@oh-my-pi/pi-catalog 17.0.8 → 17.1.0

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [17.1.0] - 2026-07-24
+
+### Added
+
+- Added Bedrock Converse prompt-cache compatibility limits, including 5-minute checkpoint support for Nova models (Lite, Micro, Pro, Premier, Nova 2 Lite) and 1-hour Claude retention.
+- Added native Meta Model API provider and Muse Spark 1.1 support, featuring Responses API reasoning replay, image input, and reasoning-effort controls.
+- Added Vercel AI Gateway integration features, including opt-in automatic prompt-cache compatibility, provider routing preferences, and Responses cache-anchor and cache-lifetime controls.
+- Added prompt-cache breakpoint capability metadata for OpenAI GPT-5.6, with opt-in support for older models and compatible endpoints.
+- Added native alibaba-token-plan provider with QwenCloud Token Plan Individual discovery and a curated chat-model fallback catalog.
+
+## [17.0.9] - 2026-07-23
+
 ### Changed
 
 - Renamed `codex-auto-review` model to `GPT-5.3 Codex Spark` with updated pricing and capabilities

package/dist/types/build.d.ts

@@ -1,3 +1,14 @@
+/**
+ * The single Model constructor. Resolution order is a dependency chain, each
+ * step materialized exactly once per spec:
+ *
+ *   1. compat   — URL/provider/id detection resolved into a complete record;
+ *   2. thinking — derived from identity + resolved compat (or trusted verbatim
+ *                 when the spec carries explicit metadata);
+ *
+ * Request handlers read fields — they never detect, parse ids, or allocate
+ * compat per request.
+ */
 import type { Api, CompatOf, Model, ModelSpec } from "./types.js";
 export declare function buildModel<TApi extends Api>(spec: ModelSpec<TApi>): Model<TApi>;
 export declare function buildCompat(spec: ModelSpec<Api>): CompatOf<Api>;

package/dist/types/compat/bedrock.d.ts

@@ -0,0 +1,3 @@
+import type { ModelSpec, ResolvedBedrockCompat } from "../types.js";
+/** Resolve Bedrock Converse prompt-cache capabilities once per model. */
+export declare function buildBedrockCompat(spec: ModelSpec<"bedrock-converse-stream">): ResolvedBedrockCompat;

package/dist/types/hosts.d.ts

@@ -62,8 +62,8 @@ export declare const KNOWN_HOSTS: {
         readonly urlMarkers: readonly ["api.kilo.ai"];
     };
     readonly alibabaDashscope: {
-        readonly providers: readonly ["alibaba-coding-plan"];
-        readonly urlMarkers: readonly ["dashscope"];
+        readonly providers: readonly ["alibaba-coding-plan", "alibaba-token-plan"];
+        readonly urlMarkers: readonly ["dashscope", "token-plan."];
     };
     readonly umans: {
         readonly providers: readonly ["umans"];

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: "alibaba-token-plan";
+    readonly defaultModel: "qwen3.7-plus";
+    readonly envVars: readonly ["ALIBABA_TOKEN_PLAN_API_KEY", "BAILIAN_TOKEN_PLAN_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "QwenCloud Token Plan";
+    };
 }, {
     readonly id: "baseten";
     readonly defaultModel: "moonshotai/Kimi-K2.7-Code";
@@ -200,6 +209,14 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly defaultModel: "devstral-medium-latest";
     readonly envVars: readonly ["MISTRAL_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+}, {
+    readonly id: "meta";
+    readonly defaultModel: "muse-spark-1.1";
+    readonly envVars: readonly ["MODEL_API_KEY", "META_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-responses", unknown>;
+    readonly catalogDiscovery: {
+        readonly label: "Meta Model API";
+    };
 }, {
     readonly id: "moonshot";
     readonly defaultModel: "kimi-k2.7-code";

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

@@ -299,6 +299,14 @@ export interface AlibabaCodingPlanModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function alibabaCodingPlanModelManagerOptions(config?: AlibabaCodingPlanModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export declare const ALIBABA_TOKEN_PLAN_BASE_URL = "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1";
+export declare const ALIBABA_TOKEN_PLAN_STATIC_MODELS: readonly ModelSpec<"openai-completions">[];
+export interface AlibabaTokenPlanModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function alibabaTokenPlanModelManagerOptions(config?: AlibabaTokenPlanModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface VercelAiGatewayModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -359,6 +367,13 @@ export interface CoreWeaveModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function coreWeaveModelManagerOptions(config?: CoreWeaveModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export declare const META_MUSE_STATIC_MODELS: readonly ModelSpec<"openai-responses">[];
+export interface MetaModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function metaModelManagerOptions(config?: MetaModelManagerConfig): ModelManagerOptions<"openai-responses">;
 export interface MoonshotModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -268,6 +268,14 @@ export interface OpenAICompat {
     promptCacheSessionHeader?: "x-grok-conv-id";
     /** Whether chat-completions payloads should include provider-specific prompt-cache markers. */
     cacheControlFormat?: "anthropic" | undefined;
+    /**
+     * Whether this endpoint/model accepts OpenAI's explicit
+     * `prompt_cache_breakpoint` content markers and `prompt_cache_options`.
+     * Defaults to the exact first-party model generation allowlist.
+     */
+    supportsPromptCacheBreakpoints?: boolean;
+    /** The only currently supported minimum lifetime for explicit OpenAI cache breakpoints. */
+    promptCacheBreakpointTtl?: "30m";
     /** Whether the provider supports the `strict` field in tool definitions. Default: auto-detected per provider/baseUrl (conservative for unknown providers). */
     supportsStrictMode?: boolean;
     /**
@@ -419,6 +427,34 @@ export interface AnthropicCompat {
      */
     escapeBuiltinToolNames?: boolean;
 }
+/**
+ * Compatibility settings for Bedrock Converse prompt caching. Cache pricing is
+ * deliberately not used to infer these request-shape capabilities.
+ */
+export interface BedrockCompat {
+    /** Whether this endpoint accepts no checkpoints, automatic caching, or explicit cachePoint blocks. */
+    promptCacheMode?: "none" | "automatic" | "explicit";
+    /** Whether explicit cachePoint blocks accept `ttl: "1h"`; omitted TTL means Bedrock's 5-minute default. */
+    supportsLongPromptCacheRetention?: boolean;
+    /**
+     * Bedrock-enforced minimum prompt-prefix tokens for an effective checkpoint.
+     * Capability metadata only: emitters must not estimate local token counts.
+     * Zero means no explicit checkpoints.
+     */
+    promptCacheMinimumTokens?: number;
+    /**
+     * Bedrock-enforced maximum explicit cache checkpoints per request.
+     * Capability metadata only; zero means no explicit checkpoints.
+     */
+    promptCacheMaximumCheckpoints?: number;
+}
+/** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
+export interface ResolvedBedrockCompat {
+    promptCacheMode: NonNullable<BedrockCompat["promptCacheMode"]>;
+    supportsLongPromptCacheRetention: boolean;
+    promptCacheMinimumTokens: number;
+    promptCacheMaximumCheckpoints: number;
+}
 /**
  * OpenRouter provider routing preferences.
  * Controls which upstream providers OpenRouter routes requests to.
@@ -440,6 +476,12 @@ export interface VercelGatewayRouting {
     only?: string[];
     /** List of provider slugs to try in order (e.g., ["anthropic", "openai"]). */
     order?: string[];
+    /** Enables Vercel AI Gateway's provider-aware automatic prompt caching. */
+    caching?: "auto";
+    /** Stable Responses input-item prefix to anchor for automatic caching. */
+    cacheAnchorItems?: number;
+    /** Requested automatic-cache lifetime for the Responses API. */
+    cacheTtl?: "5m" | "1h";
 }
 type ResolvedToolStrictMode = NonNullable<OpenAICompat["toolStrictMode"]> | "mixed";
 /**
@@ -482,6 +524,14 @@ export interface ResolvedOpenAISharedCompat {
     emptyLengthFinishIsContextError: boolean;
     usesOpenAIToolCallIdLimit: boolean;
     promptCacheSessionHeader?: OpenAICompat["promptCacheSessionHeader"];
+    /**
+     * Whether this model accepts explicit OpenAI prompt-cache breakpoints.
+     * Built catalog models always materialize this false-by-default value;
+     * optionality preserves hand-authored resolved compat fixtures.
+     */
+    supportsPromptCacheBreakpoints?: boolean;
+    /** Minimum cache lifetime supported by explicit OpenAI prompt-cache breakpoints. */
+    promptCacheBreakpointTtl?: "30m";
     /** The model sits behind OpenRouter (routing prefs and max-token omission apply). */
     isOpenRouterHost: boolean;
     /** Whether this endpoint needs a max-token field even when caller did not set one. */
@@ -500,7 +550,7 @@ 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" | "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">> & {
+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" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "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"];
@@ -520,6 +570,9 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     supportsImageDetailOriginal: boolean;
     supportsObfuscationOptOut: boolean;
     streamIdleTimeoutMs?: number;
+    vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
+    /** The model sits behind Vercel AI Gateway's Responses endpoint. */
+    isVercelGatewayHost: boolean;
 }
 /**
  * OpenRouter is a pseudo API: runtime dispatch can use either Responses
@@ -567,9 +620,9 @@ export interface DevinCompat {
 /** Fully-resolved devin-agent compat view. */
 export type ResolvedDevinCompat = Required<DevinCompat>;
 /** Sparse, user-authored compat overrides for a given API (models.json / config vocabulary). */
-export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "devin-agent" ? DevinCompat : undefined;
+export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "bedrock-converse-stream" ? BedrockCompat : TApi extends "devin-agent" ? DevinCompat : undefined;
 /** Resolved compat for a given API: complete record, materialized once by `buildModel`. */
-export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : undefined;
+export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "bedrock-converse-stream" ? ResolvedBedrockCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : undefined;
 /** Provider-native compaction endpoint configuration for one model. */
 export interface RemoteCompactionConfig<TApi extends Api = Api> {
     /** Enables provider-native compaction for providers not enabled by built-in policy. */

package/dist/types/wire/alibaba-token-plan.d.ts

@@ -0,0 +1,6 @@
+export interface AlibabaTokenPlanCredential {
+    token: string;
+    cookie?: string;
+}
+export declare function parseAlibabaTokenPlanCredential(value: string): AlibabaTokenPlanCredential | null;
+export declare function serializeAlibabaTokenPlanCredential(token: string, cookie: string): string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "17.0.8",
+	"version": "17.1.0",
 	"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": "17.0.8",
+		"@oh-my-pi/pi-utils": "17.1.0",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.0.8",
+		"@oh-my-pi/pi-ai": "17.1.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/build.ts

@@ -9,7 +9,9 @@
  * Request handlers read fields — they never detect, parse ids, or allocate
  * compat per request.
  */
+
 import { buildAnthropicCompat } from "./compat/anthropic";
+import { buildBedrockCompat } from "./compat/bedrock";
 import { buildDevinCompat } from "./compat/devin";
 import { buildOpenAICompat, buildOpenAIResponsesCompat, buildOpenRouterCompat } from "./compat/openai";
 import { resolveModelThinking } from "./model-thinking";
@@ -39,6 +41,8 @@ export function buildCompat(spec: ModelSpec<Api>): CompatOf<Api> {
 			return buildOpenAIResponsesCompat(spec as ModelSpec<"openai-responses">);
 		case "anthropic-messages":
 			return buildAnthropicCompat(spec as ModelSpec<"anthropic-messages">);
+		case "bedrock-converse-stream":
+			return buildBedrockCompat(spec as ModelSpec<"bedrock-converse-stream">);
 		case "devin-agent":
 			return buildDevinCompat(spec as ModelSpec<"devin-agent">);
 		default:

package/src/compat/bedrock.ts

@@ -0,0 +1,114 @@
+import type { ModelSpec, ResolvedBedrockCompat } from "../types";
+import { applyCompatOverrides } from "./apply";
+
+const NO_EXPLICIT_CHECKPOINTS: ResolvedBedrockCompat = {
+	promptCacheMode: "none",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 0,
+	promptCacheMaximumCheckpoints: 0,
+};
+const EXPLICIT_CHECKPOINTS_1024_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 1024,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_1024_1H: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: true,
+	promptCacheMinimumTokens: 1024,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_2048_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 2048,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_4096_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 4096,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_4096_1H: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: true,
+	promptCacheMinimumTokens: 4096,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+/**
+ * Explicit Nova cache points complement Bedrock's automatic prefix caching:
+ * AWS recommends them for consistent cache hits and input-cost savings. Keep
+ * these exact documented model and inference-profile IDs conservative rather
+ * than treating arbitrary Nova-like application profiles as checkpoint-capable.
+ */
+function detectedBedrockCompat(modelId: string): ResolvedBedrockCompat {
+	const id = modelId.toLowerCase();
+
+	if (
+		id === "amazon.nova-lite-v1:0" ||
+		id === "us.amazon.nova-lite-v1:0" ||
+		id === "amazon.nova-micro-v1:0" ||
+		id === "us.amazon.nova-micro-v1:0" ||
+		id === "amazon.nova-pro-v1:0" ||
+		id === "us.amazon.nova-pro-v1:0" ||
+		id === "amazon.nova-premier-v1:0" ||
+		id === "us.amazon.nova-premier-v1:0" ||
+		id === "amazon.nova-2-lite-v1:0" ||
+		id === "us.amazon.nova-2-lite-v1:0" ||
+		id === "eu.amazon.nova-2-lite-v1:0" ||
+		id === "jp.amazon.nova-2-lite-v1:0" ||
+		id === "global.amazon.nova-2-lite-v1:0"
+	) {
+		return EXPLICIT_CHECKPOINTS_1024_5M;
+	}
+
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
+	// This list is deliberately sourced from AWS model cards, not cache pricing:
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html
+	if (
+		id.includes("anthropic.claude-opus-4-5") ||
+		id.includes("anthropic.claude-sonnet-4-5") ||
+		id.includes("anthropic.claude-haiku-4-5") ||
+		id.includes("anthropic.claude-opus-4-7") ||
+		id.includes("anthropic.claude-opus-4-8") ||
+		id.includes("anthropic.claude-sonnet-5")
+	) {
+		return EXPLICIT_CHECKPOINTS_4096_1H;
+	}
+	if (id.includes("anthropic.claude-opus-4-6")) {
+		return EXPLICIT_CHECKPOINTS_4096_5M;
+	}
+	if (id.includes("anthropic.claude-3-5-haiku")) {
+		return EXPLICIT_CHECKPOINTS_2048_5M;
+	}
+	if (id.includes("anthropic.claude-fable-5")) {
+		return EXPLICIT_CHECKPOINTS_1024_1H;
+	}
+
+	if (
+		id.includes("anthropic.claude-opus-4-1") ||
+		id.includes("anthropic.claude-opus-4-20250514") ||
+		id.includes("anthropic.claude-sonnet-4-20250514") ||
+		id.includes("anthropic.claude-sonnet-4-6") ||
+		id.includes("anthropic.claude-3-7-sonnet") ||
+		id.includes("anthropic.claude-3-5-sonnet-20241022-v2")
+	) {
+		return EXPLICIT_CHECKPOINTS_1024_5M;
+	}
+
+	return NO_EXPLICIT_CHECKPOINTS;
+}
+
+/** Resolve Bedrock Converse prompt-cache capabilities once per model. */
+export function buildBedrockCompat(spec: ModelSpec<"bedrock-converse-stream">): ResolvedBedrockCompat {
+	const compat = { ...detectedBedrockCompat(spec.id) };
+	applyCompatOverrides(compat, spec.compat);
+	return compat;
+}

