@oneuptime/common 11.5.7 → 11.5.9

package/Tests/Server/Utils/AI/LLMServiceModelCompatibility.test.ts

@@ -0,0 +1,548 @@
+import API from "../../../../Utils/API";
+import logger from "../../../../Server/Utils/Logger";
+import HTTPErrorResponse from "../../../../Types/API/HTTPErrorResponse";
+import HTTPResponse from "../../../../Types/API/HTTPResponse";
+import { JSONObject } from "../../../../Types/JSON";
+import LlmType from "../../../../Types/LLM/LlmType";
+import LLMService, {
+  LLMProviderConfig,
+} from "../../../../Server/Utils/LLM/LLMService";
+import {
+  afterEach,
+  beforeEach,
+  describe,
+  expect,
+  jest,
+  test,
+} from "@jest/globals";
+
+type PostSpy = ReturnType<typeof jest.spyOn>;
+
+const SUCCESS_BODY: JSONObject = {
+  choices: [{ message: { content: "OK" } }],
+  usage: { prompt_tokens: 5, completion_tokens: 1, total_tokens: 6 },
+};
+
+function successResponse(): HTTPResponse<JSONObject> {
+  return { jsonData: SUCCESS_BODY } as unknown as HTTPResponse<JSONObject>;
+}
+
+/** The verbatim 400 an o-series / gpt-5 deployment returns for max_tokens. */
+function maxTokensUnsupportedResponse(): HTTPErrorResponse {
+  return new HTTPErrorResponse(
+    400,
+    {
+      error: {
+        message:
+          "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.",
+        type: "invalid_request_error",
+        param: "max_tokens",
+        code: "unsupported_parameter",
+      },
+    },
+    {},
+  );
+}
+
+/** Older models / Azure api-versions before 2024-08-01-preview. param and code are null. */
+function maxCompletionTokensUnrecognizedResponse(): HTTPErrorResponse {
+  return new HTTPErrorResponse(
+    400,
+    {
+      error: {
+        message:
+          "Unrecognized request argument supplied: max_completion_tokens",
+        type: "invalid_request_error",
+        param: null,
+        code: null,
+      },
+    },
+    {},
+  );
+}
+
+function temperatureUnsupportedResponse(): HTTPErrorResponse {
+  return new HTTPErrorResponse(
+    400,
+    {
+      error: {
+        message:
+          "Unsupported value: 'temperature' does not support 0 with this model. Only the default (1) value is supported.",
+        type: "invalid_request_error",
+        param: "temperature",
+        code: "unsupported_value",
+      },
+    },
+    {},
+  );
+}
+
+const AZURE_CONFIG: LLMProviderConfig = {
+  llmType: LlmType.AzureOpenAI,
+  apiKey: "azure-key",
+  baseUrl: "https://example.openai.azure.com/openai/deployments/my-deployment",
+  // Azure sends the *deployment* name, which need not resemble the model.
+  modelName: "my-deployment",
+};
+
+function bodyOf(spy: PostSpy, callIndex: number): JSONObject {
+  return (spy.mock.calls[callIndex]![0] as { data: JSONObject }).data;
+}
+
+beforeEach(() => {
+  LLMService.clearRequestAdaptationCache();
+  jest.spyOn(logger, "debug").mockImplementation((): void => {});
+  jest.spyOn(logger, "error").mockImplementation((): void => {});
+});
+
+afterEach(() => {
+  jest.restoreAllMocks();
+});
+
+describe("LLMService — Azure OpenAI max_tokens compatibility", () => {
+  test("retries with max_completion_tokens when the deployment rejects max_tokens", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValueOnce(successResponse()) as PostSpy;
+
+    const response: { content: string } = await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "Reply with the word: OK" }],
+      maxTokens: 16,
+    });
+
+    expect(response.content).toBe("OK");
+    expect(spy).toHaveBeenCalledTimes(2);
+
+    // First attempt used the legacy spelling...
+    expect(bodyOf(spy, 0)["max_tokens"]).toBe(16);
+    expect(bodyOf(spy, 0)["max_completion_tokens"]).toBeUndefined();
+
+    // ...the retry swapped it, preserving the cap and dropping the legacy key.
+    expect(bodyOf(spy, 1)["max_completion_tokens"]).toBe(16);
+    expect(bodyOf(spy, 1)["max_tokens"]).toBeUndefined();
+  });
+
+  test("drops temperature when the deployment rejects it, after the token swap", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValueOnce(temperatureUnsupportedResponse())
+      .mockResolvedValueOnce(successResponse()) as PostSpy;
+
+    const response: { content: string } = await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      temperature: 0,
+      maxTokens: 16,
+    });
+
+    expect(response.content).toBe("OK");
+    expect(spy).toHaveBeenCalledTimes(3);
+
+    // Errors arrive one param at a time, so the fixes accumulate.
+    expect(bodyOf(spy, 1)["max_completion_tokens"]).toBe(16);
+    expect(bodyOf(spy, 1)["temperature"]).toBe(0);
+
+    expect(bodyOf(spy, 2)["max_completion_tokens"]).toBe(16);
+    expect(bodyOf(spy, 2)["temperature"]).toBeUndefined();
+  });
+
+  test("remembers what the deployment accepted so later calls do not re-probe", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValue(successResponse()) as PostSpy;
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+    expect(spy).toHaveBeenCalledTimes(2);
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi again" }],
+      maxTokens: 32,
+    });
+
+    // One extra call only: the second completion got the shape right first try.
+    expect(spy).toHaveBeenCalledTimes(3);
+    expect(bodyOf(spy, 2)["max_completion_tokens"]).toBe(32);
+    expect(bodyOf(spy, 2)["max_tokens"]).toBeUndefined();
+  });
+
+  test("re-probes when the model name changes", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValue(successResponse()) as PostSpy;
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+    expect(spy).toHaveBeenCalledTimes(2);
+
+    await LLMService.getCompletion({
+      llmProviderConfig: { ...AZURE_CONFIG, modelName: "other-deployment" },
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+
+    // A different deployment is a different model: back to the legacy default.
+    expect(bodyOf(spy, 2)["max_tokens"]).toBe(16);
+  });
+});
+
+describe("LLMService — reasoning model detection by name", () => {
+  test.each([
+    ["gpt-5", true],
+    ["gpt-5-mini", true],
+    ["gpt5-mini", true],
+    ["o1", true],
+    ["o1-mini", true],
+    ["o3-mini", true],
+    ["o4-mini", true],
+    ["codex-mini", true],
+    ["gpt-4o", false],
+    ["gpt-4o-mini", false],
+    ["gpt-4.1", false],
+    ["gpt-4-turbo", false],
+    ["gpt-35-turbo", false],
+  ])(
+    "%s sends max_completion_tokens up front: %s",
+    async (modelName: string, expectsMaxCompletionTokens: boolean) => {
+      const spy: PostSpy = jest
+        .spyOn(API, "post")
+        .mockResolvedValue(successResponse()) as PostSpy;
+
+      await LLMService.getCompletion({
+        llmProviderConfig: {
+          llmType: LlmType.OpenAI,
+          apiKey: "test-key",
+          modelName: modelName,
+        },
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 16,
+      });
+
+      // Known reasoning models skip the wasted round trip entirely.
+      expect(spy).toHaveBeenCalledTimes(1);
+      const body: JSONObject = bodyOf(spy, 0);
+
+      if (expectsMaxCompletionTokens) {
+        expect(body["max_completion_tokens"]).toBe(16);
+        expect(body["max_tokens"]).toBeUndefined();
+      } else {
+        expect(body["max_tokens"]).toBe(16);
+        expect(body["max_completion_tokens"]).toBeUndefined();
+      }
+    },
+  );
+});
+
+describe("LLMService — legacy models that reject max_completion_tokens", () => {
+  test("falls back to max_tokens when the endpoint does not recognize max_completion_tokens", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxCompletionTokensUnrecognizedResponse())
+      .mockResolvedValueOnce(successResponse()) as PostSpy;
+
+    // A gpt-5-named deployment behind an endpoint that only knows max_tokens.
+    const response: { content: string } = await LLMService.getCompletion({
+      llmProviderConfig: { ...AZURE_CONFIG, modelName: "gpt-5-mini" },
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+
+    expect(response.content).toBe("OK");
+    expect(bodyOf(spy, 0)["max_completion_tokens"]).toBe(16);
+    expect(bodyOf(spy, 1)["max_tokens"]).toBe(16);
+    expect(bodyOf(spy, 1)["max_completion_tokens"]).toBeUndefined();
+  });
+
+  test("does not flip-flop when the endpoint rejects both spellings", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxCompletionTokensUnrecognizedResponse())
+      .mockResolvedValue(maxTokensUnsupportedResponse()) as PostSpy;
+
+    await expect(
+      LLMService.getCompletion({
+        llmProviderConfig: { ...AZURE_CONFIG, modelName: "gpt-5-mini" },
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 16,
+      }),
+    ).rejects.toThrow("Azure OpenAI API error");
+
+    // Each spelling is tried exactly once, then the provider error surfaces.
