From c4125fee6c55200e8ab3f5b340092baff0cecc44 Mon Sep 17 00:00:00 2001
From: Jeongwan Jeon <jjw@jjw.kr>
Date: Sat, 22 Aug 2026 04:36:32 +0900
Subject: [PATCH] Refuse to start the built-in Bot without a model key, so the
 healthcheck stops passing a Bot that cannot answer

---
 .github/workflows/ci.yml |  4 +++
 CHANGELOG.md             |  4 +++
 agent-bot/src/index.ts   | 19 +++++++++++++-
 tests/agent-bot.test.ts  | 57 ++++++++++++++++++++++++++++++++++++++++
 4 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 tests/agent-bot.test.ts

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 2a14feea..14225907 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -66,6 +66,10 @@ jobs:
         with:
           bun-version: 1.3.14
       - run: bun install --frozen-lockfile
+      # The Bots are not root workspaces and each keeps its own lockfile. The startup tests spawn
+      # the real entrypoint, so its dependencies have to be installed as well.
+      - run: bun install --frozen-lockfile
+        working-directory: agent-bot
       # Not the db:migrate script: that one loads ../.env, which does not exist in CI. DATABASE_URL
       # comes from the job env instead, which drizzle.config.ts already reads.
       - run: bunx drizzle-kit migrate --config=drizzle.config.ts
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 33b92b13..01a6c0d9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -30,6 +30,10 @@ digit. The same rule container and volume names have always followed. A deployme
 `AUDIT_RETENTION_DAYS` is new and unset, which keeps the audit trail forever, as before. Set it to a
 whole number of days to have old rows removed.
 
+The built-in Bot refuses to start without `OPENAI_API_KEY`. It used to start, report healthy, and
+then fail every conversation, so a missing key looked like a working deployment. The LangGraph Bot
+already refused the same way.
+
 Sessions survive and nobody signs in again.
 
 ### Added
diff --git a/agent-bot/src/index.ts b/agent-bot/src/index.ts
index 3fecd861..e48d318d 100644
--- a/agent-bot/src/index.ts
+++ b/agent-bot/src/index.ts
@@ -45,8 +45,25 @@ const MODEL = process.env.BOT_MODEL ?? "gpt-5.5";
  */
 const BASE_URL = process.env.OPENAI_BASE_URL?.trim() || undefined;
 
+/**
+ * The key that model is answered with, checked at startup rather than on the first conversation.
+ *
+ * Without the check the Bot starts, answers the healthcheck, and then fails every run, so the
+ * compose healthcheck reports a Bot that cannot answer as healthy. The LangGraph Bot already
+ * refuses to start without its provider's key, and this file already refuses without its token
+ * above; the model key was the one configuration that escaped the same posture. A missing key
+ * should fail in front of whoever is deploying, not in front of whoever is asking.
+ */
+const API_KEY = process.env.OPENAI_API_KEY?.trim();
+if (!API_KEY) {
+  console.error(
+    "OPENAI_API_KEY is not set. This Bot cannot answer without a model.",
+  );
+  process.exit(1);
+}
+
 const openai = new OpenAI({
-  apiKey: process.env.OPENAI_API_KEY,
+  apiKey: API_KEY,
   baseURL: BASE_URL,
 });
 
diff --git a/tests/agent-bot.test.ts b/tests/agent-bot.test.ts
new file mode 100644
index 00000000..201c72a1
--- /dev/null
+++ b/tests/agent-bot.test.ts
@@ -0,0 +1,57 @@
+import { expect, test } from "bun:test";
+import { join } from "node:path";
+
+/**
+ * The compose healthcheck asks `/health`, and `/health` answers without consulting the model key.
+ * The refusal to run without one therefore has to happen before the server listens: a Bot that
+ * cannot answer should not be running, let alone reporting healthy. These spawn the real
+ * entrypoint with the environment compose would hand it.
+ */
+
+async function startBot(environment: Record<string, string>) {
+  const proc = Bun.spawn(
+    ["bun", join(import.meta.dir, "..", "agent-bot", "src", "index.ts")],
+    {
+      // Every variable the repository's `.env` could inject is named explicitly: bun loads that
+      // file into the child, and a leaked token or key would let a configuration under test pass
+      // a check it is supposed to fail.
+      env: {
+        PATH: process.env.PATH ?? "",
+        MANAGED_AGENT_TOKEN: "",
+        OPENAI_API_KEY: "",
+        ...environment,
+      },
+      stdout: "pipe",
+      stderr: "pipe",
+    },
+  );
+
+  // Both configurations under test exit before the server listens. If one reaches `serve` anyway,
+  // kill it so the test fails on the missing refusal rather than on bun's test timeout.
+  const killer = setTimeout(() => proc.kill(), 5_000);
+  const exitCode = await proc.exited;
+  clearTimeout(killer);
+  const stderr = await new Response(proc.stderr).text();
+  return { exitCode, stderr };
+}
+
+test("agent-bot refuses to start when the model key is empty", async () => {
+  // Compose passes `${OPENAI_API_KEY:-}`, so an unset key arrives as an empty
+  // string rather than missing altogether.
+  const { exitCode, stderr } = await startBot({
+    MANAGED_AGENT_TOKEN: "test-token",
+    OPENAI_API_KEY: "",
+  });
+
+  expect(exitCode).toBe(1);
+  expect(stderr).toContain("OPENAI_API_KEY is not set");
+});
+
+test("agent-bot still refuses to start without its server token", async () => {
+  const { exitCode, stderr } = await startBot({
+    OPENAI_API_KEY: "sk-test",
+  });
+
+  expect(exitCode).toBe(1);
+  expect(stderr).toContain("MANAGED_AGENT_TOKEN is not set");
+});