runbios-sdk 0.2.1-dev.123 → 0.2.1-dev.128

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.123";
+export declare const VERSION = "0.2.1-dev.128";
 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.123';
+export const VERSION = '0.2.1-dev.128';
 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, LoopStats, LoopJudge, LoopJudgeParams, LoopJudgeRun, LoopJudgeWork, LoopJudgeVerdict, LoopJudgeVerdictResult } 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, LoopStats, LoopJudge, LoopJudgeParams, LoopAgentCredential, LoopAgentStatus, LoopSampleRun, LoopSampleRunParams, LoopJudgeRun, LoopJudgeWork, LoopJudgeVerdict, LoopJudgeVerdictResult } 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.
@@ -401,4 +401,41 @@ export declare class Loop {
      * abandoned rather than completed.
      */
     postVerdicts(runId: string, verdicts: LoopJudgeVerdict[], finish?: boolean): Promise<LoopJudgeVerdictResult>;
+    /**
+     * Can an agent run here, is one running, and is it on for this workspace.
+     * `available` is about the environment; `online` about the worker;
+     * `credential` about this workspace.
+     */
+    agentStatus(): Promise<LoopAgentStatus>;
+    /**
+     * Turn the agent on: mints the workspace's managed serverless key. After
+     * this, automatic judges and sample runs make model calls billed to the
+     * workspace. Idempotent.
+     */
+    enableAgent(): Promise<LoopAgentCredential>;
+    /**
+     * Turn the agent off: revokes its key and sets every automatic judge back to
+     * manual. Open runs stop where they are and continue if it is turned back
+     * on. Nothing already scored or written is removed.
+     */
+    disableAgent(): Promise<{
+        revoked: boolean;
+        judges_paused: number;
+        message: string;
+    }>;
+    /**
+     * Ask the agent to write `n` alternative answers to each conversation in a
+     * slice with `model`, score each with `judge_id`, and store them as
+     * candidates. This is how preference pairs are made without a person
+     * writing each one: the curation pass pairs the best sample against the
+     * worst wherever the gap is real.
+     *
+     * Cost: up to `n` calls to write plus `n` to judge, per conversation, at the
+     * workspace's serverless rate; `selection.sample` caps the conversations
+     * (max 200) and `n` is capped at 8. Opening a run turns the agent on if it
+     * is off.
+     */
+    createSampleRun(params: LoopSampleRunParams): Promise<LoopSampleRun>;
+    listSampleRuns(): Promise<LoopSampleRun[]>;
+    getSampleRun(runId: string): Promise<LoopSampleRun>;
 }

package/dist/resources/loop.js

@@ -564,4 +564,59 @@ export class Loop {
     async postVerdicts(runId, verdicts, finish = false) {
         return this._http.fetchPost(`/api/loop/runs/${encodeURIComponent(runId)}/verdicts`, { verdicts, finish });
     }
+    // ── the agent ─────────────────────────────────────────────────────────
+    //
+    // The worker that calls a model on the workspace's behalf: it applies
+    // automatic judges and writes sample answers. It spends through a managed
+    // serverless key OF THIS WORKSPACE, so every call is billed exactly like
+    // one of your own, and turning it off revokes that key.
+    /**
+     * Can an agent run here, is one running, and is it on for this workspace.
+     * `available` is about the environment; `online` about the worker;
+     * `credential` about this workspace.
+     */
+    async agentStatus() {
+        return this._http.fetchGet('/api/loop/agent');
+    }
+    /**
+     * Turn the agent on: mints the workspace's managed serverless key. After
+     * this, automatic judges and sample runs make model calls billed to the
+     * workspace. Idempotent.
+     */
+    async enableAgent() {
+        const res = await this._http.fetchPost('/api/loop/agent', {});
+        return res.credential;
+    }
+    /**
+     * Turn the agent off: revokes its key and sets every automatic judge back to
+     * manual. Open runs stop where they are and continue if it is turned back
+     * on. Nothing already scored or written is removed.
+     */
+    async disableAgent() {
+        return this._http.fetchDelete('/api/loop/agent');
+    }
+    /**
+     * Ask the agent to write `n` alternative answers to each conversation in a
+     * slice with `model`, score each with `judge_id`, and store them as
+     * candidates. This is how preference pairs are made without a person
+     * writing each one: the curation pass pairs the best sample against the
+     * worst wherever the gap is real.
+     *
+     * Cost: up to `n` calls to write plus `n` to judge, per conversation, at the
+     * workspace's serverless rate; `selection.sample` caps the conversations
+     * (max 200) and `n` is capped at 8. Opening a run turns the agent on if it
+     * is off.
+     */
+    async createSampleRun(params) {
+        const res = await this._http.fetchPost('/api/loop/sample-runs', params);
+        return res.run;
+    }
+    async listSampleRuns() {
+        const res = await this._http.fetchGet('/api/loop/sample-runs');
+        return res.runs;
+    }
+    async getSampleRun(runId) {
+        const res = await this._http.fetchGet(`/api/loop/sample-runs/${encodeURIComponent(runId)}`);
+        return res.run;
+    }
 }

package/dist/resources/training.js

@@ -69,6 +69,10 @@ function buildTrainingRequest(params) {
         body.network_volume_id = params.networkVolumeId;
     if (params.datasetSampleLimits !== undefined)
         body.dataset_sample_limits = params.datasetSampleLimits;
+    if (params.datasetSampling !== undefined)
+        body.dataset_sampling = params.datasetSampling;
+    if (params.datasetSamplingStrategy !== undefined)
+        body.dataset_sampling_strategy = params.datasetSamplingStrategy;
     if (params.datasetMixing !== undefined)
         body.dataset_mixing = params.datasetMixing;
     if (params.mixing !== undefined)

