@gajae-code/ai 0.4.5 → 0.5.1

package/CHANGELOG.md

@@ -2,6 +2,36 @@
 
 ## [Unreleased]
 
+## [0.5.1] - 2026-06-14
+
+### Fixed
+
+- Classified model/message limit exhaustion as persistent usage-limit errors so hosts fail fast or switch credentials instead of leaving sessions in an unbounded retry/working state.
+
+## [0.5.0] - 2026-06-13
+
+### Added
+
+- Added a generic tool-choice capability model: `toolChoiceSupport` compat enum (`none`/`auto`/`required`/`named`) available on every forced-choice-capable API, derived from the legacy `supportsToolChoice`/`supportsForcedToolChoice` booleans when absent, with a shared `resolveToolChoice` helper that clamps requested tool choices (`named` → `required` → omit) and returns structured degradation metadata.
+- Added a transparent one-shot fallback for forced `tool_choice` 400s ("tool_choice forces tool use is not compatible with this model" and equivalents): transports retry once without the forced field at a pre-content streaming boundary, record the discovery in an in-memory per-process incapability registry, and emit an internal non-rendered `toolChoiceIncapability` event. Applies to Anthropic, OpenAI Completions/Responses, Azure Responses, OpenAI code Responses, Bedrock (including event-stream `validationException`), Ollama, Google, and Gemini CLI transports.
+- Added bundled catalog entries for `kimi-code/kimi-k2.7-code`, `minimax-code/minimax-v3`, and `xai/grok-composer-2.5-fast`.
+- Added composer-harness anchor/edit discipline injection for Cursor Composer and Grok Composer models so provider-specific coding harness priors do not override GJC hashline/edit contracts.
+
+### Removed
+
+- Removed the retired `anthropic/claude-fable-5` bundled catalog entry.
+
+### Changed
+
+- Moved the Claude Mythos forced-tool-use incapability knowledge out of Anthropic request code into catalog compat defaults (`toolChoiceSupport: "auto"`), applied during catalog generation, dynamic discovery, and bundled-model loading via a shared predicate.
+- Google `toolConfig` mapping now sends `FunctionCallingConfig` mode `ANY` for both `required` and `any` requests instead of silently relaxing `required` to `AUTO`.
+- Optimized `EventStream` queue draining with a head-indexed queue to avoid repeated array shifts in hot streaming paths.
+- Clarified lazy builtin provider registration as the main provider loading path.
+
+### Fixed
+
+- Stripped `OpenAI-Beta` in the `openai-proxy` request transform profile so OpenAI-compatible proxies do not receive SDK beta headers.
+
 ## [0.4.5] - 2026-06-12
 
 ### Changed
@@ -10,7 +40,7 @@
 
 ### Fixed
 
-- Fixed direct Anthropic requests for Claude Fable/Mythos-style models that support tools but reject forced tool use by omitting forced `tool_choice` while preserving `auto`/`none` choices.
+- Fixed direct Anthropic requests for Claude Mythos-style models that support tools but reject forced tool use by omitting forced `tool_choice` while preserving `auto`/`none` choices.
 - Preserved catalog transport metadata for opencode-go `qwen3.7-max` model resolution.
 - Set SQLite auth-store `busy_timeout` before enabling WAL so initialization is reliable under contention.
 - Resolved provider credentials from inherited or GJC-owned environment sources instead of trusting the caller project's `.env` overlays.

package/dist/types/index.d.ts

@@ -33,6 +33,7 @@ export * from "./usage/claude";
 export * from "./usage/gemini";
 export * from "./usage/github-copilot";
 export * from "./usage/google-antigravity";
+export * from "./usage/grok-cli";
 export * from "./usage/kimi";
 export * from "./usage/minimax-code";
 export * from "./usage/openai-codex";
@@ -46,4 +47,5 @@ export type { OAuthCredentials, OAuthProvider, OAuthProviderId, OAuthProviderInf
 export * from "./utils/overflow";
 export * from "./utils/retry";
 export * from "./utils/schema";
+export * from "./utils/tool-choice-capability";
 export * from "./utils/validation";

