@oh-my-pi/pi-catalog 16.3.2 → 16.3.4

package/CHANGELOG.md

@@ -2,6 +2,29 @@
 
 ## [Unreleased]
 
+## [16.3.4] - 2026-07-03
+
+### Added
+
+- Added Baseten as a supported model provider
+- Added support for new models from Baseten, including DeepSeek V4 Pro and Kimi series
+- Added new Devin agent models: Claude 5 Fable variants
+- Added new Github Copilot models: Kimi K2.7 Code and MAI-Code-1-Flash
+- Added Poolside Laguna XS 2.1 models via Kilo and OpenRouter providers
+- Added support for Claude Fable 5 (Free) via Zenmux provider
+
+### Changed
+
+- Updated priority ordering to include Baseten
+- Updated pricing and limits for various existing models in the catalog
+
+## [16.3.3] - 2026-07-02
+
+### Fixed
+
+- Extended Anthropic-compatible signing-endpoint recognition to Cloudflare AI Gateway, Google Vertex, AWS Bedrock, and Azure AI Inference / Foundry to ensure consistent reasoning-replay and signature-stripping behavior, and exposed ResolvedAnthropicCompat.signingEndpoint in the public API.
+- Fixed Zhipu Coding Plan runtime discovery to prioritize account-scoped model lists over bundled fallback models, preventing routing errors for valid non-z.ai keys.
+
 ## [16.3.2] - 2026-07-02
 
 ### Fixed

package/dist/types/hosts.d.ts

@@ -86,6 +86,10 @@ export declare const KNOWN_HOSTS: {
         readonly providers: readonly ["together"];
         readonly urlMarkers: readonly ["api.together.xyz"];
     };
+    readonly baseten: {
+        readonly providers: readonly ["baseten"];
+        readonly urlMarkers: readonly ["baseten.co"];
+    };
     /** URL-only on purpose: the `fireworks`/`firepass` providers route per-model and not every model is Fireworks-shaped. */
     readonly fireworks: {
         readonly urlMarkers: readonly ["fireworks.ai"];

package/dist/types/provider-models/descriptors.d.ts

@@ -23,6 +23,15 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly catalogDiscovery: {
         readonly label: "Alibaba Coding Plan";
     };
+}, {
+    readonly id: "baseten";
+    readonly defaultModel: "moonshotai/Kimi-K2.7-Code";
+    readonly envVars: readonly ["BASETEN_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "Baseten";
+    };
 }, {
     readonly id: "amazon-bedrock";
     readonly defaultModel: "us.anthropic.claude-opus-4-8";
@@ -414,9 +423,10 @@ export declare const CATALOG_PROVIDERS: readonly [{
     };
 }, {
     readonly id: "zhipu-coding-plan";
-    readonly defaultModel: "glm-5.2";
+    readonly defaultModel: "glm-5.1";
     readonly envVars: readonly ["ZHIPU_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
     readonly catalogDiscovery: {
         readonly label: "Zhipu Coding Plan";
     };

package/dist/types/provider-models/openai-compat.d.ts

@@ -303,6 +303,12 @@ export interface VeniceModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function veniceModelManagerOptions(config?: VeniceModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export interface BasetenModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function basetenModelManagerOptions(config?: BasetenModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface TogetherModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -492,6 +492,17 @@ export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {
      * env headers, and cache-TTL shaping without per-request URL parsing.
      */
     officialEndpoint: boolean;
+    /**
+     * The configured endpoint enforces Anthropic's signature protocol on
+     * replayed thinking blocks — either the official API itself or a proxy
+     * that forwards to it (GitHub Copilot, ZenMux, Cloudflare AI Gateway's
+     * `/anthropic` route, Google Vertex's `publishers/anthropic/…`).
+     * Downstream transforms strip stale cross-model thinking signatures on
+     * these endpoints so the signing proxy doesn't 400 with
+     * `Invalid signature in thinking block` (#4297). Superset of
+     * {@link officialEndpoint}.
+     */
+    signingEndpoint: boolean;
 };
 /**
  * Compatibility settings for the devin-agent (Codeium Cascade) API. Cascade

