mirror of
https://github.com/harivansh-afk/sandbox-agent.git
synced 2026-04-16 22:03:52 +00:00
factory: rename project and handoff actors
This commit is contained in:
Nathan Flurry
parent
3022bce2ad
commit
ea7c36a8e7
147 changed files with 6313 additions and 14364 deletions
|
|
@ -1,5 +1,5 @@
|
|||
import { describe, expect, it } from "vitest";
|
||||
import type { HandoffRecord, HistoryEvent } from "@sandbox-agent/factory-shared";
|
||||
import type { TaskRecord, HistoryEvent } from "@sandbox-agent/factory-shared";
|
||||
import { createBackendClient } from "../../src/backend-client.js";
|
||||
|
||||
const RUN_E2E = process.env.HF_ENABLE_DAEMON_E2E === "1";
|
||||
|
|
@ -79,20 +79,20 @@ function parseHistoryPayload(event: HistoryEvent): Record<string, unknown> {
|
|||
}
|
||||
}
|
||||
|
||||
async function debugDump(client: ReturnType<typeof createBackendClient>, workspaceId: string, handoffId: string): Promise<string> {
|
||||
async function debugDump(client: ReturnType<typeof createBackendClient>, workspaceId: string, taskId: string): Promise<string> {
|
||||
try {
|
||||
const handoff = await client.getHandoff(workspaceId, handoffId);
|
||||
const history = await client.listHistory({ workspaceId, handoffId, limit: 80 }).catch(() => []);
|
||||
const task = await client.getTask(workspaceId, taskId);
|
||||
const history = await client.listHistory({ workspaceId, taskId, limit: 80 }).catch(() => []);
|
||||
const historySummary = history
|
||||
.slice(0, 20)
|
||||
.map((e) => `${new Date(e.createdAt).toISOString()} ${e.kind}`)
|
||||
.join("\n");
|
||||
|
||||
let sessionEventsSummary = "";
|
||||
if (handoff.activeSandboxId && handoff.activeSessionId) {
|
||||
if (task.activeSandboxId && task.activeSessionId) {
|
||||
const events = await client
|
||||
.listSandboxSessionEvents(workspaceId, handoff.providerId, handoff.activeSandboxId, {
|
||||
sessionId: handoff.activeSessionId,
|
||||
.listSandboxSessionEvents(workspaceId, task.providerId, task.activeSandboxId, {
|
||||
sessionId: task.activeSessionId,
|
||||
limit: 50,
|
||||
})
|
||||
.then((r) => r.items)
|
||||
|
|
@ -104,17 +104,17 @@ async function debugDump(client: ReturnType<typeof createBackendClient>, workspa
|
|||
}
|
||||
|
||||
return [
|
||||
"=== handoff ===",
|
||||
"=== task ===",
|
||||
JSON.stringify(
|
||||
{
|
||||
status: handoff.status,
|
||||
statusMessage: handoff.statusMessage,
|
||||
title: handoff.title,
|
||||
branchName: handoff.branchName,
|
||||
activeSandboxId: handoff.activeSandboxId,
|
||||
activeSessionId: handoff.activeSessionId,
|
||||
prUrl: handoff.prUrl,
|
||||
prSubmitted: handoff.prSubmitted,
|
||||
status: task.status,
|
||||
statusMessage: task.statusMessage,
|
||||
title: task.title,
|
||||
branchName: task.branchName,
|
||||
activeSandboxId: task.activeSandboxId,
|
||||
activeSessionId: task.activeSessionId,
|
||||
prUrl: task.prUrl,
|
||||
prSubmitted: task.prSubmitted,
|
||||
},
|
||||
null,
|
||||
2
|
||||
|
|
@ -144,7 +144,7 @@ async function githubApi(token: string, path: string, init?: RequestInit): Promi
|
|||
|
||||
describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
||||
it.skipIf(!RUN_E2E)(
|
||||
"creates a handoff, waits for agent to implement, and opens a PR",
|
||||
"creates a task, waits for agent to implement, and opens a PR",
|
||||
{ timeout: 15 * 60_000 },
|
||||
async () => {
|
||||
const endpoint =
|
||||
|
|
@ -164,7 +164,7 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
|
||||
const repo = await client.addRepo(workspaceId, repoRemote);
|
||||
|
||||
const created = await client.createHandoff({
|
||||
const created = await client.createTask({
|
||||
workspaceId,
|
||||
repoId: repo.repoId,
|
||||
task: [
|
||||
|
|
@ -187,42 +187,42 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
let lastStatus: string | null = null;
|
||||
|
||||
try {
|
||||
const namedAndProvisioned = await poll<HandoffRecord>(
|
||||
"handoff naming + sandbox provisioning",
|
||||
const namedAndProvisioned = await poll<TaskRecord>(
|
||||
"task naming + sandbox provisioning",
|
||||
// Cold Daytona snapshot/image preparation can exceed 5 minutes on first run.