package/src/compat/openai.ts

@@ -9,6 +9,7 @@
  */
 import { isFireworksFastModelId } from "../fireworks-model-id";
 import { hostMatchesUrl, modelMatchesHost } from "../hosts";
+import { bareModelId, parseOpenAIModel, semverGte } from "../identity/classify";
 import {
 	isAnthropicNamespacedModelId,
 	isClaudeModelId,
@@ -133,6 +134,17 @@ function isOfficialOpenAIEndpoint(provider: string, baseUrl: string): boolean {
 	}
 }
 
+/**
+ * Explicit prompt-cache breakpoints are a GPT-5.6+ first-party contract. Keep
+ * this intentionally narrow: compatible gateways and older OpenAI models
+ * reject the new request fields unless their catalog compat opts in.
+ */
+function supportsOfficialOpenAIPromptCacheBreakpoints(provider: string, modelId: string, baseUrl: string): boolean {
+	if (!isOfficialOpenAIEndpoint(provider, baseUrl)) return false;
+	const model = parseOpenAIModel(bareModelId(modelId));
+	return model !== null && semverGte(model.version, "5.6");
+}
+
 /**
  * OpenCode's gateways (https://opencode.ai/zen|go) gate `reasoning_content`
  * on the request's thinking state for every model they front (Kimi K2.x,
@@ -341,6 +353,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		hostMatchesUrl(baseUrl, "fireworks") ||
 		isDirectDeepseekApi;
 
+	const supportsPromptCacheBreakpoints = supportsOfficialOpenAIPromptCacheBreakpoints(provider, spec.id, baseUrl);
 	// Hosts whose chat-completions endpoints are known to accept multiple
 	// leading `system`/`developer` messages (preferred for KV-cache reuse).
 	// Anything outside this allowlist defaults to coalescing because
@@ -546,6 +559,8 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 			(thinkingFormat === "qwen" || thinkingFormat === "qwen-chat-template") && isLocalOpenAICompatBackend,
 		requiresAssistantContentForToolCalls: isKimiModel || isDirectDeepseekReasoning,
 		cacheControlFormat: isOpenRouter && spec.id.startsWith("anthropic/") ? "anthropic" : undefined,
+		supportsPromptCacheBreakpoints,
+		promptCacheBreakpointTtl: supportsPromptCacheBreakpoints ? "30m" : undefined,
 		openRouterRouting: undefined,
 		vercelGatewayRouting: undefined,
 		isOpenRouterHost: isOpenRouter,
@@ -623,7 +638,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 	const isAzure = modelMatchesHost({ provider: spec.provider, baseUrl }, "azureOpenAI");
 	const isOpenRouter = modelMatchesHost({ provider: spec.provider, baseUrl }, "openrouter");
 	const isOpenAIUrl = hostMatchesUrl(baseUrl, "openai");
+	const isVercelGateway = modelMatchesHost({ provider: spec.provider, baseUrl }, "vercelAIGateway");
 	const id = spec.id ?? "";
+	const supportsPromptCacheBreakpoints = supportsOfficialOpenAIPromptCacheBreakpoints(spec.provider, id, baseUrl);
 	const thinkingFormat: ResolvedOpenAISharedCompat["thinkingFormat"] = isOpenRouter ? "openrouter" : "openai";
 	const isKimiModel = id ? isKimiModelId(id) : false;
 	const isAnthropicModel = id ? isClaudeModelId(id) || isAnthropicNamespacedModelId(id) : false;
@@ -642,6 +659,8 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		supportsStrictMode: isAzure || detectStrictModeSupport(spec.provider, baseUrl),
 		supportsReasoningEffort: spec.provider !== "xai-oauth" || isGrokReasoningEffortCapable(id),
 		supportsLongPromptCacheRetention: isOpenAIUrl,
+		supportsPromptCacheBreakpoints,
+		promptCacheBreakpointTtl: supportsPromptCacheBreakpoints ? "30m" : undefined,
 		// Azure OpenAI and GitHub Copilot Responses paths require tool results
 		// to strictly match prior tool calls when building Responses inputs.
 		strictResponsesPairing: isAzure || spec.provider === "github-copilot",
@@ -686,7 +705,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		requiresAssistantAfterToolResult: false,
 		requiresAssistantContentForToolCalls: isKimiModel,
 		openRouterRouting: undefined,
+		vercelGatewayRouting: undefined,
 		isOpenRouterHost: isOpenRouter,
+		isVercelGatewayHost: isVercelGateway,
 		wireModelIdMode: isOpenRouter ? "openrouter" : "raw",
 		// Mirrors buildOpenAICompat: Kimi behind a Responses-capable proxy still
 		// lands on Moonshot's MFJS validator.
@@ -724,6 +745,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		strictResponsesPairing: compat.strictResponsesPairing,
 		supportsImageDetailOriginal: compat.supportsImageDetailOriginal,
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
+		isVercelGatewayHost: compat.isVercelGatewayHost,
 	} satisfies ResponsesOnlyCompat;
 }
 

