@oh-my-pi/pi-catalog 17.2.2 → 17.2.4

package/CHANGELOG.md

@@ -2,6 +2,25 @@
 
 ## [Unreleased]
 
+## [17.2.4] - 2026-08-01
+
+### Added
+
+- Added `AnthropicCompat.streamIdleTimeoutMs` and propagated it through `buildAnthropicCompat` so direct Anthropic provider streams can configure their inter-event idle watchdog.
+- Fixed Ollama Cloud DeepSeek V4 Pro/Flash models (including dated tag variants such as `deepseek-v4-flash:0731`) reporting an incorrect max-output-tokens figure by pinning it to the deployment's enforced 65536-token output ceiling ([#7266](https://github.com/can1357/oh-my-pi/issues/7266)).
+
+### Fixed
+
+- Fixed `gen:models` Codex discovery to union models across every stored OAuth account and fail closed on partial resolution, matching runtime discovery ([#6265](https://github.com/can1357/oh-my-pi/issues/6265)); restored the bundled `gpt-5.4`, `gpt-5.6-sol`, and `gpt-5.3-codex-spark` entries a single-account regen had dropped.
+- Fixed `google-antigravity` models always reporting $0 cost: Antigravity discovery carries no pricing, so the generator now back-fills each model with its Google list price (Gemini ids from the `google` provider, including `-preview` id aliases; Claude ids from `google-vertex`, falling back to `anthropic`).
+- Fixed OpenRouter `deepseek/deepseek-v4-flash-0731` exposing only `high` thinking effort by consuming the live `reasoning.supported_efforts` and `default_effort` metadata and bundling its `low`/`high`/`max` ladder. ([#7307](https://github.com/can1357/oh-my-pi/issues/7307))
+
+## [17.2.3] - 2026-08-01
+
+### Added
+
+- Added support for the ai& provider (`aiand`), an OpenAI-compatible inference API with dynamic model discovery (context windows, capabilities, reasoning efforts, and USD pricing from `/v1/models`) and API-key authentication via the `AIAND_API_KEY` environment variable.
+
 ## [17.2.2] - 2026-07-31
 
 ### Added

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

@@ -7,6 +7,15 @@
  */
 import type { ModelManagerConfig, ProviderCatalogEntry, ProviderDescriptor } from "./descriptor-types.js";
 export declare const CATALOG_PROVIDERS: readonly [{
+    readonly id: "aiand";
+    readonly defaultModel: "moonshotai/kimi-k2.7-code";
+    readonly envVars: readonly ["AIAND_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "ai&";
+    };
+}, {
     readonly id: "aimlapi";
     readonly defaultModel: "gpt-5.5-2026-04-23";
     readonly envVars: readonly ["AIMLAPI_API_KEY"];

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

@@ -5,5 +5,18 @@ export interface OllamaCloudModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * Output-token ceiling that Ollama Cloud enforces for the DeepSeek V4 Pro/Flash
+ * deployments: `/api/chat` rejects `num_predict` above it with HTTP 400
+ * (`max_tokens (...) exceeds model's maximum output tokens (65536)`) even though
+ * the model pages advertise a 1M context / 384K output. Ollama's `/api/show`
+ * never reports this cap, so the catalog pins it for the affected models
+ * (ollama/ollama#16890, #7266). The wire layer clamps `num_predict` to the same
+ * value (`OLLAMA_CLOUD_NUM_PREDICT_CAP` in `packages/ai/src/providers/ollama.ts`,
+ * #3392/#3394).
+ */
+export declare const OLLAMA_CLOUD_MAX_OUTPUT_TOKENS = 65536;
+/** Whether an Ollama Cloud model id (tagged or not) enforces the 65536 output cap. */
+export declare function isOllamaCloudOutputCapped(id: string): boolean;
 export declare function normalizeOllamaCloudBaseUrl(baseUrl?: string): string;
 export declare function ollamaCloudModelManagerOptions(config?: OllamaCloudModelManagerConfig): ModelManagerOptions<"ollama-chat">;

