Re-point POST /api/chat/generate onto runAgentWorkflow + ephemeral key (chat#1813)

app/api/chat/runs/[runId]/route.ts

@@ -0,0 +1,42 @@
+import type { NextRequest } from "next/server";
+import { NextResponse } from "next/server";
+import { getCorsHeaders } from "@/lib/networking/getCorsHeaders";
+import { handleChatRunStatus } from "@/lib/chat/runs/handleChatRunStatus";
+
+/**
+ * OPTIONS handler for CORS preflight requests.
+ *
+ * @returns A NextResponse with CORS headers.
+ */
+export async function OPTIONS() {
+  return new NextResponse(null, {
+    status: 200,
+    headers: getCorsHeaders(),
+  });
+}
+
+/**
+ * GET /api/chat/runs/{runId}
+ *
+ * Point-in-time status of an asynchronous run started via `POST /api/chat/runs`
+ * (recoupable/chat#1813). Returns `{ runId, status }` — a snapshot ("is it
+ * done?"), not the generated content. Read the content via the chat (`chatId`
+ * from the start response): `GET /api/chat/{chatId}/stream`, or the persisted
+ * messages.
+ *
+ * Authentication: x-api-key header required.
+ *
+ * @param request - The request object.
+ * @param ctx - Route context with the `runId` path param.
+ * @param ctx.params - Promise resolving to the `{ runId }` path params.
+ * @returns 200 `{ runId, status }`, 401/403 on auth, or 404 if the run is unknown.
+ */
+export async function GET(
+  request: NextRequest,
+  ctx: { params: Promise<{ runId: string }> },
+): Promise<Response> {
+  const { runId } = await ctx.params;
+  return handleChatRunStatus(request, runId);
+}
+
+export const dynamic = "force-dynamic";

app/api/chat/runs/route.ts

@@ -0,0 +1,55 @@
+import type { NextRequest } from "next/server";
+import { NextResponse } from "next/server";
+import { getCorsHeaders } from "@/lib/networking/getCorsHeaders";
+import { handleStartChatRun } from "@/lib/chat/runs/handleStartChatRun";
+
+/**
+ * OPTIONS handler for CORS preflight requests.
+ *
+ * @returns A NextResponse with CORS headers.
+ */
+export async function OPTIONS() {
+  return new NextResponse(null, {
+    status: 200,
+    headers: getCorsHeaders(),
+  });
+}
+
+/**
+ * POST /api/chat/runs
+ *
+ * Start an asynchronous, headless chat-generation run on the durable
+ * `runAgentWorkflow` (recoupable/chat#1813). Provisions a session + sandbox,
+ * starts a workflow run, and returns `{ runId, chatId, sessionId }` with **202**
+ * immediately (plus a `Location` header at the run-status resource) — generation,
+ * assistant-message persistence, and side effects happen server-side after the
+ * response.
+ *
+ * Authentication: x-api-key header required (account inferred from the key;
+ * org keys may override via body `accountId`).
+ *
+ * Request body:
+ * - prompt: String prompt (mutually exclusive with messages)
+ * - messages: Array of UIMessages (mutually exclusive with prompt)
+ * - artistId: Optional UUID of the artist account
+ * - model: Optional model ID override (default anthropic/claude-haiku-4.5)
+ * - topic: Optional session title
+ * - accountId: Optional accountId override (requires org API key)
+ *
+ * Response body (202): `{ runId, chatId, sessionId }`. Read the result later via
+ * `GET /api/chat/{chatId}/stream` (watch the stream) or the chat's persisted
+ * messages; poll `GET /api/chat/runs/{runId}` for status (status route lands in
+ * a follow-up).
+ *
+ * @param request - The request object
+ * @returns 202 `{ runId, chatId, sessionId }`, or a 4xx/5xx error
+ */
+export async function POST(request: NextRequest): Promise<Response> {
+  return handleStartChatRun(request);
+}
+
+// Provisioning (repo + session + sandbox) runs before the 202 returns, so give
+// the function headroom beyond the default. The workflow itself runs durably
+// outside this request.
+export const maxDuration = 120;
+export const dynamic = "force-dynamic";

app/lib/workflows/__tests__/runAgentWorkflow.test.ts

@@ -6,7 +6,11 @@ import { closeChatStream } from "@/app/lib/workflows/closeChatStream";
 import { generateAssistantMessageId } from "@/app/lib/workflows/generateAssistantMessageId";
 import { handleChatCredits } from "@/lib/credits/handleChatCredits";
 import { autoCommitChatTurn } from "@/lib/chat/auto-commit/autoCommitChatTurn";
+import { deleteEphemeralKeyStep } from "@/app/lib/workflows/deleteEphemeralKeyStep";
 
+vi.mock("@/app/lib/workflows/deleteEphemeralKeyStep", () => ({
+  deleteEphemeralKeyStep: vi.fn(),
+}));
 vi.mock("@/app/lib/workflows/runAgentStep", () => ({
   runAgentStep: vi.fn(),
 }));
@@ -90,6 +94,37 @@ describe("runAgentWorkflow", () => {
     expect(clearChatActiveStream).toHaveBeenCalledWith("chat-1", "wrun_from_metadata");
   });
 
+  it("deletes the ephemeral key on run end when agentContext.ephemeralKeyId is set (headless)", async () => {
+    vi.mocked(runAgentStep).mockResolvedValue({
+      finishReason: "stop",
+      aborted: false,
+      responseMessage: undefined,
+    });
+
+    await runAgentWorkflow({
+      ...baseInput,
+      agentContext: {
+        sandbox: { state: { type: "vercel" }, workingDirectory: "/sandbox/mono" },
+        ephemeralKeyId: "ephem-key-1",
+      } as never,
+    });
+
+    expect(deleteEphemeralKeyStep).toHaveBeenCalledTimes(1);
+    expect(deleteEphemeralKeyStep).toHaveBeenCalledWith("ephem-key-1");
+  });
+
+  it("does NOT delete a key for the interactive path (no ephemeralKeyId)", async () => {
+    vi.mocked(runAgentStep).mockResolvedValue({
+      finishReason: "stop",
+      aborted: false,
+      responseMessage: undefined,
+    });
+
+    await runAgentWorkflow(baseInput);
+
+    expect(deleteEphemeralKeyStep).not.toHaveBeenCalled();
+  });
+
   it("explicitly closes the chat writable after a successful run so SSE ends promptly", async () => {
     vi.mocked(runAgentStep).mockResolvedValue({
       finishReason: "stop",

app/lib/workflows/deleteEphemeralKeyStep.ts

@@ -0,0 +1,25 @@
+import { deleteApiKey } from "@/lib/supabase/account_api_keys/deleteApiKey";
+
+/**
+ * Vercel Workflow `"use step"` that deletes the ephemeral, account-scoped
+ * `recoup_sk_…` key minted for a headless `/api/chat/runs` run
+ * (recoupable/chat#1813). Called from `runAgentWorkflow`'s `finally` so the
+ * credential is revoked the moment the run ends — the key's ~15m `expires_at`
+ * TTL (enforced in `getApiKeyAccountId`) is only the backstop if this is missed.
+ *
+ * Defensively swallows its own errors: a cleanup hiccup must not fail the run,
+ * and the TTL still guarantees the key can't outlive its window.
+ *
+ * @param keyId - `account_api_keys.id` of the ephemeral key to delete.
+ */
+export async function deleteEphemeralKeyStep(keyId: string): Promise<void> {
+  "use step";
+  try {
+    const { error } = await deleteApiKey(keyId);
+    if (error) {
+      console.error(`[deleteEphemeralKeyStep] failed to delete key ${keyId}:`, error);
+    }
+  } catch (error) {
+    console.error(`[deleteEphemeralKeyStep] unhandled error deleting key ${keyId}:`, error);
+  }
+}

app/lib/workflows/runAgentWorkflow.ts

@@ -4,6 +4,7 @@ import { closeChatStream } from "@/app/lib/workflows/closeChatStream";
 import { generateAssistantMessageId } from "@/app/lib/workflows/generateAssistantMessageId";
 import { runAgentStep } from "@/app/lib/workflows/runAgentStep";
 import { clearChatActiveStream } from "@/lib/chat/clearChatActiveStream";
+import { deleteEphemeralKeyStep } from "@/app/lib/workflows/deleteEphemeralKeyStep";
 import { handleChatCredits } from "@/lib/credits/handleChatCredits";
 import { autoCommitChatTurn } from "@/lib/chat/auto-commit/autoCommitChatTurn";
 import type { AgentMessageMetadata } from "@/lib/agent/messageMetadata/AgentMessageMetadata";
@@ -151,11 +152,19 @@ export async function runAgentWorkflow(input: RunAgentWorkflowInput): Promise<vo
     //      stream-end. Mirrors open-agents'
     //      `Promise.all([clearActiveStream, sendFinish.then(closeStream)])`.
     //
-    // `Promise.all` is safe because both helpers swallow their own
-    // errors — a failure in one doesn't cancel the other.
+    // A third cleanup step runs only for headless `/api/chat/runs` runs:
+    //   3) `deleteEphemeralKeyStep` — revoke the per-run, account-scoped
+    //      `recoup_sk_…` key minted for the sandbox the moment the run ends.
+    //      The key's ~15m TTL is only the backstop if this delete is missed.
+    //
+    // `Promise.all` is safe because all helpers swallow their own errors —
+    // a failure in one doesn't cancel the others.
     await Promise.all([
       clearChatActiveStream(input.chatId, workflowRunId),
       closeChatStream(writable),
+      ...(input.agentContext.ephemeralKeyId
+        ? [deleteEphemeralKeyStep(input.agentContext.ephemeralKeyId)]
+        : []),
     ]);
   }
 }

lib/agent/tools/AgentContext.ts

@@ -65,6 +65,15 @@ export type AgentContext = {
    * Empty / undefined when the sandbox has no `skills/` directory.
    */
   skills?: SkillMetadata[];
+  /**
+   * Row id of an ephemeral, account-scoped `recoup_sk_…` key minted for a
+   * headless run (`/api/chat/runs`, recoupable/chat#1813). When set, the
+   * workflow deletes the key in its `finally` cleanup so the credential is
+   * revoked the moment the run ends (the key's ~15m `expires_at` TTL is the
+   * defense-in-depth backstop if that delete is missed). The interactive
+   * caller forwards a Privy JWT and never sets this.
+   */
+  ephemeralKeyId?: string;
   /**
    * Main agent's language model. Tools read this via `getMainModel`.
    * Set per-step by `runAgentStep` (not part of the durable input).