runbios-sdk 0.2.22-rc.320 → 0.3.0-dev.322

package/CHANGELOG.md

@@ -0,0 +1,53 @@
+# Changelog
+
+`runbios-sdk` on npm. Versions are 0.x semver: the minor moves when a public
+surface is added, renamed or removed, the patch for a fix inside an unchanged
+surface. The SDK, the Python SDK (`runbios-sdk` on PyPI) and the MCP server
+(`runbios-mcp`) are released together, at one version.
+
+## 0.3.0
+
+Added
+
+- `training.create` and `training.preflight` take `startFrom.exportStep`
+  (`'best'`, the default, or `'final'`): which step of a job trained on top of
+  earlier versions its merged model is made of. It needs `exportMerged: true`;
+  without it, or with any other value, the SDK refuses before sending, as the
+  API does. `TrainingStartFrom` echoes `export_step`.
+- Conscious Loop: `inference` chat and messages calls take
+  `loop: { pipeline, onSampleId }`, which sends `X-Loop-Pipeline` and hands
+  back the sample id (`loop_sample_id`, or to `onSampleId` before the first
+  streamed chunk).
+- Conscious Loop: `loop.listPipelineAttempts` pages a pipeline's attempts;
+  `loop.listPipelines` takes `limit`, `q` and `status`; a pipeline's training
+  settings (`training`, three-state on a write) and `training_defaults`;
+  `loop.listModels` also lists the models not offered, with why;
+  `loop.getBenchmarkHistory` takes `pipeline`; `loop.listTraces` takes `q`.
+- Feedback (`signal`, `upvote`, `downvote`, `correct`, `gold`) is sent again
+  after `Retry-After` while its sample is still being recorded
+  (`409 SAMPLE_NOT_YET_RECORDED`), at most 6 tries; `ApiError` carries
+  `retryAfter`.
+
+Changed
+
+- `LoopSignalParams` takes a person's feedback only: `verdict` is
+  `LoopFeedbackVerdict` (`accepted`, `rejected`, `edited` or `gold`) and
+  `source` is `'human'`; `score` is gone. The API refuses a `scored` verdict,
+  any other source, and a `score` on any verdict with 400.
+- A merge-chain attempt replays about half of what it trains on (was a fifth),
+  and a version built on one that learned a deleted conversation says so in
+  its lineage.
+
+Removed
+
+- `Pipeline.champion_version` (use `live_version`) and
+  `PipelineVersion.is_champion` (use `is_live`), `benchmark_id` and
+  `benchmark_name` (use `listPipelineBenchmarks`), `benchmark_min_delta` (the
+  least rise a benchmark must show is the platform's; each measurement's
+  `reason` in `promotion` says whether it was met), `auto_promote` (use
+  `promotion.policy`), and `monthly_ceiling_cents` and `month_spent_cents`:
+  the API no longer sends them. `TrainingRuleMonthlyLimitRefusal` keeps
+  `error` and `resumes_at`. The monthly safety limit's figures are not fields
+  of any answer, and no sentence states them either: `next_reason` and the
+  `MONTHLY_LIMIT_REACHED` message say the limit holds the pipeline and when
+  the next attempt can start, with no amount.

package/dist/client.js

@@ -370,7 +370,7 @@ export class HttpClient {
     constructor(config) {
         // Default host stays api.runbios.ai for now; cutover to api.runbios.ai is
         // planned once its DNS exists.
-        this.baseUrl = (config.baseUrl || envBaseUrl() || 'https://api-staging.runbios.ai').replace(/\/+$/, '');
+        this.baseUrl = (config.baseUrl || envBaseUrl() || 'https://api-dev.runbios.ai').replace(/\/+$/, '');
         this.apiKey = config.apiKey ?? envApiKey();
         this.accessToken = config.accessToken;
         this.orgId = config.orgId;

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 = "0.2.22-rc.320";
+export declare const VERSION = "0.3.0-dev.322";
 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, 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, 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 = '0.2.22-rc.320';
+export const VERSION = '0.3.0-dev.322';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/inference.js

@@ -373,7 +373,7 @@ export class Inference {
         this.key = config.inferenceKey || envInferenceKey() || envApiKey();
         // Same default host as HttpClient (api.runbios.ai); api.runbios.ai cutover
         // is planned once its DNS exists — update both call sites together.
-        this.baseUrl = (config.baseUrl || envBaseUrl() || 'https://api-staging.runbios.ai').replace(/\/+$/, '');
+        this.baseUrl = (config.baseUrl || envBaseUrl() || 'https://api-dev.runbios.ai').replace(/\/+$/, '');
         this.timeout = config.timeout ?? 900_000;
         this._http = http;
     }

