@visual-z/xeno 0.1.4 → 0.2.0

package/src/runtime/model-replay.ts

@@ -0,0 +1,10 @@
+import { createHash } from "node:crypto";
+import { replayDescriptor } from "../core/model-replay.ts";
+import { hasProviderAccess } from "./provider-access.ts";
+import type { ModelSpec } from "./model.ts";
+
+export function nativeReplay(spec: ModelSpec): boolean { return spec.driver === "pi" || hasProviderAccess(spec); }
+/** The log keeps a digest rather than endpoint values, which can contain private routing information. */
+export function providerReplayKey(spec: ModelSpec): string {
+  return createHash("sha256").update(replayDescriptor({ ...spec, driver: nativeReplay(spec) ? "pi" : "legacy" })).digest("hex");
+}

package/src/runtime/model-scope.ts

@@ -0,0 +1,20 @@
+/** Bun's existing glob implementation supplies matching; ordered scope policy remains pure. */
+import { resolveScope } from "../core/model-scope.ts";
+import { modelKey } from "../core/session-position.ts";
+import { asThinkingLevel } from "../core/thinking.ts";
+import { findModelByPattern } from "./catalog.ts";
+import type { ModelSpec } from "./model.ts";
+
+export function scopedModels(patterns: readonly string[], models: readonly ModelSpec[], strict = false) {
+  return resolveScope(patterns, pattern => {
+    const lower = pattern.toLowerCase();
+    const exact = models.filter(model => modelKey(model).toLowerCase() === lower || model.id.toLowerCase() === lower);
+    if (exact.length === 1) return exact;
+    if (!strict && /[*?\[]/.test(pattern)) {
+      if (asThinkingLevel(pattern.slice(pattern.lastIndexOf(":") + 1))) return [];
+      const glob = new Bun.Glob(lower);
+      return models.filter(model => glob.match(modelKey(model).toLowerCase()) || glob.match(model.id.toLowerCase()));
+    }
+    const found = findModelByPattern(pattern, models); return found.ok ? [found.spec] : [];
+  }, strict);
+}

