@vellumai/assistant 0.8.7 → 0.8.8-dev.202606052332.17fc8ea

package/src/__tests__/conversation-runtime-workspace.test.ts

@@ -1,11 +1,12 @@
-import { beforeEach, describe, expect, test } from "bun:test";
+import { afterEach, describe, expect, test } from "bun:test";
 
-import { applyRuntimeInjections } from "../daemon/conversation-runtime-assembly.js";
-import { defaultInjectorsPlugin } from "../plugins/defaults/injectors/register.js";
 import {
-  registerPlugin,
-  resetPluginRegistryForTests,
-} from "../plugins/registry.js";
+  clearConversations,
+  setConversation,
+} from "../daemon/conversation-registry.js";
+import { applyRuntimeInjections } from "../daemon/conversation-runtime-assembly.js";
+import type { SurfaceData, SurfaceType } from "../daemon/message-protocol.js";
+import { createApp } from "../memory/app-store.js";
 import type { Message } from "../providers/types.js";
 
 // ---------------------------------------------------------------------------
@@ -16,6 +17,46 @@ function userMsg(text: string): Message {
   return { role: "user", content: [{ type: "text", text }] };
 }
 
+// `applyRuntimeInjections` synthesizes this conversation id when no
+// `turnContext` is supplied, so the `workspace-context` injector resolves the
+// live workspace block from the registry under this key.
+const FALLBACK_CONVERSATION_ID = "runtime-assembly-fallback";
+
+// Register the fallback conversation in the live registry so the runtime
+// injectors resolve their blocks from it (the orchestrator no longer threads
+// workspace or active-surface content as options).
+function registerFallbackConversation(fields: Record<string, unknown>): void {
+  setConversation(FALLBACK_CONVERSATION_ID, {
+    conversationId: FALLBACK_CONVERSATION_ID,
+    workingDir: "/sandbox",
+    // Non-dirty empty workspace by default so the workspace-context injector
+    // skips both the filesystem rescan and the DB refresh unless a test
+    // explicitly seeds a block via `workspaceTopLevelContext`.
+    workspaceTopLevelContext: "",
+    workspaceTopLevelDirty: false,
+    ...fields,
+  } as never);
+}
+
+// Seed the live conversation registry with a pre-rendered top-level block. The
+// cache is non-dirty with non-null content, so `resolveWorkspaceTopLevelContext`
+// returns it verbatim without rescanning the filesystem.
+function seedWorkspaceContext(text: string): void {
+  registerFallbackConversation({
+    workspaceTopLevelContext: text,
+    workspaceTopLevelDirty: false,
+  });
+}
+
+// Build the conversation surface-state map that `buildActiveSurfaceContext`
+// reads to render the `<active_workspace>` block.
+function makeSurfaceState(
+  surfaceId: string,
+  data: SurfaceData,
+): Map<string, { surfaceType: SurfaceType; data: SurfaceData }> {
+  return new Map([[surfaceId, { surfaceType: "dynamic_page", data }]]);
+}
+
 // ---------------------------------------------------------------------------
 // Tests
 // ---------------------------------------------------------------------------
@@ -23,26 +64,21 @@ function userMsg(text: string): Message {
 const sampleContext =
   "<workspace>\nRoot: /sandbox\nDirectories: src, lib, tests\n</workspace>";
 
-// The standalone `injectWorkspaceTopLevelContext` helper was removed in
-// G2.1. The workspace-context default injector (registered by
-// `defaultInjectorsPlugin`) now emits the workspace block as a
-// `prepend-user-tail` placement during `applyRuntimeInjections`. The suite
-// below exercises that end-to-end path instead.
+// The workspace-context default injector emits the workspace block as a
+// `prepend-user-tail` placement during `applyRuntimeInjections`. It sources
+// the rendered block from the per-conversation workspace registry and
+// (re)injects only when the block is absent from the working messages, so the
+// suite seeds the registry and exercises that end-to-end path.
 
