@vellumai/assistant 0.8.6 → 0.8.7

package/src/__tests__/plugin-bootstrap.test.ts

@@ -32,16 +32,15 @@ mock.module("../security/secure-keys.js", () => ({
 }));
 
 import { clearFeatureFlagOverridesCache } from "../config/assistant-feature-flags.js";
-import type { AssistantConfig } from "../config/schema.js";
-import {
-  bootstrapPlugins,
-  type DaemonContext,
-} from "../daemon/external-plugins-bootstrap.js";
+import { bootstrapPlugins } from "../daemon/external-plugins-bootstrap.js";
 import { runShutdownHooks } from "../daemon/shutdown-registry.js";
 import { RiskLevel } from "../permissions/types.js";
+import { registerDefaultPlugins } from "../plugins/defaults/index.js";
 import {
+  closeRegistration,
   getInjectors,
   getMiddlewaresFor,
+  getRegisteredPlugins,
   registerPlugin,
   resetPluginRegistryForTests,
 } from "../plugins/registry.js";
@@ -51,6 +50,7 @@ import {
   PluginExecutionError,
   type PluginInitContext,
 } from "../plugins/types.js";
+import { APP_VERSION } from "../version.js";
 import { setOverridesForTesting } from "./feature-flag-test-helpers.js";
 
 // Redirect plugin storage directory creation into a per-process temp tree so
@@ -61,12 +61,6 @@ const TEST_WORKSPACE_DIR = join(
 );
 process.env.VELLUM_WORKSPACE_DIR = TEST_WORKSPACE_DIR;
 
-const fakeConfig = {} as unknown as AssistantConfig;
-const fakeCtx: DaemonContext = {
-  config: fakeConfig,
-  assistantVersion: "9.9.9-test",
-};
-
 /**
  * Test helper. Accepts the new `hooks` bag and ALSO legacy top-level
  * `init` / `onShutdown` for ergonomics — the helper merges them into a
@@ -139,7 +133,7 @@ describe("plugin bootstrap", () => {
     });
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(received).toBeDefined();
     const ctx = received!;
@@ -155,7 +149,7 @@ describe("plugin bootstrap", () => {
       join(TEST_WORKSPACE_DIR, "plugins-data", "alpha"),
     );
     expect(existsSync(ctx.pluginStorageDir)).toBe(true);
-    expect(ctx.assistantVersion).toBe("9.9.9-test");
+    expect(ctx.assistantVersion).toBe(APP_VERSION);
   });
 
   test("credential resolution: init receives the resolved value under credentials[key]", async () => {
@@ -176,7 +170,7 @@ describe("plugin bootstrap", () => {
     );
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(getSecureKeyAsyncMock).toHaveBeenCalledTimes(1);
     expect(getSecureKeyAsyncMock).toHaveBeenCalledWith("some-key");
@@ -196,7 +190,7 @@ describe("plugin bootstrap", () => {
 
     let caught: unknown;
     try {
-      await bootstrapPlugins(fakeCtx);
+      await bootstrapPlugins();
     } catch (err) {
       caught = err;
     }
@@ -228,7 +222,7 @@ describe("plugin bootstrap", () => {
 
     let caught: unknown;
     try {
-      await bootstrapPlugins(fakeCtx);
+      await bootstrapPlugins();
     } catch (err) {
       caught = err;
     }
@@ -276,7 +270,7 @@ describe("plugin bootstrap", () => {
 
     let caught: unknown;
     try {
-      await bootstrapPlugins(fakeCtx);
+      await bootstrapPlugins();
     } catch (err) {
       caught = err;
     }
@@ -304,7 +298,7 @@ describe("plugin bootstrap", () => {
       }),
     );
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     await runShutdownHooks("test-shutdown");
 
     // The last plugin to register must shut down first; the first to register
@@ -320,7 +314,38 @@ describe("plugin bootstrap", () => {
     // tool-result truncate, etc.). Just assert bootstrap completes without
     // throwing — the surface of defaults is verified in each pipeline's own
     // dedicated test file.
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
+  });
+
+  test("startup ordering: defaults registered before the window closes survive bootstrap replay", async () => {
+    /**
+     * Reproduces the daemon startup ordering the registration regression
+     * guards: defaults must register before `loadUserPlugins()` closes the
+     * window, so the `registerDefaultPlugins` replay inside `bootstrapPlugins`
+     * is swallowed by the duplicate-name check instead of throwing on the
+     * closed window and leaving every default unregistered.
+     */
+
+    // GIVEN the first-party defaults have registered while the window is open
+    // (what `initializePlugins()` does at daemon startup)
+    registerDefaultPlugins();
+
+    // AND a user plugin registers after them (what `loadUserPlugins()` does)
+    registerPlugin(buildPlugin("user-after-defaults"));
+
+    // AND the registration window has since closed
+    closeRegistration();
+
+    // WHEN bootstrap runs and replays the default registration
+    await bootstrapPlugins();
+
+    // THEN the defaults are still registered, ahead of the user plugin, so the
+    // middleware onion order is unchanged
+    const names = getRegisteredPlugins().map((p) => p.manifest.name);
+    expect(names).toContain("default-llm-call");
+    expect(names.indexOf("default-llm-call")).toBeLessThan(
+      names.indexOf("user-after-defaults"),
+    );
   });
 
