diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index e6b3deb8..3bb43438 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added `websocket-cached` transport support for OpenAI Codex Responses used with ChatGPT subscription auth. This keeps the same WebSocket open for a session and, after the first request, sends only new conversation items instead of resending the full chat history when possible.
+
 ## [0.71.0] - 2026-04-30
 
 ### Breaking Changes
diff --git a/packages/ai/src/index.ts b/packages/ai/src/index.ts
index b1e7f66e..8798b024 100644
--- a/packages/ai/src/index.ts
+++ b/packages/ai/src/index.ts
@@ -12,7 +12,15 @@ export type { GoogleOptions } from "./providers/google.js";
 export type { GoogleThinkingLevel } from "./providers/google-shared.js";
 export type { GoogleVertexOptions } from "./providers/google-vertex.js";
 export type { MistralOptions } from "./providers/mistral.js";
-export type { OpenAICodexResponsesOptions } from "./providers/openai-codex-responses.js";
+export type {
+	OpenAICodexResponsesOptions,
+	OpenAICodexWebSocketDebugStats,
+} from "./providers/openai-codex-responses.js";
+export {
+	closeOpenAICodexWebSocketSessions,
+	getOpenAICodexWebSocketDebugStats,
+	resetOpenAICodexWebSocketDebugStats,
+} from "./providers/openai-codex-responses.js";
 export type { OpenAICompletionsOptions } from "./providers/openai-completions.js";
 export type { OpenAIResponsesOptions } from "./providers/openai-responses.js";
 export * from "./providers/register-builtins.js";
diff --git a/packages/ai/src/providers/openai-codex-responses.ts b/packages/ai/src/providers/openai-codex-responses.ts
index 541cb98a..6469e741 100644
--- a/packages/ai/src/providers/openai-codex-responses.ts
+++ b/packages/ai/src/providers/openai-codex-responses.ts
@@ -74,6 +74,7 @@ interface RequestBody {
 	store?: boolean;
 	stream?: boolean;
 	instructions?: string;
+	previous_response_id?: string;
 	input?: ResponseInput;
 	tools?: OpenAITool[];
 	tool_choice?: "auto";
@@ -193,7 +194,7 @@ export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses"
 					stream.end();
 					return;
 				} catch (error) {
-					if (transport === "websocket" || websocketStarted) {
+					if (transport === "websocket" || transport === "websocket-cached" || websocketStarted) {
 						throw error;
 					}
 				}
@@ -536,13 +537,82 @@ interface WebSocketLike {
 	removeEventListener(type: WebSocketEventType, listener: WebSocketListener): void;
 }
 
+interface CachedWebSocketContinuationState {
+	lastRequestBody: RequestBody;
+	lastResponseId: string;
+	lastResponseItems: ResponseInput;
+}
+
 interface CachedWebSocketConnection {
 	socket: WebSocketLike;
 	busy: boolean;
 	idleTimer?: ReturnType<typeof setTimeout>;
+	continuation?: CachedWebSocketContinuationState;
+}
+
+export interface OpenAICodexWebSocketDebugStats {
+	requests: number;
+	connectionsCreated: number;
+	connectionsReused: number;
+	cachedContextRequests: number;
+	storeTrueRequests: number;
+	fullContextRequests: number;
+	deltaRequests: number;
+	lastInputItems: number;
+	lastDeltaInputItems?: number;
+	lastPreviousResponseId?: string;
 }
 
 const websocketSessionCache = new Map<string, CachedWebSocketConnection>();
+const websocketDebugStats = new Map<string, OpenAICodexWebSocketDebugStats>();
+
+function getOrCreateWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats {
+	let stats = websocketDebugStats.get(sessionId);
+	if (!stats) {
+		stats = {
+			requests: 0,
+			connectionsCreated: 0,
+			connectionsReused: 0,
+			cachedContextRequests: 0,
+			storeTrueRequests: 0,
+			fullContextRequests: 0,
+			deltaRequests: 0,
+			lastInputItems: 0,
+		};
+		websocketDebugStats.set(sessionId, stats);
+	}
+	return stats;
+}
+
+export function getOpenAICodexWebSocketDebugStats(sessionId: string): OpenAICodexWebSocketDebugStats | undefined {
+	const stats = websocketDebugStats.get(sessionId);
+	return stats ? { ...stats } : undefined;
+}
+
+export function resetOpenAICodexWebSocketDebugStats(sessionId?: string): void {
+	if (sessionId) {
+		websocketDebugStats.delete(sessionId);
+		return;
+	}
+	websocketDebugStats.clear();
+}
+
+export function closeOpenAICodexWebSocketSessions(sessionId?: string): void {
+	const closeEntry = (entry: CachedWebSocketConnection) => {
+		if (entry.idleTimer) clearTimeout(entry.idleTimer);
+		closeWebSocketSilently(entry.socket, 1000, "debug_close");
+	};
+	if (sessionId) {
+		const entry = websocketSessionCache.get(sessionId);
+		if (entry) closeEntry(entry);
+		websocketSessionCache.delete(sessionId);
+		return;
+	}
+	for (const entry of websocketSessionCache.values()) {
+		closeEntry(entry);
+	}
+	websocketSessionCache.clear();
+}
 
 type WebSocketConstructor = new (
 	url: string,
@@ -650,11 +720,17 @@ async function acquireWebSocket(
 	headers: Headers,
 	sessionId: string | undefined,
 	signal?: AbortSignal,
-): Promise<{ socket: WebSocketLike; release: (options?: { keep?: boolean }) => void }> {
+): Promise<{
+	socket: WebSocketLike;
+	entry?: CachedWebSocketConnection;
+	reused: boolean;
+	release: (options?: { keep?: boolean }) => void;
+}> {
 	if (!sessionId) {
 		const socket = await connectWebSocket(url, headers, signal);
 		return {
 			socket,
+			reused: false,
 			release: ({ keep } = {}) => {
 				if (keep === false) {
 					closeWebSocketSilently(socket);
@@ -675,6 +751,8 @@ async function acquireWebSocket(
 			cached.busy = true;
 			return {
 				socket: cached.socket,
+				entry: cached,
+				reused: true,
 				release: ({ keep } = {}) => {
 					if (!keep || !isWebSocketReusable(cached.socket)) {
 						closeWebSocketSilently(cached.socket);
@@ -690,6 +768,7 @@ async function acquireWebSocket(
 			const socket = await connectWebSocket(url, headers, signal);
 			return {
 				socket,
+				reused: false,
 				release: () => {
 					closeWebSocketSilently(socket);
 				},
@@ -706,6 +785,8 @@ async function acquireWebSocket(
 	websocketSessionCache.set(sessionId, entry);
 	return {
 		socket,
+		entry,
+		reused: false,
 		release: ({ keep } = {}) => {
 			if (!keep || !isWebSocketReusable(entry.socket)) {
 				closeWebSocketSilently(entry.socket);
@@ -850,6 +931,60 @@ async function* parseWebSocket(socket: WebSocketLike, signal?: AbortSignal): Asy
 	}
 }
 
+function requestBodyWithoutInput(body: RequestBody): RequestBody {
+	const { input: _input, previous_response_id: _previousResponseId, ...rest } = body;
+	return rest;
+}
+
+function responseInputsEqual(a: ResponseInput | undefined, b: ResponseInput | undefined): boolean {
+	return JSON.stringify(a ?? []) === JSON.stringify(b ?? []);
+}
+
+function requestBodiesMatchExceptInput(a: RequestBody, b: RequestBody): boolean {
+	return JSON.stringify(requestBodyWithoutInput(a)) === JSON.stringify(requestBodyWithoutInput(b));
+}
+
+function getCachedWebSocketInputDelta(
+	body: RequestBody,
+	continuation: CachedWebSocketContinuationState,
+): ResponseInput | undefined {
+	if (!requestBodiesMatchExceptInput(body, continuation.lastRequestBody)) {
+		return undefined;
+	}
+
+	const currentInput = body.input ?? [];
+	const baseline = [...(continuation.lastRequestBody.input ?? []), ...continuation.lastResponseItems];
+	if (currentInput.length < baseline.length) {
+		return undefined;
+	}
+
+	const prefix = currentInput.slice(0, baseline.length);
+	if (!responseInputsEqual(prefix, baseline)) {
+		return undefined;
+	}
+
+	return currentInput.slice(baseline.length);
+}
+
+function buildCachedWebSocketRequestBody(entry: CachedWebSocketConnection, body: RequestBody): RequestBody {
+	const continuation = entry.continuation;
+	if (!continuation) {
+		return body;
+	}
+
+	const delta = getCachedWebSocketInputDelta(body, continuation);
+	if (!delta || !continuation.lastResponseId) {
+		entry.continuation = undefined;
+		return body;
+	}
+
+	return {
+		...body,
+		previous_response_id: continuation.lastResponseId,
+		input: delta,
+	};
+}
+
 async function processWebSocketStream(
 	url: string,
 	body: RequestBody,
@@ -860,10 +995,33 @@ async function processWebSocketStream(
 	onStart: () => void,
 	options?: OpenAICodexResponsesOptions,
 ): Promise<void> {
-	const { socket, release } = await acquireWebSocket(url, headers, options?.sessionId, options?.signal);
+	const { socket, entry, reused, release } = await acquireWebSocket(url, headers, options?.sessionId, options?.signal);
 	let keepConnection = true;
+	const useCachedContext = options?.transport === "websocket-cached";
+	// ChatGPT Codex Responses rejects `store: true` ("Store must be set to false").
+	// WebSocket continuation still works via connection-scoped previous_response_id state.
+	const fullBody = body;
+	const requestBody = useCachedContext && entry ? buildCachedWebSocketRequestBody(entry, fullBody) : fullBody;
+	const stats = options?.sessionId ? getOrCreateWebSocketDebugStats(options.sessionId) : undefined;
+	if (stats) {
+		stats.requests++;
+		if (reused) stats.connectionsReused++;
+		else stats.connectionsCreated++;
+		if (useCachedContext) stats.cachedContextRequests++;
+		if (requestBody.store === true) stats.storeTrueRequests++;
+		stats.lastInputItems = requestBody.input?.length ?? 0;
+		if (requestBody.previous_response_id) {
+			stats.deltaRequests++;
+			stats.lastDeltaInputItems = requestBody.input?.length ?? 0;
+			stats.lastPreviousResponseId = requestBody.previous_response_id;
+		} else {
+			stats.fullContextRequests++;
+			stats.lastDeltaInputItems = undefined;
+			stats.lastPreviousResponseId = undefined;
+		}
+	}
 	try {
-		socket.send(JSON.stringify({ type: "response.create", ...body }));
+		socket.send(JSON.stringify({ type: "response.create", ...requestBody }));
 		onStart();
 		stream.push({ type: "start", partial: output });
 		await processResponsesStream(mapCodexEvents(parseWebSocket(socket, options?.signal)), output, stream, model, {
@@ -873,8 +1031,20 @@ async function processWebSocketStream(
 		});
 		if (options?.signal?.aborted) {
 			keepConnection = false;
+		} else if (useCachedContext && entry && output.responseId) {
+			const responseItems = convertResponsesMessages(model, { messages: [output] }, CODEX_TOOL_CALL_PROVIDERS, {
+				includeSystemPrompt: false,
+			}).filter((item) => item.type !== "function_call_output");
+			entry.continuation = {
+				lastRequestBody: fullBody,
+				lastResponseId: output.responseId,
+				lastResponseItems: responseItems,
+			};
 		}
 	} catch (error) {
+		if (entry) {
+			entry.continuation = undefined;
+		}
 		keepConnection = false;
 		throw error;
 	} finally {
diff --git a/packages/ai/src/types.ts b/packages/ai/src/types.ts
index bdca2a84..ae89baf7 100644
--- a/packages/ai/src/types.ts
+++ b/packages/ai/src/types.ts
@@ -58,7 +58,7 @@ export interface ThinkingBudgets {
 // Base options all providers share
 export type CacheRetention = "none" | "short" | "long";
 
-export type Transport = "sse" | "websocket" | "auto";
+export type Transport = "sse" | "websocket" | "websocket-cached" | "auto";
 
 export interface ProviderResponse {
 	status: number;
diff --git a/packages/ai/test/codex-websocket-cached-probe.ts b/packages/ai/test/codex-websocket-cached-probe.ts
new file mode 100644
index 00000000..ccc875f8
--- /dev/null
+++ b/packages/ai/test/codex-websocket-cached-probe.ts
@@ -0,0 +1,291 @@
+#!/usr/bin/env tsx
+/**
+ * Live probe for OpenAI Codex Responses websocket-cached mode.
+ *
+ * Runs a simple tool loop directly against the pi-ai provider source so it does not
+ * depend on built dist packages or coding-agent SDK wiring.
+ */
+
+import { tmpdir } from "node:os";
+import { join, resolve } from "node:path";
+import { Type } from "typebox";
+import { AuthStorage } from "../../coding-agent/src/core/auth-storage.js";
+import { getModel } from "../src/models.js";
+import {
+	closeOpenAICodexWebSocketSessions,
+	getOpenAICodexWebSocketDebugStats,
+	resetOpenAICodexWebSocketDebugStats,
+	streamOpenAICodexResponses,
+} from "../src/providers/openai-codex-responses.js";
+import type { AssistantMessage, Context, Message, Model, Tool, ToolResultMessage, Transport } from "../src/types.js";
+
+type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh";
+
+interface Args {
+	turns: number;
+	transport: Transport;
+	maxTokens: number;
+	reasoning: ThinkingLevel;
+	sessionId: string;
+}
+
+const DEFAULT_TURNS = 20;
+const DEFAULT_MAX_TOKENS = 64;
+
+function parseArgs(argv: string[]): Args {
+	let turns = DEFAULT_TURNS;
+	let transport: Transport = "websocket-cached";
+	let maxTokens = DEFAULT_MAX_TOKENS;
+	let reasoning: ThinkingLevel = "low";
+	let sessionId = `pi-ai-codex-ws-cached-probe-${Date.now()}`;
+
+	for (let i = 0; i < argv.length; i++) {
+		const arg = argv[i];
+		switch (arg) {
+			case "--turns":
+				turns = Number.parseInt(required(argv[++i], arg), 10);
+				break;
+			case "--transport": {
+				const value = required(argv[++i], arg);
+				if (value !== "sse" && value !== "websocket" && value !== "websocket-cached" && value !== "auto") {
+					throw new Error(`Invalid --transport: ${value}`);
+				}
+				transport = value;
+				break;
+			}
+			case "--max-tokens":
+				maxTokens = Number.parseInt(required(argv[++i], arg), 10);
+				break;
+			case "--reasoning": {
+				const value = required(argv[++i], arg);
+				if (value !== "minimal" && value !== "low" && value !== "medium" && value !== "high" && value !== "xhigh") {
+					throw new Error(`Invalid --reasoning: ${value}`);
+				}
+				reasoning = value;
+				break;
+			}
+			case "--session-id":
+				sessionId = required(argv[++i], arg);
+				break;
+			case "--help":
+				printHelp();
+				process.exit(0);
+				break;
+			default:
+				throw new Error(`Unknown argument: ${arg}`);
+		}
+	}
+
+	return { turns, transport, maxTokens, reasoning, sessionId };
+}
+
+function required(value: string | undefined, flag: string): string {
+	if (!value) throw new Error(`Missing value for ${flag}`);
+	return value;
+}
+
+function printHelp(): void {
+	console.log(`Usage: npx tsx test/codex-websocket-cached-probe.ts [options]
+
+Options:
+  --turns <n>          Number of user turns. Default: ${DEFAULT_TURNS}
+  --transport <mode>   sse | websocket | websocket-cached | auto. Default: websocket-cached
+  --reasoning <level>  minimal | low | medium | high | xhigh. Default: low
+  --max-tokens <n>     Max output tokens per model request. Default: ${DEFAULT_MAX_TOKENS}
+  --session-id <id>    Session id for websocket/cache state
+`);
+}
+
+function buildPrompt(turn: number): string {
+	const marker = `TURN-${String(turn).padStart(2, "0")}-MARKER-${(turn * 17 + 13) % 97}`;
+	const lines = [
+		"This is an automated OpenAI Codex Responses websocket cache probe.",
+		`Task for turn ${turn}: call deterministic_probe exactly once before your final answer.`,
+		`Use tool arguments: turn=${turn}, marker=${marker}`,
+		`After the tool result arrives, reply exactly: TURN ${turn} OK ${marker}`,
+		"The following repeated block is intentional benchmark padding.",
+	];
+	for (let i = 1; i <= 180; i++) {
+		lines.push(
+			`Turn ${turn} synthetic record ${String(i).padStart(3, "0")}: alpha beta gamma delta epsilon zeta eta theta iota kappa lambda mu nu xi omicron pi rho sigma tau upsilon phi chi psi omega.`,
+		);
+	}
+	return lines.join("\n");
+}
+
+function deterministicProbeTool(): Tool {
+	return {
+		name: "deterministic_probe",
+		description: "Mandatory benchmark tool. Call exactly once with the turn and marker from the user prompt.",
+		parameters: Type.Object({
+			turn: Type.Number(),
+			marker: Type.String(),
+		}),
+	};
+}
+
+function executeTool(call: Extract<AssistantMessage["content"][number], { type: "toolCall" }>): ToolResultMessage {
+	return {
+		role: "toolResult",
+		toolCallId: call.id,
+		toolName: call.name,
+		content: [{ type: "text", text: `deterministic_probe_result ${JSON.stringify(call.arguments)} fixed=OK` }],
+		details: { fixed: "OK" },
+		isError: false,
+		timestamp: Date.now(),
+	};
+}
+
+function textOf(message: AssistantMessage): string {
+	return message.content
+		.filter((block): block is Extract<AssistantMessage["content"][number], { type: "text" }> => block.type === "text")
+		.map((block) => block.text)
+		.join("\n")
+		.trim();
+}
+
+function average(values: number[]): number {
+	return values.reduce((sum, value) => sum + value, 0) / Math.max(1, values.length);
+}
+
+function percentile(values: number[], p: number): number {
+	if (values.length === 0) return 0;
+	const sorted = [...values].sort((a, b) => a - b);
+	return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil((p / 100) * sorted.length) - 1))];
+}
+
+async function main(): Promise<void> {
+	const args = parseArgs(process.argv.slice(2));
+	const model = getModel("openai-codex", "gpt-5.5") as Model<"openai-codex-responses"> | undefined;
