@oh-my-pi/pi-ai 17.2.9 → 17.2.11

package/CHANGELOG.md

@@ -2,6 +2,30 @@
 
 ## [Unreleased]
 
+## [17.2.11] - 2026-08-07
+
+### Breaking Changes
+
+- Fixed handling of GitHub Copilot's model_not_available_for_integrator error to prevent unnecessary retries, preserving the actionable available models list.
+
+### Added
+
+- Added support for reporting Cursor personal monthly USD quotas and remaining balances, labeled by verified profile email accounts.
+
+### Fixed
+
+- Fixed an issue where ANTHROPIC_BASE_URL was ignored for Anthropic chat requests, ensuring requests are routed to the configured host and forwarding ANTHROPIC_CUSTOM_HEADERS to non-official gateways.
+- Fixed an issue where a legacy pre-organization login credential could persist and cause a permanent error row in omp usage even after a successful organization-scoped re-login.
+- Fixed an issue where lazy provider streams (including Amazon Bedrock, Google, Cursor, Devin, and Ollama) ignored model-specific idle timeouts, which previously caused healthy but slow reasoning turns to prematurely time out.
+- Improved error classification for Simplified Chinese quota-exhaustion and rate-limit messages, ensuring affected credentials are correctly rotated or backed off instead of being treated as unknown errors.
+- Classified subscription and plan-cap 429 responses as rotatable usage limits rather than transient rate-limit throttles, enabling smoother credential rotation.
+
+## [17.2.10] - 2026-08-06
+
+### Breaking Changes
+
+- Removed the `zod` dependency and `z`/`ZodType` re-exports. Tool schemas now use `omptype` `type()` schemas, with Zod-style authoring still available via `@oh-my-pi/omptype/zod`.
+
 ## [17.2.9] - 2026-08-05
 
 ### Fixed

package/README.md

@@ -94,21 +94,18 @@ npm install @oh-my-pi/pi-ai
 ## Quick Start
 
 ```typescript
-import { z, getModel, stream, complete, Context, Tool } from "@oh-my-pi/pi-ai";
+import { getModel, stream, complete, Context, Tool, type } from "@oh-my-pi/pi-ai";
 
 // Fully typed with auto-complete support for both providers and models
 const model = getModel("openai", "gpt-4o-mini");
 
-// Define tools with Zod schemas for type safety and validation
+// Define tools with omptype schemas for type safety and validation
 const tools: Tool[] = [
 	{
 		name: "get_time",
 		description: "Get the current time",
-		parameters: z.object({
-			timezone: z
-				.string()
-				.optional()
-				.describe("Optional timezone (e.g., America/New_York)"),
+		parameters: type({
+			"timezone?": type("string").describe("Optional timezone (e.g., America/New_York)"),
 		}),
 	},
 ];
@@ -221,31 +218,30 @@ for (const block of response.content) {
 
 ## Tools
 
-Tools enable LLMs to interact with external systems. This library uses **Zod** schemas for type-safe tool definitions with automatic validation. Schemas are converted to JSON Schema for providers as needed.
+Tools enable LLMs to interact with external systems. Omptype schemas provide type-safe definitions, runtime validation, and JSON Schema conversion for providers.
 
 ### Defining Tools
 
 ```typescript
-import { z, Tool } from "@oh-my-pi/pi-ai";
+import { type Tool, type } from "@oh-my-pi/pi-ai";
 
-// Define tool parameters with Zod
 const weatherTool: Tool = {
 	name: "get_weather",
 	description: "Get current weather for a location",
-	parameters: z.object({
-		location: z.string().describe("City name or coordinates"),
-		units: z.enum(["celsius", "fahrenheit"]).default("celsius"),
+	parameters: type({
+		location: type("string").describe("City name or coordinates"),
+		units: type.enumerated("celsius", "fahrenheit").default("celsius"),
 	}),
 };
 
 const bookMeetingTool: Tool = {
 	name: "book_meeting",
 	description: "Schedule a meeting",
-	parameters: z.object({
-		title: z.string().min(1),
-		startTime: z.string().describe("ISO 8601 date-time"),
-		endTime: z.string().describe("ISO 8601 date-time"),
-		attendees: z.array(z.email()).min(1),
+	parameters: type({
+		title: type("string").atLeastLength(1),
+		startTime: type("string").describe("ISO 8601 date-time"),
+		endTime: type("string").describe("ISO 8601 date-time"),
+		attendees: type("string.email").array().atLeastLength(1),
 	}),
 };
 ```
@@ -345,7 +341,7 @@ for await (const event of s) {
 
 ### Validating Tool Arguments
 
-When using `agentLoop`, tool arguments are automatically validated against your Zod parameter schemas before execution. If validation fails, the error is returned to the model as a tool result, allowing it to retry.
