mirror of
https://github.com/harivansh-afk/sandbox-agent.git
synced 2026-04-16 21:03:46 +00:00
Configure lefthook formatter checks (#231)
* Add lefthook formatter checks * Fix SDK mode hydration * Stabilize SDK mode integration test
This commit is contained in:
Nathan Flurry
GitHub
parent
0471214d65
commit
d2346bafb3
282 changed files with 5840 additions and 8399 deletions
|
|
@ -1,13 +1,7 @@
|
|||
import { execFile } from "node:child_process";
|
||||
import { promisify } from "node:util";
|
||||
import { describe, expect, it } from "vitest";
|
||||
import type {
|
||||
HandoffWorkbenchSnapshot,
|
||||
WorkbenchAgentTab,
|
||||
WorkbenchHandoff,
|
||||
WorkbenchModelId,
|
||||
WorkbenchTranscriptEvent,
|
||||
} from "@openhandoff/shared";
|
||||
import type { HandoffWorkbenchSnapshot, WorkbenchAgentTab, WorkbenchHandoff, WorkbenchModelId, WorkbenchTranscriptEvent } from "@openhandoff/shared";
|
||||
import { createBackendClient } from "../../src/backend-client.js";
|
||||
|
||||
const RUN_WORKBENCH_E2E = process.env.HF_ENABLE_DAEMON_WORKBENCH_E2E === "1";
|
||||
|
|
@ -48,13 +42,7 @@ async function seedSandboxFile(workspaceId: string, handoffId: string, filePath:
|
|||
await execFileAsync("docker", ["exec", "openhandoff-backend-1", "bash", "-lc", script]);
|
||||
}
|
||||
|
||||
async function poll<T>(
|
||||
label: string,
|
||||
timeoutMs: number,
|
||||
intervalMs: number,
|
||||
fn: () => Promise<T>,
|
||||
isDone: (value: T) => boolean,
|
||||
): Promise<T> {
|
||||
async function poll<T>(label: string, timeoutMs: number, intervalMs: number, fn: () => Promise<T>, isDone: (value: T) => boolean): Promise<T> {
|
||||
const startedAt = Date.now();
|
||||
let lastValue: T;
|
||||
|
||||
|
|
@ -147,10 +135,7 @@ function extractEventText(event: WorkbenchTranscriptEvent): string {
|
|||
return JSON.stringify(payload);
|
||||
}
|
||||
|
||||
function transcriptIncludesAgentText(
|
||||
transcript: WorkbenchTranscriptEvent[],
|
||||
expectedText: string,
|
||||
): boolean {
|
||||
function transcriptIncludesAgentText(transcript: WorkbenchTranscriptEvent[], expectedText: string): boolean {
|
||||
return transcript
|
||||
.filter((event) => event.sender === "agent")
|
||||
.map((event) => extractEventText(event))
|
||||
|
|
@ -159,176 +144,164 @@ function transcriptIncludesAgentText(
|
|||
}
|
||||
|
||||
describe("e2e(client): workbench flows", () => {
|
||||
it.skipIf(!RUN_WORKBENCH_E2E)(
|
||||
"creates a handoff, adds sessions, exchanges messages, and manages workbench state",
|
||||
{ timeout: 20 * 60_000 },
|
||||
async () => {
|
||||
const endpoint =
|
||||
process.env.HF_E2E_BACKEND_ENDPOINT?.trim() || "http://127.0.0.1:7741/api/rivet";
|
||||
const workspaceId = process.env.HF_E2E_WORKSPACE?.trim() || "default";
|
||||
const repoRemote = requiredEnv("HF_E2E_GITHUB_REPO");
|
||||
const model = workbenchModelEnv("HF_E2E_MODEL", "gpt-4o");
|
||||
const runId = `wb-${Date.now().toString(36)}`;
|
||||
const expectedFile = `${runId}.txt`;
|
||||
const expectedInitialReply = `WORKBENCH_READY_${runId}`;
|
||||
const expectedReply = `WORKBENCH_ACK_${runId}`;
|
||||
it.skipIf(!RUN_WORKBENCH_E2E)("creates a handoff, adds sessions, exchanges messages, and manages workbench state", { timeout: 20 * 60_000 }, async () => {
|
||||
const endpoint = process.env.HF_E2E_BACKEND_ENDPOINT?.trim() || "http://127.0.0.1:7741/api/rivet";
|
||||
const workspaceId = process.env.HF_E2E_WORKSPACE?.trim() || "default";
|
||||
const repoRemote = requiredEnv("HF_E2E_GITHUB_REPO");
|
||||
const model = workbenchModelEnv("HF_E2E_MODEL", "gpt-4o");
|
||||
const runId = `wb-${Date.now().toString(36)}`;
|
||||
const expectedFile = `${runId}.txt`;
|
||||
const expectedInitialReply = `WORKBENCH_READY_${runId}`;
|
||||
const expectedReply = `WORKBENCH_ACK_${runId}`;
|
||||
|
||||
const client = createBackendClient({
|
||||
endpoint,
|
||||
defaultWorkspaceId: workspaceId,
|
||||
});
|
||||
const client = createBackendClient({
|
||||
endpoint,
|
||||
defaultWorkspaceId: workspaceId,
|
||||
});
|
||||
|
||||
const repo = await client.addRepo(workspaceId, repoRemote);
|
||||
const created = await client.createWorkbenchHandoff(workspaceId, {
|
||||
repoId: repo.repoId,
|
||||
title: `Workbench E2E ${runId}`,
|
||||
branch: `e2e/${runId}`,
|
||||
model,
