@vellumai/assistant 0.9.0 → 0.10.0-dev.202606192226.8ed8a6a

package/src/__tests__/injector-disk-pressure.test.ts

@@ -204,7 +204,7 @@ describe("disk-pressure-warning injector", () => {
       .run();
   });
 
-  test("emits the exact cleanup prompt during disk pressure cleanup mode", async () => {
+  test("emits the concise cleanup skill prompt during disk pressure cleanup mode", async () => {
     seedDiskPressure(true);
     const block = await diskPressureInjector.produce(makeContext());
 
@@ -217,14 +217,19 @@ describe("disk-pressure-warning injector", () => {
       DEFAULT_INJECTOR_ORDER.diskPressureWarning,
     );
     expect(DISK_PRESSURE_WARNING_PROMPT).toBe(`<disk_pressure_warning>
-Disk usage is critically low: this assistant is in storage cleanup mode because the workspace volume is critically full.
+Storage is critically low and normal work is suspended until space is freed.
 
-In your first paragraph, warn the user that storage is critically low and that normal work is suspended until space is freed.
+Your first user-visible paragraph must warn the user that storage is critically low and normal work is suspended.
 
-Then help the user clean up storage. Prefer safe inspection steps first, such as checking available space and finding large directories. Ask before deleting files or caches unless the user has already clearly approved the specific cleanup action.
+Before taking cleanup actions, call \`skill_load\` with \`skill: "system-storage-cleanup"\` and follow the cleanup skill.
 
-Do not work on unrelated tasks until enough space is freed to clear the lock or the user explicitly overrides it. Background processes and messages from trusted contacts are blocked while this cleanup mode is active.
+Unrelated work remains blocked until disk usage drops below the critical threshold or the guardian explicitly overrides the lock. Background processes and trusted-contact messages remain blocked while this cleanup mode is active.
 </disk_pressure_warning>`);
+    expect(DISK_PRESSURE_WARNING_PROMPT).toContain("skill_load");
+    expect(DISK_PRESSURE_WARNING_PROMPT).toContain("system-storage-cleanup");
+    expect(DISK_PRESSURE_WARNING_PROMPT).not.toContain(
+      "Prefer safe inspection steps first",
+    );
   });
 
   test("omits the prompt when no cleanup context is registered or it is inactive", async () => {

package/src/__tests__/injector-v3-suppression.test.ts

@@ -2,7 +2,7 @@
  * Tests for the memory-v3 step-0 branch in `applyRuntimeInjections`:
  * spotlight strip + v2 tail suppression.
  *
- * When the `memory-v3-live` flag is on AND the v3 injector (id `memory-v3`,
+ * When `memory.v3.live` is on AND the v3 injector (id `memory-v3`,
  * placement `after-memory-prefix`) produces a block — possibly EMPTY-TEXT on
  * an all-repeat turn — runtime assembly strips the v2 `<memory>` prefix from
  * the TAIL user message only before splicing the v3 block. Historical user
@@ -16,10 +16,11 @@
  * block lands at the tail.
  *
  * v2 suppression stays keyed off whether v3 produced a block, NOT off the
- * flag alone: a v3 failure (`produce()` → null) leaves v2's block intact
- * (fallback-to-v2). The flag-off path must be byte-for-byte identical to
+ * gate alone: a v3 failure (`produce()` → null) leaves v2's block intact
+ * (fallback-to-v2). The v3-off path must be byte-for-byte identical to
  * today — that is the load-bearing regression guard. `applyRuntimeInjections`
- * reads the flag itself, so these tests drive it through the override cache.
+ * reads the v3-live gate itself, so these tests drive it through the
+ * `isMemoryV3Live` mock slot.
  *
  * The strip discriminates v2's dynamic block by IDENTITY, not by prefix: v2's
  * `INJECTION_HEADER` and v3's `V3_CARDS_INJECTION_HEADER` are deliberately
@@ -42,7 +43,6 @@ import type {
   TurnContext,
 } from "../plugins/types.js";
 import type { Message } from "../providers/types.js";
-import { setOverridesForTesting } from "./feature-flag-test-helpers.js";
 
 // Drive the suppression branch by controlling the static injector chain that
 // `applyRuntimeInjections` walks. The slot is mutated per-test to stand in for
@@ -52,6 +52,13 @@ mock.module("../plugins/defaults/memory-retrieval/injector-chain.js", () => ({
   getInjectorChain: () => injectorChainSlot,
 }));
 
+// `applyRuntimeInjections` reads the v3-live gate (`config.memory.v3.live`)
+// via `isMemoryV3Live`; drive it directly through this slot per-test.
+let memoryV3LiveSlot = false;
+mock.module("../config/memory-v3-gate.js", () => ({
+  isMemoryV3Live: () => memoryV3LiveSlot,
+}));
+
 const { applyRuntimeInjections } =
   await import("../daemon/conversation-runtime-assembly.js");
 
@@ -141,9 +148,9 @@ function tailTexts(messages: Message[]): string[] {
 describe("memory-v3-live v2 suppression", () => {
   beforeEach(() => {
     injectorChainSlot.length = 0;
-    // Clean baseline: no overrides → `memory-v3-live` resolves to its registry
-    // default (off). Each test seeds the flag it needs.
-    setOverridesForTesting({});
+    // Clean baseline: v3-live off (registry/config default). Each test seeds
+    // the gate value it needs.
+    memoryV3LiveSlot = false;
   });
 
   afterEach(() => {
@@ -152,7 +159,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("flag ON + v3 produced a block → TAIL v2 stripped, historical memory blocks frozen in place", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector("net-new cards"));
     seedV2Identity("fresh recalled fact");
 
@@ -194,7 +201,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("flag ON + EMPTY-TEXT v3 block (all-repeat turn) → tail v2 stripped, nothing attached", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector(""));
     seedV2Identity("fresh recalled fact");
 
@@ -217,7 +224,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("stale spotlight blocks are stripped from EVERY user message; the new spotlight lands at the tail", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector(""), spotlightInjector("fresh sections"));
 
     const staleSpotlight = {
@@ -253,7 +260,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("convergence re-entry: a tail leading with this turn's frozen v3 cards (and <info>) is NOT stripped", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     // Re-entry shape: the memo returns the same selections, every slug is now
     // in the everInjected store, so the injector produces an EMPTY block —
     // while the tail still carries the v3 card block frozen on first entry
@@ -292,7 +299,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("first entry with a v2 prefix AND a frozen v3 block strips ONLY the v2 block", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector(""));
     seedV2Identity("fresh recalled fact");
 
@@ -319,7 +326,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("REGRESSION: a v2 block leading with the REAL summary header (byte-identical to v3's) is stripped; v3 cards and <info> survive (first entry)", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector(""));
 
     // The collision this guards: v2's router block leads with
@@ -358,7 +365,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("REGRESSION: re-entry with a header-bearing v2 identity keeps the first entry's frozen v3 cards", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     // Re-entry: produce() returns the EMPTY block (cards already claimed by
     // the store) while the tail carries the FIRST entry's v3 block — and the
     // graph handle still holds the summary-bearing v2 identity from this
@@ -391,7 +398,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("v2 memory-image groups and legacy <memory __injected> blocks are stripped even without an identity", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector("net-new cards"));
     // No live graph handle: identity unknown (null). Image groups and legacy
     // blocks are unambiguously v2's and are stripped regardless; the shared
@@ -430,7 +437,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("the v3 block's attachment-commit callback fires on a user tail", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     const commit = mock(() => {});
     injectorChainSlot.push(v3Injector("net-new cards", commit));
 
@@ -443,7 +450,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("the commit callback is SKIPPED when the tail is not a user message (block never attaches)", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     const commit = mock(() => {});
     injectorChainSlot.push(v3Injector("net-new cards", commit));
 
@@ -465,7 +472,7 @@ describe("memory-v3-live v2 suppression", () => {
   });
 
   test("flag ON but v3 produced NOTHING → v2 block left intact (fallback-to-v2)", async () => {
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     injectorChainSlot.push(v3Injector(null));
 
     const runMessages: Message[] = [
@@ -524,7 +531,7 @@ describe("memory-v3-live v2 suppression", () => {
   test("no v3 injector registered + flag ON → no stripping, messages untouched", async () => {
     // No injector named memory-v3 at all (e.g. plugin not loaded): the
     // suppression branch keys off the produced block, so nothing is stripped.
-    setOverridesForTesting({ "memory-v3-live": true });
+    memoryV3LiveSlot = true;
     expect(injectorChainSlot).toHaveLength(0);
 
     const runMessages: Message[] = [

package/src/__tests__/internal-telemetry-routes.test.ts

@@ -7,23 +7,15 @@ mock.module("../util/logger.js", () => ({
     }),
 }));
 
-// Toggle for the collectUsageData opt-out the real store consults. The store
+// Toggle for the share_analytics opt-out the real store consults. The store
 // module is intentionally NOT mocked here — it has its own DB-backed tests, and
 // Bun's `mock.module` is process-global, so mocking it would leak into those
 // tests when files share an invocation. Exercising the real store keeps every
 // auth-fallback test order-independent.
-let collectUsageData = true;
+let shareAnalytics = true;
 
-mock.module("../config/loader.js", () => ({
-  getConfig: () => ({
-    ui: {},
-    model: "test",
-    provider: "test",
-    memory: { enabled: false },
-    rateLimit: { maxRequestsPerMinute: 0 },
-    secretDetection: { enabled: false },
-    collectUsageData,
-  }),
+mock.module("../platform/consent-cache.js", () => ({
+  getCachedShareAnalytics: () => shareAnalytics,
 }));
 
 import { queryUnreportedAuthFallbackEvents } from "../memory/auth-fallback-events-store.js";
@@ -61,7 +53,7 @@ const VALID_BODY = {
 
 describe("internal-telemetry-routes: auth-fallback", () => {
   beforeEach(() => {
-    collectUsageData = true;
+    shareAnalytics = true;
     getDb().delete(authFallbackEvents).run();
   });
 
@@ -90,7 +82,7 @@ describe("internal-telemetry-routes: auth-fallback", () => {
   });
 
   test("returns skipped and persists nothing under the opt-out", () => {
-    collectUsageData = false;
+    shareAnalytics = false;
     expect(call(VALID_BODY)).toEqual({ skipped: true });
     expect(queryUnreportedAuthFallbackEvents(0, undefined, 100).length).toBe(0);
   });

package/src/__tests__/invite-redemption-service.test.ts

@@ -83,7 +83,7 @@ describe("invite-redemption-service", () => {
 
     const result = findContactChannel({
       channelType: "telegram",
-      externalUserId: "user-1",
+      address: "user-1",
     });
 
     expect(result).not.toBeNull();
@@ -112,7 +112,7 @@ describe("invite-redemption-service", () => {
 
     const result = findContactChannel({
       channelType: "telegram",
-      externalUserId: "code-user-1",
+      address: "code-user-1",
     });
 
     expect(result).not.toBeNull();
@@ -292,7 +292,6 @@ describe("invite-redemption-service", () => {
         {
           type: "telegram",
           address: "guardian-tg-id",
-          externalUserId: "guardian-tg-id",
           status: "revoked",
         },
       ],
@@ -325,7 +324,7 @@ describe("invite-redemption-service", () => {
     // Verify the redeemer's Telegram ID is now bound to Mom's contact
     const result = findContactChannel({
       channelType: "telegram",
-      externalUserId: "guardian-tg-id",
+      address: "guardian-tg-id",
     });
     expect(result).not.toBeNull();
     expect(result!.contact.id).toBe(momContact.id);
@@ -346,7 +345,6 @@ describe("invite-redemption-service", () => {
         {
           type: "telegram",
           address: "guardian-own-id",
-          externalUserId: "guardian-own-id",
           status: "revoked",
         },
       ],
@@ -381,7 +379,6 @@ describe("invite-redemption-service", () => {
         {
           type: "telegram",
           address: "guardian-code-id",
-          externalUserId: "guardian-code-id",
           status: "revoked",
         },
       ],
@@ -417,7 +414,7 @@ describe("invite-redemption-service", () => {
     // Verify the redeemer's Telegram ID is now bound to Mom's contact
     const result = findContactChannel({
       channelType: "telegram",
-      externalUserId: "guardian-code-id",
+      address: "guardian-code-id",
     });
     expect(result).not.toBeNull();
     expect(result!.contact.id).toBe(momContact.id);

package/src/__tests__/llm-callsite-catalog.test.ts

@@ -54,9 +54,7 @@ describe("LLM call-site catalog", () => {
 
   test("CALL_SITE_DEFAULTS covers every LLMCallSite enum value", () => {
     const defaultIds = new Set(Object.keys(CALL_SITE_DEFAULTS));
-    const missing = LLMCallSiteEnum.options.filter(
-      (id) => !defaultIds.has(id),
-    );
+    const missing = LLMCallSiteEnum.options.filter((id) => !defaultIds.has(id));
     expect(missing).toEqual([]);
   });
 
@@ -68,11 +66,12 @@ describe("LLM call-site catalog", () => {
     expect(extra).toEqual([]);
   });
 
-  test("every CALL_SITE_DEFAULTS entry has a profile field", () => {
+  test("every CALL_SITE_DEFAULTS profile, when present, is a non-empty string", () => {
     for (const [, config] of Object.entries(CALL_SITE_DEFAULTS)) {
-      expect(config.profile).toBeDefined();
+      // A call site may omit `profile` to inherit the workspace default config.
+      if (config.profile === undefined) continue;
       expect(typeof config.profile).toBe("string");
-      expect(config.profile!.length).toBeGreaterThan(0);
+      expect(config.profile.length).toBeGreaterThan(0);
     }
   });
 });

package/src/__tests__/llm-catalog-parity.test.ts

@@ -2,7 +2,10 @@ import { readFileSync } from "node:fs";
 import { join } from "node:path";
 import { describe, expect, test } from "bun:test";
 
-import { PROVIDER_CATALOG } from "../providers/model-catalog.js";
+import {
+  isModelInCatalog,
+  PROVIDER_CATALOG,
+} from "../providers/model-catalog.js";
 import { PLATFORM_PROVIDER_META } from "../providers/platform-proxy/constants.js";
 import { resolvePricing, resolvePricingForUsage } from "../util/pricing.js";
 
@@ -30,13 +33,6 @@ function getRepoRoot(): string {
 }
 
 const META_JSON_PATH = join(getRepoRoot(), "meta", "llm-provider-catalog.json");
-const SWIFTPM_MIRROR_PATH = join(
-  getRepoRoot(),
-  "clients",
-  "shared",
-  "Resources",
-  "llm-provider-catalog.json",
-);
 
 interface ClientCatalogCredentialsGuide {
   description: string;
@@ -309,6 +305,32 @@ describe("LLM catalog parity: daemon vs client", () => {
     });
   });
 
+  test("Fireworks catalog includes GLM 5.2", () => {
+    expect(
+      isModelInCatalog("fireworks", "accounts/fireworks/models/glm-5p2"),
+    ).toBe(true);
+
+    const fireworks = PROVIDER_CATALOG.find(
+      (entry) => entry.id === "fireworks",
+    );
+    expect(
+      fireworks?.models.find(
+        (model) => model.id === "accounts/fireworks/models/glm-5p2",
+      ),
+    ).toMatchObject({
+      displayName: "GLM 5.2",
+      contextWindowTokens: 1040000,
+      maxOutputTokens: 131072,
+      supportsToolUse: true,
+      supportsVision: false,
+      pricing: {
+        inputPer1mTokens: 1.4,
+        outputPer1mTokens: 4.4,
+        cacheReadPer1mTokens: 0.26,
+      },
+    });
+  });
+
   test("MiniMax catalog includes MiniMax M3", () => {
     const minimax = PROVIDER_CATALOG.find((entry) => entry.id === "minimax");
     expect(
@@ -457,19 +479,4 @@ describe("LLM catalog parity: daemon vs client", () => {
       longContextMode: "native-model",
     });
   });
-
-  // -----------------------------------------------------------------------
-  // Mirror byte-equality
-  // -----------------------------------------------------------------------
-
-  test("SwiftPM mirror is byte-identical to meta/ copy", () => {
-    // `sync-llm-catalog.ts` writes both files from the same serializer; this
-    // guard catches any case where one copy is regenerated without the other.
-    // Byte equality is required because SwiftPM bundles the resource verbatim
-    // into `VellumAssistantShared` while the meta/ JSON is consumed as a
-    // cross-package artifact (web codegen, etc.).
-    const metaBytes = readFileSync(META_JSON_PATH);
-    const swiftPmBytes = readFileSync(SWIFTPM_MIRROR_PATH);
-    expect(swiftPmBytes.equals(metaBytes)).toBe(true);
-  });
 });

package/src/__tests__/llm-resolver.test.ts

@@ -16,6 +16,7 @@ const fullDefault = {
   speed: "standard" as const,
   verbosity: "medium" as const,
   temperature: null,
+  topP: null,
   thinking: { enabled: true, streamThinking: true },
   contextWindow: {
     enabled: true,
@@ -203,6 +204,26 @@ describe("resolveCallSiteConfig", () => {
     expect(resolved.effort).toBe("max");
   });
 