package/dist/resources/loop.d.ts

@@ -341,9 +341,8 @@ export declare class Loop {
      * `status` says whether it is the platform working (`running`), the member
      * who has to act (`needs_review`, `needs_funds`), or nothing at all until
      * somebody does (`paused`, with `paused_reason` saying why; `complete` at
-     * `max_versions`). `month_spent_cents` is the figure the
-     * monthly limit is enforced against: what runs started this month have cost
-     * at most, not an exact spend.
+     * `max_versions`). The monthly safety limit is the platform's and has no
+     * field: when it holds the pipeline, `next_reason` says so.
      */
     getPipeline(id: string): Promise<Pipeline>;
     /**

package/dist/resources/loop.js

@@ -565,9 +565,8 @@ export class Loop {
      * `status` says whether it is the platform working (`running`), the member
      * who has to act (`needs_review`, `needs_funds`), or nothing at all until
      * somebody does (`paused`, with `paused_reason` saying why; `complete` at
-     * `max_versions`). `month_spent_cents` is the figure the
-     * monthly limit is enforced against: what runs started this month have cost
-     * at most, not an exact spend.
+     * `max_versions`). The monthly safety limit is the platform's and has no
+     * field: when it holds the pipeline, `next_reason` says so.
      */
     async getPipeline(id) {
         const res = await this._http.fetchGet(`/api/loop/pipelines/${encodeURIComponent(id)}`);

package/dist/resources/training.js

@@ -9,8 +9,9 @@ function trainingIdempotencyKey(explicit) {
 }
 /**
  * `start_from` on the wire. The adapters stay in the order given (oldest
- * first: the order they are merged in), and an empty list is refused here
- * rather than sent, since the API refuses it too.
+ * first: the order they are merged in). An empty list, an `exportStep` other
+ * than 'best' or 'final', and an `exportStep` without `exportMerged: true`
+ * are refused here rather than sent, since the API refuses each of them too.
  */
 function startFromBody(sf) {
     if (!Array.isArray(sf.adapters) || sf.adapters.length === 0) {
@@ -21,6 +22,15 @@ function startFromBody(sf) {
     };
     if (sf.exportMerged !== undefined)
         body.export_merged = sf.exportMerged;
+    if (sf.exportStep !== undefined) {
+        if (sf.exportStep !== 'best' && sf.exportStep !== 'final') {
+            throw new Error("RunBiOS: startFrom.exportStep must be 'best' (the step validation picked) or 'final' (the last step trained)");
+        }
+        if (sf.exportMerged !== true) {
+            throw new Error('RunBiOS: startFrom.exportStep names the step the merged model is made of, so it needs exportMerged: true');
+        }
+        body.export_step = sf.exportStep;
+    }
     return body;
 }
 function buildTrainingRequest(params) {

