@oh-my-pi/pi-ai 16.2.1 → 16.2.3

package/src/providers/google-shared.ts

@@ -4,8 +4,9 @@
 
 import { scheduler } from "node:timers/promises";
 import { calculateCost } from "@oh-my-pi/pi-catalog/models";
-import { extractHttpStatusFromError, readSseJson } from "@oh-my-pi/pi-utils";
-import { ProviderHttpError } from "../errors";
+import { readSseJson } from "@oh-my-pi/pi-utils";
+import { renderDemotedThinking } from "../dialect/demotion";
+import * as AIError from "../error";
 import type {
 	Api,
 	AssistantMessage,
@@ -22,7 +23,7 @@ import type {
 } from "../types";
 import { normalizeSystemPrompts } from "../utils";
 import { AssistantMessageEventStream } from "../utils/event-stream";
-import { finalizeErrorMessage, type RawHttpRequestDump } from "../utils/http-inspector";
+import type { RawHttpRequestDump } from "../utils/http-inspector";
 import { normalizeSchemaForCCA, normalizeSchemaForGoogle, toolWireSchema } from "../utils/schema";
 import type {
 	Content,
@@ -47,11 +48,6 @@ export type {
 } from "./google-types";
 export { normalizeSchemaForGoogle };
 
-/** Non-2xx response (or in-stream error chunk) from the Google Generative Language / Vertex API. */
-export class GoogleApiError extends ProviderHttpError {
-	override readonly name = "GoogleApiError";
-}
-
 type GoogleApiType = "google-generative-ai" | "google-gemini-cli" | "google-vertex";
 
 /**
@@ -234,18 +230,16 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 				} else if (block.type === "thinking") {
 					// Skip empty thinking blocks
 					if (!block.thinking || block.thinking.trim() === "") continue;
-					// Only keep as thinking block if same provider AND same model
-					// Otherwise convert to plain text (no tags to avoid model mimicking them)
-					if (isSameProviderAndModel) {
-						const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thinkingSignature);
+					const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thinkingSignature);
+					if (thoughtSignature) {
 						parts.push({
 							thought: true,
 							text: block.thinking.toWellFormed(),
-							...(thoughtSignature && { thoughtSignature }),
+							thoughtSignature,
 						});
 					} else {
 						parts.push({
-							text: block.thinking.toWellFormed(),
+							text: renderDemotedThinking(model.id, block.thinking),
 						});
 					}
 				} else if (block.type === "toolCall") {
@@ -420,7 +414,7 @@ export function mapStopReason(reason: FinishReason): StopReason {
 		case "NO_IMAGE":
 			return "error";
 		default: {
-			throw new Error(`Unhandled stop reason: ${reason satisfies never}`);
+			throw new AIError.ConfigurationError(`Unhandled stop reason: ${reason satisfies never}`);
 		}
 	}
 }
@@ -597,13 +591,14 @@ export async function consumeGoogleStream<T extends GoogleApiType>(args: {
 			const detail = chunk.error.message || chunk.error.status || "unknown error";
 			const message = `Google API stream error: ${detail}`;
 			throw typeof chunk.error.code === "number" && chunk.error.code >= 400
-				? new GoogleApiError(message, chunk.error.code)
-				: new Error(message);
+				? new AIError.GoogleApiError(message, chunk.error.code)
+				: new AIError.ProviderResponseError(message, { provider: model.provider, kind: "output" });
 		}
 		if (!chunk.candidates?.length && chunk.promptFeedback?.blockReason) {
 			const detail = chunk.promptFeedback.blockReasonMessage;
-			throw new Error(
+			throw new AIError.ProviderResponseError(
 				`Request blocked by Google (${chunk.promptFeedback.blockReason})${detail ? `: ${detail}` : ""}`,
+				{ provider: model.provider, kind: "content-blocked" },
 			);
 		}
 		const candidate = chunk.candidates?.[0];
@@ -734,15 +729,21 @@ export async function consumeGoogleStream<T extends GoogleApiType>(args: {
 	flushCurrent();
 
 	if (options?.signal?.aborted) {
-		throw new Error("Request was aborted");
+		throw new AIError.AbortError();
 	}
 
 	if (!sawFinishReason) {
-		throw new Error("Google API stream ended without a finish reason (connection dropped or response truncated)");
+		throw new AIError.ProviderResponseError(
+			"Google API stream ended without a finish reason (connection dropped or response truncated)",
+			{ provider: model.provider, kind: "incomplete-stream" },
+		);
 	}
 
 	if (output.stopReason === "aborted" || output.stopReason === "error") {
-		throw new Error(output.errorMessage ?? "An unknown error occurred");
+		throw new AIError.ProviderResponseError(output.errorMessage ?? "An unknown error occurred", {
+			provider: model.provider,
+			kind: "output",
+		});
 	}
 }
 
