[+] Charmland Hyper /models format

packages/gateway/src/repo/models-cache-contract.ts

@@ -1,4 +1,4 @@
 // Persisted ProviderModel rows contain code-derived metadata as well as the
 // upstream response. Increment this whenever that derived catalog contract or
 // its serialization changes so older rows become cold across deployments.
-export const MODEL_CATALOG_REVISION = 5;
+export const MODEL_CATALOG_REVISION = 6;

packages/provider-custom/src/fetch-models.ts

@@ -1,17 +1,26 @@
-// Custom-upstream /models response parser. Permissively accepts the three
-// shapes our `custom` provider needs to interoperate with:
-//   1. OpenAI:       { object: 'list', data: [{ id, object?, owned_by?, created? }] }
+// Custom-upstream /models response parser. Permissively accepts the
+// OpenAI, Anthropic, and Floway supersets the `custom` provider needs to
+// interoperate with:
+//   1. OpenAI:       { object: 'list', data: [{ id, object?, owned_by?, created?,
+//                      context_window?, max_output_tokens?,
+//                      pricing?: { input?, output?, cache_create?, cache_hit? } }] }
 //   2. Anthropic:    { data: [{ type: 'model', id, display_name?, created_at? }],
 //                      has_more, first_id, last_id }     (no top-level `object`)
 //   3. OpenAI/Anthropic superset with optional display_name, created_at,
 //      limits, pricing, kind on the model and a `data` array on the container.
 //
-// A model is admitted if it has a string `id`; everything else is best-
-// effort metadata. The container is admitted if `data` is an array.
+// OpenAI-compatible gateways publish model metadata at the model top level
+// (shape 1) rather than in our nested shapes:
+//   • `context_window` / `max_output_tokens`   → `limits` fields
+//   • `pricing: { input, output, ... }`        → per-token ModelPricing
+//     (dollars per million tokens; see perMillionTokenRates)
+//   • `capabilities.vision` / `reasoning`      → `chat` modalities / effort
+// A model is admitted if it has a string `id`; everything else is best-effort
+// metadata. The container is admitted if `data` is an array.
 
 import type { CustomUpstreamConfig } from './config.ts';
 import { customFetchModels } from './fetch.ts';
-import { BILLING_METRICS, canonicalizePricingSelector, type BillingMetric, type ModelKind, type ModelPricing, parseNonNegativeDecimalString, type PriceVector, type PricingSelector, validateModelPricing } from '@floway-dev/protocols/common';
+import { BILLING_METRICS, canonicalizePricingSelector, perMillionTokenRates, type BillingMetric, type ModelKind, type ModelPricing, parseNonNegativeDecimalString, type PriceVector, type PricingSelector, validateModelPricing } from '@floway-dev/protocols/common';
 import { chatField, fetchUpstreamModels, type Fetcher, type UpstreamChatModelConfig, identityWrapUpstreamCall } from '@floway-dev/provider';
 
 export interface CustomRawModel {
@@ -61,12 +70,45 @@ const parseLimits = (value: unknown): CustomRawModel['limits'] => {
   return Object.keys(limits).length > 0 ? limits : undefined;
 };
 
+// Flat OpenAI-compatible pricing block published in dollars per million tokens,
+// as documented for gateway-style upstreams at
+// https://hyper.charm.land/docs/api/list-models.html (`pricing`).
+const FLAT_PRICING_METRICS: Readonly<Record<string, BillingMetric>> = {
+  input: 'input_tokens',
+  output: 'output_tokens',
+  cache_create: 'input_cache_write_tokens',
+  cache_hit: 'input_cache_read_tokens',
+};
+
+const parseFlatPricing = (value: Record<string, unknown>): ModelPricing | undefined => {
+  try {
+    if (Object.keys(value).some(key => !(key in FLAT_PRICING_METRICS))) return undefined;
+    const published: PriceVector = {};
+    for (const [key, raw] of Object.entries(value)) {
+      if (typeof raw !== 'number' || !Number.isFinite(raw)) return undefined;
+      published[FLAT_PRICING_METRICS[key]!] = parseNonNegativeDecimalString(String(raw), `flat pricing ${key}`);
+    }
+    if (Object.keys(published).length === 0) return undefined;
+    const pricing: ModelPricing = { entries: [{ rates: perMillionTokenRates(published) }] };
+    validateModelPricing(pricing);
+    return pricing;
+  } catch {
+    return undefined;
+  }
+};
+
 const parsePricing = (value: unknown): ModelPricing | undefined => {
   // Pricing is best-effort catalog metadata: malformed pricing omits only the pricing
   // block, never the enclosing model or the rest of the catalog.
-  if (!isRecord(value) || !Array.isArray(value.entries)) return undefined;
+  if (!isRecord(value)) return undefined;
+  if ('entries' in value) return parseEntryPricing(value);
+  return parseFlatPricing(value);
+};
+
+const parseEntryPricing = (value: Record<string, unknown>): ModelPricing | undefined => {
   try {
     if (Object.keys(value).some(key => key !== 'entries')) throw new TypeError('Malformed pricing block');
+    if (!Array.isArray(value.entries)) throw new TypeError('Malformed pricing block');
     const entries: ModelPricing['entries'][number][] = [];
     for (const rawEntry of value.entries) {
       if (!isRecord(rawEntry) || !isRecord(rawEntry.rates)) throw new TypeError('Malformed pricing entry');
@@ -97,6 +139,52 @@ const parseKind = (value: unknown): ModelKind | undefined => {
   return undefined;
 };
 
