pi-inferhub 0.3.0 → 0.4.0

package/README.md

@@ -98,9 +98,14 @@ combo to one model:
   only shows an inert one. Where members advertise `reasoning_levels`, Pi's
   levels map onto that union; when none advertise a ladder the map is omitted
   so every level stays selectable.
+- **Surface** — each combo rides the API its membership implies:
+  **Anthropic messages** for Claude-only combos, **[OI] responses** for
+  GPT-only combos, and **chat completions** for everything else — including
+  combos that mix both native families. Family is decided from the model name,
+  not the serving prefix.
 - **Cost** — per-direction minimum of resolvable members' *official* rates
-  (see below). `cacheRead`/`cacheWrite` are shown as `0`: the catalog carries
-  no per-combo cache pricing, and none is fabricated.
+  (see below). Cache reads bill at **0.1x the input rate** (owner-confirmed);
+  cache writes are left at `0` because no write rate is published.
 
 Member resolution is conservative but forgiving: a member resolves directly
 from `/v1/models` when its prefixed row carries data; otherwise it joins the
@@ -130,6 +135,27 @@ member's *official* list rates (`official_in` / `official_out`) per direction.
 InferHub's marketplace bills below list, so the real `usage.cost` reported per
 request comes in **at or under** the displayed estimate.
 
+Cached input is priced at **0.1x the input rate**, the rate InferHub charges
+(confirmed by its owner). On the Anthropic surface this is not theoretical:
+sending the same prefix twice cut a request's real cost by roughly 16x in
+testing, because the second call billed its 1218-token prefix as a cache read
+instead of fresh input.
+
+## Surfaces
+
+InferHub exposes three compatible APIs. The extension picks per model:
+
+| membership | API | endpoint |
+|---|---|---|
+| Claude models only | Anthropic messages | `api.inferhub.dev/v1/messages` |
+| GPT models only | [OI] responses | `api.inferhub.dev/v1/responses` |
+| both families, or neither | chat completions | `api.inferhub.dev/v1/chat/completions` |
+
+This buys native behaviour where it is unambiguous: structured thinking
+control, `cache_control` breakpoints, and the responses surface's typed tool
+calls. Mixed combos deliberately fall back to chat completions, which InferHub
+translates for every model.
+
 ## Design rationale
 
 The full design — why combos only, why `/v1/models` supplies member metadata

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "pi-inferhub",
-	"version": "0.3.0",
+	"version": "0.4.0",
 	"description": "Pi extension for the InferHub inference provider",
 	"license": "MIT",
 	"keywords": [

package/src/catalog.ts

@@ -1,14 +1,55 @@
-import type { Model, ThinkingLevel, ThinkingLevelMap } from "@earendil-works/pi-ai";
+import type { Api, Model, ThinkingLevel, ThinkingLevelMap } from "@earendil-works/pi-ai";
 import { Type, type Static } from "typebox";
 import { Compile } from "typebox/compile";
 import { fetchJson, type FetchJsonOptions } from "./http.js";
-import { INFERENCE_BASE_URL, MGMT_BASE_URL, PROVIDER_NAME } from "./inferhub.js";
+import { ANTHROPIC_BASE_URL, INFERENCE_BASE_URL, MGMT_BASE_URL, PROVIDER_NAME } from "./inferhub.js";
 
 /** pi thinking levels, in ladder order; `off` is handled separately. */
 const PI_THINKING_LEVELS = ["minimal", "low", "medium", "high", "xhigh", "max"] as const satisfies readonly ThinkingLevel[];
 