|
||||
8 * 60_000,
|
||||
1_000,
|
||||
async () => client.getHandoff(workspaceId, created.handoffId),
|
||||
async () => client.getTask(workspaceId, created.taskId),
|
||||
(h) => Boolean(h.title && h.branchName && h.activeSandboxId),
|
||||
(h) => {
|
||||
if (h.status !== lastStatus) {
|
||||
lastStatus = h.status;
|
||||
}
|
||||
if (h.status === "error") {
|
||||
throw new Error("handoff entered error state during provisioning");
|
||||
throw new Error("task entered error state during provisioning");
|
||||
}
|
||||
}
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
});
|
||||
|
||||
branchName = namedAndProvisioned.branchName!;
|
||||
sandboxId = namedAndProvisioned.activeSandboxId!;
|
||||
|
||||
const withSession = await poll<HandoffRecord>(
|
||||
"handoff to create active session",
|
||||
const withSession = await poll<TaskRecord>(
|
||||
"task to create active session",
|
||||
3 * 60_000,
|
||||
1_500,
|
||||
async () => client.getHandoff(workspaceId, created.handoffId),
|
||||
async () => client.getTask(workspaceId, created.taskId),
|
||||
(h) => Boolean(h.activeSessionId),
|
||||
(h) => {
|
||||
if (h.status === "error") {
|
||||
throw new Error("handoff entered error state while waiting for active session");
|
||||
throw new Error("task entered error state while waiting for active session");
|
||||
}
|
||||
}
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
});
|
||||
|
||||
|
|
@ -241,23 +241,23 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
).items,
|
||||
(events) => events.length > 0
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
});
|
||||
|
||||
await poll<HandoffRecord>(
|
||||
"handoff to reach idle state",
|
||||
await poll<TaskRecord>(
|
||||
"task to reach idle state",
|
||||
8 * 60_000,
|
||||
2_000,
|
||||
async () => client.getHandoff(workspaceId, created.handoffId),
|
||||
async () => client.getTask(workspaceId, created.taskId),
|
||||
(h) => h.status === "idle",
|
||||
(h) => {
|
||||
if (h.status === "error") {
|
||||
throw new Error("handoff entered error state while waiting for idle");
|
||||
throw new Error("task entered error state while waiting for idle");
|
||||
}
|
||||
}
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
});
|
||||
|
||||
|
|
@ -265,14 +265,14 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
"PR creation history event",
|
||||
3 * 60_000,
|
||||
2_000,
|
||||
async () => client.listHistory({ workspaceId, handoffId: created.handoffId, limit: 200 }),
|
||||
(events) => events.some((e) => e.kind === "handoff.pr_created")
|
||||
async () => client.listHistory({ workspaceId, taskId: created.taskId, limit: 200 }),
|
||||
(events) => events.some((e) => e.kind === "task.pr_created")
|
||||
)
|
||||
.catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
})
|
||||
.then((events) => events.find((e) => e.kind === "handoff.pr_created")!);
|
||||
.then((events) => events.find((e) => e.kind === "task.pr_created")!);
|
||||
|
||||
const payload = parseHistoryPayload(prCreatedEvent);
|
||||
prNumber = Number(payload.prNumber);
|
||||
|
|
@ -293,17 +293,17 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
const prFiles = (await prFilesRes.json()) as Array<{ filename: string }>;
|
||||
expect(prFiles.some((f) => f.filename === expectedFile)).toBe(true);
|
||||
|
||||
// Close the handoff and assert the sandbox is released (stopped).
|
||||
await client.runAction(workspaceId, created.handoffId, "archive");
|
||||
// Close the task and assert the sandbox is released (stopped).
|
||||
await client.runAction(workspaceId, created.taskId, "archive");
|
||||
|
||||
await poll<HandoffRecord>(
|
||||
"handoff to become archived (session released)",
|
||||
await poll<TaskRecord>(
|
||||
"task to become archived (session released)",
|
||||
60_000,
|
||||
1_000,
|
||||
async () => client.getHandoff(workspaceId, created.handoffId),
|
||||
async () => client.getTask(workspaceId, created.taskId),
|
||||
(h) => h.status === "archived" && h.activeSessionId === null
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
throw new Error(`${err instanceof Error ? err.message : String(err)}\n${dump}`);
|
||||
});
|
||||
|
||||
|
|
@ -318,7 +318,7 @@ describe("e2e: backend -> sandbox-agent -> git -> PR", () => {
|
|||
return st.includes("stopped") || st.includes("suspended") || st.includes("paused");
|
||||
}
|
||||
).catch(async (err) => {
|
||||
const dump = await debugDump(client, workspaceId, created.handoffId);
|
||||
const dump = await debugDump(client, workspaceId, created.taskId);
|
||||
const state = await client
|
||||
.sandboxProviderState(workspaceId, "daytona", sandboxId!)
|
||||
.catch(() => null);
|
||||
|
|
|
|||
|
|
@ -3,10 +3,10 @@ import { mkdir, writeFile } from "node:fs/promises";
|
|||
import { promisify } from "node:util";
|
||||
import { describe, expect, it } from "vitest";
|
||||
import type {
|
||||
HandoffRecord,
|
||||
HandoffWorkbenchSnapshot,
|
||||
TaskRecord,
|
||||
TaskWorkbenchSnapshot,
|
||||
WorkbenchAgentTab,
|
||||
WorkbenchHandoff,
|
||||
WorkbenchTask,
|
||||
WorkbenchModelId,
|
||||
WorkbenchTranscriptEvent,
|
||||
} from "@sandbox-agent/factory-shared";
|
||||
|
|
@ -76,18 +76,18 @@ async function resolveBackendContainerName(endpoint: string): Promise<string | n
|
|||
return containerName ?? null;
|
||||
}
|
||||
|
||||
function sandboxRepoPath(record: HandoffRecord): string {
|
||||
function sandboxRepoPath(record: TaskRecord): string {
|
||||
const activeSandbox =
|
||||
record.sandboxes.find((sandbox) => sandbox.sandboxId === record.activeSandboxId) ??