   // ── requiresFlag gating (G2.2) ──────────────────────────────────────────
@@ -351,7 +376,7 @@ describe("plugin bootstrap", () => {
     );
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(initFired).toBe(true);
   });
@@ -372,6 +397,7 @@ describe("plugin bootstrap", () => {
           {
             name: "gated-off-tool",
             description: "should not be registered",
+            category: "test",
             defaultRiskLevel: RiskLevel.Low,
             executionTarget: "sandbox",
             input_schema: { type: "object", properties: {}, required: [] },
@@ -407,7 +433,7 @@ describe("plugin bootstrap", () => {
     const { matchSkillRoute } =
       await import("../runtime/skill-route-registry.js");
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // init must not have fired.
     expect(initFired).toBe(false);
@@ -430,7 +456,7 @@ describe("plugin bootstrap", () => {
     });
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(initFired).toBe(true);
   });
@@ -455,7 +481,7 @@ describe("plugin bootstrap", () => {
     );
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(initFired).toBe(false);
   });
@@ -491,7 +517,7 @@ describe("plugin bootstrap", () => {
     );
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Neither the middleware slot nor the injector list should expose the
     // flag-gated plugin's contributions. The default plugins also contribute
@@ -519,7 +545,7 @@ describe("plugin bootstrap", () => {
     );
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     await runShutdownHooks("test-shutdown");
 
     // The shutdown hook is a single registered callback that walks a

package/src/__tests__/plugin-route-contribution.test.ts

@@ -43,11 +43,7 @@ mock.module("../security/secure-keys.js", () => ({
   getSecureKeyAsync: getSecureKeyAsyncMock,
 }));
 
-import type { AssistantConfig } from "../config/schema.js";
-import {
-  bootstrapPlugins,
-  type DaemonContext,
-} from "../daemon/external-plugins-bootstrap.js";
+import { bootstrapPlugins } from "../daemon/external-plugins-bootstrap.js";
 import { runShutdownHooks } from "../daemon/shutdown-registry.js";
 import {
   registerPlugin,
@@ -69,12 +65,6 @@ const TEST_WORKSPACE_DIR = join(
 );
 process.env.VELLUM_WORKSPACE_DIR = TEST_WORKSPACE_DIR;
 
-const fakeConfig = {} as unknown as AssistantConfig;
-const fakeCtx: DaemonContext = {
-  config: fakeConfig,
-  assistantVersion: "9.9.9-test",
-};
-
 /** Build a minimal valid plugin with optional route contributions. */
 /**
  * Test helper. Accepts the new `hooks` bag and ALSO legacy top-level
@@ -149,7 +139,7 @@ describe("plugin route contributions", () => {
       }),
     );
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // init() must have run — route registration is gated on init success.
     expect(initFired).toBe(true);
@@ -179,7 +169,7 @@ describe("plugin route contributions", () => {
 
     registerPlugin(buildPlugin("echo-plugin", { routes: [route] }));
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Sanity: route is live after bootstrap.
     expect(matchSkillRoute("/_plugin/echo", "GET")).not.toBeNull();
@@ -197,7 +187,7 @@ describe("plugin route contributions", () => {
     // route handling entirely (the guard is `if plugin.routes && length > 0`).
     registerPlugin(buildPlugin("no-routes-plugin", { async init() {} }));
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     await runShutdownHooks("test-shutdown");
 
     // Nothing to verify beyond "neither throws" — an empty `routes` must not
@@ -228,7 +218,7 @@ describe("plugin route contributions", () => {
       }),
     );
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Simulate an external wipe before the shutdown hook runs — e.g. a
     // different subsystem calling `resetSkillRoutesForTests` or a hot-reload
@@ -301,7 +291,7 @@ describe("plugin route contributions", () => {
       }),
     );
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Both plugins' routes landed in the registry; matching returns one of
     // them (order defined by registration, but we only care that *some*

package/src/__tests__/plugin-skill-contribution.test.ts

@@ -31,12 +31,8 @@ mock.module("../security/secure-keys.js", () => ({
   getSecureKeyAsync: getSecureKeyAsyncMock,
 }));
 
