Compare commits

...
Author SHA1 Message Date
paseo-bot[bot] 21a25c39aa test(codex): assert the steer rewind is refused 2026-10-04 02:16:02 +02:00
paseo-bot[bot] c644da16d1 test(codex): move the held-response fixture into a helper 2026-10-04 02:15:45 +02:00
paseo-bot[bot]and3ae3ae dbf41b059e fix(codex): rewind whole turns so a steer never drops earlier prompts
The turn-boundary rule now lives in the Codex rewind helper. It refuses a
steer message on both history paths and counts turns rather than user
messages for thread/rollback, so a steer in a later turn no longer rolls
back one turn too many on Codex versions that still roll back.

Co-authored-by: 3ae3ae <ogyshjsm@naver.com>
2026-10-04 02:11:43 +02:00
paseo-bot[bot] b4bf345a30 Merge main into fix/codex-steer-rewind 2026-10-04 02:09:45 +02:00
Bae Seongmin fe9b3d693e Merge branch 'main' into fix/codex-steer-rewind 2026-09-29 23:25:59 +09:00
3ae3ae 7624140141 Merge main into fix/codex-steer-rewind 2026-09-29 16:01:29 +09:00
3ae3ae 59009a2c2a fix(codex): reject rewinds inside a steered turn 2026-09-29 13:41:45 +09:00
5 changed files with 274 additions and 85 deletions
@@ -2,7 +2,7 @@ import { describe, expect, test, vi } from "vitest";
import { execFileSync } from "node:child_process";
import os from "node:os";
import path from "node:path";
import { createServer } from "node:http";
import { createServer, type ServerResponse } from "node:http";
import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
import { once } from "node:events";
@@ -171,6 +171,62 @@ async function startMockResponsesServer(sequence: string[]): Promise<{
};
}
async function startHeldFirstResponseServer(): Promise<{
url: string;
requestBodies: string[];
releaseFirstResponse: () => void;
close: () => Promise<void>;
}> {
const requestBodies: string[] = [];
let heldResponse: ServerResponse | null = null;
const reply = (res: ServerResponse) => {
res.statusCode = 200;
res.setHeader("content-type", "text/event-stream");
res.end(assistantMessageSse("done"));
};
const server = createServer((req, res) => {
const chunks: Buffer[] = [];
req.on("data", (chunk) => chunks.push(Buffer.from(chunk)));
req.on("end", () => {
if (req.method !== "POST" || req.url !== "/v1/responses") {
res.statusCode = 404;
res.end("not found");
return;
}
requestBodies.push(Buffer.concat(chunks).toString("utf8"));
if (requestBodies.length === 1) {
heldResponse = res;
} else {
reply(res);
}
});
});
server.listen(0, "127.0.0.1");
await once(server, "listening");
const address = server.address();
if (!address || typeof address === "string") {
throw new Error("Expected TCP address for mock responses server");
}
return {
url: `http://127.0.0.1:${address.port}`,
requestBodies,
releaseFirstResponse: () => {
if (!heldResponse) throw new Error("No held response to release");
reply(heldResponse);
heldResponse = null;
},
close: () =>
new Promise<void>((resolve, reject) => {
server.closeAllConnections();
server.close((error) => {
if (error) reject(error);
else resolve();
});
}),
};
}
function writeMockCodexConfig(codexHome: string, serverUrl: string): void {
writeFileSync(
path.join(codexHome, "config.toml"),
@@ -217,6 +273,72 @@ function waitForEvent<TEvent extends AgentStreamEvent>(params: {
}
describe("Codex app-server provider (local e2e)", () => {
test.runIf(isCodexInstalled())(
"rewinding a steer message keeps the earlier prompt of its turn in model context",
async () => {
const cwd = mkdtempSync(path.join(os.tmpdir(), "codex-steer-rewind-cwd-"));
const codexHome = mkdtempSync(path.join(os.tmpdir(), "codex-steer-rewind-home-"));