 describe("applyRuntimeInjections — workspace top-level context", () => {
-  beforeEach(() => {
-    // Post-G2.1: workspace injection is driven by the `workspace-context`
-    // default injector, so the plugin must be registered for the chain to
-    // produce a block. Each test gets a clean registry.
-    resetPluginRegistryForTests();
-    registerPlugin(defaultInjectorsPlugin);
+  afterEach(() => {
+    clearConversations();
   });
 
-  test("injects workspace context when provided", async () => {
+  test("injects workspace context when registered", async () => {
+    seedWorkspaceContext(sampleContext);
     const messages: Message[] = [userMsg("Hello")];
-    const { messages: result } = await applyRuntimeInjections(messages, {
-      workspaceTopLevelContext: sampleContext,
-    });
+    const { messages: result } = await applyRuntimeInjections(messages, {});
 
     expect(result).toHaveLength(1);
     expect(result[0].content).toHaveLength(2);
@@ -50,22 +86,38 @@ describe("applyRuntimeInjections — workspace top-level context", () => {
     expect((result[0].content[1] as { text: string }).text).toBe("Hello");
   });
 
-  test("does not inject when workspace context is null", async () => {
+  test("does not inject when no workspace context is registered", async () => {
     const messages: Message[] = [userMsg("Hello")];
-    const { messages: result } = await applyRuntimeInjections(messages, {
-      workspaceTopLevelContext: null,
-    });
+    const { messages: result } = await applyRuntimeInjections(messages, {});
 
     expect(result).toHaveLength(1);
     expect(result[0].content).toHaveLength(1);
   });
 
+  test("does not re-inject when the workspace block is already present", async () => {
+    // GIVEN the registry holds a workspace block AND the working messages
+    // already carry that block (a normal cached turn, post-injection).
+    seedWorkspaceContext(sampleContext);
+    const messages: Message[] = [userMsg(sampleContext), userMsg("Hello")];
+
+    // WHEN injections are applied
+    const { messages: result } = await applyRuntimeInjections(messages, {});
+
+    // THEN presence detection skips the block to keep the prefix stable.
+    expect(result).toHaveLength(2);
+    expect(result[1].content).toHaveLength(1);
+    expect((result[1].content[0] as { text: string }).text).toBe("Hello");
+  });
+
   test("workspace context appears before active surface context in content", async () => {
-    const messages: Message[] = [userMsg("Hello")];
-    const { messages: result } = await applyRuntimeInjections(messages, {
-      activeSurface: { surfaceId: "sf_1", html: "<div>test</div>" },
+    registerFallbackConversation({
       workspaceTopLevelContext: sampleContext,
+      workspaceTopLevelDirty: false,
+      currentActiveSurfaceId: "sf_1",
+      surfaceState: makeSurfaceState("sf_1", { html: "<div>test</div>" }),
     });
+    const messages: Message[] = [userMsg("Hello")];
+    const { messages: result } = await applyRuntimeInjections(messages, {});
 
     // Workspace is injected last (in applyRuntimeInjections order) so it
     // prepends to whatever was already prepended by activeSurface.
@@ -79,17 +131,20 @@ describe("applyRuntimeInjections — workspace top-level context", () => {
   });
 
   test("app-backed active surface tells the model to load app-builder with the right argument", async () => {
-    const messages: Message[] = [userMsg("Edit this app")];
-    const { messages: result } = await applyRuntimeInjections(messages, {
-      activeSurface: {
-        surfaceId: "sf_1",
+    const app = createApp({
+      name: "Example App",
+      schemaJson: "{}",
+      htmlDefinition: "<div>test</div>",
+    });
+    registerFallbackConversation({
+      currentActiveSurfaceId: "sf_1",
+      surfaceState: makeSurfaceState("sf_1", {
         html: "<div>test</div>",
-        appId: "app-1",
-        appName: "Example App",
-        appFiles: [],
-      },
-      workspaceTopLevelContext: null,
+        appId: app.id,
+      }),
     });
+    const messages: Message[] = [userMsg("Edit this app")];
+    const { messages: result } = await applyRuntimeInjections(messages, {});
 
