@oneuptime/common 12.0.32 → 12.0.33

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

@@ -4,8 +4,25 @@ import {
 } from "../../../../Server/Utils/AI/Chat/ObservabilityChatPrompt";
 import AIChatPageContextType, {
   AIChatPageContext,
+  AIChatPageContextHelper,
 } from "../../../../Types/AI/AIChatPageContext";
 import AIChatPermissionMode from "../../../../Types/AI/AIChatPermissionMode";
+/*
+ * Toolbox/Index MUST be imported before any individual tool module. The
+ * toolbox sits in an import cycle (tool -> service -> ... -> AIToolbox), so
+ * entering through a tool module first re-enters Index before its tools array
+ * is assigned and leaves undefined holes in it. Production always enters
+ * through Index, which is the order reproduced here.
+ */
+import AIToolbox from "../../../../Server/Utils/AI/Toolbox/Index";
+import {
+  QueryIncidentsTool,
+  SearchIncidentsTool,
+} from "../../../../Server/Utils/AI/Toolbox/IncidentTools";
+import { QueryAlertsTool } from "../../../../Server/Utils/AI/Toolbox/AlertTools";
+import { QueryMonitorsTool } from "../../../../Server/Utils/AI/Toolbox/MonitorTools";
+import { ObservabilityTool } from "../../../../Server/Utils/AI/Toolbox/ToolTypes";
+import { JSONObject } from "../../../../Types/JSON";
 import { describe, expect, test } from "@jest/globals";
 
 const ENTITY_ID: string = "0f10509d-6656-4a08-b957-235fd4e8c52e";
@@ -102,6 +119,58 @@ describe("buildPageContextSection", () => {
     },
   );
 
+  /*
+   * Regression for #3552. The incidents list ships a suggested prompt card
+   * asking "which incidents are currently active or unresolved, and what
+   * state is each in?" — and the model replied that it had no tool for that
+   * and could only act on one incident once given an incidentId. It was not
+   * hallucinating: this guidance summarised the tool as "query_incidents
+   * (recent incidents, or one by incidentId)", a NARROWER capability claim
+   * than query_incidents' own schema, injected on that exact page. The model
+   * paraphrased the parenthetical back as a denial. Page guidance must name
+   * the affordance the page's own suggested prompts depend on.
+   */
+  test('incidents list guidance offers state="active" for unresolved incidents', () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.IncidentsList,
+    });
+
+    expect(section).toContain('state="active"');
+    expect(section).toContain("unresolved");
+    // The exact phrasing the model quoted back as "I have no such tool".
+    expect(section).not.toContain("(recent incidents, or one by incidentId)");
+  });
+
+  /*
+   * The alerts list carries the same "what is firing right now?" suggested
+   * prompt, and query_alerts has the same state="active" affordance. Its
+   * guidance used to stop dead at the tool name ("answered with
+   * query_alerts."), which is the same under-claim one page over.
+   */
+  test('alerts list guidance offers state="active" for unresolved alerts', () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.AlertsList,
+    });
+
+    expect(section).toContain('state="active"');
+    expect(section).toContain("unresolved");
+    expect(section).not.toContain("answered with query_alerts.");
+  });
+
+  /*
+   * "What is down right now?" on the monitors list is query_monitors'
+   * problemsOnly flag. Naming only the tool left the model to guess that a
+   * plain listing was the best it could do.
+   */
+  test("monitors list guidance offers problemsOnly for what is down now", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.MonitorsList,
+    });
+
+    expect(section).toContain("problemsOnly");
+    expect(section).not.toContain("answered with query_monitors.");
+  });
+
   test("every section reminds the model that context is not evidence", () => {
     const section: string = buildPageContextSection({
       type: AIChatPageContextType.Incident,
@@ -168,6 +237,26 @@ describe("buildObservabilityChatSystemPrompt with page context", () => {
     expect(prompt).toContain("## Hard rules");
     expect(prompt).toContain("Cite your sources.");
   });
+
+  /*
+   * buildPageContextSection passing is worthless if the section never makes
+   * it into the prompt: the assembled string is what actually reaches the
+   * model. #3552 was reported by a user sitting on the incidents list, so
+   * that context is asserted end to end.