|
||||
task: `Reply with exactly: ${expectedInitialReply}`,
|
||||
});
|
||||
const repo = await client.addRepo(workspaceId, repoRemote);
|
||||
const created = await client.createWorkbenchHandoff(workspaceId, {
|
||||
repoId: repo.repoId,
|
||||
title: `Workbench E2E ${runId}`,
|
||||
branch: `e2e/${runId}`,
|
||||
model,
|
||||
task: `Reply with exactly: ${expectedInitialReply}`,
|
||||
});
|
||||
|
||||
const provisioned = await poll(
|
||||
"handoff provisioning",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => handoff.branch === `e2e/${runId}` && handoff.tabs.length > 0,
|
||||
);
|
||||
const provisioned = await poll(
|
||||
"handoff provisioning",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => handoff.branch === `e2e/${runId}` && handoff.tabs.length > 0,
|
||||
);
|
||||
|
||||
const primaryTab = provisioned.tabs[0]!;
|
||||
const primaryTab = provisioned.tabs[0]!;
|
||||
|
||||
const initialCompleted = await poll(
|
||||
"initial agent response",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, primaryTab.id);
|
||||
return (
|
||||
handoff.status === "idle" &&
|
||||
tab.status === "idle" &&
|
||||
transcriptIncludesAgentText(tab.transcript, expectedInitialReply)
|
||||
);
|
||||
const initialCompleted = await poll(
|
||||
"initial agent response",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, primaryTab.id);
|
||||
return handoff.status === "idle" && tab.status === "idle" && transcriptIncludesAgentText(tab.transcript, expectedInitialReply);
|
||||
},
|
||||
);
|
||||
|
||||
expect(findTab(initialCompleted, primaryTab.id).sessionId).toBeTruthy();
|
||||
expect(transcriptIncludesAgentText(findTab(initialCompleted, primaryTab.id).transcript, expectedInitialReply)).toBe(true);
|
||||
|
||||
await seedSandboxFile(workspaceId, created.handoffId, expectedFile, runId);
|
||||
|
||||
const fileSeeded = await poll(
|
||||
"seeded sandbox file reflected in workbench",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => handoff.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
expect(fileSeeded.fileChanges.some((file) => file.path === expectedFile)).toBe(true);
|
||||
|
||||
await client.renameWorkbenchHandoff(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
value: `Workbench E2E ${runId} Renamed`,
|
||||
});
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: primaryTab.id,
|
||||
title: "Primary Session",
|
||||
});
|
||||
|
||||
const secondTab = await client.createWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
model,
|
||||
});
|
||||
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
title: "Follow-up Session",
|
||||
});
|
||||
|
||||
await client.updateWorkbenchDraft(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [
|
||||
{
|
||||
id: `${expectedFile}:1`,
|
||||
filePath: expectedFile,
|
||||
lineNumber: 1,
|
||||
lineContent: runId,
|
||||
},
|
||||
);
|
||||
],
|
||||
});
|
||||
|
||||
expect(findTab(initialCompleted, primaryTab.id).sessionId).toBeTruthy();
|
||||
expect(transcriptIncludesAgentText(findTab(initialCompleted, primaryTab.id).transcript, expectedInitialReply)).toBe(true);
|
||||
const drafted = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.text).toContain(expectedReply);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.attachments).toHaveLength(1);
|
||||
|
||||
await seedSandboxFile(workspaceId, created.handoffId, expectedFile, runId);
|
||||
await client.sendWorkbenchMessage(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [],
|
||||
});
|
||||
|
||||
const fileSeeded = await poll(
|
||||
"seeded sandbox file reflected in workbench",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => handoff.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
expect(fileSeeded.fileChanges.some((file) => file.path === expectedFile)).toBe(true);
|
||||
const withSecondReply = await poll(
|
||||
"follow-up session response",
|
||||
10 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, secondTab.tabId);
|
||||
return tab.status === "idle" && transcriptIncludesAgentText(tab.transcript, expectedReply);
|
||||
},
|
||||
);
|
||||
|
||||
await client.renameWorkbenchHandoff(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
value: `Workbench E2E ${runId} Renamed`,
|
||||
});
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: primaryTab.id,
