Compare commits

...
Author SHA1 Message Date
Mohamed Boudra 757bf81cb9 docs(hub): clarify credential storage and retention 2026-09-10 08:21:13 +02:00
Mohamed Boudra c56c4bb390 docs(hub): explain execution recovery across restarts 2026-09-10 00:21:01 +02:00
Mohamed Boudra 92504cd525 docs(hub): document configurable startup timeout (#4601)
* docs(hub): document configurable startup timeout

* docs: keep startup timeout guidance in one place
2026-09-09 23:37:54 +02:00
Tom GringauzandTom Gringauz 726067b437 fix(claude): open a single compaction marker per compaction (#4391)
Suppress repeated Claude compaction heartbeat markers at the provider and re-arm marker creation at compaction and turn boundaries. Verify emitted events through the public session interface, including subsequent compactions after interrupted or abandoned turns.

Closes #4390.

Co-authored-by: Tom Gringauz <17593920+tomgrin10@users.noreply.github.com>
2026-09-09 22:45:30 +02:00
Mohamed Boudra ec4e81ce0f Show Codex questions while the agent keeps working (#4587)
* fix(codex): show asynchronous questions and preserve their lifecycle

* fix(server): reject concurrent replies to the same question

* test(codex): use native temporary paths in question fixtures
2026-09-09 20:32:21 +02:00
Christoph Leiter bc385b5930 fix(app): don't steal focus when workspace creation finishes late (#2986)
Preserve the selected workspace when delayed creation finishes, while still starting the requested agent or terminal. Keep normal navigation when the initiating screen remains active, and preserve newer draft text and context.

Verified with before/after recordings, ten passing real-browser journeys using isolated daemons, focused regression tests, green CI, and independent review.

QA: https://github.com/getpaseo/paseo/tree/qa/workspace-focus-2986

Co-authored-by: Christoph Leiter <christoph@leiter.io>
2026-09-09 19:45:24 +02:00
Mohamed Boudra ec5ab3fc4d revert(app): undo file Find merge pending maintainer review
Reverts 69510ac72f. The feature must return to an open pull request for maintainer review.
2026-09-09 19:39:14 +02:00
Mohamed Boudra 69510ac72f feat(app): add pane Find to file source (#4586)
Make Find discoverable at the top right of editable and read-only web file panes, using shared controls backed by CodeMirror search. Preserve literal case-insensitive buffer search, keyboard focus and navigation, optional replacement, Undo, and save.
2026-09-09 19:28:36 +02:00
22 changed files with 1671 additions and 61 deletions
@@ -0,0 +1,22 @@
import { test } from "../support/fixtures";
import { verifyDelayedWorkspaceCreation } from "../support/helpers/new-workspace-navigation";
test.describe("Delayed workspace creation", () => {
test.describe.configure({ timeout: 120_000 });
for (const launch of ["chat", "terminal", "empty"] as const) {
test(`${launch}: keeps the workspace chosen during creation`, async ({ page }) => {
await verifyDelayedWorkspaceCreation(page, launch, "leave");
});
test(`worktree ${launch}: keeps the workspace chosen during creation`, async ({ page }) => {
await verifyDelayedWorkspaceCreation(page, launch, "leave", "worktree");
});
test(`${launch}: opens the created workspace when the user stays`, async ({ page }) => {
await verifyDelayedWorkspaceCreation(page, launch, "stay");
});
}
test("preserves a newer draft opened while creation is pending", async ({ page }) => {
await verifyDelayedWorkspaceCreation(page, "chat", "new-draft");
});
});
@@ -0,0 +1,197 @@
import { expect, test, type Page } from "../fixtures";
import { readFile } from "node:fs/promises";
import { buildHostWorkspaceRoute } from "@/utils/host-routes";
import { gotoAppShell } from "./app";
import { scrollTimelineToOldestLoadedEdge } from "./timeline-pagination";
import {
archiveLocalWorkspaceFromDaemon,
archiveWorkspaceFromDaemon,
connectNewWorkspaceDaemonClient,
countWorkspaceAgents,
countWorkspaceTerminals,
delayBrowserWorkspaceCreatedResponse,
expectNewWorkspaceDraft,
fillNewWorkspaceDraft,
openNewWorkspaceComposer,
openProjectViaDaemon,
selectWorkspaceIsolation,
submitNewWorkspaceEmpty,
submitNewWorkspacePrompt,
waitForCreatedWorkspace,
} from "./new-workspace";
import { createTempGitRepo } from "./workspace";
import { expectAppRoute } from "./route-assertions";
import { getServerId } from "./server-id";
import {
expectTerminalOutputContains,
expectWorkspaceOpensWithTerminalTab,
fillTerminalPrompt,
seedTerminalProfiles,
selectLaunchOption,
submitTerminalLaunch,
type TerminalProfile,
} from "./new-workspace-launch";
import { switchWorkspaceViaSidebar, waitForSidebarHydration } from "./workspace-ui";
import { dropFileOnComposer, expectAttachmentPill } from "./composer";
const PROMPT = "Hello from the navigation guard";
const NEXT_PROMPT = "Keep this newer draft";
const CONTEXT = {
name: "context.json",
mimeType: "application/json",
buffer: Buffer.from('{"preserve":"attachment"}'),
};
const PROFILE: TerminalProfile = {
id: "e2e-nav-guard-echo",
name: "Nav Guard Echo",
command: "/bin/sh",
args: ["-c", 'echo captured: "$0"; sleep 120', "{{{prompt}}}"],
};
export async function verifyDelayedWorkspaceCreation(
page: Page,
launch: "chat" | "terminal" | "empty",
destination: "leave" | "stay" | "new-draft",
isolation: "local" | "worktree" = "local",
): Promise<void> {
const client = await connectNewWorkspaceDaemonClient();
const repo = await createTempGitRepo("workspace-focus-");
const profileSeed = await seedTerminalProfiles([PROFILE]);
const delay = await delayBrowserWorkspaceCreatedResponse(page);
const serverId = getServerId();
let localWorkspaceId: string | undefined;
let createdDirectory: string | undefined;
let newDraftRoute = "";
try {
const project = await openProjectViaDaemon(client, repo.path);
localWorkspaceId = project.workspaceId;
const knownIds = new Set((await client.fetchWorkspaces()).entries.map((entry) => entry.id));
await gotoAppShell(page);
await waitForSidebarHydration(page);
await switchWorkspaceViaSidebar({ page, serverId, workspaceId: project.workspaceId });
await test.step("Submit workspace creation with the requested content", async () => {
await openNewWorkspaceComposer(page, project);
await selectWorkspaceIsolation(page, isolation);
if (launch === "chat") {
await dropFileOnComposer(page, CONTEXT);
await expectAttachmentPill(page, "composer-file-attachment-pill");
await submitNewWorkspacePrompt(page, PROMPT);
} else if (launch === "terminal") {
await selectLaunchOption(page, PROFILE.id);
await fillTerminalPrompt(page, PROMPT);
await submitTerminalLaunch(page);
} else {
await submitNewWorkspaceEmpty(page);
}
await delay.waitForCreateRequest();
});
await test.step("Choose where to work while the creation response is held", async () => {
if (destination !== "stay") {
await switchWorkspaceViaSidebar({ page, serverId, workspaceId: project.workspaceId });
}
if (destination === "new-draft") {
await openNewWorkspaceComposer(page, project);
await fillNewWorkspaceDraft(page, NEXT_PROMPT);
const url = new URL(page.url());
newDraftRoute = url.pathname + url.search;
}
// Recording pacing only; the response is held deterministically above.
if (process.env.E2E_RECORD_VIDEO === "1") {
await page.mouse.move(1100, 80);
await page.waitForTimeout(1500);
}
delay.release();
});
const created = await waitForCreatedWorkspace(client, knownIds);
createdDirectory = created.workspaceDirectory;
expect(created.workspaceKind).toBe(isolation === "worktree" ? "worktree" : "local_checkout");
await test.step("Completion creates exactly the requested resource and preserves focus", async () => {
await expect
.poll(() => countWorkspaceAgents(client, created.id))
.toBe(launch === "chat" ? 1 : 0);
await expect
.poll(() => countWorkspaceTerminals(client, created.id))
.toBe(launch === "terminal" ? 1 : 0);
// A negative navigation assertion needs an observation window after completion.