|
||||
record.sandboxes.find((sandbox) => typeof sandbox.cwd === "string" && sandbox.cwd.length > 0);
|
||||
const cwd = activeSandbox?.cwd?.trim();
|
||||
if (!cwd) {
|
||||
throw new Error(`No sandbox cwd is available for handoff ${record.handoffId}`);
|
||||
throw new Error(`No sandbox cwd is available for task ${record.taskId}`);
|
||||
}
|
||||
return cwd;
|
||||
}
|
||||
|
||||
async function seedSandboxFile(endpoint: string, record: HandoffRecord, filePath: string, content: string): Promise<void> {
|
||||
async function seedSandboxFile(endpoint: string, record: TaskRecord, filePath: string, content: string): Promise<void> {
|
||||
const repoPath = sandboxRepoPath(record);
|
||||
const containerName = await resolveBackendContainerName(endpoint);
|
||||
if (!containerName) {
|
||||
|
|
@ -128,18 +128,18 @@ async function poll<T>(
|
|||
}
|
||||
}
|
||||
|
||||
function findHandoff(snapshot: HandoffWorkbenchSnapshot, handoffId: string): WorkbenchHandoff {
|
||||
const handoff = snapshot.handoffs.find((candidate) => candidate.id === handoffId);
|
||||
if (!handoff) {
|
||||
throw new Error(`handoff ${handoffId} missing from snapshot`);
|
||||
function findTask(snapshot: TaskWorkbenchSnapshot, taskId: string): WorkbenchTask {
|
||||
const task = snapshot.tasks.find((candidate) => candidate.id === taskId);
|
||||
if (!task) {
|
||||
throw new Error(`task ${taskId} missing from snapshot`);
|
||||
}
|
||||
return handoff;
|
||||
return task;
|
||||
}
|
||||
|
||||
function findTab(handoff: WorkbenchHandoff, tabId: string): WorkbenchAgentTab {
|
||||
const tab = handoff.tabs.find((candidate) => candidate.id === tabId);
|
||||
function findTab(task: WorkbenchTask, tabId: string): WorkbenchAgentTab {
|
||||
const tab = task.tabs.find((candidate) => candidate.id === tabId);
|
||||
if (!tab) {
|
||||
throw new Error(`tab ${tabId} missing from handoff ${handoff.id}`);
|
||||
throw new Error(`tab ${tabId} missing from task ${task.id}`);
|
||||
}
|
||||
return tab;
|
||||
}
|
||||
|
|
@ -218,7 +218,7 @@ function transcriptIncludesAgentText(
|
|||
|
||||
describe("e2e(client): workbench flows", () => {
|
||||
it.skipIf(!RUN_WORKBENCH_E2E)(
|
||||
"creates a handoff, adds sessions, exchanges messages, and manages workbench state",
|
||||
"creates a task, adds sessions, exchanges messages, and manages workbench state",
|
||||
{ timeout: 20 * 60_000 },
|
||||
async () => {
|
||||
const endpoint =
|
||||
|
|
@ -237,7 +237,7 @@ describe("e2e(client): workbench flows", () => {
|
|||
});
|
||||
|
||||
const repo = await client.addRepo(workspaceId, repoRemote);
|
||||
const created = await client.createWorkbenchHandoff(workspaceId, {
|
||||
const created = await client.createWorkbenchTask(workspaceId, {
|
||||
repoId: repo.repoId,
|
||||
title: `Workbench E2E ${runId}`,
|
||||
branch: `e2e/${runId}`,
|
||||
|
|
@ -246,11 +246,11 @@ describe("e2e(client): workbench flows", () => {
|
|||
});
|
||||
|
||||
const provisioned = await poll(
|
||||
"handoff provisioning",
|
||||
"task provisioning",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => handoff.branch === `e2e/${runId}` && handoff.tabs.length > 0,
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => task.branch === `e2e/${runId}` && task.tabs.length > 0,
|
||||
);
|
||||
|
||||
const primaryTab = provisioned.tabs[0]!;
|
||||
|
|
@ -259,11 +259,11 @@ describe("e2e(client): workbench flows", () => {
|
|||
"initial agent response",
|
||||
12 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, primaryTab.id);
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => {
|
||||
const tab = findTab(task, primaryTab.id);
|
||||
return (
|
||||
handoff.status === "idle" &&
|
||||
task.status === "idle" &&
|
||||
tab.status === "idle" &&
|
||||
transcriptIncludesAgentText(tab.transcript, expectedInitialReply)
|
||||
);
|
||||
|
|
@ -273,41 +273,41 @@ describe("e2e(client): workbench flows", () => {
|
|||
expect(findTab(initialCompleted, primaryTab.id).sessionId).toBeTruthy();
|
||||
expect(transcriptIncludesAgentText(findTab(initialCompleted, primaryTab.id).transcript, expectedInitialReply)).toBe(true);
|
||||
|
||||
const detail = await client.getHandoff(workspaceId, created.handoffId);
|
||||
const detail = await client.getTask(workspaceId, created.taskId);
|
||||
await seedSandboxFile(endpoint, detail, 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),
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => task.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