-import type { AssistantConfig } from "../config/schema.js";
 import { loadSkillBySelector, loadSkillCatalog } from "../config/skills.js";
-import {
-  bootstrapPlugins,
-  type DaemonContext,
-} from "../daemon/external-plugins-bootstrap.js";
+import { bootstrapPlugins } from "../daemon/external-plugins-bootstrap.js";
 import { runShutdownHooks } from "../daemon/shutdown-registry.js";
 import {
   getPluginContributedSkillDefinition,
@@ -64,12 +60,6 @@ const TEST_WORKSPACE_DIR = join(
 );
 process.env.VELLUM_WORKSPACE_DIR = TEST_WORKSPACE_DIR;
 
-const fakeConfig = {} as unknown as AssistantConfig;
-const fakeCtx: DaemonContext = {
-  config: fakeConfig,
-  assistantVersion: "9.9.9-test",
-};
-
 /** Build a plugin that contributes one or more skills. */
 function buildSkillPlugin(
   name: string,
@@ -106,7 +96,7 @@ describe("plugin skill contributions", () => {
     };
 
     registerPlugin(buildSkillPlugin("demo-plugin", [skill]));
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Ref count bumped to exactly 1 so we can tell register and unregister
     // are balanced downstream.
@@ -138,7 +128,7 @@ describe("plugin skill contributions", () => {
     };
 
     registerPlugin(buildSkillPlugin("catalog-plugin", [skill]));
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // loadSkillCatalog is the exact entry point `skill_load` consults via
     // `loadSkillBySelector` -> `resolveSkillSelector`.
@@ -171,7 +161,7 @@ describe("plugin skill contributions", () => {
     };
 
     registerPlugin(buildSkillPlugin("ephemeral-plugin", [skill]));
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     // Sanity: present before shutdown.
     expect(loadSkillCatalog().some((s) => s.id === "ephemeral-skill")).toBe(
@@ -206,7 +196,7 @@ describe("plugin skill contributions", () => {
       },
     });
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     expect(getPluginSkillRefCount("no-skills-plugin")).toBe(0);
     expect(getPluginContributedSkillSummaries()).toEqual([]);
@@ -231,7 +221,7 @@ describe("plugin skill contributions", () => {
 
     let caught: unknown;
     try {
-      await bootstrapPlugins(fakeCtx);
+      await bootstrapPlugins();
     } catch (err) {
       caught = err;
     }
@@ -347,7 +337,7 @@ describe("plugin skill contributions", () => {
     registerPlugin(buildSkillPlugin("shadow-plugin", [pluginSkill]));
 
     try {
-      await bootstrapPlugins(fakeCtx);
+      await bootstrapPlugins();
 
       const catalog = loadSkillCatalog();
       const entry = catalog.find((s) => s.id === "shared-id");

package/src/__tests__/plugin-tool-contribution.test.ts

@@ -39,11 +39,7 @@ mock.module("../security/secure-keys.js", () => ({
   getSecureKeyAsync: getSecureKeyAsyncMock,
 }));
 
-import type { AssistantConfig } from "../config/schema.js";
-import {
-  bootstrapPlugins,
-  type DaemonContext,
-} from "../daemon/external-plugins-bootstrap.js";
+import { bootstrapPlugins } from "../daemon/external-plugins-bootstrap.js";
 import { runShutdownHooks } from "../daemon/shutdown-registry.js";
 import { RiskLevel } from "../permissions/types.js";
 import {
@@ -61,11 +57,7 @@ import {
   registerPluginTools,
   unregisterPluginTools,
 } from "../tools/registry.js";
-import type {
-  LoadedTool,
-  ToolContext,
-  ToolExecutionResult,
-} from "../tools/types.js";
+import type { Tool, ToolContext, ToolExecutionResult } from "../tools/types.js";
 
 // Redirect plugin-storage-directory creation into a per-process temp tree so
 // the test doesn't touch the developer's real ~/.vellum. This matches the
@@ -76,22 +68,14 @@ const TEST_WORKSPACE_DIR = join(
 );
 process.env.VELLUM_WORKSPACE_DIR = TEST_WORKSPACE_DIR;
 
-const fakeConfig = {} as unknown as AssistantConfig;
-const fakeCtx: DaemonContext = {
-  config: fakeConfig,
-  assistantVersion: "9.9.9-test",
-};
-
-function makeFakeTool(
-  name: string,
-  extras: Partial<LoadedTool> = {},
-): LoadedTool {
+function makeFakeTool(name: string, extras: Partial<Tool> = {}): Tool {
   return {
     name,
     description: `Fake ${name}`,
     defaultRiskLevel: RiskLevel.Low,
     executionTarget: "sandbox",
     input_schema: { type: "object", properties: {}, required: [] },
+    category: "",
     async execute(
       _input: Record<string, unknown>,
       _context: ToolContext,
@@ -163,7 +147,7 @@ describe("plugin tool contributions", () => {
     });
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
 
     const retrieved = getTool("plugin-contrib-tool");
     expect(retrieved).toBeDefined();
@@ -192,7 +176,7 @@ describe("plugin tool contributions", () => {
     });
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     expect(getTool("bravo-tool")).toBeDefined();
 
     await runShutdownHooks("test-shutdown");
@@ -205,7 +189,7 @@ describe("plugin tool contributions", () => {
     const plugin = buildPlugin("no-tools", { async init() {} });
     registerPlugin(plugin);
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     // No tool should have been registered.
     expect(getAllTools()).toHaveLength(0);
 
@@ -229,7 +213,7 @@ describe("plugin tool contributions", () => {
 
     expect(getTool("charlie-tool")).toBeUndefined();
 
-    await bootstrapPlugins(fakeCtx);
+    await bootstrapPlugins();
     expect(getTool("charlie-tool")).toBeDefined();
   });
 
@@ -245,7 +229,7 @@ describe("plugin tool contributions", () => {
     });
     registerPlugin(plugin);
 
-    await expect(bootstrapPlugins(fakeCtx)).rejects.toThrow(/delta-broken/);
+    await expect(bootstrapPlugins()).rejects.toThrow(/delta-broken/);
     expect(getTool("delta-tool")).toBeUndefined();
   });
 });
@@ -336,7 +320,7 @@ describe("registerPluginTools / unregisterPluginTools helpers", () => {
       ...makeFakeTool("pt_spoof"),
       origin: "skill",
       owner: { kind: "skill", id: "some-other-skill" },
-    } as unknown as LoadedTool;
+    } as unknown as Tool;
     registerPluginTools("my-plugin", [spoofed]);
     expect(getTool("pt_spoof")).toBeDefined();
     expect(getToolOwner("pt_spoof")).toEqual({

package/src/__tests__/plugin-types.test.ts

@@ -12,6 +12,10 @@ import { describe, expect, test } from "bun:test";
 
 import type { TrustContext } from "../daemon/trust-context.js";
 import { RiskLevel } from "../permissions/types.js";
+import type {
+  ToolResultTruncateArgs,
+  ToolResultTruncateResult,
+} from "../plugins/defaults/tool-result-truncate/types.js";
 import {
   type CircuitBreakerArgs,
   type CircuitBreakerResult,
@@ -42,11 +46,9 @@ import {
   type ToolErrorDecision,
   type ToolExecuteArgs,
   type ToolExecuteResult,
-  type ToolResultTruncateArgs,
-  type ToolResultTruncateResult,
   type TurnContext,
 } from "../plugins/types.js";
-import type { LoadedTool } from "../tools/types.js";
+import type { Tool } from "../tools/types.js";
 
 const sampleTrust: TrustContext = {
   sourceChannel: "vellum",
@@ -75,18 +77,10 @@ describe("plugin core types", () => {
     // arg/result types have diverged from the early `{input: unknown}` /
     // `{output: unknown}` placeholders as individual pipeline wrap-up PRs
     // land.
-    function passthroughFor<A, R>(): Middleware<A, R> {
-      return async (args, next, _ctx) => next(args);
-    }
     const passthrough: Middleware<
       { input: unknown },
       { output: unknown }
     > = async (args, next, _ctx) => next(args);
-    const passthroughHistoryRepair = passthroughFor<
-      import("../plugins/types.js").HistoryRepairArgs,
-      import("../plugins/types.js").HistoryRepairResult
-    >();
-
     // `llmCall` has concrete arg/result types (upgraded in PR 15).
     const llmCallPassthrough: Middleware<LLMCallArgs, LLMCallResult> = async (
       args,
@@ -207,12 +201,13 @@ describe("plugin core types", () => {
       },
     };
 