+When using `agentLoop`, tool arguments are automatically validated against their omptype schemas before execution. Validation failures are returned to the model as tool results so it can retry.
 
 When implementing your own tool execution loop with `stream()` or `complete()`, use `validateToolCall` to validate arguments before passing them to your tools:
 

package/dist/types/error/flags.d.ts

@@ -65,10 +65,10 @@ export declare function isGrammarError(error: unknown): boolean;
  */
 export declare function isFastModeUnsupported(error: unknown): boolean;
 /**
- * GitHub Copilot 400 rejecting a model its own `/models` catalog advertises —
- * transient fleet skew, not a malformed request. Reads the structural `code`
- * through the SDK/body envelopes, then falls back to the stringified body both
- * SDK families put in `message` (shapes drift; the wire text does not).
+ * GitHub Copilot 400 `model_not_supported` response for a model advertised by
+ * `/models` — transient fleet skew, not a malformed request. Reads the
+ * structural `code` through the SDK/body envelopes, then falls back to the
+ * stringified body both SDK families put in `message`.
  */
 export declare function isCopilotTransientModelError(error: unknown): boolean;
 export declare function classifyMessage(message: {

package/dist/types/index.d.ts

@@ -1,5 +1,4 @@
 export { type Type, type } from "@oh-my-pi/omptype";
-export { type ZodType, z } from "zod/v4";
 export * from "./api-registry.js";
 export type * from "./auth-broker/index.js";
 export type { AuthGatewayBootOptions, ModelResolver } from "./auth-gateway/server.js";

package/dist/types/registry/oauth/cursor.d.ts

@@ -12,4 +12,5 @@ export declare function pollCursorAuth(uuid: string, verifier: string): Promise<
 }>;
 export declare function loginCursor(onAuthUrl: (url: string) => void, onPollStart?: () => void): Promise<OAuthCredentials>;
 export declare function refreshCursorToken(apiKeyOrRefreshToken: string): Promise<OAuthCredentials>;
+export declare function extractCursorAccessTokenUserId(accessToken: string): string | undefined;
 export declare function isCursorTokenExpiringSoon(token: string, thresholdSeconds?: number): boolean;

package/dist/types/registry/oauth/index.d.ts

@@ -6,6 +6,10 @@ export type * from "./types.js";
  * Register a custom OAuth provider.
  */
 export declare function registerOAuthProvider(provider: OAuthProviderInterface): void;
+/**
+ * Remove a custom OAuth provider by ID.
+ */
+export declare function unregisterOAuthProvider(id: string): void;
 /**
  * Get a custom OAuth provider by ID.
  */

package/dist/types/types.d.ts

@@ -4,7 +4,6 @@ import type { Type } from "@oh-my-pi/omptype";
 import type { DeleteArgs, DeleteResult, DiagnosticsArgs, DiagnosticsResult, GrepArgs, GrepResult, LsArgs, LsResult, McpResult, PiBashExecArgs, PiBashExecResult, PiEditExecArgs, PiEditExecResult, PiFindExecArgs, PiFindExecResult, PiGrepExecArgs, PiGrepExecResult, PiLsExecArgs, PiLsExecResult, PiReadExecArgs, PiReadExecResult, PiWriteExecArgs, PiWriteExecResult, ReadArgs, ReadResult, ShellArgs, ShellResult, WriteArgs, WriteResult } from "@oh-my-pi/pi-catalog/discovery/cursor-gen/agent_pb";
 import type { Effort } from "@oh-my-pi/pi-catalog/effort";
 import type { Api, FetchImpl, Model, Provider, ThinkingBudgets, Usage } from "@oh-my-pi/pi-catalog/types";
-import type { ZodType, z } from "zod/v4";
 import type { ApiKey } from "./auth-retry.js";
 import type { BedrockOptions } from "./providers/amazon-bedrock.js";
 import type { AnthropicOptions } from "./providers/anthropic.js";
@@ -650,8 +649,10 @@ export interface DeveloperMessage {
     providerPayload?: ProviderPayload;
     timestamp: number;
 }
+/** How an automatic retry recovered or ultimately settled a failed attempt. */
 export type AssistantRetryRecoveryKind = "credential" | "model" | "wait" | "plain";
-export interface AssistantRetryRecovery {
+/** Persisted presentation state for an assistant error superseded by an automatic retry saga. */
+export type AssistantRetryRecovery = {
     kind: "auto-retry";
     status: "recovered";
     attempt: number;
@@ -664,7 +665,13 @@ export interface AssistantRetryRecovery {
         provider: string;
         model: string;
     };
-}
+} | {
+    kind: "auto-retry";
+    status: "superseded";
+    attempt: number;
+    recovery: AssistantRetryRecoveryKind;
+    note: string;
+};
 export interface ContextSnapshot {
     promptTokens: number;
     nonMessageTokens: number;
@@ -928,12 +935,12 @@ export type TJsonSchema = Record<string, unknown>;
 /**
  * Schema type accepted by the {@link Tool} interface.
  *
- * Canonical authoring uses Zod or ArkType. Extension compat may supply a JSON
- * Schema object (including TypeBox static schema objects).
+ * Canonical authoring uses ArkType. Extension compat may supply a JSON Schema
+ * object (including TypeBox static schema objects).
  */
-export type TSchema = ZodType | Type | TJsonSchema;
+export type TSchema = Type | TJsonSchema;
 /** Resolve parameter types for tool execution / handlers. */
-export type Static<S> = S extends ZodType ? z.infer<S> : S extends Type ? S["infer"] : S extends {
+export type Static<S> = S extends Type ? S["infer"] : S extends {
     static: infer T;
 } ? T : unknown;
 export interface ToolCallExample<TArgs = Record<string, unknown>> {

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

@@ -1,3 +1,4 @@
 import type { UsageProvider, UsageReport } from "../usage.js";
+export declare function parseCursorIndividualUsage(payload: unknown, fetchedAt?: number): UsageReport | null;
 export declare function parseCursorUsage(payload: unknown, fetchedAt?: number): UsageReport | null;
 export declare const cursorUsageProvider: UsageProvider;

package/dist/types/utils/http-inspector.d.ts

@@ -38,11 +38,10 @@ export declare function finalizeErrorMessage(error: unknown, rawRequestDump: Raw
  * Rewrite error message for GitHub Copilot request failures.
  * Must run AFTER finalizeErrorMessage since it replaces the message entirely.
  *
- * 400 model-unavailable = Copilot fleet skew. A model that `/models` advertises
- *        (claude-sonnet-4.6, claude-opus-4.6, gpt-5.4, gpt-5.3-codex, ...)
- *        flaps between 200 and 400 because only part of Copilot's fleet has it
- *        in the integrator allowlist. After the in-request retry exhausts,
- *        surface guidance rather than the raw error.
+ * 400 `model_not_supported` = Copilot fleet skew. A model that `/models`
+ *        advertises can flap between 200 and 400 because only part of
+ *        Copilot's fleet has it in the integrator allowlist. After the
+ *        in-request retry exhausts, surface guidance rather than the raw error.
  * 401 = token invalid/expired → credential removal is safe, prompt re-login.
  * 403 = token valid but access denied (plan, model policy, org restriction) →
  *       do NOT reuse the auth-failed string (which triggers credential removal).

package/dist/types/utils/retry.d.ts

@@ -1,9 +1,8 @@
 import { isCopilotTransientModelError } from "../error/flags.js";
 export { isCopilotTransientModelError };
 /**
- * Wrap an initial Copilot request so transient model-availability 400s
- * (`model_not_supported`, `model_not_available_for_integrator`) are retried a
- * small number of times. No-op for non-Copilot providers.
+ * Wrap an initial Copilot request so transient `model_not_supported` 400s are
+ * retried a small number of times. No-op for non-Copilot providers.
  *
  * The callback **MUST** create a fresh in-flight request each invocation — a
  * once-consumed AsyncIterable cannot be re-iterated.

package/dist/types/utils/schema/index.d.ts

@@ -12,4 +12,3 @@ export * from "./strict-tool-validation.js";
 export * from "./types.js";
 export * from "./typescript.js";
 export * from "./wire.js";
-export * from "./zod-decontaminate.js";

package/dist/types/utils/schema/wire.d.ts

@@ -1,76 +1,36 @@
 /**
  * Compute the wire (JSON Schema) representation of a tool's parameters.