expect(fileSeeded.fileChanges.some((file) => file.path === expectedFile)).toBe(true);
|
||||
|
||||
await client.renameWorkbenchHandoff(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
await client.renameWorkbenchTask(workspaceId, {
|
||||
taskId: created.taskId,
|
||||
value: `Workbench E2E ${runId} Renamed`,
|
||||
});
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: primaryTab.id,
|
||||
title: "Primary Session",
|
||||
});
|
||||
|
||||
const secondTab = await client.createWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
model,
|
||||
});
|
||||
|
||||
await client.renameWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: secondTab.tabId,
|
||||
title: "Follow-up Session",
|
||||
});
|
||||
|
||||
await client.updateWorkbenchDraft(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [
|
||||
|
|
@ -320,12 +320,12 @@ describe("e2e(client): workbench flows", () => {
|
|||
],
|
||||
});
|
||||
|
||||
const drafted = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
const drafted = findTask(await client.getWorkbench(workspaceId), created.taskId);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.text).toContain(expectedReply);
|
||||
expect(findTab(drafted, secondTab.tabId).draft.attachments).toHaveLength(1);
|
||||
|
||||
await client.sendWorkbenchMessage(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: secondTab.tabId,
|
||||
text: `Reply with exactly: ${expectedReply}`,
|
||||
attachments: [],
|
||||
|
|
@ -335,9 +335,9 @@ describe("e2e(client): workbench flows", () => {
|
|||
"follow-up session response",
|
||||
10 * 60_000,
|
||||
2_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, secondTab.tabId);
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => {
|
||||
const tab = findTab(task, secondTab.tabId);
|
||||
return (
|
||||
tab.status === "idle" &&
|
||||
transcriptIncludesAgentText(tab.transcript, expectedReply)
|
||||
|
|
@ -349,17 +349,17 @@ describe("e2e(client): workbench flows", () => {
|
|||
expect(transcriptIncludesAgentText(secondTranscript, expectedReply)).toBe(true);
|
||||
|
||||
await client.setWorkbenchSessionUnread(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: secondTab.tabId,
|
||||
unread: false,
|
||||
});
|
||||
await client.markWorkbenchUnread(workspaceId, { handoffId: created.handoffId });
|
||||
await client.markWorkbenchUnread(workspaceId, { taskId: created.taskId });
|
||||
|
||||
const unreadSnapshot = findHandoff(await client.getWorkbench(workspaceId), created.handoffId);
|
||||
const unreadSnapshot = findTask(await client.getWorkbench(workspaceId), created.taskId);
|
||||
expect(unreadSnapshot.tabs.some((tab) => tab.unread)).toBe(true);
|
||||
|
||||
await client.closeWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: secondTab.tabId,
|
||||
});
|
||||
|
||||
|
|
@ -367,13 +367,13 @@ describe("e2e(client): workbench flows", () => {
|
|||
"secondary session closed",
|
||||
30_000,
|
||||
1_000,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => !handoff.tabs.some((tab) => tab.id === secondTab.tabId),
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => !task.tabs.some((tab) => tab.id === secondTab.tabId),
|
||||
);
|
||||
expect(closedSnapshot.tabs).toHaveLength(1);
|
||||
|
||||
await client.revertWorkbenchFile(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
path: expectedFile,
|
||||
});
|
||||
|
||||
|
|
@ -381,8 +381,8 @@ describe("e2e(client): workbench flows", () => {
|
|||
"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),
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => !task.fileChanges.some((file) => file.path === expectedFile),
|
||||
);
|
||||
|
||||
expect(revertedSnapshot.fileChanges.some((file) => file.path === expectedFile)).toBe(false);
|
||||
|
|
|
|||
|
|
@ -1,8 +1,8 @@
|
|||
import { describe, expect, it } from "vitest";
|
||||
import type {
|
||||
HandoffWorkbenchSnapshot,
|
||||
TaskWorkbenchSnapshot,
|
||||
WorkbenchAgentTab,
|
||||
WorkbenchHandoff,
|
||||
WorkbenchTask,
|
||||
WorkbenchModelId,
|
||||
WorkbenchTranscriptEvent,
|
||||
} from "@sandbox-agent/factory-shared";
|
||||
|
|
@ -70,18 +70,18 @@ async function poll<T>(
|
|||
}
|
||||
}
|
||||
|
||||
function findHandoff(snapshot: HandoffWorkbenchSnapshot, handoffId: string): WorkbenchHandoff {
|
||||
const handoff = snapshot.handoffs.find((candidate) => candidate.id === handoffId);
|
||||
if (!handoff) {
|
||||
throw new Error(`handoff ${handoffId} missing from snapshot`);
|
||||