package/dist/types/providers/amazon-bedrock.d.ts

@@ -7,15 +7,12 @@
  * Bun's native `HTTPS_PROXY` support.
  */
 import type { Effort } from "../model-thinking";
-import type { StreamFunction, StreamOptions, ThinkingBudgets } from "../types";
+import type { StreamFunction, StreamOptions, ThinkingBudgets, Tool, ToolChoice } from "../types";
 export type BedrockThinkingDisplay = "summarized" | "omitted";
 export interface BedrockOptions extends StreamOptions {
     region?: string;
     profile?: string;
-    toolChoice?: "auto" | "any" | "none" | {
-        type: "tool";
-        name: string;
-    };
+    toolChoice?: ToolChoice;
     reasoning?: Effort;
     thinkingBudgets?: ThinkingBudgets;
     interleavedThinking?: boolean;
@@ -33,4 +30,31 @@ export interface BedrockOptions extends StreamOptions {
      */
     thinkingDisplay?: BedrockThinkingDisplay;
 }
+interface WireToolSpec {
+    toolSpec: {
+        name: string;
+        description: string;
+        inputSchema: {
+            json: unknown;
+        };
+    };
+}
+interface WireToolChoice {
+    auto?: Record<string, never>;
+    any?: Record<string, never>;
+    tool?: {
+        name: string;
+    };
+}
+interface WireToolConfig {
+    tools: WireToolSpec[];
+    toolChoice?: WireToolChoice;
+}
 export declare const streamBedrock: StreamFunction<"bedrock-converse-stream">;
+export declare function stripBedrockForcedToolChoiceForRetry<T extends {
+    toolConfig?: {
+        toolChoice?: unknown;
+    };
+}>(body: T): T;
+export declare function convertToolConfig(tools: Tool[] | undefined, toolChoice: BedrockOptions["toolChoice"]): WireToolConfig | undefined;
+export {};

package/dist/types/providers/composer-discipline.d.ts

@@ -0,0 +1,27 @@
+/**
+ * Anchor/edit discipline for composer-harness models (xai grok-composer-*,
+ * cursor composer-*).
+ *
+ * Composer models are trained on a proprietary coding-agent harness
+ * (Cursor / Grok Build) and carry habits that break this agent's hashline
+ * edit workflow when driven through a generic provider. Observed in live
+ * sessions with grok-composer-2.5-fast:
+ *
+ *  - they print files with shell commands (`sed -n`, `cat`, `grep -n`) or
+ *    python heredocs whose output carries NO line anchors, then FABRICATE the
+ *    2-char anchor hash the edit tool requires (e.g. guessed "617hp" where
+ *    the file had "617ca" → "Edit rejected: N anchors do not match");
+ *  - they mutate files out-of-band via python heredocs (pathlib write_text /
+ *    str.replace), which invalidates every previously seen anchor and defeats
+ *    the read-cache snapshot that powers stale-anchor recovery;
+ *  - they arithmetically renumber anchors after their own edits instead of
+ *    copying them from the latest tool output;
+ *  - they leak reasoning prose into heredoc bodies, producing shell/python
+ *    syntax errors.
+ *
+ * This prompt is the per-request countermeasure, pinned ahead of the host
+ * system prompt on both the openai-completions path and the cursor RPC path.
+ */
+/** Matches composer-harness model ids on any provider (xai grok-composer-*, cursor composer-*). */
+export declare function isComposerHarnessModel(modelId: string): boolean;
+export declare const COMPOSER_EDIT_DISCIPLINE_PROMPT = "File-editing discipline for this harness (this OVERRIDES contrary habits from your training):\n\n- Read file contents ONLY with the provided read/search tools. NEVER print files through shell commands (sed, cat, awk, head, grep) or scripts \u2014 that output carries no line anchors, and the edit tool accepts ONLY anchors.\n- Modify files ONLY with the provided edit/write tools. NEVER mutate files through shell redirection, sed -i, or inline python scripts \u2014 out-of-band writes invalidate every known anchor and break edit recovery.\n- A line anchor (e.g. \"42sr\") is a line number plus a 2-char content hash. You CANNOT compute the hash yourself: copy anchors verbatim from the MOST RECENT read/search/edit output of that exact file. NEVER guess, renumber, or arithmetically shift an anchor.\n- After ANY edit to a file (including your own), anchors you saw earlier are stale. Re-read the edited region, or copy the fresh anchors printed in the edit result, before issuing the next edit.\n- If an edit is rejected with \"anchors do not match\", the rejection message prints the current lines WITH fresh anchors. Retry using exactly those printed anchors.\n- A shell command string must contain only the command itself. NEVER interleave reasoning or commentary into command strings or heredocs.";