     const activeWorkspaceText = (result[0].content[0] as { text: string }).text;
     expect(activeWorkspaceText).toContain('skill: "app-builder"');
@@ -98,15 +153,14 @@ describe("applyRuntimeInjections — workspace top-level context", () => {
 });
 
 describe("applyRuntimeInjections — minimal mode skips workspace blocks", () => {
-  beforeEach(() => {
-    resetPluginRegistryForTests();
-    registerPlugin(defaultInjectorsPlugin);
+  afterEach(() => {
+    clearConversations();
   });
 
   test("minimal mode skips workspace top-level context", async () => {
+    seedWorkspaceContext(sampleContext);
     const messages: Message[] = [userMsg("Hello")];
     const { messages: result } = await applyRuntimeInjections(messages, {
-      workspaceTopLevelContext: sampleContext,
       mode: "minimal",
     });
 
@@ -116,9 +170,12 @@ describe("applyRuntimeInjections — minimal mode skips workspace blocks", () =>
   });
 
   test("minimal mode skips active surface context", async () => {
+    registerFallbackConversation({
+      currentActiveSurfaceId: "sf_1",
+      surfaceState: makeSurfaceState("sf_1", { html: "<div>test</div>" }),
+    });
     const messages: Message[] = [userMsg("Hello")];
     const { messages: result } = await applyRuntimeInjections(messages, {
-      activeSurface: { surfaceId: "sf_1", html: "<div>test</div>" },
       mode: "minimal",
     });
 
@@ -128,11 +185,14 @@ describe("applyRuntimeInjections — minimal mode skips workspace blocks", () =>
   });
 
   test("full mode (default) still includes workspace blocks", async () => {
-    const messages: Message[] = [userMsg("Hello")];
-    const { messages: result } = await applyRuntimeInjections(messages, {
+    registerFallbackConversation({
       workspaceTopLevelContext: sampleContext,
-      activeSurface: { surfaceId: "sf_1", html: "<div>test</div>" },
+      workspaceTopLevelDirty: false,
+      currentActiveSurfaceId: "sf_1",
+      surfaceState: makeSurfaceState("sf_1", { html: "<div>test</div>" }),
     });
+    const messages: Message[] = [userMsg("Hello")];
+    const { messages: result } = await applyRuntimeInjections(messages, {});
 
     expect(result[0].content).toHaveLength(3);
     expect((result[0].content[0] as { text: string }).text).toBe(sampleContext);

package/src/__tests__/conversation-slash-commands.test.ts

@@ -140,53 +140,19 @@ describe("resolveSlash command contract", () => {
   });
 });
 
-describe("resolveSlash /compact target override", () => {
-  test("plain /compact returns no override", async () => {
+describe("resolveSlash /compact", () => {
+  test("plain /compact resolves to kind=compact", async () => {
     const result = await resolveSlash("/compact");
     expect(result).toEqual({ kind: "compact" });
   });
 
