test(app): add a golden path for mocked e2e prompts (#20593)

6 files changed, 111 insertions, 240 deletions

diff --git a/packages/app/e2e/prompt/mock.ts b/packages/app/e2e/prompt/mock.ts
index bd09af266..c7eb54b52 100644
--- a/packages/app/e2e/prompt/mock.ts
+++ b/packages/app/e2e/prompt/mock.ts
@@ -1,21 +1,9 @@
-import { createSdk } from "../utils"
-
-export const openaiModel = { providerID: "openai", modelID: "gpt-5.3-chat-latest" }
-
 type Hit = { body: Record<string, unknown> }
 
 export function bodyText(hit: Hit) {
   return JSON.stringify(hit.body)
 }
 
-export function titleMatch(hit: Hit) {
-  return bodyText(hit).includes("Generate a title for this conversation")
-}
-
-export function promptMatch(token: string) {
-  return (hit: Hit) => bodyText(hit).includes(token)
-}
-
 /**
  * Match requests whose body contains the exact serialized tool input.
  * The seed prompts embed JSON.stringify(input) in the prompt text, which
@@ -25,32 +13,3 @@ export function inputMatch(input: unknown) {
   const escaped = JSON.stringify(JSON.stringify(input)).slice(1, -1)
   return (hit: Hit) => bodyText(hit).includes(escaped)
 }
-
-export async function withMockOpenAI<T>(input: { serverUrl: string; llmUrl: string; fn: () => Promise<T> }) {
-  const sdk = createSdk(undefined, input.serverUrl)
-  const prev = await sdk.global.config.get().then((res) => res.data ?? {})
-
-  try {
-    await sdk.global.config.update({
-      config: {
-        ...prev,
-        model: `${openaiModel.providerID}/${openaiModel.modelID}`,
-        enabled_providers: ["openai"],
-        provider: {
-          ...prev.provider,
-          openai: {
-            ...prev.provider?.openai,
-            options: {
-              ...prev.provider?.openai?.options,
-              apiKey: "test-key",
-              baseURL: input.llmUrl,
-            },
-          },
-        },
-      },
-    })
-    return await input.fn()
-  } finally {
-    await sdk.global.config.update({ config: prev })
-  }
-}
diff --git a/packages/app/e2e/prompt/prompt-async.spec.ts b/packages/app/e2e/prompt/prompt-async.spec.ts
index a9a12cb95..403369947 100644
--- a/packages/app/e2e/prompt/prompt-async.spec.ts
+++ b/packages/app/e2e/prompt/prompt-async.spec.ts
@@ -1,52 +1,25 @@
 import { test, expect } from "../fixtures"
 import { promptSelector } from "../selectors"
-import { assistantText, sessionIDFromUrl, withSession } from "../actions"
-import { openaiModel, promptMatch, titleMatch, withMockOpenAI } from "./mock"
+import { assistantText, withSession } from "../actions"
 
 const text = (value: string | null) => (value ?? "").replace(/\u200B/g, "").trim()
 
 // Regression test for Issue #12453: the synchronous POST /message endpoint holds
 // the connection open while the agent works, causing "Failed to fetch" over
 // VPN/Tailscale. The fix switches to POST /prompt_async which returns immediately.
-test("prompt succeeds when sync message endpoint is unreachable", async ({
-  page,
-  llm,
-  backend,
-  withBackendProject,
-}) => {
+test("prompt succeeds when sync message endpoint is unreachable", async ({ page, project, assistant }) => {
   test.setTimeout(120_000)
 
   // Simulate Tailscale/VPN killing the long-lived sync connection
   await page.route("**/session/*/message", (route) => route.abort("connectionfailed"))
 
-  await withMockOpenAI({
-    serverUrl: backend.url,
-    llmUrl: llm.url,
-    fn: async () => {
-      const token = `E2E_ASYNC_${Date.now()}`
-      await llm.textMatch(titleMatch, "E2E Title")
-      await llm.textMatch(promptMatch(token), token)
+  const token = `E2E_ASYNC_${Date.now()}`
+  await project.open()
+  await assistant.reply(token)
+  const sessionID = await project.prompt(`Reply with exactly: ${token}`)
 