package/dist/types/providers/cursor.d.ts

@@ -34,7 +34,7 @@ export declare function resolveExecHandler<TArgs, TResult>(args: TArgs, handler:
  * When no system prompts are provided, returns a single default greeting so we never emit
  * an empty `rootPromptMessagesJson` head.
  */
-export declare function buildCursorSystemPromptJsons(systemPrompt: readonly string[] | undefined): string[];
+export declare function buildCursorSystemPromptJsons(systemPrompt: readonly string[] | undefined, modelId?: string): string[];
 /** Exported for tests: decodes Cursor history blobs built from conversation messages. */
 export declare function buildCursorHistoryForTest(messages: Message[]): {
     rootPromptMessagesJson: unknown[];

package/dist/types/providers/google-gemini-cli.d.ts

@@ -7,7 +7,7 @@ import { type GoogleThinkingLevel } from "./google-shared";
  */
 export type { GoogleThinkingLevel };
 export interface GoogleGeminiCliOptions extends StreamOptions {
-    toolChoice?: "auto" | "none" | "any";
+    toolChoice?: "auto" | "none" | "any" | "required";
     /**
      * Thinking/reasoning configuration.
      * - Gemini 2.x models: use `budgetTokens` to set the thinking budget

package/dist/types/providers/google-shared.d.ts

@@ -19,7 +19,7 @@ export type GoogleThinkingLevel = "THINKING_LEVEL_UNSPECIFIED" | "MINIMAL" | "LO
  * `google-gemini-cli` uses a different transport and request shape — do not extend this for it.
  */
 export interface GoogleSharedStreamOptions extends StreamOptions {
-    toolChoice?: "auto" | "none" | "any";
+    toolChoice?: "auto" | "none" | "any" | "required";
     thinking?: {
         enabled: boolean;
         budgetTokens?: number;
@@ -161,3 +161,13 @@ export declare function streamGoogleGenAI<T extends "google-generative-ai" | "go
     retainTextSignature?: boolean;
     prepare: () => GoogleGenAIRequestPlan | Promise<GoogleGenAIRequestPlan>;
 }): AssistantMessageEventStream;
+/**
+ * Lift the SDK's `params.config` fields out of `config` and place them where the
+ * Gemini / Vertex AI REST API expects them on the request body. Mirrors the
+ * generateContentParametersTo{Mldev,Vertex} transformation in @google/genai
+ * for the subset of fields this codebase actually sets.
+ *
+ * `abortSignal` is intentionally dropped — the SDK propagates it via `fetch.signal`,
+ * which our caller already wires up through `options.signal`.
+ */
+export declare function paramsToWireBody(params: GenerateContentParameters): Record<string, unknown>;

package/dist/types/providers/ollama.d.ts

@@ -1,6 +1,41 @@
-import type { StreamFunction, StreamOptions, ToolChoice } from "../types";
+import type { Context, Model, StreamFunction, StreamOptions, ToolChoice } from "../types";
 export interface OllamaChatOptions extends StreamOptions {
     reasoning?: "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
     toolChoice?: ToolChoice;
 }
+type OllamaFunctionTool = {
+    type: "function";
+    function: {
+        name: string;
+        description: string;
+        parameters: Record<string, unknown>;
+    };
+};
+type OllamaMessage = {
+    role: "system" | "user" | "assistant" | "tool";
+    content: string;
+    images?: string[];
+    thinking?: string;
+    tool_calls?: Array<{
+        type: "function";
+        function: {
+            index?: number;
+            name: string;
+            arguments: Record<string, unknown>;
+        };
+    }>;
+    tool_name?: string;
+};
+export declare function createChatBody(model: Model<"ollama-chat">, context: Context, options: OllamaChatOptions | undefined): {
+    model: string;
+    messages: OllamaMessage[];
+    tools?: OllamaFunctionTool[] | undefined;
+    think?: "high" | "low" | "medium" | boolean | undefined;
+    tool_choice?: "auto" | "none" | "required" | undefined;
+    options?: {
+        num_predict: number;
+    } | undefined;
+    stream: boolean;
+};
 export declare const streamOllama: StreamFunction<"ollama-chat">;
+export {};