+// OpenAI-compatible top-level chat metadata: `capabilities.vision` turns on
+// image input, and `reasoning.effort_levels` / `reasoning.default_effort_level`
+// (a `[{value, display}]` list with one display string per level) project onto
+// our unified chat `effort` shape. Best-effort: anything malformed contributes
+// nothing rather than dropping the model.
+const parseOpenAICompatReasoningEffort = (reasoning: Record<string, unknown>): UpstreamChatModelConfig['reasoning'] | undefined => {
+  try {
+    if (!Array.isArray(reasoning.effort_levels) || reasoning.effort_levels.length === 0) return undefined;
+    const supported: string[] = [];
+    for (const raw of reasoning.effort_levels) {
+      if (!isRecord(raw) || typeof raw.value !== 'string' || raw.value === '') return undefined;
+      if (!supported.includes(raw.value)) supported.push(raw.value);
+    }
+    if (typeof reasoning.default_effort_level !== 'string' || !supported.includes(reasoning.default_effort_level)) return undefined;
+    return { effort: { supported, default: reasoning.default_effort_level } };
+  } catch {
+    return undefined;
+  }
+};
+
+const parseOpenAICompatChat = (value: Record<string, unknown>): UpstreamChatModelConfig | undefined => {
+  const out: UpstreamChatModelConfig = {};
+  if (isRecord(value.capabilities) && value.capabilities.vision === true) {
+    out.modalities = { input: ['text', 'image'], output: ['text'] };
+  }
+  if (isRecord(value.reasoning)) {
+    const reasoning = parseOpenAICompatReasoningEffort(value.reasoning);
+    if (reasoning !== undefined) out.reasoning = reasoning;
+  }
+  return out.modalities === undefined && out.reasoning === undefined ? undefined : out;
+};
+
+// Explicit Floway-shaped `chat` metadata wins per-field; OpenAI-compat
+// top-level derivation fills whatever the explicit block leaves out.
+const mergeChat = (explicit: UpstreamChatModelConfig | undefined, derived: UpstreamChatModelConfig | undefined): UpstreamChatModelConfig | undefined => {
+  if (explicit === undefined) return derived;
+  if (derived === undefined) return explicit;
+  const out: UpstreamChatModelConfig = {};
+  if (explicit.modalities !== undefined) out.modalities = explicit.modalities;
+  else if (derived.modalities !== undefined) out.modalities = derived.modalities;
+  if (explicit.reasoning !== undefined || derived.reasoning !== undefined) {
+    out.reasoning = { ...derived.reasoning, ...explicit.reasoning };
+  }
+  return out.modalities === undefined && out.reasoning === undefined ? undefined : out;
+};
+
 const parseRawModel = (value: unknown): CustomRawModel | null => {
   if (!isRecord(value)) return null;
   if (typeof value.id !== 'string' || value.id === '') return null;
@@ -111,16 +199,23 @@ const parseRawModel = (value: unknown): CustomRawModel | null => {
   if (name !== undefined) model.name = name;
   const owned_by = optionalStringField(value.owned_by);
   if (owned_by !== undefined) model.owned_by = owned_by;
-  const limits = parseLimits(value.limits);
-  if (limits !== undefined) model.limits = limits;
+  // OpenAI-compat gateways publish context_window / max_output_tokens at the
+  // model top level instead of inside a `limits` object; the nested limits
+  // (Floway superset) win when both are present.
+  const limits = parseLimits(value.limits) ?? {};
+  const context_window = optionalNumberField(value.context_window);
+  if (context_window !== undefined && limits.max_context_window_tokens === undefined) limits.max_context_window_tokens = context_window;
+  const max_output_tokens = optionalNumberField(value.max_output_tokens);
+  if (max_output_tokens !== undefined && limits.max_output_tokens === undefined) limits.max_output_tokens = max_output_tokens;
+  if (Object.keys(limits).length > 0) model.limits = limits;
   const pricing = parsePricing(value.pricing);
   if (pricing !== undefined) model.pricing = pricing;
   const kind = parseKind(value.kind);
   if (kind !== undefined) model.kind = kind;
   // Attempt to parse chat metadata; silently skip on malformed data.
   try {
-    const chat = chatField(value.chat, `${value.id}.chat`);
-    if (chat !== undefined) model.chat = chat;
+    const explicit = chatField(value.chat, `${value.id}.chat`);
+    model.chat = mergeChat(explicit, parseOpenAICompatChat(value));
   } catch { /* skip */ }
   return model;
 };

packages/provider-custom/src/provider.ts

@@ -32,6 +32,7 @@ const customRawToProviderModel = (model: CustomRawModel): Omit<ProviderModel, 'k
   const display = model.display_name ?? model.name;
   if (display !== undefined) partial.display_name = display;
   if (model.pricing) partial.pricing = model.pricing;
+  if (model.chat) partial.chat = model.chat;
   return partial;
 };