+	if (!model) throw new Error("Model openai-codex/gpt-5.5 not found");
+	const modelWithMaxTokens = { ...model, maxTokens: args.maxTokens };
+	const authStorage = AuthStorage.create();
+	const apiKey = (await authStorage.getApiKey("openai-codex")) ?? (await authStorage.getApiKey("openai"));
+	if (!apiKey) {
+		throw new Error("No OpenAI Codex API key found in coding-agent auth storage.");
+	}
+	const context: Context = {
+		systemPrompt:
+			"You are participating in a benchmark. For each benchmark turn, call deterministic_probe exactly once before the final answer. Keep final answers minimal.",
+		messages: [],
+		tools: [deterministicProbeTool()],
+	};
+	const elapsed: number[] = [];
+	resetOpenAICodexWebSocketDebugStats(args.sessionId);
+
+	console.log(`provider openai-codex, model gpt-5.5`);
+	console.log(`sessionId ${args.sessionId}`);
+	console.log(
+		`turns ${args.turns}, transport ${args.transport}, reasoning ${args.reasoning}, maxTokens ${args.maxTokens}`,
+	);
+	console.log(`scratch ${resolve(join(tmpdir(), args.sessionId))}`);
+	console.log("");
+
+	for (let turn = 1; turn <= args.turns; turn++) {
+		context.messages.push({ role: "user", content: buildPrompt(turn), timestamp: Date.now() });
+		const beforeStats = getOpenAICodexWebSocketDebugStats(args.sessionId);
+		const started = Date.now();
+		let requests = 0;
+		let assistantCount = 0;
+		let toolResults = 0;
+		let finalText = "";
+		let turnInput = 0;
+		let turnOutput = 0;
+		let turnCacheRead = 0;
+		let turnCacheWrite = 0;
+
+		while (true) {
+			requests++;
+			const message = await streamOpenAICodexResponses(modelWithMaxTokens, context, {
+				apiKey,
+				sessionId: args.sessionId,
+				transport: args.transport,
+				reasoningEffort: args.reasoning,
+				maxTokens: args.maxTokens,
+			}).result();
+			assistantCount++;
+			context.messages.push(message);
+			turnInput += message.usage.input;
+			turnOutput += message.usage.output;
+			turnCacheRead += message.usage.cacheRead;
+			turnCacheWrite += message.usage.cacheWrite;
+			const toolCalls = message.content.filter(
+				(block): block is Extract<AssistantMessage["content"][number], { type: "toolCall" }> =>
+					block.type === "toolCall",
+			);
+			console.log(
+				[
+					`turn ${String(turn).padStart(2, "0")}.${requests}`,
+					`stop ${message.stopReason}`,
+					`in ${message.usage.input}`,
+					`out ${message.usage.output}`,
+					`cache ${message.usage.cacheRead}/${message.usage.cacheWrite}`,
+					`tools ${toolCalls.length}`,
+				].join(" | "),
+			);
+			if (message.stopReason === "error" || message.stopReason === "aborted") {
+				throw new Error(message.errorMessage ?? `request failed on turn ${turn}.${requests}`);
+			}
+			if (toolCalls.length === 0) {
+				finalText = textOf(message);
+				break;
+			}
+			for (const call of toolCalls) {
+				context.messages.push(executeTool(call) as Message);
+				toolResults++;
+			}
+			if (requests > 4) throw new Error(`Too many requests for turn ${turn}`);
+		}
+
+		const elapsedMs = Date.now() - started;
+		elapsed.push(elapsedMs);
+		const afterStats = getOpenAICodexWebSocketDebugStats(args.sessionId);
+		const statLine = afterStats
+			? `ws requests ${afterStats.requests - (beforeStats?.requests ?? 0)} | new/reused ${afterStats.connectionsCreated - (beforeStats?.connectionsCreated ?? 0)}/${afterStats.connectionsReused - (beforeStats?.connectionsReused ?? 0)} | cached ${afterStats.cachedContextRequests - (beforeStats?.cachedContextRequests ?? 0)} | store ${afterStats.storeTrueRequests - (beforeStats?.storeTrueRequests ?? 0)} | full/delta ${afterStats.fullContextRequests - (beforeStats?.fullContextRequests ?? 0)}/${afterStats.deltaRequests - (beforeStats?.deltaRequests ?? 0)}`
+			: "ws none";
+		console.log(
+			[
+				`turn ${String(turn).padStart(2, "0")} agg`,
+				`elapsed ${(elapsedMs / 1000).toFixed(1)}s`,
+				`assistant ${assistantCount}`,
+				`toolResults ${toolResults}`,
+				`in ${turnInput}`,
+				`out ${turnOutput}`,
+				`cache ${turnCacheRead}/${turnCacheWrite}`,
+				statLine,
+				`final ${JSON.stringify(finalText).slice(0, 80)}`,
+			].join(" | "),
+		);
+	}
+
+	const stats = getOpenAICodexWebSocketDebugStats(args.sessionId);
+	console.log("");
+	console.log(
+		[
+			"timing",
+			`turns ${elapsed.length}`,
+			`total ${(elapsed.reduce((sum, value) => sum + value, 0) / 1000).toFixed(1)}s`,
+			`avg ${(average(elapsed) / 1000).toFixed(2)}s`,
+			`p50 ${(percentile(elapsed, 50) / 1000).toFixed(2)}s`,
+			`p95 ${(percentile(elapsed, 95) / 1000).toFixed(2)}s`,
+			`max ${(Math.max(...elapsed) / 1000).toFixed(2)}s`,
+		].join(" | "),
+	);
+	console.log(
+		[
+			"transport summary",
+			`requested ${args.transport}`,
+			`observed ${stats && stats.requests > 0 ? "websocket" : "sse/no-websocket"}`,
+			`storeTrue ${stats ? `${stats.storeTrueRequests}/${stats.requests}` : "0/0"}`,
+			`full/delta ${stats ? `${stats.fullContextRequests}/${stats.deltaRequests}` : "0/0"}`,
+			`connections created/reused ${stats ? `${stats.connectionsCreated}/${stats.connectionsReused}` : "0/0"}`,
+			`lastPreviousResponseId ${stats?.lastPreviousResponseId ?? "n/a"}`,
+		].join(" | "),
+	);
+	closeOpenAICodexWebSocketSessions(args.sessionId);
+}
+
+main().catch((error: unknown) => {
+	console.error(error instanceof Error ? error.message : String(error));
+	process.exitCode = 1;
+});
diff --git a/packages/ai/test/openai-codex-stream.test.ts b/packages/ai/test/openai-codex-stream.test.ts
index 90be500b..ab51333a 100644
--- a/packages/ai/test/openai-codex-stream.test.ts
+++ b/packages/ai/test/openai-codex-stream.test.ts
@@ -3,21 +3,26 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { afterEach, describe, expect, it, vi } from "vitest";
 import {
+	getOpenAICodexWebSocketDebugStats,
+	resetOpenAICodexWebSocketDebugStats,
 	streamOpenAICodexResponses,
 	streamSimpleOpenAICodexResponses,
 } from "../src/providers/openai-codex-responses.js";
 import type { Context, Model } from "../src/types.js";
 
 const originalFetch = global.fetch;
+const originalWebSocket = globalThis.WebSocket;
 const originalAgentDir = process.env.PI_CODING_AGENT_DIR;
 
 afterEach(() => {
 	global.fetch = originalFetch;
+	globalThis.WebSocket = originalWebSocket;
 	if (originalAgentDir === undefined) {
 		delete process.env.PI_CODING_AGENT_DIR;
 	} else {
 		process.env.PI_CODING_AGENT_DIR = originalAgentDir;
 	}
+	resetOpenAICodexWebSocketDebugStats();
 	vi.restoreAllMocks();
 });
 
@@ -746,4 +751,150 @@ describe("openai-codex streaming", () => {
 		const streamResult = streamOpenAICodexResponses(model, context, { apiKey: token });
 		await streamResult.result();
 	});
+	it("sends only response input deltas in websocket-cached mode", async () => {
+		const token = mockToken();
+		const sentBodies: unknown[] = [];
+		const responses = [
+			{ responseId: "resp_1", messageId: "msg_1", text: "Hello" },
+			{ responseId: "resp_2", messageId: "msg_2", text: "Done" },
+		];
+
+		class MockWebSocket {
+			static OPEN = 1;
+			readyState = MockWebSocket.OPEN;
+			private listeners = new Map<string, Set<(event: unknown) => void>>();
+
+			constructor(_url: string, _protocols?: string | string[] | { headers?: Record<string, string> }) {
+				queueMicrotask(() => this.dispatch("open", {}));
+			}
+
+			addEventListener(type: string, listener: (event: unknown) => void): void {
+				let listeners = this.listeners.get(type);
+				if (!listeners) {
+					listeners = new Set();
+					this.listeners.set(type, listeners);
+				}
+				listeners.add(listener);
+			}
+
+			removeEventListener(type: string, listener: (event: unknown) => void): void {
+				this.listeners.get(type)?.delete(listener);
+			}
+
+			send(data: string): void {
+				sentBodies.push(JSON.parse(data));
+				const response = responses.shift();
+				if (!response) throw new Error("unexpected websocket request");
+				const events = [
+					{ type: "response.created", response: { id: response.responseId } },
+					{
+						type: "response.output_item.added",
+						item: {
+							type: "message",
+							id: response.messageId,
+							role: "assistant",
+							status: "in_progress",
+							content: [],
+						},
+					},
+					{ type: "response.content_part.added", part: { type: "output_text", text: "" } },
+					{ type: "response.output_text.delta", delta: response.text },
+					{
+						type: "response.output_item.done",
+						item: {
+							type: "message",
+							id: response.messageId,
+							role: "assistant",
+							status: "completed",
+							content: [{ type: "output_text", text: response.text }],
+						},
+					},
+					{
+						type: "response.completed",
+						response: {
+							id: response.responseId,
+							status: "completed",
+							usage: {
+								input_tokens: 5,
+								output_tokens: 3,
+								total_tokens: 8,
+								input_tokens_details: { cached_tokens: 0 },
+							},
+						},
+					},
+				];
+				queueMicrotask(() => {
+					for (const event of events) {
+						this.dispatch("message", { data: JSON.stringify(event) });
+					}
+				});
+			}
+
+			close(): void {
+				this.readyState = 3;
+			}
+
+			private dispatch(type: string, event: unknown): void {
+				for (const listener of this.listeners.get(type) ?? []) {
+					listener(event);
+				}
+			}
+		}
+
+		globalThis.WebSocket = MockWebSocket as unknown as typeof WebSocket;
+
+		const model: Model<"openai-codex-responses"> = {
+			id: "gpt-5.1-codex",
+			name: "GPT-5.1 Codex",
+			api: "openai-codex-responses",
+			provider: "openai-codex",
+			baseUrl: "https://chatgpt.com/backend-api",
+			reasoning: true,
+			input: ["text"],
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+			contextWindow: 400000,
+			maxTokens: 128000,
+		};
+		const firstContext: Context = {
+			systemPrompt: "You are a helpful assistant.",
+			messages: [{ role: "user", content: "Say hello", timestamp: 1 }],
+		};
+
+		const first = await streamOpenAICodexResponses(model, firstContext, {
+			apiKey: token,
