@prestyj/core 5.28.0 → 5.29.0

package/dist/model-registry.d.cts

@@ -16,8 +16,8 @@ interface ModelInfo {
     /**
      * Vendor-declared default reasoning level (Codex models.json
      * `default_reasoning_level`). When present, fresh sessions start here rather
-     * than at the ceiling: the deep-reasoning flagships (Astra ships "low",
-     * GPT-6 Sol/Luna "medium") think dramatically longer per rung, so defaulting to
+     * than at the ceiling: the deep-reasoning models (Astra and GPT-6.1 Sol ship
+     * "low", GPT-6 Luna "medium") think dramatically longer per rung, so defaulting to
      * `maxThinkingLevel` made new sessions pathologically slow.
      */
     defaultThinkingLevel?: ThinkingLevel;
@@ -39,7 +39,7 @@ interface ModelInfo {
     /**
      * The top reasoning tier this model genuinely uses. Used when thinking is
      * enabled to pick the strongest setting per model:
-     *   - OpenAI GPT-6 Astra / Sol: `ultra` (Codex orchestration preset above `max`)
+     *   - OpenAI GPT-6 Astra / GPT-6.1 Sol: `ultra` (Codex orchestration preset above `max`)
      *   - OpenAI GPT-6 Luna: `max`
      *   - OpenAI Pro/Codex/old: clamped to what the model accepts
      *   - Claude Fable 5.1 / Fable 5 / Mythos 5, Opus 5.5 / Opus 5 and Sonnet 5.5: `max`
@@ -134,25 +134,17 @@ declare function getDefaultThinkingLevel(modelId: string, options?: {
     baseUrl?: string;
 }): ThinkingLevel;
 /**
- * Get the model to use for compaction summarization.
- * - Anthropic: always Sonnet 5.5
- * - OpenAI: cheapest (Codex Mini)
- * - Gemini: use the current model
- * - GLM: GLM-5.3-Flash (the registered low-cost sibling)
- * - Moonshot: use the current model (no cheap alternative registered)
- */
-declare function getSummaryModel(provider: Provider, currentModelId: string): ModelInfo;
-/**
- * Fastest/cheapest sibling within the SAME provider, for scout-style read-only
- * sub-agents (recon, research) where a low-latency model is enough and the
- * frontier model is wasted spend + latency.
+ * Get the model to use for compaction summarization — the ONLY place a
+ * cheaper model is used. Everything else (sub-agents, reviewers, judges)
+ * runs on the parent's model.
  *
- * Routes off each model's `costTier` — the single source of truth that already
- * travels with the registry entry — so a model rename/bump needs no change
- * here. Providers with no low-tier sibling (Moonshot, MiniMax, Sakana,
- * OpenRouter) gracefully keep the parent model, so there's never a
- * crash or a cross-provider jump to a login the user may not have.
+ * Picks the provider's first model tagged with its summary tier
+ * ({@link SUMMARY_COST_TIER}). When none is registered — the tier was removed,
+ * or the provider has no cheap sibling — it summarizes on the current model.
+ * Never throws. The caller still retries on the current model when the
+ * provider rejects the chosen one at runtime (retired upstream, or not on the
+ * user's plan), since the registry cannot know that.
  */
-declare function getFastModel(provider: Provider, currentModelId: string): ModelInfo;
+declare function getSummaryModel(provider: Provider, currentModelId: string): ModelInfo;
 
-export { type ContextWindowOptions, DEFAULT_MAX_VIDEO_BYTES, MODELS, type ModelInfo, clearRuntimeModels, getAllModels, getAuthStorageKey, getAuthStorageKeys, getContextWindow, getDefaultModel, getDefaultThinkingLevel, getFastModel, getMaxThinkingLevel, getModel, getModelsForProvider, getSummaryModel, getToolResultCharLimit, getVideoByteLimit, registerRuntimeModels, usesOpenAICodexTransport };
+export { type ContextWindowOptions, DEFAULT_MAX_VIDEO_BYTES, MODELS, type ModelInfo, clearRuntimeModels, getAllModels, getAuthStorageKey, getAuthStorageKeys, getContextWindow, getDefaultModel, getDefaultThinkingLevel, getMaxThinkingLevel, getModel, getModelsForProvider, getSummaryModel, getToolResultCharLimit, getVideoByteLimit, registerRuntimeModels, usesOpenAICodexTransport };