package/src/runtime/model-sources.ts

@@ -0,0 +1,151 @@
+/** Provider wire metadata is interpreted here, not in the SDK-free core (D13). */
+import { getBuiltinModels, getBuiltinProviders } from "@earendil-works/pi-ai/providers/all";
+import { mergeCatalogs } from "../core/catalog-merge.ts";
+import { THINKING_LEVELS } from "../core/thinking.ts";
+import type { Api, ModelSpec, Price, Effort } from "./model.ts";
+import { isProviderApi, SDK_APIS } from "./provider-protocols.ts";
+
+export interface ProviderDefinition { name: string; env: string[]; api?: Api; baseUrl?: string }
+export interface CatalogExclusion { provider: string; id: string; reason: string }
+export interface SourceCatalog {
+  models: ModelSpec[];
+  providers: Record<string, ProviderDefinition>;
+  excluded: CatalogExclusion[];
+}
+export function record(value: unknown): value is Record<string, unknown> {
+  return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+export function finitePositive(value: unknown): value is number { return typeof value === "number" && Number.isFinite(value) && value > 0; }
+export function finiteNonnegative(value: unknown): value is number { return typeof value === "number" && Number.isFinite(value) && value >= 0; }
+
+/** Published unknown prices remain unknown, even when an SDK requires numeric placeholders. */
+function sourcePrice(value: unknown): Price | undefined {
+  if (!record(value) || !finiteNonnegative(value["input"]) || !finiteNonnegative(value["output"])) return undefined;
+  return { input: value["input"], output: value["output"],
+    ...(finiteNonnegative(value["cache_read"]) ? { cacheRead: value["cache_read"] } : {}),
+    ...(finiteNonnegative(value["cache_write"]) ? { cacheWrite: value["cache_write"] } : {}) };
+}
+function priced(spec: ModelSpec, price: Price | undefined): ModelSpec {
+  if (!price) return spec;
+  return { ...spec, price, ...(price.input > 0 && price.cacheRead !== undefined && price.cacheWrite !== undefined
+    ? { cache: spec.api === "anthropic-messages" ? "explicit" as const : "automatic" as const,
+      rates: { read: price.cacheRead / price.input, write: price.cacheWrite / price.input } } : {}) };
+}
+
+export function piSourceCatalog(): SourceCatalog {
+  const out: SourceCatalog = { models: [], providers: {}, excluded: [] };
+  for (const provider of getBuiltinProviders()) {
+    for (const model of getBuiltinModels(provider)) {
+      if (!isProviderApi(model.api)) { out.excluded.push({ provider, id: model.id, reason: `unsupported Pi API ${model.api}` }); continue; }
+      const spec: ModelSpec = { id: model.id, provider, name: model.name, api: model.api, driver: "pi",
+        baseUrl: model.baseUrl.replace(/(?<!\$)\{([A-Z][A-Z_0-9]*)\}/g, "$${$1}"),
+        input: model.input, reasoning: model.reasoning, contextWindow: model.contextWindow, maxTokens: model.maxTokens,
+        ...(model.compat ? { nativeCompat: model.compat as Record<string, unknown> } : {}),
+        ...(model.thinkingLevelMap ? { thinkingLevelMap: model.thinkingLevelMap } : {}),
+        ...(model.samplingParams ? { samplingParams: model.samplingParams } : {}),
+        ...(model.headers ? { defaultHeaders: model.headers } : {}) };
+      // Pi fills missing cost slots with zero. Do not turn those placeholders into measured prices.
+      const cost = model.cost;
+      out.models.push(priced(spec, finiteNonnegative(cost.input) && finiteNonnegative(cost.output) && (cost.input || cost.output) ? { input: cost.input, output: cost.output,
+        ...(finitePositive(cost.cacheRead) ? { cacheRead: cost.cacheRead } : {}), ...(finitePositive(cost.cacheWrite) ? { cacheWrite: cost.cacheWrite } : {}) } : undefined));
+      out.providers[provider] ??= { name: provider, env: [], api: spec.api, ...(spec.baseUrl ? { baseUrl: spec.baseUrl } : {}) };
+    }
+  }
+  return out;
+}
+
+// SDK defaults missing from Models.dev. These are URLs, not protocol guesses.
+// DeepInfra docs/quickstart.mdx; the other defaults are Pi's published provider models.
+const ENDPOINTS: Readonly<Record<string, string>> = {
+  deepinfra: "https://api.deepinfra.com/v1/openai", vercel: "https://ai-gateway.vercel.sh",
+  togetherai: "https://api.together.xyz/v1", azure: "", "azure-cognitive-services": "",
+};
+
+// Dedicated SDK names need not imply dedicated protocols. Only documented compatible endpoints
+// are mapped here (provider-union task card); unknown SDKs remain explicit exclusions.
+const COMPATIBLE_ENDPOINTS: Readonly<Record<string, string>> = {
+  "@ai-sdk/cohere": "https://api.cohere.ai/compatibility/v1",
+  "venice-ai-sdk-provider": "https://api.venice.ai/api/v1",
+  "@aihubmix/ai-sdk-provider": "https://aihubmix.com/v1",
+};
+
+function reasoningMap(value: unknown): ModelSpec["thinkingLevelMap"] {
+  if (!Array.isArray(value)) return undefined;
+  const effort = value.find((entry: unknown) => record(entry) && entry["type"] === "effort");
+  if (!record(effort) || !Array.isArray(effort["values"])) return undefined;
+  const values = effort["values"].filter((x): x is string => typeof x === "string");
+  const map: Partial<Record<Effort, string | null>> = {};
+  for (const level of THINKING_LEVELS) map[level] = values.includes(level) ? level : null;
+  return map;
+}
+
+/** The same normalization feeds the bundled snapshot and explicit refresh. No whitelist of providers. */
