@linxiraos/pi-catalog 1.1.7 → 1.1.8

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

@@ -8,6 +8,12 @@ export interface ModelsDevModel {
     name?: string;
     tool_call?: boolean;
     reasoning?: boolean;
+    reasoning_options?: Array<{
+        type?: string;
+        values?: string[];
+        min?: number;
+        max?: number;
+    }>;
     limit?: {
         context?: number;
         output?: number;
@@ -64,30 +70,6 @@ export interface UmansModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function umansModelManagerOptions(config?: UmansModelManagerConfig): ModelManagerOptions<"anthropic-messages">;
-/** GPT-5.6 rates applied when a first-party request exceeds 272K input tokens. */
-export declare const OPENAI_GPT_56_LONG_CONTEXT_COSTS: {
-    readonly luna: {
-        readonly inputThreshold: 272000;
-        readonly input: 0.4;
-        readonly output: 1.8;
-        readonly cacheRead: 0.04;
-        readonly cacheWrite: 0.5;
-    };
-    readonly sol: {
-        readonly inputThreshold: 272000;
-        readonly input: 10;
-        readonly output: 45;
-        readonly cacheRead: 1;
-        readonly cacheWrite: 12.5;
-    };
-    readonly terra: {
-        readonly inputThreshold: 272000;
-        readonly input: 4;
-        readonly output: 18;
-        readonly cacheRead: 0.4;
-        readonly cacheWrite: 5;
-    };
-};
 export interface OpenAIModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -183,8 +165,9 @@ export interface XaiModelManagerConfig {
     fetch?: FetchImpl;
 }
 /**
- * Applies xAI's long-context rate card and mirrors exact public-model prices
- * onto matching SuperGrok catalog rows.
+ * Mirrors exact public-model prices onto matching SuperGrok catalog rows.
+ * The >200K long-context tier itself is rule-owned (`classes/xai.kdl`
+ * `long-context-cost` multiplier axis) and derives at build time.
  */
 export declare function applyXaiCatalogPricing(models: readonly ModelSpec[]): ModelSpec[];
 export declare function xaiModelManagerOptions(config?: XaiModelManagerConfig): ModelManagerOptions<"openai-responses">;
@@ -202,9 +185,8 @@ interface XAICuratedModel {
     /**
      * Whether xAI accepts the `reasoning.effort` wire param for this model.
      * Default true. When false: the picker hides the effort dial (via
-     * getSupportedEfforts in model-thinking.ts) AND the wire omits the param —
-     * both derive from `isGrokReasoningEffortCapable` (identity/family.ts), the
-     * single allowlist shared by this curated layer and the compat builder.
+     * getSupportedEfforts in model-thinking.ts) AND the wire omits the param.
+     * Both read the compat engine's resolved reasoning-effort axes.
      */
     supportsReasoningEffort?: boolean;
     /**
@@ -220,8 +202,8 @@ export declare const XAI_OAUTH_CURATED_MODELS: readonly XAICuratedModel[];
  * Bake first-party xAI Responses effort-dial metadata onto a catalog spec.
  *
  * models.dev marks many Grok SKUs as reasoners and the thinking rebake would
- * otherwise emit a default `minimal/low/medium/high` dial. api.x.ai only
- * accepts `reasoning.effort` for {@link isGrokReasoningEffortCapable} ids —
+ * otherwise emit a default `minimal/low/medium/high` dial. api.x.ai
+ * accepts `reasoning.effort` only for ids selected by the compat rule —
  * off-allowlist reasoners (`grok-code-fast-1`, `grok-build-0.1`,
  * `grok-4.20-0309-reasoning`, …) 400 if the param is sent. SuperGrok
  * (`xai-oauth`) already curates this via {@link mergeCuratedIntoModel}; paid
@@ -269,6 +251,12 @@ export interface SiliconFlowModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * SiliconFlow's `/v1/models` lists every served model — including embeddings,
+ * rerankers, image, audio, and video generators that cannot serve chat
+ * completions — and carries no per-model type field, so non-chat entries are
+ * dropped by id to keep the picker usable.
+ */
 export declare function isLikelySiliconFlowChatModelId(id: string): boolean;
 export declare function siliconflowModelManagerOptions(config?: SiliconFlowModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export declare function siliconflowCnModelManagerOptions(config?: SiliconFlowModelManagerConfig): ModelManagerOptions<"openai-completions">;
@@ -355,6 +343,17 @@ export interface FirepassModelManagerConfig {
  * See https://docs.fireworks.ai/firepass.
  */
 export declare function firepassModelManagerOptions(_config?: FirepassModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Cline's OpenAI-compatible `/models` route is absent, but its public
+ * recommended-models endpoint is the authoritative ClinePass roster: the
+ * `clinePass` bucket lists subscription models and the optional `free` bucket
+ * adds $0 usage-billing models selectable with the same key. Metadata resolves
+ * in three tiers: the current Cline-published metadata snapshot first, bundled
+ * upstream references and OpenRouter enrichment for future ids second, then
+ * conservative constants. Newly added ids remain usable before the next regen;
+ * known ids retain exact Cline limits, plan pricing, and reasoning controls.
+ */
+export declare function clinePassModelManagerOptions(config?: ModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface WaferModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -408,17 +407,6 @@ export interface AlibabaCodingPlanModelManagerConfig {
 export declare function alibabaCodingPlanModelManagerOptions(config?: AlibabaCodingPlanModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export { ALIBABA_TOKEN_PLAN_BASE_URL };
 export declare const ALIBABA_TOKEN_PLAN_STATIC_MODELS: readonly ModelSpec<"openai-completions">[];
-/**
- * Metadata for Alibaba Token Plan models that are dynamically discovered but not
- * in the static catalog. Context window and max tokens are sourced from
- * official model documentation and provider catalogs. Unknown future models
- * remain available with null limits instead of assigning unsafe guessed limits.
- */
-export interface AlibabaTokenPlanModelLimits {
-    contextWindow: number;
-    maxTokens: number;
-}
-export declare const ALIBABA_TOKEN_PLAN_DISCOVERED_MODEL_LIMITS: Readonly<Record<string, AlibabaTokenPlanModelLimits>>;
 export interface AlibabaTokenPlanModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -442,13 +430,12 @@ export interface KimiCodeModelManagerConfig {
  * per-family caps instead of a blanket constant. Values match models.dev's
  * `kimi-for-coding` and `moonshotai` kimi-k3 entries. See #6711.
  */
-export declare const KIMI_CODE_K3_MAX_TOKENS = 131072;
-export declare const KIMI_CODE_FOR_CODING_MAX_TOKENS = 32768;
 /** Fallback output cap for Kimi Code families without a documented ceiling (legacy K2 discovery rows). */
 export declare const KIMI_CODE_DEFAULT_MAX_TOKENS = 32000;
 /**
- * Resolve a Kimi Code model's output ceiling from its id: `k3` / `k3-256k` ->
- * 131072, `kimi-for-coding[-highspeed]` -> 32768, everything else -> `fallback`.
+ * Resolve a Kimi Code model's output ceiling; the per-family caps are
+ * rule-owned (`runtime/behavior.kdl` model-limits), everything else ->
+ * `fallback`.
  */
 export declare function kimiCodeMaxTokens(modelId: string, fallback?: number): number;
 export declare function kimiCodeMaxTokens(modelId: string, fallback: number | null): number | null;
@@ -501,6 +488,12 @@ export interface CoreWeaveModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function coreWeaveModelManagerOptions(config?: CoreWeaveModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Muse Spark revisions served by Meta's first-party Responses API. Meta's
+ * `/v1/models` lists bare ids with no capability metadata, so every revision
+ * must be seeded here or discovery yields a text-only, non-reasoning model
+ * with an unknown context window.
+ */
 export declare const META_MUSE_STATIC_MODELS: readonly ModelSpec<"openai-responses">[];
 /**
  * OpenAI frontier models served exclusively through Bedrock Mantle's Responses
@@ -544,6 +537,21 @@ export interface AiandModelManagerConfig {
  * pricing metadata, so discovery is authoritative over the bundled seed.
  */
 export declare function aiandModelManagerOptions(config?: AiandModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Documented abliteration.ai catalog (docs.abliteration.ai/models, 2026-09)
+ * bundled so the provider is usable when generation and first boot have no
+ * live key. The `/v1/models` response is authoritative once discovery runs.
+ * Rows carry only the documented limits and pricing; the per-model reasoning
+ * ladders, alias map and wire quirks come from `rules/classes/glm.kdl` and
+ * `rules/providers/abliteration.kdl` via `buildModel`.
+ */
+export declare const ABLITERATION_STATIC_MODELS: readonly ModelSpec<"openai-responses">[];
+export interface AbliterationModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function abliterationModelManagerOptions(config?: AbliterationModelManagerConfig): ModelManagerOptions<"openai-responses">;
 /**
  * Documented Yolo-Auto catalog (yolo-auto.com/docs, 2026-08) bundled so the
  * provider is usable when generation and first boot have no live key. The
@@ -686,6 +694,11 @@ export interface ModelsDevProviderDescriptor {
 }
 /** Generic mapper that converts models.dev data using provider descriptors. */
 export declare function mapModelsDevToModels(data: Record<string, unknown>, descriptors: readonly ModelsDevProviderDescriptor[]): ModelSpec<Api>[];
+/** Resolves the transport and endpoint for one Z.AI model catalog entry. */
+export declare function resolveZaiApi(modelId: string): {
+    api: "anthropic-messages" | "openai-completions";
+    baseUrl: string;
+};
 /** All provider descriptors for models.dev data mapping in generate-models.ts. */
 export declare const MODELS_DEV_PROVIDER_DESCRIPTORS: readonly ModelsDevProviderDescriptor[];
 /** Providers whose bundled catalog can receive additive models.dev updates at runtime. */

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

@@ -1,6 +1,6 @@
 import type { DevinModelDiscoveryOptions } from "../discovery/devin.js";
 import type { ModelManagerOptions } from "../model-manager.js";
-import type { FetchImpl } from "../types.js";
+import type { Api, FetchImpl, Model, ModelSpec } from "../types.js";
 /** One Codex OAuth account to fetch a catalog for. */
 export interface OpenAICodexAccount {
     /** OAuth access token used for `Authorization: Bearer ...`. */
@@ -33,6 +33,13 @@ export interface CursorModelManagerConfig {
     clientVersion?: string;
 }
 export declare function cursorModelManagerOptions(config?: CursorModelManagerConfig): ModelManagerOptions<"cursor-agent">;
+export type GitLabDuoModelIdentity = {
+    upstreamModelId: string;
+    referenceProvider: "anthropic" | "openai";
+    referenceModelId: string;
+};
+export declare function resolveGitLabDuoModelIdentity(modelId: string): GitLabDuoModelIdentity | undefined;
+export declare function getGitLabDuoModels(): Model<Api>[];
 export interface GitLabDuoWorkflowModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -47,7 +54,21 @@ export interface DevinModelManagerConfig {
     baseUrl?: string;
     fetch?: DevinModelDiscoveryOptions["fetch"];
 }
+/**
+ * Curated Devin seed — the entire bundled surface for the provider. The
+ * Cascade catalog is credential-scoped (gated per account/team), so catalog
+ * generation never fetches it: baking one account's roster into the shared
+ * bundle would misstate every other account's entitlements and leave zombie
+ * rows behind (see CREDENTIAL_SCOPED_PROVIDERS in generate-models.ts). Both
+ * SWE-1.6 lanes are verified live against `GetCliModelConfigs`; the
+ * descriptor's `defaultModel` (`swe-1-6`) must resolve synchronously at
+ * boot, before credential-scoped runtime discovery replaces the seed. Field
+ * shape mirrors `devinModelSpec` so seeded and discovered rows are
+ * indistinguishable downstream.
+ */
+export declare const DEVIN_STATIC_MODELS: readonly ModelSpec<"devin-agent">[];
 export declare function devinModelManagerOptions(config?: DevinModelManagerConfig): ModelManagerOptions<"devin-agent">;
 export interface ZaiModelManagerConfig {
 }
-export declare function zaiModelManagerOptions(_config?: ZaiModelManagerConfig): ModelManagerOptions<"anthropic-messages">;
+/** Creates model-manager options for Z.AI's mixed native and Anthropic transports. */
+export declare function zaiModelManagerOptions(_config?: ZaiModelManagerConfig): ModelManagerOptions<"anthropic-messages" | "openai-completions">;

package/dist/types/types.d.ts

@@ -1,3 +1,4 @@
+import type { ModelIdentity } from "./compat/types.js";
 import type { Effort } from "./effort.js";
 export type { FetchImpl } from "@linxiraos/pi-utils";
 export type { KnownProvider } from "./provider-models/descriptors.js";
@@ -28,9 +29,14 @@ export interface ThinkingConfig {
      * 5). Also implies native interleaved thinking — no beta header needed.
      */
     supportsDisplay?: boolean;
+    /**
+     * Thinking signatures bind each block to its preceding conversation prefix.
+     * Requests that rewrite the prefix must opt into the provider's drop/error control.
+     */
+    prefixBinding?: boolean;
     /**
      * Per-effort upstream wire-id routing for collapsed effort-tier variants
-     * (`variant-collapse.ts`). Keyed by pi effort; `"off"` applies when
+     * (`compat/collapse.ts`). Keyed by pi effort; `"off"` applies when
      * thinking is disabled. Missing keys fall back to `requestModelId ?? id`.
      */
     effortRouting?: Readonly<Partial<Record<Effort | "off", string>>>;
@@ -113,6 +119,15 @@ export interface Usage {
         webSearch?: number;
         webFetch?: number;
     };
+    /** Provider-reported credit meter values for credit-based billing APIs. */
+    credits?: {
+        /** Uncommitted per-response credit cost, when reported. */
+        cost?: number;
+        /** Final committed credit cost, when reported. */
+        committedCost?: number;
+        /** Final committed ACU cost, when reported. */
+        acuCost?: number;
+    };
     cost: {
         input: number;
         output: number;
@@ -122,7 +137,7 @@ export interface Usage {
     };
 }
 export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "kimi" | "qwen" | "qwen-chat-template" | "chat-template";
-export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "venice-disable-thinking" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false" | "chat-template-thinking-false";
+export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "cline-enabled-false" | "venice-disable-thinking" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false" | "chat-template-thinking-false";
 export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "qwen" | "thinking";
 /**
  * Compatibility settings for openai-completions API.
@@ -267,6 +282,14 @@ export interface OpenAICompat {
     openRouterRouting?: OpenRouterRouting;
     /** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */
     vercelGatewayRouting?: VercelGatewayRouting;
+    /**
+     * Provider-specific wire model-id transform applied to the base id, overriding
+     * the provider-detected default. Discovery layers set this per model when the
+     * upstream roster declares the id's namespace (e.g. ClinePass free-tier models
+     * arrive pre-namespaced and use `"raw"`), so wire form follows data rather
+     * than id shape.
+     */
+    wireModelIdMode?: "raw" | "cline-pass" | "firepass" | "fireworks" | "openrouter";
     /** Extra fields to include in request body (e.g. gateway routing hints for OpenClaw-style proxies). */
     extraBody?: Record<string, unknown>;
     /** Request-session header that should mirror the normalized prompt-cache key. Default: unset. */
@@ -325,6 +348,8 @@ 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 Responses requests may include `reasoning.summary`. Default: true except on known incompatible hosts. */
+    supportsReasoningSummary?: boolean;
     /**
      * Whether the endpoint accepts explicit sampling parameters (`temperature`,
      * `top_p`, `top_k`, `min_p`, penalties). OpenAI proprietary reasoning models
@@ -365,6 +390,16 @@ export interface OpenAICompat {
      * spread. Default: auto-detected (OpenCode gateways, #1071/#1484).
      */
     whenThinking?: Partial<Omit<OpenAICompat, "whenThinking">>;
+    /** Kimi K3 drives reasoning through native `reasoning_effort` on Moonshot-native hosts. */
+    nativeKimiK3Reasoning?: boolean;
+    /** GLM-5.2 accepts Z.ai's `reasoning_effort` dialect alongside binary thinking. */
+    zaiReasoningEffortDialect?: boolean;
+    /** Clamp the requested max output tokens to the model's advertised ceiling. */
+    clampOutputToModelMax?: boolean;
+    /** Strip image inputs before encoding (text-only serving of a multimodal id). */
+    stripImageInput?: boolean;
+    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
+    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
 }
 /**
  * Compatibility settings for anthropic-messages API.
@@ -372,6 +407,13 @@ export interface OpenAICompat {
  * that proxy gateways (Vertex AI, AWS Bedrock-style fronts, etc.) reject.
  */
 export interface AnthropicCompat {
+    /** Whether thinking requests may include `context_management` and its beta header. Default: true. */
+    supportsContextManagement?: boolean;
+    /**
+     * Whether requests may carry `output_config.effort` (and its effort beta
+     * header). Vertex AI rejects the field/header. Default: true.
+     */
+    supportsOutputEffort?: boolean;
     /**
      * Stream-watchdog idle-timeout fallback in ms for slow reasoning hosts.
      * Set to 0 to disable the inter-event idle watchdog entirely, matching
@@ -400,12 +442,20 @@ export interface AnthropicCompat {
     supportsLongCacheRetention?: boolean;
     /**
      * Whether mid-conversation `role: "system"` messages are accepted in the
-     * `messages` array (Claude Opus 4.8+ and Claude Fable/Mythos 5 on the
-     * first-party Claude API and Claude Platform on AWS). When unset,
-     * auto-detected from the model id and base URL. Not available on Bedrock,
-     * Vertex AI, or Microsoft Foundry.
+     * `messages` array. When unset, auto-detected from model and deployment policy.
      */
     supportsMidConversationSystem?: boolean;
+    /** Whether turn-scoped system messages accept `clear_at`. */
+    supportsTurnScopedSystem?: boolean;
+    /** Whether tool availability can change through system-message tool references. */
+    supportsMidConversationToolChanges?: boolean;
+    /** Whether effort can change through a per-message `output_config`. */
+    supportsPerMessageEffort?: boolean;
+    /**
+     * Whether the endpoint accepts `thinking.block_binding` and reports
+     * `input_transformations` under the thinking-binding-controls beta.
+     */
+    supportsThinkingBindingControls?: boolean;
     /**
      * Whether the model accepts a forced `tool_choice` (`{ type: "any" }` or
      * `{ type: "tool", name }`). Claude Fable/Mythos 5 reject forced tool use
@@ -467,6 +517,15 @@ export interface AnthropicCompat {
      * {@link ResolvedAnthropicCompat.officialEndpoint}.
      */
     signingEndpoint?: boolean;
+    /**
+     * Inject the Claude Code identity instruction on OAuth requests. Haiku 3.5
+     * predates the framing and rejects it; class rules set `false` there.