+    expect(spy).toHaveBeenCalledTimes(2);
+  });
+
+  test("does not flip-flop when the name heuristic misses and additionalParams chose the new spelling", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxCompletionTokensUnrecognizedResponse())
+      .mockResolvedValue(maxTokensUnsupportedResponse()) as PostSpy;
+
+    /*
+     * modelName does not match the heuristic, so the adaptation starts at the
+     * Default sentinel while additionalParams puts max_completion_tokens on
+     * the wire. Both spellings must still be tried at most once each.
+     */
+    await expect(
+      LLMService.getCompletion({
+        llmProviderConfig: { ...AZURE_CONFIG, modelName: "prod" },
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 16,
+        additionalParams: { max_completion_tokens: 4000 },
+      }),
+    ).rejects.toThrow("Azure OpenAI API error");
+
+    expect(spy).toHaveBeenCalledTimes(2);
+    expect(bodyOf(spy, 0)["max_completion_tokens"]).toBeDefined();
+    expect(bodyOf(spy, 1)["max_tokens"]).toBeDefined();
+  });
+
+  test("adapts a chain longer than four params rather than giving up short", async () => {
+    const reject: (param: string) => HTTPErrorResponse = (
+      param: string,
+    ): HTTPErrorResponse => {
+      return new HTTPErrorResponse(
+        400,
+        {
+          error: {
+            message: `Unsupported parameter: '${param}' is not supported with this model.`,
+            type: "invalid_request_error",
+            param: param,
+            code: "unsupported_parameter",
+          },
+        },
+        {},
+      );
+    };
+
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValueOnce(reject("temperature"))
+      .mockResolvedValueOnce(reject("top_p"))
+      .mockResolvedValueOnce(reject("frequency_penalty"))
+      .mockResolvedValueOnce(reject("presence_penalty"))
+      .mockResolvedValueOnce(successResponse()) as PostSpy;
+
+    // Five reshapes are needed; the attempt cap must not cut this short.
+    const response: { content: string } = await LLMService.getCompletion({
+      llmProviderConfig: { ...AZURE_CONFIG, modelName: "prod" },
+      messages: [{ role: "user", content: "hi" }],
+      temperature: 0,
+      maxTokens: 16,
+      additionalParams: {
+        top_p: 0.9,
+        frequency_penalty: 0.2,
+        presence_penalty: 0.2,
+      },
+    });
+
+    expect(response.content).toBe("OK");
+    expect(spy).toHaveBeenCalledTimes(6);
+
+    const finalBody: JSONObject = bodyOf(spy, 5);
+    expect(finalBody["max_completion_tokens"]).toBe(16);
+    expect(finalBody["temperature"]).toBeUndefined();
+    expect(finalBody["top_p"]).toBeUndefined();
+    expect(finalBody["frequency_penalty"]).toBeUndefined();
+    expect(finalBody["presence_penalty"]).toBeUndefined();
+  });
+
+  test("remembers dropped params, not just the token spelling", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValueOnce(temperatureUnsupportedResponse())
+      .mockResolvedValue(successResponse()) as PostSpy;
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      temperature: 0,
+      maxTokens: 16,
+    });
+    expect(spy).toHaveBeenCalledTimes(3);
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "again" }],
+      temperature: 0,
+      maxTokens: 16,
+    });
+
+    // Both halves of what was learned must survive: the spelling and the drop.
+    expect(spy).toHaveBeenCalledTimes(4);
+    expect(bodyOf(spy, 3)["max_completion_tokens"]).toBe(16);
+    expect(bodyOf(spy, 3)["temperature"]).toBeUndefined();
+  });
+
+  test("keeps two Azure resources that share a deployment name apart", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValue(successResponse()) as PostSpy;
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+    expect(spy).toHaveBeenCalledTimes(2);
+
+    await LLMService.getCompletion({
+      llmProviderConfig: {
+        ...AZURE_CONFIG,
+        baseUrl:
+          "https://other.openai.azure.com/openai/deployments/my-deployment",
+      },
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+
+    /*
+     * Same deployment name, different resource: what one accepts says nothing
+     * about the other, so it must re-probe.