  *
- * Tools may author parameters in three shapes:
- *   1. Zod (canonical) — converted to JSON Schema on demand.
- *   2. ArkType — converted to JSON Schema via its native `toJsonSchema`.
- *   3. TypeBox / plain JSON Schema (legacy + extension compat) — upgraded to
- *      draft 2020-12 without converting.
- *
- * All three are normalized at the boundary so providers and validators see the same
- * JSON Schema dialect.
+ * Tools may author parameters as ArkType schemas or legacy TypeBox / plain JSON
+ * Schema documents. Both are normalized at the boundary so providers and
+ * validators see the same JSON Schema dialect.
  */
 import type { Type } from "@oh-my-pi/omptype";
-import { type ZodType } from "zod/v4";
 import type { Tool } from "../../types.js";
-/**
- * True when `value` is a live Zod schema instance.
- *
- * The check is stricter than "has a `_zod` property" because a JSON
- * round-trip preserves the `_zod` key as a plain object and would otherwise
- * fool the predicate — see issue #1101, where MCP servers ship
- * `JSON.stringify(zodSchemaInstance)` as a tool's `inputSchema` and the
- * resulting plain object then explodes `z.toJSONSchema` because the prototype
- * (and every Zod parsing method) is gone.
- *
- * Live Zod instances always carry a `.parse` function on the prototype;
- * impostors do not.
- */
-export declare function isZodSchema(value: unknown): value is ZodType;
 /**
  * True when `value` is a live ArkType schema instance.
  *
  * ArkType schemas are callable functions carrying `toJsonSchema`/`assert`
- * methods. Zod v4 instances are non-callable objects (keyed off `_zod`), and
- * raw JSON Schema is a plain object — the three are disjoint. We deliberately
- * avoid the Standard Schema `~standard` marker because Zod v4 implements it too.
+ * methods, while raw JSON Schema is a plain object.
  */
 export declare function isArkSchema(value: unknown): value is Type;
 /**
- * Normalize `{}` (empty JSON Schema = `z.unknown()` / unconstrained value) to
- * boolean `true` in every schema-valued position. JSON Schema draft 2020-12
- * §4.3.1: `{}` and `true` are semantically equivalent ("any JSON value").
- * Grammar-constrained samplers (llama.cpp, etc.) treat the object form as
- * "generate an empty object" rather than "any JSON value", causing open-typed
- * fields like `extra.title` (from `z.record(z.string(), z.unknown())`) to
- * always emit `{}` instead of the intended string/number/etc. (issue #1179).
+ * Normalize `{}` (an unconstrained schema) to boolean `true` in every
+ * schema-valued position. JSON Schema draft 2020-12 §4.3.1 defines them as
+ * semantically equivalent. Grammar-constrained samplers often treat the object
+ * form as "generate an empty object" rather than "any JSON value".
  *
- * Mutates in place. Provider-agnostic — applied to every tool wire schema so
- * Anthropic, Google, OpenAI, Ollama, Bedrock, and Cursor all see the
- * normalized form, regardless of whether the source was Zod or TypeBox.
+ * Mutates in place and applies to every tool wire schema.
  */
 export declare function normalizeEmptySchemas(node: unknown): void;
-/** Convert a Zod schema into the JSON Schema shape providers consume. */
-export declare function zodToWireSchema(schema: ZodType): Record<string, unknown>;
 /**
  * Convert an ArkType schema into the JSON Schema shape providers consume.
- *
- * Mirrors {@link zodToWireSchema}: emit draft-2020-12, drop the `$schema`
- * metadata, run the JSON-schema post-process (NOT the Zod-only cleanup), then
- * close declared objects so the wire is `additionalProperties: false` like Zod.
- *
- * The `fallback` degrades any un-emittable node (a `.narrow()` predicate or a
- * morph) to its underlying base schema instead of throwing — matching Zod,
- * whose `.refine()`/`.transform()` likewise never appear in the wire schema.
  */
 export declare function arkToWireSchema(schema: Type): Record<string, unknown>;
 /**
  * Resolve a tool's parameters to a JSON Schema object suitable for sending
- * over the wire. Zod schemas are converted (and cached); legacy TypeBox / raw
- * JSON Schema parameters are upgraded to draft 2020-12 (and cached).
- *
- * Zod schemas also receive Zod-artifact cleanup; both branches normalize
- * schema-valued positions and nullable scalar unions.
+ * over the wire. ArkType schemas are converted and cached; legacy TypeBox /
+ * raw JSON Schema parameters are upgraded to draft 2020-12 and cached.
  */
 export declare function toolWireSchema(tool: Tool): Record<string, unknown>;
 /**

package/dist/types/utils/validation.d.ts

@@ -1,22 +1,10 @@
 /**
  * Tool-call argument validation pipeline.
  *
- * Tools may declare their parameters as either Zod schemas (canonical) or
- * plain JSON Schema (legacy / extensions). This module is the single
- * entrypoint the agent calls before dispatching a tool — it:
- *
- *   1. Builds (or fetches from cache) a `ValidationContext` for the tool —
- *      the Zod schema if available plus the equivalent wire JSON Schema, or
- *      just the JSON Schema for non-Zod tools.
- *   2. Normalizes LLM quirks (null / "null" → omit-or-default substitution)
- *      against the JSON Schema before validation.
- *   3. Validates with the Zod or JSON-Schema validator.
- *   4. On failure, walks the resulting issues and coerces common LLM type
- *      drift (JSON-stringified values, boolean/number/string scalar drift),
- *      drops unrecognized keys, and retries up to `MAX_COERCION_PASSES` times.
- *   5. Throws a formatted error if reconciliation fails; otherwise returns
- *      the parsed arguments with original unknown root fields preserved (so
- *      hallucinated top-level keys still surface to the caller).
+ * Tools may declare ArkType schemas or plain JSON Schema. This module builds a
+ * cached validation context, normalizes common LLM quirks against the wire
+ * schema, validates, performs conservative schema-directed coercions, and
+ * returns parsed arguments while preserving unknown root fields.
  *
  * The goal is to be conservative: every coercion is a structural rewrite that
  * keeps the schema in charge of acceptance — we never invent values, only
@@ -32,10 +20,8 @@ import type { Tool, ToolCall } from "../types.js";
  */
 export declare function validateToolCall(tools: Tool[], toolCall: ToolCall): ToolCall["arguments"];
 /**
- * Validates tool call arguments against the tool's schema (Zod or plain JSON
- * Schema). Applies LLM-quirk coercions (numeric strings, JSON-string
- * containers, null/invalid-empty-string-for-optional, null-for-default) before
- * declaring failure.
+ * Validates tool call arguments against an ArkType or plain JSON Schema schema.
+ * Applies conservative LLM-quirk normalization before declaring failure.
  *
  * @throws Error with a formatted message when validation cannot be reconciled.
  */

package/dist/types/utils.d.ts

@@ -17,6 +17,15 @@ interface OpenAIResponsesReplaySanitizeOptions {
     supportsImageDetailOriginal?: boolean;
     supportsComputerUse?: boolean;
 }
+/**
+ * Removes response-only lifecycle status from item types that reject it when replayed as input.
+ *
+ * Returns the original array when no item needs sanitization.
+ */
+export declare function stripOpenAIResponsesOutputOnlyStatusesForReplay<TItem extends {
+    type?: unknown;
+    status?: unknown;
+}>(items: TItem[]): TItem[];
 export declare function sanitizeOpenAIResponsesHistoryItemsForReplay(items: Array<Record<string, unknown>>, options?: OpenAIResponsesReplaySanitizeOptions): ResponseInput;
 /** Strip reasoning IDs whose only linked native output is a computer call that will be demoted. */
 export declare function stripOpenAIResponsesComputerLinkedReasoningIdsForReplay(items: ResponseInput): ResponseInput;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "17.2.9",
+	"version": "17.2.11",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -38,11 +38,10 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.1",
-		"@oh-my-pi/omptype": "17.2.9",
-		"@oh-my-pi/pi-catalog": "17.2.9",
-		"@oh-my-pi/pi-utils": "17.2.9",
-		"@oh-my-pi/pi-wire": "17.2.9",
-		"zod": "^4"
+		"@oh-my-pi/omptype": "17.2.11",
+		"@oh-my-pi/pi-catalog": "17.2.11",
+		"@oh-my-pi/pi-utils": "17.2.11",
+		"@oh-my-pi/pi-wire": "17.2.11"
 	},
 	"devDependencies": {
 		"@bufbuild/protoc-gen-es": "^2.12.1",

package/src/auth/sqlite-credential-store.ts

@@ -1374,11 +1374,13 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 		try {
 			let hasActiveApiKey = false;
 			const activeIdentityKeys = new Set<string>();
+			const activeOAuthCredentials: AuthCredential[] = [];
 			for (const row of activeRows) {
 				if (row.credential.type === "api_key") {
 					hasActiveApiKey = true;
 					continue;
 				}
+				activeOAuthCredentials.push(row.credential);
 				const identityKey = resolveCredentialIdentityKey(provider, row.credential);
 				if (identityKey) activeIdentityKeys.add(identityKey);
 			}
@@ -1393,7 +1395,22 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 				const identityKey = resolveRowCredentialIdentityKey(provider, row);
 				if (identityKey && activeIdentityKeys.has(identityKey)) {
 					this.#hardDeleteStmt.run(row.id);
+					continue;
 				}
+				// Exact key equality misses a tombstone whose key predates a format
+				// the active row now uses (pre-org `<b>` vs `<b>|org:<o>`). An active
+				// credential that WOULD have replaced this row had it still been
+				// active supersedes its tombstone too, so mirror the replacement
+				// matcher rather than restating a weaker rule. The one-way upgrade
+				// and shared-workspace guards in matchesReplacementCredential carry
+				// over, so this never over-deletes another member's or subscription's
+				// row.