 const CONTEXT_WINDOW_FALLBACK = 262_144;
 const MAX_TOKENS_FALLBACK = 131_072;
+/**
+ * Owner-confirmed 2026-10-18 (and matching the marketplace cache rule): InferHub
+ * bills cached input at 0.1x the input rate. Cache WRITES have no published
+ * rate, so they stay 0 rather than being guessed.
+ */
+const CACHE_READ_MULTIPLIER = 0.1;
+
+type ChatModel = Model<Api>;
+
+/** Which native surface a model name belongs to. */
+export type NativeFamily = "anthropic" | "openai" | "other";
+
+/**
+ * Family of the model NAME (last path segment), never the serving prefix:
+ * live catalog 2026-10-18 shows `kiro/` serving both claude-* and gpt-* rows.
+ * `gpt-oss` is an open-weight model, not an [OI] native one, so it stays other.
+ */
+export function nativeFamily(id: string): NativeFamily {
+	const name = (id.split("/").pop() ?? id).toLowerCase();
+	if (name.includes("claude")) return "anthropic";
+	if (/^(gpt(?!-oss)|o\d|codex)/.test(name)) return "openai";
+	return "other";
+}
+
+interface Surface {
+	api: Api;
+	baseUrl: string;
+}
+
+/**
+ * One native family present → its own surface. Both native families present
+ * (or none) → chat completions, which InferHub translates for every model.
+ */
+export function selectSurface(names: readonly string[]): Surface {
+	const families = new Set(names.map(nativeFamily));
+	const anthropic = families.has("anthropic");
+	const openai = families.has("openai");
+	if (anthropic && !openai) return { api: "anthropic-messages", baseUrl: ANTHROPIC_BASE_URL };
+	if (openai && !anthropic) return { api: "openai-responses", baseUrl: INFERENCE_BASE_URL };
+	return { api: "openai-completions", baseUrl: INFERENCE_BASE_URL };
+}
 
 export const MemberSchema = Type.Union([
 	Type.Object({ kind: Type.Literal("model"), model: Type.String() }),
@@ -72,7 +113,7 @@ export async function fetchComboCatalog(
 	token: string,
 	signal?: AbortSignal,
 	opts?: FetchCatalogOptions,
-): Promise<Model<"openai-completions">[]> {
+): Promise<ChatModel[]> {
 	const [combos, models] = await Promise.all([
 		fetchJson<Combo[]>(`${MGMT_BASE_URL}/combos`, token, CombosValidator, signal, opts),
 		fetchJson<Static<typeof ModelsEnvelope>>(`${INFERENCE_BASE_URL}/models`, token, ModelsValidator, signal, opts),
@@ -87,7 +128,7 @@ export async function fetchComboCatalog(
 export function combosToModels(
 	combos: Combo[],
 	catalogModels: CatalogModel[],
-): Model<"openai-completions">[] {
+): ChatModel[] {
 	const catalogById = new Map(catalogModels.map((model) => [model.id, model]));
 	return combos.map((combo) => comboToModel(combo, catalogById, catalogModels));
 }
@@ -96,7 +137,7 @@ function comboToModel(
 	combo: Combo,
 	catalogById: Map<string, CatalogModel>,
 	catalogModels: CatalogModel[],
-): Model<"openai-completions"> {
+): ChatModel {
 	// Alias rows (`owned_by: "alias"`, no data) and alias-kind members resolve
 	// through prefixed sibling servings of the same model name; min-of over the
 	// siblings bounds the combo exactly like directly-joined members.
@@ -136,7 +177,8 @@ function comboToModel(
 	// including level-less resolved members (live-verified 2026-10-18 on
 	// combo/deepseek-v4-pro: no advertised levels, honors reasoning_effort).
 	const name = combo.name.trim().length > 0 ? combo.name : combo.slug;
-	return buildModel(`combo/${combo.slug}`, name, resolvable);
+	const surface = selectSurface(combo.members.map((member) => (member.kind === "model" ? member.model : member.alias)));
+	return buildModel(`combo/${combo.slug}`, name, resolvable, surface);
 }
 
 /**
@@ -144,7 +186,7 @@ function comboToModel(
  * Skipped: alias rows (information-free duplicates of a prefixed serving),
  * image-output rows (not chat models), and blank ids (live shape `cc/`).
  */
-export function catalogToModels(catalogModels: CatalogModel[]): Model<"openai-completions">[] {
+export function catalogToModels(catalogModels: CatalogModel[]): ChatModel[] {
 	return catalogModels
 		.filter(
 			(row) =>
@@ -159,16 +201,12 @@ export function catalogToModels(catalogModels: CatalogModel[]): Model<"openai-co
 		)
 		.map((row) => {
 			const label = row.upstream_label?.trim();
-			return buildModel(row.id, label && label.length > 0 ? label : row.id, [row]);
+			return buildModel(row.id, label && label.length > 0 ? label : row.id, [row], selectSurface([row.id]));
 		});
 }
 