package/dist/types.d.ts

@@ -8,13 +8,13 @@ export interface BiOSConfig {
     orgId?: string;
     /** Workspace ID. Optional when using API keys (resolved from the key). Can override for multi-workspace keys. */
     workspaceId?: string;
-    /** Base URL for the API. Falls back to the RUNBIOS_BASE_URL environment variable (legacy: BIOS_BASE_URL), then the canonical https://api-staging.runbios.ai hostname. */
+    /** Base URL for the API. Falls back to the RUNBIOS_BASE_URL environment variable (legacy: BIOS_BASE_URL), then the canonical https://api-dev.runbios.ai hostname. */
     baseUrl?: string;
     /** Request timeout in milliseconds. Defaults to 30000. */
     timeout?: number;
     /** Default per-deployment inference key. Can be overridden per inference call. */
     inferenceKey?: string;
-    /** Inference base URL. Defaults to baseUrl, then https://api-staging.runbios.ai. */
+    /** Inference base URL. Defaults to baseUrl, then https://api-dev.runbios.ai. */
     inferenceBaseUrl?: string;
     /** End-to-end inference timeout in milliseconds. Defaults to 15 minutes. */
     inferenceTimeout?: number;
@@ -565,6 +565,11 @@ export interface GPUChoice {
  * adapter merged into one model: a checkpoint of format `merged`, which is
  * what deploys (the job's own adapters are not deployable alone).
  *
+ * `exportStep` says which of this job's steps that merged model is made of:
+ * `'best'` (the default) is the step the validation loss picked, `'final'`
+ * the last step trained. It needs `exportMerged: true`; without it the SDK
+ * refuses before sending, as the API does (400 on `start_from.export_step`).
+ *
  * Refusals, all before anything is charged: 404 for an unknown, deleted or
  * other-workspace checkpoint; 422 START_FROM_BASE_MISMATCH,
  * START_FROM_CHAIN_BROKEN or START_FROM_LINK_UNREADABLE naming the `index`;
@@ -579,14 +584,22 @@ export interface TrainingStartFromParams {
         checkpointId: string;
     }[];
     exportMerged?: boolean;
+    exportStep?: TrainingExportStep;
 }
-/** `start_from` as the API echoes it on create, preflight and the job, with the digest of the chain's exact bytes. */
+/** Which step of a `start_from` job its merged model is made of: the step validation picked, or the last one trained. */
+export type TrainingExportStep = 'best' | 'final';
+/**
+ * `start_from` as the API echoes it on create, preflight and the job, with the
+ * digest of the chain's exact bytes. `export_step` is echoed only when the
+ * request named it; absent, the merged model is of the `'best'` step.
+ */
 export interface TrainingStartFrom {
     adapters: {
         job_id: string;
         checkpoint_id: string;
     }[];
     export_merged: boolean;
+    export_step?: TrainingExportStep;
     chain_digest?: string;
 }
 /** Parameters for creating a training job. */
@@ -2403,7 +2416,11 @@ export interface ChatCompletionUsage {
     total_tokens: number;
     prompt_tokens_details?: PromptTokensDetails;
 }
-/** Who gave a verdict. Feedback you send is `human` unless you say otherwise. */
+/**
+ * Who gave a verdict a sample carries. Feedback you send is always `human`:
+ * `verifier`, `judge` and `behavioural` are written by the platform's own
+ * scoring, and the feedback route refuses them (400).
+ */
 export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
 /**
  * The judgement recorded on an answer.
@@ -2414,8 +2431,11 @@ export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
  * trained on in place of the model's.
  * `gold` records the reference answer for the question, making no claim about
  * whether the model was right; it is trained on the same way.
+ * `scored` is a score the platform's own scoring wrote; it is read, never sent.
  */
 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';
 /** A function an assistant turn asked to invoke. `arguments` is JSON-encoded. */
 export interface LoopToolCall {
     id?: string;
@@ -2642,12 +2662,11 @@ export interface LoopSignal {
     created_at: string;
 }
 export interface LoopSignalParams {
-    verdict: LoopVerdict;
-    source?: LoopSource;
-    /** Required when verdict is `edited`. The answer the model should have given. */
+    verdict: LoopFeedbackVerdict;
+    /** 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. */
     correction?: string;
-    /** Required when verdict is `scored`: above zero trains the answer as it stands, otherwise it is left out. */
-    score?: number;
     /** 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;
     reason?: string;
@@ -3974,10 +3993,10 @@ export interface BenchmarkRetireParams {
  *   after a deleted conversation (`chain.rebuild`) is not held by the limit:
  *   Train now starts it, and each one that goes live is a version, so the
  *   pipeline can end past its limit.
- * - `MONTHLY_LIMIT_REACHED` (409): the most this month's attempts can have
- *   cost plus the most one more may cost would pass the pipeline's internal
- *   monthly limit. The body carries the figures:
- *   see {@link TrainingRuleMonthlyLimitRefusal}.
+ * - `MONTHLY_LIMIT_REACHED` (409): the pipeline has reached its monthly
+ *   safety limit. The body carries `resumes_at`, and `error.message` says
+ *   what holds the pipeline and when the next attempt can start, with no
+ *   amount: see {@link TrainingRuleMonthlyLimitRefusal}.
  * - `SCORING_CAPPED` (409): the key the comparison runs under is at its
  *   monthly cap, so what this attempt trained could not be scored. The body
  *   carries `resumes_at`, when the cap lifts (the next UTC month, or sooner
@@ -4063,34 +4082,22 @@ export type TrainingRuleRunRefusalCode = 'NOT_ENOUGH_SAMPLES' | 'PIPELINE_NEEDS_