+export function modelsDevCatalog(json: unknown, pi: SourceCatalog = piSourceCatalog()): SourceCatalog {
+  const out: SourceCatalog = { models: [], providers: {}, excluded: [] };
+  if (!record(json)) return out;
+  const piModels = new Map(pi.models.map(model => [`${model.provider}/${model.id}`, model]));
+  for (const [rawProvider, value] of Object.entries(json)) {
+    if (!record(value) || !record(value["models"])) continue;
+    // Normalize the upstream spelling to the existing xeno/Pi provider identity before merging.
+    const provider = rawProvider === "togetherai" ? "together" : rawProvider;
+    const piDefault = pi.providers[provider];
+    const env = Array.isArray(value["env"]) ? value["env"].filter((x): x is string => typeof x === "string") : [];
+    const definition: ProviderDefinition = { name: typeof value["name"] === "string" ? value["name"] : provider, env };
+    out.providers[provider] = definition;
+    for (const [key, raw] of Object.entries(value["models"])) {
+      if (!record(raw)) continue;
+      const id = typeof raw["id"] === "string" ? raw["id"] : key;
+      const skip = (reason: string) => { out.excluded.push({ provider, id, reason }); };
+      const modalities = record(raw["modalities"]) ? raw["modalities"] : {};
+      if (raw["tool_call"] === false) { skip("model declares no tool calling"); continue; }
+      if (Array.isArray(modalities["output"]) && !modalities["output"].includes("text")) { skip("non-text output model"); continue; }
+      const modelProvider = record(raw["provider"]) ? raw["provider"] : {};
+      const npm = typeof modelProvider["npm"] === "string" ? modelProvider["npm"] : typeof value["npm"] === "string" ? value["npm"] : "@ai-sdk/openai-compatible";
+      const matched = piModels.get(`${provider}/${id}`);
+      const compatibleEndpoint = COMPATIBLE_ENDPOINTS[npm];
+      const api = matched?.api ?? SDK_APIS[npm] ?? (compatibleEndpoint ? "openai-completions" : undefined);
+      if (!api) { skip(`requires dedicated SDK ${npm}`); continue; }
+      const baseUrl = matched?.baseUrl ?? compatibleEndpoint ?? (typeof modelProvider["api"] === "string" ? modelProvider["api"] : typeof value["api"] === "string" ? value["api"] : ENDPOINTS[rawProvider] ?? piDefault?.baseUrl);
+      if (!baseUrl && !["bedrock-converse-stream", "google-vertex", "azure-openai-responses"].includes(api)) { skip("provider has no verified endpoint; configure baseUrl"); continue; }
+      const limit = record(raw["limit"]) ? raw["limit"] : {};
+      const effort = reasoningMap(raw["reasoning_options"]);
+      const spec: ModelSpec = { ...(matched ?? {}), id, provider, name: typeof raw["name"] === "string" ? raw["name"] : id,
+        api, driver: "pi", reasoning: raw["reasoning"] === true, credentialEnv: env,
+        ...(baseUrl !== undefined ? { baseUrl } : {}),
+        ...(Array.isArray(modalities["input"]) ? { input: modalities["input"].filter((x): x is "text" | "image" => x === "text" || x === "image") } : {}),
+        ...(finitePositive(limit["context"]) ? { contextWindow: limit["context"] } : {}),
+        ...(finitePositive(limit["output"]) ? { maxTokens: limit["output"] } : {}),
+        ...(!matched?.thinkingLevelMap && effort ? { thinkingLevelMap: effort } : {}),
+        ...(!matched && api === "openai-completions" ? { nativeCompat: {
+          supportsStore: false, supportsDeveloperRole: false, maxTokensField: "max_tokens",
+          supportsReasoningEffort: effort !== undefined || provider === "cohere",
+          ...(provider === "cohere" ? { supportsUsageInStreaming: false } : {}),
+        } } : {}),
+        ...(!matched && provider === "cohere" && raw["reasoning"] === true ? {
+          thinkingLevelMap: { off: "none", minimal: "high", low: "high", medium: "high", high: "high", xhigh: "high", max: "high" },
+        } : {}) };
+      // Native Pi compatibility wins over less expressive directory metadata.
+      out.models.push(priced(spec, sourcePrice(raw["cost"])));
+      definition.api ??= api;
+      if (baseUrl) definition.baseUrl ??= baseUrl;
+    }
+  }
+  return out;
+}
+
+export function unionSourceCatalog(json: unknown): SourceCatalog {
+  const pi = piSourceCatalog();
+  const remote = modelsDevCatalog(json, pi);
+  const denied = new Set(remote.excluded.filter(row => row.reason.includes("no tool") || row.reason.includes("non-text")).map(row => `${row.provider}/${row.id}`));
+  const directory = new Map(remote.models.map(model => [`${model.provider}/${model.id}`, model]));
+  const models = mergeCatalogs(remote.models, pi.models.filter(model => !denied.has(`${model.provider}/${model.id}`)))
+    .map(model => {
+      // Pi supplies dialects; the directory supplies public credential names and explicit prices.
+      const metadata = directory.get(`${model.provider}/${model.id}`);
+      return { ...model, ...(metadata?.credentialEnv ? { credentialEnv: metadata.credentialEnv } : {}),
+        ...(metadata?.price ? { price: metadata.price } : {}) };
+    });
+  const included = new Set(models.map(model => `${model.provider}/${model.id}`));
+  return { models, providers: { ...pi.providers, ...remote.providers },
+    excluded: [...pi.excluded, ...remote.excluded].filter(row => !included.has(`${row.provider}/${row.id}`)) };
+}