+   */
+  test('an incidents list context carries state="active" into the prompt', () => {
+    const prompt: string = buildObservabilityChatSystemPrompt({
+      currentTime: new Date("2026-07-16T00:00:00Z"),
+      permissionMode: AIChatPermissionMode.ReadOnly,
+      pageContext: {
+        type: AIChatPageContextType.IncidentsList,
+      },
+    });
+
+    expect(prompt).toContain("## Current page context");
+    expect(prompt).toContain('state="active"');
+    expect(prompt).not.toContain("(recent incidents, or one by incidentId)");
+  });
 });
 
 /*
@@ -344,3 +433,167 @@ describe("buildObservabilityChatSystemPrompt — action guidance per mode", () =
     },
   );
 });
+
+/*
+ * The page guidance is a SUMMARY of tools the model also receives in full.
+ * When the summary claims LESS than the schema, the model believes the
+ * summary — that is #3552. These tests assert both halves at once: that the
+ * tool really does take the argument, and that the page whose suggested
+ * prompts depend on it says so. Adding an area page means adding a row here,
+ * or deciding explicitly that the page has no "what is wrong right now?"
+ * affordance to advertise.
+ *
+ * A type alias rather than an interface on purpose: jest's object-form
+ * test.each is typed against Record<string, unknown>, which an interface does
+ * not satisfy (it has no implicit index signature).
+ */
+type AreaAffordanceCase = {
+  type: AIChatPageContextType;
+  tool: ObservabilityTool;
+  // The tool argument that answers "what is wrong RIGHT NOW?" on this page.
+  argument: string;
+  // The enum value the model must pass, or null when the argument is a flag.
+  enumValue: string | null;
+  // The literal text the page guidance must use to name the affordance.
+  guidancePhrase: string;
+};
+
+function getSchemaProperty(
+  tool: ObservabilityTool,
+  argument: string,
+): JSONObject | undefined {
+  const properties: JSONObject =
+    (tool.inputSchema["properties"] as JSONObject | undefined) || {};
+
+  return properties[argument] as JSONObject | undefined;
+}
+
+describe("page guidance stays in sync with the tool schemas it summarises", () => {
+  test.each<AreaAffordanceCase>([
+    {
+      type: AIChatPageContextType.IncidentsList,
+      tool: QueryIncidentsTool,
+      argument: "state",
+      enumValue: "active",
+      guidancePhrase: 'state="active"',
+    },
+    {
+      type: AIChatPageContextType.AlertsList,
+      tool: QueryAlertsTool,
+      argument: "state",
+      enumValue: "active",
+      guidancePhrase: 'state="active"',
+    },
+    {
+      type: AIChatPageContextType.MonitorsList,
+      tool: QueryMonitorsTool,
+      argument: "problemsOnly",
+      enumValue: null,
+      guidancePhrase: "problemsOnly",
+    },
+  ])(
+    "$type guidance names the $argument affordance its tool really has",
+    (testCase: AreaAffordanceCase) => {
+      // Half one: the tool's own schema documents the affordance.
+      const property: JSONObject | undefined = getSchemaProperty(
+        testCase.tool,
+        testCase.argument,
+      );
+
+      expect(property).toBeDefined();
+
+      if (testCase.enumValue) {
+        const allowedValues: Array<string> =
+          (property?.["enum"] as Array<string> | undefined) || [];
+
+        expect(allowedValues).toContain(testCase.enumValue);
+      }
+
+      // Half two: the page the user is standing on is told about it.
+      const section: string = buildPageContextSection({ type: testCase.type });
+
+      expect(section).toContain(testCase.tool.name);
+      expect(section).toContain(testCase.guidancePhrase);
+    },
+  );
+
+  /*
+   * The incidents list owns the historical question too, and search_incidents
+   * is the only tool that searches incident text. Dropping it here would push
+   * "have we seen this before?" onto query_incidents, which cannot search at
+   * all. This one is a drift guard: it holds before and after the fix.
+   */
+  test("the incidents list still routes free-text history to search_incidents", () => {
+    const section: string = buildPageContextSection({
+      type: AIChatPageContextType.IncidentsList,
+    });
+
+    expect(getSchemaProperty(SearchIncidentsTool, "searchText")).toBeDefined();
+    expect(section).toContain(SearchIncidentsTool.name);
+  });
+});
+
+/*
+ * The other direction of the same drift: guidance naming a tool the model was
+ * never given. Extracting tool names from prose is safe here because this
+ * prompt keeps a strict convention — tool names are snake_case, tool
+ * ARGUMENTS are camelCase (incidentId, problemsOnly, createdWithinHours), and
+ * everything else is plain English or a UUID. So a snake_case token in the
+ * guidance is always a capability claim, and every one of them is checked
+ * against the toolbox the runner actually offers. A fresh regex per call
+ * keeps the global flag's lastIndex out of the picture.
+ */
+function extractToolNameTokens(text: string): Array<string> {
+  const tokens: Array<string> =
+    text.match(/[a-z][a-z0-9]*(?:_[a-z0-9]+)+/g) || [];
+
+  return Array.from(new Set(tokens));
+}
+
+const TOOLBOX_TOOL_NAMES: Set<string> = new Set(
+  AIToolbox.getTools().map((tool: ObservabilityTool): string => {
+    return tool.name;
+  }),
+);
+
+function buildSectionForType(type: AIChatPageContextType): string {
+  if (!AIChatPageContextHelper.isEntityType(type)) {
+    return buildPageContextSection({ type: type });
+  }
+
+  return buildPageContextSection({
+    type: type,
+    entityId: type === AIChatPageContextType.Trace ? TRACE_ID : ENTITY_ID,
+  });
+}
+
+describe("page guidance only names tools that exist in the toolbox", () => {
+  /*
+   * Driven off the enum rather than a hand-written list, so a new page type
+   * fails here until someone decides what it routes to.
+   */
+  test.each<[AIChatPageContextType]>(
+    Object.values(AIChatPageContextType).map(
+      (type: AIChatPageContextType): [AIChatPageContextType] => {
+        return [type];
+      },
+    ),
+  )(
+    "%s names at least one tool and every tool it names is real",
+    (type: AIChatPageContextType) => {
+      const section: string = buildSectionForType(type);
+      const mentionedTools: Array<string> = extractToolNameTokens(section);
+
+      // Guidance that routes the model nowhere is guidance it cannot use.
+      expect(mentionedTools.length).toBeGreaterThan(0);
+
+      const notInToolbox: Array<string> = mentionedTools.filter(
+        (name: string): boolean => {
+          return !TOOLBOX_TOOL_NAMES.has(name);
+        },
+      );
+
+      expect(notInToolbox).toEqual([]);
+    },
+  );
+});

package/Tests/Server/Utils/Captcha.test.ts

@@ -0,0 +1,422 @@
+import { describe, expect, it, jest, beforeEach } from "@jest/globals";
+
+/*
+ * Captcha verification, which is a gate rather than a feature: everything it
+ * protects (signup, password reset, the public forms) is reachable by anyone on
+ * the internet, so the only behaviour that matters is what it does when
+ * something is wrong.
+ *
+ * Three ways it could fail open, none of them visible from the outside — the
+ * page looks identical whether the check ran or not:
+ *
+ *   1. Enabled with no secret key. There is nothing to verify against, and
+ *      "cannot verify" must mean "reject", not "allow".
+ *   2. The provider is unreachable, times out, or answers with something other
+ *      than success. An hCaptcha outage must not become an open door.
+ *   3. A blank or whitespace-only token passed off as a solved challenge.
+ *
+ * `CaptchaEnabled` and `CaptchaSecretKey` are read from the environment at
+ * module load, so each configuration below gets a freshly required copy of the
+ * module. axios is mocked throughout: these assertions are about the decision,
+ * and a test that reached hcaptcha.com would be neither hermetic nor honest.
+ */
+
+const SITEVERIFY_URL: string = "https://hcaptcha.com/siteverify";
+
+interface CaptchaModule {
+  default: {
+    isCaptchaEnabled: () => boolean;
+    verifyCaptcha: (options: {
+      token: string | null | undefined;
+      remoteIp?: string | null;
+    }) => Promise<void>;
+  };
+}
+
+interface PostCall {
+  url: string;
+  body: string;
+}
+
+interface LoadedCaptcha {
+  captcha: CaptchaModule["default"];
+  posts: Array<PostCall>;
+  setResponse: (response: unknown) => void;
+  setRejection: (error: Error | null) => void;
+}
+
+/*
+ * Load Captcha against one environment configuration, with axios stubbed. The
+ * real EnvironmentConfig is spread in first so only the two captcha values
+ * differ from production — mocking the module wholesale would quietly blank
+ * every other setting the import graph reads.
+ */
+function loadCaptcha(config: {
+  captchaEnabled: boolean;
+  captchaSecretKey: string;
+}): LoadedCaptcha {
+  const posts: Array<PostCall> = [];
+  let response: unknown = { data: { success: true } };
+  let rejection: Error | null = null;
+
+  let loaded: CaptchaModule["default"] | null = null;
+
+  jest.isolateModules((): void => {
+    jest.doMock("../../../Server/EnvironmentConfig", (): unknown => {
+      return {
+        ...(jest.requireActual("../../../Server/EnvironmentConfig") as Record<
+          string,
+          unknown
+        >),
+        CaptchaEnabled: config.captchaEnabled,
+        CaptchaSecretKey: config.captchaSecretKey,
+      };
+    });
+
+    jest.doMock("axios", (): unknown => {
+      const post: unknown = (url: string, body: string): Promise<unknown> => {
+        posts.push({ url, body });
+
+        if (rejection) {
+          return Promise.reject(rejection);
+        }
+
+        return Promise.resolve(response);
+      };
+
+      return {
+        __esModule: true,
+        default: {
+          post,
+          isAxiosError: (): boolean => {
+            return false;
+          },
+        },
+        post,
+        isAxiosError: (): boolean => {
+          return false;
+        },
+      };
+    });
+
+    /*
+     * requireActual, not require: it loads Captcha itself unmocked while its
+     * own imports - the two doMock calls above - still resolve through this
+     * isolated registry, which is exactly the split these tests need.
+     */
+    loaded = (
+      jest.requireActual("../../../Server/Utils/Captcha") as CaptchaModule
+    ).default;
+  });
+
+  return {
+    captcha: loaded!,
+    posts,
+    setResponse: (next: unknown): void => {
+      response = next;
+    },
+    setRejection: (next: Error | null): void => {
+      rejection = next;
+    },
+  };
+}
+
+beforeEach((): void => {
+  jest.resetModules();
+});
+
+/*
+ * Assert a rejection is a BadDataException carrying `message`.
+ *
+ * By constructor NAME, not identity. Each configuration above is loaded through
+ * jest.isolateModules, which gives the module under test its own registry — so
+ * the BadDataException it throws is a different class object from the one this
+ * file would import, and `toThrow(BadDataException)` fails on two constructors
+ * that print the same name. The name and the message are what the API surface
+ * actually depends on.
+ */
+async function expectBadDataRejection(
+  work: Promise<void>,
+  message: string | RegExp,
+): Promise<void> {
+  let thrown: unknown = null;
+
+  try {
+    await work;
+  } catch (err) {
+    thrown = err;
+  }
+
+  expect(thrown).toBeInstanceOf(Error);
+  expect((thrown as Error).constructor.name).toBe("BadDataException");
+
+  if (message instanceof RegExp) {
+    expect((thrown as Error).message).toMatch(message);
+  } else {
+    expect((thrown as Error).message).toBe(message);
+  }
+}
+
+const GENERIC_ERROR_MESSAGE: string =
+  "Captcha verification failed. Please try again.";
+
+describe("CaptchaUtil.isCaptchaEnabled", () => {
+  it("is on only when the feature flag and a secret key agree", () => {
+    expect(
+      loadCaptcha({
+        captchaEnabled: true,
+        captchaSecretKey: "a-secret",
+      }).captcha.isCaptchaEnabled(),
+    ).toBe(true);
+  });
+
+  it("is off when the flag is off", () => {
+    expect(
+      loadCaptcha({
+        captchaEnabled: false,
+        captchaSecretKey: "a-secret",
+      }).captcha.isCaptchaEnabled(),
+    ).toBe(false);
+  });
+
+  it("is off when the flag is on but no secret key is configured", () => {
+    /*
+     * A half-configured deployment must report itself as off, so the UI does
+     * not render a widget whose answer can never be checked.
+     */
+    expect(
+      loadCaptcha({
+        captchaEnabled: true,
+        captchaSecretKey: "",
+      }).captcha.isCaptchaEnabled(),
+    ).toBe(false);
+  });
+});
+
+describe("CaptchaUtil.verifyCaptcha when captcha is switched off", () => {
+  it("lets the request through without calling the provider", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: false,
+      captchaSecretKey: "a-secret",
+    });
+
+    await expect(
+      loaded.captcha.verifyCaptcha({ token: null }),
+    ).resolves.toBeUndefined();
+
+    // No outbound call at all — an off switch must not cost a round trip.
+    expect(loaded.posts).toHaveLength(0);
+  });
+});
+
+describe("CaptchaUtil.verifyCaptcha when captcha is on", () => {
+  it("resolves when the provider says the challenge was solved", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await expect(
+      loaded.captcha.verifyCaptcha({ token: "solved-token" }),
+    ).resolves.toBeUndefined();
+
+    expect(loaded.posts).toHaveLength(1);
+    expect(loaded.posts[0]!.url).toBe(SITEVERIFY_URL);
+  });
+
+  it("sends the secret and the token the visitor supplied", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await loaded.captcha.verifyCaptcha({ token: "solved-token" });
+
+    const body: URLSearchParams = new URLSearchParams(loaded.posts[0]!.body);
+
+    expect(body.get("secret")).toBe("a-secret");
+    expect(body.get("response")).toBe("solved-token");
+  });
+
+  it("forwards the caller's IP when one is known", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await loaded.captcha.verifyCaptcha({
+      token: "solved-token",
+      remoteIp: "203.0.113.7",
+    });
+
+    expect(new URLSearchParams(loaded.posts[0]!.body).get("remoteip")).toBe(
+      "203.0.113.7",
+    );
+  });
+
+  it("omits remoteip rather than sending an empty one", async () => {
+    /*
+     * hCaptcha scores a blank remoteip differently from an absent one, so a
+     * request with no known client IP must leave the field off entirely.
+     */
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await loaded.captcha.verifyCaptcha({ token: "solved-token", remoteIp: "" });
+
+    expect(new URLSearchParams(loaded.posts[0]!.body).has("remoteip")).toBe(
+      false,
+    );
+  });
+
+  it("trims the token before sending it", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await loaded.captcha.verifyCaptcha({ token: "  solved-token  " });
+
+    expect(new URLSearchParams(loaded.posts[0]!.body).get("response")).toBe(
+      "solved-token",
+    );
+  });
+});
+
+describe("CaptchaUtil.verifyCaptcha fails closed", () => {
+  it("rejects when enabled with no secret key, and never calls the provider", async () => {
+    /*
+     * The misconfiguration that would otherwise fail OPEN: no key means no
+     * verification is possible, and "cannot verify" has to mean "reject".
+     */
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "",
+    });
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "solved-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+
+    expect(loaded.posts).toHaveLength(0);
+  });
+
+  it.each([
+    ["null", null],
+    ["undefined", undefined],
+    ["empty", ""],
+    ["whitespace only", "   "],
+  ])(
+    "rejects a %s token without calling the provider",
+    async (_label: string, token: string | null | undefined) => {
+      const loaded: LoadedCaptcha = loadCaptcha({
+        captchaEnabled: true,
+        captchaSecretKey: "a-secret",
+      });
+
+      await expectBadDataRejection(
+        loaded.captcha.verifyCaptcha({ token }),
+        /complete the verification challenge/i,
+      );
+
+      expect(loaded.posts).toHaveLength(0);
+    },
+  );
+
+  it("names the missing challenge, since that one is the visitor's to fix", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "" }),
+      /complete the verification challenge/i,
+    );
+  });
+
+  it("rejects when the provider answers success: false", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    loaded.setResponse({
+      data: { success: false, "error-codes": ["invalid-input-response"] },
+    });
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "forged-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+  });
+
+  it("rejects when the provider answers with no success field at all", async () => {
+    // A shape change at the provider must not read as a pass.
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    loaded.setResponse({ data: {} });
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "some-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+  });
+
+  it("rejects when the provider answers with no body", async () => {
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    loaded.setResponse({ data: null });
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "some-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+  });
+
+  it("rejects when the provider is unreachable", async () => {
+    /*
+     * An hCaptcha outage is the moment a fail-open bug would be exploited, and
+     * the moment nobody is watching for one.
+     */
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    loaded.setRejection(new Error("connect ETIMEDOUT"));
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "some-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+  });
+
+  it("does not leak the provider's error text to the visitor", async () => {
+    /*
+     * The message reaches an anonymous caller, so it says only that
+     * verification failed — not which secret, provider or error code was
+     * involved.
+     */
+    const loaded: LoadedCaptcha = loadCaptcha({
+      captchaEnabled: true,
+      captchaSecretKey: "a-secret",
+    });
+
+    loaded.setRejection(new Error("secret a-secret rejected by upstream"));
+
+    await expectBadDataRejection(
+      loaded.captcha.verifyCaptcha({ token: "some-token" }),
+      GENERIC_ERROR_MESSAGE,
+    );
+  });
+});