const mockServer = await startHeldFirstResponseServer();
try {
writeMockCodexConfig(codexHome, mockServer.url);
const client = new CodexAppServerAgentClient(createTestLogger());
const session = await client.createSession(
{ provider: "codex", cwd, modeId: "auto", model: "mock-model" },
{ env: { CODEX_HOME: codexHome } },
);
try {
const userMessageIds = new Map<string, string>();
session.subscribe((event) => {
if (
event.type === "timeline" &&
event.item.type === "user_message" &&
event.item.messageId
) {
userMessageIds.set(event.item.text, event.item.messageId);
}
});
const turnFinished = waitForEvent({
session,
timeoutMs: 15_000,
label: "turn completion",
predicate: (event): event is Extract<AgentStreamEvent, { type: "turn_completed" }> =>
event.type === "turn_completed",
});
const started = await session.startTurn("ORIGINAL_PROMPT");
await vi.waitFor(() => expect(mockServer.requestBodies).toHaveLength(1), {
timeout: 15_000,
});
await expect(
session.steerActiveTurn!("STEER_PROMPT", { expectedTurnId: started.turnId }),
).resolves.toEqual({ status: "accepted" });
mockServer.releaseFirstResponse();
await turnFinished;
await vi.waitFor(() => expect(userMessageIds.get("STEER_PROMPT")).toBeDefined(), {
timeout: 15_000,
});
await expect(
session.revertConversation!({ messageId: userMessageIds.get("STEER_PROMPT")! }),
).rejects.toThrow("Select the first message in the turn instead");
await session.run("AFTER_REWIND");
expect(mockServer.requestBodies.at(-1)).toContain("AFTER_REWIND");
expect(mockServer.requestBodies.at(-1)).toContain("ORIGINAL_PROMPT");
} finally {
await session.close();
}
} finally {
await mockServer.close();
rmSync(cwd, { recursive: true, force: true });
rmSync(codexHome, { recursive: true, force: true });
}
},
45_000,
);
test.runIf(isCodexInstalled())(
"reloads a persisted idle thread repeatedly without overlapping writers",
async () => {
@@ -105,10 +105,7 @@ interface CollaborationModeRecord {
}
interface CodexSessionTestAccess {
codexUserMessageTurns(): {
resolve(messageId: string): { index: number; turnId: string | null } | null;
count(): number;
};
codexUserMessages(): { messageId: string; turnId: string | null }[];
ensureThreadLoaded(): Promise<void>;
handleToolApprovalRequest(params: unknown): Promise<unknown>;
handleNotification(method: string, params: unknown): void;
@@ -2203,6 +2200,60 @@ describe("Codex app-server provider", () => {
await session.close();
});
test.each(["legacy", "paginated"] as const)(
"rejects rewinding a steer message in a %s thread before forking",
async (historyMode) => {
const appServer = createFakeCodexAppServer({
"thread/read": () => ({ thread: { id: "thread-1", historyMode, turns: [] } }),
"turn/steer": () => ({ turnId: "turn-first" }),
});
const session = new CodexAppServerAgentSession(
createConfig({ cwd: "/workspace/project" }),
null,
createTestLogger(),
async () => appServer.child,
);
try {
const started = await session.startTurn("original prompt");
appServer.startsTurn({ threadId: "thread-1", turnId: "turn-first" });
const originalMessage = waitForNextTimelineItem(session, "user_message");
emitCodexUserMessage(appServer, {
id: "codex-first",
text: "original prompt",
turnId: "turn-first",
});
await originalMessage;
await expect(
session.steerActiveTurn("steer prompt", { expectedTurnId: started.turnId }),
).resolves.toEqual({ status: "accepted" });
const steerMessage = waitForNextTimelineItem(session, "user_message");
emitCodexUserMessage(appServer, {
id: "codex-steer",
text: "steer prompt",
turnId: "turn-first",
});
await steerMessage;
appServer.completeTurn();