 /** Shared pi-model construction for combo and direct-catalog entries. */
-function buildModel(
-	id: string,
-	name: string,
-	sources: CatalogModel[],
-): Model<"openai-completions"> {
+function buildModel(id: string, name: string, sources: CatalogModel[], surface: Surface): ChatModel {
 	const advertised = new Set<string>();
 	for (const model of sources) {
 		for (const level of model.reasoning_levels ?? []) advertised.add(level);
@@ -177,50 +215,77 @@ function buildModel(
 	const image = sources.length > 0 && sources.every((model) => (model.modality ?? "").includes("image"));
 
 	const priced = sources.flatMap((model) => (model.pricing ? [model.pricing] : []));
+	const inputRate = minOf(priced.map((pricing) => pricing.official_in)) ?? 0;
 
-	const model: Model<"openai-completions"> = {
+	const model: ChatModel = {
 		id,
 		name,
-		api: "openai-completions",
+		api: surface.api,
 		provider: PROVIDER_NAME,
-		baseUrl: INFERENCE_BASE_URL,
+		baseUrl: surface.baseUrl,
 		input: image ? ["text", "image"] : ["text"],
 		cost: {
-			input: minOf(priced.map((pricing) => pricing.official_in)) ?? 0,
+			input: inputRate,
 			output: minOf(priced.map((pricing) => pricing.official_out)) ?? 0,
-			cacheRead: 0,
+			cacheRead: roundRate(inputRate * CACHE_READ_MULTIPLIER),
 			cacheWrite: 0,
 		},
 		reasoning: true,
 		contextWindow: minPositiveOf(sources.map((source) => source.input_token_limit)) ?? CONTEXT_WINDOW_FALLBACK,
 		maxTokens: minPositiveOf(sources.map((source) => source.max_output_tokens)) ?? MAX_TOKENS_FALLBACK,
+	};
+
+	if (surface.api === "openai-completions") {
 		// InferHub's [OI]-compatible surface ignores max_completion_tokens (verified
 		// live 2026-10-18: a 20-token cap sent that way produced 107 tokens; the
 		// same cap as max_tokens stopped at 20). pi-ai auto-detects per-provider
 		// and inferhub is not on its list, so pin the field explicitly.
-		compat: { maxTokensField: "max_tokens" },
-	};
+		model.compat = { maxTokensField: "max_tokens" };
+	}
+
 	// thinkingLevelMap: built only from an advertised union; a level-less union
-	// omits the map so pi defaults keep thinking levels selectable — marking
-	// them all `null` would hide the knob entirely.
-	if (advertised.size > 0) model.thinkingLevelMap = buildThinkingLevelMap(advertised);
+	// omits the level keys so pi defaults keep thinking levels selectable —
+	// marking them all `null` would hide the knob entirely.
+	if (surface.api === "anthropic-messages") {
+		// These Claude models reject thinking:{type:"disabled"} (400, live
+		// 2026-10-18) and require adaptive thinking. `off: null` makes pi-ai omit
+		// the field instead; the map is therefore always present, even when no
+		// ladder is advertised.
+		model.compat = { forceAdaptiveThinking: true };
+		// With no advertised ladder only `off` is pinned: the level keys stay
+		// absent so pi keeps every level selectable (optimistic by design).
+		model.thinkingLevelMap = advertised.size > 0 ? buildThinkingLevelMap(advertised, null) : { off: null };
+	} else if (surface.api === "openai-responses") {
+		// [OI] reasoning-off is effort "none"; the value "off" is accepted but
+		// leaves the effort unset (live-verified 2026-10-18).
+		if (advertised.size > 0) model.thinkingLevelMap = buildThinkingLevelMap(advertised, "none");
+		else model.thinkingLevelMap = { off: "none" };
+	} else if (advertised.size > 0) {
+		model.thinkingLevelMap = buildThinkingLevelMap(advertised, "off");
+	}
+
 	return model;
 }
 