|
||||
title: "Primary Session",
|
||||
});
|
||||
const secondTranscript = findTab(withSecondReply, secondTab.tabId).transcript;
|
||||
expect(transcriptIncludesAgentText(secondTranscript, expectedReply)).toBe(true);
|
||||
|
||||
const secondTab = await client.createWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
model,
|
||||
});
|
||||
await client.setWorkbenchSessionUnread(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
unread: false,
|
||||
});
|
||||
await client.markWorkbenchUnread(workspaceId, { handoffId: created.handoffId });
|
||||
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
title: "Follow-up Session",
|
||||
});
|
||||
const unreadSnapshot = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
expect(unreadSnapshot.tabs.some((tab) => tab.unread)).toBe(true);
|
||||
|
||||
await client.updateWorkbenchDraft(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [
|
||||
{
|
||||
id: `${expectedFile}:1`,
|
||||
filePath: expectedFile,
|
||||
lineNumber: 1,
|
||||
lineContent: runId,
|
||||
},
|
||||
],
|
||||
});
|
||||
await client.closeWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
});
|
||||
|
||||
const drafted = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.text).toContain(expectedReply);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.attachments).toHaveLength(1);
|
||||
const closedSnapshot = await poll(
|
||||
"secondary session closed",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => !handoff.tabs.some((tab) => tab.id === secondTab.tabId),
|
||||
);
|
||||
expect(closedSnapshot.tabs).toHaveLength(1);
|
||||
|
||||
await client.sendWorkbenchMessage(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [],
|
||||
});
|
||||
await client.revertWorkbenchFile(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
path: expectedFile,
|
||||
});
|
||||
|
||||
const withSecondReply = await poll(
|
||||
"follow-up session response",
|
||||
10 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, secondTab.tabId);
|
||||
return (
|
||||
tab.status === "idle" &&
|
||||
transcriptIncludesAgentText(tab.transcript, expectedReply)
|
||||
);
|
||||
},
|
||||
);
|
||||
const revertedSnapshot = await poll(
|
||||
"file revert reflected in workbench",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => !handoff.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
|
||||
const secondTranscript = findTab(withSecondReply, secondTab.tabId).transcript;
|
||||
expect(transcriptIncludesAgentText(secondTranscript, expectedReply)).toBe(true);
|
||||
|
||||
await client.setWorkbenchSessionUnread(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
unread: false,
|
||||
});
|
||||
await client.markWorkbenchUnread(workspaceId, { handoffId: created.handoffId });
|
||||
|
||||
const unreadSnapshot = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
expect(unreadSnapshot.tabs.some((tab) => tab.unread)).toBe(true);
|
||||
|
||||
await client.closeWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
tabId: secondTab.tabId,
|
||||
});
|
||||
|
||||
const closedSnapshot = await poll(
|
||||
"secondary session closed",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => !handoff.tabs.some((tab) => tab.id === secondTab.tabId),
|
||||
);
|
||||
expect(closedSnapshot.tabs).toHaveLength(1);
|
||||
|
||||
await client.revertWorkbenchFile(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
path: expectedFile,
|
||||
});
|
||||
|
||||
const revertedSnapshot = await poll(
|
||||
"file revert reflected in workbench",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => !handoff.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
|
||||
expect(revertedSnapshot.fileChanges.some((file) => file.path === expectedFile)).toBe(false);
|
||||
expect(revertedSnapshot.title).toBe(`Workbench E2E ${runId} Renamed`);
|
||||
expect(findTab(revertedSnapshot, primaryTab.id).sessionName).toBe("Primary Session");
|
||||
},
|
||||
);
|
||||
expect(revertedSnapshot.fileChanges.some((file) => file.path === expectedFile)).toBe(false);
|
||||
expect(revertedSnapshot.title).toBe(`Workbench E2E ${runId} Renamed`);
|
||||
expect(findTab(revertedSnapshot, primaryTab.id).sessionName).toBe("Primary Session");
|
||||
});
|
||||
});
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue