@crossworks/client-types 0.232.110 → 0.232.114

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@crossworks/client-types",
-  "version": "0.232.110",
+  "version": "0.232.114",
   "type": "module",
   "main": "./src/index.ts",
   "types": "./src/index.ts",

package/src/model-pools-data.ts

@@ -6,6 +6,11 @@
  * at curation time). Do NOT hand-edit entries here: curate at /models/pools
  * (by hand or via the Curator) and re-export with GET /api/model-pools/export.
  *
+ * One hand-correction on 2026-09-02: the vision ("Read images") pool carried
+ * Nano Banana Pro (`google/gemini-3-pro-image`), which is an image GENERATOR
+ * (`text+image->text+image`). It is out, and `poolModelIssue` in
+ * ./model-pools.ts now rejects that class of entry on every write path.
+ *
  * Seeded into `curated_models` for owners who have no curated entries yet
  * (fresh installs at onboarding; empty existing brains on upgrade). Owners who
  * have curated ANYTHING are never touched — their pools are their judgment.
@@ -2138,27 +2143,6 @@ export const CURATED_MODEL_POOLS: readonly CuratedTemplateEntry[] = [
   {
     pool: 'vision',
     position: 1,
-    name: 'Gemini 3 Pro Image (Nano Banana Pro)',
-    vendor: 'Google',
-    routes: [
-      {
-        provider: 'openrouter',
-        model: 'google/gemini-3-pro-image',
-      },
-    ],
-    pricing: {
-      source: 'openrouter',
-      currency: 'USD',
-      inputPerM: 2,
-      capturedAt: '2026-08-22T15:31:39.522Z',
-      outputPerM: 12,
-    },
-    rating: 5,
-    note: 'purpose-built image model, flagship OCR/extraction quality, can also edit/generate images; 131K context',
-  },
-  {
-    pool: 'vision',
-    position: 2,
     name: 'Claude Sonnet 5',
     vendor: 'Anthropic',
     routes: [
@@ -2183,7 +2167,7 @@ export const CURATED_MODEL_POOLS: readonly CuratedTemplateEntry[] = [
   },
   {
     pool: 'vision',
-    position: 3,
+    position: 2,
     name: 'Qwen3 VL 235B A22B Instruct',
     vendor: 'Alibaba/Qwen',
     routes: [
@@ -2204,7 +2188,7 @@ export const CURATED_MODEL_POOLS: readonly CuratedTemplateEntry[] = [
   },
   {
     pool: 'vision',
-    position: 4,
+    position: 3,
     name: 'Qwen3 VL 30B A3B Instruct',
     vendor: 'Alibaba/Qwen',
     routes: [
@@ -2225,7 +2209,7 @@ export const CURATED_MODEL_POOLS: readonly CuratedTemplateEntry[] = [
   },
   {
     pool: 'vision',
-    position: 5,
+    position: 4,
     name: 'Nemotron Nano 12B V2 VL (free)',
     vendor: 'NVIDIA',
     routes: [