-    const sampleTool: LoadedTool = {
+    const sampleTool: Tool = {
       name: "sample-tool",
       description: "Sample plugin tool",
       defaultRiskLevel: RiskLevel.Low,
       executionTarget: "sandbox",
       input_schema: { type: "object", properties: {}, required: [] },
+      category: "",
       async execute() {
         return { content: "ok", isError: false };
       },
@@ -255,7 +250,6 @@ describe("plugin core types", () => {
         llmCall: llmCallPassthrough,
         toolExecute: toolExecutePassthrough,
         memoryRetrieval: memoryPassthrough,
-        historyRepair: passthroughHistoryRepair,
         tokenEstimate: tokenEstimatePassthrough,
         compaction: compactionPassthrough,
         overflowReduce: overflowReducePassthrough,
@@ -271,7 +265,6 @@ describe("plugin core types", () => {
     // Minimal runtime check so the test body is non-empty.
     expect(plugin.manifest.name).toBe("sample-plugin");
     expect(plugin.middleware.turn).toBe(passthrough);
-    expect(plugin.middleware.historyRepair).toBe(passthroughHistoryRepair);
   });
 
   test("PluginTimeoutError carries pipeline, plugin, and elapsed fields", () => {

package/src/__tests__/prechat-onboarding-contract.test.ts

@@ -299,6 +299,29 @@ describe("pre-chat onboarding contract", () => {
       expect(bootstrap).not.toContain(expectedJson);
     });
 
+    test("renders prior assistants in first-run context", () => {
+      writeFileSync(
+        join(TEST_DIR, "BOOTSTRAP.md"),
+        "# Bootstrap\n\nOnboarding.",
+      );
+
+      const context: OnboardingContext = {
+        tools: [],
+        tasks: [],
+        tone: "warm",
+        userName: "Alex",
+        priorAssistants: ["chatgpt", "claude", "perplexity"],
+      };
+
+      const result = buildSystemPrompt({ onboardingContext: context });
+      const bootstrap = bootstrapBlock(result);
+
+      expect(bootstrap).toContain("## First-Run User Context");
+      expect(bootstrap).toContain(
+        "- Prior AI assistants used: ChatGPT, Claude, Perplexity",
+      );
+    });
+
     test("empty tools/tasks arrays result in no Daily tools / Common work lines", () => {
       writeFileSync(
         join(TEST_DIR, "BOOTSTRAP.md"),

package/src/__tests__/process-message-background-slack.test.ts

@@ -254,12 +254,11 @@ describe("processMessageInBackground Slack option propagation", () => {
     const result = await processMessageInBackground(
       "conv-background-slack",
       "Reply from Slack",
-      undefined,
       {
         slackInbound,
+        sourceChannel: "slack",
+        sourceInterface: "slack",
       },
-      "slack",
-      "slack",
     );
 
     expect(result).toEqual({ messageId: "persisted-user-message-id" });
@@ -280,21 +279,21 @@ describe("processMessageInBackground Slack option propagation", () => {
     const processing = processMessage(
       "conv-background-slack",
       "Reply from Slack",
-      undefined,
       {
         onEvent: (msg) => {
           observedMessages.push(msg);
         },
+        sourceChannel: "slack",
+        sourceInterface: "slack",
       },
-      "slack",
-      "slack",
     );
 
     await waitForRunAgentLoopCall();
 
-    const loopOnEvent = activeConversation.runAgentLoop.mock.calls[0][2] as
-      | ((msg: unknown) => void)
+    const loopOptions = activeConversation.runAgentLoop.mock.calls[0][2] as
+      | { onEvent?: (msg: unknown) => void }
       | undefined;
+    const loopOnEvent = loopOptions?.onEvent;
     const delta = {
       type: "assistant_text_delta",
       text: "Working on it.",
@@ -321,20 +320,22 @@ describe("processMessageInBackground Slack option propagation", () => {
     await processMessageInBackground(
       "conv-background-slack",
       "Regular background wake",
-      undefined,
-      undefined,
-      "vellum",
-      "web",
+      {
+        sourceChannel: "vellum",
+        sourceInterface: "web",
+      },
     );
 
     expect(activeConversation.persistUserMessage).toHaveBeenCalledTimes(1);
     expect(
       activeConversation.persistUserMessage.mock.calls[0][0].metadata,
     ).toBeUndefined();
-    expect(activeConversation.runAgentLoop.mock.calls[0][3]).toEqual({
-      isInteractive: false,
-      isUserMessage: true,
-    });
+    expect(activeConversation.runAgentLoop.mock.calls[0][2]).toEqual(
+      expect.objectContaining({
+        isInteractive: false,
+        isUserMessage: true,
+      }),
+    );
 
     activeConversation.__loopDeferred.resolve();
   });