@oh-my-pi/pi-ai 17.4.1 → 17.4.2

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [17.4.2] - 2026-08-21
+
+### Added
+
+- Image content blocks accept an optional `url` mirror: providers whose APIs fetch remote images (Anthropic url sources, OpenAI/xAI Responses and Chat Completions `image_url`, Google `fileData`) send the URL instead of the inline base64 payload.
+
+### Fixed
+
+- Fixed Cursor thinking-effort selection being cosmetic: collapsed effort-routed families (GPT-5.6 Luna/Sol/Terra, Grok 4.5/4.6) now send the effort-routed wire model id instead of always pinning the `-none` off tier ([#9246](https://github.com/can1357/oh-my-pi/issues/9246)).
+- Fixed OAuth preflight refresh stranding a peer-rotated credential: when a concurrent process rotated a rotating-refresh-token grant (e.g. Anthropic) during preflight, the resolve pass skipped the freshly reloaded row and failed the request with no credentials for single-account setups ([#9194](https://github.com/can1357/oh-my-pi/issues/9194)).
+- Fixed Cursor reasoning-sibling models (e.g. `gpt-5.4-mini-low`, `gpt-5.6-sol-xhigh`) failing with `resource_exhausted` (errorId 528384): the per-effort GPT slug is now split into its base model id plus a `{ id: "reasoning", value: <effort> }` request parameter, matching the official `cursor-agent` wire shape, instead of sending the sibling slug as the wire model id with no parameters ([#9164](https://github.com/can1357/oh-my-pi/issues/9164)).
+
 ## [17.4.1] - 2026-08-21
 
 ### Added

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

@@ -12,6 +12,8 @@ export interface CursorOptions extends StreamOptions {
     conversationId?: string;
     execHandlers?: CursorExecHandlers;
     onToolResult?: CursorToolResultHandler;
+    /** Wire model id selected after thinking-effort routing (`resolveWireModelId`). */
+    wireModelId?: string;
 }
 /**
  * Maps an opaque HTTP/2 negotiation failure into an actionable error.

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

@@ -42,9 +42,15 @@ export interface Part {
     thought?: boolean;
     thoughtSignature?: string;
     inlineData?: InlineDataPart;
+    fileData?: FileDataPart;
     functionCall?: FunctionCallPart;
     functionResponse?: FunctionResponsePart;
 }
+/** Remote media reference; the backend fetches `fileUri` server-side. */
+export interface FileDataPart {
+    fileUri: string;
+    mimeType?: string;
+}
 /** Conversation turn. Roles: `"user"`, `"model"`, optionally absent for system instructions. */
 export interface Content {
     role?: string;

package/dist/types/types.d.ts

@@ -542,6 +542,13 @@ export interface AnthropicServerToolContent {
         [key: string]: unknown;
     };
 }
+/** Provider-native uploaded file reference for image reuse without retransmitting bytes. */
+export interface ProviderFileReference {
+    provider: "openai" | "anthropic" | "google";
+    id?: string;
+    uri?: string;
+    expiresAt?: number;
+}
 export interface ImageContent {
     type: "image";
     data: string;
@@ -552,6 +559,18 @@ export interface ImageContent {
      * default `auto` downscale). Providers without a detail knob ignore it.
      */
     detail?: "auto" | "low" | "high" | "original";
+    /** Provider-native file reference preferred only by its matching provider. */
+    providerFile?: ProviderFileReference;
+    /**
+     * Optional https mirror of `data`, served by a caller-run blob server.
+     * Providers whose API fetches remote images send this URL instead of the
+     * base64 payload; every other provider ignores it. `data` remains the
+     * source of truth — the URL must serve exactly those bytes, and callers
+     * are responsible for keeping it stable across turns (prefix caches hash
+     * the URL string, and Anthropic silently forgets images when a resent
+     * turn differs byte-wise).
+     */
+    url?: string;
 }
 export type ComputerAction = {
     type: "click";

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "17.4.1",
+	"version": "17.4.2",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -37,10 +37,10 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "17.4.1",
-		"@oh-my-pi/pi-catalog": "17.4.1",
-		"@oh-my-pi/pi-utils": "17.4.1",
-		"@oh-my-pi/pi-wire": "17.4.1"
+		"@oh-my-pi/omptype": "17.4.2",
+		"@oh-my-pi/pi-catalog": "17.4.2",
+		"@oh-my-pi/pi-utils": "17.4.2",
+		"@oh-my-pi/pi-wire": "17.4.2"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/auth-storage.ts