package/dist/types/providers/openai-completions-compat.d.ts

@@ -1,10 +1,12 @@
 import type { Model, OpenAICompat } from "../types";
 type ResolvedToolStrictMode = NonNullable<OpenAICompat["toolStrictMode"]> | "mixed";
-export type ResolvedOpenAICompat = Required<Omit<OpenAICompat, "openRouterRouting" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode">> & {
+export type ResolvedOpenAICompat = Required<Omit<OpenAICompat, "openRouterRouting" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolChoiceSupport">> & {
     openRouterRouting?: OpenAICompat["openRouterRouting"];
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     toolStrictMode: ResolvedToolStrictMode;
+    /** Optional explicit capability override; resolved via deriveToolChoiceSupport. */
+    toolChoiceSupport?: OpenAICompat["toolChoiceSupport"];
 };
 /**
  * Detect compatibility settings from provider and baseUrl for known providers.

package/dist/types/providers/register-builtins.d.ts

@@ -6,9 +6,9 @@
  * openai) at startup. The loaded module promise is cached so subsequent calls
  * reuse the same import.
  *
- * NOTE: stream.ts currently imports providers directly, so this file is not yet
- * wired into the main streaming path. It provides the infrastructure for lazy
- * loading that can be integrated when stream.ts is refactored.
+ * stream.ts imports its provider stream functions from this module (see the
+ * lazy wrappers below), so this file IS the main streaming path's provider
+ * loader: heavy SDKs stay out of the CLI startup parse graph.
  */
 import type { AssistantMessageEventStream, Context, Model, OptionsForApi } from "../types";
 import { AssistantMessageEventStream as EventStreamImpl } from "../utils/event-stream";

package/dist/types/types.d.ts

@@ -68,6 +68,16 @@ export type ToolChoice = "auto" | "none" | "any" | "required" | {
     type: "tool";
     name: string;
 };
+export type ToolChoiceSupport = "none" | "auto" | "required" | "named";
+export type ToolChoiceSupportSource = "static" | "derived" | "runtime";
+export interface ToolChoiceCompat {
+    /** Maximum supported tool_choice level. */
+    toolChoiceSupport?: ToolChoiceSupport;
+    /** Legacy flag for accepting the tool_choice parameter. */
+    supportsToolChoice?: boolean;
+    /** Legacy flag for forced tool_choice support. */
+    supportsForcedToolChoice?: boolean;
+}
 export type CacheRetention = "none" | "short" | "long";
 /**
  * Service tier hint for processing priority / cost control.
@@ -576,12 +586,22 @@ export type AssistantMessageEvent = {
     contentIndex?: undefined;
     reason: Extract<StopReason, "aborted" | "error">;
     error: AssistantMessage;
+} | {
+    type: "toolChoiceIncapability";
+    contentIndex?: undefined;
+    api: string;
+    provider: string;
+    model: string;
+    requestedLevel: ToolChoiceSupport;
+    resolvedLevel: ToolChoiceSupport;
+    reason: string;
+    registryKey: string;
 };
 /**
  * Compatibility settings for openai-completions API.
  * Use this to override URL-based auto-detection for custom providers.
  */
-export interface OpenAICompat {
+export interface OpenAICompat extends ToolChoiceCompat {
     /** Whether the provider supports the `store` field. Default: auto-detected from URL. */
     supportsStore?: boolean;
     /** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
@@ -627,6 +647,8 @@ export interface OpenAICompat {
     requiresAssistantContentForToolCalls?: boolean;
     /** Whether the provider supports the `tool_choice` parameter. Default: true. */
     supportsToolChoice?: boolean;
+    /** Whether `tool_choice` may force a tool (`required` / named tool). Default: true. */
+    supportsForcedToolChoice?: boolean;
     /**
      * Drop reasoning fields (`reasoning_effort`, OpenRouter `reasoning`) for
      * the request when `tool_choice` forces a tool call. Mirrors the Anthropic
@@ -658,7 +680,7 @@ export interface OpenAICompat {
  * Use this to disable features that strict-by-default Anthropic accepts but
  * that proxy gateways (Vertex AI, AWS Bedrock-style fronts, etc.) reject.
  */
-export interface AnthropicCompat {
+export interface AnthropicCompat extends ToolChoiceCompat {
     /**
      * Drop the top-level `strict: true` field on tool definitions. Vertex AI's
      * Anthropic-compatible endpoint rejects unknown tool fields with
@@ -770,7 +792,7 @@ export interface Model<TApi extends Api = any> {
     /** Canonical thinking capability metadata for this model. */
     thinking?: ThinkingConfig;
     /** Compatibility overrides per API. If not set, auto-detected from baseUrl. */
-    compat?: TApi extends "openai-completions" | "openai-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : never;
+    compat?: TApi extends "openai-completions" | "openai-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "bedrock-converse-stream" | "google-generative-ai" | "google-gemini-cli" | "google-vertex" | "ollama-chat" | "azure-openai-responses" | "openai-codex-responses" ? ToolChoiceCompat : never;
     /**
      * Which shape to use when exposing the OpenAI code backend `apply_patch` tool to this model.
      * Generated catalog policy sets `"freeform"` for first-party GPT-5 Responses

package/dist/types/usage/grok-cli.d.ts

@@ -0,0 +1,10 @@
+import type { CredentialRankingStrategy, UsageProvider } from "../usage";
+interface BillingUsage {
+    monthlyLimit: number;
+    used: number;
+    billingPeriodEnd: string;
+}
+export declare function parseGrokCliBillingUsage(payload: unknown): BillingUsage;
+export declare const grokCliUsageProvider: UsageProvider;
+export declare const grokCliRankingStrategy: CredentialRankingStrategy;
+export {};

package/dist/types/utils/event-stream.d.ts

@@ -1,7 +1,6 @@
 import type { AssistantMessage, AssistantMessageEvent } from "../types";
 export declare class EventStream<T, R = T> implements AsyncIterable<T> {
     #private;
-    queue: T[];
     waiting: Array<{
         resolve: (value: IteratorResult<T>) => void;
         reject: (err: unknown) => void;
@@ -13,6 +12,12 @@ export declare class EventStream<T, R = T> implements AsyncIterable<T> {
     isComplete: (event: T) => boolean;
     extractResult: (event: T) => R;
     constructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R);
+    /**
+     * Read-only snapshot of the not-yet-consumed events. Always a fresh copy:
+     * external code can never mutate internal queue state or observe head-index
+     * tombstones, so the deque cannot desynchronize.
+     */
+    get queue(): T[];
     push(event: T): void;
     deliver(event: T): void;
     end(result?: R): void;

package/dist/types/utils/oauth/xai.d.ts

@@ -8,16 +8,23 @@ interface XaiDiscovery {
     authorizationEndpoint: string;
     tokenEndpoint: string;
 }
+export interface XaiOAuthFlowOptions {
+    extraAuthorizeParams?: Readonly<Record<string, string>>;
+}
+export interface XaiOAuthRefreshOptions {
+    signal?: AbortSignal;
+    extraTokenParams?: Readonly<Record<string, string>>;
+}
 export declare function discoverXaiOAuthEndpoints(signal?: AbortSignal): Promise<XaiDiscovery>;
 export declare class XaiOAuthFlow extends OAuthCallbackFlow {
     #private;
-    constructor(ctrl: OAuthController);
+    constructor(ctrl: OAuthController, options?: XaiOAuthFlowOptions);
     generateAuthUrl(state: string, redirectUri: string): Promise<{
         url: string;
         instructions?: string;
     }>;
     exchangeToken(code: string, _state: string, redirectUri: string): Promise<OAuthCredentials>;
 }
-export declare function loginXai(ctrl: OAuthController): Promise<OAuthCredentials>;
-export declare function refreshXaiToken(refreshToken: string, signal?: AbortSignal): Promise<OAuthCredentials>;
+export declare function loginXai(ctrl: OAuthController, options?: XaiOAuthFlowOptions): Promise<OAuthCredentials>;
+export declare function refreshXaiToken(refreshToken: string, options?: AbortSignal | XaiOAuthRefreshOptions): Promise<OAuthCredentials>;
 export {};