@@ -826,7 +827,7 @@ export function buildGoogleGenerateContentParams<T extends "google-generative-ai
 
 	if (options.signal) {
 		if (options.signal.aborted) {
-			throw new Error("Request aborted");
+			throw new AIError.AbortError("Request aborted");
 		}
 		config.abortSignal = options.signal;
 	}
@@ -864,7 +865,7 @@ export function streamGoogleGenAI<T extends "google-generative-ai" | "google-ver
 	const stream = new AssistantMessageEventStream();
 
 	(async () => {
-		const startTime = Date.now();
+		const startTime = performance.now();
 		let firstTokenTime: number | undefined;
 
 		const output: AssistantMessage = {
@@ -914,14 +915,17 @@ export function streamGoogleGenAI<T extends "google-generative-ai" | "google-ver
 				});
 				if (!response.ok) {
 					const errorText = await response.text().catch(() => "");
-					throw new GoogleApiError(
+					throw new AIError.GoogleApiError(
 						`Google API error (${response.status}): ${extractGoogleErrorMessage(errorText)}`,
 						response.status,
 						{ headers: response.headers },
 					);
 				}
 				if (!response.body) {
-					throw new Error("Google API returned an empty response body");
+					throw new AIError.ProviderResponseError("Google API returned an empty response body", {
+						provider: model.provider,
+						kind: "empty-body",
+					});
 				}
 				return response.body as ReadableStream<Uint8Array>;
 			};
@@ -944,39 +948,37 @@ export function streamGoogleGenAI<T extends "google-generative-ai" | "google-ver
 					options,
 					retainTextSignature,
 					onFirstToken: () => {
-						firstTokenTime = Date.now();
+						firstTokenTime = performance.now();
 					},
 				});
 
 				if (output.stopReason !== "stop" || hasMeaningfulGoogleContent(output)) break;
 				if (emptyAttempt >= MAX_EMPTY_STREAM_RETRIES) {
-					throw new Error(
+					throw new AIError.ProviderResponseError(
 						`Google API returned an empty response (finishReason STOP with no content) after ${MAX_EMPTY_STREAM_RETRIES + 1} attempts`,
+						{ provider: model.provider, kind: "empty-body" },
 					);
 				}
 				try {
 					await scheduler.wait(EMPTY_STREAM_BASE_DELAY_MS * 2 ** emptyAttempt, { signal: options?.signal });
 				} catch {
-					throw new Error("Request was aborted");
+					throw new AIError.AbortError();
 				}
 				resetGoogleStreamOutputForRetry(output);
 				body = await openStream();
 			}
 
-			output.duration = Date.now() - startTime;
+			output.duration = performance.now() - startTime;
 			if (firstTokenTime) output.ttft = firstTokenTime - startTime;
 			stream.push({ type: "done", reason: output.stopReason as "length" | "stop" | "toolUse", message: output });
 			stream.end();
 		} catch (error) {
-			for (const block of output.content) {
-				if ("index" in block) {
-					delete (block as { index?: number }).index;
-				}
-			}
-			output.stopReason = options?.signal?.aborted ? "aborted" : "error";
-			output.errorStatus = extractHttpStatusFromError(error);
-			output.errorMessage = await finalizeErrorMessage(error, rawRequestDump);
-			output.duration = Date.now() - startTime;
+			const result = await AIError.finalize(error, { api: model.api, signal: options?.signal, rawRequestDump });
+			output.stopReason = result.stopReason;
+			output.errorStatus = result.status;
+			output.errorId = result.id;
+			output.errorMessage = result.message;
+			output.duration = performance.now() - startTime;
 			if (firstTokenTime) output.ttft = firstTokenTime - startTime;
 			stream.push({ type: "error", reason: output.stopReason, error: output });
 			stream.end();