await page.waitForTimeout(5000);
const expectedRoute =
destination === "new-draft"
? newDraftRoute
: buildHostWorkspaceRoute(
serverId,
destination === "stay" ? created.id : project.workspaceId,
);
await expectAppRoute(page, expectedRoute);
if (destination === "new-draft") await expectNewWorkspaceDraft(page, NEXT_PROMPT);
if (launch === "chat") {
expect(delay.agentRequests).toHaveLength(1);
const upload = delay.agentRequests[0]?.attachments?.find(
(item) => item.type === "uploaded_file",
);
if (!upload) throw new Error("The create request did not contain the submitted file");
expect(await readFile(upload.path)).toEqual(CONTEXT.buffer);
expect(delay.agentRequests[0]).toMatchObject({
workspaceId: created.id,
initialPrompt: PROMPT,
config: {
provider: "mock",
model: "ten-second-stream",
modeId: "load-test",
thinkingOptionId: "low",
cwd: created.workspaceDirectory,
},
attachments: [
expect.objectContaining({
fileName: CONTEXT.name,
mimeType: CONTEXT.mimeType,
size: CONTEXT.buffer.length,
}),
],
});
}
});
await test.step("Visit the new workspace and return without submitting a second draft", async () => {
await switchWorkspaceViaSidebar({ page, serverId, workspaceId: created.id });
if (launch === "chat") {
await scrollTimelineToOldestLoadedEdge(page);
await expect(page.getByText(PROMPT, { exact: true }).first()).toBeVisible();
} else if (launch === "terminal") {
await expectWorkspaceOpensWithTerminalTab(page);
await expectTerminalOutputContains(page, `captured: ${PROMPT}`);
}
await switchWorkspaceViaSidebar({ page, serverId, workspaceId: project.workspaceId });
await switchWorkspaceViaSidebar({ page, serverId, workspaceId: created.id });
await expect
.poll(() => countWorkspaceAgents(client, created.id))
.toBe(launch === "chat" ? 1 : 0);
await expect
.poll(() => countWorkspaceTerminals(client, created.id))
.toBe(launch === "terminal" ? 1 : 0);
if (launch === "chat") {
expect(delay.agentRequests).toHaveLength(1);
await openNewWorkspaceComposer(page, project);
await expectNewWorkspaceDraft(page, destination === "new-draft" ? NEXT_PROMPT : "");
}
if (process.env.E2E_RECORD_VIDEO === "1") {
await page.mouse.move(1100, 80);
await page.waitForTimeout(1500);
}
});
} catch (error) {
await page.screenshot({ path: test.info().outputPath("before-cleanup.png") });
console.error(error);
throw error;
} finally {
delay.release();
if (createdDirectory) await archiveWorkspaceFromDaemon(client, createdDirectory);
if (localWorkspaceId) await archiveLocalWorkspaceFromDaemon(client, localWorkspaceId);
await profileSeed.restore();
await client.close();
await repo.cleanup();
}
}
@@ -1,4 +1,5 @@
import { expect, type BrowserContext, type Page } from "@playwright/test";
import type { CreateAgentRequestMessage } from "@getpaseo/protocol/messages";
import type { DaemonClient as InternalDaemonClient } from "@getpaseo/client/internal/daemon-client";
import { decodeWorkspaceIdFromPathSegment } from "@/utils/host-routes";
import { connectDaemonClient } from "./daemon-client-loader";
@@ -16,6 +17,7 @@ type NewWorkspaceDaemonClient = Pick<
| "connect"
| "createPaseoWorktree"
| "createWorkspace"
| "fetchAgents"
| "fetchWorkspaces"
| "getPaseoWorktreeList"
| "getDaemonConfig"
@@ -24,6 +26,7 @@ type NewWorkspaceDaemonClient = Pick<
| "enablePlugin"
| "inspectWorkspaceRecovery"
| "listProjects"
| "listTerminals"
| "on"
| "patchDaemonConfig"
| "removeProject"
@@ -622,3 +625,120 @@ export async function delayBrowserAgentCreatedStatus(
waitForDelayedCreatedStatus: () => delayedCreatedStatusSeen,
};
}
export interface WorkspaceCreatedDelayControl {
agentRequests: readonly CreateAgentRequestMessage[];
release(): void;
waitForCreateRequest(): Promise<void>;
}
/**
* Holds `workspace.create.response` so a test can act as the user does while the daemon is still
* running `git worktree add` — most importantly, navigate somewhere else.
*/
export async function delayBrowserWorkspaceCreatedResponse(
page: Page,
): Promise<WorkspaceCreatedDelayControl> {
const daemonPortPattern = daemonWsRoutePattern();
const agentRequests: CreateAgentRequestMessage[] = [];
const createRequestIds = new Set<string>();
const delayedForwards: Array<() => void> = [];
let releaseRequested = false;
let resolveCreateRequest: (() => void) | null = null;
const createRequestSeen = new Promise<void>((resolve) => {
resolveCreateRequest = resolve;
});
await page.routeWebSocket(daemonPortPattern, (ws) => {
const server = ws.connectToServer();
ws.onMessage((message) => {
const sessionMessage = getSessionMessage(message);
if (sessionMessage?.type === "create_agent_request")
agentRequests.push(sessionMessage as CreateAgentRequestMessage);
if (sessionMessage?.type === "workspace.create.request") {
const requestId = getStringField(sessionMessage, "requestId");
if (requestId) {
createRequestIds.add(requestId);
resolveCreateRequest?.();
}
}
server.send(message);
});
server.onMessage((message) => {
const sessionMessage = getSessionMessage(message);
const payload =
sessionMessage?.type === "workspace.create.response" &&
typeof sessionMessage.payload === "object"
? (sessionMessage.payload as Record<string, unknown>)
: null;
const requestId = payload ? getStringField(payload, "requestId") : null;
if (requestId && createRequestIds.has(requestId) && !releaseRequested) {
delayedForwards.push(() => ws.send(message));
return;
}
ws.send(message);
});
});
return {
agentRequests,
release() {
releaseRequested = true;
for (const forward of delayedForwards.splice(0)) {
forward();
}
},
waitForCreateRequest: () => createRequestSeen,
};
}
type WorkspaceEntry = Awaited<
ReturnType<NewWorkspaceDaemonClient["fetchWorkspaces"]>
>["entries"][number];
/**
* Waits for a workspace the caller did not already know about, and returns it. Tests that create a
* workspace without navigating to it have no route to assert on, so the daemon's own list is the
* signal that creation finished.
*/
export async function waitForCreatedWorkspace(
client: NewWorkspaceDaemonClient,
knownWorkspaceIds: ReadonlySet<string>,
options?: { timeout?: number },
): Promise<WorkspaceEntry> {
let created: WorkspaceEntry | undefined;
await expect
.poll(
async () => {
const payload = await client.fetchWorkspaces();
created = payload.entries.find((entry) => !knownWorkspaceIds.has(entry.id));
return created !== undefined;
},
{ timeout: options?.timeout ?? 60_000 },
)
.toBe(true);
if (!created) {
throw new Error("Workspace list reported a new workspace but it could not be read back");
}
return created;
}
export async function countWorkspaceAgents(
client: NewWorkspaceDaemonClient,
workspaceId: string,
): Promise<number> {
const payload = await client.fetchAgents({ filter: { includeArchived: true } });
return payload.entries.filter((entry) => entry.agent.workspaceId === workspaceId).length;
}
export async function countWorkspaceTerminals(
client: NewWorkspaceDaemonClient,
workspaceId: string,
): Promise<number> {
const payload = await client.listTerminals(undefined, undefined, { workspaceId });
return payload.terminals.length;
}
@@ -80,8 +80,11 @@ function QuestionOptionRow({
);
const optionLabelStyle = useMemo(
() => [styles.optionLabel, { color: theme.colors.foreground }],
[theme.colors.foreground],
() => [
styles.optionLabel,
{ color: isSelected ? theme.colors.foreground : theme.colors.foregroundMuted },
],
[isSelected, theme.colors.foreground, theme.colors.foregroundMuted],
);
const optionDescriptionStyle = useMemo(
() => [styles.optionDescription, { color: theme.colors.foregroundMuted }],
@@ -96,11 +99,11 @@ function QuestionOptionRow({
styles.selectionControl,
multiSelect ? styles.selectionControlCheckbox : styles.selectionControlRadio,
{
borderColor: isSelected ? theme.colors.accent : theme.colors.foregroundMuted,
borderColor: isSelected ? theme.colors.accent : theme.colors.foregroundExtraMuted,
backgroundColor: isSelected && multiSelect ? theme.colors.accent : "transparent",
},
],
[isSelected, multiSelect, theme.colors.accent, theme.colors.foregroundMuted],
[isSelected, multiSelect, theme.colors.accent, theme.colors.foregroundExtraMuted],
);
const radioDotStyle = useMemo(
() => [styles.selectionRadioDot, { backgroundColor: theme.colors.accent }],
@@ -628,7 +631,7 @@ const styles = StyleSheet.create((theme) => ({
questionText: {
flex: 1,
fontSize: theme.fontSize.base,
fontWeight: theme.fontWeight.medium,
fontWeight: theme.fontWeight.normal,
lineHeight: 22,
},
optionsWrap: {
@@ -653,7 +656,7 @@ const styles = StyleSheet.create((theme) => ({
},
questionNavText: {
fontSize: theme.fontSize.base,
fontWeight: theme.fontWeight.medium,
fontWeight: theme.fontWeight.normal,
},
optionItem: {
flexDirection: "row",
@@ -677,7 +680,7 @@ const styles = StyleSheet.create((theme) => ({
},
optionLabel: {
fontSize: theme.fontSize.base,
fontWeight: theme.fontWeight.semibold,
fontWeight: theme.fontWeight.normal,
lineHeight: 22,
},
optionDescription: {
@@ -0,0 +1,35 @@
import type { AgentSessionConfig } from "@getpaseo/protocol/agent-types";
import type { AgentSnapshotPayload, CreateAgentRequestMessage } from "@getpaseo/protocol/messages";
import type { DaemonClient } from "@getpaseo/client/internal/daemon-client";
import { encodeImages } from "@/utils/encode-images";
import type { UserMessageImageAttachment } from "@/types/stream";
export interface WorkspaceDraftAgentRequest {
workspaceId: string;
config: AgentSessionConfig;
text: string;
clientMessageId: string;
images?: UserMessageImageAttachment[];
attachments?: CreateAgentRequestMessage["attachments"];
}
/**
* Shared by the workspace draft tab and by the new-workspace screen when it finishes creation
* after the user has already navigated away and no draft tab will ever mount.
*/
export async function requestWorkspaceDraftAgent(
client: DaemonClient,
request: WorkspaceDraftAgentRequest,
): Promise<AgentSnapshotPayload> {
const images = await encodeImages(request.images);
return await client.createAgent({
config: request.config,
workspaceId: request.workspaceId,
clientMessageId: request.clientMessageId,
...(request.text ? { initialPrompt: request.text } : {}),
...(images && images.length > 0 ? { images } : {}),
...(request.attachments && request.attachments.length > 0
? { attachments: request.attachments }
: {}),
});
}
@@ -24,7 +24,7 @@ import type { Agent } from "@/stores/session-store";
import { useWorkspaceFields } from "@/stores/session-store-hooks";
import { useWorkspaceDraftSubmissionStore } from "@/stores/workspace-draft-submission-store";
import { useAgentControlCommandCenterActions } from "@/command-center/agent-control-registration";
import { encodeImages } from "@/utils/encode-images";
import { requestWorkspaceDraftAgent } from "@/composer/draft/create-agent-request";
import type { WorkspaceFileOpenRequest } from "@/workspace/file-open";
import { shouldAutoFocusWorkspaceDraftComposer } from "@/screens/workspace/workspace-draft-pane-focus";
import {
@@ -194,15 +194,14 @@ async function submitDraftCreateRequest(input: {
featureValues: autoSubmitConfig?.featureValues ?? composerState.featureValues,
});
const imagesData = await encodeImages(images);
const attachmentsArray = Array.isArray(attachments) ? attachments : undefined;
const result = await client.createAgent({
const result = await requestWorkspaceDraftAgent(client, {
config,
workspaceId,
...(text ? { initialPrompt: text } : {}),
text,
clientMessageId: attempt.clientMessageId,
...(imagesData && imagesData.length > 0 ? { images: imagesData } : {}),
...(attachmentsArray && attachmentsArray.length > 0 ? { attachments: attachmentsArray } : {}),
...(images ? { images } : {}),
...(attachmentsArray ? { attachments: attachmentsArray } : {}),
});
return {
+108 -12
View File
@@ -85,10 +85,15 @@ import { useProjectIcons } from "@/projects/icons";
import { ICON_SIZE, type Theme } from "@/styles/theme";
import type { ComposerAttachment } from "@/attachments/types";
import { useDraftWorkspaceAttachmentScopeKey } from "@/attachments/workspace-attachments-store";
import { requestWorkspaceDraftAgent } from "@/composer/draft/create-agent-request";
import { buildWorkspaceDraftAgentConfig } from "@/screens/workspace/workspace-draft-agent-config";
import type { MessagePayload } from "@/composer/types";
import type { UserComposerAttachment } from "@/attachments/types";
import type { AgentAttachment, ForgeSearchItem } from "@getpaseo/protocol/messages";
import type { CreatePaseoWorktreeInput } from "@getpaseo/client/internal/daemon-client";
import type {
CreatePaseoWorktreeInput,
DaemonClient,
} from "@getpaseo/client/internal/daemon-client";
import type { AgentProvider } from "@getpaseo/protocol/agent-types";
import type { WorkspaceDraftTabSetup, WorkspaceTabTarget } from "@/workspace-tabs/model";
import { isEmptyWorkspaceSubmission, runCreateEmptyWorkspace } from "./new-workspace-empty";
@@ -123,6 +128,11 @@ import {
type ListTerminalsPayload,
upsertCreatedTerminalPayload,
} from "./workspace/terminals/state";
import {
captureWorkspaceDraftCleanup,
createWorkspaceAgentInBackground,
} from "./new-workspace/background-handoff";
import { useNewWorkspaceScreenPresence } from "./new-workspace/screen-presence";
const ThemedFolderPlus = withUnistyles(FolderPlus);
const foregroundMutedColorMapping = (theme: Theme) => ({ color: theme.colors.foregroundMuted });
@@ -747,10 +757,19 @@ function normalizeBranchDetails(
return names.map((name) => ({ name, committerDate: 0 }));
}
/**
* "background" means the user left the New workspace screen mid-creation, so nothing navigated
* and the screen — if still mounted under another route — has to drop its pending state itself.
*/
type SubmitOutcome = "navigated" | "background";
interface SubmitDraftInput {
clearConsumedDraft: () => void;
serverId: string;
draftKey: string;
clearDraft: (lifecycle: "sent" | "abandoned") => void;
draftId?: string;
draftContextScopeKey: string | null;
initialSetup?: WorkspaceDraftTabSetup;
workspaceId: string;
workspaceDirectory: string;
@@ -759,6 +778,8 @@ interface SubmitDraftInput {
provider: AgentProvider;
composerState: NewWorkspaceComposerState;
supportsForgeSearch: boolean;
resolveClient: () => DaemonClient;
isStillOnCreateScreen: () => boolean;
}
type NewWorkspaceComposerState = NonNullable<
@@ -860,9 +881,13 @@ interface CreateChatAgentInput {
withInitialAgent: boolean;
}) => Promise<ReturnType<typeof normalizeWorkspaceDescriptor>>;
serverId: string;
draftKey: string;
clearDraft: (lifecycle: "sent" | "abandoned") => void;
draftId?: string;
draftContextScopeKey: string | null;
supportsForgeSearch: boolean;
resolveClient: () => DaemonClient;
isStillOnCreateScreen: () => boolean;
labels: {
composerStateRequired: string;
selectModel: string;
@@ -918,8 +943,9 @@ function buildComposerInitialValues(input: {
return undefined;
}
async function runCreateChatAgent(input: CreateChatAgentInput): Promise<void> {
async function runCreateChatAgent(input: CreateChatAgentInput): Promise<SubmitOutcome> {
const { payload, composerState, ensureWorkspace, serverId, clearDraft } = input;
const clearConsumedDraft = captureWorkspaceDraftCleanup(input);
const { text, attachments, cwd } = payload;
if (!composerState) {
throw new Error(input.labels.composerStateRequired);
@@ -947,10 +973,13 @@ async function runCreateChatAgent(input: CreateChatAgentInput): Promise<void> {
provider,
composerState,
});
submitWorkspaceDraft({
return await submitWorkspaceDraft({
clearConsumedDraft,
serverId,
clearDraft,
draftKey: input.draftKey,
draftId: input.draftId,
draftContextScopeKey: input.draftContextScopeKey,
initialSetup,
workspaceId: ensuredWorkspace.id,
workspaceDirectory: ensuredWorkspace.workspaceDirectory,
@@ -959,6 +988,8 @@ async function runCreateChatAgent(input: CreateChatAgentInput): Promise<void> {
provider,
composerState,
supportsForgeSearch: input.supportsForgeSearch,
resolveClient: input.resolveClient,
isStillOnCreateScreen: input.isStillOnCreateScreen,
});
}
@@ -1021,7 +1052,7 @@ function resolveWorkspaceDraftSubmissionConfig(input: {
};
}
function submitWorkspaceDraft(input: SubmitDraftInput): void {
async function submitWorkspaceDraft(input: SubmitDraftInput): Promise<SubmitOutcome> {
const {
serverId,
clearDraft,
@@ -1049,6 +1080,33 @@ function submitWorkspaceDraft(input: SubmitDraftInput): void {
composerState,
initialSetup,
});
// Creation blocks on a slow daemon RPC. If the user moved on while it ran, the destination
// screen's draft tab will never mount to issue create_agent, so this path does it instead.
if (!input.isStillOnCreateScreen()) {
await createWorkspaceAgentInBackground({
clearConsumedDraft: input.clearConsumedDraft,
createAgent: () =>
requestWorkspaceDraftAgent(input.resolveClient(), {
workspaceId,
config: buildWorkspaceDraftAgentConfig({
provider: submission.provider,
cwd: submission.cwd,
...(submission.modeId ? { modeId: submission.modeId } : {}),
...(submission.model ? { model: submission.model } : {}),
...(submission.thinkingOptionId
? { thinkingOptionId: submission.thinkingOptionId }
: {}),
...(submission.featureValues ? { featureValues: submission.featureValues } : {}),
}),
text: text.trim(),
clientMessageId,
...(wirePayload.images.length > 0 ? { images: wirePayload.images } : {}),
...(wirePayload.attachments.length > 0 ? { attachments: wirePayload.attachments } : {}),
}),
});
return "background";
}
useCreateFlowStore.getState().setPending({
serverId,
draftId,
@@ -1082,6 +1140,7 @@ function submitWorkspaceDraft(input: SubmitDraftInput): void {
workspaceId,
target: submission.target,
});
return "navigated";
}
function useNewWorkspaceHostSelector(input: {
@@ -1591,6 +1650,7 @@ export function NewWorkspaceScreen({
const isolationPickerAnchorRef = useRef<View>(null);
const hostPickerAnchorRef = useRef<View | null>(null);
const isDraftHandoffActive = useIsNewWorkspaceDraftHandoffActive({ draftId, selectedServerId });
const isStillOnCreateScreen = useNewWorkspaceScreenPresence();
// Launch target: what the composer submits to (chat agent, or a terminal
// profile). Mirrors useWorkspaceIsolation's pattern below: the derived
@@ -1696,11 +1756,12 @@ export function NewWorkspaceScreen({
}, []);
const withConnectedClient = useCallback(() => {
if (!client || !isConnected) {
const connectedClient = getHostRuntimeStore().getClient(selectedServerId);
if (!connectedClient?.isConnected) {
throw new Error(t("newWorkspace.errors.hostDisconnected"));
}
return client;
}, [client, isConnected, t]);
return connectedClient;
}, [selectedServerId, t]);
const clientReady = isConnected && Boolean(client);
const hasSelectedSourceDirectory = selectedSourceDirectory !== null;
@@ -2043,31 +2104,49 @@ export function NewWorkspaceScreen({
await updateFormPreferences({ launchTarget });
if (isEmptyWorkspaceSubmission(payload)) {
setPendingAction("empty");
let outcome: SubmitOutcome = "background";
await runCreateEmptyWorkspace({
payload,
ensureWorkspace,
serverId: selectedServerId,
navigate: (targetServerId, workspaceId) =>
navigateToWorkspace({ serverId: targetServerId, workspaceId }),
navigate: (targetServerId, workspaceId) => {
if (!isStillOnCreateScreen()) {
return;
}
outcome = "navigated";
navigateToWorkspace({ serverId: targetServerId, workspaceId });
},
});
// Nothing navigated, so this screen may still be mounted under another route. Release
// the pending lock it would otherwise keep forever.
if (outcome === "background") {
setPendingAction(null);
}
return;
}
setPendingAction("chat");
await runCreateChatAgent({
const outcome = await runCreateChatAgent({
payload,
composerState,
forkDraftSetup,
ensureWorkspace,
serverId: selectedServerId,
clearDraft: chatDraft.clear,
draftKey,
draftId,
draftContextScopeKey,
supportsForgeSearch,
resolveClient: withConnectedClient,
isStillOnCreateScreen,
labels: {
composerStateRequired: t("newWorkspace.errors.composerStateRequired"),
selectModel: t("newWorkspace.errors.selectModel"),
},
});
if (outcome === "background") {
setPendingAction(null);
}
} catch (error) {
const message = toErrorMessage(error);
setPendingAction(null);
@@ -2077,16 +2156,20 @@ export function NewWorkspaceScreen({
},
[
composerState,
draftContextScopeKey,
draftId,
chatDraft.clear,
draftKey,
ensureWorkspace,
forkDraftSetup,
isStillOnCreateScreen,
launchTarget,
selectedServerId,
supportsForgeSearch,
t,
toast,
updateFormPreferences,
withConnectedClient,
],
);
@@ -2095,6 +2178,7 @@ export function NewWorkspaceScreen({
setErrorMessage(null);
await updateFormPreferences({ launchTarget });
setPendingAction("terminal");
let outcome: SubmitOutcome = "background";
await runCreateTerminalWorkspace({
cwd: selectedSourceDirectory ?? "",
prompt: terminalPromptText,
@@ -2128,9 +2212,20 @@ export function NewWorkspaceScreen({
withConnectedClient().sendTerminalInput(terminalId, { type: "input", data });
},
serverId: selectedServerId,
navigate: (targetServerId, workspaceId, target) =>
navigateToWorkspace({ serverId: targetServerId, workspaceId, target }),
// The terminal is spawned and fed its command before this runs, so skipping the
// navigation costs nothing: it is standalone, and `reconcileTabs` auto-opens standalone
// terminals when the workspace is next visited.