+     */
+    injectClaudeCodeInstruction?: boolean;
+    /** Strip image inputs before encoding (text-only serving of a multimodal id). */
+    stripImageInput?: boolean;
+    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
+    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
 }
 /**
  * Compatibility settings for Bedrock Converse prompt caching. Cache pricing is
@@ -596,9 +655,17 @@ export interface ResolvedOpenAISharedCompat {
     alwaysSendMaxTokens: boolean;
     openRouterRouting?: OpenAICompat["openRouterRouting"];
     /** Provider-specific wire model-id transform applied to the base id. */
-    wireModelIdMode: "raw" | "firepass" | "fireworks" | "openrouter";
+    wireModelIdMode: "raw" | "cline-pass" | "firepass" | "fireworks" | "openrouter";
     /** See {@link OpenAICompat.toolSchemaFlavor}. Read by both wire paths when converting tools. */
     toolSchemaFlavor?: OpenAICompat["toolSchemaFlavor"];
+    /** Strip image inputs before encoding (text-only serving of a multimodal id). */
+    stripImageInput: boolean;
+    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
+    thinkingLoopGuard?: OpenAICompat["thinkingLoopGuard"];
+    /** Flatten/reject leftover root `anyOf`/`oneOf` unions in strict tool schemas (xAI's function-calling validator 400s on them). */
+    rejectRootObjectUnion: boolean;
+    /** Retry without strict tools when the host rejects a strict grammar as too large (OpenRouter-Anthropic compiled-grammar overflow). */
+    retryWithoutStrictOnGrammarError: boolean;
 }
 /**
  * Fully-resolved chat-completions compat view: every detected default
@@ -606,7 +673,7 @@ 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" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -615,6 +682,8 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<Op
     toolStrictMode: ResolvedToolStrictMode;
     /** The model sits behind Vercel AI Gateway. */
     isVercelGatewayHost: boolean;
+    /** Send the normalized prompt-cache key as top-level `prompt_cache_key` on chat completions. */
+    supportsPromptCacheKey: boolean;
     dropThinkingWhenReasoningEffort: boolean;
     /** Complete alternate view for thinking-engaged requests; swap pointers, never spread. */
     whenThinking?: ResolvedOpenAICompat;
@@ -625,6 +694,14 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     strictResponsesPairing: boolean;
     supportsImageDetailOriginal: boolean;
     supportsObfuscationOptOut: boolean;
+    /**
+     * Whether `reasoning.context: "all_turns"` (full cross-turn reasoning
+     * replay) is accepted. Rule-owned: gpt-5.4+ wire generation on the Codex
+     * transport; earlier ids reject the value.
+     */
+    supportsAllTurnsReasoningContext: boolean;
+    /** Inject the `# Juice: 0 !important` developer item when reasoning is forced off (gpt-5.6+). */
+    requiresReasoningOffJuiceInstruction: boolean;
     /**
      * Whether `reasoning.summary` may be sent. First-party xAI `/v1/responses`
      * rejects the field; handlers pass `null` so the wire omits it instead of
@@ -635,6 +712,17 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     /** The model sits behind Vercel AI Gateway's Responses endpoint. */
     isVercelGatewayHost: boolean;
+    /**
+     * The configured endpoint is first-party OpenAI (`provider === "openai"` on
+     * an `api.openai.com` or unset baseUrl). Gates official-only Responses
+     * behavior: default-on stateful `previous_response_id` chaining and the
+     * `text.verbosity` field.
+     */
+    officialEndpoint: boolean;
+    /** Run Harmony-protocol leak detection/mitigation on streamed output. */
+    harmonyLeakMitigation: boolean;
+    /** Responses-surface prompt-cache marker dialect (OpenRouter-Anthropic `cache_control`). */
+    cacheControlFormat?: OpenAICompat["cacheControlFormat"];
 }
 /**
  * OpenRouter is a pseudo API: runtime dispatch can use either Responses
@@ -643,7 +731,9 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
  */
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs">> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard">> & {
+    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
+    thinkingLoopGuard?: AnthropicCompat["thinkingLoopGuard"];
     /**
      * 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
@@ -661,7 +751,7 @@ export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdle
 /**
  * Compatibility settings for the devin-agent (Codeium Cascade) API. Cascade
  * selects reasoning effort only by routing to a sibling model id (the
- * `thinking.effortRouting` baked by variant-collapse), never by a wire
+ * `thinking.effortRouting` baked by the compat collapse engine), never by a wire
  * reasoning/effort field, so the model-thinking deriver must not invent an
  * effort ladder from identity for these models.
  */
@@ -673,13 +763,62 @@ export interface DevinCompat {
      * effort) instead of a fabricated minimal/low/medium/high ladder.
      */
     trustExplicitThinkingOnly?: boolean;
+    /**
+     * Server-side router model. Providers must resolve the router through
+     * `AssignModel` and send the returned assignment JWT with the chat request.
+     */
+    modelRouter?: boolean;
+    /** Whether the upstream model supports native parallel tool calls. */
+    supportsParallelToolCalls?: boolean;
 }
 /** Fully-resolved devin-agent compat view. */
 export type ResolvedDevinCompat = Required<DevinCompat>;
+/**
+ * Compatibility settings for the Google API family (google-generative-ai,
+ * google-vertex, google-gemini-cli). Class-driven defaults come from the
+ * compat cascade; sparse overrides follow the same shape.
+ */
+export interface GoogleCompat {
+    /** Whether functionCall/functionResponse parts carry the `id` field. */
+    supportsFunctionPartId?: boolean;
+    /** Add the bypass sentinel to every unsigned Gemini function call. */
+    requiresSkipThoughtSignature?: boolean;
+    /** Add the bypass sentinel when a Gemini turn's first function call is unsigned. */
+    requiresSkipThoughtSignatureOnFirstFunctionCall?: boolean;
+    /** Drop unsigned thinking blocks from replayed history (Antigravity Claude). */
+    dropUnsignedThinking?: boolean;
+    /** Cloud Code Assist legacy `parameters` schema field instead of `parametersJsonSchema`. */
+    ccaLegacyParametersSchema?: boolean;
+    /** Whether multimodal (non-text) functionResponse parts are accepted (Gemini 3+). */
+    multimodalFunctionResponse?: boolean;
+    /** Stream-watchdog first-event timeout in ms; 0 disables it. */
+    streamFirstEventTimeoutMs?: number;
+    /** Stream-watchdog idle-timeout floor in ms. */
+    streamIdleTimeoutMs?: number;
+    /** Work around the Flash streaming leak (duplicate first chunk) on gemini-cli. */
+    flashStreamLeakWorkaround?: boolean;
+    /** Send the Claude thinking beta header on Antigravity Claude requests. */
+    claudeThinkingBetaHeader?: boolean;
+    /** Antigravity Claude tool-mode request framing. */
+    antigravityClaudeToolMode?: boolean;
+    /** Usage-label bucket Antigravity reports this model under. */
+    antigravityUsageLabel?: string;
+    /** Strip image inputs before encoding (text-only serving of a multimodal id). */
+    stripImageInput?: boolean;
+    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
+    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
+}
+/** Fully-resolved google-API compat view, materialized once by `buildModel`. */
+export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel">> & {
+    streamFirstEventTimeoutMs?: number;
+    streamIdleTimeoutMs?: number;
+    thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
+    antigravityUsageLabel?: string;
+};
 /** 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 "bedrock-converse-stream" ? BedrockCompat : 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 : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? GoogleCompat : 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 "bedrock-converse-stream" ? ResolvedBedrockCompat : 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 : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? ResolvedGoogleCompat : 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. */
@@ -730,12 +869,20 @@ export interface ModelCost extends TokenCost {
 export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5";
 export interface Model<TApi extends Api = Api> {
     id: string;
+    /**
+     * Structured model identity resolved by the compat engine: vendor lineage
+     * class, product family, and revision. Baked into models.json rows and
+     * materialized by `buildModel` for discovered/custom specs.
+     */
+    identity: ModelIdentity;
     /**
      * Whether provider-bound private-use glyphs require reversible ASCII tokenization.
      * Materialized by `buildModel`; request handlers read this capability instead of
      * inferring it from the transport API.