-      await withBackendProject(
-        async (project) => {
-          await page.locator(promptSelector).click()
-          await page.keyboard.type(`Reply with exactly: ${token}`)
-          await page.keyboard.press("Enter")
-
-          await expect(page).toHaveURL(/\/session\/[^/?#]+/, { timeout: 30_000 })
-          const sessionID = sessionIDFromUrl(page.url())!
-          project.trackSession(sessionID)
-
-          await expect.poll(() => llm.calls()).toBeGreaterThanOrEqual(1)
-
-          await expect.poll(() => assistantText(project.sdk, sessionID), { timeout: 90_000 }).toContain(token)
-        },
-        {
-          model: openaiModel,
-        },
-      )
-    },
-  })
+  await expect.poll(() => assistant.calls()).toBeGreaterThanOrEqual(1)
+  await expect.poll(() => assistantText(project.sdk, sessionID), { timeout: 90_000 }).toContain(token)
 })
 
 test("failed prompt send restores the composer input", async ({ page, sdk, gotoSession }) => {
diff --git a/packages/app/e2e/prompt/prompt-history.spec.ts b/packages/app/e2e/prompt/prompt-history.spec.ts
index f2d15914d..55cb0c9aa 100644
--- a/packages/app/e2e/prompt/prompt-history.spec.ts
+++ b/packages/app/e2e/prompt/prompt-history.spec.ts
@@ -1,10 +1,9 @@
 import type { ToolPart } from "@opencode-ai/sdk/v2/client"
 import type { Page } from "@playwright/test"
 import { test, expect } from "../fixtures"
-import { assistantText, sessionIDFromUrl } from "../actions"
+import { assistantText } from "../actions"
 import { promptSelector } from "../selectors"
 import { createSdk } from "../utils"
-import { openaiModel, promptMatch, titleMatch, withMockOpenAI } from "./mock"
 
 const text = (value: string | null) => (value ?? "").replace(/\u200B/g, "").trim()
 type Sdk = ReturnType<typeof createSdk>
@@ -43,73 +42,45 @@ async function shell(sdk: Sdk, sessionID: string, cmd: string, token: string) {
     .toContain(token)
 }
 
-test("prompt history restores unsent draft with arrow navigation", async ({
-  page,
-  llm,
-  backend,
-  withBackendProject,
-}) => {
+test("prompt history restores unsent draft with arrow navigation", async ({ page, project, assistant }) => {
   test.setTimeout(120_000)
 