navigate: (targetServerId, workspaceId, target) => {
if (!isStillOnCreateScreen()) {
return;
}
outcome = "navigated";
navigateToWorkspace({ serverId: targetServerId, workspaceId, target });
},
});
if (outcome === "background") {
setPendingAction(null);
}
} catch (error) {
const message = toErrorMessage(error);
setPendingAction(null);
@@ -2139,6 +2234,7 @@ export function NewWorkspaceScreen({
}
}, [
ensureWorkspace,
isStillOnCreateScreen,
launchTarget,
queryClient,
selectedServerId,
@@ -0,0 +1,176 @@
import { beforeEach, describe, expect, it } from "vitest";
import { useCreateFlowStore } from "@/stores/create-flow-store";
import { useDraftStore } from "@/stores/draft-store";
import { useWorkspaceAttachmentsStore } from "@/attachments/workspace-attachments-store";
import { useWorkspaceDraftSubmissionStore } from "@/stores/workspace-draft-submission-store";
import {
captureWorkspaceDraftCleanup,
createWorkspaceAgentInBackground,
} from "./background-handoff";
const DRAFT_ID = "draft_1";
const DRAFT_KEY = "new-workspace";
const SCOPE_KEY = "draft:draft_1";
function saveDraft(text: string) {
useDraftStore
.getState()
.saveDraftInput({ draftKey: DRAFT_KEY, draft: { text, attachments: [] } });
}
function draftText(): string | undefined {
return useDraftStore.getState().drafts[DRAFT_KEY]?.input.text;
}
function captureCleanup() {
return captureWorkspaceDraftCleanup({
draftId: DRAFT_ID,
draftKey: DRAFT_KEY,
clearDraft: (lifecycle) =>
useDraftStore.getState().clearDraftInput({ draftKey: DRAFT_KEY, lifecycle }),
draftContextScopeKey: SCOPE_KEY,
});
}
function run(createAgent: () => Promise<unknown>, clearConsumedDraft = captureCleanup()) {
return createWorkspaceAgentInBackground({ createAgent, clearConsumedDraft });
}
beforeEach(() => {
useWorkspaceDraftSubmissionStore.setState({ pendingByDraftId: {}, setupByDraftId: {} });
useCreateFlowStore.setState({ pendingByDraftId: {} });
useWorkspaceAttachmentsStore.setState({ attachmentsByScope: { [SCOPE_KEY]: [] } });
saveDraft("do the thing");
});
describe("createWorkspaceAgentInBackground", () => {
it("preserves context added to a newer draft without editing its text", async () => {
await run(async () => {
useWorkspaceAttachmentsStore.getState().addWorkspaceAttachment({
scopeKey: SCOPE_KEY,
attachment: {
kind: "chat_history",
id: "fork-context",
attachment: {
type: "text",
mimeType: "text/plain",
contextKind: "chat_history",
title: "Chat history",
text: "Preserve this context",
},
source: { serverId: "local", agentId: "agent-1" },
},
});
useWorkspaceDraftSubmissionStore.getState().setDraftSetup({
draftId: DRAFT_ID,
setup: {
cwd: "/work/new-repo",
provider: "codex",
modeId: null,
model: null,
thinkingOptionId: null,
featureValues: {},
},
});
});
expect(useWorkspaceAttachmentsStore.getState().attachmentsByScope[SCOPE_KEY]).toHaveLength(1);
expect(useWorkspaceDraftSubmissionStore.getState().setupByDraftId[DRAFT_ID]?.setup.cwd).toBe(
"/work/new-repo",
);
});
it("creates one agent and clears the consumed draft setup, attachments and input", async () => {
let calls = 0;
useWorkspaceDraftSubmissionStore.getState().setDraftSetup({
draftId: DRAFT_ID,
setup: {
cwd: "/work/repo",
provider: "claude",
modeId: null,
model: null,
thinkingOptionId: null,
featureValues: {},
},
});
await run(async () => {
calls += 1;
});
expect(calls).toBe(1);
expect(useWorkspaceDraftSubmissionStore.getState().setupByDraftId).toEqual({});
expect(useWorkspaceAttachmentsStore.getState().attachmentsByScope[SCOPE_KEY]).toBeUndefined();
expect(draftText()).toBe("");
});
it("preserves context changed while workspace creation was pending", async () => {
const clearConsumedDraft = captureCleanup();
useWorkspaceAttachmentsStore.getState().addWorkspaceAttachment({
scopeKey: SCOPE_KEY,
attachment: {
kind: "chat_history",
id: "fork-context",
attachment: {
type: "text",
mimeType: "text/plain",
contextKind: "chat_history",
title: "Chat history",
text: "Preserve this context",
},
source: { serverId: "local", agentId: "agent-1" },
},
});
await run(async () => undefined, clearConsumedDraft);
expect(useWorkspaceAttachmentsStore.getState().attachmentsByScope[SCOPE_KEY]).toHaveLength(1);
expect(draftText()).toBe("do the thing");
});
it("leaves no pending submission behind, so no draft tab can create a second agent", async () => {
await run(async () => undefined);
expect(useWorkspaceDraftSubmissionStore.getState().pendingByDraftId).toEqual({});
expect(useCreateFlowStore.getState().pendingByDraftId).toEqual({});
});
it("rethrows a failed creation and keeps the prompt for retry", async () => {
const failure = new Error("daemon exploded");
await expect(
run(async () => {
throw failure;
}),
).rejects.toThrow(failure);
expect(draftText()).toBe("do the thing");
});
it("keeps a prompt typed into the shared draft while creation was in flight", async () => {
await run(async () => {
saveDraft("a different idea");
});
expect(draftText()).toBe("a different idea");
});
it("keeps a newer draft written while workspace creation was pending", async () => {
useWorkspaceAttachmentsStore.getState().addWorkspaceAttachment({
scopeKey: SCOPE_KEY,
attachment: {
kind: "chat_history",
id: "fork-context",
attachment: {
type: "text",
mimeType: "text/plain",
contextKind: "chat_history",
title: "Chat history",
text: "Preserve this context",
},
source: { serverId: "local", agentId: "agent-1" },
},
});
const clearConsumedDraft = captureCleanup();
saveDraft("a newer workspace idea");
await run(async () => undefined, clearConsumedDraft);
expect(draftText()).toBe("a newer workspace idea");
expect(useWorkspaceAttachmentsStore.getState().attachmentsByScope[SCOPE_KEY]).toHaveLength(1);
});
});
@@ -0,0 +1,54 @@
import { useWorkspaceAttachmentsStore } from "@/attachments/workspace-attachments-store";
import { useDraftStore } from "@/stores/draft-store";
import { useWorkspaceDraftSubmissionStore } from "@/stores/workspace-draft-submission-store";
interface WorkspaceDraftCleanupInput {
draftId?: string;
draftKey: string;
clearDraft: (lifecycle: "sent" | "abandoned") => void;
draftContextScopeKey: string | null;
}
/** Capture ownership before workspace creation starts, not when its response arrives. */
export function captureWorkspaceDraftCleanup(input: WorkspaceDraftCleanupInput): () => void {
const draftId = input.draftId?.trim() ?? "";
const scopeKey = input.draftContextScopeKey ?? "";
const draftVersion = useDraftStore.getState().drafts[input.draftKey]?.version;
const setup = useWorkspaceDraftSubmissionStore.getState().setupByDraftId[draftId];
const attachments = useWorkspaceAttachmentsStore.getState().attachmentsByScope[scopeKey];
return () => {
// These stores update independently. A context-only edit does not bump the text version.
// Any newer draft state owns the whole draft, including its unchanged context.
if (
useDraftStore.getState().drafts[input.draftKey]?.version !== draftVersion ||
useWorkspaceDraftSubmissionStore.getState().setupByDraftId[draftId] !== setup ||
useWorkspaceAttachmentsStore.getState().attachmentsByScope[scopeKey] !== attachments
) {
return;
}
useWorkspaceDraftSubmissionStore.getState().clearDraftSetup({ draftId });
if (scopeKey) {
useWorkspaceAttachmentsStore.getState().clearWorkspaceAttachments({ scopeKey });
}
input.clearDraft("sent");
};
}
/**
* The workspace draft tab normally issues create_agent once it mounts on the destination screen.
* When the user leaves the New workspace screen before creation resolves, that tab never mounts,
* so this path issues the request itself. The agent tab opens on its own when the workspace is
* next visited — `reconcileTabs` auto-opens agents.
*
* The caller must not write a pending entry to the draft-submission or create-flow stores on this
* path: those drive the draft tab's auto-submit, and the daemon does not dedupe create_agent by
* clientMessageId, so a leftover entry would create a second agent when the workspace is opened.
*/
export async function createWorkspaceAgentInBackground(input: {
clearConsumedDraft: () => void;
createAgent: () => Promise<unknown>;
}): Promise<void> {
await input.createAgent();
input.clearConsumedDraft();
}
@@ -0,0 +1,22 @@
import { describe, expect, it } from "vitest";
import { isNewWorkspaceScreenActive } from "./screen-presence";
describe("isNewWorkspaceScreenActive", () => {
it("is active while the screen is mounted and the route is /new", () => {
expect(isNewWorkspaceScreenActive({ isMounted: true, pathname: "/new" })).toBe(true);
});
it("is inactive once the screen has been popped off the stack", () => {
expect(isNewWorkspaceScreenActive({ isMounted: false, pathname: "/new" })).toBe(false);
});
it("is inactive while another route sits on top of the screen", () => {
expect(isNewWorkspaceScreenActive({ isMounted: true, pathname: "/settings" })).toBe(false);
});
it("is inactive on a workspace route", () => {
expect(isNewWorkspaceScreenActive({ isMounted: true, pathname: "/host/workspace/wks_1" })).toBe(
false,
);
});
});
@@ -0,0 +1,41 @@
import { usePathname } from "expo-router";
import { useCallback, useEffect, useRef } from "react";
const NEW_WORKSPACE_PATHNAME = "/new";
export interface NewWorkspaceScreenPresence {
isMounted: boolean;
pathname: string;
}
/**
* Workspace creation blocks on a slow daemon RPC, so by the time it resolves the user may have
* moved on. Both signals are needed: popping to a workspace unmounts the screen while its last
* observed pathname stays "/new", and pushing a route on top of it keeps it mounted underneath.