package/src/providers/google-vertex.ts

@@ -1,4 +1,5 @@
 import { $env } from "@oh-my-pi/pi-utils";
+import * as AIError from "../error";
 import type { Context, Model, StreamFunction } from "../types";
 import type { AssistantMessageEventStream } from "../utils/event-stream";
 import { getVertexAccessToken } from "./google-auth";
@@ -69,7 +70,7 @@ function resolveApiKey(options?: GoogleVertexOptions): string | undefined {
 function resolveProject(options?: GoogleVertexOptions): string {
 	const project = options?.project || $env.GOOGLE_CLOUD_PROJECT || $env.GCP_PROJECT || $env.GCLOUD_PROJECT;
 	if (!project) {
-		throw new Error(
+		throw new AIError.ConfigurationError(
 			"Vertex AI requires a project ID. Set GOOGLE_CLOUD_PROJECT/GCP_PROJECT/GCLOUD_PROJECT or pass project in options.",
 		);
 	}
@@ -83,7 +84,7 @@ function resolveLocation(options?: GoogleVertexOptions): string {
 	const location =
 		options?.location || $env.GOOGLE_VERTEX_LOCATION || $env.GOOGLE_CLOUD_LOCATION || $env.VERTEX_LOCATION;
 	if (!location) {
-		throw new Error(
+		throw new AIError.ConfigurationError(
 			"Vertex AI requires a location. Set GOOGLE_VERTEX_LOCATION/GOOGLE_CLOUD_LOCATION/VERTEX_LOCATION or pass location in options.",
 		);
 	}

package/src/providers/google.ts

@@ -1,3 +1,4 @@
+import * as AIError from "../error";
 import { getEnvApiKey } from "../stream";
 import type { Context, Model, StreamFunction } from "../types";
 import type { AssistantMessageEventStream } from "../utils/event-stream";
@@ -24,7 +25,10 @@ export const streamGoogle: StreamFunction<"google-generative-ai"> = (
 		prepare: (): GoogleGenAIRequestPlan => {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider);
 			if (!apiKey) {
-				throw new Error("Google Generative AI requires an API key (GEMINI_API_KEY or options.apiKey).");
+				throw new AIError.MissingApiKeyError(
+					undefined,
+					"Google Generative AI requires an API key (GEMINI_API_KEY or options.apiKey).",
+				);
 			}
 			const params = buildGoogleGenerateContentParams(model, context, options ?? {});
 			// `model.baseUrl` already includes the API version segment when set (mirrors the

package/src/providers/mock.ts

@@ -43,6 +43,7 @@
  */
 
 import { registerCustomApi } from "../api-registry";
+import * as AIError from "../error";
 import type {
 	Api,
 	AssistantMessage,
@@ -242,7 +243,7 @@ export function streamMock(
 	if (!isMockModel(model)) {
 		queueMicrotask(() => {
 			stream.fail(
-				new Error(
+				new AIError.ValidationError(
 					"streamMock called with a model not produced by createMockModel(). " + "Pass a MockModel instance.",
 				),
 			);
@@ -288,6 +289,7 @@ async function runMock(
 	options: SimpleStreamOptions | undefined,
 ): Promise<void> {
 	const startedAt = Date.now();
+	const perfStart = performance.now();
 
 	let handler: MockHandler | undefined;
 	try {
@@ -299,7 +301,7 @@ async function runMock(
 
 	if (handler === undefined) {
 		stream.fail(
-			new Error(
+			new AIError.ValidationError(
 				`Mock model "${model.id}" received call ${model.calls.length} but no response or handler is configured.`,
 			),
 		);
@@ -338,7 +340,7 @@ async function runMock(
 		try {
 			await sleep(response.delayMs, options?.signal);
 		} catch {
-			emitTerminalError(stream, model, startedAt, "aborted", "Mock aborted during delay.");
+			emitTerminalError(stream, model, startedAt, perfStart, "aborted", "Mock aborted during delay.");
 			return;
 		}
 	}
@@ -350,7 +352,7 @@ async function runMock(
 				: response.throw instanceof Error
 					? response.throw.message
 					: String(response.throw);
-		emitTerminalError(stream, model, startedAt, "error", message);
+		emitTerminalError(stream, model, startedAt, perfStart, "error", message);
 		return;
 	}
 