      */
     requiresGlyphTokenization?: boolean;
+    /** Whether this model requires Cursor's tool-schema combiner projection. */
+    requiresCursorToolSchemaProjection?: boolean;
     /**
      * Model id to send on the wire when it differs from `id`. Used by catalog
      * variants that present one upstream model under several local entries —
@@ -831,6 +978,23 @@ export interface Model<TApi extends Api = Api> {
     remoteCompaction?: RemoteCompactionConfig<TApi>;
     /** Provider-assigned priority value (lower = higher priority). */
     priority?: number;
+    /**
+     * Per-service-tier cost multipliers baked from the `service-tier-cost`
+     * catalog axis (e.g. `{ priority: 2.5 }`). Absent tiers use the API-generic
+     * defaults.
+     */
+    serviceTierCost?: Readonly<Partial<Record<"flex" | "priority", number>>>;
+    /**
+     * Provider-supplied one-line blurb for this model. Set only when an upstream
+     * ships one (Devin's `GetCliModelConfigs`); never synthesized locally.
+     */
+    description?: string;
+    /** Upstream marks this model as newly released. */
+    isNew?: boolean;
+    /** Upstream marks this model as beta / preview quality. */
+    isBeta?: boolean;
+    /** Upstream marks this model as one of its recommended picks. */
+    isRecommended?: boolean;
     /** Canonical thinking capability metadata for this model. */
     thinking?: ThinkingConfig;
     /**
@@ -873,7 +1037,7 @@ export interface Model<TApi extends Api = Api> {
  * vocabulary of `buildModel`. Identical to `Model` except `compat` carries the
  * sparse override shape and nothing is resolved yet.
  */
-export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "compatConfig" | "requiresGlyphTokenization" | "supportsComputerUseConfig"> {
+export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "identity" | "compatConfig" | "requiresGlyphTokenization" | "requiresCursorToolSchemaProjection" | "supportsComputerUseConfig"> {
     /** Sparse compatibility overrides; resolved into `Model.compat` by `buildModel`. */
     compat?: CompatConfigOf<TApi>;
 }

package/dist/types/wire/cline-pass.d.ts

@@ -0,0 +1,21 @@
+/**
+ * Cline API base shared by ClinePass subscription inference, free-tier models,
+ * roster discovery, login validation, and usage reporting. One constant so a
+ * host migration touches a single module.
+ */
+export declare const CLINEPASS_API_BASE_URL = "https://api.cline.bot/api/v1";
+/**
+ * Client-identity headers mirroring the official Cline CLI's request layer
+ * (sdk/packages/llms/src/providers/request-headers.ts). Cline's gateway gates
+ * some roster entries — certain free-tier models — to Cline product surfaces,
+ * and this is the contract that identifies a Cline client. Mirrored with Cline's
+ * blessing for a native integration.
+ *
+ * The gate accepted an arbitrary version string when probed (0.0.1 passed on
+ * 2026-08-13), so the pinned versions are mirror fidelity, not a minimum — but
+ * the full set is sent deliberately: partial mirrors are one gateway change
+ * away from a 403. `X-Task-ID` carries OMP's stable prompt-cache/session key
+ * when available; account and discovery calls omit it rather than inventing a
+ * per-request identity.
+ */
+export declare function clinePassClientHeaders(taskId?: string): Record<string, string>;

package/dist/types/wire/devin-proto.d.ts

@@ -0,0 +1,14 @@
+/**
+ * Protobuf decoding helpers for Devin's unary Connect RPCs. Kept separate from
+ * `wire/devin.ts`: that module stays protobuf-free for the synchronous catalog
+ * boot path, while this one is shared by discovery, chat, and usage callers
+ * that already load the generated schemas.
+ */
+import { type MessageCodec, type ProtoMessage } from "../discovery/protobuf.js";
+/**
+ * Decode a unary Devin Connect response. Edges variously return bare protobuf
+ * or a gzipped protobuf body; Bun's `fetch` normally decompresses first, so the
+ * direct decode is attempted before the gzip fallback. Returns `null` when
+ * neither representation decodes against `schema`.
+ */
+export declare function decodeDevinUnaryMessage<TMessage extends ProtoMessage>(schema: MessageCodec<TMessage>, payload: Uint8Array): TMessage | null;