*/
export function isNewWorkspaceScreenActive(input: NewWorkspaceScreenPresence): boolean {
return input.isMounted && input.pathname === NEW_WORKSPACE_PATHNAME;
}
export function useNewWorkspaceScreenPresence(): () => boolean {
const pathname = usePathname();
const pathnameRef = useRef(pathname);
pathnameRef.current = pathname;
const isMountedRef = useRef(true);
useEffect(() => {
isMountedRef.current = true;
return () => {
isMountedRef.current = false;
};
}, []);
return useCallback(
() =>
isNewWorkspaceScreenActive({
isMounted: isMountedRef.current,
pathname: pathnameRef.current,
}),
[],
);
}
@@ -2885,6 +2885,9 @@ export class AgentManager {
response: AgentPermissionResponse,
): Promise<AgentPermissionResult | void> {
const agent = this.requireAgent(agentId);
if (agent.inFlightPermissionResponses.has(requestId)) {
throw new Error("A response to this permission request is already being submitted");
}
agent.inFlightPermissionResponses.add(requestId);
try {
@@ -3079,8 +3082,11 @@ export class AgentManager {
epoch: this.timelineStore.getEpoch(agentId),
});
}
await this.refreshRuntimeInfo(agent);
// Rewind stages provider events under the run lock; publish its final state directly.
this.refreshSessionPersistence(agent);
await this.refreshSessionState(agent, { emit: false });
await this.persistSnapshot(agent);
this.emitState(agent, { persist: false });
this.logger.info(
{ agentId, provider: agent.provider, messageId, mode },
"agent.rewind.complete",
@@ -4231,14 +4237,18 @@ export class AgentManager {
private onStreamThreadStarted(agent: ActiveManagedAgent): void {
const previousSessionId = agent.persistence?.sessionId ?? null;
this.refreshSessionPersistence(agent);
if (agent.persistence?.sessionId !== previousSessionId) {
this.emitState(agent);
}
void this.refreshRuntimeInfo(agent);
}
private refreshSessionPersistence(agent: ActiveManagedAgent): void {
const handle = agent.session.describePersistence();
if (handle) {
agent.persistence = attachPersistenceCwd(handle, agent.cwd);
if (agent.persistence?.sessionId !== previousSessionId) {
this.emitState(agent);
}
}
void this.refreshRuntimeInfo(agent);
}
private async onStreamTimelineEvent(params: {
@@ -4443,6 +4453,7 @@ export class AgentManager {
): void {
const hadPendingPermissions = agent.pendingPermissions.size > 0;
agent.pendingPermissions.set(event.request.id, event.request);
this.refreshSessionPersistence(agent);
if (!hadPendingPermissions && !agent.internal) {
this.broadcastAgentAttention(agent, "permission");
}
@@ -4457,6 +4468,7 @@ export class AgentManager {
}): void {
const { agent, event, options, flags } = params;
agent.pendingPermissions.delete(event.requestId);
this.refreshSessionPersistence(agent);
if (!options?.fromHistory && agent.inFlightPermissionResponses.has(event.requestId)) {
agent.bufferedPermissionResolutions.set(event.requestId, event);
flags.shouldDispatchEvent = false;
@@ -24,6 +24,12 @@ interface TestClaudeSession {
close(): Promise<void>;
}
function isLoadingCompactionEvent(event: AgentStreamEvent): boolean {
return (
event.type === "timeline" && event.item.type === "compaction" && event.item.status === "loading"
);
}
function isPermissionResolvedEvent(
event: AgentStreamEvent,
): event is Extract<AgentStreamEvent, { type: "permission_resolved" }> {
@@ -1981,6 +1987,15 @@ describe("ClaudeAgentSession context window usage", () => {
};
}
function createCompactingStatus(): Record<string, unknown> {
return {
type: "system",
subtype: "status",
status: "compacting",
session_id: "session-1",
};
}
test("emits turn_started before the submitted user message", async () => {
const session = await createSessionForTurns([[]]);
const events: AgentStreamEvent[] = [];
@@ -2857,6 +2872,110 @@ describe("ClaudeAgentSession context window usage", () => {
}
});
test("repeated compacting statuses open a single compaction marker", async () => {
const session = await createSessionForTurns([
[
createCompactingStatus(),
createCompactingStatus(),
createCompactingStatus(),
createCompactBoundary(),
createCompactingStatus(),
createCompactingStatus(),
createCompactBoundary(),
createSuccessResult(),
],
]);
try {
const events = await collectStreamEvents(session, "compact twice");
const compactions = events.flatMap((event) =>
event.type === "timeline" && event.item.type === "compaction" ? [event.item.status] : [],
);
expect(compactions).toEqual(["loading", "completed", "loading", "completed"]);
} finally {
await session.close();
}
});
test("a compaction abandoned mid-turn does not suppress the next compaction marker", async () => {
// The first turn starts compacting and then ends without ever reaching a
// compact_boundary, so the marker it opened is never resolved.
const session = await createSessionForTurns([
[createCompactingStatus(), createSuccessResult()],
[createCompactingStatus(), createSuccessResult()],
]);
try {
const abandonedTurn = await collectStreamEvents(session, "abandoned compaction");
expect(abandonedTurn.filter(isLoadingCompactionEvent)).toHaveLength(1);
const nextTurn = await collectStreamEvents(session, "next compaction");
expect(nextTurn.filter(isLoadingCompactionEvent)).toHaveLength(1);
} finally {
await session.close();
}
});
test("an interrupted compaction does not suppress the next compaction marker", async () => {
// The first turn starts compacting and is then interrupted, so it never reaches a
// compact_boundary and the marker it opened is never resolved.
const session = await createSessionForTurns([
[createCompactingStatus()],
[createCompactingStatus(), createSuccessResult()],
]);
try {
const interruptedTurn: AgentStreamEvent[] = [];
const streaming = (async () => {
for await (const event of streamSession(session, "interrupted compaction")) {
interruptedTurn.push(event);
}
})();
await vi.waitFor(() => {
expect(interruptedTurn.filter(isLoadingCompactionEvent)).toHaveLength(1);
});
await session.interrupt();
await streaming;
expect(interruptedTurn).toContainEqual(
expect.objectContaining({ type: "turn_canceled", provider: "claude" }),
);
const nextTurn = await collectStreamEvents(session, "next compaction");
expect(nextTurn.filter(isLoadingCompactionEvent)).toHaveLength(1);
} finally {
await session.close();
}
});
test("a compaction abandoned in an autonomous turn does not suppress the next marker", async () => {
// Trailing output after the foreground result opens an autonomous turn, which starts
// compacting and is then ended by the next foreground turn, never reaching a boundary.
const session = await createSessionForTurns([
[createSuccessResult(), createMessageStartEvent(), createCompactingStatus()],
[createCompactingStatus(), createSuccessResult()],
]);
try {
const observed: AgentStreamEvent[] = [];
const unsubscribe = session.subscribe((event) => {
observed.push(event);
});
await collectStreamEvents(session, "foreground turn");
await vi.waitFor(() => {
expect(observed.filter(isLoadingCompactionEvent)).toHaveLength(1);
});
unsubscribe();
const nextTurn = await collectStreamEvents(session, "next compaction");
expect(nextTurn.filter(isLoadingCompactionEvent)).toHaveLength(1);
} finally {
await session.close();
}
});
test("result.result is surfaced as an assistant message when no model output was produced", async () => {
const session = await createSessionForTest();
@@ -2095,6 +2095,7 @@ class ClaudeAgentSession implements AgentSession {
private lastOptionsModel: string | null = null;
private lastRuntimeModel: string | null = null;
private compacting = false;
private compactionMarkerOpen = false;
private queryPumpPromise: Promise<void> | null = null;
private queryRestartNeeded = false;
private pendingInterruptAbort = false;
@@ -3572,6 +3573,7 @@ class ClaudeAgentSession implements AgentSession {
this.activeForegroundInput = null;
this.cancelCurrentTurn = null;
this.activeTurnHasAssistantText = false;
this.compactionMarkerOpen = false;
this.syncTurnState("foreground turn terminal");
}
@@ -3583,6 +3585,7 @@ class ClaudeAgentSession implements AgentSession {
}
if (terminalSeen) {
this.compactionMarkerOpen = false;
if (this.activeForegroundTurnId) {
this.activeForegroundTurnId = null;
this.activeForegroundQuery = null;
@@ -3624,6 +3627,7 @@ class ClaudeAgentSession implements AgentSession {
this.activeForegroundQuery = null;
this.activeForegroundInput = null;
this.activeTurnHasAssistantText = false;
this.compactionMarkerOpen = false;
this.syncTurnState("autonomous turn completed");
}
@@ -4260,15 +4264,22 @@ class ClaudeAgentSession implements AgentSession {
const status = toObjectRecord(message)?.status;
if (status === "compacting") {
this.compacting = true;
events.push({
type: "timeline",
item: { type: "compaction", status: "loading" },
provider: "claude",
});
// Claude Code repeats this status every 30 seconds until the compaction
// finishes. Each repeat used to open its own marker, and the app only ever
// resolves one of them, so the rest stayed on "Compacting..." forever.