@@ -397,7 +399,7 @@ async function runMock(
 	partial.stopDetails = response.stopDetails;
 	partial.errorMessage = response.errorMessage;
 	partial.usage = mergeUsage(response.usage);
-	partial.duration = Date.now() - startedAt;
+	partial.duration = performance.now() - perfStart;
 
 	if (reason === "aborted" || reason === "error") {
 		stream.push({
@@ -460,6 +462,7 @@ function emitTerminalError(
 	stream: AssistantMessageEventStream,
 	model: Model<Api>,
 	startedAt: number,
+	perfStart: number,
 	reason: "aborted" | "error",
 	message: string,
 ): void {
@@ -473,7 +476,7 @@ function emitTerminalError(
 		stopReason: reason as StopReason,
 		errorMessage: message,
 		timestamp: startedAt,
-		duration: Date.now() - startedAt,
+		duration: performance.now() - perfStart,
 	};
 	stream.push({ type: "start", partial: failure });
 	stream.push({ type: "error", reason, error: failure });
@@ -488,7 +491,7 @@ function sleep(ms: number, signal?: AbortSignal): Promise<void> {
 	const onAbort = () => {
 		clearTimeout(timer);
 		signal?.removeEventListener("abort", onAbort);
-		reject(signal?.reason ?? new Error("aborted"));
+		reject(signal?.reason ?? new AIError.AbortError("aborted"));
 	};
 	const timer = setTimeout(() => {
 		signal?.removeEventListener("abort", onAbort);

package/src/providers/ollama.ts

@@ -1,5 +1,5 @@
-import { extractHttpStatusFromError, fetchWithRetry, parseStreamingJson } from "@oh-my-pi/pi-utils";
-import { ProviderHttpError } from "../errors";
+import { fetchWithRetry, parseStreamingJson } from "@oh-my-pi/pi-utils";
+import * as AIError from "../error";
 import { getEnvApiKey } from "../stream";
 import type {
 	Api,
@@ -15,8 +15,9 @@ import type {
 	ToolChoice,
 } from "../types";
 import { normalizeSystemPrompts } from "../utils";
+import { clearStreamingPartialJson, kStreamingPartialJson } from "../utils/block-symbols";
 import { AssistantMessageEventStream } from "../utils/event-stream";
-import { type CapturedHttpErrorResponse, finalizeErrorMessage, type RawHttpRequestDump } from "../utils/http-inspector";
+import type { CapturedHttpErrorResponse, RawHttpRequestDump } from "../utils/http-inspector";
 import {
 	armPreResponseTimeout,
 	getOpenAIStreamFirstEventTimeoutMs,
@@ -32,11 +33,6 @@ import {
 import { transformMessages } from "./transform-messages";
 import { joinTextWithImagePlaceholder, partitionVisionContent } from "./vision-guard";
 
-/** Non-2xx response from the Ollama `/api/chat` endpoint. */
-export class OllamaApiError extends ProviderHttpError {
-	override readonly name = "OllamaApiError";
-}
-
 export interface OllamaChatOptions extends StreamOptions {
 	reasoning?: "minimal" | "low" | "medium" | "high" | "xhigh";
 	disableReasoning?: boolean;
@@ -90,7 +86,7 @@ type OllamaChatChunk = {
 
 type InternalToolCallBlock = AssistantMessage["content"][number] & {
 	type: "toolCall";
-	partialJson?: string;
+	[kStreamingPartialJson]?: string;
 };
 