package/src/discovery/codex.ts

@@ -8,11 +8,11 @@ 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. OpenAI's Codex model registry
- * declares context_window = max_context_window = 372000 (#5705), but Codex
- * discovery under-reports it — omitting the field for some accounts and
- * actively returning 272000 for others (#6259). Applied as a floor for these
- * SKUs so the reported/absent value never regresses the real window.
+ * 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 = {
@@ -223,18 +223,14 @@ function normalizeCodexModelEntry(entry: unknown, baseUrl: string): NormalizedCo
 	}
 
 	const name = toNonEmptyString(payload.display_name) ?? slug;
-	// GPT-5.6 luna/sol/terra have a 372000 hard window, but Codex discovery
-	// under-reports it: for some accounts the field is omitted, for others it is
-	// actively returned as 272000 (#6259). Treat GPT_5_6_CONTEXT_WINDOW as a
-	// floor for these SKUs so neither the omission nor the active under-report
-	// regresses the real capacity; other models honor the reported value with
-	// the generic 272000 fallback.
+	// 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 isGpt56 = parsed.family === "openai" && semverEqual(parsed.version, "5.6");
-	const reportedContextWindow = toPositiveInt(payload.context_window);
-	const contextWindow = isGpt56
-		? Math.max(GPT_5_6_CONTEXT_WINDOW, reportedContextWindow ?? 0)
-		: (reportedContextWindow ?? DEFAULT_CONTEXT_WINDOW);
+	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/hosts.ts

@@ -41,7 +41,10 @@ export const KNOWN_HOSTS = {
 	zai: { providers: ["zai"], urlMarkers: ["api.z.ai"] },
 	zhipu: { providers: ["zhipu-coding-plan"], urlMarkers: ["open.bigmodel.cn"] },
 	kilo: { providers: ["kilo"], urlMarkers: ["api.kilo.ai"] },
-	alibabaDashscope: { providers: ["alibaba-coding-plan"], urlMarkers: ["dashscope"] },
+	alibabaDashscope: {
+		providers: ["alibaba-coding-plan", "alibaba-token-plan"],
+		urlMarkers: ["dashscope", "token-plan."],
+	},
 	umans: { providers: ["umans"], urlMarkers: ["api.code.umans.ai"] },
 	xiaomi: { providers: ["xiaomi"], providerPrefixes: ["xiaomi-token-plan-"], urlMarkers: ["xiaomimimo.com"] },
 	xai: { providers: ["xai"], urlMarkers: ["api.x.ai"] },

package/src/identity/priority.ts

@@ -10,6 +10,7 @@ const DEFAULT_MODEL_PROVIDER_ORDER = [
 	"kimi-code",
 	"moonshot",
 	"qwen-portal",
+	"alibaba-token-plan",
 	"zai",
 	"xai-oauth",
 	"xai",