@@ -4965,13 +4965,29 @@ export class AuthStorage {
 						// so if this branch only blocked the row (like the transient case), the
 						// definitive failure would never reach `#tryOAuthCredential`'s own
 						// disable logic and the row would be retried forever instead of torn down.
-						await this.#disableDefinitiveOAuthFailure(
+						const outcome = await this.#disableDefinitiveOAuthFailure(
 							provider,
 							credentialId,
 							candidate.selection.credential,
 							candidate.selection.index,
 							errorMsg,
 						);
+						if (
+							outcome !== "disabled" &&
+							credentialId !== undefined &&
+							this.#syncOAuthSelectionFromStore(provider, candidate.selection, credentialId)
+						) {
+							// A peer rotated this row (or won the disable CAS) between our
+							// snapshot and the refresh; the helper reloaded storage and the row
+							// still exists, so it now holds a valid, freshly rotated credential.
+							// Re-sync the candidate onto it and leave it eligible so the final
+							// pass retries with the live token instead of stranding it (mirrors
+							// #tryOAuthCredential's peer-rotated re-resolve). If the peer instead
+							// deleted/disabled the row, the re-sync fails and we fall through to
+							// preflightFailures — leaving a stale index could rebind the candidate
+							// to a sibling account with the wrong prefetched usage/plan.
+							return;
+						}
 					} else if (credentialId !== undefined) {
 						const latestIndex = this.#getStoredCredentials(provider).findIndex(
 							entry => entry.id === credentialId,

package/src/providers/anthropic.ts

@@ -914,7 +914,15 @@ async function resizeAnthropicManyImageContent(
 	let changed = false;
 	const next = await Promise.all(
 		content.map(async block => {
-			if (block.type !== "image") return block;
+			// Remotely referenced blocks never put base64 on the wire, so their size
+			// cannot violate the many-image request budget — and resizing would
+			// desync fallback bytes from the advertised remote image.
+			if (
+				block.type !== "image" ||
+				block.url ||
+				(block.providerFile?.provider === "anthropic" && block.providerFile.id)
+			)
+				return block;
 			let resized = anthropicManyImageResizeCache.get(block);
 			if (resized === undefined) {
 				resized = await limit(() => resizeAnthropicManyImageBlock(block));
@@ -971,19 +979,14 @@ async function prepareAnthropicManyImageContext(context: Context, supportsImages
 	return { ...context, messages };
 }
 
+type AnthropicImageSource =
+	| { type: "base64"; media_type: AnthropicImageMediaType; data: string }
+	| { type: "url"; url: string }
+	| { type: "file"; file_id: string };
+
 type AnthropicToolResultContent =
 	| string
-	| Array<
-			| { type: "text"; text: string }
-			| {
-					type: "image";
-					source: {
-						type: "base64";
-						media_type: AnthropicImageMediaType;
-						data: string;
-					};
-			  }
-	  >;
+	| Array<{ type: "text"; text: string } | { type: "image"; source: AnthropicImageSource }>;
 
 /**
  * Convert content blocks to Anthropic API format
@@ -992,17 +995,7 @@ function convertContentBlocks(
 	content: (TextContent | ImageContent)[],
 	supportsImages = true,
 ): AnthropicToolResultContent {
-	const blocks: Array<
-		| { type: "text"; text: string }
-		| {
-				type: "image";
-				source: {
-					type: "base64";
-					media_type: AnthropicImageMediaType;
-					data: string;
-				};
-		  }
-	> = [];
+	const blocks: Array<{ type: "text"; text: string } | { type: "image"; source: AnthropicImageSource }> = [];
 	let sawText = false;
 	let sawImage = false;
 
@@ -1020,21 +1013,22 @@ function convertContentBlocks(
 			continue;
 		}
 
-		const mediaType = normalizeAnthropicImageMediaType(block.mimeType);
-		if (!mediaType) {
-			blocks.push({ type: "text", text: `[unsupported image: ${block.mimeType}]` });
-			continue;
+		let source: AnthropicImageSource;
+		if (block.providerFile?.provider === "anthropic" && block.providerFile.id) {
+			source = { type: "file", file_id: block.providerFile.id };
+		} else if (block.url) {
+			source = { type: "url", url: block.url };
+		} else {
+			const mediaType = normalizeAnthropicImageMediaType(block.mimeType);
+			if (!mediaType) {
+				blocks.push({ type: "text", text: `[unsupported image: ${block.mimeType}]` });
+				continue;
+			}
+			source = { type: "base64", media_type: mediaType, data: block.data };
 		}
 
