@vellumai/assistant 0.8.5 → 0.8.7

package/src/__tests__/edit-propagation.test.ts

@@ -76,8 +76,7 @@ async function seedSlackMessage(opts: {
     inboundResult.conversationId,
     "user",
     initialContent,
-    { userMessageChannel: "slack" },
-    { skipIndexing: true },
+    { metadata: { userMessageChannel: "slack" }, skipIndexing: true },
   );
 
   linkMessage(inboundResult.eventId, inserted.id);

package/src/__tests__/empty-response-pipeline.test.ts

@@ -19,10 +19,8 @@
 import { beforeEach, describe, expect, test } from "bun:test";
 
 import type { TrustContext } from "../daemon/trust-context.js";
-import {
-  defaultEmptyResponsePlugin,
-  defaultEmptyResponseTerminal,
-} from "../plugins/defaults/empty-response.js";
+import { defaultEmptyResponsePlugin } from "../plugins/defaults/empty-response/register.js";
+import { defaultEmptyResponseTerminal } from "../plugins/defaults/empty-response/terminal.js";
 import { DEFAULT_TIMEOUTS, runPipeline } from "../plugins/pipeline.js";
 import {
   getMiddlewaresFor,
@@ -73,10 +71,21 @@ function makeArgs(
     emptyResponseRetries: 0,
     maxEmptyResponseRetries: 1,
     priorAssistantHadVisibleText: false,
+    // Default to `null` (no stop reason reported) so existing fixtures
+    // exercise the "organic empty turn" path. The refusal branch
+    // dedicated tests below set this to `"refusal"` explicitly.
+    stopReason: null,
     ...overrides,
   };
 }
 
+/**
+ * Refusal-specific nudge text — keep in sync with `register.ts`. Clients
+ * (and the model) may match on this exact text.
+ */
+const CANONICAL_REFUSAL_NUDGE_TEXT =
+  '<system_notice>Your previous response was empty because the upstream provider returned stop_reason="refusal". Please answer the user\'s last message directly with a plain-text response. Do not use any tools — just respond with text.</system_notice>';
+
 async function runEmpty(
   args: EmptyResponseArgs,
 ): Promise<EmptyResponseDecision> {
@@ -167,15 +176,128 @@ describe("emptyResponse pipeline — default decisions", () => {
 
   test("no prior tool-use turn (toolUseTurns === 0) → accept", async () => {
     // Empty first assistant response with no tools is not the pattern the
-    // nudge guards against. Default accepts.
+    // organic-empty-turn nudge guards against. Default accepts (unless the
+    // stop reason is `"refusal"` — see the refusal-specific tests below).
+    const decision = await runEmpty(
+      makeArgs({
+        responseContent: [],
+        toolUseTurns: 0,
+      }),
+    );
+    expect(decision.action).toBe("accept");
+  });
+
+  // ─── Refusal stop ────────────────────────────────────────────────────────
+
+  test("stopReason='refusal' on turn 0 with no content → nudge with refusal text", async () => {
+    // The canonical failure mode this branch exists to catch: Anthropic's
+    // safety classifier zeros the response on the very first model call,
+    // returning a single thinking block and `stopReason: "refusal"`. Without
+    // this branch, the terminal would `accept` and the loop would persist
+    // an empty assistant bubble to the user.
+    const decision = await runEmpty(
+      makeArgs({
+        stopReason: "refusal",
+        responseContent: [],
+        toolUseBlocksLength: 0,
+        toolUseTurns: 0,
+        emptyResponseRetries: 0,
+        priorAssistantHadVisibleText: false,
+      }),
+    );
+    expect(decision.action).toBe("nudge");
+    expect(decision.nudgeText).toBe(CANONICAL_REFUSAL_NUDGE_TEXT);
+  });
+
+  test("stopReason='refusal' with a thinking-only block still nudges", async () => {
+    // Thinking blocks aren't visible text — the user sees nothing. A
+    // refusal with only thinking content matches the same shape the
+    // production log captured (`contentBlocks: 1, toolUseCount: 0`).
+    const decision = await runEmpty(
+      makeArgs({
+        stopReason: "refusal",
+        responseContent: [
+          {
+            type: "thinking",
+            thinking: "...",
+            signature: "sig",
+          } as ContentBlock,
+        ],
+        toolUseBlocksLength: 0,
+        toolUseTurns: 0,
+      }),
+    );
+    expect(decision.action).toBe("nudge");
+    expect(decision.nudgeText).toBe(CANONICAL_REFUSAL_NUDGE_TEXT);
+  });
+
+  test("stopReason='refusal' but visible text present → accept (model recovered)", async () => {
+    // The classifier can flag a partial response; if the model already
+    // delivered some visible text before refusing, the user has something
+    // to see. Accept.
+    const decision = await runEmpty(
+      makeArgs({
+        stopReason: "refusal",
+        responseContent: [{ type: "text", text: "partial answer" }],
+      }),
+    );
+    expect(decision.action).toBe("accept");
+  });
+
+  test("stopReason='refusal' but tool_use blocks present → accept", async () => {
+    // A refusal with tool_use blocks is unusual (the model wouldn't normally
+    // issue tools after a classifier hit) but we still shouldn't nudge —
+    // the loop will execute the tools and the model will get another shot.
     const decision = await runEmpty(
       makeArgs({
+        stopReason: "refusal",
+        responseContent: [
+          {
+            type: "tool_use",
+            id: "tu-1",
+            name: "read",
+            input: { path: "/tmp/x" },
+          } as ContentBlock,
+        ],
+        toolUseBlocksLength: 1,
+      }),
+    );
+    expect(decision.action).toBe("accept");
+  });
+
+  test("stopReason='refusal' but retries exhausted → accept (no infinite loop)", async () => {
+    // Persistent classifier hit shouldn't burn turns indefinitely. Once
+    // we've used our retry budget, accept (the user will see an empty
+    // bubble, but the loop terminates).
+    const decision = await runEmpty(
+      makeArgs({
+        stopReason: "refusal",
         responseContent: [],
         toolUseTurns: 0,
+        emptyResponseRetries: 1,
+        maxEmptyResponseRetries: 1,
       }),
     );
     expect(decision.action).toBe("accept");
   });
+
+  test("stopReason='refusal' beats post-tool-empty nudge text (refusal-specific wording)", async () => {
+    // When both branches would fire, refusal wins because the refusal
+    // text is more accurate ("safety classifier zeroed the response"
+    // vs. "summary of what you found or did"). This guards against a
+    // future refactor that orders the branches differently.
+    const decision = await runEmpty(
+      makeArgs({
+        stopReason: "refusal",
+        responseContent: [],
+        toolUseBlocksLength: 0,
+        toolUseTurns: 2, // would trip the post-tool branch too
+        priorAssistantHadVisibleText: false,
+      }),
+    );
+    expect(decision.action).toBe("nudge");
+    expect(decision.nudgeText).toBe(CANONICAL_REFUSAL_NUDGE_TEXT);
+  });
 });
 
 describe("emptyResponse pipeline — custom middleware overrides", () => {

package/src/__tests__/encrypted-store-test-helpers.ts

@@ -0,0 +1,56 @@
+/**
+ * Test-only utilities for overriding the encrypted credential store paths.
+ *
+ * Replaces the removed `_setStorePath` and `_setStoreKeyPath` exports
+ * from `encrypted-store.ts`. Lives here (not in the source module)
+ * because production modules should not expose test backdoors.
+ *
+ * No source-module imports
+ * ------------------------
+ * This file has ZERO imports from `src/`. It accesses the store-path
+ * override state via the shared `globalThis.vellumAssistant.storePathOverride`
+ * slot that `src/security/store-path-override.ts` also reads/writes. The
+ * slot shape is duplicated here on purpose: keeping this file off the
+ * production import graph is what protects the test preload from a
+ * broken `node_modules` symlink (DB ghost #3). The two declarations MUST
+ * stay in sync — if you change one, change the other.
+ *
+ * Most tests no longer need these overrides: the test preload places
+ * `VELLUM_WORKSPACE_DIR` at `<tmpRoot>/workspace`, so `getProtectedDir()`
+ * resolves to `<tmpRoot>/protected` per process. The setters here exist
+ * for the small set of tests that exercise specific path scenarios
+ * (env-var fallbacks, migration corner cases, etc.).
+ */
+
+// Mirrors `src/security/store-path-override.ts`. Duplicated by design — see
+// the "No source-module imports" section above.
+type PathSlot = {
+  storePath: string | null;
+  storeKeyPath: string | null;
+};
+
+type VellumAssistantNamespace = {
+  storePathOverride?: PathSlot;
+};
+
+function pathSlot(): PathSlot {
+  const g = globalThis as { vellumAssistant?: VellumAssistantNamespace };
+  const ns = (g.vellumAssistant ??= {});
+  return (ns.storePathOverride ??= { storePath: null, storeKeyPath: null });
+}
+
+/**
+ * Override the encrypted store file path. Pass `null` to reset to the
+ * default (`<protectedDir>/keys.enc`).
+ */
+export function setStorePathForTesting(path: string | null): void {
+  pathSlot().storePath = path;
+}
+
+/**
+ * Override the store-key file path. Pass `null` to reset to the default
+ * (`<dirname(storePath)>/store.key`).
+ */
+export function setStoreKeyPathForTesting(path: string | null): void {
+  pathSlot().storeKeyPath = path;
+}

package/src/__tests__/encrypted-store.test.ts

@@ -24,7 +24,7 @@ import {
 } from "bun:test";
 
 // ---------------------------------------------------------------------------
-// Mock only the logger (not platform -- we use _setStorePath instead)
+// Mock only the logger (not platform -- we use setStorePathForTesting instead)
 // ---------------------------------------------------------------------------
 
 mock.module("../util/logger.js", () => ({
@@ -35,13 +35,15 @@ mock.module("../util/logger.js", () => ({
 }));
 
 import {
-  _setStoreKeyPath,
-  _setStorePath,
   deleteKey,
   getKey,
   listKeys,
   setKey,
 } from "../security/encrypted-store.js";
+import {
+  setStoreKeyPathForTesting,
+  setStorePathForTesting,
+} from "./encrypted-store-test-helpers.js";
 
 // ---------------------------------------------------------------------------
 // Use a temp directory so tests don't touch the real ~/.vellum
@@ -141,13 +143,13 @@ describe("encrypted-store", () => {
     for (const entry of readdirSync(TEST_DIR)) {
       rmSync(join(TEST_DIR, entry), { recursive: true, force: true });
     }
-    _setStorePath(STORE_PATH);
-    _setStoreKeyPath(STORE_KEY_PATH);
+    setStorePathForTesting(STORE_PATH);
+    setStoreKeyPathForTesting(STORE_KEY_PATH);
   });
 
   afterEach(() => {
-    _setStorePath(null);
-    _setStoreKeyPath(null);
+    setStorePathForTesting(null);
+    setStoreKeyPathForTesting(null);
   });
 
   afterAll(() => {
@@ -508,8 +510,8 @@ describe("encrypted-store", () => {
       // Point to a path in a non-existent subdirectory
       const nestedPath = join(TEST_DIR, "sub", "dir", "keys.enc");
       const nestedKeyPath = join(TEST_DIR, "sub", "dir", "store.key");
-      _setStorePath(nestedPath);
-      _setStoreKeyPath(nestedKeyPath);
+      setStorePathForTesting(nestedPath);
+      setStoreKeyPathForTesting(nestedKeyPath);
       const result = setKey("test", "value");
       expect(result).toBe(true);
       expect(getKey("test")).toBe("value");