package/src/runtime/model.ts

@@ -14,15 +14,16 @@
  */
 
 /**
- * The wire formats worth abstracting. Four is the number pi arrived at after
- * supporting dozens of providers (research appendix, line 176); this project
- * needs two of them so far, and adding a third should be adding a member here
- * plus an adapter, never a conditional inside an existing adapter.
+ * The protocol vocabulary is local even when the runtime delegates wire IO to Pi.
  */
-export type Api = "anthropic-messages" | "openai-completions" | "openai-responses";
+import type { ProviderApi } from "./provider-protocols.ts";
+import { SOURCE_MODELS } from "./model-data.ts";
+import { mergeCatalogs } from "../core/catalog-merge.ts";
+export type Api = ProviderApi;
 
-/** How hard the model should think. Mapped per-model: not every provider takes all five. */
-export type Effort = "low" | "medium" | "high" | "xhigh" | "max";
+/** How hard the model should think. Capability and wire vocabulary are mapped per model. */
+import type { ThinkingLevel } from "../core/thinking.ts";
+export type Effort = ThinkingLevel;
 
 /**
  * Who decides what gets cached.
@@ -140,6 +141,7 @@ export const KIMI_K3_COMPAT: Compat = {
   thinkingFormat: "openai",
 };
 export const KIMI_K3_EFFORT: Partial<Record<Effort, string | null>> = {
+  off: null, minimal: "low",
   low: "low", medium: "high", high: "high", xhigh: "max", max: "max",
 };
 
@@ -160,6 +162,7 @@ export const GEMINI_COMPAT: Compat = {
   thinkingFormat: "openai",
 };
 export const GEMINI_EFFORT: Partial<Record<Effort, string | null>> = {
+  off: null, minimal: "low",
   low: "low", medium: "medium", high: "high", xhigh: "high", max: "high",
 };
 