 function normalizeBaseUrl(baseUrl?: string): string {
@@ -416,9 +412,9 @@ function endToolCallBlock(stream: AssistantMessageEventStream, output: Assistant
 		return;
 	}
 	const toolCall = block as InternalToolCallBlock;
-	if (toolCall.partialJson) {
-		toolCall.arguments = parseStreamingJson<Record<string, unknown>>(toolCall.partialJson);
-		delete toolCall.partialJson;
+	if (toolCall[kStreamingPartialJson]) {
+		toolCall.arguments = parseStreamingJson<Record<string, unknown>>(toolCall[kStreamingPartialJson]);
+		clearStreamingPartialJson(toolCall);
 	}
 	stream.push({ type: "toolcall_end", contentIndex: index, toolCall, partial: output });
 }
@@ -456,7 +452,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
 	void (async () => {
-		const startTime = Date.now();
+		const startTime = performance.now();
 		let firstTokenTime: number | undefined;
 		const output = createEmptyOutput(model);
 		let rawRequestDump: RawHttpRequestDump | undefined;
@@ -469,6 +465,10 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 			? new StreamMarkupHealing({ pattern: streamMarkupHealingPattern })
 			: undefined;
 		let healedToolCallEmitted = false;
+		// Once the provider streams native reasoning (`message.thinking`), drop any
+		// thinking the text-channel healer also recovers so a model that emits both
+		// does not double-count its reasoning.
+		let suppressHealedThinking = false;
 		const endActiveTextBlock = (): void => {
 			if (activeTextIndex === undefined) return;
 			endTextBlock(stream, output, activeTextIndex);
@@ -497,7 +497,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 					partial: output,
 				});
 			}
-			if (!firstTokenTime) firstTokenTime = Date.now();
+			if (!firstTokenTime) firstTokenTime = performance.now();
 		};
 		const appendVisibleThinking = (thinking: string): void => {
 			if (thinking.length === 0) return;
@@ -517,7 +517,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 					partial: output,
 				});
 			}
-			if (!firstTokenTime) firstTokenTime = Date.now();
+			if (!firstTokenTime) firstTokenTime = performance.now();
 		};
 		const emitHealedToolCall = (call: HealedToolCall): void => {
 			endActiveThinkingBlock();
@@ -527,7 +527,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 				id: call.id,
 				name: call.name,
 				arguments: parseStreamingJson<Record<string, unknown>>(call.arguments),
-				partialJson: call.arguments,
+				[kStreamingPartialJson]: call.arguments,
 			};
 			output.content.push(toolCall);
 			const index = output.content.length - 1;
@@ -540,13 +540,13 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 			});
 			endToolCallBlock(stream, output, index);
 			healedToolCallEmitted = true;
-			if (!firstTokenTime) firstTokenTime = Date.now();
+			if (!firstTokenTime) firstTokenTime = performance.now();
 		};
 		const emitHealingEvent = (event: StreamMarkupHealingEvent): void => {
 			if (event.type === "text") {
 				appendVisibleText(event.text);
 			} else if (event.type === "thinking") {
-				appendVisibleThinking(event.thinking);
+				if (!suppressHealedThinking) appendVisibleThinking(event.thinking);
 			} else {
 				emitHealedToolCall(event.call);
 			}
@@ -558,7 +558,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 		try {
 			const apiKey = options.apiKey || getEnvApiKey(model.provider);
 			if (!apiKey) {
-				throw new Error(`No API key for provider: ${model.provider}`);
+				throw new AIError.MissingApiKeyError(model.provider);
 			}
 			const baseUrl = normalizeBaseUrl(model.baseUrl);
 			let body = createChatBody(model, context, options);
@@ -606,16 +606,19 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 			}
 			if (!response.ok) {
 				capturedErrorResponse = await captureHttpErrorResponse(response);
-				throw new OllamaApiError(`HTTP ${response.status} from ${baseUrl}/api/chat`, response.status, {
+				throw new AIError.OllamaApiError(`HTTP ${response.status} from ${baseUrl}/api/chat`, response.status, {
 					headers: response.headers,
 				});
 			}
 			if (!response.body) {
-				throw new Error("Ollama returned an empty response body");
+				throw new AIError.OllamaApiError("Ollama returned an empty response body", response.status, {
+					headers: response.headers,
+				});
 			}
 			stream.push({ type: "start", partial: output });
 			for await (const chunk of iterateNdjson(response.body)) {
 				if (chunk.message?.thinking) {
+					suppressHealedThinking = true;
 					endActiveTextBlock();
 					if (activeThinkingIndex === undefined) {
 						output.content.push({ type: "thinking", thinking: "" });
@@ -633,7 +636,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 						});
 					}
 					if (!firstTokenTime) {
-						firstTokenTime = Date.now();
+						firstTokenTime = performance.now();
 					}
 				}
 				const chunkContent = chunk.message?.content;
@@ -662,7 +665,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 							id: `ollama:${output.content.length}:${name}`,
 							name,
 							arguments: parseStreamingJson<Record<string, unknown>>(partialJson),
-							partialJson,
+							[kStreamingPartialJson]: partialJson,
 						};
 						output.content.push(toolCall);
 						const index = output.content.length - 1;
@@ -675,7 +678,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 							partial: output,
 						});
 						if (!firstTokenTime) {
-							firstTokenTime = Date.now();
+							firstTokenTime = performance.now();
 						}
 					}
 				}
@@ -723,7 +726,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 			if (output.stopReason === "stop" && output.content.some(block => block.type === "toolCall")) {
 				output.stopReason = "toolUse";
 			}
-			output.duration = Date.now() - startTime;
+			output.duration = performance.now() - startTime;
 			if (firstTokenTime) {
 				output.ttft = firstTokenTime - startTime;
 			}
@@ -739,13 +742,20 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 		} catch (error) {
 			for (const block of output.content) {
 				if (block.type === "toolCall") {
-					delete (block as InternalToolCallBlock).partialJson;
+					clearStreamingPartialJson(block);
 				}
 			}
-			output.stopReason = options.signal?.aborted ? "aborted" : "error";
-			output.errorStatus = extractHttpStatusFromError(error);
-			output.errorMessage = await finalizeErrorMessage(error, rawRequestDump, capturedErrorResponse);
-			output.duration = Date.now() - startTime;
+			const result = await AIError.finalize(error, {
+				api: model.api,
+				signal: options.signal,
+				rawRequestDump,
+				capturedErrorResponse,
+			});
+			output.stopReason = result.stopReason;
+			output.errorStatus = result.status;
+			output.errorId = result.id;
+			output.errorMessage = result.message;
+			output.duration = performance.now() - startTime;
 			if (firstTokenTime) {
 				output.ttft = firstTokenTime - startTime;
 			}