+     */
+    expect(bodyOf(spy, 2)["max_tokens"]).toBe(16);
+  });
+});
+
+describe("LLMService — errors that are not parameter-compatibility problems", () => {
+  test("a max_tokens value error surfaces as-is instead of triggering a rename", async () => {
+    const spy: PostSpy = jest.spyOn(API, "post").mockResolvedValue(
+      new HTTPErrorResponse(
+        400,
+        {
+          error: {
+            message:
+              "max_tokens is too large: 200000. This model supports at most 16384 completion tokens.",
+            type: "invalid_request_error",
+            param: "max_tokens",
+            code: null,
+          },
+        },
+        {},
+      ),
+    ) as PostSpy;
+
+    await expect(
+      LLMService.getCompletion({
+        llmProviderConfig: {
+          llmType: LlmType.OpenAI,
+          apiKey: "test-key",
+          modelName: "gpt-4o",
+        },
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 200000,
+      }),
+      /*
+       * Renaming the param would not fix a too-large value; it would only
+       * bury the real complaint under one about a param the caller never set.
+       */
+    ).rejects.toThrow("max_tokens is too large");
+
+    expect(spy).toHaveBeenCalledTimes(1);
+  });
+
+  test("the plural unrecognized-argument wording still triggers the fallback", async () => {
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(
+        new HTTPErrorResponse(
+          400,
+          {
+            error: {
+              message:
+                "Unrecognized request arguments supplied: max_completion_tokens, reasoning_effort",
+              type: "invalid_request_error",
+              param: null,
+              code: null,
+            },
+          },
+          {},
+        ),
+      )
+      .mockResolvedValueOnce(successResponse()) as PostSpy;
+
+    const response: { content: string } = await LLMService.getCompletion({
+      llmProviderConfig: { ...AZURE_CONFIG, modelName: "gpt-5-mini" },
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+
+    expect(response.content).toBe("OK");
+    expect(bodyOf(spy, 1)["max_tokens"]).toBe(16);
+  });
+
+  test("an auth failure is surfaced without any retry", async () => {
+    const spy: PostSpy = jest.spyOn(API, "post").mockResolvedValue(
+      new HTTPErrorResponse(
+        401,
+        {
+          error: {
+            message: "Incorrect API key provided.",
+            type: "invalid_request_error",
+            param: null,
+            code: "invalid_api_key",
+          },
+        },
+        {},
+      ),
+    ) as PostSpy;
+
+    await expect(
+      LLMService.getCompletion({
+        llmProviderConfig: AZURE_CONFIG,
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 16,
+      }),
+    ).rejects.toThrow("Azure OpenAI API error");
+
+    expect(spy).toHaveBeenCalledTimes(1);
+  });
+
+  test("a failed request is not cached as a working shape", async () => {
+    jest
+      .spyOn(API, "post")
+      .mockResolvedValueOnce(maxTokensUnsupportedResponse())
+      .mockResolvedValueOnce(
+        new HTTPErrorResponse(500, { error: { message: "upstream boom" } }, {}),
+      );
+
+    await expect(
+      LLMService.getCompletion({
+        llmProviderConfig: AZURE_CONFIG,
+        messages: [{ role: "user", content: "hi" }],
+        maxTokens: 16,
+      }),
+    ).rejects.toThrow("Azure OpenAI API error");
+
+    jest.restoreAllMocks();
+    jest.spyOn(logger, "debug").mockImplementation((): void => {});
+    jest.spyOn(logger, "error").mockImplementation((): void => {});
+
+    const spy: PostSpy = jest
+      .spyOn(API, "post")
+      .mockResolvedValue(successResponse()) as PostSpy;
+
+    await LLMService.getCompletion({
+      llmProviderConfig: AZURE_CONFIG,
+      messages: [{ role: "user", content: "hi" }],
+      maxTokens: 16,
+    });
+
+    // The swap was never confirmed to work, so it must not be remembered.
+    expect(bodyOf(spy, 0)["max_tokens"]).toBe(16);
+  });
+});

package/Tests/Server/Utils/AI/ObservabilityChatPrompt.test.ts

@@ -0,0 +1,142 @@
+import {
+  buildObservabilityChatSystemPrompt,
+  buildPageContextSection,
+} from "../../../../Server/Utils/AI/Chat/ObservabilityChatPrompt";
+import AIChatPageContextType, {
+  AIChatPageContext,
+} from "../../../../Types/AI/AIChatPageContext";
+import AIChatPermissionMode from "../../../../Types/AI/AIChatPermissionMode";
+import { describe, expect, test } from "@jest/globals";
+
+const ENTITY_ID: string = "0f10509d-6656-4a08-b957-235fd4e8c52e";
+const TRACE_ID: string = "4bf92f3577b34da6a3ce929d0e0e4736";
+
+describe("buildPageContextSection", () => {
+  test("returns an empty string without a context", () => {
+    expect(buildPageContextSection(undefined)).toBe("");
+  });
+
+  /*
+   * Each entity type must name the tool (and its id argument) that fetches
+   * the entity — that pairing is what lets the model resolve "this incident"
+   * without guessing.