package/src/model-pools.test.ts

@@ -0,0 +1,74 @@
+/**
+ * Pool/model fit. Pins the 2026-09-02 bug: the shipped vision ("Read images")
+ * pool carried Nano Banana Pro, an image GENERATOR. Generators accept image
+ * input exactly like readers do, so nothing on the input side caught it —
+ * the model billed image-generation tokens and returned a picture where the
+ * vision worker wanted text.
+ */
+import { describe, expect, it } from 'vitest';
+import { MODEL_POOLS, poolModelIssue } from './model-pools';
+import { CURATED_MODEL_POOLS } from './model-pools-data';
+
+const READER = { input: ['text', 'image'], output: ['text'] };
+const GENERATOR = { input: ['image', 'text'], output: ['image', 'text'] };
+const TEXT_ONLY = { input: ['text'], output: ['text'] };
+
+describe('poolModelIssue', () => {
+  it('keeps an image generator out of the vision pool', () => {
+    expect(poolModelIssue('vision', GENERATOR)).toMatch(/OUTPUTS images/);
+  });
+
+  it('allows a real image reader in the vision pool', () => {
+    expect(poolModelIssue('vision', READER)).toBeNull();
+  });
+
+  it('keeps a blind model out of the vision pool', () => {
+    expect(poolModelIssue('vision', TEXT_ONLY)).toMatch(/does not accept image input/);
+  });
+
+  it('keeps an image generator out of every text-out pool', () => {
+    for (const pool of MODEL_POOLS.filter((p) => p.modality.output === 'text')) {
+      expect(poolModelIssue(pool.id, GENERATOR), pool.id).toMatch(/OUTPUTS images/);
+    }
+  });
+
+  it('wants a generator in the image_gen pool, and nothing else', () => {
+    expect(poolModelIssue('image_gen', GENERATOR)).toBeNull();
+    expect(poolModelIssue('image_gen', READER)).toMatch(/does not output images/);
+  });
+
+  it('never checks the audio pools — their models are not in the chat catalog', () => {
+    expect(poolModelIssue('tts', GENERATOR)).toBeNull();
+    expect(poolModelIssue('stt', TEXT_ONLY)).toBeNull();
+  });
+
+  it('fails OPEN on an unloaded catalog and on an unknown pool', () => {
+    expect(poolModelIssue('vision', null)).toBeNull();
+    expect(poolModelIssue('vision', { input: [], output: [] })).toBeNull();
+    expect(poolModelIssue('not-a-pool', GENERATOR)).toBeNull();
+  });
+});
+
+describe('the shipped curated template', () => {
+  const slugs = (pool: string) =>
+    new Set(
+      CURATED_MODEL_POOLS.filter((e) => e.pool === pool).flatMap((e) =>
+        e.routes.map((r) => r.model),
+      ),
+    );
+
+  it('never lists the same model as both an image reader and an image generator', () => {
+    const overlap = [...slugs('vision')].filter((s) => slugs('image_gen').has(s));
+    expect(overlap).toEqual([]);
+  });
+
+  it('gives every pool contiguous positions from 0', () => {
+    for (const pool of MODEL_POOLS) {
+      const positions = CURATED_MODEL_POOLS.filter((e) => e.pool === pool.id)
+        .map((e) => e.position)
+        .sort((a, b) => a - b);
+      if (positions.length === 0) continue;
+      expect(positions, pool.id).toEqual(positions.map((_, i) => i));
+    }
+  });
+});

package/src/model-pools.ts

@@ -17,8 +17,33 @@ export type ModelPoolDef = {
   description: string;
   /** Which side of the split it belongs to. */
   group: 'agents' | 'workers';
+  /** What the pool's consumer needs the model to DO, in catalog terms. */
+  modality: PoolModality;
 };
 