function findTask(snapshot: TaskWorkbenchSnapshot, taskId: string): WorkbenchTask {
|
||||
const task = snapshot.tasks.find((candidate) => candidate.id === taskId);
|
||||
if (!task) {
|
||||
throw new Error(`task ${taskId} missing from snapshot`);
|
||||
}
|
||||
return handoff;
|
||||
return task;
|
||||
}
|
||||
|
||||
function findTab(handoff: WorkbenchHandoff, tabId: string): WorkbenchAgentTab {
|
||||
const tab = handoff.tabs.find((candidate) => candidate.id === tabId);
|
||||
function findTab(task: WorkbenchTask, tabId: string): WorkbenchAgentTab {
|
||||
const tab = task.tabs.find((candidate) => candidate.id === tabId);
|
||||
if (!tab) {
|
||||
throw new Error(`tab ${tabId} missing from handoff ${handoff.id}`);
|
||||
throw new Error(`tab ${tabId} missing from task ${task.id}`);
|
||||
}
|
||||
return tab;
|
||||
}
|
||||
|
|
@ -156,12 +156,12 @@ async function measureWorkbenchSnapshot(
|
|||
avgMs: number;
|
||||
maxMs: number;
|
||||
payloadBytes: number;
|
||||
handoffCount: number;
|
||||
taskCount: number;
|
||||
tabCount: number;
|
||||
transcriptEventCount: number;
|
||||
}> {
|
||||
const durations: number[] = [];
|
||||
let snapshot: HandoffWorkbenchSnapshot | null = null;
|
||||
let snapshot: TaskWorkbenchSnapshot | null = null;
|
||||
|
||||
for (let index = 0; index < iterations; index += 1) {
|
||||
const startedAt = performance.now();
|
||||
|
|
@ -173,13 +173,13 @@ async function measureWorkbenchSnapshot(
|
|||
workspaceId,
|
||||
repos: [],
|
||||
projects: [],
|
||||
handoffs: [],
|
||||
tasks: [],
|
||||
};
|
||||
const payloadBytes = Buffer.byteLength(JSON.stringify(finalSnapshot), "utf8");
|
||||
const tabCount = finalSnapshot.handoffs.reduce((sum, handoff) => sum + handoff.tabs.length, 0);
|
||||
const transcriptEventCount = finalSnapshot.handoffs.reduce(
|
||||
(sum, handoff) =>
|
||||
sum + handoff.tabs.reduce((tabSum, tab) => tabSum + tab.transcript.length, 0),
|
||||
const tabCount = finalSnapshot.tasks.reduce((sum, task) => sum + task.tabs.length, 0);
|
||||
const transcriptEventCount = finalSnapshot.tasks.reduce(
|
||||
(sum, task) =>
|
||||
sum + task.tabs.reduce((tabSum, tab) => tabSum + tab.transcript.length, 0),
|
||||
0,
|
||||
);
|
||||
|
||||
|
|
@ -187,7 +187,7 @@ async function measureWorkbenchSnapshot(
|
|||
avgMs: Math.round(average(durations)),
|
||||
maxMs: Math.round(Math.max(...durations, 0)),
|
||||
payloadBytes,
|
||||
handoffCount: finalSnapshot.handoffs.length,
|
||||
taskCount: finalSnapshot.tasks.length,
|
||||
tabCount,
|
||||
transcriptEventCount,
|
||||
};
|
||||
|
|
@ -202,7 +202,7 @@ describe("e2e(client): workbench load", () => {
|
|||
const workspaceId = process.env.HF_E2E_WORKSPACE?.trim() || "default";
|
||||
const repoRemote = requiredRepoRemote();
|
||||
const model = workbenchModelEnv("HF_E2E_MODEL", "gpt-4o");
|
||||
const handoffCount = intEnv("HF_LOAD_HANDOFF_COUNT", 3);
|
||||
const taskCount = intEnv("HF_LOAD_HANDOFF_COUNT", 3);
|
||||
const extraSessionCount = intEnv("HF_LOAD_EXTRA_SESSION_COUNT", 2);
|
||||
const pollIntervalMs = intEnv("HF_LOAD_POLL_INTERVAL_MS", 2_000);
|
||||
|
||||
|
|
@ -212,12 +212,12 @@ describe("e2e(client): workbench load", () => {
|
|||
});
|
||||
|
||||
const repo = await client.addRepo(workspaceId, repoRemote);
|
||||
const createHandoffLatencies: number[] = [];
|
||||
const createTaskLatencies: number[] = [];
|
||||
const provisionLatencies: number[] = [];
|
||||
const createSessionLatencies: number[] = [];
|
||||
const messageRoundTripLatencies: number[] = [];
|
||||
const snapshotSeries: Array<{
|
||||
handoffCount: number;
|
||||
taskCount: number;
|
||||
avgMs: number;
|
||||
maxMs: number;
|
||||
payloadBytes: number;
|
||||
|
|
@ -227,31 +227,31 @@ describe("e2e(client): workbench load", () => {
|
|||
|
||||
snapshotSeries.push(await measureWorkbenchSnapshot(client, workspaceId, 2));
|
||||
|
||||
for (let handoffIndex = 0; handoffIndex < handoffCount; handoffIndex += 1) {
|
||||
const runId = `load-${handoffIndex}-${Date.now().toString(36)}`;
|
||||
for (let taskIndex = 0; taskIndex < taskCount; taskIndex += 1) {
|
||||
const runId = `load-${taskIndex}-${Date.now().toString(36)}`;
|
||||
const initialReply = `LOAD_INIT_${runId}`;
|
||||
|
||||
const createStartedAt = performance.now();
|
||||
const created = await client.createWorkbenchHandoff(workspaceId, {
|
||||
const created = await client.createWorkbenchTask(workspaceId, {
|
||||
repoId: repo.repoId,
|
||||
title: `Workbench Load ${runId}`,
|
||||
branch: `load/${runId}`,
|
||||