+   */
+  test.each<[AIChatPageContextType, string]>([
+    [
+      AIChatPageContextType.Incident,
+      `query_incidents using incidentId="${ENTITY_ID}"`,
+    ],
+    [AIChatPageContextType.Alert, `query_alerts using alertId="${ENTITY_ID}"`],
+    [
+      AIChatPageContextType.Monitor,
+      `query_monitors using monitorId="${ENTITY_ID}"`,
+    ],
+    [AIChatPageContextType.TelemetryService, `serviceId="${ENTITY_ID}"`],
+    [
+      AIChatPageContextType.Exception,
+      `find_code_for_exception with exceptionId="${ENTITY_ID}"`,
+    ],
+  ])(
+    "%s guidance names the fetch tool and id",
+    (type: AIChatPageContextType, expected: string) => {
+      const section: string = buildPageContextSection({
+        type: type,
+        entityId: ENTITY_ID,
+      });
+
+      expect(section).toContain("## Current page context");
+      expect(section).toContain(expected);
+    },
+  );
+
+  test("trace guidance uses get_trace with the hex trace id", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.Trace,
+      entityId: TRACE_ID,
+    });
+
+    expect(section).toContain(`get_trace using traceId="${TRACE_ID}"`);
+  });
+
+  test("scheduled maintenance guidance names query_scheduled_maintenance", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.ScheduledMaintenanceEvent,
+      entityId: ENTITY_ID,
+    });
+
+    expect(section).toContain(
+      `query_scheduled_maintenance using scheduledMaintenanceId="${ENTITY_ID}"`,
+    );
+  });
+
+  test("the entity title is included in quotes when present", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.Incident,
+      entityId: ENTITY_ID,
+      entityTitle: "#42 Payment API down",
+    });
+
+    expect(section).toContain('titled "#42 Payment API down"');
+  });
+
+  test.each<[AIChatPageContextType, string]>([
+    [AIChatPageContextType.IncidentsList, "query_incidents"],
+    [AIChatPageContextType.AlertsList, "query_alerts"],
+    [AIChatPageContextType.MonitorsList, "query_monitors"],
+    [
+      AIChatPageContextType.ScheduledMaintenanceList,
+      "query_scheduled_maintenance",
+    ],
+    [AIChatPageContextType.LogsExplorer, "log_histogram"],
+    [AIChatPageContextType.TracesExplorer, "query_traces"],
+    [AIChatPageContextType.MetricsExplorer, "query_metrics"],
+    [AIChatPageContextType.ExceptionsList, "top_exceptions"],
+  ])(
+    "area context %s names its primary tool",
+    (type: AIChatPageContextType, expectedTool: string) => {
+      const context: AIChatPageContext = { type: type };
+      const section: string = buildPageContextSection(context);
+
+      expect(section).toContain("## Current page context");
+      expect(section).toContain(expectedTool);
+    },
+  );
+
+  test("every section reminds the model that context is not evidence", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.Incident,
+      entityId: ENTITY_ID,
+    });
+
+    expect(section).toContain("It is not evidence");
+  });
+});
+
+describe("buildObservabilityChatSystemPrompt with page context", () => {
+  test("omits the section when no context is passed", () => {
+    const prompt: string = buildObservabilityChatSystemPrompt({
+      currentTime: new Date("2026-07-16T00:00:00Z"),
+      permissionMode: AIChatPermissionMode.AskForApproval,
+    });
+
+    expect(prompt).not.toContain("## Current page context");
+  });
+
+  test("includes the section without displacing the hard rules", () => {
+    const prompt: string = buildObservabilityChatSystemPrompt({
+      currentTime: new Date("2026-07-16T00:00:00Z"),
+      permissionMode: AIChatPermissionMode.AskForApproval,
+      pageContext: {
+        type: AIChatPageContextType.Incident,
+        entityId: ENTITY_ID,
+        entityTitle: "Payment API down",
+      },
+    });
+
+    expect(prompt).toContain("## Current page context");
+    expect(prompt).toContain(`incidentId="${ENTITY_ID}"`);
+    // The binding trust rules survive untouched.
+    expect(prompt).toContain("## Hard rules");
+    expect(prompt).toContain("Cite your sources.");
+  });
+});