+				const disabledCredential = deserializeCredential(row);
+				if (disabledCredential === null) continue;
+				const superseded = activeOAuthCredentials.some(active =>
+					matchesReplacementCredential(provider, disabledCredential, identityKey, active),
+				);
+				if (superseded) this.#hardDeleteStmt.run(row.id);
 			}
 		} catch {
 			// Best-effort cleanup; don't let it break the main operation

package/src/error/flags.ts

@@ -113,17 +113,15 @@ const STALE_RESPONSE_ITEM_DETAIL_PATTERN = /not[ _]?found|invalid|expired|stale|
 export const LLAMA_CPP_TOOL_CALL_PARSE_PATTERN =
 	/failed to parse tool call arguments as json|\[json\.exception\.parse_error\.101\]/i;
 
-// Copilot fleet skew: HTTP 400 rejecting a model that `/models` advertised on
-// the very same host. Two codes appear in the wild — `model_not_supported`
-// (per-OAuth-client rollout gap) and `model_not_available_for_integrator`
-// (replicas whose integrator allowlist predates the model). Both flap
-// request-to-request, so a retry usually lands on a backend that has the model.
+// Copilot fleet skew: HTTP 400 `model_not_supported` can reject a model that
+// `/models` advertised on the same host when the request lands on a stale
+// replica. `model_not_available_for_integrator` is deliberately excluded:
+// GitHub also uses it for stable per-integrator entitlement denials and includes
+// that integrator's actionable `Available models` list in the response.
 const COPILOT_TRANSIENT_MODEL_CODES: Record<string, true> = {
 	model_not_supported: true,
-	model_not_available_for_integrator: true,
 };
-const COPILOT_MODEL_UNAVAILABLE_PATTERN =
-	/model_not_supported|model_not_available_for_integrator|not available for integrator/i;
+const COPILOT_TRANSIENT_MODEL_PATTERN = /model_not_supported/i;
 // Anthropic strict-tool grammar too large / schema too complex (400 invalid_request_error).
 // Feature-gated deployments (Azure Foundry, Baseten, …) reject `strict: true`
 // tools outright when the hosted model lacks structured outputs, e.g.
@@ -377,8 +375,8 @@ function classifyText(errorMessage: string | undefined, errorStatus: number | un
 			kinds |= Flag.StaleResponsesItem;
 		}
 
-		// Copilot fleet-skew model rejection is transient.
-		if (statusClean === 400 && COPILOT_MODEL_UNAVAILABLE_PATTERN.test(cleanMessage)) kinds |= Flag.Transient;
+		// Copilot's `model_not_supported` fleet-skew rejection is transient.
+		if (statusClean === 400 && COPILOT_TRANSIENT_MODEL_PATTERN.test(cleanMessage)) kinds |= Flag.Transient;
 		if (matchesStrictToolsRejection(cleanMessage, statusClean)) kinds |= Flag.Grammar;
 		if (matchesFastModeUnsupported(cleanMessage, statusClean)) kinds |= Flag.FastModeUnsupported;
 	}
@@ -513,10 +511,10 @@ function providerErrorCode(error: object): string | undefined {
 }
 
 /**
- * GitHub Copilot 400 rejecting a model its own `/models` catalog advertises —
- * transient fleet skew, not a malformed request. Reads the structural `code`
- * through the SDK/body envelopes, then falls back to the stringified body both
- * SDK families put in `message` (shapes drift; the wire text does not).