package/dist/types.d.ts

@@ -548,6 +548,21 @@ export interface TrainingCreateParams {
      * large dataset without importing a trimmed copy.
      */
     datasetSampleLimits?: Record<string, number>;
+    /**
+     * Per-dataset sampling (key = a dataset ID from datasetIds): exactly one of
+     * `rows` or `percent` (of the dataset's usable rows for this method). Asking
+     * for more than the dataset holds uses every usable row and the run says
+     * so. Resolved to a row count identically by preflight and create, recorded
+     * in the mix manifest, and reproduced exactly on resume.
+     */
+    datasetSampling?: Record<string, DatasetSampling>;
+    /**
+     * How sampled rows are chosen: `first` (default) keeps the first N usable
+     * rows in file order -- what a curriculum wants; `random` draws a seeded
+     * uniform sample of the usable rows (kept in file order, seed = the mixing
+     * plan's seed, so a resume draws the same rows).
+     */
+    datasetSamplingStrategy?: 'first' | 'random';
     /** Training method. */
     method: TrainingMethod;
     /** Adapter type. */
@@ -1211,6 +1226,12 @@ export interface GPUPricingResponse {
     stale?: boolean;
     stale_message?: string;
 }
+/** One dataset's sampling request: exactly one of rows or percent. */
+export interface DatasetSampling {
+    rows?: number;
+    /** Percent (0-100) of the dataset's usable rows for the training method. */
+    percent?: number;
+}
 /** Parameters understood by the authenticated training GPU-options endpoint. */
 export interface GPUOptionsParams {
     modelId: string;
@@ -2439,6 +2460,8 @@ export interface LoopJudge {
     model: string | null;
     write_gold: boolean;
     enabled: boolean;
+    /** The platform's agent runs this judge, on the workspace's own serverless account. */
+    auto: boolean;
     created_by: string | null;
     created_at: string;
     updated_at: string;
@@ -2456,6 +2479,14 @@ export interface LoopJudgeParams {
      */
     write_gold?: boolean;
     enabled?: boolean;
+    /**
+     * Hand the running of this judge to the platform's agent: new conversations
+     * in its slice are scored as they arrive, one model call each, on THIS
+     * WORKSPACE'S OWN serverless account (the agent spends through a managed key
+     * of yours, "Conscious Loop" in your key list). Requires `model`. Off, you
+     * run the model yourself with startRun / takeWork / postVerdicts.
+     */
+    auto?: boolean;
 }
 /** One pass of one rubric over one slice. */
 export interface LoopJudgeRun {
@@ -2469,6 +2500,80 @@ export interface LoopJudgeRun {
     selected: number;
     scored: number;
     failed: number;
+    /** Who drains it: your own code ('caller') or the platform's agent ('platform'). */
+    runner: 'caller' | 'platform';
+    /** The agent's last complaint about this run, or null while it is working. */
+    last_error: string | null;
+    last_activity_at: string | null;
+    created_by: string | null;
+    created_at: string;
+    finished_at: string | null;
+}
+/** The workspace's managed serverless key the agent spends through. Never the secret. */
+export interface LoopAgentCredential {
+    workspace_id: string;
+    key_id: string;
+    key_prefix: string;
+    created_by: string | null;
+    created_at: string;
+    updated_at: string;
+    last_used_at: string | null;
+    last_error: string | null;
+    revoked_at: string | null;
+}
+export interface LoopAgentStatus {
+    /** An agent can exist in this environment at all. */
+    available: boolean;
+    /** Which piece is missing when it cannot. */
+    reason: string;
+    /** A worker has checked in within the last minute. */
+    online: boolean;
+    agent: {
+        seen_at: string;
+        passes: number;
+        judge_items: number;
+        samples: number;
+    } | null;
+    credential: LoopAgentCredential | null;
+    open_judge_runs: number;
+    open_sample_runs: number;
+    auto_judges: number;
+}
+export interface LoopSampleSelection extends LoopJudgeSelection {
+    /** Skip conversations that already have alternatives. */
+    only_unsampled?: boolean;
+}
+export interface LoopSampleRunParams {
+    /** The model that writes the alternatives. For distillation, the teacher. */
+    model: string;
+    /** Alternatives per conversation, 1-8. Default 4. */
+    n?: number;
+    /** 0-2. Default 0.8; 0 makes every sample the same answer. */
+    temperature?: number;
+    /** 16-8192. Default 1024. */
+    max_tokens?: number;
+    /** Score each sample with this judge as it is written. */
+    judge_id?: string;
+    selection?: LoopSampleSelection;
+}
+/** "Write N alternatives to each conversation in this slice, and score them." */
+export interface LoopSampleRun {
+    id: string;
+    workspace_id: string;
+    model: string;
+    n: number;
+    temperature: number;
+    max_tokens: number;
+    judge_id: string | null;
+    judge_name: string | null;
+    selection: LoopSampleSelection;
+    status: 'open' | 'done';
+    selected: number;
+    done: number;
+    failed: number;
+    samples: number;
+    last_error: string | null;
+    last_activity_at: string | null;
     created_by: string | null;
     created_at: string;
     finished_at: string | null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.123",
+  "version": "0.2.1-dev.128",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",