@oh-my-pi/pi-catalog 17.0.0 → 17.0.2

package/CHANGELOG.md

@@ -2,6 +2,46 @@
 
 ## [Unreleased]
 
+## [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
+
+- Added GPT-5.6 Luna, Sol, and Terra entries for Amazon Bedrock, Azure, and Cloudflare
+- Added KAT-Coder Air/Pro V2.5 entries across Kilo, OpenRouter, NanoGPT, and Vercel
+- Added Inkling model entries for Baseten and Vercel AI Gateway
+- Added Umans DeepSeek V4 Pro DSpark as an experimental model listing
+- Added Claude Opus 4.7 Fast and 4.8 Fast on Vercel AI Gateway
+- Added Workers AI GLM-5.2, Muse Spark 1.1, Stealth GPT-5.6 Sol, and nano-gpt-help entries
+
+### Changed
+
+- Added image input and reasoning support to several existing Codeium and Kilo GPT-5.6 models
+- Enabled image input and reasoning for Gemini Flash Latest and Grok 4.5
+- Renamed many model labels for consistency, including Claude, Grok, DeepSeek, GLM, and Gemi­ni names
+- Updated pricing for many existing models, including input, output, and cache cost values
+- Updated context window and max token limits for many catalog models across providers
+
+### Fixed
+
+- Fixed Z.AI (GLM) coding-plan token costs all showing as "Free" in `/models`: the `zai` provider descriptor sourced the models.dev `zai-coding-plan` key (all-$0 subscription rates) instead of the `zai` pay-as-you-go key, which carries the real per-token rates for the identical GLM ids ([#5598](https://github.com/can1357/oh-my-pi/issues/5598)).
+- Fixed custom Anthropic endpoints receiving the first-party-only `eager_input_streaming` tool field by default ([#5572](https://github.com/can1357/oh-my-pi/issues/5572)).
+- Added resolved OpenAI sampling-parameter compatibility metadata for o-series and GPT-5+ models.
+- Fixed GitHub Copilot `mai-code-1-flash-picker` (and other `mai-*` models) to route through the `/responses` endpoint instead of `/chat/completions`, which rejected them with `400 unsupported_api_for_model` ([#5612](https://github.com/can1357/oh-my-pi/issues/5612)).
+- Extended the reasoning `streamIdleTimeoutMs` floor (300s) to native Kimi K2.7 Code (`kimi-k2.7-code` / `kimi-k2.7-code-highspeed`), which previously fell through to the 120s default and aborted on long reasoning turns ([#4836](https://github.com/can1357/oh-my-pi/issues/4836)).
+- Fixed GLM-5.x coding-plan streams via the OpenCode Go/Zen gateways (`opencode.ai/zen/…`) timing out with `OpenAI completions stream stalled while waiting for the next event` during slow plan-writing/reasoning phases. The 600s idle-timeout floor for GLM coding-plan SKUs was gated to the native Z.AI/Zhipu hosts only, so OpenCode-fronted GLM fell back to the 120s default watchdog. ([#4758](https://github.com/can1357/oh-my-pi/issues/4758))
+
 ## [16.5.2] - 2026-07-14
 
 ### Fixed

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-…`,
@@ -76,6 +83,18 @@ export declare const supportsAllTurnsReasoningContext: (modelId: string) => bool
  * the human-readable summary stream.
  */
 export declare const supportsCodexReasoningSummary: (modelId: string) => boolean;
+/**
+ * OpenAI proprietary models whose serving path rejects explicit sampling
+ * parameters (`temperature`, `top_p`, `top_k`, …) with
+ * `400 Unsupported parameter: 'temperature' is not supported with this model`.
+ * Covers the o-series and the entire gpt-5+ generation — base, `mini`, `nano`,
+ * `codex*`, the `luna`/`sol`/`terra` SKUs, and the `-chat-latest` variants,
+ * since even the non-reasoning gpt-5 chat models reject sampling params (see
+ * litellm#13781). Holds regardless of which OpenAI-serving host proxies the
+ * model (official, Azure, GitHub Copilot). Version floor (not an allowlist) so
+ * 6.x inherits automatically. Issue #5606.
+ */
+export declare const isOpenAISamplingRestrictedModelId: (modelId: string) => boolean;
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
  * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning

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

@@ -289,6 +289,15 @@ export interface OpenAICompat {
     toolStrictMode?: "all_strict" | "none";
     /** Whether request shaping may send reasoning params at all. Default: auto-detected (disabled for GitHub Copilot chat-completions). */
     supportsReasoningParams?: boolean;
+    /**
+     * Whether the endpoint accepts explicit sampling parameters (`temperature`,
+     * `top_p`, `top_k`, `min_p`, penalties). OpenAI proprietary reasoning models
+     * (o-series, gpt-5+) reject them with `400 Unsupported parameter:
+     * 'temperature' is not supported with this model` on every serving host
+     * (official, Azure, GitHub Copilot). When unset, auto-detected from the
+     * model id. Default: true. Issue #5606.
+     */
+    supportsSamplingParams?: boolean;
     /** Always send a max-token field when the caller did not provide one. Default: auto-detected (Kimi-family models derive TPM limits from max_tokens). */
     alwaysSendMaxTokens?: boolean;
     /** Whether Responses-API tool-call/result history must be strictly paired. Default: auto-detected (Azure OpenAI, GitHub Copilot). */
@@ -333,7 +342,7 @@ export interface AnthropicCompat {
      * tags: 'disabled', 'enabled'`.
      */
     disableAdaptiveThinking?: boolean;
-    /** Whether tools may include Anthropic's per-tool eager_input_streaming flag. Default: true. */
+    /** Whether tools may include Anthropic's per-tool eager_input_streaming flag. Default: true for the canonical Anthropic API. */
     supportsEagerToolInputStreaming?: boolean;
     /** Whether long prompt-cache retention (`ttl: "1h"`) is supported. Default: true for canonical Anthropic API. */
     supportsLongCacheRetention?: boolean;
@@ -421,6 +430,7 @@ export interface ResolvedOpenAISharedCompat {
     supportsReasoningEffort: boolean;
     reasoningEffortMap: Partial<Record<Effort, string>>;
     supportsReasoningParams: boolean;
+    supportsSamplingParams: boolean;
     thinkingFormat: OpenAIReasoningFormat;
     reasoningDisableMode: OpenAIReasoningDisableMode;
     omitReasoningEffort: boolean;
@@ -464,7 +474,7 @@ 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" | "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" | "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": "17.0.0",
+	"version": "17.0.2",
 	"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": "17.0.0",
+		"@oh-my-pi/pi-utils": "17.0.2",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.0.0",
+		"@oh-my-pi/pi-ai": "17.0.2",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -109,7 +109,7 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		signingEndpoint,
 		disableStrictTools: isAzure,
 		disableAdaptiveThinking: false,
-		supportsEagerToolInputStreaming: !isCopilot,
+		supportsEagerToolInputStreaming: official,
 		// Long cache retention is only sent to the official API by default;
 		// proxies opt in explicitly via `compat.supportsLongCacheRetention: true`.
 		supportsLongCacheRetention: official,