+ * GitHub Copilot 400 `model_not_supported` response for a model advertised by
+ * `/models` — transient fleet skew, not a malformed request. Reads the
+ * structural `code` through the SDK/body envelopes, then falls back to the
+ * stringified body both SDK families put in `message`.
  */
 export function isCopilotTransientModelError(error: unknown): boolean {
 	if (!error || typeof error !== "object" || status(error) !== 400) return false;
@@ -525,7 +523,7 @@ export function isCopilotTransientModelError(error: unknown): boolean {
 	// prototype key (`__proto__`, `toString`, …) would otherwise read truthy.
 	if (code !== undefined && Object.hasOwn(COPILOT_TRANSIENT_MODEL_CODES, code)) return true;
 	const message: unknown = "message" in error ? error.message : undefined;
-	return typeof message === "string" && COPILOT_MODEL_UNAVAILABLE_PATTERN.test(message);
+	return typeof message === "string" && COPILOT_TRANSIENT_MODEL_PATTERN.test(message);
 }
 
 export function classifyMessage(message: {

package/src/error/rate-limit.ts

@@ -22,6 +22,13 @@ const ACCOUNT_RATE_LIMIT_PATTERN =
 	/\baccount(?:'s)?\b[^\n]{0,80}\brate.?limit\b|\brate.?limit\b[^\n]{0,80}\baccount\b/i;
 const INSUFFICIENT_BALANCE_PATTERN = /insufficient.?balance/i;
 const SPEND_LIMIT_PATTERN = /spend.?limit/i;
+const SUBSCRIPTION_CAP_PATTERN =
+	/\b(?:subscription|plan|membership)\b[^\n]{0,80}\b(?:rate.?limits?|quota|cap)\b|\b(?:rate.?limits?|quota|cap)\b[^\n]{0,80}\b(?:subscription|plan|membership)\b/i;
+const TRANSIENT_INTERVAL_RATE_LIMIT_PATTERN = /\bper\s+(?:second|minute)\b/i;
+
+function matchesSubscriptionCapText(errorMessage: string): boolean {
+	return SUBSCRIPTION_CAP_PATTERN.test(errorMessage) && !TRANSIENT_INTERVAL_RATE_LIMIT_PATTERN.test(errorMessage);
+}
 const OPENROUTER_DAILY_FREE_LIMIT_PATTERN = /\bfree[-_ ]models[-_ ]per[-_ ]day\b/i;
 // gRPC/Connect end-streams carry the status as its name (`resource_exhausted`),
 // while HTTP bodies use the phrase ("resource exhausted"). Strip either form
@@ -40,6 +47,25 @@ const ACCOUNT_SCOPED_403_PATTERN =
 	// "Your limit will reset in …"); the overall/account qualifiers arm above
 	// already covers the rest.
 	/\b(?:overall|account|organization|team|workspace)\b[^\n]{0,40}\b(?:message |request )?rate.?limit\b|\byour\b[^\n]{0,30}\b(?:limit )?will reset\b/i;
+// Simplified Chinese account-quota exhaustion phrasing. Zhipu Coding Plan
+// returns e.g. "429 已达到 5 小时的使用上限。您的限额将在 2026-08-06 20:06:00 重置。"
+// (type=1308) when the 5h window is spent; other CN providers use 额度已用完 /
+// 配额已耗尽 / 余额不足. These are persistent account-local caps that must
+// rotate to a sibling credential, not transient rate limits, so they are
+// matched before the RATE_LIMIT_EXCEEDED branch. The 上限 arm is anchored on
+// the 使用 token: a rate/concurrency cap phrased as 每分钟请求数已达上限 /
+// 并发请求数已达上限 / 速率达到上限 (no 使用) must NOT match, or it would burn a
+// healthy sibling credential as a false quota. "速率限制" is absent for the
+// same reason.
+const CN_QUOTA_EXHAUSTED_PATTERN = /使用.{0,30}?上限|(?:额度|配额)已?(?:用|耗)(?:完|尽)|限额.{0,30}重置|余额不足/;
+// Simplified Chinese rate/concurrency caps can contain both 使用 and 上限, but
+// remain transient rather than account quota exhaustion.
+const CN_TRANSIENT_CAP_PATTERN =
+	/速率.{0,30}上限|频率.{0,30}上限|每分钟.{0,30}上限|并发.{0,30}上限|使用.{0,30}(?:速率|频率|每分钟|并发).{0,30}上限/;
+// Common Simplified Chinese throttle phrasing. Consulted by
+// isOpaqueStatusBody so CN transients stay in the provider backoff lane instead
+// of rotating through the opaque-429 fallback.
+const CN_THROTTLE_PATTERN = /速率(?:限制|过快)|频率(?:过高|过快)|过于频繁|稍后[重再]试/;
 
 /**
  * Classify a rate-limit error message into a reason category.
@@ -64,6 +90,13 @@ export function parseRateLimitReason(errorMessage: string): RateLimitReason {
 		return "QUOTA_EXHAUSTED";
 	}
 
+	// Simplified Chinese quota-exhaustion phrasing (Zhipu Coding Plan and other
+	// CN providers). Must precede the MODEL_CAPACITY / RATE_LIMIT branches so an
+	// account-local cap rotates instead of backing off as a transient.
+	if (CN_QUOTA_EXHAUSTED_PATTERN.test(errorMessage) && !CN_TRANSIENT_CAP_PATTERN.test(errorMessage)) {
+		return "QUOTA_EXHAUSTED";
+	}
+
 	if (CONCURRENT_LIMIT_PATTERN.test(errorMessage)) {
 		return "CONCURRENT_LIMIT";
 	}
@@ -80,6 +113,10 @@ export function parseRateLimitReason(errorMessage: string): RateLimitReason {
 		return "QUOTA_EXHAUSTED";
 	}
 
+	if (matchesSubscriptionCapText(errorMessage)) {
+		return "QUOTA_EXHAUSTED";
+	}
+
 	if (OPENROUTER_DAILY_FREE_LIMIT_PATTERN.test(errorMessage)) {
 		return "QUOTA_EXHAUSTED";
 	}
@@ -212,7 +249,20 @@ export function isOpaqueStatusBody(message: string): boolean {
 	const cleaned = message
 		.replace(/\b(?:429|402)\b/g, "")
 		.replace(/\b(?:http|https|status|error|code|response|message)\b/gi, "");
-	return !/[a-z\d]{3,}/i.test(cleaned);
+	// A body is informative when the text classifier can act on it. Any Latin
+	// word or Simplified Chinese phrasing the classifier recognizes (quota
+	// exhaustion or a throttle) defers to parseRateLimitReason; a body that
+	// is only status digits / HTTP framing is opaque and rotates conservatively.
+	// A Han-only body the classifier cannot interpret (e.g. Japanese Kanji
+	// quota text, since Japanese is out of scope) must stay opaque so the
+	// opaque-429 fallback still rotates. This keeps the exception scoped to
+	// text we actually classify, rather than to any Han ideograph.
+	return (
+		!/[a-z\d]{3,}/i.test(cleaned) &&
+		!CN_QUOTA_EXHAUSTED_PATTERN.test(cleaned) &&
+		!CN_TRANSIENT_CAP_PATTERN.test(cleaned) &&
+		!CN_THROTTLE_PATTERN.test(cleaned)
+	);
 }
 
 /**
@@ -224,8 +274,10 @@ export function isOpaqueStatusBody(message: string): boolean {
 export function matchesUsageLimitText(errorMessage: string): boolean {
 	return (
 		USAGE_LIMIT_PATTERN.test(errorMessage) ||
+		(CN_QUOTA_EXHAUSTED_PATTERN.test(errorMessage) && !CN_TRANSIENT_CAP_PATTERN.test(errorMessage)) ||
 		SPEND_LIMIT_PATTERN.test(errorMessage) ||
 		ACCOUNT_RATE_LIMIT_PATTERN.test(errorMessage) ||
+		matchesSubscriptionCapText(errorMessage) ||
 		OPENROUTER_DAILY_FREE_LIMIT_PATTERN.test(errorMessage)
 	);
 }

package/src/index.ts

@@ -1,5 +1,4 @@
 export { type Type, type } from "@oh-my-pi/omptype";
-export { type ZodType, z } from "zod/v4";
 export * from "./api-registry";
 export type * from "./auth-broker";
 export type { AuthGatewayBootOptions, ModelResolver } from "./auth-gateway/server";