 		sawImage = true;
-		blocks.push({
-			type: "image",
-			source: {
-				type: "base64",
-				media_type: mediaType,
-				data: block.data,
-			},
-		});
+		blocks.push({ type: "image", source });
 	}
 
 	if (!supportsImages) {

package/src/providers/cursor.ts

@@ -1,7 +1,12 @@
 import { createHash } from "node:crypto";
 import * as fs from "node:fs/promises";
 import http2 from "node:http2";
-import type { ConversationStep, CursorRule, McpToolDefinition } from "@oh-my-pi/pi-catalog/discovery/cursor-proto";
+import type {
+	ConversationStep,
+	CursorRule,
+	McpToolDefinition,
+	RequestedModel_ModelParameterbytes,
+} from "@oh-my-pi/pi-catalog/discovery/cursor-proto";
 import {
 	AgentClientMessageSchema,
 	AgentConversationTurnStructureSchema,
@@ -106,6 +111,7 @@ import {
 	RequestContextResultSchema,
 	RequestContextSchema,
 	RequestContextSuccessSchema,
+	RequestedModel_ModelParameterbytesSchema,
 	RequestedModelSchema,
 	ResumeActionSchema,
 	SelectedContextSchema,
@@ -151,7 +157,8 @@ import {
 	toBinary,
 	toJson,
 } from "@oh-my-pi/pi-catalog/discovery/protobuf";
-import { isKimiK3ModelId } from "@oh-my-pi/pi-catalog/identity";
+import { THINKING_EFFORTS } from "@oh-my-pi/pi-catalog/effort";
+import { isKimiK3ModelId, parseOpenAIModel } from "@oh-my-pi/pi-catalog/identity";
 import { calculateCost } from "@oh-my-pi/pi-catalog/models";
 import {
 	$env,
@@ -330,6 +337,8 @@ export interface CursorOptions extends StreamOptions {
 	conversationId?: string;
 	execHandlers?: CursorExecHandlers;
 	onToolResult?: CursorToolResultHandler;
+	/** Wire model id selected after thinking-effort routing (`resolveWireModelId`). */
+	wireModelId?: string;
 }
 
 const CONNECT_END_STREAM_FLAG = 0b00000010;
@@ -5011,6 +5020,45 @@ function extractImages(content: (TextContent | ImageContent)[]) {
 		);
 }
 
+/**
+ * Resolve the Cursor Run wire model id and its parameter list.
+ *
+ * Cursor's `GetUsableModels` lists reasoning models as per-effort sibling
+ * slugs (`gpt-5.4-mini-low`, `gpt-5.6-sol-high`), and OMP copies those ids 1:1.
+ * The Run endpoint rejects a sibling slug as the wire `model_id` with
+ * `resource_exhausted` (errorId 528384); the official `cursor-agent` splits the
+ * slug into its base model id plus a `reasoning` effort parameter. Mirror that
+ * for OpenAI-family ids: strip a trailing effort tier and emit
+ * `{ id: "reasoning", value: <effort> }`.
+ *
+ * Non-OpenAI ids pass through unchanged — Cursor-native ids (`composer-*`,
+ * `cursor-grok-*`, `default`) carry no effort suffix, and Claude/other siblings
+ * need additional parameters (`thinking`, `context`) whose per-model values are
+ * not exposed by the decoded `GetUsableModels` schema, so guessing them would
+ * re-trigger 528384.
+ */
+function resolveCursorWireModel(
+	model: Model<"cursor-agent">,
+	requestModelId?: string,
+): {
+	modelId: string;
+	parameters: RequestedModel_ModelParameterbytes[];
+} {
+	const wireModelId = requestModelId ?? model.requestModelId ?? model.id;
+	// Cursor's fast lane follows the effort token (`-high-fast`), while the
+	// standard lane ends at it (`-high`). Preserve the lane in the base id.
+	const match = /^(.*)-(minimal|low|medium|high|xhigh|max)(-fast)?$/.exec(wireModelId);
+	const base = match?.[1];
+	const effort = match?.[2];
+	if (base && effort && (THINKING_EFFORTS as readonly string[]).includes(effort) && parseOpenAIModel(base) !== null) {
+		return {
+			modelId: `${base}${match[3] ?? ""}`,
+			parameters: [create(RequestedModel_ModelParameterbytesSchema, { id: "reasoning", value: effort })],
+		};
+	}
+	return { modelId: wireModelId, parameters: [] };
+}
+
 export async function buildGrpcRequest(
 	model: Model<"cursor-agent">,
 	context: Context,
@@ -5117,7 +5165,7 @@ export async function buildGrpcRequest(
 		turns,
 	});
 