+/**
+ * A pool's modality contract, expressed the way OpenRouter's catalog does
+ * (`architecture.input_modalities` / `output_modalities`).
+ *
+ * This exists because of one specific trap: "Read images" and "Image
+ * generation" BOTH accept image input, so the input side alone cannot tell
+ * them apart. A generator like Nano Banana Pro (`google/gemini-3-pro-image`,
+ * `text+image->text+image`) looks like a perfect vision model on inputs
+ * alone, and a curator reading names picks it for the reader pool — where it
+ * bills image-generation tokens and hands back a picture instead of the text
+ * the vision worker parses. The OUTPUT side is the decider: an image READER
+ * is just a capable text-out model that happens to accept pictures.
+ */
+export type PoolModality = {
+  /** Modalities the model must ACCEPT. Empty = text-only is fine. */
+  input: readonly ('image' | 'file')[];
+  /** What the consumer reads back. `audio` pools live in the provider voice
+   *  catalogs, not OpenRouter's chat catalog, so they are never checked. */
+  output: 'text' | 'image' | 'audio';
+};
+
+const TEXT_OUT: PoolModality = { input: [], output: 'text' };
+
 export const MODEL_POOLS: readonly ModelPoolDef[] = [
   {
     id: 'agents',
@@ -26,6 +51,7 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'The shared premium pool: the assistant persona, team responder, and every specialist (pages, tables, coder, appsmith, researcher…). Frontier chat models with strong tool use.',
     group: 'agents',
+    modality: TEXT_OUT,
   },
   {
     id: 'extractor',
@@ -33,18 +59,21 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Reads every ingested item and produces the summary, facts, and entities. Highest call volume in the system — needs cheap, fast, reliable structured output.',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'summarizer',
     label: 'Summarizer',
     description: 'Condenses text wherever a summary is needed. Cheap, fast workhorse.',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'reflector',
     label: 'Reflector',
     description: 'Periodic memory-reflection passes over recent activity. Cheap, fast workhorse.',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'document',
@@ -52,18 +81,21 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Reads whole documents natively (PDF understanding). Needs a multimodal model that accepts document input.',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'vision',
     label: 'Read images',
     description: 'Pulls text and meaning out of images (uploads, extracted document images).',
     group: 'workers',
+    modality: { input: ['image'], output: 'text' },
   },
   {
     id: 'image_gen',
     label: 'Image generation',
     description: 'Generates images. Provider-specific catalog (Gemini image, DALL-E, …).',
     group: 'workers',
+    modality: { input: [], output: 'image' },
   },
   {
     id: 'tts',
@@ -71,6 +103,7 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Turns replies into speech. Provider-specific catalog (Grok voice, GPT-4o TTS voices, ElevenLabs).',
     group: 'workers',
+    modality: { input: [], output: 'audio' },
   },
   {
     id: 'stt',
@@ -78,6 +111,7 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Speech to text for voice notes and video ingest (Whisper family, grok-stt, gpt-4o-mini-transcribe).',
     group: 'workers',
+    modality: { input: [], output: 'audio' },
   },
   {
     id: 'search',
@@ -85,12 +119,14 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'The standard web-search answer tier. Must be a search-native model (Perplexity Sonar family).',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'search_advanced',
     label: 'Deep web search',
     description: 'The strong search tier for hard or conflicting questions (sonar-pro class).',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'narrator',
@@ -98,6 +134,7 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Turns tool outcomes and events into short prose for the user. Off the critical path — cheap.',
     group: 'workers',
+    modality: TEXT_OUT,
   },
   {
     id: 'suggester',
@@ -105,7 +142,58 @@ export const MODEL_POOLS: readonly ModelPoolDef[] = [
     description:
       'Generates the follow-up suggestion chips after a reply. Cheapest of all; has a fallback chain (suggester → narrator → summarizer).',
     group: 'workers',
+    modality: TEXT_OUT,
   },
 ];
 
 export const MODEL_POOL_IDS = new Set(MODEL_POOLS.map((p) => p.id));
+
+const POOL_BY_ID = new Map(MODEL_POOLS.map((p) => [p.id, p]));
+
+/** One model's modalities as OpenRouter reports them. */
+export type ModelModalities = {
+  input: readonly string[];
+  output: readonly string[];
+};
+
+/**
+ * Does this model belong in this pool? Returns the reason it does NOT, or
+ * null when it fits.
+ *
+ * Fail-open by design (same rule as the worker-config catalog check): a
+ * `null` modalities argument means the catalog never loaded, and an outage
+ * must never block a curator from recording their judgment. Only positive
+ * catalog evidence rejects. Audio pools (tts/stt) are never checked — their
+ * models live in the provider voice catalogs, not OpenRouter's chat catalog,
+ * so any "evidence" about them here would be an absence, not a fact.
+ */
+export function poolModelIssue(
+  poolId: string,
+  modalities: ModelModalities | null | undefined,
+): string | null {
+  const pool = POOL_BY_ID.get(poolId);
+  if (!pool || !modalities) return null;
+  const want = pool.modality;
+  if (want.output === 'audio') return null;
+  const outputs = modalities.output ?? [];
+  const inputs = modalities.input ?? [];
+  if (outputs.length === 0 && inputs.length === 0) return null;
+
+  const makesImages = outputs.includes('image');
+  if (want.output === 'text' && makesImages) {
+    return (
+      `this model OUTPUTS images (${outputs.join('+')}) — it is an image generator, ` +
+      `and the ${pool.label} pool needs a text-out model. Put it in the Image generation ` +
+      `pool instead. Reading images is just a capable text-out model that accepts pictures.`
+    );
+  }
+  if (want.output === 'image' && !makesImages && outputs.length > 0) {
+    return `this model does not output images (${outputs.join('+')}) — the ${pool.label} pool needs a generator.`;
+  }
+  for (const need of want.input) {
+    if (inputs.length > 0 && !inputs.includes(need)) {
+      return `this model does not accept ${need} input (accepts ${inputs.join('+')}) — the ${pool.label} pool needs one that does.`;
+    }
+  }
+  return null;
+}