-  await withMockOpenAI({
-    serverUrl: backend.url,
-    llmUrl: llm.url,
-    fn: async () => {
-      const firstToken = `E2E_HISTORY_ONE_${Date.now()}`
-      const secondToken = `E2E_HISTORY_TWO_${Date.now()}`
-      const first = `Reply with exactly: ${firstToken}`
-      const second = `Reply with exactly: ${secondToken}`
-      const draft = `draft ${Date.now()}`
-
-      await llm.textMatch(titleMatch, "E2E Title")
-      await llm.textMatch(promptMatch(firstToken), firstToken)
-      await llm.textMatch(promptMatch(secondToken), secondToken)
-
-      await withBackendProject(
-        async (project) => {
-          const prompt = page.locator(promptSelector)
-
-          await prompt.click()
-          await page.keyboard.type(first)
-          await page.keyboard.press("Enter")
-          await wait(page, "")
-
-          await expect(page).toHaveURL(/\/session\/[^/?#]+/, { timeout: 30_000 })
-          const sessionID = sessionIDFromUrl(page.url())!
-          project.trackSession(sessionID)
-          await reply(project.sdk, sessionID, firstToken)
-
-          await prompt.click()
-          await page.keyboard.type(second)
-          await page.keyboard.press("Enter")
-          await wait(page, "")
-          await reply(project.sdk, sessionID, secondToken)
-
-          await prompt.click()
-          await page.keyboard.type(draft)
-          await wait(page, draft)
-
-          await prompt.fill("")
-          await wait(page, "")
-
-          await page.keyboard.press("ArrowUp")
-          await wait(page, second)
-
-          await page.keyboard.press("ArrowUp")
-          await wait(page, first)
-
-          await page.keyboard.press("ArrowDown")
-          await wait(page, second)
-
-          await page.keyboard.press("ArrowDown")
-          await wait(page, "")
-        },
-        {
-          model: openaiModel,
-        },
-      )
-    },
-  })
+  const firstToken = `E2E_HISTORY_ONE_${Date.now()}`
+  const secondToken = `E2E_HISTORY_TWO_${Date.now()}`
+  const first = `Reply with exactly: ${firstToken}`
+  const second = `Reply with exactly: ${secondToken}`
+  const draft = `draft ${Date.now()}`
+
+  await project.open()
+  await assistant.reply(firstToken)
+  const sessionID = await project.prompt(first)
+  await wait(page, "")
+  await reply(project.sdk, sessionID, firstToken)
+
+  await assistant.reply(secondToken)
+  await project.prompt(second)
+  await wait(page, "")
+  await reply(project.sdk, sessionID, secondToken)
+
+  const prompt = page.locator(promptSelector)
+  await prompt.click()
+  await page.keyboard.type(draft)
+  await wait(page, draft)
+
+  await prompt.fill("")
+  await wait(page, "")
+
+  await page.keyboard.press("ArrowUp")
+  await wait(page, second)
+
+  await page.keyboard.press("ArrowUp")
+  await wait(page, first)
+
+  await page.keyboard.press("ArrowDown")
+  await wait(page, second)
+
+  await page.keyboard.press("ArrowDown")
+  await wait(page, "")
 })
 
 test.fixme("shell history stays separate from normal prompt history", async ({ page, sdk, gotoSession }) => {
diff --git a/packages/app/e2e/prompt/prompt-shell.spec.ts b/packages/app/e2e/prompt/prompt-shell.spec.ts
index 7c39a2db3..d81f1d4c4 100644
--- a/packages/app/e2e/prompt/prompt-shell.spec.ts
+++ b/packages/app/e2e/prompt/prompt-shell.spec.ts
@@ -1,7 +1,6 @@
 import type { ToolPart } from "@opencode-ai/sdk/v2/client"
 import { test, expect } from "../fixtures"
-import { sessionIDFromUrl } from "../actions"
-import { promptSelector } from "../selectors"
+import { withSession } from "../actions"
 
 const isBash = (part: unknown): part is ToolPart => {
   if (!part || typeof part !== "object") return false
@@ -10,33 +9,35 @@ const isBash = (part: unknown): part is ToolPart => {
   return "state" in part
 }
 
-test("shell mode runs a command in the project directory", async ({ page, withBackendProject }) => {
-  test.setTimeout(120_000)
-
-  await withBackendProject(async ({ directory, gotoSession, trackSession, sdk }) => {
-    const prompt = page.locator(promptSelector)
-    const cmd = process.platform === "win32" ? "dir" : "command ls"
-
-    await gotoSession()
-    await prompt.click()
-    await page.keyboard.type("!")
-    await expect(prompt).toHaveAttribute("aria-label", /enter shell command/i)
+async function setAutoAccept(page: Parameters<typeof test>[0]["page"], enabled: boolean) {
+  const button = page.locator('[data-action="prompt-permissions"]').first()
+  await expect(button).toBeVisible()
+  const pressed = (await button.getAttribute("aria-pressed")) === "true"
+  if (pressed === enabled) return
+  await button.click()
+  await expect(button).toHaveAttribute("aria-pressed", enabled ? "true" : "false")
+}
 
-    await page.keyboard.type(cmd)
-    await page.keyboard.press("Enter")
+test("shell mode runs a command in the project directory", async ({ page, project }) => {
+  test.setTimeout(120_000)
 
-    await expect(page).toHaveURL(/\/session\/[^/?#]+/, { timeout: 30_000 })
+  await project.open()
+  const cmd = process.platform === "win32" ? "dir" : "command ls"
 
-    const id = sessionIDFromUrl(page.url())
-    if (!id) throw new Error(`Failed to parse session id from url: ${page.url()}`)
-    trackSession(id, directory)
+  await withSession(project.sdk, `e2e shell ${Date.now()}`, async (session) => {
+    project.trackSession(session.id)
+    await project.gotoSession(session.id)
+    await setAutoAccept(page, true)
+    await project.shell(cmd)
 
     await expect
       .poll(
         async () => {
-          const list = await sdk.session.messages({ sessionID: id, limit: 50 }).then((x) => x.data ?? [])
+          const list = await project.sdk.session
+            .messages({ sessionID: session.id, limit: 50 })
+            .then((x) => x.data ?? [])
           const msg = list.findLast(
-            (item) => item.info.role === "assistant" && "path" in item.info && item.info.path.cwd === directory,
+            (item) => item.info.role === "assistant" && "path" in item.info && item.info.path.cwd === project.directory,
           )
           if (!msg) return
 
@@ -49,12 +50,10 @@ test("shell mode runs a command in the project directory", async ({ page, withBa
             typeof part.state.metadata?.output === "string" ? part.state.metadata.output : part.state.output
           if (!output.includes("README.md")) return
 
-          return { cwd: directory, output }
+          return { cwd: project.directory, output }
         },
         { timeout: 90_000 },
       )
-      .toEqual(expect.objectContaining({ cwd: directory, output: expect.stringContaining("README.md") }))
-
-    await expect(prompt).toHaveText("")
+      .toEqual(expect.objectContaining({ cwd: project.directory, output: expect.stringContaining("README.md") }))
   })
 })
diff --git a/packages/app/e2e/prompt/prompt-slash-share.spec.ts b/packages/app/e2e/prompt/prompt-slash-share.spec.ts
index 5371d8a91..f3eeceee5 100644
--- a/packages/app/e2e/prompt/prompt-slash-share.spec.ts
+++ b/packages/app/e2e/prompt/prompt-slash-share.spec.ts
@@ -22,46 +22,45 @@ async function seed(sdk: Parameters<typeof withSession>[0], sessionID: string) {
     .toBeGreaterThan(0)
 }
 
-test("/share and /unshare update session share state", async ({ page, withBackendProject }) => {
+test("/share and /unshare update session share state", async ({ page, project }) => {
   test.skip(shareDisabled, "Share is disabled in this environment (OPENCODE_DISABLE_SHARE).")
 
-  await withBackendProject(async (project) => {
-    await withSession(project.sdk, `e2e slash share ${Date.now()}`, async (session) => {
-      project.trackSession(session.id)
-      const prompt = page.locator(promptSelector)
+  await project.open()
+  await withSession(project.sdk, `e2e slash share ${Date.now()}`, async (session) => {
+    project.trackSession(session.id)
+    const prompt = page.locator(promptSelector)
 
-      await seed(project.sdk, session.id)
-      await project.gotoSession(session.id)
+    await seed(project.sdk, session.id)
+    await project.gotoSession(session.id)
 
-      await prompt.click()
-      await page.keyboard.type("/share")
-      await expect(page.locator('[data-slash-id="session.share"]').first()).toBeVisible()
-      await page.keyboard.press("Enter")
+    await prompt.click()
+    await page.keyboard.type("/share")
+    await expect(page.locator('[data-slash-id="session.share"]').first()).toBeVisible()
+    await page.keyboard.press("Enter")
 
-      await expect
-        .poll(
-          async () => {
-            const data = await project.sdk.session.get({ sessionID: session.id }).then((r) => r.data)
-            return data?.share?.url || undefined
-          },
-          { timeout: 30_000 },
-        )
-        .not.toBeUndefined()
+    await expect
+      .poll(
+        async () => {
+          const data = await project.sdk.session.get({ sessionID: session.id }).then((r) => r.data)
+          return data?.share?.url || undefined
+        },
+        { timeout: 30_000 },
+      )
+      .not.toBeUndefined()
 