model,
|
||||
task: `Reply with exactly: ${initialReply}`,
|
||||
});
|
||||
createHandoffLatencies.push(performance.now() - createStartedAt);
|
||||
createTaskLatencies.push(performance.now() - createStartedAt);
|
||||
|
||||
const provisionStartedAt = performance.now();
|
||||
const provisioned = await poll(
|
||||
`handoff ${runId} provisioning`,
|
||||
`task ${runId} provisioning`,
|
||||
12 * 60_000,
|
||||
pollIntervalMs,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = handoff.tabs[0];
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => {
|
||||
const tab = task.tabs[0];
|
||||
return Boolean(
|
||||
tab &&
|
||||
handoff.status === "idle" &&
|
||||
task.status === "idle" &&
|
||||
tab.status === "idle" &&
|
||||
transcriptIncludesAgentText(tab.transcript, initialReply),
|
||||
);
|
||||
|
|
@ -267,13 +267,13 @@ describe("e2e(client): workbench load", () => {
|
|||
const expectedReply = `LOAD_REPLY_${runId}_${sessionIndex}`;
|
||||
const createSessionStartedAt = performance.now();
|
||||
const createdSession = await client.createWorkbenchSession(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
model,
|
||||
});
|
||||
createSessionLatencies.push(performance.now() - createSessionStartedAt);
|
||||
|
||||
await client.sendWorkbenchMessage(workspaceId, {
|
||||
handoffId: created.handoffId,
|
||||
taskId: created.taskId,
|
||||
tabId: createdSession.tabId,
|
||||
text: `Run pwd in the repo, then reply with exactly: ${expectedReply}`,
|
||||
attachments: [],
|
||||
|
|
@ -281,12 +281,12 @@ describe("e2e(client): workbench load", () => {
|
|||
|
||||
const messageStartedAt = performance.now();
|
||||
const withReply = await poll(
|
||||
`handoff ${runId} session ${sessionIndex} reply`,
|
||||
`task ${runId} session ${sessionIndex} reply`,
|
||||
10 * 60_000,
|
||||
pollIntervalMs,
|
||||
async () => findHandoff(await client.getWorkbench(workspaceId), created.handoffId),
|
||||
(handoff) => {
|
||||
const tab = findTab(handoff, createdSession.tabId);
|
||||
async () => findTask(await client.getWorkbench(workspaceId), created.taskId),
|
||||
(task) => {
|
||||
const tab = findTab(task, createdSession.tabId);
|
||||
return tab.status === "idle" && transcriptIncludesAgentText(tab.transcript, expectedReply);
|
||||
},
|
||||
);
|
||||
|
|
@ -300,7 +300,7 @@ describe("e2e(client): workbench load", () => {
|
|||
console.info(
|
||||
"[workbench-load-snapshot]",
|
||||
JSON.stringify({
|
||||
handoffIndex: handoffIndex + 1,
|
||||
taskIndex: taskIndex + 1,
|
||||
...snapshotMetrics,
|
||||
}),
|
||||
);
|
||||
|
|
@ -309,9 +309,9 @@ describe("e2e(client): workbench load", () => {
|
|||
const firstSnapshot = snapshotSeries[0]!;
|
||||
const lastSnapshot = snapshotSeries[snapshotSeries.length - 1]!;
|
||||
const summary = {
|
||||
handoffCount,
|
||||
taskCount,
|
||||
extraSessionCount,
|
||||
createHandoffAvgMs: Math.round(average(createHandoffLatencies)),
|
||||
createTaskAvgMs: Math.round(average(createTaskLatencies)),
|
||||
provisionAvgMs: Math.round(average(provisionLatencies)),
|
||||
createSessionAvgMs: Math.round(average(createSessionLatencies)),
|
||||
messageRoundTripAvgMs: Math.round(average(messageRoundTripLatencies)),
|
||||
|
|
@ -326,10 +326,10 @@ describe("e2e(client): workbench load", () => {
|
|||
|
||||
console.info("[workbench-load-summary]", JSON.stringify(summary));
|
||||
|
||||
expect(createHandoffLatencies.length).toBe(handoffCount);
|
||||
expect(provisionLatencies.length).toBe(handoffCount);
|
||||
expect(createSessionLatencies.length).toBe(handoffCount * extraSessionCount);
|
||||
expect(messageRoundTripLatencies.length).toBe(handoffCount * extraSessionCount);
|
||||
expect(createTaskLatencies.length).toBe(taskCount);
|
||||
expect(provisionLatencies.length).toBe(taskCount);
|
||||
expect(createSessionLatencies.length).toBe(taskCount * extraSessionCount);
|
||||
expect(messageRoundTripLatencies.length).toBe(taskCount * extraSessionCount);
|
||||
},
|
||||
);
|
||||
});
|
||||
|
|
|
|||
|
|
@ -1,11 +1,11 @@
|
|||
import { describe, expect, it } from "vitest";
|
||||
import {
|
||||
handoffKey,
|
||||
handoffStatusSyncKey,
|
||||
taskKey,
|
||||
taskStatusSyncKey,
|
||||
historyKey,
|
||||
projectBranchSyncKey,
|
||||
projectKey,
|
||||
projectPrSyncKey,
|
||||
repoBranchSyncKey,
|
||||
repoKey,
|
||||
repoPrSyncKey,
|
||||
sandboxInstanceKey,
|
||||
workspaceKey
|
||||
} from "../src/keys.js";
|
||||
|
|
@ -14,13 +14,13 @@ describe("actor keys", () => {
|
|||
it("prefixes every key with workspace namespace", () => {
|
||||
const keys = [
|
||||
workspaceKey("default"),
|
||||
projectKey("default", "repo"),
|
||||
handoffKey("default", "repo", "handoff"),
|
||||
repoKey("default", "repo"),
|
||||
taskKey("default", "task"),
|
||||
sandboxInstanceKey("default", "daytona", "sbx"),
|
||||
historyKey("default", "repo"),
|
||||
projectPrSyncKey("default", "repo"),
|
||||
projectBranchSyncKey("default", "repo"),
|
||||
handoffStatusSyncKey("default", "repo", "handoff", "sandbox-1", "session-1")
|
||||
repoPrSyncKey("default", "repo"),
|
||||
repoBranchSyncKey("default", "repo"),
|
||||
taskStatusSyncKey("default", "repo", "task", "sandbox-1", "session-1")
|
||||
];
|
||||
|
||||
for (const key of keys) {
|
||||
|
|
|
|||
|
|
@ -1,17 +1,17 @@
|
|||
import { describe, expect, it } from "vitest";
|
||||
import type { HandoffRecord } from "@sandbox-agent/factory-shared";
|
||||
import type { TaskRecord } from "@sandbox-agent/factory-shared";
|
||||
import {
|
||||
filterHandoffs,
|
||||
filterTasks,
|
||||
formatRelativeAge,
|
||||
fuzzyMatch,
|
||||
summarizeHandoffs
|
||||
summarizeTasks
|
||||
} from "../src/view-model.js";
|
||||
|
||||
const sample: HandoffRecord = {
|
||||
const sample: TaskRecord = {
|
||||
workspaceId: "default",
|
||||
repoId: "repo-a",
|
||||
repoRemote: "https://example.com/repo-a.git",
|
||||
handoffId: "handoff-1",
|
||||
taskId: "task-1",
|
||||
branchName: "feature/test",
|
||||
title: "Test Title",
|
||||
task: "Do test",
|
||||
|
|
@ -53,19 +53,19 @@ describe("search helpers", () => {
|
|||
});
|
||||
|
||||
it("filters rows across branch and title", () => {
|
||||
const rows: HandoffRecord[] = [
|
||||
const rows: TaskRecord[] = [
|
||||
sample,
|
||||
{
|
||||
...sample,
|
||||
handoffId: "handoff-2",
|
||||
taskId: "task-2",
|
||||
branchName: "docs/update-intro",
|
||||
title: "Docs Intro Refresh",
|
||||
status: "idle"
|
||||
}
|
||||
];
|
||||
expect(filterHandoffs(rows, "doc")).toHaveLength(1);
|
||||
expect(filterHandoffs(rows, "h2")).toHaveLength(1);
|
||||
expect(filterHandoffs(rows, "test")).toHaveLength(2);
|
||||
expect(filterTasks(rows, "doc")).toHaveLength(1);
|
||||
expect(filterTasks(rows, "h2")).toHaveLength(1);
|
||||
expect(filterTasks(rows, "test")).toHaveLength(2);
|
||||
});
|
||||
});
|
||||
|
||||
|
|
@ -76,13 +76,13 @@ describe("summary helpers", () => {
|
|||
});
|
||||
|
||||
it("summarizes by status and provider", () => {
|
||||
const rows: HandoffRecord[] = [
|
||||
const rows: TaskRecord[] = [
|
||||
sample,
|
||||
{ ...sample, handoffId: "handoff-2", status: "idle", providerId: "daytona" },
|
||||
{ ...sample, handoffId: "handoff-3", status: "error", providerId: "daytona" }
|
||||
{ ...sample, taskId: "task-2", status: "idle", providerId: "daytona" },
|
||||
{ ...sample, taskId: "task-3", status: "error", providerId: "daytona" }
|
||||
];
|
||||
|
||||
const summary = summarizeHandoffs(rows);
|
||||
const summary = summarizeTasks(rows);
|
||||
expect(summary.total).toBe(3);
|
||||
expect(summary.byStatus.running).toBe(1);
|
||||
expect(summary.byStatus.idle).toBe(1);
|
||||
|
|
|
|||
|
|
@ -1,18 +1,18 @@
|
|||
import { describe, expect, it } from "vitest";
|
||||
import type { BackendClient } from "../src/backend-client.js";
|
||||
import { createHandoffWorkbenchClient } from "../src/workbench-client.js";
|
||||
import { createTaskWorkbenchClient } from "../src/workbench-client.js";
|
||||
|
||||
async function sleep(ms: number): Promise<void> {
|
||||
await new Promise((resolve) => setTimeout(resolve, ms));
|
||||
}
|
||||
|
||||
describe("createHandoffWorkbenchClient", () => {
|
||||
describe("createTaskWorkbenchClient", () => {
|
||||
it("scopes mock clients by workspace", async () => {
|
||||
const alpha = createHandoffWorkbenchClient({
|
||||
const alpha = createTaskWorkbenchClient({
|
||||
mode: "mock",
|
||||
workspaceId: "mock-alpha",
|
||||
});
|
||||
const beta = createHandoffWorkbenchClient({
|
||||
const beta = createTaskWorkbenchClient({
|
||||
mode: "mock",
|
||||
workspaceId: "mock-beta",
|
||||
});
|
||||
|
|
@ -22,24 +22,24 @@ describe("createHandoffWorkbenchClient", () => {
|
|||
expect(alphaInitial.workspaceId).toBe("mock-alpha");
|
||||
expect(betaInitial.workspaceId).toBe("mock-beta");
|
||||
|
||||
await alpha.createHandoff({
|
||||
await alpha.createTask({
|
||||
repoId: alphaInitial.repos[0]!.id,
|
||||
task: "Ship alpha-only change",
|
||||
title: "Alpha only",
|
||||
});
|
||||
|
||||