@@ -173,8 +176,18 @@ export const DEFAULT_COMPAT: Compat = {
 };
 
 export interface ModelSpec {
+  /** Public variable names from the source directory, never their values. */
+  credentialEnv?: readonly string[];
+  /** Native Pi transport is runtime-only. Existing models retain their verified adapter. */
+  driver?: "pi";
+  /** Public catalog metadata only. User credentials/headers live in provider-access's sidecar. */
+  nativeCompat?: Record<string, unknown>;
+  samplingParams?: Record<string, unknown>;
+  defaultHeaders?: Record<string, string>;
   /** The id sent on the wire. */
   id: string;
+  /** Optional upstream deployment/model id when a user gives it a local catalog alias. */
+  wireId?: string;
   /** For humans, in `/model` listings. */
   name: string;
   api: Api;
@@ -187,6 +200,8 @@ export interface ModelSpec {
    */
   baseUrl?: string;
   reasoning: boolean;
+  /** Explicit capability from the model catalog. Unknown input capability is text-only. */
+  input?: readonly ("text" | "image")[];
   /**
    * Absent when the provider's caching terms are not in the source we
    * trust. Filling `automatic` + `write: 0` here would be DeepSeek's
@@ -206,9 +221,8 @@ export interface ModelSpec {
   /** Only meaningful for `openai-completions`. Absent means `DEFAULT_COMPAT`. */
   compat?: Compat;
   /**
-   * Our five effort levels onto what this model accepts. `null` means "send
-   * no effort at all and let the model default"; an absent entry means send
-   * our level through unchanged.
+   * Pi's effort names onto the wire vocabulary. `null` disables a level.
+   * Unmapped xhigh/max are unsupported; other missing names pass through.
    */
   thinkingLevelMap?: Partial<Record<Effort, string | null>>;
 }
@@ -223,10 +237,14 @@ export const ANTHROPIC_RATES: CacheRates = { read: 0.1, write: 1.25 };
 
 export const CLAUDE_OPUS_5: ModelSpec = {
   id: "claude-opus-5",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
   name: "Claude Opus 5",
   api: "anthropic-messages",
   provider: "anthropic",
   reasoning: true,
+  // Preserve the built-in request vocabulary while the generic capability policy becomes explicit.
+  thinkingLevelMap: { minimal: "low", xhigh: "xhigh", max: "max" },
   cache: "explicit",
   rates: ANTHROPIC_RATES,
   // Cost from models.dev on 2026-09-21. Ratios match ANTHROPIC_RATES.
@@ -257,6 +275,7 @@ const DEEPSEEK_COMPAT: Compat = {
  * sending a level the API has to guess about.
  */
 const DEEPSEEK_EFFORT: Partial<Record<Effort, string | null>> = {
+  minimal: "low",
   low: "low",
   medium: "high",
   high: "high",
@@ -275,6 +294,8 @@ const DEEPSEEK_EFFORT: Partial<Record<Effort, string | null>> = {
  */
 export const DEEPSEEK_V4_PRO: ModelSpec = {
   id: "deepseek-v4-pro",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
   name: "DeepSeek V4 Pro",
   api: "openai-completions",
   provider: "deepseek",
@@ -291,6 +312,8 @@ export const DEEPSEEK_V4_PRO: ModelSpec = {
 