-/** Last path segment, lowercased, alphanumerics only: "cp/cline-pass/Kimi-K3" → "kimik3". */
-function normalizeModelName(id: string): string {
-	return (id.split("/").pop() ?? id).toLowerCase().replace(/[^a-z0-9]/g, "");
+/** Cache rates are derived; trim float noise (3 * 0.1 = 0.30000000000000004). */
+function roundRate(value: number): number {
+	return Math.round(value * 1e6) / 1e6;
 }
 
 /** Each pi level maps to itself when the combo advertises it, else `null`. */
-function buildThinkingLevelMap(advertised: ReadonlySet<string>): ThinkingLevelMap {
-	const map: ThinkingLevelMap = { off: "off" };
+function buildThinkingLevelMap(advertised: ReadonlySet<string>, off: "off" | "none" | null): ThinkingLevelMap {
+	const map: ThinkingLevelMap = { off };
 	for (const level of PI_THINKING_LEVELS) {
 		map[level] = advertised.has(level) ? level : null;
 	}
 	return map;
 }
 
+/** Last path segment, lowercased, alphanumerics only: "cp/cline-pass/Kimi-K3" → "kimik3". */
+function normalizeModelName(id: string): string {
+	return (id.split("/").pop() ?? id).toLowerCase().replace(/[^a-z0-9]/g, "");
+}
+
 function minOf(values: Array<number | undefined>): number | undefined {
 	const present = values.filter((value): value is number => value !== undefined);
 	return present.length > 0 ? Math.min(...present) : undefined;

package/src/index.ts

@@ -1,7 +1,7 @@
-import { createProvider, type Provider } from "@earendil-works/pi-ai";
+import { createProvider, type Api, type Provider } from "@earendil-works/pi-ai";
 import type { ExtensionAPI, ExtensionCommandContext } from "@earendil-works/pi-coding-agent";
 // Pi's extension loader doesn't virtualize lazy API implementation subpaths yet
-import { openAICompletionsApi } from "@earendil-works/pi-ai/compat";
+import { anthropicMessagesApi, openAICompletionsApi, openAIResponsesApi } from "@earendil-works/pi-ai/compat";
 import { inferhubApiKeyAuth, loginInferhub } from "./auth.js";
 import { fetchComboCatalog } from "./catalog.js";
 import { type InferhubConfig, loadInferhubConfig } from "./config.js";
@@ -23,9 +23,9 @@ export interface InferhubRegistrationOptions {
 	loadConfig?: () => InferhubConfig;
 }
 
-function inferhubProvider(options: InferhubRegistrationOptions = {}): Provider<"openai-completions"> {
+function inferhubProvider(options: InferhubRegistrationOptions = {}): Provider<Api> {
 	const loadConfig = options.loadConfig ?? (() => loadInferhubConfig());
-	return createProvider({
+	return createProvider<Api>({
 		id: PROVIDER_NAME,
 		name: PROVIDER_DISPLAY_NAME,
 		baseUrl: INFERENCE_BASE_URL,
@@ -45,7 +45,15 @@ function inferhubProvider(options: InferhubRegistrationOptions = {}): Provider<"
 			if (!token) return [];
 			return fetchComboCatalog(token, context.signal, { publishUpstream: loadConfig().publishUpstream });
 		},
-		api: openAICompletionsApi(),
+		api: {
+			// Combos carry the surface their membership selects: anthropic-messages
+			// for Claude-only, openai-responses for GPT-only, chat completions for
+			// everything else (including mixed membership). createProvider
+			// dispatches on model.api.
+			"openai-completions": openAICompletionsApi(),
+			"openai-responses": openAIResponsesApi(),
+			"anthropic-messages": anthropicMessagesApi(),
+		},
 	});
 }
 

package/src/inferhub.ts

@@ -1,4 +1,6 @@
 export const PROVIDER_NAME = "inferhub";
 export const PROVIDER_DISPLAY_NAME = "InferHub";
 export const INFERENCE_BASE_URL = "https://api.inferhub.dev/v1";
+/** Anthropic surface lives on the bare host: pi-ai appends /v1/messages itself. */
+export const ANTHROPIC_BASE_URL = "https://api.inferhub.dev";
 export const MGMT_BASE_URL = "https://inferhub.dev/api";