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.3.2",
+	"version": "16.3.4",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,12 +34,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "16.3.2",
+		"@oh-my-pi/pi-utils": "16.3.4",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.3.2",
+		"@oh-my-pi/pi-ai": "16.3.4",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -36,6 +36,52 @@ function matchesKimiK27CodeFamily(spec: ModelSpec<"anthropic-messages">): boolea
 	return spec.id === "kimi-for-coding" && /k2\.?7 code/i.test(spec.name ?? "");
 }
 
+const CLOUDFLARE_ANTHROPIC_GATEWAY_URL_MARKER = /gateway\.ai\.cloudflare\.com\/.+\/anthropic(?:\/|$)/i;
+const VERTEX_ANTHROPIC_URL_MARKER = /aiplatform\.googleapis\.com\/.+\/publishers\/anthropic\//i;
+const BEDROCK_ANTHROPIC_URL_MARKER = /(?:^|\/\/|\.)bedrock-runtime\.[a-z0-9-]+\.amazonaws\.com/i;
+const AZURE_ANTHROPIC_URL_MARKER = /(?:^|\/\/|\.)[a-z0-9-]+\.(?:inference|services)\.ai\.azure\.com/i;
+
+/**
+ * Cloudflare AI Gateway's `/anthropic` route forwards to signature-enforcing
+ * Anthropic (same failure class as GitHub Copilot #2851 / ZenMux #4192).
+ * Detection is by baseUrl marker rather than provider id: users routinely
+ * declare `provider: "custom"` (or other free-form ids) in `models.yml` for
+ * their own Cloudflare gateway account.
+ */
+function isCloudflareAnthropicGateway(baseUrl?: string): boolean {
+	return baseUrl !== undefined && CLOUDFLARE_ANTHROPIC_GATEWAY_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * Google Vertex's `publishers/anthropic/models/…:streamRawPredict` route
+ * proxies Claude through Google's identity layer and returns full thinking
+ * signatures, so it is a SIGNING endpoint.
+ */
+function isVertexAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && VERTEX_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * AWS Bedrock's Anthropic route (`bedrock-runtime.<region>.amazonaws.com`)
+ * forwards Claude requests through Anthropic's signature protocol. Users can
+ * front Bedrock with a custom `anthropic-messages` provider entry in
+ * `models.yml`; the URL marker makes those signing by default without
+ * requiring a provider-id list.
+ */
+function isBedrockAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && BEDROCK_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * Azure AI Inference / Foundry Anthropic route
+ * (`<resource>.inference.ai.azure.com`, `<resource>.services.ai.azure.com`).
+ * Fronts Claude behind Azure identity and enforces Anthropic signatures on
+ * replay.
+ */
+function isAzureAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && AZURE_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
 /** Build the resolved anthropic-messages compat record for a model spec. */
 export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): ResolvedAnthropicCompat {
 	const baseUrl = spec.baseUrl;
@@ -53,8 +99,17 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 	// (issue #4192).
 	const isZenmux = modelMatchesHost(spec, "zenmux");
 	const requiresThinkingEnabled = modelMatchesHost(spec, "moonshotNative") && matchesKimiK27CodeFamily(spec);
+	const signingEndpoint =
+		official ||
+		isCopilot ||
+		isZenmux ||
+		isCloudflareAnthropicGateway(baseUrl) ||
+		isVertexAnthropicRoute(baseUrl) ||
+		isBedrockAnthropicRoute(baseUrl) ||
+		isAzureAnthropicRoute(baseUrl);
 	const compat: ResolvedAnthropicCompat = {
 		officialEndpoint: official,
+		signingEndpoint,
 		disableStrictTools: false,
 		disableAdaptiveThinking: false,