 export const DEEPSEEK_FLASH: ModelSpec = {
   id: "deepseek-flash",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
   name: "DeepSeek V4.1 Flash",
   api: "openai-completions",
   provider: "deepseek",
@@ -356,10 +379,12 @@ function responsesModel(spec: Omit<ModelSpec, "api">): ModelSpec {
  * *rates* (multiples) stay omitted unless both cache dollars are present,
  * except DeepSeek (verified against its own page).
  */
-export const CATALOG: readonly ModelSpec[] = [
+export const CURATED_CATALOG: readonly ModelSpec[] = [
   CLAUDE_OPUS_5,
   anthropicModel({
     id: "claude-fable-5-1",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Claude Fable 5.1",
     reasoning: true,
     cache: "explicit",
@@ -371,6 +396,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   responsesModel({
     id: "gpt-5.5-pro",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "GPT-5.5 Pro",
     provider: "openai",
     baseUrl: "https://api.openai.com/v1",
@@ -382,6 +409,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   responsesModel({
     id: "gpt-5.4-pro",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "GPT-5.4 Pro",
     provider: "openai",
     baseUrl: "https://api.openai.com/v1",
@@ -394,6 +423,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "gemini-3.1-pro-preview",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Gemini 3.1 Pro Preview",
     provider: "google",
     baseUrl: "https://generativelanguage.googleapis.com/v1beta/openai",
@@ -411,6 +442,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "grok-4.6",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Grok 4.6",
     provider: "xai",
     baseUrl: "https://api.x.ai/v1",
@@ -422,6 +455,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "grok-4.5",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Grok 4.5",
     provider: "xai",
     baseUrl: "https://api.x.ai/v1",
@@ -434,6 +469,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "llama-3.3-70b-versatile",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "Llama 3.3 70B",
     provider: "groq",
     baseUrl: "https://api.groq.com/openai/v1",
@@ -444,6 +481,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "qwen/qwen3.6-27b",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Qwen3.6 27B",
     provider: "groq",
     baseUrl: "https://api.groq.com/openai/v1",
@@ -455,6 +494,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "gpt-oss-120b",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "GPT OSS 120B",
     provider: "cerebras",
     baseUrl: "https://api.cerebras.ai/v1",
@@ -465,6 +506,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "qwen-3.8-27b",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Qwen3.8 27B",
     provider: "cerebras",
     baseUrl: "https://api.cerebras.ai/v1",
@@ -476,6 +519,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "magistral-medium-latest",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "Magistral Medium (latest)",
     provider: "mistral",
     baseUrl: "https://api.mistral.ai/v1",
@@ -486,6 +531,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "mistral-large-2411",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "Mistral Large 2.1",
     provider: "mistral",
     baseUrl: "https://api.mistral.ai/v1",
@@ -497,6 +544,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "openai/gpt-5.5-pro",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "GPT-5.5 Pro",
     provider: "openrouter",
     baseUrl: "https://openrouter.ai/api/v1",
@@ -508,6 +557,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "openai/gpt-5.4-pro",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "GPT-5.4 Pro",
     provider: "openrouter",
     baseUrl: "https://openrouter.ai/api/v1",
@@ -520,6 +571,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "glm-5.3",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "GLM-5.3",
     provider: "zai",
     baseUrl: "https://api.z.ai/api/paas/v4",
@@ -532,6 +585,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "glm-5.2",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text"],
     name: "GLM-5.2",
     provider: "zai",
     baseUrl: "https://api.z.ai/api/paas/v4",
@@ -545,6 +600,8 @@ export const CATALOG: readonly ModelSpec[] = [
 
   completionsModel({
     id: "kimi-k3",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Kimi K3",
     provider: "moonshotai",
     baseUrl: "https://api.moonshot.ai/v1",
@@ -557,6 +614,8 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
   completionsModel({
     id: "kimi-k2.6",
+  // Input modalities from the existing models.dev catalog source, 2026-09-30.
+  input: ["text","image"],
     name: "Kimi K2.6",
     provider: "moonshotai",
     baseUrl: "https://api.moonshot.ai/v1",
@@ -594,6 +653,10 @@ export const CATALOG: readonly ModelSpec[] = [
   }),
 ];
 
+// Keep existing default/provider order and verified wire dialects; add the source union once.
+export const CATALOG: readonly ModelSpec[] = mergeCatalogs(CURATED_CATALOG,
+  SOURCE_MODELS.filter(model => !CURATED_CATALOG.some(curated => curated.provider === model.provider && curated.id === model.id)));
+
 /**
  * Output tokens to request, which is also the output reserve every window
  * budget subtracts.