-	const wireModelId = model.requestModelId ?? model.id;
+	const { modelId: wireModelId, parameters: wireParameters } = resolveCursorWireModel(model, options?.wireModelId);
 	const cursorMaxMode = model.cursorMaxMode === true;
 	const modelDetails = create(ModelDetailsSchema, {
 		modelId: wireModelId,
@@ -5128,6 +5176,7 @@ export async function buildGrpcRequest(
 	const requestedModel = create(RequestedModelSchema, {
 		modelId: wireModelId,
 		maxMode: cursorMaxMode,
+		parameters: wireParameters,
 	});
 
 	let runRequest = create(AgentRunRequestSchema, {

package/src/providers/google-shared.ts

@@ -53,6 +53,15 @@ export { normalizeSchemaForGoogle };
 
 type GoogleApiType = "google-generative-ai" | "google-gemini-cli" | "google-vertex";
 
+function convertGoogleImagePart(image: ImageContent): Part {
+	if (image.providerFile?.provider === "google" && image.providerFile.uri) {
+		return { fileData: { fileUri: image.providerFile.uri, mimeType: image.mimeType } };
+	}
+	return image.url
+		? { fileData: { fileUri: image.url, mimeType: image.mimeType } }
+		: { inlineData: { mimeType: image.mimeType, data: image.data } };
+}
+
 /**
  * Thinking level for Gemini 3 models. Mirrors Google's `ThinkingLevel` enum values.
  * Defined here (not in any specific provider) so all Google providers can reference it
@@ -211,12 +220,7 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 						if (text.trim().length === 0) continue;
 						parts.push({ text });
 					} else if (supportsImages) {
-						parts.push({
-							inlineData: {
-								mimeType: item.mimeType,
-								data: item.data,
-							},
-						});
+						parts.push(convertGoogleImagePart(item));
 					} else {
 						omittedImages = true;
 					}
@@ -315,12 +319,7 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 						? "(see attached image)"
 						: "";
 
-			const imageParts: Part[] = imageContent.map(imageBlock => ({
-				inlineData: {
-					mimeType: imageBlock.mimeType,
-					data: imageBlock.data,
-				},
-			}));
+			const imageParts = imageContent.map(convertGoogleImagePart);
 
 			const includeId = supportsFunctionPartId(model);
 			const emittedName = emittedToolCallNames.get(msg.toolCallId);

package/src/providers/google-types.ts

@@ -65,10 +65,17 @@ export interface Part {
 	thought?: boolean;
 	thoughtSignature?: string;
 	inlineData?: InlineDataPart;
+	fileData?: FileDataPart;
 	functionCall?: FunctionCallPart;
 	functionResponse?: FunctionResponsePart;
 }
 
+/** Remote media reference; the backend fetches `fileUri` server-side. */
+export interface FileDataPart {
+	fileUri: string;
+	mimeType?: string;
+}
+
 /** Conversation turn. Roles: `"user"`, `"model"`, optionally absent for system instructions. */
 export interface Content {
 	role?: string;

package/src/providers/openai-codex-responses.ts

@@ -4503,7 +4503,9 @@ function convertMessages(model: Model<"openai-codex-responses">, context: Contex
 
 function normalizeInputMessageContent(
 	model: Model<"openai-codex-responses">,
-	content: string | Array<{ type: "text"; text: string } | { type: "image"; mimeType: string; data: string }>,
+	content:
+		| string
+		| Array<{ type: "text"; text: string } | { type: "image"; mimeType: string; data: string; url?: string }>,
 ): ResponseInputContent[] {
 	// gpt-5.x codex rejects reserved Harmony control-token spellings in input
 	// data; escape the transport copy of untrusted user text so ordinary docs or

package/src/providers/openai-completions.ts

@@ -1950,7 +1950,7 @@ export function convertMessages(
 						content.push({
 							type: "image_url",
 							image_url: {
-								url: `data:${item.mimeType};base64,${item.data}`,
+								url: item.url ?? `data:${item.mimeType};base64,${item.data}`,
 								// Chat Completions has no "original"; omit it (provider default).
 								...(item.detail && item.detail !== "original" ? { detail: item.detail } : {}),
 							},
@@ -2257,7 +2257,7 @@ export function convertMessages(
 							imageBlocks.push({
 								type: "image_url",
 								image_url: {
-									url: `data:${block.mimeType};base64,${block.data}`,
+									url: block.url ?? `data:${block.mimeType};base64,${block.data}`,
 								},
 							});
 						}

package/src/providers/openai-shared.ts

@@ -1630,6 +1630,18 @@ function clampResponsesImageDetail(
 	return resolved === "original" && !supportsImageDetailOriginal ? "auto" : resolved;
 }
 
+function convertResponsesInputImage(image: ImageContent, supportsImageDetailOriginal: boolean): ResponseInputImage {
+	const detail = clampResponsesImageDetail(image.detail, supportsImageDetailOriginal);
+	if (image.providerFile?.provider === "openai" && image.providerFile.id) {
+		return { type: "input_image", detail, file_id: image.providerFile.id };
+	}
+	return {
+		type: "input_image",
+		detail,
+		image_url: image.url ?? `data:${image.mimeType};base64,${image.data}`,
+	};
+}
+
 export function convertResponsesInputContent(
 	content: string | Array<TextContent | ImageContent>,
 	supportsImages: boolean,
@@ -1659,11 +1671,7 @@ export function convertResponsesInputContent(
 		} satisfies ResponseInputText);
 	}
 	for (const item of imageBlocks) {
-		normalizedContent.push({
-			type: "input_image",
-			detail: clampResponsesImageDetail(item.detail, supportsImageDetailOriginal),
-			image_url: `data:${item.mimeType};base64,${item.data}`,
-		} satisfies ResponseInputImage);
+		normalizedContent.push(convertResponsesInputImage(item, supportsImageDetailOriginal));
 	}
 	if (omittedImages) {
 		normalizedContent.push({
@@ -2338,11 +2346,7 @@ export function appendResponsesToolResultMessages<TApi extends Api>(
 	];
 	for (const block of toolResult.content) {
 		if (block.type === "image") {
-			contentParts.push({
-				type: "input_image",
-				detail: clampResponsesImageDetail(block.detail, supportsImageDetailOriginal),
-				image_url: `data:${block.mimeType};base64,${block.data}`,
-			} satisfies ResponseInputImage);
+			contentParts.push(convertResponsesInputImage(block, supportsImageDetailOriginal));
 		}
 	}
 	const imageMessage = { role: "user", content: contentParts } satisfies ResponseInput[number];

package/src/stream.ts

@@ -2309,10 +2309,16 @@ function mapOptionsForApi<TApi extends Api>(
 		case "cursor-agent": {
 			const execHandlers = options?.cursorExecHandlers ?? options?.execHandlers;
 			const onToolResult = options?.cursorOnToolResult ?? execHandlers?.onToolResult;
+			const cursorModel = model as Model<"cursor-agent">;
+			const effort =
+				options?.reasoning && !options.disableReasoning && !options.forceReasoningOff && cursorModel.reasoning
+					? requireSupportedEffort(cursorModel, options.reasoning)
+					: undefined;
 			return castApi<"cursor-agent">({
 				...base,
 				execHandlers,
 				onToolResult,
+				wireModelId: resolveWireModelId(cursorModel, effort),
 			});
 		}
 

package/src/types.ts

@@ -740,6 +740,14 @@ export interface AnthropicServerToolContent {
 		  };
 }
 
+/** Provider-native uploaded file reference for image reuse without retransmitting bytes. */
+export interface ProviderFileReference {
+	provider: "openai" | "anthropic" | "google";
+	id?: string;
+	uri?: string;
+	expiresAt?: number;
+}
+
 export interface ImageContent {
 	type: "image";
 	data: string; // base64 encoded image data
@@ -750,6 +758,18 @@ export interface ImageContent {
 	 * default `auto` downscale). Providers without a detail knob ignore it.
 	 */
 	detail?: "auto" | "low" | "high" | "original";
+	/** Provider-native file reference preferred only by its matching provider. */
+	providerFile?: ProviderFileReference;
+	/**
+	 * Optional https mirror of `data`, served by a caller-run blob server.
+	 * Providers whose API fetches remote images send this URL instead of the
+	 * base64 payload; every other provider ignores it. `data` remains the
+	 * source of truth — the URL must serve exactly those bytes, and callers
+	 * are responsible for keeping it stable across turns (prefix caches hash
+	 * the URL string, and Anthropic silently forgets images when a resent
+	 * turn differs byte-wise).
+	 */
+	url?: string;
 }
 
 export type ComputerAction =