package/src/__tests__/feature-flag-test-helpers.ts

@@ -0,0 +1,53 @@
+/**
+ * Test-only utilities for seeding the feature flag override cache.
+ *
+ * Replaces the removed `_setOverridesForTesting` export from
+ * `assistant-feature-flags.ts`. Lives here (not in the source module)
+ * because production modules should not expose test backdoors.
+ *
+ * No source-module imports
+ * ------------------------
+ * This file has ZERO imports from `src/`. It accesses the feature flag
+ * cache's state via the shared `globalThis.vellumAssistant.featureFlagCache`
+ * slot that `src/config/feature-flag-cache.ts` also reads/writes. The slot
+ * shape is duplicated here on purpose: keeping this file off the
+ * production import graph is what protects the test preload from a
+ * broken `node_modules` symlink (DB ghost #3). The two declarations MUST
+ * stay in sync — if you change one, change the other.
+ */
+
+// Mirrors `src/config/feature-flag-cache.ts`. Duplicated by design — see
+// the "No source-module imports" section above.
+type FlagSlot = {
+  overrides: Record<string, boolean> | null;
+  fromGateway: boolean;
+};
+
+type VellumAssistantNamespace = {
+  featureFlagCache?: FlagSlot;
+};
+
+function flagSlot(): FlagSlot {
+  const g = globalThis as { vellumAssistant?: VellumAssistantNamespace };
+  const ns = (g.vellumAssistant ??= {});
+  return (ns.featureFlagCache ??= { overrides: null, fromGateway: false });
+}
+
+/**
+ * Synchronously seed the feature flag override cache for the current test.
+ *
+ * Sets the cache to a clone of `overrides` and marks it as
+ * gateway-populated, so subsequent `initFeatureFlagOverrides()` calls are
+ * no-ops (preventing the production retry loop from running during tests).
+ *
+ * Tests that want the gateway IPC retry path to actually run should not
+ * call this — they should leave the cache empty or call
+ * `clearFeatureFlagOverridesCache()` from `assistant-feature-flags.ts`.
+ */
+export function setOverridesForTesting(
+  overrides: Record<string, boolean>,
+): void {
+  const s = flagSlot();
+  s.overrides = { ...overrides };
+  s.fromGateway = true;
+}

package/src/__tests__/filing-service.test.ts

@@ -83,6 +83,7 @@ mock.module("../memory/conversation-crud.js", () => ({
     createdConversations.push(opts);
     return { id: `conv-${++conversationIdCounter}`, ...opts };
   },
+  reserveMessage: mock(async () => ({ id: "msg-reserve" })),
 }));
 
 // Mock logger
@@ -154,7 +155,7 @@ describe("FilingService", () => {
         conversationId: args[0] as string,
         content: args[1] as string,
         options:
-          (args[3] as { speed?: string; callSite?: string } | undefined) ??
+          (args[2] as { speed?: string; callSite?: string } | undefined) ??
           undefined,
       });
       return { messageId: "msg-1" };
@@ -318,7 +319,7 @@ describe("FilingService", () => {
       let compactionCalls = 0;
 
       setTestProcessMessage((...args: unknown[]) => {
-        const callSite = (args[3] as { callSite?: string } | undefined)
+        const callSite = (args[2] as { callSite?: string } | undefined)
           ?.callSite;
         if (callSite === "filingAgent") {
           filingCalls += 1;

package/src/__tests__/first-greeting.test.ts

@@ -8,6 +8,7 @@ const {
   isWakeUpGreeting,
   getCannedFirstGreeting,
   buildScanFirstMessage,
+  buildSelfIntroMessage,
   CANNED_FIRST_GREETING,
 } = await import("../daemon/first-greeting.js");
 import type { OnboardingGreetingContext } from "../daemon/first-greeting.js";
@@ -96,7 +97,7 @@ describe("first-greeting", () => {
         assistantName: "Pax",
       });
       expect(greeting).toBe(
-        "Hey Alice, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
+        "Hey Alice, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
       );
     });
 
@@ -108,7 +109,7 @@ describe("first-greeting", () => {
         assistantName: "Remy",
       });
       expect(greeting).toBe(
-        "Hey Alice, I'm Remy. Good to meet you.\n\nWe can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here.",
+        "Hey Alice, I'm Remy. Good to meet you.\n\nWe can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here. And you don't have to start me from scratch — if there's a ChatGPT or Claude that already knows you, bring it over and I'll get up to speed fast. Honestly, it's the best head start you could give me.",
       );
     });
 
@@ -119,7 +120,7 @@ describe("first-greeting", () => {
         assistantName: "Pax",
       });
       expect(greeting).toBe(
-        "Hey, I'm Pax. Let's see what you've got.\n\nWe can jump straight into whatever you've got, or take a few minutes to just talk first. What sounds right?",
+        "Hey, I'm Pax. Let's see what you've got.\n\nWe can jump straight into whatever you've got, or take a few minutes to just talk first. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll get up to speed fast. Best head start you can give me — want to start there?",
       );
     });
 
@@ -131,7 +132,7 @@ describe("first-greeting", () => {
         assistantName: "Pax",
       });
       expect(greeting).toBe(
-        "Hey Alice, I'm Pax.\n\nWe can start with whatever's in front of you, or just talk for a bit first. Either way.",
+        "Hey Alice, I'm Pax.\n\nWe can start with whatever's in front of you, or just talk for a bit first. Either way. And you don't have to start me from nothing — if there's a ChatGPT or Claude that already knows you, bring it over and I'll learn from it. The best head start you could give me.",
       );
     });
 
@@ -142,7 +143,7 @@ describe("first-greeting", () => {
         userName: "Alice",
       });
       expect(greeting).toBe(
-        "Hey Alice,\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
+        "Hey Alice,\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
       );
     });
 
@@ -153,7 +154,7 @@ describe("first-greeting", () => {
         assistantName: "Pax",
       });
       expect(greeting).toBe(
-        "Hey, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
+        "Hey, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
       );
     });
 
@@ -173,7 +174,7 @@ describe("first-greeting", () => {
         tone: "warm",
       });
       expect(greeting).toBe(
-        "Hey,\n\nWe can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here.",
+        "Hey,\n\nWe can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here. And you don't have to start me from scratch — if there's a ChatGPT or Claude that already knows you, bring it over and I'll get up to speed fast. Honestly, it's the best head start you could give me.",
       );
     });
 
@@ -185,6 +186,17 @@ describe("first-greeting", () => {
       expect(unique.size).toBe(4);
     });
 
+    it("google-connected greeting points to Gmail only", () => {
+      const greeting = getCannedFirstGreeting({
+        ...base,
+        tone: "grounded",
+        googleConnected: true,
+      });
+      expect(greeting).toContain("scan Gmail");
+      expect(greeting).not.toContain("calendar");
+      expect(greeting).not.toContain("drive");
+    });
+
     it("unknown tone falls back to grounded defaults", () => {
       const greeting = getCannedFirstGreeting({
         ...base,
@@ -193,7 +205,7 @@ describe("first-greeting", () => {
         assistantName: "Pax",
       });
       expect(greeting).toBe(
-        "Hey Alice, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
+        "Hey Alice, I'm Pax.\n\nWe can get into whatever you've got, or just talk first — that tends to go better. Up to you. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
       );
     });
 
@@ -223,7 +235,7 @@ describe("first-greeting", () => {
       const [intro, invite] = greeting.split("\n\n");
       expect(intro).toBe("Hey Bob, I'm Pax.");
       expect(invite).toBe(
-        "We can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
+        "We can get into whatever you've got, or just talk first — that tends to go better. Up to you. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
       );
     });
 
@@ -232,7 +244,7 @@ describe("first-greeting", () => {
       const [intro, invite] = greeting.split("\n\n");
       expect(intro).toBe("Hey Bob, I'm Pax. Good to meet you.");
       expect(invite).toBe(
-        "We can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here.",
+        "We can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here. And you don't have to start me from scratch — if there's a ChatGPT or Claude that already knows you, bring it over and I'll get up to speed fast. Honestly, it's the best head start you could give me.",
       );
     });
 
@@ -241,7 +253,7 @@ describe("first-greeting", () => {
       const [intro, invite] = greeting.split("\n\n");
       expect(intro).toBe("Hey Bob, I'm Pax. Let's see what you've got.");
       expect(invite).toBe(
-        "We can jump straight into whatever you've got, or take a few minutes to just talk first. What sounds right?",
+        "We can jump straight into whatever you've got, or take a few minutes to just talk first. And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll get up to speed fast. Best head start you can give me — want to start there?",
       );
     });
 
@@ -250,7 +262,7 @@ describe("first-greeting", () => {
       const [intro, invite] = greeting.split("\n\n");
       expect(intro).toBe("Hey Bob, I'm Pax.");
       expect(invite).toBe(
-        "We can start with whatever's in front of you, or just talk for a bit first. Either way.",
+        "We can start with whatever's in front of you, or just talk for a bit first. Either way. And you don't have to start me from nothing — if there's a ChatGPT or Claude that already knows you, bring it over and I'll learn from it. The best head start you could give me.",
       );
     });
 
@@ -274,6 +286,45 @@ describe("first-greeting", () => {
     });
   });
 
+  describe("migration offer is present in every variant", () => {
+    const MIGRATION_MARKER = "head start";
+
+    it("no-onboarding greeting includes the migration offer", () => {
+      expect(getCannedFirstGreeting(undefined)).toContain(MIGRATION_MARKER);
+      expect(CANNED_FIRST_GREETING).toContain(MIGRATION_MARKER);
+    });
+
+    it("minimal onboarding (falls back to CANNED) includes the migration offer", () => {
+      expect(
+        getCannedFirstGreeting({ tools: [], tasks: [], tone: "" }),
+      ).toContain(MIGRATION_MARKER);
+    });
+
+    it("every tone variant includes the migration offer", () => {
+      for (const tone of ["grounded", "warm", "energetic", "poetic"]) {
+        const greeting = getCannedFirstGreeting({
+          tools: [],
+          tasks: [],
+          tone,
+          userName: "Alice",
+          assistantName: "Pax",
+        });
+        expect(greeting).toContain(MIGRATION_MARKER);
+        expect(greeting).toContain("ChatGPT or Claude");
+      }
+    });
+
+    it("google-connected greeting still includes the migration offer", () => {
+      const greeting = getCannedFirstGreeting({
+        tools: [],
+        tasks: [],
+        tone: "warm",
+        googleConnected: true,
+      });
+      expect(greeting).toContain(MIGRATION_MARKER);
+    });
+  });
+
   describe("buildScanFirstMessage", () => {
     it("website variant includes 'my website' and the URL", () => {
       const msg = buildScanFirstMessage("https://acme.com", "website");
@@ -325,4 +376,44 @@ describe("first-greeting", () => {
       expect(greeting).not.toContain("personal");
     });
   });
+
+  describe("buildSelfIntroMessage", () => {
+    const ctx = (
+      over: Partial<OnboardingGreetingContext> = {},
+    ): OnboardingGreetingContext => ({
+      tools: [],
+      tasks: [],
+      tone: "grounded",
+      ...over,
+    });
+
+    it("uses both names when present", () => {
+      expect(
+        buildSelfIntroMessage(ctx({ assistantName: "Vela", userName: "alex" })),
+      ).toBe("Hi Vela, I'm alex. Nice to meet you.");
+    });
+
+    it("drops the missing user name", () => {
+      expect(buildSelfIntroMessage(ctx({ assistantName: "Vela" }))).toBe(
+        "Hi Vela. Nice to meet you.",
+      );
+    });
+
+    it("drops the missing assistant name", () => {
+      expect(buildSelfIntroMessage(ctx({ userName: "alex" }))).toBe(
+        "Hi, I'm alex. Nice to meet you.",
+      );
+    });
+
+    it("treats whitespace-only names as missing", () => {
+      expect(
+        buildSelfIntroMessage(ctx({ assistantName: "  ", userName: "  " })),
+      ).toBeUndefined();
+    });
+
+    it("returns undefined when neither name is known (caller keeps canned greeting)", () => {
+      expect(buildSelfIntroMessage(ctx())).toBeUndefined();
+      expect(buildSelfIntroMessage(undefined)).toBeUndefined();
+    });
+  });
 });

package/src/__tests__/gateway-flag-listener.test.ts

@@ -33,7 +33,6 @@ mock.module("../config/assistant-feature-flags.js", () => ({
   initFeatureFlagOverrides: async () => {},
   clearFeatureFlagOverridesCache: () => {},
   isAssistantFeatureFlagEnabled: () => true,
-  _setOverridesForTesting: () => {},
 }));
 
 // ---------------------------------------------------------------------------

package/src/__tests__/gemini-inline-media.test.ts

@@ -0,0 +1,78 @@
+import { describe, expect, test } from "bun:test";
+
+import {
+  base64ByteLength,
+  estimateGeminiAudioTokens,
+  GEMINI_MAX_INLINE_AUDIO_BYTES,
+  normalizeGeminiAudioMime,
+} from "../providers/gemini/inline-media.js";
+
+describe("normalizeGeminiAudioMime", () => {
+  test("maps audio/mpeg onto Gemini's audio/mp3 spelling", () => {
+    expect(normalizeGeminiAudioMime("audio/mpeg")).toBe("audio/mp3");
+  });
+
+  test("passes supported audio types through unchanged", () => {
+    for (const mime of [
+      "audio/wav",
+      "audio/mp3",
+      "audio/aiff",
+      "audio/aac",
+      "audio/ogg",
+      "audio/flac",
+    ]) {
+      expect(normalizeGeminiAudioMime(mime)).toBe(mime);
+    }
+  });
+
+  test("is case-insensitive and strips parameters", () => {
+    expect(normalizeGeminiAudioMime("AUDIO/MPEG")).toBe("audio/mp3");
+    expect(normalizeGeminiAudioMime("audio/ogg; codecs=opus")).toBe(
+      "audio/ogg",
+    );
+  });
+
+  test("returns null for types Gemini cannot take inline", () => {
+    for (const mime of [
+      "audio/x-m4a",
+      "audio/mp4",
+      "audio/opus",
+      "application/pdf",
+      "image/png",
+      "",
+    ]) {
+      expect(normalizeGeminiAudioMime(mime)).toBeNull();
+    }
+  });
+});
+
+describe("base64ByteLength", () => {
+  test("approximates raw bytes from base64 length", () => {
+    // 8 base64 chars → 6 raw bytes
+    expect(base64ByteLength("QUJDREVG")).toBe(6);
+    expect(base64ByteLength("")).toBe(0);
+  });
+});
+
+describe("estimateGeminiAudioTokens", () => {
+  test("scales with payload size and stays far below the base64-as-text count", () => {
+    // ~3 MB of base64 ≈ ~2.25 MB raw ≈ ~140s at 16 KB/s ≈ ~4.5k tokens.
+    const data = "A".repeat(3 * 1024 * 1024);
+    const tokens = estimateGeminiAudioTokens(data);
+    expect(tokens).toBeGreaterThan(1_000);
+    // Must be a small fraction of the naive base64-length/4 estimate (~786k).
+    expect(tokens).toBeLessThan(data.length / 40);
+  });
+
+  test("returns zero for empty data", () => {
+    expect(estimateGeminiAudioTokens("")).toBe(0);
+  });
+});
+
+test("inline audio cap leaves headroom under Gemini's 20 MB request limit", () => {
+  expect(GEMINI_MAX_INLINE_AUDIO_BYTES).toBe(12 * 1024 * 1024);
+  // Base64 of the cap (~16 MB) must stay under the 20 MB wire limit.
+  expect(GEMINI_MAX_INLINE_AUDIO_BYTES * (4 / 3)).toBeLessThan(
+    20 * 1024 * 1024,
+  );
+});