runbios-sdk 0.2.1-dev.107 → 0.2.1-dev.109

package/dist/index.d.ts

@@ -35,7 +35,7 @@ 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.107";
+export declare const VERSION = "0.2.1-dev.109";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

package/dist/index.js

@@ -35,7 +35,7 @@ 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.107';
+export const VERSION = '0.2.1-dev.109';
 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, LoopSignal, LoopSignalParams, LoopCandidate, LoopCandidateParams, LoopGrader, LoopGraderParams, LoopGradeResult, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopStats } from '../types.js';
+import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopSignal, LoopSignalParams, LoopCandidate, LoopCandidateParams, LoopGrader, LoopGraderParams, LoopLabel, LoopLabelCount, LoopGradeResult, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopStats } 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.
@@ -88,6 +88,42 @@ export declare class Loop {
      * behavioural hint.
      */
     signal(traceId: string, params: LoopSignalParams): Promise<LoopSignal>;
+    /**
+     * Thumbs up: this answer was good. Shorthand for `signal(id, { verdict:
+     * 'accepted' })` — there is one stored vocabulary behind both, so a thumbs-up
+     * and an "accepted" can never disagree in your corpus.
+     */
+    upvote(traceId: string, reason?: string): Promise<LoopSignal>;
+    /**
+     * Thumbs down: this answer was bad, and you have nothing better to offer.
+     *
+     * On its own this REMOVES the answer from training rather than teaching
+     * anything — there is no better version to learn. If you know what it should
+     * have said, `correct()` is worth far more.
+     */
+    downvote(traceId: string, reason?: string): Promise<LoopSignal>;
+    /**
+     * The model was wrong; here is what it should have said.
+     *
+     * The most valuable feedback there is, because it produces BOTH halves of a
+     * preference pair from one action: the model learns your answer and learns to
+     * avoid its own.
+     */
+    correct(traceId: string, answer: string, reason?: string): Promise<LoopSignal>;
+    /**
+     * Record the GOLD answer for this question — the reference, regardless of
+     * what the model happened to say.
+     *
+     * Different from `correct()` in a way that matters: a correction asserts the
+     * model was wrong, so the original becomes the rejected half of a pair. A gold
+     * answer asserts nothing about the model, so if it MATCHES what was said, no
+     * preference pair is invented — but SFT still learns it.
+     *
+     * It is the most reusable thing you can record: it trains SFT, forms a DPO
+     * pair when it differs from the answer, and stands in as the GRPO reference
+     * when you have not supplied a separate ground truth.
+     */
+    gold(traceId: string, answer: string, reason?: string): Promise<LoopSignal>;
     /** Every verdict on one conversation, oldest first. */
     listSignals(traceId: string): Promise<LoopSignal[]>;
     /**
@@ -121,6 +157,27 @@ export declare class Loop {
     addCandidate(traceId: string, params: LoopCandidateParams): Promise<LoopCandidate>;
     /** Every alternative answer recorded for one prompt, oldest first. */
     listCandidates(traceId: string): Promise<LoopCandidate[]>;
+    /**
+     * Label a conversation, so it can be selected later.
+     *
+     * An average over everything is the least useful thing to train on. A model
+     * weak at refunds is fixed with refund examples, and you can only select
+     * those if you said so when the conversation arrived.
+     *
+     * A label may name a `parent`, which is what makes a MASTER GROUP: labelling
+     * something `refunds` with parent `billing` makes it selectable as either,
+     * and selecting `billing` later gathers every child without you maintaining
+     * a list of them.
+     *
+     * Labels are lowercased and trimmed, so `Refunds` and `refunds` are one label.
+     */
+    label(traceId: string, labels: string[], parent?: string): Promise<LoopLabel[]>;
+    unlabel(traceId: string, label: string): Promise<{
+        deleted: boolean;
+        label: string;
+    }>;
+    /** Every label in the workspace, with how many conversations carry it. */
+    listLabels(): Promise<LoopLabelCount[]>;
     /**
      * Build a training set from the feedback recorded so far.
      *