if (!this.compactionMarkerOpen) {
this.compactionMarkerOpen = true;
events.push({
type: "timeline",
item: { type: "compaction", status: "loading" },
provider: "claude",
});
}
}
return;
}
if (message.subtype === "compact_boundary") {
this.compactionMarkerOpen = false;
const compactMetadata = readCompactionMetadata(message);
events.push({
type: "timeline",
@@ -52,6 +52,7 @@ import { z } from "zod";
import { renderPromptAttachmentAsText } from "../prompt-attachments.js";
import { composeSystemPromptParts } from "../system-prompt.js";
import { curateAgentActivity } from "../activity-curator.js";
import { CodexAsyncQuestions, codexAsyncQuestionToTimeline } from "./codex/async-questions.js";
import {
mapCodexToolCallEnvelope,
mapCodexToolCallFromThreadItem,
@@ -1845,6 +1846,17 @@ function mapCodexThreadImageItem(
);
}
function mapCodexAgentMessage(item: Record<string, unknown>): AgentTimelineItem {
const question = codexAsyncQuestionToTimeline(item);
if (question) return question;
const messageId = nonEmptyString(item.id);
return {
type: "assistant_message",
text: typeof item.text === "string" ? item.text : "",
...(messageId ? { messageId } : {}),
};
}
export function threadItemToTimeline(
item: unknown,
options?: { includeUserMessage?: boolean; cwd?: string | null },
@@ -1871,14 +1883,8 @@ export function threadItemToTimeline(
switch (normalizedType) {
case "userMessage":
return mapCodexThreadUserMessageItem(normalizedItem, includeUserMessage);
case "agentMessage": {
const messageId = nonEmptyString(normalizedItem.id);
return {
type: "assistant_message",
text: typeof normalizedItem.text === "string" ? normalizedItem.text : "",
...(messageId ? { messageId } : {}),
};
}
case "agentMessage":
return mapCodexAgentMessage(normalizedItem);
case "plan":
return mapCodexThreadPlanItem(normalizedItem);
case "reasoning":
@@ -3293,6 +3299,7 @@ export class CodexAppServerAgentSession implements AgentSession {
private readonly logger: Logger;
private readonly config: AgentSessionConfig;
private readonly asyncQuestions: CodexAsyncQuestions;
private currentMode: string;
private hasWorkflowModeOverride: boolean;
private readonly providerOptions: CodexProviderOptions;
@@ -3405,6 +3412,7 @@ export class CodexAppServerAgentSession implements AgentSession {
this.currentMode = config.modeId ?? DEFAULT_CODEX_MODE_ID;
this.providerOptions = CodexProviderOptionsSchema.parse(config.providerOptions ?? {});
this.config = config;
this.asyncQuestions = new CodexAsyncQuestions(resumeHandle?.metadata?.asyncQuestions);
this.config.thinkingOptionId = normalizeCodexThinkingOptionId(this.config.thinkingOptionId);
if (this.config.featureValues?.fast_mode && codexModelSupportsFastMode(this.config.model)) {
this.serviceTier = "fast";
@@ -3799,6 +3807,9 @@ export class CodexAppServerAgentSession implements AgentSession {
}
this.resetCodexUserMessageTurns();
for (const entry of timeline) {
if (entry.item.type === "tool_call" && entry.item.name === "request_user_input_async") {
entry.item = this.asyncQuestions.timeline(entry.item.callId) ?? entry.item;
}
if (entry.item.type === "user_message") {
this.rememberCodexUserMessageTurn(entry.item.messageId, entry.providerTurnId);
}
@@ -4464,13 +4475,16 @@ export class CodexAppServerAgentSession implements AgentSession {
}
getPendingPermissions(): AgentPermissionRequest[] {
return Array.from(this.pendingPermissions.values());
return [...this.pendingPermissions.values(), ...this.asyncQuestions.pending()];
}
async respondToPermission(
requestId: string,
response: AgentPermissionResponse,
): Promise<AgentPermissionResult | void> {
if (this.asyncQuestions.hasPending(requestId)) {
return this.respondToAsyncQuestion(requestId, response);
}
const pending = this.pendingPermissionHandlers.get(requestId);
if (!pending) {
throw new Error(`No pending Codex app-server permission request with id '${requestId}'`);
@@ -4564,6 +4578,33 @@ export class CodexAppServerAgentSession implements AgentSession {
pending.resolve({ answers: {} });
}
private async respondToAsyncQuestion(
requestId: string,
response: AgentPermissionResponse,
): Promise<AgentPermissionResult | void> {
const prepared = this.asyncQuestions.prepareResponse(requestId, response);
let followUpPrompt = prepared.prompt;
const expectedTurnId = this.activeForegroundTurnId;
if (prepared.prompt && expectedTurnId) {
const result = await this.steerActiveTurn(prepared.prompt, {
expectedTurnId,
clientMessageId: randomUUID(),
});
if (result.status !== "accepted") {
throw new Error("The active Codex turn changed. Retry sending your answer.");
}
followUpPrompt = undefined;
}
this.emitEvent({ type: "timeline", provider: CODEX_PROVIDER, item: prepared.complete() });
this.emitEvent({
type: "permission_resolved",
provider: CODEX_PROVIDER,
requestId,
resolution: response,
});
return followUpPrompt ? { followUpPrompt } : undefined;
}
private handlePlanPermissionResponse(params: {
requestId: string;
response: AgentPermissionResponse;
@@ -4700,6 +4741,7 @@ export class CodexAppServerAgentSession implements AgentSession {
toolPolicy: this.config.toolPolicy,
systemPrompt: this.config.systemPrompt,
mcpServers: this.config.mcpServers,
asyncQuestions: this.asyncQuestions.serialize(),
},
};
}
@@ -4729,6 +4771,7 @@ export class CodexAppServerAgentSession implements AgentSession {
this.persistedHistory = [];
this.historyPending = false;
await this.loadPersistedHistory();
this.reconcileAsyncQuestionsAfterRewind();
},
});
}
@@ -5903,6 +5946,7 @@ export class CodexAppServerAgentSession implements AgentSession {
error: parsed.errorMessage ?? "Codex turn failed",
});
} else if (parsed.status === "interrupted") {
this.dismissInterruptedAsyncQuestions();
this.emitEvent({ type: "turn_canceled", provider: CODEX_PROVIDER, reason: "interrupted" });
} else {
if (this.planModeEnabled && this.latestPlanResult?.text) {
@@ -6255,6 +6299,43 @@ export class CodexAppServerAgentSession implements AgentSession {
}
}
private receiveAsyncQuestion(threadId: string | null, item: unknown): void {
if (threadId !== this.currentThreadId) return;
const request = this.asyncQuestions.receive(item);
if (request)
this.emitEvent({ type: "permission_requested", provider: CODEX_PROVIDER, request });
}
private dismissInterruptedAsyncQuestions(): void {
const resolution: AgentPermissionResponse = { behavior: "deny", message: "Interrupted" };
for (const request of this.asyncQuestions.pending()) {
const prepared = this.asyncQuestions.prepareResponse(request.id, resolution);
this.emitEvent({ type: "timeline", provider: CODEX_PROVIDER, item: prepared.complete() });
this.emitEvent({
type: "permission_resolved",
provider: CODEX_PROVIDER,
requestId: request.id,
resolution,
});
}
}
private reconcileAsyncQuestionsAfterRewind(): void {
const retainedIds = new Set(
this.persistedHistory.flatMap(({ item }) =>
item.type === "tool_call" && item.name === "request_user_input_async" ? [item.callId] : [],
),
);
for (const requestId of this.asyncQuestions.retain(retainedIds)) {
this.emitEvent({
type: "permission_resolved",
provider: CODEX_PROVIDER,
requestId,
resolution: { behavior: "deny", message: "Removed by rewind" },
});
}
}
private handleItemCompletedNotification(
parsed: Extract<ParsedCodexNotification, { kind: "item_completed" }>,
): void {
@@ -6265,6 +6346,7 @@ export class CodexAppServerAgentSession implements AgentSession {
if (shouldIgnoreMirroredLifecycleItem(parsed.source, parsed.item)) {
return;
}
this.receiveAsyncQuestion(parsed.threadId, parsed.item);
if (this.isUserMessageItem(parsed.item)) {
this.handleUserMessageItem(parsed);
return;
@@ -0,0 +1,436 @@
import { tmpdir } from "node:os";
import { expect, test } from "vitest";
import { CodexAppServerAgentSession } from "./codex-app-server-agent.js";
import {
createFakeCodexAppServer,
waitForNextEvent,
waitForTimelineToolCall,
} from "./codex/test-utils/fake-app-server.js";
import { createTestLogger } from "../../../test-utils/test-logger.js";
import type { AgentStreamEvent } from "../agent-sdk-types.js";
import { AgentManager } from "../agent-manager.js";
const questionItem = {
type: "agentMessage",
id: "async-question-1",
text: "Which color?\n- Blue\n- Green",
phase: "final_answer",
delivery: "async",
questions: [{ title: "Which color?", options: ["Blue", "Green"] }],
};
async function setup(metadata?: Record<string, unknown>, rejectSteer = false) {
const appServer = createFakeCodexAppServer({
"turn/interrupt": () => ({}),
"turn/steer": () => {
if (rejectSteer) return { __jsonRpcError: { code: -32000, message: "Delivery failed" } };
return { turnId: "native-turn" };
},
"thread/read": () => ({ thread: { id: "thread-1", turns: [] } }),
});
const session = new CodexAppServerAgentSession(
{ provider: "codex", cwd: tmpdir(), model: "gpt-5.4", modeId: "full-access" },
metadata ? { sessionId: "thread-1", metadata } : null,
createTestLogger(),
async () => appServer.child,
);
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
await session.startTurn("Help me choose a color while you inspect the project.");
const started = waitForNextEvent(session, "turn_started");
appServer.startsTurn({ threadId: "thread-1", turnId: "native-turn" });
await started;
async function ask() {
const shown = waitForTimelineToolCall(session, questionItem.id);
appServer.child.stdout.write(
JSON.stringify({
method: "item/completed",
params: { threadId: "thread-1", turnId: "native-turn", item: questionItem },
}) + "\n",
);
await shown;
}
async function say(text: string) {
const shown = waitForNextEvent(
session,
"timeline",
(event) => event.item.type === "assistant_message" && event.item.text.includes(text),
);
appServer.says({ threadId: "thread-1", text });
await shown;
}
async function finish(status: "completed" | "interrupted" = "completed") {
const finished = waitForNextEvent(
session,
status === "completed" ? "turn_completed" : "turn_canceled",
);
appServer.completeTurn({ status });
await finished;
}
function allowAnswers() {
rejectSteer = false;
}
return { session, appServer, events, ask, say, finish, allowAnswers };
}
const answer = { behavior: "allow" as const, updatedInput: { answers: { "Question 1": "Green" } } };
async function setupRewind(fail = false) {
const records = [
{ item: { ...questionItem, id: "earlier-pending" } },
{ item: { ...questionItem, id: "earlier-answered" }, resolution: ["Blue"] },
{ item: { ...questionItem, id: "removed-pending" } },
{ item: { ...questionItem, id: "removed-answered" }, resolution: ["Green"] },
];
function userMessage(id: string) {
return { type: "userMessage", id, content: [{ type: "text", text: id }] };
}
const turns = [
{
id: "earlier-turn",
items: [userMessage("earlier"), ...records.slice(0, 2).map((r) => r.item)],
},
{
id: "removed-turn",
items: [userMessage("rewind-here"), ...records.slice(2).map((r) => r.item)],
