runbios-sdk 0.2.19-rc.279 → 0.2.20-dev.283

package/dist/client.js

@@ -344,7 +344,7 @@ 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 = "0.2.19-rc.279";
+export declare const VERSION = "0.2.20-dev.283";
 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 = '0.2.19-rc.279';
+export const VERSION = '0.2.20-dev.283';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/inference.js

@@ -336,7 +336,7 @@ 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

@@ -402,7 +402,9 @@ export declare class Loop {
      * numbers `minimums` shows. 503 `AGENT_COULD_NOT_START` means the key the
      * attempt would be scored with could not be renewed just then: nothing was
      * started, so press train now again in a minute. Every refusal is a
-     * `TrainingRuleRunRefusalCode`.
+     * `TrainingRuleRunRefusalCode`. When the attempt in flight is waiting for
+     * GPUs to compare (`active_run.state` `waiting_gpus`), this books its two
+     * comparison machines again at once instead of starting another attempt.
      */
     runPipeline(id: string): Promise<PipelineMutationResponse>;
     /**

package/dist/resources/loop.js

@@ -555,7 +555,9 @@ export class Loop {
      * numbers `minimums` shows. 503 `AGENT_COULD_NOT_START` means the key the
      * attempt would be scored with could not be renewed just then: nothing was
      * started, so press train now again in a minute. Every refusal is a
-     * `TrainingRuleRunRefusalCode`.
+     * `TrainingRuleRunRefusalCode`. When the attempt in flight is waiting for
+     * GPUs to compare (`active_run.state` `waiting_gpus`), this books its two
+     * comparison machines again at once instead of starting another attempt.
      */
     async runPipeline(id) {
         return this._http.fetchPost(`/api/loop/pipelines/${encodeURIComponent(id)}/run`, {});