 export type TrainingRuleSaveRefusalCode = 'PIPELINE_NAME_TAKEN' | 'OWN_TAG_TAKEN' | 'REVISION_MISMATCH' | 'RUN_ACTIVE' | 'NOT_A_SIMPLE_PIPELINE' | 'MODEL_NOT_TRAINABLE' | 'NO_GPU_FITS' | 'TRAINING_SETTING_OUT_OF_RANGE' | 'MODEL_REVISION_UNAVAILABLE' | 'NO_JUDGE_MODEL' | 'PROMOTION_POLICY_UNKNOWN' | 'PROMOTION_K_REQUIRED' | 'PROMOTION_K_UNREACHABLE' | 'PROMOTION_POLICY_NEEDS_BENCHMARK' | 'TRAINING_NOT_ENABLED' | 'METHOD_NOT_ENABLED' | 'CEILING_TOO_LOW' | 'MODELS_UNAVAILABLE' | 'GPU_OPTIONS_UNAVAILABLE' | 'AGENT_COULD_NOT_START';
 /**
  * The body of a manual run refused `409 MONTHLY_LIMIT_REACHED`, as
- * `ApiError.body` carries it.
- *
- * The test is the worst case, not the average: a run may start only if the
- * most the rule's runs this UTC calendar month can have cost
- * (`month_spent_cents`), plus the most the run about to start may spend
- * (`run_max_cents`), fits under the limit.
+ * `ApiError.body` carries it: the pipeline's monthly safety limit holds it.
+ * The limit is the platform's guard, set from the model's prices; the member
+ * never sets it, so its figures are neither fields of the answer nor in
+ * `message`, which says what holds the pipeline and when the next attempt can
+ * start.
  */
 export interface TrainingRuleMonthlyLimitRefusal {
     error: {
         code: 'MONTHLY_LIMIT_REACHED';
         message: string;
     };
-    /** What this rule's runs have cost at most this UTC calendar month; see {@link Pipeline.month_spent_cents}. */
-    month_spent_cents: number;
-    /** The rule's monthly limit. */
-    monthly_ceiling_cents: number;
     /**
-     * The most one attempt of this rule may spend: its training, its comparison
-     * machine and its judging. While no version is live that machine also
-     * answers as the base model, so a pipeline's own attempt counts one; a
-     * challenger's attempt also counts the base model's own machine beside it.
-     */
-    run_max_cents: number;
-    /**
-     * The first instant of the next UTC month, when the counted spend starts
-     * again from nothing. When `run_max_cents` alone exceeds the limit, no month
-     * will ever fit: saving the pipeline sets the limit again from its model's
-     * prices, and nothing else changes it; `message` says which.
+     * The first instant of the next UTC month, when the month's count starts
+     * again from nothing. When one attempt alone can cost more than the whole
+     * limit, no month will ever fit: saving the pipeline sets the limit again
+     * from its model's prices, and nothing else changes it; `message` says which.
      */
     resumes_at: string;
 }
@@ -4168,10 +4175,8 @@ export interface PipelineVersion {
     verdict: TrainingVerdict | null;
     decision: TrainingDecision | null;
     decided_at: string | null;
-    /** True for the version the app is served by today (the same as `is_champion`). */
-    is_live: boolean;
     /** True for the version the app is served by today. */
-    is_champion: boolean;
+    is_live: boolean;
     /** When it was first put live. */
     promoted_at: string | null;
     checkpoint_id: string | null;
@@ -4524,8 +4529,6 @@ export interface Pipeline {
     base_model_revision: string;
     /** The version serving now; null when what serves is not a version of this pipeline. */
     live_version: number | null;
-    /** The same number under its older name. */
-    champion_version: number | null;
     /**
      * Whether the deployment the live version answers from is up, as the
      * platform last checked it. Null when there is no live version, when what
@@ -4544,10 +4547,6 @@ export interface Pipeline {
      * does. Null when nothing has taken the name over.
      */
     serving_taken_over_by: string | null;
-    benchmark_id: string | null;
-    benchmark_name: string | null;
-    /** The least a benchmark's score has to rise by to count as an improvement under a policy that reads benchmarks. */
-    benchmark_min_delta: number;
     status: PipelineStatus;
     /**
      * The rule's own sentence about what it is waiting for or why it stopped,
@@ -4557,7 +4556,6 @@ export interface Pipeline {
     next_reason: string | null;
     last_checked_at: string | null;
     next_due_at: string | null;
-    auto_promote: boolean;
     /**
      * Why the platform stopped firing this rule, or null. `consent_invalid` is
      * a pipeline whose settings changed since it was last saved: saving it
@@ -4571,22 +4569,6 @@ export interface Pipeline {
      * trains on its own: an attempt would train and could not be compared.
      */
     scoring_capped_until: string | null;
-    /** The rule's monthly limit. Null means it has none. */
-    monthly_ceiling_cents: number | null;
-    /**
-     * What this rule's runs have cost at most, counted toward the monthly limit
-     * this UTC calendar month -- the figure the limit is enforced against. An
-     * upper bound, not an exact spend. A run counts toward the month it was
-     * created in, and a run still in progress also counts toward the current
-     * month, at the most it may cost (as `run_max_cents` counts it, a base
-     * model's machine included where the attempt has one of its own) or what it
-     * has been billed when that is more. A finished run counts what it was billed, and its comparison
-     * machines are billed at the most they could have cost (each machine's
-     * hourly cap for the time it was up, never more than its own amount). A run created in
-     * an earlier month that finishes in this one counts here only while it is
-     * still running.
-     */
-    month_spent_cents: number;
     active_run: PipelineActiveRun | null;
     /** How many versions it has made: every one is in `versions` from `getPipeline`, only the latest few and the live one on `listPipelines`. */
     versions_total: number;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.22-rc.320",
+  "version": "0.3.0-dev.322",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",
@@ -16,7 +16,8 @@
     "dist/**/*.d.ts",
     "!dist/**/*.map",
     "!dist/**/*.d.ts.map",
-    "README.md"
+    "README.md",
+    "CHANGELOG.md"
   ],
   "scripts": {
     "build": "tsc",