+			sessionId: "session-1",
+			transport: "websocket-cached",
+		}).result();
+
+		const secondContext: Context = {
+			systemPrompt: "You are a helpful assistant.",
+			messages: [...firstContext.messages, first, { role: "user", content: "Now finish", timestamp: 2 }],
+		};
+		await streamOpenAICodexResponses(model, secondContext, {
+			apiKey: token,
+			sessionId: "session-1",
+			transport: "websocket-cached",
+		}).result();
+
+		expect(sentBodies).toHaveLength(2);
+		const firstBody = sentBodies[0] as { input: unknown[]; previous_response_id?: string; store?: boolean };
+		const secondBody = sentBodies[1] as { input: unknown[]; previous_response_id?: string; store?: boolean };
+		expect(firstBody.store).toBe(false);
+		expect(firstBody.previous_response_id).toBeUndefined();
+		expect(firstBody.input).toEqual([{ role: "user", content: [{ type: "input_text", text: "Say hello" }] }]);
+		expect(secondBody.store).toBe(false);
+		expect(secondBody.previous_response_id).toBe("resp_1");
+		expect(secondBody.input).toEqual([{ role: "user", content: [{ type: "input_text", text: "Now finish" }] }]);
+		expect(getOpenAICodexWebSocketDebugStats("session-1")).toMatchObject({
+			requests: 2,
+			connectionsCreated: 1,
+			connectionsReused: 1,
+			cachedContextRequests: 2,
+			storeTrueRequests: 0,
+			fullContextRequests: 1,
+			deltaRequests: 1,
+			lastDeltaInputItems: 1,
+			lastPreviousResponseId: "resp_1",
+		});
+	});
 });
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index ff91ea43..987224c1 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added `websocket-cached` to the transport setting options for the OpenAI Codex provider used with ChatGPT subscription auth. This keeps the same WebSocket open for a session and, after the first request, sends only the new conversation items instead of resending the full chat history when possible.
+
 ## [0.71.0] - 2026-04-30
 
 ### Breaking Changes
diff --git a/packages/coding-agent/src/modes/interactive/components/settings-selector.ts b/packages/coding-agent/src/modes/interactive/components/settings-selector.ts
index 0bc92052..dff07822 100644
--- a/packages/coding-agent/src/modes/interactive/components/settings-selector.ts
+++ b/packages/coding-agent/src/modes/interactive/components/settings-selector.ts
@@ -235,7 +235,7 @@ export class SettingsSelectorComponent extends Container {
 				label: "Transport",
 				description: "Preferred transport for providers that support multiple transports",
 				currentValue: config.transport,
-				values: ["sse", "websocket", "auto"],
+				values: ["sse", "websocket", "websocket-cached", "auto"],
 			},
 			{
 				id: "hide-thinking",
diff --git a/packages/coding-agent/test/sdk-codex-cache-probe-tool-loop.ts b/packages/coding-agent/test/sdk-codex-cache-probe-tool-loop.ts
index 96a9b77d..19aa1bac 100644
--- a/packages/coding-agent/test/sdk-codex-cache-probe-tool-loop.ts
+++ b/packages/coding-agent/test/sdk-codex-cache-probe-tool-loop.ts
@@ -11,7 +11,20 @@ import { mkdirSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { dirname, join, resolve } from "node:path";
 import process from "node:process";
-import { type AssistantMessage, getModel, Type } from "@mariozechner/pi-ai";
+import {
+	type Api,
+	type AssistantMessage,
+	type AssistantMessageEventStream,
+	type Context,
+	getModel,
+	type Model,
+	type SimpleStreamOptions,
+	Type,
+} from "@mariozechner/pi-ai";
+import {
+	getOpenAICodexWebSocketDebugStats,
+	streamSimpleOpenAICodexResponses,
+} from "../../ai/src/providers/openai-codex-responses.js";
 import { AuthStorage } from "../src/core/auth-storage.js";
 import { createExtensionRuntime } from "../src/core/extensions/loader.js";
 import type { ToolDefinition } from "../src/core/extensions/types.js";
@@ -21,7 +34,7 @@ import { createAgentSession } from "../src/core/sdk.js";
 import { SessionManager } from "../src/core/session-manager.js";
 import { SettingsManager } from "../src/core/settings-manager.js";
 
-type Transport = "sse" | "websocket" | "auto";
+type Transport = "sse" | "websocket" | "websocket-cached" | "auto";
 
 interface Args {
 	turns: number;
@@ -30,6 +43,16 @@ interface Args {
 	maxTokens: number;
 }
 
+interface WebSocketStatsSnapshot {
+	requests: number;
+	connectionsCreated: number;
+	connectionsReused: number;
+	cachedContextRequests: number;
+	storeTrueRequests: number;
+	fullContextRequests: number;
+	deltaRequests: number;
+}
+
 interface SubrequestRecord {
 	turn: number;
 	subrequest: number;
@@ -67,7 +90,7 @@ function parseArgs(argv: string[]): Args {
 			}
 			case "--transport": {
 				const value = argv[++i];
-				if (value !== "sse" && value !== "websocket" && value !== "auto") {
+				if (value !== "sse" && value !== "websocket" && value !== "websocket-cached" && value !== "auto") {
 					throw new Error(`Invalid --transport value: ${value}`);
 				}
 				transport = value;
@@ -105,14 +128,14 @@ function printHelp(): void {
 Options:
   --turns <n>         Number of turns to run. Must be between ${MIN_TURNS} and ${MAX_TURNS}. Default: ${DEFAULT_TURNS}
   --session <path>    Specific session jsonl file to write
-  --transport <mode>  sse | websocket | auto. Default: sse
+  --transport <mode>  sse | websocket | websocket-cached | auto. Default: sse
   --max-tokens <n>    Max output tokens per subrequest. Default: ${DEFAULT_MAX_TOKENS}
   --help              Show this message
 
 Notes:
   - Uses createAgentSession() from the coding-agent SDK
-  - Provider/model fixed to openai-codex/gpt-5.4
-  - Thinking level fixed to medium
+  - Provider/model fixed to openai-codex/gpt-5.5
+  - Thinking level fixed to low
   - Activates exactly one deterministic custom tool
   - Prompts are intentionally > 1024 tokens and explicitly describe the test
 `);
@@ -161,6 +184,54 @@ function createMinimalResourceLoader(systemPrompt: string): ResourceLoader {
 	};
 }
 
+function average(values: number[]): number {
+	return values.length === 0 ? 0 : values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+
+function percentile(values: number[], percentileValue: number): number {
+	if (values.length === 0) return 0;
+	const sorted = [...values].sort((a, b) => a - b);
+	const index = Math.min(sorted.length - 1, Math.max(0, Math.ceil((percentileValue / 100) * sorted.length) - 1));
+	return sorted[index];
+}
+
+function getWebSocketStatsSnapshot(sessionId: string): WebSocketStatsSnapshot {
+	const stats = getOpenAICodexWebSocketDebugStats(sessionId);
+	return {
+		requests: stats?.requests ?? 0,
+		connectionsCreated: stats?.connectionsCreated ?? 0,
+		connectionsReused: stats?.connectionsReused ?? 0,
+		cachedContextRequests: stats?.cachedContextRequests ?? 0,
+		storeTrueRequests: stats?.storeTrueRequests ?? 0,
+		fullContextRequests: stats?.fullContextRequests ?? 0,
+		deltaRequests: stats?.deltaRequests ?? 0,
+	};
+}
+
+function diffWebSocketStats(after: WebSocketStatsSnapshot, before: WebSocketStatsSnapshot): WebSocketStatsSnapshot {
+	return {
+		requests: after.requests - before.requests,
+		connectionsCreated: after.connectionsCreated - before.connectionsCreated,
+		connectionsReused: after.connectionsReused - before.connectionsReused,
+		cachedContextRequests: after.cachedContextRequests - before.cachedContextRequests,
+		storeTrueRequests: after.storeTrueRequests - before.storeTrueRequests,
+		fullContextRequests: after.fullContextRequests - before.fullContextRequests,
+		deltaRequests: after.deltaRequests - before.deltaRequests,
+	};
+}
+
+function formatWebSocketStats(label: string, stats: WebSocketStatsSnapshot): string {
+	if (stats.requests === 0) return `${label} websocket none`;
+	return [
+		`${label} websocket`,
+		`requests ${stats.requests}`,
+		`new/reused ${stats.connectionsCreated}/${stats.connectionsReused}`,
+		`cached ${stats.cachedContextRequests}`,
+		`store ${stats.storeTrueRequests}`,
+		`full/delta ${stats.fullContextRequests}/${stats.deltaRequests}`,
+	].join(" | ");
+}
+
 function getAssistantText(message: AssistantMessage): string {
 	return message.content
 		.filter((block): block is Extract<AssistantMessage["content"][number], { type: "text" }> => block.type === "text")
@@ -206,11 +277,24 @@ async function main(): Promise<void> {
 	const authStorage = AuthStorage.create();
 	const modelRegistry = ModelRegistry.create(authStorage);
 
-	const model = getModel("openai-codex", "gpt-5.4");
+	const model = getModel("openai-codex", "gpt-5.5");
 	if (!model) {
-		throw new Error("Model openai-codex/gpt-5.4 not found");
+		throw new Error("Model openai-codex/gpt-5.5 not found");
 	}
 	const baseModel = { ...model, maxTokens: args.maxTokens };
+	const streamSimpleOpenAICodexResponsesForRegistry = (
+		registryModel: Model<Api>,
+		context: Context,
+		options?: SimpleStreamOptions,
+	): AssistantMessageEventStream =>
+		streamSimpleOpenAICodexResponses(registryModel as Model<"openai-codex-responses">, context, options);
+	modelRegistry.registerProvider("openai-codex", {
+		api: "openai-codex-responses",
+		baseUrl: baseModel.baseUrl,
+		apiKey: "!echo source-provider-override-uses-auth-storage",
+		streamSimple: streamSimpleOpenAICodexResponsesForRegistry,
+		models: [baseModel],
+	});
 
 	const settingsManager = SettingsManager.inMemory({
 		compaction: { enabled: false },
@@ -226,7 +310,7 @@ async function main(): Promise<void> {
 		cwd: process.cwd(),
 		agentDir: dirname(args.sessionPath),
 		model: baseModel,
-		thinkingLevel: "medium",
+		thinkingLevel: "low",
 		customTools: [deterministicProbeTool() as unknown as ToolDefinition],
 		resourceLoader,
 		sessionManager: SessionManager.open(args.sessionPath),
@@ -239,20 +323,23 @@ async function main(): Promise<void> {
 	const unsubscribe = session.subscribe(() => {});
 
 	const records: SubrequestRecord[] = [];
+	const turnElapsedMs: number[] = [];
 	let previousCacheRead: number | null = null;
 
-	console.log(`provider openai-codex, model gpt-5.4`);
+	console.log(`provider openai-codex, model gpt-5.5`);
 	console.log(`session ${session.sessionFile}`);
-	console.log(`turns ${args.turns}, transport ${args.transport}, reasoning medium, maxTokens ${args.maxTokens}`);
+	console.log(`turns ${args.turns}, transport ${args.transport}, reasoning low, maxTokens ${args.maxTokens}`);
 	console.log("");
 
 	for (let turn = 1; turn <= args.turns; turn++) {
 		const prompt = buildPrompt(turn);
 		const promptTokens = estimateTokens(prompt);
 		const previousMessagesLength = session.messages.length;
+		const websocketStatsBefore = getWebSocketStatsSnapshot(session.sessionId);
 		const startedAt = Date.now();
 		await session.prompt(prompt);
 		const elapsedMs = Date.now() - startedAt;
+		turnElapsedMs.push(elapsedMs);
 
 		const newMessages = session.messages.slice(previousMessagesLength);
 		const assistantMessages = newMessages.filter((message): message is AssistantMessage =>
@@ -316,6 +403,8 @@ async function main(): Promise<void> {
 			previousCacheRead = assistant.usage.cacheRead;
 		}
 
+		const websocketStatsAfter = getWebSocketStatsSnapshot(session.sessionId);
+		const websocketStatsForTurn = diffWebSocketStats(websocketStatsAfter, websocketStatsBefore);
 		console.log(
 			[
 				`turn ${String(turn).padStart(2, "0")} agg`,
@@ -327,6 +416,7 @@ async function main(): Promise<void> {
 				`total ${turnTotal}`,
 			].join(" | "),
 		);
+		console.log(formatWebSocketStats(`turn ${String(turn).padStart(2, "0")}`, websocketStatsForTurn));
 	}
 
 	const violations = records
@@ -343,7 +433,50 @@ async function main(): Promise<void> {
 		})
 		.filter((value): value is NonNullable<typeof value> => value !== null);
 
+	const totalElapsedMs = turnElapsedMs.reduce((sum, value) => sum + value, 0);
 	console.log("");
+	console.log(
+		[
+			"timing",
+			`turns ${turnElapsedMs.length}`,
+			`total ${(totalElapsedMs / 1000).toFixed(1)}s`,
+			`avg ${(average(turnElapsedMs) / 1000).toFixed(2)}s`,
+			`p50 ${(percentile(turnElapsedMs, 50) / 1000).toFixed(2)}s`,
+			`p95 ${(percentile(turnElapsedMs, 95) / 1000).toFixed(2)}s`,
+			`max ${(Math.max(...turnElapsedMs) / 1000).toFixed(2)}s`,
+		].join(" | "),
+	);
+	const websocketStats = getOpenAICodexWebSocketDebugStats(session.sessionId);
+	const requestedWebsocket =
+		args.transport === "websocket" || args.transport === "websocket-cached" || args.transport === "auto";
+	const observedWebsocket = Boolean(websocketStats && websocketStats.requests > 0);
+	console.log(
+		[
+			"transport summary",
+			`requested ${args.transport}`,
+			`observed ${observedWebsocket ? "websocket" : "sse/no-websocket"}`,
+			`sseFallbackSuspected ${requestedWebsocket && !observedWebsocket ? "yes" : "no"}`,
+			`cachedContext ${websocketStats?.cachedContextRequests ? "yes" : "no"}`,
+			`storeTrue ${websocketStats ? `${websocketStats.storeTrueRequests}/${websocketStats.requests}` : "0/0"}`,
+			`delta ${websocketStats ? `${websocketStats.deltaRequests}/${websocketStats.requests}` : "0/0"}`,
+			`full ${websocketStats ? `${websocketStats.fullContextRequests}/${websocketStats.requests}` : "0/0"}`,
+		].join(" | "),
+	);
+	if (websocketStats) {
+		console.log(
+			[
+				"websocket details",
+				`requests ${websocketStats.requests}`,
+				`connections created/reused ${websocketStats.connectionsCreated}/${websocketStats.connectionsReused}`,
+				`cachedContext ${websocketStats.cachedContextRequests}`,
+				`storeTrue ${websocketStats.storeTrueRequests}`,
+				`full/delta ${websocketStats.fullContextRequests}/${websocketStats.deltaRequests}`,
+				`lastInputItems ${websocketStats.lastInputItems}`,
+				`lastDeltaItems ${websocketStats.lastDeltaInputItems ?? "n/a"}`,
+				`lastPreviousResponseId ${websocketStats.lastPreviousResponseId ?? "n/a"}`,
+			].join(" | "),
+		);
+	}
 	console.log(`subrequest cache read monotonic: ${violations.length === 0 ? "yes" : "NO"}`);
 	if (violations.length > 0) {
 		console.log("violations:");