runbios-sdk 0.2.1-dev.141 → 0.2.1-dev.145

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 = "0.2.1-dev.141";
+export declare const VERSION = "0.2.1-dev.145";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

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 = '0.2.1-dev.141';
+export const VERSION = '0.2.1-dev.145';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -1,5 +1,5 @@
 import type { HttpClient } from '../client.js';
-import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopImportParams, LoopImportResult, LoopSignal, LoopSignalParams, LoopCandidate, LoopCandidateParams, LoopGrader, LoopGraderParams, LoopLabel, LoopLabelCount, LoopGradeResult, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopBuildRule, LoopBuildRuleParams, LoopBuildRuleUpdateParams, LoopStats, LoopJudge, LoopJudgeParams, LoopAgentCredential, LoopAgentStatus, LoopSampleRun, LoopSampleRunParams, LoopJudgeRun, LoopJudgeWork, LoopJudgeVerdict, LoopJudgeVerdictResult, TrainingRule, TrainingRulePreflight, TrainingRulePreflightRequest, TrainingRuleCreateRequest, TrainingRuleUpdateRequest, TrainingRuleConsentRequest, TrainingRuleListParams, TrainingRuleListResponse, TrainingRuleResponse, TrainingRuleMutationResponse, TrainingRuleDeleteResponse, TrainingRun, TrainingRunListParams, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, Evaluation, EvaluationItemListParams, EvaluationItemsResponse, JudgeAgreement, JudgeAgreementParams, AgentSettings, AgentSettingsRequest } from '../types.js';
+import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopImportParams, LoopImportResult, LoopSignal, LoopSignalParams, LoopCandidate, LoopCandidateParams, LoopGrader, LoopGraderParams, LoopLabel, LoopLabelCount, LoopGradeResult, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopBuildRule, LoopBuildRuleParams, LoopBuildRuleUpdateParams, LoopStats, LoopJudge, LoopJudgeParams, LoopAgentCredential, LoopAgentStatus, LoopSampleRun, LoopSampleRunParams, LoopJudgeRun, LoopJudgeRunItems, LoopJudgeRunItemStatus, LoopJudgeWork, LoopJudgeVerdict, LoopJudgeVerdictResult, TrainingRule, TrainingRulePreflight, TrainingRulePreflightRequest, TrainingRuleCreateRequest, TrainingRuleUpdateRequest, TrainingRuleConsentRequest, TrainingRuleListParams, TrainingRuleListResponse, TrainingRuleResponse, TrainingRuleMutationResponse, TrainingRuleDeleteResponse, TrainingRun, TrainingRunListParams, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, Evaluation, EvaluationItemListParams, EvaluationItemsResponse, JudgeAgreement, JudgeAgreementParams, AgentSettings, AgentSettingsRequest } from '../types.js';
 /**
  * The Conscious Loop -- capture what your model was asked and answered, record
  * whether it was right, and turn those judgements into training data.
@@ -398,6 +398,25 @@ export declare class Loop {
      * nothing: ask again and the same items come back.
      */
     takeWork(runId: string, limit?: number): Promise<LoopJudgeWork>;
+    /**
+     * What happened to each conversation in a run, and why.
+     *
+     * `takeWork` hands out what is still PENDING, so a finished run answers it
+     * with an empty list. This answers with every item and the outcome on it:
+     * `status`, `scored_at`, and `error` — the reason the caller gave for an
+     * item it could not score, which is where a wrong model slug or a refused
+     * key actually shows up. A run that ends "scored 0, failed 3" is otherwise a
+     * number with no detail behind it.
+     *
+     * Pass `status: 'failed'` for the usual question. At most 500 items come
+     * back at a time; when `has_more` is true, call again with the `next_offset`
+     * from the reply.
+     */
+    listRunItems(runId: string, opts?: {
+        status?: LoopJudgeRunItemStatus;
+        limit?: number;
+        offset?: number;
+    }): Promise<LoopJudgeRunItems>;
     /**
      * Hand the scores back. A verdict for a conversation outside this run, or
      * scoring something the rubric never asked for, is refused and reported in
@@ -420,6 +439,12 @@ export declare class Loop {
      * automatic judges and sample runs make model calls billed to the
      * workspace. Idempotent.
      *
+     * A judge whose model the serving gateway will not route is NOT resumed:
+     * making it automatic would buy a run that fails every conversation. It
+     * stays paused, `judges_still_paused` counts those, and each one carries
+     * `auto_pause_reason` saying so. Point it at a model that is served and
+     * turn the agent on again.
+     *
      * `monthly_spend_cap_cents` caps what that key may spend on model calls in
      * a calendar month. Omitted = no cap on a fresh key, and an existing cap is
      * left as it is; sent while the agent is already on, it moves the cap on