-      await prompt.click()
-      await page.keyboard.type("/unshare")
-      await expect(page.locator('[data-slash-id="session.unshare"]').first()).toBeVisible()
-      await page.keyboard.press("Enter")
+    await prompt.click()
+    await page.keyboard.type("/unshare")
+    await expect(page.locator('[data-slash-id="session.unshare"]').first()).toBeVisible()
+    await page.keyboard.press("Enter")
 
-      await expect
-        .poll(
-          async () => {
-            const data = await project.sdk.session.get({ sessionID: session.id }).then((r) => r.data)
-            return data?.share?.url || undefined
-          },
-          { timeout: 30_000 },
-        )
-        .toBeUndefined()
-    })
+    await expect
+      .poll(
+        async () => {
+          const data = await project.sdk.session.get({ sessionID: session.id }).then((r) => r.data)
+          return data?.share?.url || undefined
+        },
+        { timeout: 30_000 },
+      )
+      .toBeUndefined()
   })
 })
diff --git a/packages/app/e2e/prompt/prompt.spec.ts b/packages/app/e2e/prompt/prompt.spec.ts
index 3c9ed51dc..b5dc02bad 100644
--- a/packages/app/e2e/prompt/prompt.spec.ts
+++ b/packages/app/e2e/prompt/prompt.spec.ts
@@ -1,9 +1,7 @@
 import { test, expect } from "../fixtures"
-import { promptSelector } from "../selectors"
-import { assistantText, sessionIDFromUrl } from "../actions"
-import { openaiModel, promptMatch, titleMatch, withMockOpenAI } from "./mock"
+import { assistantText } from "../actions"
 
-test("can send a prompt and receive a reply", async ({ page, llm, backend, withBackendProject }) => {
+test("can send a prompt and receive a reply", async ({ page, project, assistant }) => {
   test.setTimeout(120_000)
 
   const pageErrors: string[] = []
@@ -13,41 +11,13 @@ test("can send a prompt and receive a reply", async ({ page, llm, backend, withB
   page.on("pageerror", onPageError)
 
   try {
-    await withMockOpenAI({
-      serverUrl: backend.url,
-      llmUrl: llm.url,
-      fn: async () => {
-        const token = `E2E_OK_${Date.now()}`
+    const token = `E2E_OK_${Date.now()}`
+    await project.open()
+    await assistant.reply(token)
+    const sessionID = await project.prompt(`Reply with exactly: ${token}`)
 
-        await llm.textMatch(titleMatch, "E2E Title")
-        await llm.textMatch(promptMatch(token), token)
-
-        await withBackendProject(
-          async (project) => {
-            const prompt = page.locator(promptSelector)
-            await prompt.click()
-            await page.keyboard.type(`Reply with exactly: ${token}`)
-            await page.keyboard.press("Enter")
-
-            await expect(page).toHaveURL(/\/session\/[^/?#]+/, { timeout: 30_000 })
-
-            const sessionID = (() => {
-              const id = sessionIDFromUrl(page.url())
-              if (!id) throw new Error(`Failed to parse session id from url: ${page.url()}`)
-              return id
-            })()
-            project.trackSession(sessionID)
-
-            await expect.poll(() => llm.calls()).toBeGreaterThanOrEqual(1)
-
-            await expect.poll(() => assistantText(project.sdk, sessionID), { timeout: 30_000 }).toContain(token)
-          },
-          {
-            model: openaiModel,
-          },
-        )
-      },
-    })
+    await expect.poll(() => assistant.calls()).toBeGreaterThanOrEqual(1)
+    await expect.poll(() => assistantText(project.sdk, sessionID), { timeout: 30_000 }).toContain(token)
   } finally {
     page.off("pageerror", onPageError)
   }