package/src/runtime/models-fetch.ts

@@ -9,7 +9,7 @@
  * starting".
  */
 
-import { PROVIDER_WIRE, type Api, type ModelSpec, type Price } from "./model.ts";
+import { modelsDevCatalog, unionSourceCatalog } from "./model-sources.ts";
 
 export const MODELS_DEV_URL = "https://models.dev/api.json";
 
@@ -43,113 +43,15 @@ export async function refreshRemoteCatalog(opts: {
   } catch (err) {
     return { ok: false, reason: `could not refresh models: body is not JSON (${(err as Error).message})` };
   }
-  const models = modelsFromRemote(json);
-  if (models.length === 0) {
+  if (modelsDevCatalog(json).models.length === 0) {
     return { ok: false, reason: `could not refresh models: ${url} named no provider we can speak` };
   }
+  const catalog = unionSourceCatalog(json);
   try {
-    await opts.write({ models: models.map(serializeSpec) });
+    // Provider metadata is not the user's executable provider configuration.
+    await opts.write({ schemaVersion: 1, models: catalog.models, providerMetadata: catalog.providers, excluded: catalog.excluded });
   } catch (err) {
     return { ok: false, reason: `could not write models: ${(err as Error).message}` };
   }
-  return { ok: true, count: models.length };
-}
-
-function modelsFromRemote(json: unknown): ModelSpec[] {
-  if (typeof json !== "object" || json === null) return [];
-  const out: ModelSpec[] = [];
-  for (const [rawId, value] of Object.entries(json as Record<string, unknown>)) {
-    const providerId = rawId === "togetherai" ? "together" : rawId;
-    const wire = PROVIDER_WIRE[rawId] ?? PROVIDER_WIRE[providerId];
-    if (wire === undefined) continue;
-    if (typeof value !== "object" || value === null) continue;
-    const provider = value as Record<string, unknown>;
-    const models = provider["models"];
-    if (typeof models !== "object" || models === null) continue;
-    for (const model of Object.values(models as Record<string, unknown>)) {
-      const spec = specFromRemote(model, providerId, wire);
-      if (spec !== null) out.push(spec);
-    }
-  }
-  return out;
-}
-
-function specFromRemote(
-  raw: unknown,
-  provider: string,
-  wire: { api: Api; baseUrl?: string },
-): ModelSpec | null {
-  if (typeof raw !== "object" || raw === null) return null;
-  const m = raw as Record<string, unknown>;
-  if (m["tool_call"] === false) return null;
-  if (typeof m["id"] !== "string" || m["id"].length === 0) return null;
-  const cost = m["cost"];
-  const limit = m["limit"];
-  const ctx = isRecord(limit) && typeof limit["context"] === "number" ? limit["context"] : undefined;
-  const max = isRecord(limit) && typeof limit["output"] === "number" ? limit["output"] : undefined;
-  const price = priceFromCost(cost);
-  const cache = cacheFromCost(provider, cost);
-  const spec: ModelSpec = {
-    id: m["id"],
-    name: typeof m["name"] === "string" && m["name"].length > 0 ? m["name"] : m["id"],
-    api: wire.api,
-    provider,
-    reasoning: m["reasoning"] === true,
-    ...(wire.baseUrl !== undefined ? { baseUrl: wire.baseUrl } : {}),
-    ...(ctx !== undefined ? { contextWindow: ctx } : {}),
-    ...(max !== undefined ? { maxTokens: max } : {}),
-    ...(price !== undefined ? { price } : {}),
-    ...(cache !== undefined ? cache : {}),
-  };
-  return spec;
-}
-
-function priceFromCost(cost: unknown): Price | undefined {
-  if (!isRecord(cost)) return undefined;
-  const input = cost["input"];
-  const output = cost["output"];
-  if (typeof input !== "number" || typeof output !== "number") return undefined;
-  // Cache dollars are optional. Missing is missing — a 0 here would be
-  // DeepSeek's term, not a default for everyone else.
-  const price: Price = { input, output };
-  if (typeof cost["cache_read"] === "number") price.cacheRead = cost["cache_read"];
-  if (typeof cost["cache_write"] === "number") price.cacheWrite = cost["cache_write"];
-  return price;
-}
-
-function cacheFromCost(
-  provider: string,
-  cost: unknown,
-): { cache: "explicit" | "automatic"; rates: { read: number; write: number } } | undefined {
-  if (!isRecord(cost)) return undefined;
-  const input = cost["input"];
-  const cacheRead = cost["cache_read"];
-  const cacheWrite = cost["cache_write"];
-  if (typeof input !== "number" || input === 0) return undefined;
-  if (typeof cacheRead !== "number" || typeof cacheWrite !== "number") return undefined;
-  return {
-    cache: provider === "anthropic" ? "explicit" : "automatic",
-    rates: { read: cacheRead / input, write: cacheWrite / input },
-  };
-}
-
-function serializeSpec(spec: ModelSpec): Record<string, unknown> {
-  return {
-    id: spec.id,
-    name: spec.name,
-    api: spec.api,
-    provider: spec.provider,
-    reasoning: spec.reasoning,
-    ...(spec.baseUrl !== undefined ? { baseUrl: spec.baseUrl } : {}),
-    ...(spec.cache !== undefined ? { cache: spec.cache } : {}),
-    ...(spec.rates !== undefined ? { rates: spec.rates } : {}),
-    ...(spec.price !== undefined ? { price: spec.price } : {}),
-    ...(spec.contextWindow !== undefined ? { contextWindow: spec.contextWindow } : {}),
-    ...(spec.maxTokens !== undefined ? { maxTokens: spec.maxTokens } : {}),
-    ...(spec.compat !== undefined ? { compat: spec.compat } : {}),
-  };
-}
-
-function isRecord(v: unknown): v is Record<string, unknown> {
-  return typeof v === "object" && v !== null && !Array.isArray(v);
+  return { ok: true, count: catalog.models.length };
 }