package/dist/resources/loop.js

@@ -559,6 +559,31 @@ export class Loop {
     async takeWork(runId, limit = 20) {
         return this._http.fetchGet(`/api/loop/runs/${encodeURIComponent(runId)}/work?limit=${limit}`);
     }
+    /**
+     * What happened to each conversation in a run, and why.
+     *
+     * `takeWork` hands out what is still PENDING, so a finished run answers it
+     * with an empty list. This answers with every item and the outcome on it:
+     * `status`, `scored_at`, and `error` — the reason the caller gave for an
+     * item it could not score, which is where a wrong model slug or a refused
+     * key actually shows up. A run that ends "scored 0, failed 3" is otherwise a
+     * number with no detail behind it.
+     *
+     * Pass `status: 'failed'` for the usual question. At most 500 items come
+     * back at a time; when `has_more` is true, call again with the `next_offset`
+     * from the reply.
+     */
+    async listRunItems(runId, opts = {}) {
+        const qs = new URLSearchParams();
+        if (opts.status)
+            qs.set('status', opts.status);
+        if (opts.limit != null)
+            qs.set('limit', String(opts.limit));
+        if (opts.offset)
+            qs.set('offset', String(opts.offset));
+        const suffix = qs.toString() ? `?${qs.toString()}` : '';
+        return this._http.fetchGet(`/api/loop/runs/${encodeURIComponent(runId)}/items${suffix}`);
+    }
     /**
      * Hand the scores back. A verdict for a conversation outside this run, or
      * scoring something the rubric never asked for, is refused and reported in
@@ -591,6 +616,12 @@ export class Loop {
      * automatic judges and sample runs make model calls billed to the
      * workspace. Idempotent.
      *
+     * A judge whose model the serving gateway will not route is NOT resumed:
+     * making it automatic would buy a run that fails every conversation. It
+     * stays paused, `judges_still_paused` counts those, and each one carries
+     * `auto_pause_reason` saying so. Point it at a model that is served and
+     * turn the agent on again.
+     *
      * `monthly_spend_cap_cents` caps what that key may spend on model calls in
      * a calendar month. Omitted = no cap on a fresh key, and an existing cap is
      * left as it is; sent while the agent is already on, it moves the cap on

package/dist/types.d.ts

@@ -2350,7 +2350,11 @@ export interface LoopDatasetCreateParams {
     to?: string;
     /** Narrow to one slice. A label with children selects them too. */
     label?: string;
-    /** Take this many at random from what the filters matched. */
+    /**
+     * Take this many of what the filters matched. Which ones is arbitrary but
+     * REPEATABLE: the same conversations always give the same slice, so two
+     * builds of one spec describe the same set.
+     */
     sample?: number;
     /** Holds back your most recent work, not a random slice. 0-50. */
     holdout_percent?: number;
@@ -2486,6 +2490,19 @@ export interface LoopJudge {
     enabled: boolean;
     /** The platform's agent runs this judge, on the workspace's own serverless account. */
     auto: boolean;
+    /**
+     * Was automatic when the agent was turned off. Turning the agent back on
+     * restores exactly these judges.
+     */
+    auto_paused: boolean;
+    /**
+     * Why the last turn-on did NOT restore this judge. Null is the ordinary
+     * state. Set when the resume declined to make it automatic because the model
+     * it names is not one the serving gateway will route — handing it back would
+     * buy a run that fails every conversation. It stays paused, so pointing it at
+     * a model that routes and turning the agent on again brings it back.
+     */
+    auto_pause_reason: string | null;
     created_by: string | null;
     created_at: string;
     updated_at: string;
@@ -2517,7 +2534,13 @@ export interface LoopJudgeRun {
     id: string;
     judge_id: string;
     judge_name?: string;
-    status: 'open' | 'done';
+    /**
+     * `abandoned` is a run you opened and never drained: nothing was scored on
+     * it for a week and items were still waiting. Nothing is deleted and posting
+     * verdicts to it still works and still closes it as done. It exists so that
+     * `open` keeps meaning "somebody is working on this".
+     */
+    status: 'open' | 'done' | 'abandoned';
     instructions: string;
     dimensions: LoopJudgeDimension[];
     model: string | null;
@@ -2630,6 +2653,27 @@ export interface LoopJudgeWork {
     items: LoopJudgeWorkItem[];
     remaining: number;
 }
+export type LoopJudgeRunItemStatus = 'pending' | 'scored' | 'failed';
+/** What happened to one conversation in one run. */
+export interface LoopJudgeRunItem {
+    trace_id: string;
+    status: LoopJudgeRunItemStatus;
+    /**
+     * Why this one could not be scored, as the caller reported it. This is where
+     * a wrong model name or a refused key shows up; null for anything that is
+     * not failed.
+     */
+    error: string | null;
+    scored_at: string | null;
+}
+export interface LoopJudgeRunItems {
+    run: LoopJudgeRun;
+    items: LoopJudgeRunItem[];
+    /** True when the page ended before the run did. */
+    has_more: boolean;
+    /** Where to carry on from. Present only when `has_more`. */
+    next_offset?: number;
+}
 /** One scored conversation going back. */
 export interface LoopJudgeVerdict {
     trace_id: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.141",
+  "version": "0.2.1-dev.145",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",