await expect(session.revertConversation({ messageId: "codex-steer" })).rejects.toThrow(
"Select the first message in the turn instead",
);
expect(appServer.requests().filter((request) => request.method === "thread/fork")).toEqual(
[],
);
expect(appServer.recordedRollbacks).toEqual([]);
expect(session.id).toBe("thread-1");
await session.revertConversation({ messageId: "codex-first" });
expect(session.id).toBe("forked-thread");
appServer.assertNoErrors();
} finally {
await session.close();
}
},
);
test.each(["legacy", "paginated"] as const)(
"rewinds a %s thread onto a fork that keeps the custom provider and runtime MCP servers",
async (historyMode) => {
@@ -4638,10 +4689,9 @@ describe("Codex app-server provider", () => {
await asInternals(session).loadPersistedHistory(session.client);
expect(asInternals(session).codexUserMessageTurns().resolve("message-history")).toEqual({
index: 0,
turnId: "native-turn-1",
});
expect(asInternals(session).codexUserMessages()).toEqual([
{ messageId: "message-history", turnId: "native-turn-1" },
]);
});
test("loads mixed legacy and MultiAgentV2 sub-agent history", async () => {
@@ -91,7 +91,7 @@ import {
type CodexThreadRollbackResponse,
type CodexAppServerTraceContext,
} from "./codex/app-server-transport.js";
import { type CodexUserMessageTurnIndex, revertCodexConversation } from "./codex/rewind.js";
import { type CodexUserMessage, revertCodexConversation } from "./codex/rewind.js";
import {
materializeProviderImage,
renderProviderImageOutputAsAssistantMarkdown,
@@ -4529,16 +4529,11 @@ export class CodexAppServerAgentSession implements AgentSession {
});
}
private codexUserMessageTurns(): CodexUserMessageTurnIndex {
return {
resolve: (messageId) => {
const index = this.userMessageTurnIndexes.get(messageId);
return index === undefined
? null
: { index, turnId: this.userMessageProviderTurnIds.get(messageId) ?? null };
},
count: () => this.userMessageTurnIds.length,
};
private codexUserMessages(): CodexUserMessage[] {
return this.userMessageTurnIds.map((messageId) => ({
messageId,
turnId: this.userMessageProviderTurnIds.get(messageId) ?? null,
}));
}
subscribe(callback: (event: AgentStreamEvent) => void): () => void {
@@ -4955,7 +4950,7 @@ export class CodexAppServerAgentSession implements AgentSession {
model: this.config.model ?? null,
serviceTier: this.serviceTier,
config: this.buildCodexInnerConfig(),
userMessageTurns: this.codexUserMessageTurns(),
userMessages: this.codexUserMessages(),
threadRollbackAvailable: this.threadRollbackAvailable,
setThreadId: async (threadId) => {
this.currentThreadId = threadId;
@@ -6,11 +6,7 @@ import type {
CodexThreadRollbackParams,
CodexThreadRollbackResponse,
} from "./app-server-transport.js";
import {
type CodexUserMessageTurnIndex,
type CodexRewindClient,
revertCodexConversation,
} from "./rewind.js";
import { type CodexRewindClient, revertCodexConversation } from "./rewind.js";
class FakeCodex implements CodexRewindClient {
readonly recordedForks: CodexThreadForkParams[] = [];
@@ -59,33 +55,13 @@ class FakeCodex implements CodexRewindClient {
}
}
class CodexMessageTurns implements CodexUserMessageTurnIndex {
constructor(
private readonly indexesByMessageId: Map<string, number>,
private readonly turnIdsByMessageId: Map<string, string> = new Map(),
) {}
resolve(messageId: string): { index: number; turnId: string | null } | null {
const index = this.indexesByMessageId.get(messageId);
return index === undefined
? null
: { index, turnId: this.turnIdsByMessageId.get(messageId) ?? null };
}
count(): number {
return this.indexesByMessageId.size;
}
}
describe("Codex Rewind", () => {
test("rewinds the conversation by forking the thread and rolling back past the native user message", async () => {
const codex = new FakeCodex();
const userMessageTurns = new CodexMessageTurns(
new Map([
["codex-first", 0],
["codex-second", 1],
]),
);
const userMessages = [
{ messageId: "codex-first", turnId: null },
{ messageId: "codex-second", turnId: null },
];
let reboundThreadId: string | null = null;
await revertCodexConversation({
@@ -95,7 +71,7 @@ describe("Codex Rewind", () => {
cwd: "/workspace/project",
model: "gpt-5.4-mini",
serviceTier: null,
userMessageTurns,
userMessages,
threadRollbackAvailable: true,
setThreadId: (threadId) => {
reboundThreadId = threadId;
@@ -118,20 +94,18 @@ describe("Codex Rewind", () => {
test("rewinds the conversation using native user message ids hydrated from app-server history", async () => {
const codex = new FakeCodex();
const userMessageTurns = new CodexMessageTurns(
new Map([
["codex-first", 0],
["codex-second", 1],
["codex-third", 2],
]),
);
const userMessages = [
{ messageId: "codex-first", turnId: null },
{ messageId: "codex-second", turnId: null },
{ messageId: "codex-third", turnId: null },
];
let reboundThreadId: string | null = null;
await revertCodexConversation({
client: codex,
threadId: "source-thread",
messageId: "codex-second",
userMessageTurns,
userMessages,
threadRollbackAvailable: true,
setThreadId: (threadId) => {
reboundThreadId = threadId;
@@ -157,16 +131,10 @@ describe("Codex Rewind", () => {
}
const codex = new PaginatedCodex();
const userMessageTurns = new CodexMessageTurns(
new Map([
["codex-first", 0],
["codex-second", 1],
]),
new Map([
["codex-first", "turn-first"],
["codex-second", "turn-second"],
]),
);
const userMessages = [
{ messageId: "codex-first", turnId: "turn-first" },
{ messageId: "codex-second", turnId: "turn-second" },
];
let reboundThreadId: string | null = null;
await revertCodexConversation({
@@ -176,7 +144,7 @@ describe("Codex Rewind", () => {
cwd: "/workspace/project",
model: "gpt-5.4-mini",
serviceTier: null,
userMessageTurns,
userMessages,
threadRollbackAvailable: true,
setThreadId: (threadId) => {
reboundThreadId = threadId;
@@ -209,7 +177,7 @@ describe("Codex Rewind", () => {
}
const codex = new PaginatedCodex();
const userMessageTurns = new CodexMessageTurns(new Map([["codex-first", 0]]));
const userMessages = [{ messageId: "codex-first", turnId: null }];
let reboundThreadId: string | null = null;
await expect(
@@ -220,7 +188,7 @@ describe("Codex Rewind", () => {
cwd: "/workspace/project",
model: "gpt-5.4-mini",
serviceTier: null,
userMessageTurns,
userMessages,
threadRollbackAvailable: true,
setThreadId: (threadId) => {
reboundThreadId = threadId;
@@ -235,14 +203,14 @@ describe("Codex Rewind", () => {
test("declines to rewind when the user message is not in the Codex thread", async () => {
const codex = new FakeCodex();
const userMessageTurns = new CodexMessageTurns(new Map([["codex-first", 0]]));
const userMessages = [{ messageId: "codex-first", turnId: null }];
await expect(
revertCodexConversation({
client: codex,
threadId: "source-thread",
messageId: "missing-message",
userMessageTurns,
userMessages,
threadRollbackAvailable: true,
setThreadId: () => undefined,
}),
@@ -250,4 +218,47 @@ describe("Codex Rewind", () => {
expect(codex.recordedForks).toEqual([]);
expect(codex.recordedRollbacks).toEqual([]);
});
test("rolls back whole turns when a later turn holds a steer message", async () => {
const codex = new FakeCodex();
await revertCodexConversation({
client: codex,
threadId: "source-thread",
messageId: "codex-second",
userMessages: [
{ messageId: "codex-first", turnId: "turn-first" },
{ messageId: "codex-second", turnId: "turn-second" },
{ messageId: "codex-steer", turnId: "turn-second" },
],
threadRollbackAvailable: true,
setThreadId: () => undefined,
});
expect(codex.recordedRollbacks).toEqual([{ threadId: "forked-thread", numTurns: 1 }]);
});
test("declines to rewind a steer message without forking", async () => {
const codex = new FakeCodex();
let reboundThreadId: string | null = null;
await expect(
revertCodexConversation({
client: codex,
threadId: "source-thread",
messageId: "codex-steer",
userMessages: [
{ messageId: "codex-first", turnId: "turn-first" },
{ messageId: "codex-steer", turnId: "turn-first" },
],
threadRollbackAvailable: true,
setThreadId: (threadId) => {
reboundThreadId = threadId;
},
}),
).rejects.toThrow("Select the first message in the turn instead");
expect(codex.recordedForks).toEqual([]);
expect(codex.recordedRollbacks).toEqual([]);
expect(reboundThreadId).toBeNull();
});
});
@@ -15,9 +15,9 @@ export interface CodexRewindClient {
request(method: string, params?: unknown, timeoutMs?: number): Promise<unknown>;
}
export interface CodexUserMessageTurnIndex {
resolve(messageId: string): { index: number; turnId: string | null } | null;
count(): number;
export interface CodexUserMessage {
messageId: string;
turnId: string | null;
}
type CodexThreadHistoryMode = "legacy" | "paginated";
@@ -60,6 +60,12 @@ async function rollbackCodexThread(
return parseCodexThreadRollbackResponse(await client.request("thread/rollback", params));
}
function countTurns(userMessages: readonly CodexUserMessage[]): number {
return userMessages.filter(
(message, index) => !message.turnId || message.turnId !== userMessages[index - 1]?.turnId,
).length;
}
export async function revertCodexConversation(input: {
client: CodexRewindClient;
threadId: string | null;
@@ -68,7 +74,7 @@ export async function revertCodexConversation(input: {
model?: string | null;
serviceTier?: string | null;
config?: Record<string, unknown> | null;
userMessageTurns: CodexUserMessageTurnIndex;
userMessages: readonly CodexUserMessage[];
threadRollbackAvailable: boolean;
setThreadId: (threadId: string) => void | Promise<void>;
}): Promise<void> {
@@ -76,16 +82,21 @@ export async function revertCodexConversation(input: {
throw new Error("Codex thread is not ready for rewind");
}
const targetTurn = input.userMessageTurns.resolve(input.messageId);
if (targetTurn === null) {
const targetIndex = input.userMessages.findIndex(
(message) => message.messageId === input.messageId,
);
if (targetIndex === -1) {
throw new Error(`Codex could not find user message ${input.messageId} in the current thread`);
}
const currentUserTurnCount = input.userMessageTurns.count();
const numTurns = currentUserTurnCount - targetTurn.index;
if (numTurns < 0) {
throw new Error(`Codex user message ${input.messageId} is outside the current thread`);
const targetTurn = input.userMessages[targetIndex];
// Codex rewinds whole turns; a steer cannot be removed without the earlier
// messages of its turn.
if (targetTurn.turnId && input.userMessages[targetIndex - 1]?.turnId === targetTurn.turnId) {
throw new Error(
"Codex cannot rewind a message sent during a turn without removing earlier messages. Select the first message in the turn instead.",
);
}
const numTurns = countTurns(input.userMessages.slice(targetIndex));
// Codex does not carry the parent thread's config into a fork; without it the
// forked thread falls back to the default model provider.