package/dist/model-registry.d.ts

@@ -16,8 +16,8 @@ interface ModelInfo {
     /**
      * Vendor-declared default reasoning level (Codex models.json
      * `default_reasoning_level`). When present, fresh sessions start here rather
-     * than at the ceiling: the deep-reasoning flagships (Astra ships "low",
-     * GPT-6 Sol/Luna "medium") think dramatically longer per rung, so defaulting to
+     * than at the ceiling: the deep-reasoning models (Astra and GPT-6.1 Sol ship
+     * "low", GPT-6 Luna "medium") think dramatically longer per rung, so defaulting to
      * `maxThinkingLevel` made new sessions pathologically slow.
      */
     defaultThinkingLevel?: ThinkingLevel;
@@ -39,7 +39,7 @@ interface ModelInfo {
     /**
      * The top reasoning tier this model genuinely uses. Used when thinking is
      * enabled to pick the strongest setting per model:
-     *   - OpenAI GPT-6 Astra / Sol: `ultra` (Codex orchestration preset above `max`)
+     *   - OpenAI GPT-6 Astra / GPT-6.1 Sol: `ultra` (Codex orchestration preset above `max`)
      *   - OpenAI GPT-6 Luna: `max`
      *   - OpenAI Pro/Codex/old: clamped to what the model accepts
      *   - Claude Fable 5.1 / Fable 5 / Mythos 5, Opus 5.5 / Opus 5 and Sonnet 5.5: `max`
@@ -134,25 +134,17 @@ declare function getDefaultThinkingLevel(modelId: string, options?: {
     baseUrl?: string;
 }): ThinkingLevel;
 /**
- * Get the model to use for compaction summarization.
- * - Anthropic: always Sonnet 5.5
- * - OpenAI: cheapest (Codex Mini)
- * - Gemini: use the current model
- * - GLM: GLM-5.3-Flash (the registered low-cost sibling)
- * - Moonshot: use the current model (no cheap alternative registered)
- */
-declare function getSummaryModel(provider: Provider, currentModelId: string): ModelInfo;
-/**
- * Fastest/cheapest sibling within the SAME provider, for scout-style read-only
- * sub-agents (recon, research) where a low-latency model is enough and the
- * frontier model is wasted spend + latency.
+ * Get the model to use for compaction summarization — the ONLY place a
+ * cheaper model is used. Everything else (sub-agents, reviewers, judges)
+ * runs on the parent's model.
  *
- * Routes off each model's `costTier` — the single source of truth that already
- * travels with the registry entry — so a model rename/bump needs no change
- * here. Providers with no low-tier sibling (Moonshot, MiniMax, Sakana,
- * OpenRouter) gracefully keep the parent model, so there's never a
- * crash or a cross-provider jump to a login the user may not have.
+ * Picks the provider's first model tagged with its summary tier
+ * ({@link SUMMARY_COST_TIER}). When none is registered — the tier was removed,
+ * or the provider has no cheap sibling — it summarizes on the current model.
+ * Never throws. The caller still retries on the current model when the
+ * provider rejects the chosen one at runtime (retired upstream, or not on the
+ * user's plan), since the registry cannot know that.
  */
-declare function getFastModel(provider: Provider, currentModelId: string): ModelInfo;
+declare function getSummaryModel(provider: Provider, currentModelId: string): ModelInfo;
 
-export { type ContextWindowOptions, DEFAULT_MAX_VIDEO_BYTES, MODELS, type ModelInfo, clearRuntimeModels, getAllModels, getAuthStorageKey, getAuthStorageKeys, getContextWindow, getDefaultModel, getDefaultThinkingLevel, getFastModel, getMaxThinkingLevel, getModel, getModelsForProvider, getSummaryModel, getToolResultCharLimit, getVideoByteLimit, registerRuntimeModels, usesOpenAICodexTransport };
+export { type ContextWindowOptions, DEFAULT_MAX_VIDEO_BYTES, MODELS, type ModelInfo, clearRuntimeModels, getAllModels, getAuthStorageKey, getAuthStorageKeys, getContextWindow, getDefaultModel, getDefaultThinkingLevel, getMaxThinkingLevel, getModel, getModelsForProvider, getSummaryModel, getToolResultCharLimit, getVideoByteLimit, registerRuntimeModels, usesOpenAICodexTransport };

package/dist/model-registry.js

@@ -8,7 +8,6 @@ import {
   getContextWindow,
   getDefaultModel,
   getDefaultThinkingLevel,
-  getFastModel,
   getMaxThinkingLevel,
   getModel,
   getModelsForProvider,
@@ -17,7 +16,7 @@ import {
   getVideoByteLimit,
   registerRuntimeModels,
   usesOpenAICodexTransport
-} from "./chunk-ZCXAIMZF.js";
+} from "./chunk-N73LFQ3N.js";
 import "./chunk-CRU3SSNX.js";
 export {
   DEFAULT_MAX_VIDEO_BYTES,
@@ -29,7 +28,6 @@ export {
   getContextWindow,
   getDefaultModel,
   getDefaultThinkingLevel,
-  getFastModel,
   getMaxThinkingLevel,
   getModel,
   getModelsForProvider,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@prestyj/core",
-  "version": "5.28.0",
+  "version": "5.29.0",
   "type": "module",
   "description": "Provider-agnostic, UI-free shared foundation: model registry, auth, paths, telegram, voice",
   "license": "MIT",
@@ -33,7 +33,7 @@
     "dist"
   ],
   "dependencies": {
-    "@prestyj/ai": "5.28.0"
+    "@prestyj/ai": "5.29.0"
   },
   "optionalDependencies": {
     "@huggingface/transformers": "4.2.0",