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;
@@ -781,10 +781,24 @@ export interface TrainingJob {
     checkpoint_count?: number;
     /** The current machine attempt's per-phase records, in canonical order. */
     phases?: TrainingJobPhase[];
-    /** Estimated seconds until the job finishes (absent when unknown). */
+    /**
+     * Estimated seconds remaining (absent when unknown). What it covers is
+     * `eta_scope`: before step 1 only the time until training starts.
+     */
     eta_seconds?: number;
-    /** How the ETA was derived: measured on the machine, or from historical medians. */
+    /**
+     * How the ETA was derived: `measured` from this job's own live rate
+     * (training: its own steps since the first), or `historical` from past runs
+     * of the same model on the same GPU (else the GPU).
+     */
     eta_confidence?: 'measured' | 'historical';
+    /**
+     * What `eta_seconds` covers: `until_training` before step 1 (startup
+     * phases only; training time is measured from the run's own steps, never
+     * guessed), `run` once training has started (remaining training plus what
+     * follows it).
+     */
+    eta_scope?: 'until_training' | 'run';
     /** Most recent per-phase update time. */
     last_phase_update_at?: string;
     /**
@@ -1663,6 +1677,12 @@ export interface InferenceDeployment {
     base_model_revision?: string;
     immutable_base_model_source?: string;
     serving_mode?: 'full' | 'adapter' | 'merged';
+    /**
+     * Only on an adapter deployment: the `model` that asks its BASE weights with
+     * no adapter, `<name>:base`, on /chat/completions and /completions. A
+     * deployment with no adapter has no separate base and refuses `:base`.
+     */
+    base_weights_model?: string;
     supports_tool_calls?: boolean;
     tool_call_parser?: InferenceToolCallParser | null;
     reasoning_parser?: InferenceReasoningParser | null;
@@ -2866,7 +2886,18 @@ export interface TrainingRecipe {
     lora_rank: number | null;
     lora_alpha: number | null;
 }
-export type TrainingRunState = 'built' | 'registering' | 'dataset_validating' | 'preflighting' | 'job_creating' | 'training' | 'checkpoint_ready' | 'candidate_booking' | 'candidate_running' | 'evaluating' | 'reported' | 'awaiting_review' | 'promoting' | 'waiting_funds' | 'promoted' | 'rejected' | 'expired' | 'failed' | 'budget_stopped' | 'cancelled' | 'superseded' | 'rolled_back';
+export type TrainingRunState = 'built' | 'registering' | 'dataset_validating' | 'preflighting' | 'job_creating' | 'training' | 'checkpoint_ready' | 'candidate_booking' | 'candidate_running' | 'evaluating' | 'reported' | 'awaiting_review' | 'promoting' | 'waiting_funds'
+/**
+ * Trained, and waiting for GPUs to run the new model and the one it is
+ * compared with side by side: one machine got a GPU and the other did not
+ * within 20 minutes, so both were stopped. Each try can bill the machine
+ * that got a GPU for up to those 20 minutes; nothing bills between tries.
+ * It keeps what it trained, books both again after 30 minutes, then 1, 2
+ * and 4 hours, and ends (`failed`, `COMPARISON_NO_GPUS`) after four tries or
+ * 24 hours (`COMPARISON_WORKSPACE_FULL` when what was missing was room in
+ * the workspace). Train now books them again at once.
+ */
+ | 'waiting_gpus' | 'promoted' | 'rejected' | 'expired' | 'failed' | 'budget_stopped' | 'cancelled' | 'superseded' | 'rolled_back';
 /** The closed set a run never leaves. */
 export declare const TERMINAL_RUN_STATES: readonly TrainingRunState[];
 export type TrainingVerdict = 'better' | 'not_better' | 'inconclusive' | 'not_evaluated';
@@ -2957,7 +2988,19 @@ export interface TrainingRun {
      */
     error_next_step: string | null;
     billed_training_cents: number;
+    /**
+     * The comparison machines: each from the first time it was on a GPU to its
+     * stop, less any time back in the GPU queue, at the price the platform
+     * placed it at.
+     */
     billed_candidate_cents: number;
+    /**
+     * True when a comparison machine was charged: the figure is then the most
+     * it can have cost ("up to"), since it counts each machine from its first
+     * boot where the platform bills from its first answer (and at the hourly
+     * cap when no placement price was reported).
+     */
+    billed_candidate_up_to: boolean;
     spent_eval_cents: number;
     train_rows: number | null;
     holdout_rows: number | null;
@@ -3049,6 +3092,8 @@ export interface TrainingRunSummary {
     holdout_rows: number | null;
     billed_training_cents: number;
     billed_candidate_cents: number;
+    /** See {@link TrainingRun.billed_candidate_up_to}. */
+    billed_candidate_up_to: boolean;
     spent_eval_cents: number;
     /** The fourth money column. A row that leaves it out adds up short. */
     benchmark_spent_cents: number;
@@ -3528,8 +3573,9 @@ export interface TrainingRuleMonthlyLimitRefusal {
     monthly_ceiling_cents: number;
     /**
      * The most one attempt of this rule may spend: its training, its comparison
-     * machine and its judging, and -- on a pipeline, while no version is live --
-     * the base model's own comparison machine beside the new model's.
+     * 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;
     /**
@@ -3574,6 +3620,11 @@ export interface PipelineActiveRun {
     gpu_seconds: number | null;
     /** What it has been billed so far, in cents. */
     cost_cents: number;
+    /**
+     * True when a comparison machine was charged in `cost_cents`: the figure is
+     * the most it can have cost, so show it as "up to".
+     */
+    cost_up_to: boolean;
     /** The version number it will take if it is put live. */
     will_be_version: number;
     /** Always null: nothing in flight is a version yet. */
@@ -3646,6 +3697,14 @@ export interface PipelineVersion {
      * not what it did.
      */
     cost_includes_candidate_ceiling: boolean;
+    /**
+     * True when `cost_cents` is the most the version can have cost rather than
+     * what it did: its comparison machines at their ceilings
+     * (`cost_includes_candidate_ceiling`), or one charged (counted from its
+     * first boot, where the platform bills from its first answer). Show it as
+     * "up to".
+     */
+    cost_up_to: boolean;
     /** Machine time the attempt held (training and comparison); null when unknown. */
     gpu_seconds: number | null;
     created_at: string;
@@ -3694,6 +3753,8 @@ export interface PipelineAttempt {
     gpu_seconds: number | null;
     /** As the versions are priced once it has ended; as recorded so far while it runs. */
     cost_cents: number;
+    /** See {@link PipelineVersion.cost_up_to}. */
+    cost_up_to: boolean;
     verdict: TrainingVerdict | null;
     win_rate: number | null;
     evaluation_id: string | null;
@@ -3823,9 +3884,9 @@ export interface Pipeline {
      * 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, the base
-     * model's comparison machine included) or what it has been billed when that
-     * is more. A finished run counts what it was billed, and its comparison
+     * 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

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.19-rc.279",
+  "version": "0.2.20-dev.283",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",