+  test("topP defaults to null when no profile or override sets it", () => {
+    const llm = LLMSchema.parse({ default: fullDefault });
+    const resolved = resolveCallSiteConfig("mainAgent", llm);
+    expect(resolved.topP).toBeNull();
+  });
+
+  test("profile-level topP resolves onto the merged config", () => {
+    const llm = LLMSchema.parse({
+      default: fullDefault,
+      profiles: {
+        nucleus: { topP: 0.9 },
+      },
+      callSites: {
+        memoryExtraction: { profile: "nucleus" },
+      },
+    });
+    const resolved = resolveCallSiteConfig("memoryExtraction", llm);
+    expect(resolved.topP).toBe(0.9);
+  });
+
   test("returns isolated nested objects (not aliased to llm.default)", () => {
     // Resolve a call site that has no override touching `thinking` or
     // `contextWindow` — the bug being guarded against would have those
@@ -670,30 +691,6 @@ describe("resolveCallSiteConfig", () => {
     expect(resolved.effort).toBe("medium");
   });
 
-  test("workflowLeaf defaults to the cost-optimized profile", () => {
-    const llm = LLMSchema.parse({
-      default: fullDefault,
-      profiles: {
-        balanced: {
-          model: "claude-sonnet-4-7",
-          effort: "medium",
-        },
-        "cost-optimized": {
-          model: "claude-haiku-4-5-20251001",
-          effort: "low",
-        },
-      },
-    });
-
-    expect(resolveDefaultProfileKey("workflowLeaf", llm)).toBe(
-      "cost-optimized",
-    );
-    const resolved = resolveCallSiteConfig("workflowLeaf", llm);
-    expect(resolved.model).toBe("claude-haiku-4-5-20251001");
-    expect(resolved.effort).toBe("low");
-    expect(resolved.thinking.enabled).toBe(false);
-  });
-
   test("explicit callSites config overrides CALL_SITE_DEFAULTS", () => {
     const llm = LLMSchema.parse({
       default: fullDefault,
@@ -1437,3 +1434,52 @@ describe("resolveCallSiteConfig logitBias provenance", () => {
     expect(resolveCallSiteConfig("mainAgent", llm).logitBias).toBeUndefined();
   });
 });
+
+describe("resolveCallSiteConfig — workflowLeaf default", () => {
+  test("inherits the workspace default config rather than pinning cost-optimized", () => {
+    const llm = LLMSchema.parse({
+      default: fullDefault,
+      profiles: {
+        "cost-optimized": {
+          provider: "anthropic",
+          model: "claude-haiku-4-5-20251001",
+        },
+      },
+    });
+    const resolved = resolveCallSiteConfig("workflowLeaf", llm);
+    // No pinned profile → the model comes from llm.default, NOT the
+    // `cost-optimized` profile (which is uncredentialed on a BYOK install).
+    expect(resolved.model).toBe("claude-opus-4-7");
+    // Call-site tuning still applies.
+    expect(resolved.effort).toBe("low");
+    expect(resolved.thinking?.enabled).toBe(false);
+    // The call site has no implicit default profile key.
+    expect(resolveDefaultProfileKey("workflowLeaf", llm)).toBeUndefined();
+  });
+
+  test("honors an explicit workflowLeaf call-site override", () => {
+    const llm = LLMSchema.parse({
+      default: fullDefault,
+      profiles: {
+        cheap: { provider: "anthropic", model: "claude-haiku-4-5-20251001" },
+      },
+      callSites: { workflowLeaf: { profile: "cheap" } },
+    });
+    expect(resolveCallSiteConfig("workflowLeaf", llm).model).toBe(
+      "claude-haiku-4-5-20251001",
+    );
+  });
+
+  test("honors a per-call override profile (an explicit per-leaf profile)", () => {
+    const llm = LLMSchema.parse({
+      default: fullDefault,
+      profiles: {
+        fancy: { provider: "anthropic", model: "claude-sonnet-4-7" },
+      },
+    });
+    expect(
+      resolveCallSiteConfig("workflowLeaf", llm, { overrideProfile: "fancy" })
+        .model,
+    ).toBe("claude-sonnet-4-7");
+  });
+});

package/src/__tests__/llm-schema.test.ts

@@ -78,6 +78,7 @@ describe("LLMSchema", () => {
       speed: "standard",
       verbosity: "medium",
       temperature: null,
+      topP: null,
       thinking: { enabled: true, streamThinking: true },
       contextWindow: {
         enabled: true,