-  test("/compact <integer> sets explicit token target", async () => {
+  test("/compact rejects arguments with usage hint", async () => {
     const result = await resolveSlash("/compact 30000");
-    expect(result).toEqual({
-      kind: "compact",
-      targetInputTokensOverride: 30000,
-    });
-  });
-
-  test("/compact <n>k expands to thousands", async () => {
-    const result = await resolveSlash("/compact 30k");
-    expect(result).toEqual({
-      kind: "compact",
-      targetInputTokensOverride: 30_000,
-    });
-  });
-
-  test("/compact <n>m expands to millions", async () => {
-    const result = await resolveSlash("/compact 1.5M");
-    expect(result).toEqual({
-      kind: "compact",
-      targetInputTokensOverride: 1_500_000,
-    });
-  });
-
-  test("/compact rejects malformed args with usage hint", async () => {
-    const result = await resolveSlash("/compact bogus");
     expect(result.kind).toBe("unknown");
     if (result.kind !== "unknown") throw new Error("expected unknown");
-    expect(result.message).toContain("`bogus`");
+    expect(result.message).toContain("does not take arguments");
     expect(result.message).toContain("/compact");
   });
-
-  test("/compact rejects zero", async () => {
-    const result = await resolveSlash("/compact 0");
-    expect(result.kind).toBe("unknown");
-  });
-
-  test("/compact rejects negative numbers", async () => {
-    const result = await resolveSlash("/compact -50");
-    expect(result.kind).toBe("unknown");
-  });
 });
 
 describe("resolveSlash /clean", () => {
@@ -227,9 +193,9 @@ describe("classifySlash is a pure classifier matching resolveSlash kinds", () =>
     { input: "/status", kind: "unknown" },
     { input: "/commands", kind: "unknown" },
     { input: "/compact", kind: "compact" },
-    { input: "/compact 30000", kind: "compact" },
-    { input: "/compact 30k", kind: "compact" },
-    { input: "/compact 1.5M", kind: "compact" },
+    { input: "/compact 30000", kind: "unknown" },
+    { input: "/compact 30k", kind: "unknown" },
+    { input: "/compact 1.5M", kind: "unknown" },
     { input: "/compact bogus", kind: "unknown" },
     { input: "/clean", kind: "clean" },
     { input: "  /clean  ", kind: "clean" },
@@ -239,7 +205,7 @@ describe("classifySlash is a pure classifier matching resolveSlash kinds", () =>
     { input: "/opus", kind: "unknown" },
     { input: "hello", kind: "passthrough" },
     { input: "  /compact  ", kind: "compact" },
-    { input: "  /compact 50k  ", kind: "compact" },
+    { input: "  /compact 50k  ", kind: "unknown" },
     { input: "/models foo", kind: "passthrough" },
   ];
 

package/src/__tests__/conversation-slash-queue.test.ts

@@ -228,7 +228,12 @@ mock.module("../agent/loop.js", () => ({
       const history = await new Promise<Message[]>((resolve) => {
         pendingRuns.push({ resolve, messages, onEvent });
       });
-      return { history, exitReason: null };
+      return {
+        history,
+        exitReason: null,
+        appendedNewMessages: history.length > messages.length,
+        newMessages: history.slice(messages.length),
+      };
     }
   },
 }));

package/src/__tests__/conversation-starter-routes.test.ts

@@ -339,8 +339,8 @@ describe("GET /v1/conversation-starters", () => {
 
   test("does not re-enqueue for invalid items when within cooldown period", async () => {
     const now = Date.now();
-    // Generation happened 30 seconds ago (within the 60s cooldown)
-    const recentGenAt = now - 30_000;
+    // Last attempt 30 seconds ago (within the 5-minute cooldown)
+    const recentAttemptAt = now - 30_000;
     insertStarter({
       label: "Let me check calendar",
       prompt: "Let me check what Alice has today.",
@@ -353,9 +353,13 @@ describe("GET /v1/conversation-starters", () => {
       category: "productivity",
       createdAt: now - 1,
     });
+    setCheckpoint(
+      "conversation_starters:last_attempt_at:default",
+      String(recentAttemptAt),
+    );
     setCheckpoint(
       "conversation_starters:last_gen_at:default",
-      String(recentGenAt),
+      String(recentAttemptAt),
     );
     setCheckpoint("conversation_starters:item_count_at_last_gen:default", "1");
     insertMemoryItem();
@@ -375,8 +379,8 @@ describe("GET /v1/conversation-starters", () => {
 
   test("re-enqueues for invalid items after cooldown expires", async () => {
     const now = Date.now();
-    // Generation happened 90 seconds ago (past the 60s cooldown)
-    const oldGenAt = now - 90_000;
+    // Last attempt 6 minutes ago (past the 5-minute cooldown)
+    const oldAttemptAt = now - 6 * 60_000;
     insertStarter({
       label: "Let me check calendar",
       prompt: "Let me check what Alice has today.",
@@ -389,9 +393,13 @@ describe("GET /v1/conversation-starters", () => {
       category: "productivity",
       createdAt: now - 1,
     });
+    setCheckpoint(
+      "conversation_starters:last_attempt_at:default",
+      String(oldAttemptAt),
+    );
     setCheckpoint(
       "conversation_starters:last_gen_at:default",
-      String(oldGenAt),
+      String(oldAttemptAt),
     );
     setCheckpoint("conversation_starters:item_count_at_last_gen:default", "1");
     insertMemoryItem();

package/src/__tests__/conversation-surfaces-action-delivery.test.ts

@@ -366,6 +366,90 @@ describe("surface action delivery to assistant", () => {
     expect(JSON.stringify(completeMsg)).not.toContain(largeBase64);
   });
 
+  test("choice surface broadcasts ui_surface_complete on action", async () => {
+    const sent: ServerMessage[] = [];
+    const ctx = makeContext(sent);
+
+    const showResult = await surfaceProxyResolver(ctx, "ui_show", {
+      surface_type: "choice",
+      title: "Pick an outcome",
+      data: {
+        options: [
+          { id: "inbox", title: "Clean up my inbox" },
+          { id: "calendar", title: "Plan my week" },
+        ],
+      },
+    });
+
+    expect(showResult.isError).toBe(false);
+    expect(showResult.yieldToUser).toBe(true);
+
+    const showMessage = sent.find(
+      (msg): msg is UiSurfaceShow => msg.type === "ui_surface_show",
+    ) as UiSurfaceShow;
+    const surfaceId = showMessage.surfaceId;
+    expect(ctx.pendingSurfaceActions.has(surfaceId)).toBe(true);
+
+    await handleSurfaceAction(ctx, surfaceId, "inbox", {
+      choiceId: "inbox",
+      choiceTitle: "Clean up my inbox",
+      selectedIds: ["inbox"],
+      selectedTitles: ["Clean up my inbox"],
+    });
+
+    const completeMsg = broadcastedMessages.find(
+      (m) =>
+        (m as unknown as Record<string, unknown>).type ===
+          "ui_surface_complete" &&
+        (m as unknown as Record<string, unknown>).surfaceId === surfaceId,
+    ) as unknown as Record<string, unknown> | undefined;
+    expect(completeMsg).toBeDefined();
+    expect(completeMsg?.conversationId).toBe("conv-1");
+    expect(completeMsg?.summary).toBe('User chose: "Clean up my inbox"');
+    expect(ctx.pendingSurfaceActions.has(surfaceId)).toBe(false);
+  });
+
+  test("oauth_connect surface broadcasts ui_surface_complete on action", async () => {
+    const sent: ServerMessage[] = [];
+    const ctx = makeContext(sent);
+
+    const showResult = await surfaceProxyResolver(ctx, "ui_show", {
+      surface_type: "oauth_connect",
+      title: "Connect Google",
+      data: {
+        providerKey: "google",
+        displayName: "Google",
+      },
+    });
+
+    expect(showResult.isError).toBe(false);
+    expect(showResult.yieldToUser).toBe(true);
+
+    const showMessage = sent.find(
+      (msg): msg is UiSurfaceShow => msg.type === "ui_surface_show",
+    ) as UiSurfaceShow;
+    const surfaceId = showMessage.surfaceId;
+    expect(ctx.pendingSurfaceActions.has(surfaceId)).toBe(true);
+
+    await handleSurfaceAction(ctx, surfaceId, "connect", {
+      status: "connected",
+      providerKey: "google",
+      providerLabel: "Google",
+      accountLabel: "user@example.com",
+    });
+
+    const completeMsg = broadcastedMessages.find(
+      (m) =>
+        (m as unknown as Record<string, unknown>).type ===
+          "ui_surface_complete" &&
+        (m as unknown as Record<string, unknown>).surfaceId === surfaceId,
+    ) as unknown as Record<string, unknown> | undefined;
+    expect(completeMsg).toBeDefined();
+    expect(completeMsg?.conversationId).toBe("conv-1");
+    expect(completeMsg?.summary).toBe("Connected Google: user@example.com");
+    expect(ctx.pendingSurfaceActions.has(surfaceId)).toBe(false);
+  });
+
   test("table surface does NOT broadcast ui_surface_complete (not one-shot)", async () => {
     const sent: ServerMessage[] = [];
     const ctx = makeContext(sent);

package/src/__tests__/conversation-sync-tags.test.ts

@@ -1,5 +1,3 @@
-import { readFileSync } from "node:fs";
-import { join } from "node:path";
 import { afterAll, beforeEach, describe, expect, test } from "bun:test";
 
 import {
@@ -18,6 +16,7 @@ import { assistantEventHub } from "../runtime/assistant-event-hub.js";
 import { ROUTES as CONVERSATION_LIST_ROUTES } from "../runtime/routes/conversation-list-routes.js";
 import { ROUTES as CONVERSATION_MANAGEMENT_ROUTES } from "../runtime/routes/conversation-management-routes.js";
 import type { RouteDefinition } from "../runtime/routes/types.js";
+import { publishConversationTitleChanged } from "../runtime/sync/resource-sync-events.js";
 import { resetDbForTesting } from "./db-test-helpers.js";
 import { waitFor } from "./helpers/wait-for.js";
 
@@ -122,20 +121,33 @@ describe("conversation sync tags", () => {
     ).toBe(false);
   });
 
-  test("agent-loop title updates emit metadata-only sync tags (no list umbrella)", () => {
-    const source = readFileSync(
-      join(import.meta.dir, "..", "daemon", "conversation-agent-loop.ts"),
-      "utf-8",
+  test("auto-title generation emits the typed title event and a metadata-only sync tag (no list umbrella)", async () => {
+    // Auto-title generation (first-pass on prompt submit, second-pass
+    // regeneration, bootstrap, and voice) persists via the title service and
+    // broadcasts through `publishConversationTitleChanged` — the same helper
+    // the rename route uses. Like a rename, generation is content-only: the
+    // row stays in place and only the title flips, so web patches the cached
+    // row from the typed `conversation_title_updated` event and the per-
+    // conversation metadata tag is the belt-and-suspenders signal. The list
+    // umbrella is deliberately omitted so web never redrains the paginated
+    // list for a title change.
+    const conversation = createConversation("Generating…");
+
+    const received = await captureEvents(() => {
+      publishConversationTitleChanged(conversation.id, "Generated title");
+    }, 2);
+
+    expect(received.map((event) => event.message.type)).toEqual([
+      "conversation_title_updated",
+      "sync_changed",
+    ]);
+    expect(received[1]!.message).toEqual({
+      type: "sync_changed",
+      tags: [conversationMetadataSyncTag(conversation.id)],
+    });
+    expect((received[1]!.message as { tags: string[] }).tags).not.toContain(
+      SYNC_TAGS.conversationsList,
     );
-    const titleUpdateBlocks =
-      source.match(
-        /type: "conversation_title_updated"[\s\S]{0,500}?type: "sync_changed"[\s\S]{0,250}?tags: \[[\s\S]*?\]/g,
-      ) ?? [];
-
-    expect(titleUpdateBlocks.length).toBeGreaterThanOrEqual(2);
-    for (const block of titleUpdateBlocks) {
-      expect(block).not.toContain("SYNC_TAGS.conversationsList");
-    }
   });
 
   test("create emits a sync_changed with the conversationsList umbrella tag", async () => {