 		supportsEagerToolInputStreaming: !isCopilot,
@@ -73,26 +128,23 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		// into a class that reads `.id`.
 		requiresToolResultId: isZai,
 		requiresThinkingEnabled,
-		// Official Anthropic enforces signature-based thinking-chain integrity, so
-		// unsigned thinking blocks must stay text there. Anthropic-compatible
-		// reasoning endpoints commonly emit unsigned thinking blocks while still
-		// expecting them back as `type: "thinking"` on continuation; demoting them
-		// loses the reasoning chain and can destabilize the next tool-call
-		// arguments (#2005). Known non-signing hosts (Z.AI, DeepSeek) are also
-		// preserved for compatibility.
+		// Official Anthropic and Anthropic-compatible signing proxies enforce
+		// signature-based thinking-chain integrity, so unsigned thinking blocks
+		// must stay text there. Every other `anthropic-messages` reasoning
+		// endpoint replays unsigned thinking natively so the reasoning chain
+		// survives continuation and doesn't destabilize the next tool-call
+		// arguments (#2005, #2257, #2265, #3288, #3433, #3434). Opaque custom
+		// signing proxies remain the reporter's responsibility to mark with
+		// `compat.replayUnsignedThinking: false`; the transport surfaces a
+		// pointed remediation the first time the signing 400 fires (#4297).
 		//
-		// GitHub Copilot's `anthropic-messages` proxy and ZenMux's Anthropic route
-		// are excluded: both forward to signature-enforcing Anthropic and return
-		// full thinking signatures, so they are SIGNING endpoints. Replaying a
-		// stripped/unsigned thinking block as `signature: ""` there 400s the whole
-		// request ("Invalid signature") — most visibly when a checkpoint/branch-
-		// return turn's end_turn-bound signature is stripped on replay (issues
-		// #2851, #4192). Treating them like official Anthropic degrades such
-		// blocks to text instead, which the API accepts.
-		replayUnsignedThinking:
-			!isCopilot &&
-			!isZenmux &&
-			(isZai || modelMatchesHost(spec, "deepseekFamily") || (spec.reasoning && !official)),
+		// Known signing Anthropic-messages hosts (Copilot, ZenMux, Cloudflare
+		// AI Gateway `/anthropic`, Google Vertex `publishers/anthropic`, AWS
+		// Bedrock `bedrock-runtime.<region>.amazonaws.com`, and Azure
+		// AI Inference / Foundry `<res>.(inference|services).ai.azure.com`)
+		// are excluded automatically because they can be recognised by provider
+		// id or baseUrl marker.
+		replayUnsignedThinking: !signingEndpoint && (Boolean(spec.reasoning) || modelMatchesHost(spec, "deepseekFamily")),
 		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
 	};
 	applyCompatOverrides(compat, spec.compat);

package/src/hosts.ts

@@ -47,6 +47,7 @@ export const KNOWN_HOSTS = {
 	xai: { providers: ["xai"], urlMarkers: ["api.x.ai"] },
 	mistral: { providers: ["mistral"], urlMarkers: ["mistral.ai"] },
 	together: { providers: ["together"], urlMarkers: ["api.together.xyz"] },
+	baseten: { providers: ["baseten"], urlMarkers: ["baseten.co"] },
 	/** URL-only on purpose: the `fireworks`/`firepass` providers route per-model and not every model is Fireworks-shaped. */
 	fireworks: { urlMarkers: ["fireworks.ai"] },
 	groq: { providers: ["groq"], urlMarkers: ["api.groq.com"] },

package/src/identity/priority.ts

@@ -20,6 +20,7 @@ const DEFAULT_MODEL_PROVIDER_ORDER = [
 	// High-quality aggregators / hosted inference providers.
 	"fireworks",
 	"cerebras",
+	"baseten",
 	"openrouter",
 	"aimlapi",
 	"together",