package/dist/types/utils/tool-choice-capability.d.ts

@@ -0,0 +1,41 @@
+import type { Api, Model, ToolChoice, ToolChoiceCompat, ToolChoiceSupport, ToolChoiceSupportSource } from "../types";
+/**
+ * Claude Mythos accepts tools but rejects forced tool use (Anthropic 400:
+ * "tool_choice forces tool use is not compatible with this model"). Catalog
+ * generation and dynamic discovery use this to default `toolChoiceSupport`.
+ */
+export declare function isClaudeForcedToolChoiceIncapableModelId(modelId: string): boolean;
+/** Derives the effective static tool-choice support from compatibility flags. */
+export declare function deriveToolChoiceSupport(compat: ToolChoiceCompat | undefined): {
+    support: ToolChoiceSupport;
+    source: "static" | "derived";
+};
+/** Returns the registry key used for runtime tool-choice capability overrides. */
+export declare function toolChoiceRegistryKey(model: Model<Api>): string;
+/** Returns the current runtime tool-choice capability override for a model. */
+export declare function getToolChoiceCapabilityOverride(model: Model<Api>): ToolChoiceSupport | undefined;
+/** Clears runtime tool-choice capability overrides for tests. */
+export declare function clearToolChoiceIncapabilityRegistryForTests(): void;
+/** Records a discovered maximum supported tool-choice level for a model. */
+export declare function markToolChoiceIncapability(model: Model<Api>, maxSupport: ToolChoiceSupport, reason?: string): void;
+/**
+ * Resolves a requested tool_choice against static and runtime capability limits.
+ * `compat` overrides `model.compat` for transports that layer URL/provider
+ * detection on top of explicit model overrides (e.g. resolveOpenAICompat).
+ */
+export declare function resolveToolChoice(model: Model<Api>, requested: ToolChoice | undefined, compat?: ToolChoiceCompat): ResolveToolChoiceResult;
+/** Detects provider errors indicating forced tool_choice is unsupported. */
+export declare function isForcedToolChoiceUnsupportedError(error: unknown, sentForcedToolChoice: boolean): boolean;
+export type { ToolChoiceCompat, ToolChoiceSupport, ToolChoiceSupportSource } from "../types";
+export interface ResolveToolChoiceResult {
+    requestedChoice: ToolChoice | undefined;
+    requestedLevel: ToolChoiceSupport;
+    resolvedChoice: ToolChoice | undefined;
+    resolvedLevel: ToolChoiceSupport;
+    support: ToolChoiceSupport;
+    supportSource: ToolChoiceSupportSource;
+    degraded: boolean;
+    reason?: string;
+    registryKey: string;
+    targetToolName?: string;
+}

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/ai",
-	"version": "0.4.5",
+	"version": "0.5.1",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://gaebal-gajae.dev",
 	"author": "Yeachan-Heo",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@gajae-code/utils": "0.4.5",
+		"@gajae-code/utils": "0.5.1",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/auth-storage.ts

@@ -27,6 +27,7 @@ import { claudeRankingStrategy, claudeUsageProvider } from "./usage/claude";
 import { googleGeminiCliUsageProvider } from "./usage/gemini";
 import { githubCopilotUsageProvider } from "./usage/github-copilot";
 import { antigravityUsageProvider } from "./usage/google-antigravity";
+import { grokCliRankingStrategy, grokCliUsageProvider } from "./usage/grok-cli";
 import { kimiUsageProvider } from "./usage/kimi";
 import { codexRankingStrategy, openaiCodexUsageProvider } from "./usage/openai-codex";
 import { zaiUsageProvider } from "./usage/zai";
@@ -370,6 +371,7 @@ const DEFAULT_USAGE_PROVIDERS: UsageProvider[] = [
 	claudeUsageProvider,
 	zaiUsageProvider,
 	githubCopilotUsageProvider,
+	grokCliUsageProvider,
 ];
 
 const DEFAULT_USAGE_PROVIDER_MAP = new Map<Provider, UsageProvider>(
@@ -498,6 +500,7 @@ function resolveDefaultUsageProvider(provider: Provider): UsageProvider | undefi
 const DEFAULT_RANKING_STRATEGIES = new Map<Provider, CredentialRankingStrategy>([
 	["openai-codex", codexRankingStrategy],
 	["anthropic", claudeRankingStrategy],
+	["grok-build", grokCliRankingStrategy],
 ]);
 
 function resolveDefaultRankingStrategy(provider: Provider): CredentialRankingStrategy | undefined {