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

@@ -443,6 +443,23 @@ export interface SakanaModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function sakanaModelManagerOptions(config?: SakanaModelManagerConfig): ModelManagerOptions<"openai-responses">;
+/**
+ * Documented ai& catalog (docs.aiand.com/models/catalog, 2026-08) bundled so
+ * the provider is usable when generation and first boot have no live key.
+ * The org-scoped `/v1/models` response is authoritative once discovery runs.
+ */
+export declare const AIAND_STATIC_MODELS: readonly ModelSpec<"openai-completions">[];
+export interface AiandModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+/**
+ * ai& (aiand.com) model manager: OpenAI-compatible chat completions with an
+ * org-scoped `/v1/models` catalog carrying context, capability, effort, and
+ * pricing metadata, so discovery is authoritative over the bundled seed.
+ */
+export declare function aiandModelManagerOptions(config?: AiandModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface QwenPortalModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -362,6 +362,15 @@ export interface OpenAICompat {
  * that proxy gateways (Vertex AI, AWS Bedrock-style fronts, etc.) reject.
  */
 export interface AnthropicCompat {
+    /**
+     * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts.
+     * Set to 0 to disable the inter-event idle watchdog entirely, matching
+     * `OpenAICompat.streamIdleTimeoutMs`.
+     *
+     * When unset, direct Anthropic streams use `PI_STREAM_IDLE_TIMEOUT_MS`,
+     * then the legacy `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+     */
+    streamIdleTimeoutMs?: number;
     /**
      * Drop the top-level `strict: true` field on tool definitions. Vertex AI's
      * Anthropic-compatible endpoint rejects unknown tool fields with
@@ -605,7 +614,13 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
  */
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs">> & {
+    /**
+     * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts; 0 disables the idle watchdog.
+     * Undefined defers to `PI_STREAM_IDLE_TIMEOUT_MS`, then the legacy
+     * `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then 300s.
+     */
+    streamIdleTimeoutMs?: number;
     /**
      * The configured endpoint is the official first-party Anthropic API
      * (https + exact `api.anthropic.com` host; a missing baseUrl counts as

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "17.2.2",
+	"version": "17.2.4",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,12 +35,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.1",
-		"@oh-my-pi/pi-utils": "17.2.2",
+		"@oh-my-pi/pi-utils": "17.2.4",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.2.2",
+		"@oh-my-pi/pi-ai": "17.2.4",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -172,6 +172,7 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		// id or baseUrl marker.
 		replayUnsignedThinking: !signingEndpoint && (Boolean(spec.reasoning) || modelMatchesHost(spec, "deepseekFamily")),
 		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
+		streamIdleTimeoutMs: spec.compat?.streamIdleTimeoutMs,
 	};
 	applyCompatOverrides(compat, spec.compat);
 	return compat;

package/src/model-thinking.ts

@@ -62,8 +62,8 @@ const GEMINI_3_FLASH_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort.Low, E
 const GPT_5_2_PLUS_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
 const GPT_5_1_CODEX_MINI_EFFORTS: readonly Effort[] = [Effort.Medium, Effort.High];
 const LOW_MEDIUM_HIGH_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
-/** Kimi K3's wire-exact mandatory reasoning scale. */
-const KIMI_K3_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
+/** Wire-exact `low`/`high`/`max` scale used by Kimi K3 and OpenRouter DeepSeek V4 Flash 0731. */
+const LOW_HIGH_MAX_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
 /** Wire-exact two-tier scale (`high`/`max`): GLM-5.2 on Z.ai/Umans/Ollama Cloud/Baseten, Sakana Fugu, DeepSeek. */
 const HIGH_MAX_REASONING_EFFORTS: readonly Effort[] = [Effort.High, Effort.Max];
 /** OpenRouter's DeepSeek route accepts only `high`. */
@@ -339,7 +339,7 @@ function getModelDefinedEfforts<TApi extends Api>(
 		}
 	}
 	if (isKimiK3ModelId(spec.id)) {
-		return KIMI_K3_REASONING_EFFORTS;
+		return LOW_HIGH_MAX_REASONING_EFFORTS;
 	}
 	if (isSakanaFuguReasoningModel(spec)) {
 		return HIGH_MAX_REASONING_EFFORTS;
@@ -366,9 +366,14 @@ function getModelDefinedEfforts<TApi extends Api>(
 		return OLLAMA_REASONING_EFFORTS;
 	}
 	if (isOpenAICompatReasoningApi(spec.api) && isDeepseekReasoningModel(spec)) {
-		// DeepSeek's reasoning_effort accepts only high/max; OpenRouter's
-		// DeepSeek route tops out at high.
-		return isOpenRouterThinkingFormat(compat) ? HIGH_ONLY_REASONING_EFFORTS : HIGH_MAX_REASONING_EFFORTS;
+		// OpenRouter generally exposes only high for DeepSeek, but V4 Flash 0731
+		// advertises and accepts the wire-exact low/high/max ladder.
+		if (isOpenRouterThinkingFormat(compat)) {
+			return bareModelId(spec.id) === "deepseek-v4-flash-0731"
+				? LOW_HIGH_MAX_REASONING_EFFORTS
+				: HIGH_ONLY_REASONING_EFFORTS;
+		}
+		return HIGH_MAX_REASONING_EFFORTS;
 	}
 	if (spec.provider === "baseten" && isOpenAIGptOssModelId(spec.id)) {
 		// Baseten's gpt-oss router mirrors its GLM route: high/max only.