package/src/providers/openai-chat-server.ts

@@ -6,6 +6,7 @@ import { resolvePromptCacheKey } from "../auth-gateway/http";
  * `stream(model, context, options)`.
  */
 import type { AuthGatewayStreamControl, AuthGatewayParsedRequest as ParsedRequest } from "../auth-gateway/types";
+import * as AIError from "../error";
 import type {
 	AssistantMessage,
 	AssistantMessageEventStream,
@@ -53,7 +54,7 @@ export function parseRequest(body: unknown, headers?: Headers): ParsedRequest {
 	// vendor-neutral headers when the body doesn't carry one.
 	const parsed = openaiChatRequestSchema(body);
 	if (parsed instanceof type.errors) {
-		throw new Error(`openai-chat: ${parsed.summary}`);
+		throw new AIError.ValidationError(`openai-chat: ${parsed.summary}`);
 	}
 	const data = parsed;
 

package/src/providers/openai-codex/request-transformer.ts

@@ -14,11 +14,11 @@ export interface ReasoningConfig {
 export interface CodexRequestOptions {
 	reasoningEffort?: ReasoningConfig["effort"];
 	reasoningSummary?: ReasoningConfig["summary"] | null;
-	/** Explicit `reasoning.context` override. Defaults to `all_turns` under {@link CodexRequestOptions.responsesLite}, otherwise omitted (server default is `current_turn`). */
+	/** Explicit `reasoning.context` override; defaults to `all_turns` for every Codex request when unset. */
 	reasoningContext?: CodexReasoningContext;
 	textVerbosity?: "low" | "medium" | "high";
 	include?: string[];
-	/** Responses Lite transport contract: strips image detail and defaults `reasoning.context` to `all_turns`, mirroring codex-rs. */
+	/** Responses Lite transport contract: strips image detail and disables parallel tool calling, mirroring codex-rs. */
 	responsesLite?: boolean;
 }
 
@@ -195,8 +195,13 @@ function stripImageDetails(input: InputItem[]): void {
 		for (const collection of [item.content, item.output]) {
 			if (!Array.isArray(collection)) continue;
 			for (const part of collection) {
-				if (part && typeof part === "object" && (part as { type?: unknown }).type === "input_image") {
-					delete (part as { detail?: unknown }).detail;
+				if (
+					part &&
+					typeof part === "object" &&
+					(part as { type?: unknown }).type === "input_image" &&
+					"detail" in part
+				) {
+					part.detail = undefined;
 				}
 			}
 		}
@@ -249,20 +254,16 @@ export async function transformRequestBody(
 			...body.reasoning,
 			...reasoningConfig,
 		};
-		// Responses Lite keeps reasoning replay server-side; codex-rs requests
-		// `all_turns` there and otherwise omits context so the server default
-		// (currently `current_turn`) applies.
-		const reasoningContext = options.reasoningContext ?? (responsesLite ? "all_turns" : undefined);
-		if (reasoningContext !== undefined) {
-			body.reasoning.context = reasoningContext;
-		}
+		// Default reasoning replay to `all_turns` for every Codex request,
+		// mirroring codex-rs; an explicit `reasoningContext` overrides it.
+		body.reasoning.context = options.reasoningContext ?? "all_turns";
 	} else {
 		delete body.reasoning;
 	}
 
 	body.text = {
 		...body.text,
-		verbosity: options.textVerbosity || "low",
+		verbosity: options.textVerbosity || "high",
 	};
 
 	const include = Array.isArray(options.include) ? [...options.include] : [];

package/src/providers/openai-codex/response-handler.ts

@@ -1,5 +1,5 @@
 import { toNumber } from "@oh-my-pi/pi-catalog/utils";
-import { ProviderHttpError } from "../../errors";
+import { ProviderHttpError } from "../../error";
 
 export type CodexRateLimit = {
 	used_percent?: number;