package/src/runtime/oauth.ts

@@ -17,7 +17,7 @@ import { createServer, type Server } from "node:http";
 import { pollDeviceCode } from "./oauth-device.ts";
 
 /** Matches package.json. Sent as User-Agent so the server can see who we are (D18). */
-export const XENO_VERSION = "0.1.4";
+export const XENO_VERSION = "0.2.0";
 
 /**
  * Minutes subtracted from server expiry so a request issued at the edge
@@ -31,7 +31,7 @@ export const SUBSCRIPTION_NOTICE =
 
 export const SUBSCRIPTION_OPTIONS = [
   { id: "claude-pro", provider: "anthropic", label: "Claude Pro/Max" },
-  { id: "codex", provider: "openai", label: "ChatGPT Plus/Pro (Codex)" },
+  { id: "codex", provider: "openai-codex", label: "ChatGPT Plus/Pro (Codex)" },
   { id: "github-copilot", provider: "github-copilot", label: "GitHub Copilot" },
 ] as const;
 
@@ -501,6 +501,7 @@ export async function refreshOAuthCredential(
     case "anthropic":
       return refreshAnthropic(credential.refresh, opts);
     case "openai":
+    case "openai-codex":
       return refreshOpenAI(credential.refresh, opts);
     case "github-copilot":
       return refreshCopilot(credential.refresh, opts, credential.enterpriseUrl);
@@ -748,7 +749,7 @@ export async function loginSubscription(
     case "claude-pro":
       return { provider: "anthropic", credential: await loginAnthropic(interaction) };
     case "codex":
-      return { provider: "openai", credential: await loginOpenAICodex(interaction) };
+      return { provider: "openai-codex", credential: await loginOpenAICodex(interaction) };
     case "github-copilot":
       return { provider: "github-copilot", credential: await loginGitHubCopilot(interaction) };
     default: {