},
];
const appServer = createFakeCodexAppServer({
"thread/read": (params) => {
const { threadId } = params as { threadId: string };
return {
thread: { id: threadId, turns: threadId === "thread-1" ? turns : turns.slice(0, 1) },
};
},
...(fail
? {
"thread/rollback": () => ({ __jsonRpcError: { code: -32000, message: "Rewind failed" } }),
}
: {}),
});
const session = new CodexAppServerAgentSession(
{ provider: "codex", cwd: tmpdir(), model: "gpt-5.4", modeId: "full-access" },
{ sessionId: "thread-1", metadata: { asyncQuestions: records } },
createTestLogger(),
async () => appServer.child,
);
const events: AgentStreamEvent[] = [];
session.subscribe((event) => events.push(event));
await session.connect();
return { session, events, records };
}
test("rewind removes only questions outside the remaining history, including after resume", async () => {
const { session, events, records } = await setupRewind();
let metadata: Record<string, unknown> | undefined;
try {
await session.revertConversation({ messageId: "rewind-here" });
expect(session.getPendingPermissions().map((p) => p.id)).toEqual([
"permission-earlier-pending",
]);
metadata = session.describePersistence()!.metadata;
expect(metadata.asyncQuestions).toEqual(
records.slice(0, 2).map((record) => ({
resolution: record.resolution,
item: {
type: "agentMessage",
id: record.item.id,
delivery: "async",
questions: questionItem.questions,
},
})),
);
expect(events.filter((e) => e.type === "permission_resolved")).toEqual([
expect.objectContaining({ requestId: "permission-removed-pending" }),
]);
const history = [];
for await (const event of session.streamHistory()) history.push(event);
expect(
history.some(
(e) =>
e.type === "timeline" &&
e.item.type === "tool_call" &&
e.item.callId === "removed-pending",
),
).toBe(false);
} finally {
await session.close();
}
const resumed = await setup(metadata);
try {
expect(resumed.session.getPendingPermissions().map((p) => p.id)).toEqual([
"permission-earlier-pending",
]);
} finally {
await resumed.session.close();
}
});
test("failed rewind preserves question state", async () => {
const { session, events } = await setupRewind(true);
try {
const before = session.describePersistence();
await expect(session.revertConversation({ messageId: "rewind-here" })).rejects.toThrow(
"Rewind failed",
);
expect(session.describePersistence()).toEqual(before);
expect(session.getPendingPermissions().map((p) => p.id)).toEqual([
"permission-earlier-pending",
"permission-removed-pending",
]);
expect(events.some((e) => e.type === "permission_resolved")).toBe(false);
} finally {
await session.close();
}
});
async function manage(session: CodexAppServerAgentSession) {
const manager = new AgentManager({
clients: {
codex: {
provider: "codex",
capabilities: session.capabilities,
createSession: async () => session,
resumeSession: async () => session,
isAvailable: async () => true,
fetchCatalog: async () => ({ models: [], modes: [] }),
},
},
logger: createTestLogger(),
});
const agent = await manager.createAgent(
{ provider: "codex", cwd: tmpdir(), model: "gpt-5.4" },
undefined,
{ workspaceId: undefined },
);
return { manager, agent };
}
test("manager publishes and saves the provider state after rewind", async () => {
const { session } = await setupRewind();
const { manager, agent } = await manage(session);
try {
await manager.rewind(agent.id, "rewind-here", "conversation");
const snapshot = manager.getAgent(agent.id)!;
expect(Array.from(snapshot.pendingPermissions.keys())).toEqual(["permission-earlier-pending"]);
expect(snapshot.persistence).toEqual(session.describePersistence());
expect(snapshot.persistence?.sessionId).toBe("forked-thread");
} finally {
await manager.closeAgent(agent.id);
}
});
test("manager snapshots capture pending and answered question state before the turn ends", async () => {
const { session, ask } = await setup();
const { manager, agent } = await manage(session);
try {
await ask();
const [permission] = session.getPendingPermissions();
expect(manager.getAgent(agent.id)?.persistence?.metadata?.asyncQuestions).toEqual([
{
item: {
type: "agentMessage",
id: questionItem.id,
delivery: "async",
questions: questionItem.questions,
},
},
]);
await manager.respondToPermission(agent.id, permission.id, answer);
expect(manager.getAgent(agent.id)?.persistence?.metadata?.asyncQuestions).toEqual([
{
item: {
type: "agentMessage",
id: questionItem.id,
delivery: "async",
questions: questionItem.questions,
},
resolution: ["Green"],
},
]);
} finally {
await manager.closeAgent(agent.id);
}
});
test("concurrent answers deliver only one response to the active Codex turn", async () => {
const { session, appServer, ask } = await setup();
const { manager, agent } = await manage(session);
try {
await ask();
const [permission] = session.getPendingPermissions();
const results = await Promise.allSettled([
manager.respondToPermission(agent.id, permission.id, answer),
manager.respondToPermission(agent.id, permission.id, answer),
]);
expect(results.map((result) => result.status)).toEqual(["fulfilled", "rejected"]);
expect(appServer.requests().filter((request) => request.method === "turn/steer")).toHaveLength(
1,
);
expect(session.getPendingPermissions()).toEqual([]);
} finally {
await manager.closeAgent(agent.id);
}
});
test("a failed submission releases the request so the user can retry", async () => {
const { session, ask, allowAnswers } = await setup(undefined, true);
const { manager, agent } = await manage(session);
try {
await ask();
const [permission] = session.getPendingPermissions();
await expect(manager.respondToPermission(agent.id, permission.id, answer)).rejects.toThrow(
"Delivery failed",
);
expect(session.getPendingPermissions().map((request) => request.id)).toEqual([permission.id]);
allowAnswers();
await manager.respondToPermission(agent.id, permission.id, answer);
expect(session.getPendingPermissions()).toEqual([]);
} finally {
await manager.closeAgent(agent.id);
}
});
test("shows an async question, keeps streaming, and delivers its answer without interrupting", async () => {
const { session, appServer, events, ask, say } = await setup();
try {
await ask();
const [permission] = session.getPendingPermissions();
expect(permission).toMatchObject({
kind: "question",
input: {
questions: [
{
header: "Question 1",
question: "Which color?",
isOther: true,
options: [{ label: "Blue" }, { label: "Green" }],
},
],
},
});
await say("I am still inspecting the project.");
expect(
events.some(
(event) =>
event.type === "timeline" &&
event.item.type === "assistant_message" &&
event.item.text.includes("still inspecting"),
),
).toBe(true);
await session.respondToPermission(permission.id, answer);
expect(session.getPendingPermissions()).toEqual([]);
expect(appServer.requests().filter((request) => request.method === "turn/steer")).toMatchObject(
[
{
params: {
expectedTurnId: "native-turn",
clientUserMessageId: expect.any(String),
input: [{ type: "text", text: expect.stringContaining("Which color?\nGreen") }],
},
},
],
);
expect(appServer.requests().filter((request) => request.method === "turn/interrupt")).toEqual(
[],
);
expect(
events.some(
(event) => event.type === "permission_resolved" && event.requestId === permission.id,
),
).toBe(true);
} finally {
await session.close();
}
});
test("keeps a failed answer pending for retry", async () => {
const { session, ask } = await setup(undefined, true);
try {
await ask();
const [permission] = session.getPendingPermissions();
expect(permission).toBeDefined();
await expect(session.respondToPermission(permission.id, answer)).rejects.toThrow();
expect(session.getPendingPermissions()).toHaveLength(1);
} finally {
await session.close();
}
});
test("Stop dismisses async questions before cancellation and keeps them dismissed after resume", async () => {
const first = await setup();
let metadata: Record<string, unknown> | undefined;
try {
await first.ask();
expect(first.session.getPendingPermissions()).toHaveLength(1);
first.session.subscribe((event) => {
if (event.type === "turn_canceled") {
metadata = first.session.describePersistence()!.metadata;
}
});
await first.session.interrupt();
await first.finish("interrupted");
expect(first.session.getPendingPermissions()).toEqual([]);
expect(metadata?.asyncQuestions).toEqual([
expect.objectContaining({ resolution: "dismissed" }),
]);
expect(first.events).toContainEqual(
expect.objectContaining({
type: "permission_resolved",
requestId: "permission-async-question-1",
resolution: { behavior: "deny", message: "Interrupted" },
}),
);
} finally {
await first.session.close();
}
const resumed = await setup(metadata);
try {
await resumed.ask();
expect(resumed.session.getPendingPermissions()).toEqual([]);
} finally {
await resumed.session.close();
}
});
test("late answers use the existing follow-up prompt and dismissal does not interrupt", async () => {
const { session, appServer, ask, finish } = await setup();
try {
await ask();
const [permission] = session.getPendingPermissions();
expect(permission).toBeDefined();
await finish();
expect(session.getPendingPermissions()).toHaveLength(1);
expect(await session.respondToPermission(permission.id, answer)).toEqual({
followUpPrompt: "Answers to your questions:\n\nWhich color?\nGreen",
});
expect(appServer.requests().some((request) => request.method === "turn/interrupt")).toBe(false);
} finally {
await session.close();
}
});
test("restores unanswered questions and does not reopen answered questions on duplicate events", async () => {
const first = await setup();
let metadata: Record<string, unknown> | undefined;
try {
await first.ask();
expect(first.session.getPendingPermissions()).toHaveLength(1);
metadata = first.session.describePersistence()!.metadata;
} finally {
await first.session.close();
}
const resumed = await setup(metadata);
try {
const [permission] = resumed.session.getPendingPermissions();
expect(permission).toBeDefined();
await resumed.session.respondToPermission(permission.id, { behavior: "deny" });
await resumed.ask();
expect(resumed.session.getPendingPermissions()).toEqual([]);
expect(
resumed.appServer.requests().some((request) => request.method === "turn/interrupt"),
).toBe(false);
} finally {
await resumed.session.close();
}
});
@@ -0,0 +1,155 @@
import { z } from "zod";
import type {
AgentPermissionRequest,
AgentPermissionResponse,
ToolCallTimelineItem,
} from "../../agent-sdk-types.js";
const QuestionSchema = z.object({
title: z.string().trim().min(1),
options: z.array(z.string().min(1)).nullish(),
});
const ItemSchema = z.object({
type: z.literal("agentMessage"),
id: z.string().min(1),
delivery: z.literal("async"),
questions: z.array(QuestionSchema).min(1),
});
const RecordSchema = z.object({
item: ItemSchema,
resolution: z.union([z.literal("dismissed"), z.array(z.string())]).optional(),
});
type QuestionRecord = z.infer<typeof RecordSchema>;
function requestId(itemId: string): string {
return `permission-${itemId}`;
}
function toPermission(record: QuestionRecord): AgentPermissionRequest {
return {
id: requestId(record.item.id),
provider: "codex",
name: "request_user_input_async",
kind: "question",
title: "Question",
input: {
questions: record.item.questions.map((question, index) => ({
id: String(index),
header: `Question ${index + 1}`,
question: question.title,
options: (question.options ?? []).map((label) => ({ label })),
isOther: true,
})),
},
};
}
function toTimeline(record: QuestionRecord): ToolCallTimelineItem {
const answers = Array.isArray(record.resolution) ? record.resolution : undefined;
return {
type: "tool_call",
callId: record.item.id,
name: "request_user_input_async",
status: "completed",
error: null,
detail: {
type: "plain_text",
icon: "brain",
text:
record.item.questions
.map((question, index) =>
[question.title, answers ? answers[index] : (question.options ?? []).join(", ")]
.filter(Boolean)
.join("\n"),
)
.join("\n\n") + (record.resolution === "dismissed" ? "\n\nDismissed" : ""),
},
};
}
export function codexAsyncQuestionToTimeline(item: unknown): ToolCallTimelineItem | null {
const parsed = ItemSchema.safeParse(item);
return parsed.success ? toTimeline({ item: parsed.data }) : null;
}
/** Codex emits these as completed messages; the outstanding answer belongs to the session. */
export class CodexAsyncQuestions {
private readonly records = new Map<string, QuestionRecord>();
constructor(saved: unknown) {
const parsed = z.array(RecordSchema).safeParse(saved);
if (parsed.success) {
for (const record of parsed.data) this.records.set(requestId(record.item.id), record);
}
}
receive(item: unknown): AgentPermissionRequest | null {
const parsed = ItemSchema.safeParse(item);
if (!parsed.success || this.records.has(requestId(parsed.data.id))) return null;
const record = { item: parsed.data };
this.records.set(requestId(parsed.data.id), record);
return toPermission(record);
}
pending(): AgentPermissionRequest[] {
return Array.from(this.records.values())
.filter((record) => record.resolution === undefined)
.map(toPermission);
}
hasPending(id: string): boolean {
const record = this.records.get(id);
return record !== undefined && record.resolution === undefined;
}
prepareResponse(
id: string,
response: AgentPermissionResponse,
): { prompt?: string; complete: () => ToolCallTimelineItem } {
const record = this.records.get(id);
if (!record || record.resolution !== undefined)
throw new Error("Question is no longer pending");
let resolution: QuestionRecord["resolution"] = "dismissed";
let prompt: string | undefined;
if (response.behavior === "allow") {
const answers = z.record(z.string(), z.string()).parse(response.updatedInput?.answers);
resolution = record.item.questions.map((_, index) => {
const answer = answers[`Question ${index + 1}`]?.trim();
if (!answer) throw new Error(`Answer Question ${index + 1} before submitting`);
return answer;
});
const values = resolution;
prompt =
"Answers to your questions:\n\n" +
record.item.questions
.map((question, index) => `${question.title}\n${values[index]}`)
.join("\n\n");
}
return {
prompt,
complete: () => {
record.resolution = resolution;
return toTimeline(record);
},
};
}
timeline(itemId: string): ToolCallTimelineItem | null {
const record = this.records.get(requestId(itemId));
return record ? toTimeline(record) : null;
}
retain(itemIds: ReadonlySet<string>): string[] {
const removedPendingIds: string[] = [];
for (const [id, record] of this.records) {
if (itemIds.has(record.item.id)) continue;
this.records.delete(id);
if (record.resolution === undefined) removedPendingIds.push(id);
}
return removedPendingIds;
}
serialize(): unknown {
return Array.from(this.records.values());
}
}
@@ -618,7 +618,7 @@ function toJsonObject(value: unknown): JsonObject {
type StreamEventType = AgentStreamEvent["type"];
type StreamEventOfType<TType extends StreamEventType> = Extract<AgentStreamEvent, { type: TType }>;
function waitForNextEvent<TType extends StreamEventType>(
export function waitForNextEvent<TType extends StreamEventType>(
session: AgentSession,
type: TType,
accepts?: (event: StreamEventOfType<TType>) => boolean,
+15 -14
View File
@@ -188,20 +188,21 @@ An environment or dynamic named-agent expression must have a finite set of possi
### Steps
| Field | Required | Notes |
| --------------- | -------- | ---------------------------------------------------------------------------------------- |
| `id` | yes | Unique within the workflow. |
| `environment` | yes | Literal environment name or finite expression resolving to one. |
| `max_runtime` | yes | Step hard limit. |
| `idle_timeout` | yes | Idle limit no longer than `max_runtime`. |
| `agent` | yes | Named agent, finite expression selecting a named agent, or complete static inline agent. |
| `prompt` | yes | Ordered `text` and `include` blocks. |
| `if` | no | Expression deciding whether the step runs. |
| `env` | no | Environment variables from connection values. |
| `output.schema` | no | JSON Schema for structured step output. |
| `allow_outputs` | no | Provider output capabilities with optional `max` and `required`. |
| `auto_archive` | no | Archive the agent after the step ends. |
| `github` | no | Explicit GitHub authority for this step. |
| Field | Required | Notes |
| ----------------- | -------- | -------------------------------------------------------------------------------------------------------------------- |
| `id` | yes | Unique within the workflow. |
| `environment` | yes | Literal environment name or finite expression resolving to one. |
| `max_runtime` | yes | Step hard limit. |
| `idle_timeout` | yes | Idle limit no longer than `max_runtime`. |
| `startup_timeout` | no | Step startup waiting budget. See [startup timeout](/docs/hub/configuration#startup-timeout) for defaults and limits. |
| `agent` | yes | Named agent, finite expression selecting a named agent, or complete static inline agent. |
| `prompt` | yes | Ordered `text` and `include` blocks. |
| `if` | no | Expression deciding whether the step runs. |
| `env` | no | Environment variables from connection values. |
| `output.schema` | no | JSON Schema for structured step output. |
| `allow_outputs` | no | Provider output capabilities with optional `max` and `required`. |
| `auto_archive` | no | Archive the agent after the step ends. |
| `github` | no | Explicit GitHub authority for this step. |
An inline agent is static and complete:
+17
View File
@@ -64,6 +64,23 @@ A Discord starter uses `discord.mention`, your Discord user ID, and `discord.rep
Setup asks before replacing the selected trigger file. It preserves other triggers and any existing legacy bundle. Read [Hub security](/docs/hub/security) before widening `from_users` or the agent's authority.
## Startup timeout
Hub waits up to **two minutes** for agent startup, including worktree creation, provider startup, and initial prompt acceptance. For a slower machine, set `run.startup_timeout` in the trigger YAML:
```yaml
name: inspect
on:
manual.run: {}
run:
target: { daemon: devbox, cwd: /workspace/project }
agent: { provider: codex, mode: full-access }
startup_timeout: 5m
prompt: Inspect this repository and summarize its current state.
```
Use a positive duration in `ms`, `s`, `m`, or `h`, up to `24h`. Omitting the field uses `2m`. The existing `max_runtime` and `idle_timeout` limits still apply during startup and can expire sooner. This setting changes Hub's waiting budget; provider-specific timeouts remain in effect.
## Deploy from the CLI
Run from the repository root:
+2
View File
@@ -96,6 +96,8 @@ worktree:
For agents it dispatched, Hub owns creation, reconnect recovery, output observation, and completion. Agents you start yourself are untouched.
Restarting Hub lets active agents continue working. Hub reconnects to the same agents and resumes tracking their output and completion. [GitHub credentials retain their original lifetime](/docs/hub/github#restarts-and-credential-lifetime).
If Hub loses the create response, or the daemon restarts mid-execution, Hub resends the same create intent with the same execution id. The daemon returns the existing agent rather than running the prompt again. An agent that closed or errored is recorded as interrupted; Hub never silently starts a second one.
## Status
+12 -2
View File
@@ -37,7 +37,7 @@ steps:
${{ paseo.prompt }}
```
The agent can use `git` and `gh` within the declared repositories and permissions. Hub mints the token when the step starts and revokes it when execution ends.
The agent can use `git` and `gh` within the declared repositories and permissions. Hub mints the token when the step starts and revokes it when execution ends or its configured duration elapses.
## Fields
@@ -50,6 +50,12 @@ The agent can use `git` and `gh` within the declared repositories and permission
Requested authority cannot exceed the GitHub App installation. Activation and dispatch fail clearly when the connection, repository, or permissions cannot be resolved.
## Restarts and credential lifetime
Restarting Hub preserves active executions and their existing credentials. Hub reconnects to the same agent and restores the original credential deadlines. A restart does not mint a replacement token or extend its lifetime.
If Hub is unavailable when a shorter duration elapses, it revokes the overdue token when it returns. GitHub’s own expiry still limits the token lifetime to one hour. Revocation failures are retried after recovery.
## Agent environment
Hub supplies `GH_TOKEN` and process-scoped git configuration through environment variables:
@@ -73,4 +79,8 @@ env:
SOME_TOKEN: "${{ paseo.connections.some-connection.token }}"
```
Hub resolves the value for the step and does not persist it. See [Hub security](/docs/hub/security) for provider and host boundaries.
Hub persists resolved values as private execution data so it can recover after a restart. Authored configuration retains the expressions.
These values are stored in the Hub database without application-level encryption. Protect database access, storage, and backups as credentials. Hub deletes the execution authority record when the execution becomes terminal; token lease records remain until revocation succeeds or the token expires, so interrupted cleanup can resume. This cleanup is not a guarantee that every copy is erased: retained agent/session data and database backups may still contain resolved values. Token revocation ends access; it does not erase those copies. Backup retention and deletion remain the operator’s responsibility.
See [Hub security](/docs/hub/security) for provider and host boundaries.