package/src/compat/openai.ts

@@ -15,9 +15,11 @@ import {
 	isDeepseekModelIdOrName,
 	isGlm52ReasoningEffortModelId,
 	isGrokReasoningEffortCapable,
+	isKimiK3ModelId,
 	isKimiK26ModelId,
 	isKimiModelId,
 	isMimoModelIdOrName,
+	isOpenAISamplingRestrictedModelId,
 	isQwenModelId,
 	modelFamilyToken,
 } from "../identity/family";
@@ -37,8 +39,8 @@ const GLM_CODING_PLAN_MODEL_PATTERN = /(^|\/)glm-5(?:[.-]|$)/i;
 const GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS = 600_000;
 /** Direct DeepSeek reasoning models stall between thinking and answer phases. */
 const DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS = 300_000;
-/** Kimi K2.6 can spend several minutes reasoning before the first visible token. */
-const KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS = 300_000;
+/** Kimi K2.6 and native K2.7 Code can spend several minutes reasoning before the first visible token. */
+const KIMI_REASONING_STREAM_IDLE_TIMEOUT_MS = 300_000;
 /**
  * Native Kimi K2.7 Code requires `thinking.type: "enabled"` and rejects
  * disabled thinking. Match the public id, its Fast variant, and the
@@ -246,6 +248,11 @@ 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 =
@@ -357,14 +364,17 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 	// for minutes while reasoning or cold-loading weights; widen the idle
 	// timeout so warm-ups stop aborting and retrying.
 	const streamIdleTimeoutMs =
-		GLM_CODING_PLAN_MODEL_PATTERN.test(spec.id) && (isZai || isZhipu)
+		GLM_CODING_PLAN_MODEL_PATTERN.test(spec.id) && (isZai || isZhipu || isOpenCodeHost)
 			? GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
 			: provider === "alibaba-coding-plan"
 				? ALIBABA_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
 				: isXiaomiMimo
 					? XIAOMI_MIMO_STREAM_IDLE_TIMEOUT_MS
-					: spec.reasoning && isKimiK26ModelId(spec.id)
-						? KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS
+					: spec.reasoning &&
+							(isKimiK26ModelId(spec.id) ||
+								isMoonshotKimiK3 ||
+								(isMoonshotKimi && matchesKimiK27CodeFamily(spec)))
+						? KIMI_REASONING_STREAM_IDLE_TIMEOUT_MS
 						: spec.reasoning && isDirectDeepseekApi
 							? DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS
 							: isLocalOpenAICompatBackend
@@ -384,7 +394,7 @@ 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"
@@ -409,6 +419,9 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		supportsReasoningEffort: !isGrok && !isXiaomiMimo && (!(isZai || isZhipu) || supportsZaiReasoningEffort),
 		// GitHub Copilot's chat-completions endpoint rejects reasoning params wholesale.
 		supportsReasoningParams: provider !== "github-copilot",
+		// OpenAI proprietary reasoning models (o-series, gpt-5+) reject explicit
+		// temperature/top_p/… with a 400 on every serving host (#5606).
+		supportsSamplingParams: !isOpenAISamplingRestrictedModelId(spec.id),
 		reasoningEffortMap: isMimoReasoningEffortModel ? MIMO_REASONING_EFFORT_MAP : {},
 		supportsUsageInStreaming: !isCerebras,
 		// pi-ai's thinking-loop guard is gemini-only; default the flag from the
@@ -422,7 +435,12 @@ 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,
@@ -604,6 +622,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 			spec.provider !== "xai-oauth" && !modelMatchesHost({ provider: spec.provider, baseUrl }, "githubCopilot"),
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
+		// OpenAI proprietary reasoning models (o-series, gpt-5+) reject explicit
+		// temperature/top_p/… with a 400 on every serving host (#5606).
+		supportsSamplingParams: !isOpenAISamplingRestrictedModelId(id),
 		thinkingFormat,
 		reasoningDisableMode: resolveReasoningDisableMode(thinkingFormat),
 		omitReasoningEffort: false,

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-…`,
@@ -163,6 +173,32 @@ export const supportsAllTurnsReasoningContext = isOpenAIWireGen54Plus;
  */
 export const supportsCodexReasoningSummary = isOpenAIWireGen54Plus;
 
+/** OpenAI proprietary reasoning families keyed off the parsed gpt version (gpt-5+). */
+const isOpenAIWireGen5Plus = memo((modelId: string): boolean => {
+	const parsed = parseOpenAIModel(bareModelId(modelId));
+	if (!parsed) return false;
+	return semverGte(parsed.version, "5");
+});
+
+/** o-series reasoning ids (`o1`, `o1-pro`, `o3`, `o3-mini`, `o4-mini`, `openai/o3`, …). */
+const O_SERIES_REASONING_RE = /(^|\/)o[134](?:[-.]|$)/i;
+
+/**
+ * OpenAI proprietary models whose serving path rejects explicit sampling
+ * parameters (`temperature`, `top_p`, `top_k`, …) with
+ * `400 Unsupported parameter: 'temperature' is not supported with this model`.
+ * Covers the o-series and the entire gpt-5+ generation — base, `mini`, `nano`,
+ * `codex*`, the `luna`/`sol`/`terra` SKUs, and the `-chat-latest` variants,
+ * since even the non-reasoning gpt-5 chat models reject sampling params (see
+ * litellm#13781). Holds regardless of which OpenAI-serving host proxies the
+ * model (official, Azure, GitHub Copilot). Version floor (not an allowlist) so
+ * 6.x inherits automatically. Issue #5606.
+ */
+export const isOpenAISamplingRestrictedModelId = memo((modelId: string): boolean => {
+	const bare = bareModelId(modelId);
+	return isOpenAIWireGen5Plus(modelId) || O_SERIES_REASONING_RE.test(bare);
+});
+
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
  * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning