@vellumai/assistant 0.10.12 → 0.11.0-staging.2

package/src/__tests__/conversation-agent-loop.test.ts

@@ -12,11 +12,11 @@ import {
 } from "bun:test";
 
 import type { LoopToolExecutor } from "../agent/loop.js";
+import type { AssistantEvent } from "../api/index.js";
 import {
   queueConversationNotice,
   resetConversationNoticesForTests,
 } from "../daemon/conversation-notices.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
 import { getConversationDirName } from "../persistence/conversation-directories.js";
 import type { UserPromptSubmitContext } from "../plugin-api/types.js";
 import { resetPluginRegistryAndRegisterDefaults } from "../plugins/defaults/index.js";
@@ -804,7 +804,16 @@ function makeCtx(
     getQueueDepth: () => 0,
     hasQueuedMessages: () => false,
     canHandoffAtCheckpoint: () => false,
-    drainQueue: () => {},
+    drainQueue: (_reason?: string) => {},
+    // Forwards to drainQueue so tests that spy the drain observe the agent
+    // loop's post-turn kick through the guarded entry point.
+    kickDrainQueue(
+      this: { drainQueue: (reason?: string) => unknown },
+      reason: string = "loop_complete",
+      _origin?: string,
+    ) {
+      return this.drainQueue(reason);
+    },
     getTurnInterfaceContext: () => null,
     getTurnChannelContext: () => ({
       userMessageChannel: "vellum" as const,
@@ -1017,7 +1026,7 @@ describe("session-agent-loop", () => {
         },
       });
 
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const ctx = makeCtx({ providerResponses: [textResponse("ok")] });
       const runSpy = spyOn(ctx.agentLoop, "run");
 
@@ -1041,7 +1050,7 @@ describe("session-agent-loop", () => {
 
   describe("conversation notices", () => {
     test("emits queued billing notices after a successful turn", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const ctx = makeCtx({ providerResponses: [textResponse("ok")] });
       queueConversationNotice(ctx.conversationId, "memory-v3-test", {
         source: "memory_v3",
@@ -1078,7 +1087,7 @@ describe("session-agent-loop", () => {
       resolveAssistantAttachmentsMock.mockImplementation(async () => {
         throw new Error("attachment resolution failed");
       });
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const ctx = makeCtx({ providerResponses: [textResponse("ok")] });
       queueConversationNotice(ctx.conversationId, "memory-v3-test", {
         source: "memory_v3",
@@ -1289,7 +1298,7 @@ describe("session-agent-loop", () => {
         action: "block",
         reason: "trusted-contact",
       };
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const activityStates: unknown[][] = [];
       const ctx = makeCtx({
         emitActivityState: (...args: unknown[]) => {
@@ -1361,7 +1370,7 @@ describe("session-agent-loop", () => {
 
   describe("tool execution errors via agent loop", () => {
     test("error events from agent loop are classified and emitted", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       // The model calls a tool whose executor throws, surfacing an `error`
       // event from the loop's catch handler.
@@ -1380,7 +1389,7 @@ describe("session-agent-loop", () => {
     });
 
     test("non-error agent loop completion does not emit conversation_error", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx({
         providerResponses: [textResponse("All good")],
@@ -1398,7 +1407,7 @@ describe("session-agent-loop", () => {
 
   describe("LLM request log persistence", () => {
     test("record request log captures the actual provider name", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const rawRequest = {
         model: "gpt-4.1",
         messages: [{ role: "user", content: "Hello" }],
@@ -1508,7 +1517,7 @@ describe("session-agent-loop", () => {
     });
 
     test("record request log handles Responses API shaped payloads", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const rawRequest = {
         model: "gpt-5.4",
         instructions: "Be helpful.",
@@ -1571,7 +1580,7 @@ describe("session-agent-loop", () => {
 
   describe("usage accounting", () => {
     test("records the actual provider for usage accounting", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx({
         providerResponses: [
@@ -1627,7 +1636,7 @@ describe("session-agent-loop", () => {
     });
 
     test("persists the served model onto the assistant row's metadata at finalize", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx({
         providerResponses: [
@@ -1654,7 +1663,7 @@ describe("session-agent-loop", () => {
 
   describe("checkpoint handoff (infinite loop prevention)", () => {
     test("yields at checkpoint when canHandoffAtCheckpoint returns true", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       // A tool turn drives the loop to its first mid-loop checkpoint, where the
       // orchestrator yields for a queued handoff.
@@ -1682,7 +1691,7 @@ describe("session-agent-loop", () => {
     });
 
     test("continues when canHandoffAtCheckpoint returns false", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       // The tool turn reaches a checkpoint, but with handoff disabled the loop
       // continues to the next turn and completes normally.
@@ -1717,7 +1726,7 @@ describe("session-agent-loop", () => {
 
   describe("user cancellation", () => {
     test("emits generation_cancelled when abort signal fires", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const abortController = new AbortController();
 
       // The provider completes its response but the user cancels mid-turn, so
@@ -1739,7 +1748,7 @@ describe("session-agent-loop", () => {
     });
 
     test("handles AbortError thrown from agent loop as user cancellation", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const abortController = new AbortController();
 
       // The provider rejects with an AbortError after the user cancels.
@@ -1764,7 +1773,7 @@ describe("session-agent-loop", () => {
     });
 
     test("skips resolveAssistantAttachments when cancelled", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const abortController = new AbortController();
       resolveAssistantAttachmentsMock.mockClear();
 
@@ -1821,7 +1830,7 @@ describe("session-agent-loop", () => {
 
     test("clears state and surfaces a processing error when the provider call fails", async () => {
       // GIVEN a real loop whose provider rejects with an unexpected error
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const ctx = makeCtx({
         loopProvider: {
           name: "mock-provider",
@@ -1868,7 +1877,7 @@ describe("session-agent-loop", () => {
       // GIVEN a provider whose call wedges: it acknowledges the user cancel
       // (aborts the signal) but its promise never settles and never observes
       // the signal — the exact condition that latched `processing` true.
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const abortController = new AbortController();
       let drainReason: string | undefined;
       // The provider's call wedges on this promise. It settles only on test
@@ -1925,7 +1934,7 @@ describe("session-agent-loop", () => {
 
   describe("stale pending surface cleanup", () => {
     test("auto-completes non-dynamic_page pending surfaces on regular user message", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx();
       // Pre-populate a stale pending table surface
@@ -1956,7 +1965,7 @@ describe("session-agent-loop", () => {
     });
 
     test("does not auto-complete surfaces when request is a surface action", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx();
       ctx.pendingSurfaceActions.set("active-table-1", { surfaceType: "table" });
@@ -1982,7 +1991,7 @@ describe("session-agent-loop", () => {
     });
 
     test("no-op when no pending surfaces exist", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx();
       // No pending surfaces
@@ -1998,7 +2007,7 @@ describe("session-agent-loop", () => {
     });
 
     test("does not auto-complete surfaces for internal/subagent turns (no isUserMessage)", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx();
       ctx.pendingSurfaceActions.set("active-table-1", { surfaceType: "table" });
@@ -2024,7 +2033,7 @@ describe("session-agent-loop", () => {
       const ctx = makeCtx();
       ctx.pendingSurfaceActions.set("stale-table-1", { surfaceType: "table" });
 
-      const throwingOnEvent = (msg: ServerMessage) => {
+      const throwingOnEvent = (msg: AssistantEvent) => {
         _eventCount++;
         if (msg.type === "ui_surface_complete") {
           throw new Error("onEvent sink failed");
@@ -2130,7 +2139,7 @@ describe("session-agent-loop", () => {
         throw new Error("simulated DB failure");
       });
 
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const ctx = makeCtx();
 
       // Should not throw; agent loop continues and emits message_complete.
@@ -2145,7 +2154,7 @@ describe("session-agent-loop", () => {
 
   describe("error-only response with no assistant text", () => {
     test("synthesizes error assistant message when provider returns no response", async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       // GIVEN a real loop whose provider rejects with a generic error
       // (non-ordering, non-context-too-large) so the loop emits `error` and
@@ -2179,7 +2188,7 @@ describe("session-agent-loop", () => {
       // its `llm_request_logs` row orphaned. Without the backfill call in
       // the synthetic-message branch, a later turn's `handleMessageComplete`
       // sweep would wrong-attach this row to the wrong assistant message.
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       // GIVEN a real loop whose provider rejects: the loop emits
       // `provider_error` (writing an `llm_request_logs` row with
@@ -2225,7 +2234,7 @@ describe("session-agent-loop", () => {
         retryable: false,
         errorCategory: "managed_key_invalid",
       };
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       const ctx = makeCtx({
         loopProvider: {
@@ -2351,7 +2360,7 @@ describe("session-agent-loop", () => {
         metadata: null,
       };
 
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let messageCompleteSeenWhenIndexed: boolean | undefined;
       indexMessageNowMock.mockImplementationOnce(async () => {
         messageCompleteSeenWhenIndexed = events.some(
@@ -3370,7 +3379,7 @@ describe("session-agent-loop", () => {
 
     test("applyCompactionResult records Slack timestamp watermark when provided", async () => {
       const ctx = makeCtx();
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
 
       await applyCompactionResult(
         ctx,

package/src/__tests__/conversation-confirmation-signals.test.ts

@@ -13,7 +13,7 @@ import { describe, expect, mock, test } from "bun:test";
 
 import { CompactionCircuit } from "../agent/compaction-circuit.js";
 import type { AgentEvent } from "../agent/loop.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
+import type { AssistantEvent } from "../api/index.js";
 import type { Message, ProviderResponse } from "../providers/types.js";
 
 // ---------------------------------------------------------------------------
@@ -176,7 +176,7 @@ function makeProvider() {
 }
 
 function makeConversation(
-  sendToClient?: (msg: ServerMessage) => void,
+  sendToClient?: (msg: AssistantEvent) => void,
 ): Conversation {
   return new Conversation(
     "conv-signals-test",
@@ -212,7 +212,7 @@ function seedPendingConfirmation(
 
 describe("centralized confirmation emissions", () => {
   test("handleConfirmationResponse emits confirmation_state_changed with approved state for allow decision", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     seedPendingConfirmation(conversation, "req-allow-1");
@@ -240,7 +240,7 @@ describe("centralized confirmation emissions", () => {
   });
 
   test("handleConfirmationResponse emits confirmation_state_changed with denied state for deny decision", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     seedPendingConfirmation(conversation, "req-deny-1");
@@ -264,7 +264,7 @@ describe("centralized confirmation emissions", () => {
   });
 
   test("handleConfirmationResponse emits assistant_activity_state with thinking phase", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     seedPendingConfirmation(conversation, "req-activity-1");
@@ -287,7 +287,7 @@ describe("centralized confirmation emissions", () => {
   });
 
   test("handleConfirmationResponse passes emissionContext source", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     seedPendingConfirmation(conversation, "req-ctx-1");
@@ -314,7 +314,7 @@ describe("centralized confirmation emissions", () => {
 
 describe("activity version ordering", () => {
   test("emitActivityState produces monotonically increasing activityVersion", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     conversation.emitActivityState("thinking", "message_dequeued");
@@ -326,7 +326,7 @@ describe("activity version ordering", () => {
 
     const activityMsgs = emitted.filter(
       (m) => m.type === "assistant_activity_state",
-    ) as Array<ServerMessage & { activityVersion: number }>;
+    ) as Array<AssistantEvent & { activityVersion: number }>;
 
     expect(activityMsgs).toHaveLength(4);
 
@@ -342,7 +342,7 @@ describe("activity version ordering", () => {
   });
 
   test("handleConfirmationResponse increments activityVersion for its activity emission", () => {
-    const emitted: ServerMessage[] = [];
+    const emitted: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => emitted.push(msg));
 
     // Emit a baseline activity state
@@ -350,7 +350,7 @@ describe("activity version ordering", () => {
 
     const baselineMsg = emitted.find(
       (m) => m.type === "assistant_activity_state",
-    ) as ServerMessage & { activityVersion: number };
+    ) as AssistantEvent & { activityVersion: number };
     const baselineVersion = baselineMsg.activityVersion;
 
     // Now handle a confirmation
@@ -359,7 +359,7 @@ describe("activity version ordering", () => {
 
     const activityMsgs = emitted.filter(
       (m) => m.type === "assistant_activity_state",
-    ) as Array<ServerMessage & { activityVersion: number; reason: string }>;
+    ) as Array<AssistantEvent & { activityVersion: number; reason: string }>;
 
     // The confirmation_resolved activity message should have a higher version
     const resolvedMsg = activityMsgs.find(
@@ -372,7 +372,7 @@ describe("activity version ordering", () => {
 
 describe("sendToClient receives state signals", () => {
   test("emitActivityState delivers to sendToClient", () => {
-    const clientMsgs: ServerMessage[] = [];
+    const clientMsgs: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => clientMsgs.push(msg));
 
     conversation.emitActivityState("thinking", "message_dequeued");
@@ -383,7 +383,7 @@ describe("sendToClient receives state signals", () => {
   });
 
   test("emitConfirmationStateChanged delivers to sendToClient", () => {
-    const clientMsgs: ServerMessage[] = [];
+    const clientMsgs: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => clientMsgs.push(msg));
 
     conversation.emitConfirmationStateChanged({
@@ -399,7 +399,7 @@ describe("sendToClient receives state signals", () => {
   });
 
   test("handleConfirmationResponse delivers state signals to sendToClient", () => {
-    const clientMsgs: ServerMessage[] = [];
+    const clientMsgs: AssistantEvent[] = [];
     const conversation = makeConversation((msg) => clientMsgs.push(msg));
 
     seedPendingConfirmation(conversation, "req-signal-confirm");

package/src/__tests__/conversation-evictor.test.ts

@@ -19,12 +19,16 @@ import {
 
 function createMockSession(
   processing = false,
+  queued = false,
 ): EvictableConversation & { disposed: boolean } {
   return {
     disposed: false,
     isProcessing() {
       return processing;
     },
+    hasQueuedMessages() {
+      return queued;
+    },
     dispose() {
       this.disposed = true;
     },
@@ -98,6 +102,21 @@ describe("ConversationEvictor", () => {
       expect(sessions.has("a")).toBe(true);
       expect(s1.disposed).toBe(false);
     });
+
+    test("skips sessions with queued messages even when not processing", () => {
+      // Models the gap between a turn's `finally` and the async queue-drain
+      // dispatch: isProcessing() is false but a queued turn (e.g. a subagent
+      // completion wake) is still pending.
+      const s1 = createMockSession(false, true);
+      sessions.set("a", s1);
+
+      const result = evictor.sweep();
+
+      expect(result.ttlEvicted).toBe(0);
+      expect(result.skipped).toBe(1);
+      expect(sessions.has("a")).toBe(true);
+      expect(s1.disposed).toBe(false);
+    });
   });
 
   describe("LRU eviction", () => {
@@ -132,6 +151,36 @@ describe("ConversationEvictor", () => {
       expect(sessions.has("s4")).toBe(true);
     });
 
+    test("skips sessions with queued messages during LRU eviction", () => {
+      // maxConversations = 3, add 4 sessions; the LRU one has a queued turn.
+      const s0 = createMockSession(false, true); // queued — should not be evicted
+      const s1 = createMockSession();
+      const s2 = createMockSession();
+      const s3 = createMockSession();
+      sessions.set("s0", s0);
+      sessions.set("s1", s1);
+      sessions.set("s2", s2);
+      sessions.set("s3", s3);
+
+      const now = Date.now();
+      const lastAccess = (
+        evictor as unknown as { lastAccess: Map<string, number> }
+      ).lastAccess;
+      lastAccess.set("s0", now - 50); // old but has queued messages
+      lastAccess.set("s1", now - 40);
+      lastAccess.set("s2", now);
+      lastAccess.set("s3", now);
+
+      const result = evictor.sweep();
+
+      // s1 is the LRU evictable session, gets evicted
+      expect(result.lruEvicted).toBe(1);
+      expect(sessions.has("s0")).toBe(true); // kept: queued messages
+      expect(sessions.has("s1")).toBe(false); // evicted: LRU
+      expect(s0.disposed).toBe(false);
+      expect(s1.disposed).toBe(true);
+    });
+
     test("skips processing sessions during LRU eviction", () => {
       // maxConversations = 3, add 4 sessions, one processing
       const s0 = createMockSession(true); // processing — should not be evicted

package/src/__tests__/conversation-fork-crud.test.ts

@@ -483,7 +483,7 @@ describe("forkConversation", () => {
     // message — same-millisecond inserts would otherwise make the event
     // "at-or-before" the branch point and inherit.
     const db = getDb();
-    const base = Date.now();
+    const base = Date.now() - 10_000;
     db.run(
       `UPDATE messages SET created_at = ${base} WHERE id = '${compactedBranchPoint.id}'`,
     );