package/src/index.ts

@@ -33,6 +33,7 @@ export * from "./usage/claude";
 export * from "./usage/gemini";
 export * from "./usage/github-copilot";
 export * from "./usage/google-antigravity";
+export * from "./usage/grok-cli";
 export * from "./usage/kimi";
 export * from "./usage/minimax-code";
 export * from "./usage/openai-codex";
@@ -51,4 +52,5 @@ export type {
 export * from "./utils/overflow";
 export * from "./utils/retry";
 export * from "./utils/schema";
+export * from "./utils/tool-choice-capability";
 export * from "./utils/validation";

package/src/model-thinking.ts

@@ -1,5 +1,6 @@
 import { resolveOpenAICompat } from "./providers/openai-completions-compat";
 import type { Api, Model as ApiModel, ThinkingConfig } from "./types";
+import { isClaudeForcedToolChoiceIncapableModelId } from "./utils/tool-choice-capability";
 
 /** User-facing thinking levels, ordered least to most intensive. */
 export const enum Effort {
@@ -382,6 +383,14 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	} else {
 		delete model.applyPatchToolType;
 	}
+	if (
+		(model.api === "anthropic-messages" || model.api === "bedrock-converse-stream") &&
+		isClaudeForcedToolChoiceIncapableModelId(model.id)
+	) {
+		// Claude Mythos accepts tools but rejects forced tool use (Anthropic
+		// 400: "tool_choice forces tool use is not compatible with this model").
+		model.compat = { ...(model.compat ?? {}), toolChoiceSupport: "auto" } as typeof model.compat;
+	}
 	if (parsedModel.family === "anthropic") {
 		applyAnthropicCatalogPolicy(model, parsedModel);
 	}

package/src/models.json

@@ -18021,6 +18021,40 @@
 				"minLevel": "minimal",
 				"maxLevel": "high"
 			}
+		},
+		"kimi-k2.7-code": {
+			"id": "kimi-k2.7-code",
+			"name": "Kimi K2.7 Code",
+			"api": "openai-completions",
+			"provider": "kimi-code",
+			"baseUrl": "https://api.kimi.com/coding/v1",
+			"headers": {
+				"User-Agent": "KimiCLI/1.0",
+				"X-Msh-Platform": "kimi_cli"
+			},
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 65536,
+			"compat": {
+				"thinkingFormat": "zai",
+				"reasoningContentField": "reasoning_content",
+				"supportsDeveloperRole": false
+			},
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
 		}
 	},
 	"litellm": {
@@ -35153,6 +35187,37 @@
 				"minLevel": "minimal",
 				"maxLevel": "high"
 			}
+		},
+		"minimax-v3": {
+			"id": "minimax-v3",
+			"name": "MiniMax-V3",
+			"api": "openai-completions",
+			"provider": "minimax-code",
+			"baseUrl": "https://api.minimax.io/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 512000,
+			"maxTokens": 128000,
+			"compat": {
+				"supportsStore": false,
+				"supportsDeveloperRole": false,
+				"supportsReasoningEffort": false,
+				"reasoningContentField": "reasoning_content"
+			},
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
 		}
 	},
 	"minimax-code-cn": {
@@ -70520,6 +70585,33 @@
 				"maxLevel": "high"
 			}
 		},
+		"grok-composer-2.5-fast": {
+			"id": "grok-composer-2.5-fast",
+			"name": "Grok Composer 2.5 Fast",
+			"api": "openai-completions",
+			"provider": "xai",
+			"baseUrl": "https://api.x.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 200000,
+			"maxTokens": 64000,
+			"compat": {
+				"supportsReasoningEffort": false
+			},
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
+		},
 		"grok-vision-beta": {
 			"id": "grok-vision-beta",
 			"name": "Grok Vision Beta",
@@ -70956,6 +71048,30 @@
 				"maxLevel": "xhigh"
 			}
 		},
+		"glm-5.2": {
+			"id": "glm-5.2",
+			"name": "GLM-5.2",
+			"api": "anthropic-messages",
+			"provider": "zai",
+			"baseUrl": "https://api.z.ai/api/anthropic",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 200000,
+			"maxTokens": 131072,
+			"thinking": {
+				"mode": "budget",
+				"minLevel": "minimal",
+				"maxLevel": "xhigh"
+			}
+		},
 		"glm-5v-turbo": {
 			"id": "glm-5v-turbo",
 			"name": "GLM-5V-Turbo",