wrong; switching it back on starts it again.
  */
  | 'training_not_enabled';
 /**
@@ -3212,11 +3278,20 @@ export interface TrainingRun {
     next_poll_at: string | null;
     last_reason: string | null;
     /**
-     * Why it failed or was set aside, as a code. Two belong to a pipeline that
+     * Why it failed or was set aside, as a code. Three belong to a pipeline that
      * builds on its live version: `CHAIN_MOVED` (it was built on a version that
      * was rolled back while it trained; it is superseded, no comparison was
-     * paid for, and its samples go to the next attempt) and `NO_MERGED_EXPORT`
-     * (its training finished without the merged model it is served as).
+     * paid for, and its samples go to the next attempt), `NO_MERGED_EXPORT`
+     * (its training finished without the merged model it is served as) and
+     * `TRAINING_SERVICE_BEHIND` (the platform's training service was being
+     * updated when it asked for its training, so it stopped before any
+     * training started; nothing was charged and the next attempt asks again).
+     * `STEP_STALLED:TRAINING_SERVICE_BEHIND` is the same for one that waited
+     * six hours for that update; the pipeline tries again on its own.
+     * `ENV_GATED:SWITCHED_OFF`: the Conscious Loop was switched off on the
+     * platform before its training started, so it stopped there; no GPU time
+     * was charged, and the pipeline trains again on its own once the loop is
+     * back.
      */
     error_code: string | null;
     last_error: string | null;
@@ -3295,6 +3370,12 @@ export interface TrainingRun {
      * machine bills beside it), so the attempt can stay in candidate_booking
      * past this time, by at most one more window. An attempt that ran out
      * gives the deadline it was held to on its expiry event (deadline_at).
+     * Null while the machine has never had a GPU: that wait is free and its
+     * `last_reason` says since when. A day after the machine was booked with no
+     * GPU it is stopped (never started, not charged) and the attempt waits for
+     * GPUs (`waiting_gpus`) keeping its trained model, then books the machine
+     * again; if no GPU is free by that wait's end too, the attempt ends
+     * (`expired`, `CANDIDATE_NOT_STARTED:NO_GPU`).
      */
     candidate_deadline_at: string | null;
     candidate_cleanup_at: string | null;
@@ -4065,11 +4146,14 @@ export type TrainingRuleRunRefusalCode = 'NOT_ENOUGH_SAMPLES' | 'PIPELINE_NEEDS_
  *   benchmark the pipeline does not have yet.
  *
  * The platform:
- * - `TRAINING_NOT_ENABLED` (409): training is in beta and this organisation
- *   does not hold the grant. Nothing in the request can lift it, so do not
- *   retry: ask the platform team to enable training for the organisation. A
- *   pipeline that meets it at fire time is paused `training_not_enabled`
- *   instead (see {@link TrainingRulePausedReason}).
+ * - `TRAINING_NOT_ENABLED` (409): the platform's training service would not
+ *   take the pipeline's training while the platform was being updated (the
+ *   Conscious Loop trains for every organisation; this answer only appears
+ *   while the platform's services are being updated). Nothing was saved and
+ *   nothing in the request is wrong: try again in a few minutes. A pipeline
+ *   that meets it at fire time is paused `training_not_enabled` instead (see
+ *   {@link TrainingRulePausedReason}); switching it back on later starts it
+ *   again.
  * - `METHOD_NOT_ENABLED` (422): the platform has switched off the training
  *   method pipelines train with.
  * - `CEILING_TOO_LOW` (422): the platform's own runaway limits cannot cover
@@ -4502,6 +4586,15 @@ export interface Pipeline {
     own_tag: string | null;
     /** A conversation carrying ANY of these tags is this pipeline's data. */
     tags: string[];
+    /**
+     * The calls sent with `X-Loop-Pipeline` in the last `days` days that were
+     * not recorded: under this pipeline's tags (an answer too large to keep, or
+     * cut short), and under any tag no pipeline had (`no_pipeline`: a misspelt
+     * tag, or a deleted pipeline's). One row per tag: a call naming two tags is
+     * counted under each, so rows are not added up as calls. Null on the list,
+     * and when it could not be read: only `pipelines.get` reads it.
+     */
+    capture: LoopCaptureHealth | null;
     /**
      * A second base model, trained from its own base on all samples after each
      * attempt (it never builds on this pipeline's versions, which are another

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.3.0-dev.322",
+  "version": "0.3.0-dev.324",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",