expect(alpha.getSnapshot().handoffs).toHaveLength(alphaInitial.handoffs.length + 1);
|
||||
expect(beta.getSnapshot().handoffs).toHaveLength(betaInitial.handoffs.length);
|
||||
expect(alpha.getSnapshot().tasks).toHaveLength(alphaInitial.tasks.length + 1);
|
||||
expect(beta.getSnapshot().tasks).toHaveLength(betaInitial.tasks.length);
|
||||
});
|
||||
|
||||
it("uses the initial task to bootstrap a new mock handoff session", async () => {
|
||||
const client = createHandoffWorkbenchClient({
|
||||
it("uses the initial task to bootstrap a new mock task session", async () => {
|
||||
const client = createTaskWorkbenchClient({
|
||||
mode: "mock",
|
||||
workspaceId: "mock-onboarding",
|
||||
});
|
||||
const snapshot = client.getSnapshot();
|
||||
|
||||
const created = await client.createHandoff({
|
||||
const created = await client.createTask({
|
||||
repoId: snapshot.repos[0]!.id,
|
||||
task: "Reply with exactly: MOCK_WORKBENCH_READY",
|
||||
title: "Mock onboarding",
|
||||
|
|
@ -47,22 +47,22 @@ describe("createHandoffWorkbenchClient", () => {
|
|||
model: "gpt-4o",
|
||||
});
|
||||
|
||||
const runningHandoff = client.getSnapshot().handoffs.find((handoff) => handoff.id === created.handoffId);
|
||||
expect(runningHandoff).toEqual(
|
||||
const runningTask = client.getSnapshot().tasks.find((task) => task.id === created.taskId);
|
||||
expect(runningTask).toEqual(
|
||||
expect.objectContaining({
|
||||
title: "Mock onboarding",
|
||||
branch: "feat/mock-onboarding",
|
||||
status: "running",
|
||||
}),
|
||||
);
|
||||
expect(runningHandoff?.tabs[0]).toEqual(
|
||||
expect(runningTask?.tabs[0]).toEqual(
|
||||
expect.objectContaining({
|
||||
id: created.tabId,
|
||||
created: true,
|
||||
status: "running",
|
||||
}),
|
||||
);
|
||||
expect(runningHandoff?.tabs[0]?.transcript).toEqual([
|
||||
expect(runningTask?.tabs[0]?.transcript).toEqual([
|
||||
expect.objectContaining({
|
||||
sender: "client",
|
||||
payload: expect.objectContaining({
|
||||
|
|
@ -73,26 +73,26 @@ describe("createHandoffWorkbenchClient", () => {
|
|||
|
||||
await sleep(2_700);
|
||||
|
||||
const completedHandoff = client.getSnapshot().handoffs.find((handoff) => handoff.id === created.handoffId);
|
||||
expect(completedHandoff?.status).toBe("idle");
|
||||
expect(completedHandoff?.tabs[0]).toEqual(
|
||||
const completedTask = client.getSnapshot().tasks.find((task) => task.id === created.taskId);
|
||||
expect(completedTask?.status).toBe("idle");
|
||||
expect(completedTask?.tabs[0]).toEqual(
|
||||
expect.objectContaining({
|
||||
status: "idle",
|
||||
unread: true,
|
||||
}),
|
||||
);
|
||||
expect(completedHandoff?.tabs[0]?.transcript).toEqual([
|
||||
expect(completedTask?.tabs[0]?.transcript).toEqual([
|
||||
expect.objectContaining({ sender: "client" }),
|
||||
expect.objectContaining({ sender: "agent" }),
|
||||
]);
|
||||
});
|
||||
|
||||
it("routes remote push actions through the backend boundary", async () => {
|
||||
const actions: Array<{ workspaceId: string; handoffId: string; action: string }> = [];
|
||||
const actions: Array<{ workspaceId: string; taskId: string; action: string }> = [];
|
||||
let snapshotReads = 0;
|
||||
const backend = {
|
||||
async runAction(workspaceId: string, handoffId: string, action: string): Promise<void> {
|
||||
actions.push({ workspaceId, handoffId, action });
|
||||
async runAction(workspaceId: string, taskId: string, action: string): Promise<void> {
|
||||
actions.push({ workspaceId, taskId, action });
|
||||
},
|
||||
async getWorkbench(workspaceId: string) {
|
||||
snapshotReads += 1;
|
||||
|
|
@ -100,7 +100,7 @@ describe("createHandoffWorkbenchClient", () => {
|
|||
workspaceId,
|
||||
repos: [],
|
||||
projects: [],
|
||||
handoffs: [],
|
||||
tasks: [],
|
||||
};
|
||||
},
|
||||
subscribeWorkbench(): () => void {
|
||||
|
|
@ -108,18 +108,18 @@ describe("createHandoffWorkbenchClient", () => {
|
|||
},
|
||||
} as unknown as BackendClient;
|
||||
|
||||
const client = createHandoffWorkbenchClient({
|
||||
const client = createTaskWorkbenchClient({
|
||||
mode: "remote",
|
||||
backend,
|
||||
workspaceId: "remote-ws",
|
||||
});
|
||||
|
||||
await client.pushHandoff({ handoffId: "handoff-123" });
|
||||
await client.pushTask({ taskId: "task-123" });
|
||||
|
||||
expect(actions).toEqual([
|
||||
{
|
||||
workspaceId: "remote-ws",
|
||||
handoffId: "handoff-123",
|
||||
taskId: "task-123",
|
||||
action: "push",
|
||||
},
|
||||
]);
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue