@oh-my-pi/pi-catalog 17.0.1 → 17.0.3

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [17.0.3] - 2026-07-17
+
+### Fixed
+
+- Logged LiteLLM rich-metadata endpoint failures once with their endpoint and status before falling back to incomplete `/v1/models` data ([#5801](https://github.com/can1357/oh-my-pi/issues/5801)).
+- Fixed authenticated Kimi Code discovery to preserve live effort levels, default effort, mandatory-thinking state, and per-model protocol metadata ([#5893](https://github.com/can1357/oh-my-pi/issues/5893)).
+- Fixed LiteLLM provider ignoring per-model pricing: `mapLiteLLMRichEntry` now reads `input_cost_per_token` / `output_cost_per_token` (plus cache costs) from LiteLLM rich metadata and maps them to `cost.input` / `cost.output`, falling back to the bundled reference only when LiteLLM omits cost, so proxied models no longer display as free ([#5818](https://github.com/can1357/oh-my-pi/issues/5818)).
+
+## [17.0.2] - 2026-07-17
+
+### Changed
+
+- Increased the maximum output tokens (maxTokens) from 32,768 to 65,536 for Kimi K2.7-Code models on Fireworks.
+
+### Fixed
+
+- Fixed a regression where the context window for openai-codex GPT-5.6 models (Luna, Sol, Terra) incorrectly fell back to 272,000 instead of preserving its 372,000 capacity.
+- Fixed Umans PAYG models incorrectly displaying as "Free" in /models by correctly sourcing their published per-token rates.
+- Fixed native moonshot/kimi-k3 capabilities and pricing, ensuring it correctly reflects its official pricing, 1M context window, image input support, reasoning capabilities, and 128k output token limit.
+
 ## [17.0.1] - 2026-07-16
 
 ### Added

package/dist/types/identity/family.d.ts

@@ -10,6 +10,13 @@
 export declare const isKimiModelId: (modelId: string) => boolean;
 /** Kimi K2.6 specifically, including router ids that spell the version `k2p6`. */
 export declare const isKimiK26ModelId: (modelId: string) => boolean;
+/**
+ * Kimi K3 in any namespace form (`kimi-k3`, `kimi-k3.1`, `kimi-k3-turbo`,
+ * `moonshotai/kimi-k3`). K3 always reasons and drives thinking via OpenAI-style
+ * `reasoning_effort: "max"`, not the K2.x binary `thinking: { type }` block —
+ * see the moonshot discovery mapper and `buildOpenAICompat`.
+ */
+export declare const isKimiK3ModelId: (modelId: string) => boolean;
 /**
  * Claude ids in any namespace form: bare (`claude-*`), path-namespaced
  * (`anthropic/claude.x`), or dot-prefixed (`us.anthropic.claude-…`,

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

@@ -181,10 +181,26 @@ export declare function zhipuCodingPlanModelManagerOptions(config?: ZhipuCodingP
  */
 export declare const FIREWORKS_KIMI_MAX_TOKENS = 32768;
 /**
- * Returns true for any Kimi K2.x public model id served by Fireworks-backed
- * providers (`fireworks` direct, `firepass` router). Matches both the public
- * catalog id (`kimi-k2.5`, `kimi-k2.6`, `kimi-k2.6-turbo`) and the canonical
- * Fireworks wire id (`accounts/fireworks/{models,routers}/kimi-k2…`).
+ * Fireworks' output ceiling for Kimi K2.7-Code specifically. Its `/v1/models`
+ * generic `max_completion_tokens` is 65,536 and Fireworks serves it in full —
+ * verified with a single completion emitting 58,971 output tokens and
+ * `max_tokens: 200000` accepted without error. Unlike the older K2.5/K2.6
+ * family (see {@link FIREWORKS_KIMI_MAX_TOKENS}), K2.7-Code is not clamped to
+ * 32,768; that ceiling only truncated it.
+ */
+export declare const FIREWORKS_KIMI_K27_CODE_MAX_TOKENS = 65536;
+/**
+ * Returns true for the Kimi K2.5 / K2.6 family served by Fireworks-backed
+ * providers (`fireworks` direct, `firepass` router) that share the 32,768
+ * `maxTokens` ceiling. Matches both the public catalog id (`kimi-k2.5`,
+ * `kimi-k2.6`, `kimi-k2.6-turbo`) and the canonical Fireworks wire id
+ * (`accounts/fireworks/{models,routers}/kimi-k2…`).
+ *
+ * K2.7-Code (incl. `-fast` / `-highspeed`) is deliberately excluded: unlike the
+ * earlier K2 family it serves its full context on Fireworks — verified with a
+ * single completion emitting 58,971 output tokens and `max_tokens: 200000`
+ * accepted without error — so the 32,768 cap would only truncate it. It inherits
+ * Fireworks' reported `max_completion_tokens` (65,536) instead.
  */
 export declare function isFireworksKimiK2ModelId(modelId: string): boolean;
 /**

package/dist/types/types.d.ts

@@ -119,7 +119,7 @@ export interface Usage {
         total: number;
     };
 }
-export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "qwen" | "qwen-chat-template";
+export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "kimi" | "qwen" | "qwen-chat-template";
 export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "openrouter-enabled-false" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false";
 export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "thinking";
 /**
@@ -167,8 +167,10 @@ export interface OpenAICompat {
     requiresThinkingAsText?: boolean;
     /** Whether tool call IDs must be normalized to Mistral format (exactly 9 alphanumeric chars). Default: auto-detected from URL. */
     requiresMistralToolIds?: boolean;
-    /** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "zai" uses thinking: { type: "enabled" | "disabled" } (also used by Moonshot Kimi), "qwen" uses top-level enable_thinking, and "qwen-chat-template" uses chat_template_kwargs.enable_thinking. Default: "openai". */
+    /** Format for reasoning/thinking parameter. `"kimi"` uses `thinking: { type, effort }`; other values select their provider-native reasoning fields. Default: `"openai"`. */
     thinkingFormat?: OpenAIReasoningFormat;
+    /** Kimi Code transport selected by live per-model protocol metadata. User settings take precedence. */
+    kimiApiFormat?: "openai" | "anthropic";
     /** Request-time disable encoding for the selected reasoning/thinking format. Default: derived from `thinkingFormat`. */
     reasoningDisableMode?: OpenAIReasoningDisableMode;
     /** Whether the provider rejects `reasoning.effort`/`reasoning_effort` even when the model reasons natively. Default: false unless reasoning effort is unsupported. */
@@ -432,6 +434,8 @@ export interface ResolvedOpenAISharedCompat {
     supportsReasoningParams: boolean;
     supportsSamplingParams: boolean;
     thinkingFormat: OpenAIReasoningFormat;
+    /** Kimi Code transport selected by live per-model protocol metadata. */
+    kimiApiFormat?: OpenAICompat["kimiApiFormat"];
     reasoningDisableMode: OpenAIReasoningDisableMode;
     omitReasoningEffort: boolean;
     includeEncryptedReasoning: boolean;
@@ -474,7 +478,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsSamplingParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsSamplingParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "17.0.1",
+	"version": "17.0.3",
 	"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": "^2.12.1",
-		"@oh-my-pi/pi-utils": "17.0.1",
+		"@oh-my-pi/pi-utils": "17.0.3",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.0.1",
+		"@oh-my-pi/pi-ai": "17.0.3",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -15,6 +15,7 @@ import {
 	isDeepseekModelIdOrName,
 	isGlm52ReasoningEffortModelId,
 	isGrokReasoningEffortCapable,
+	isKimiK3ModelId,
 	isKimiK26ModelId,
 	isKimiModelId,
 	isMimoModelIdOrName,
@@ -85,6 +86,7 @@ function resolveReasoningDisableMode(
 		case "openrouter":
 			return "openrouter-enabled-false";
 		case "zai":
+		case "kimi":
 			return "zai-thinking-disabled";
 		case "qwen":
 			return "qwen-enable-thinking-false";
@@ -247,6 +249,11 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 	const isKimiModel = isKimiModelId(spec.id);
 	const isMoonshotNative = modelMatchesHost(hostModel, "moonshotNative");
 	const isMoonshotKimi = isKimiModel && isMoonshotNative;
+	// Kimi K3 (native) always reasons via OpenAI-style `reasoning_effort: "max"`
+	// and does NOT accept the K2.x binary `thinking: { type }` block, so it must
+	// stay on the "openai" thinking dialect even though it is a Moonshot-native
+	// Kimi model (#5756).
+	const isMoonshotKimiK3 = isMoonshotKimi && isKimiK3ModelId(spec.id);
 	const requiresEnabledThinking = isMoonshotKimi && matchesKimiK27CodeFamily(spec);
 	const usesMoonshotKimiPreservedThinking = isMoonshotKimi && isKimiK26ModelId(spec.id);
 	const isAnthropicModel =
@@ -364,7 +371,10 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 				? ALIBABA_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
 				: isXiaomiMimo
 					? XIAOMI_MIMO_STREAM_IDLE_TIMEOUT_MS
-					: spec.reasoning && (isKimiK26ModelId(spec.id) || (isMoonshotKimi && matchesKimiK27CodeFamily(spec)))
+					: spec.reasoning &&
+							(isKimiK26ModelId(spec.id) ||
+								isMoonshotKimiK3 ||
+								(isMoonshotKimi && matchesKimiK27CodeFamily(spec)))
 						? KIMI_REASONING_STREAM_IDLE_TIMEOUT_MS
 						: spec.reasoning && isDirectDeepseekApi
 							? DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS
@@ -385,7 +395,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 					? "openrouter"
 					: "raw";
 	const thinkingFormat: ResolvedOpenAISharedCompat["thinkingFormat"] =
-		isZai || isZhipu || isMoonshotKimi || isXiaomiMimo
+		(isMoonshotKimi && !isMoonshotKimiK3) || isZai || isZhipu || isXiaomiMimo
 			? "zai"
 			: isOpenRouter
 				? "openrouter"
@@ -426,7 +436,12 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		// every call since the family can otherwise emit very long reasoning traces
 		// before the final answer.
 		alwaysSendMaxTokens: isKimiModel,
-		disableReasoningOnForcedToolChoice: isKimiModel || isAnthropicModel,
+		// Native Kimi K3 always reasons via `reasoning_effort: "max"` (never the
+		// K2.x binary `thinking` block that #827's forced-tool-choice conflict is
+		// about), so suppressing its effort would strip the mandatory `max` from
+		// normal forced-tool turns (e.g. plan-mode `toolChoice: "required"`) and
+		// leave K3 in an unsupported mode (#5758 review).
+		disableReasoningOnForcedToolChoice: (isKimiModel && !isMoonshotKimiK3) || isAnthropicModel,
 		disableReasoningOnToolChoice: isDeepseekFamily && Boolean(spec.reasoning) && !isOpenRouter,
 		supportsToolChoice: !isDirectDeepseekReasoning,
 		supportsForcedToolChoice: !requiresEnabledThinking,
@@ -446,6 +461,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		// is rejected by NIM's `additionalProperties: false` request schema
 		// (issue #2299).
 		thinkingFormat,
+		kimiApiFormat: undefined,
 		reasoningDisableMode: resolveReasoningDisableMode(thinkingFormat),
 		omitReasoningEffort: false,
 		includeEncryptedReasoning: true,

package/src/discovery/codex.ts

@@ -1,4 +1,5 @@
 import { type } from "arktype";
+import { parseKnownModel, semverEqual } from "../identity/classify";
 import type { ModelSpec } from "../types";
 import { discoveryFetch } from "../utils";
 import { CODEX_BASE_URL, CODEX_CLIENT_VERSION, OPENAI_HEADER_VALUES, OPENAI_HEADERS } from "../wire/codex";
@@ -6,6 +7,14 @@ import { CODEX_BASE_URL, CODEX_CLIENT_VERSION, OPENAI_HEADER_VALUES, OPENAI_HEAD
 const DEFAULT_MODEL_LIST_PATHS = ["/codex/models", "/models"] as const;
 const DEFAULT_CONTEXT_WINDOW = 272_000;
 const DEFAULT_MAX_TOKENS = 128_000;
+/**
+ * GPT-5.6 luna/sol/terra hard context capacity. Codex discovery omits
+ * `context_window` for these SKUs, so the generic {@link DEFAULT_CONTEXT_WINDOW}
+ * (272000) would understate the real window — OpenAI's Codex model registry
+ * declares context_window = max_context_window = 372000 (#5705). Used as the
+ * fallback only when upstream reports no value.
+ */
+const GPT_5_6_CONTEXT_WINDOW = 372_000;
 const CODEX_REMOTE_COMPACTION = {
 	enabled: true,
 	api: "openai-codex-responses",
@@ -214,7 +223,14 @@ function normalizeCodexModelEntry(entry: unknown, baseUrl: string): NormalizedCo
 	}
 
 	const name = toNonEmptyString(payload.display_name) ?? slug;
-	const contextWindow = toPositiveInt(payload.context_window) ?? DEFAULT_CONTEXT_WINDOW;
+	// Codex discovery omits `context_window` for GPT-5.6 luna/sol/terra; the
+	// generic 272000 fallback understates their real 372000 window (#5705).
+	const parsed = parseKnownModel(slug);
+	const fallbackContextWindow =
+		parsed.family === "openai" && semverEqual(parsed.version, "5.6")
+			? GPT_5_6_CONTEXT_WINDOW
+			: DEFAULT_CONTEXT_WINDOW;
+	const contextWindow = toPositiveInt(payload.context_window) ?? fallbackContextWindow;
 	const maxTokens = Math.min(DEFAULT_MAX_TOKENS, contextWindow);
 	const reasoning = supportsReasoning(payload.default_reasoning_level, payload.supported_reasoning_levels);
 	const input = normalizeInputModalities(payload.input_modalities);

package/src/identity/family.ts

@@ -41,6 +41,16 @@ export const isKimiK26ModelId = memo((modelId: string): boolean => {
 	return /(^|\/)kimi-k2(?:\.6|p6)(?:[-:]|$)/i.test(modelId);
 });
 
+/**
+ * Kimi K3 in any namespace form (`kimi-k3`, `kimi-k3.1`, `kimi-k3-turbo`,
+ * `moonshotai/kimi-k3`). K3 always reasons and drives thinking via OpenAI-style
+ * `reasoning_effort: "max"`, not the K2.x binary `thinking: { type }` block —
+ * see the moonshot discovery mapper and `buildOpenAICompat`.
+ */
+export const isKimiK3ModelId = memo((modelId: string): boolean => {
+	return /(^|\/)kimi-k3(?:\.\d+)?(?:[-.:_]|$)/i.test(modelId);
+});
+
 /**
  * Claude ids in any namespace form: bare (`claude-*`), path-namespaced
  * (`anthropic/claude.x`), or dot-prefixed (`us.anthropic.claude-…`,

package/src/model-cache.ts

@@ -7,7 +7,8 @@ import { getModelDbPath } from "@oh-my-pi/pi-utils";
 import type { Api, Model, ModelSpec } from "./types";
 
 // Rows persist ModelSpec JSON (sparse `compat`, never the resolved record);
-// the model manager rebuilds via `buildModel` on load. v8 invalidates Codex
+// the model manager rebuilds via `buildModel` on load. v9 invalidates Kimi
+// Code rows predating live effort and protocol metadata; v8 invalidated Codex
 // discovery rows predating provider-native V2 compaction metadata; v7
 // invalidated rows predating the Antigravity Gemini budget-mode migration
 // (cached specs still carrying `thinking.mode: "google-level"` and the old
@@ -15,7 +16,7 @@ import type { Api, Model, ModelSpec } from "./types";
 // unknown-limit sentinels (222222/8888); v5 invalidated rows predating
 // effort-tier variant collapsing (raw `-low`/`-high`/`-thinking` member ids);
 // v4 dropped the pre-efforts ThinkingConfig shape.
-const CACHE_SCHEMA_VERSION = 8;
+const CACHE_SCHEMA_VERSION = 9;
 
 interface CacheRow {
 	provider_id: string;