package/dist/resources/loop.js

@@ -116,6 +116,50 @@ export class Loop {
         const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/signals`, params);
         return res.signal;
     }
+    /**
+     * Thumbs up: this answer was good. Shorthand for `signal(id, { verdict:
+     * 'accepted' })` — there is one stored vocabulary behind both, so a thumbs-up
+     * and an "accepted" can never disagree in your corpus.
+     */
+    async upvote(traceId, reason) {
+        return this.signal(traceId, { verdict: 'accepted', reason });
+    }
+    /**
+     * Thumbs down: this answer was bad, and you have nothing better to offer.
+     *
+     * On its own this REMOVES the answer from training rather than teaching
+     * anything — there is no better version to learn. If you know what it should
+     * have said, `correct()` is worth far more.
+     */
+    async downvote(traceId, reason) {
+        return this.signal(traceId, { verdict: 'rejected', reason });
+    }
+    /**
+     * The model was wrong; here is what it should have said.
+     *
+     * The most valuable feedback there is, because it produces BOTH halves of a
+     * preference pair from one action: the model learns your answer and learns to
+     * avoid its own.
+     */
+    async correct(traceId, answer, reason) {
+        return this.signal(traceId, { verdict: 'edited', correction: answer, reason });
+    }
+    /**
+     * Record the GOLD answer for this question — the reference, regardless of
+     * what the model happened to say.
+     *
+     * Different from `correct()` in a way that matters: a correction asserts the
+     * model was wrong, so the original becomes the rejected half of a pair. A gold
+     * answer asserts nothing about the model, so if it MATCHES what was said, no
+     * preference pair is invented — but SFT still learns it.
+     *
+     * It is the most reusable thing you can record: it trains SFT, forms a DPO
+     * pair when it differs from the answer, and stands in as the GRPO reference
+     * when you have not supplied a separate ground truth.
+     */
+    async gold(traceId, answer, reason) {
+        return this.signal(traceId, { verdict: 'gold', correction: answer, reason });
+    }
     /** Every verdict on one conversation, oldest first. */
     async listSignals(traceId) {
         const res = await this._http.fetchGet(`/api/loop/traces/${encodeURIComponent(traceId)}/signals`);
@@ -159,6 +203,33 @@ export class Loop {
         const res = await this._http.fetchGet(`/api/loop/traces/${encodeURIComponent(traceId)}/candidates`);
         return res.candidates;
     }
+    // ── labels ────────────────────────────────────────────────────────────
+    /**
+     * Label a conversation, so it can be selected later.
+     *
+     * An average over everything is the least useful thing to train on. A model
+     * weak at refunds is fixed with refund examples, and you can only select
+     * those if you said so when the conversation arrived.
+     *
+     * A label may name a `parent`, which is what makes a MASTER GROUP: labelling
+     * something `refunds` with parent `billing` makes it selectable as either,
+     * and selecting `billing` later gathers every child without you maintaining
+     * a list of them.
+     *
+     * Labels are lowercased and trimmed, so `Refunds` and `refunds` are one label.
+     */
+    async label(traceId, labels, parent) {
+        const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/labels`, { labels, parent });
+        return res.labels;
+    }
+    async unlabel(traceId, label) {
+        return this._http.fetchDelete(`/api/loop/traces/${encodeURIComponent(traceId)}/labels/${encodeURIComponent(label)}`);
+    }
+    /** Every label in the workspace, with how many conversations carry it. */
+    async listLabels() {
+        const res = await this._http.fetchGet('/api/loop/labels');
+        return res.labels;
+    }
     // ── training sets ─────────────────────────────────────────────────────
     /**
      * Build a training set from the feedback recorded so far.

package/dist/types.d.ts

@@ -1960,7 +1960,15 @@ export type LoopMethod = 'sft' | 'dpo' | 'grpo';
 /** Who judged an answer. */
 export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
 /** The judgement itself. */
-export type LoopVerdict = 'accepted' | 'rejected' | 'edited' | 'scored';
+/**
+ * The judgement recorded on an answer.
+ *
+ * `accepted` / `rejected` are the thumbs (see `upvote` / `downvote`).
+ * `edited` says the model was WRONG and carries the better answer.
+ * `gold` records the reference answer for the question, making no claim about
+ * whether the model was right — which is why it is separate from `edited`.
+ */
+export type LoopVerdict = 'accepted' | 'rejected' | 'edited' | 'scored' | 'gold';
 /** A function an assistant turn asked to invoke. `arguments` is JSON-encoded. */
 export interface LoopToolCall {
     id?: string;
@@ -2077,6 +2085,10 @@ export interface LoopDatasetCreateParams {
     deployment_id?: string;
     from?: string;
     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. */
+    sample?: number;
     /** Holds back your most recent work, not a random slice. 0-50. */
     holdout_percent?: number;
     max_items?: number;
@@ -2229,3 +2241,15 @@ export interface LoopGradeResult {
     signal_id: string | null;
     candidates_graded: number;
 }
+export interface LoopLabel {
+    label: string;
+    /** The master group this label belongs to, if any. */
+    parent: string | null;
+    created_by: string | null;
+    created_at: string;
+}
+export interface LoopLabelCount {
+    label: string;
+    parent: string | null;
+    traces: number;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.107",
+  "version": "0.2.1-dev.109",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",