@vellumai/assistant 0.8.9-staging.1 → 0.8.9-staging.3

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

@@ -102,23 +102,26 @@ mock.module("../context/token-estimator.js", () => ({
   estimateToolsTokens: () => 0,
 }));
 
-mock.module("../daemon/context-overflow-reducer.js", () => ({
-  createInitialReducerState: () => ({
-    appliedTiers: [],
-    injectionMode: "full" as const,
-    exhausted: false,
-  }),
-  reduceContextOverflow: async (msgs: Message[]) => ({
-    messages: msgs,
-    tier: "forced_compaction",
-    state: {
-      appliedTiers: ["forced_compaction"],
-      injectionMode: "full",
-      exhausted: true,
-    },
-    estimatedTokens: 1000,
+mock.module(
+  "../plugins/defaults/compaction/context-overflow-reducer.js",
+  () => ({
+    createInitialReducerState: () => ({
+      appliedTiers: [],
+      injectionMode: "full" as const,
+      exhausted: false,
+    }),
+    reduceContextOverflow: async (msgs: Message[]) => ({
+      messages: msgs,
+      tier: "forced_compaction",
+      state: {
+        appliedTiers: ["forced_compaction"],
+        injectionMode: "full",
+        exhausted: true,
+      },
+      estimatedTokens: 1000,
+    }),
   }),
-}));
+);
 
 mock.module("../daemon/context-overflow-policy.js", () => ({
   resolveOverflowAction: () => "fail_gracefully",

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

@@ -39,6 +39,20 @@ const conversationRuntimeAssemblyRealSnapshot = {
 
 // ── Module mocks (must precede imports of the module under test) ─────
 
+// The real AgentLoop resolves the per-conversation ContextWindowManager from
+// the compaction store keyed by conversationId. These overflow tests build
+// fake conversations whose manager is a canned stub, so register each stub in a
+// map the mocked store reads from.
+const fakeContextWindowManagers = new Map<string, unknown>();
+mock.module("../plugins/defaults/compaction/manager-store.js", () => ({
+  createContextWindowManager: () => undefined,
+  getContextWindowManager: (conversationId: string) =>
+    fakeContextWindowManagers.get(conversationId),
+  disposeContextWindowManager: (conversationId: string) => {
+    fakeContextWindowManagers.delete(conversationId);
+  },
+}));
+
 mock.module("../util/logger.js", () => ({
   getLogger: () =>
     new Proxy({} as Record<string, unknown>, { get: () => () => {} }),
@@ -137,35 +151,38 @@ mock.module("../context/token-estimator.js", () => ({
 let mockReducerStepFn:
   | ((msgs: Message[], cfg: unknown, state: unknown) => unknown)
   | null = null;
-mock.module("../daemon/context-overflow-reducer.js", () => ({
-  createInitialReducerState: () => ({
-    appliedTiers: [],
-    injectionMode: "full" as const,
-    exhausted: false,
+mock.module(
+  "../plugins/defaults/compaction/context-overflow-reducer.js",
+  () => ({
+    createInitialReducerState: () => ({
+      appliedTiers: [],
+      injectionMode: "full" as const,
+      exhausted: false,
+    }),
+    reduceContextOverflow: async (
+      msgs: Message[],
+      cfg: unknown,
+      state: unknown,
+    ) => {
+      if (mockReducerStepFn) return mockReducerStepFn(msgs, cfg, state);
+      return {
+        messages: msgs,
+        tier: "forced_compaction",
+        state: {
+          appliedTiers: [
+            "forced_compaction",
+            "tool_result_truncation",
+            "media_stubbing",
+            "injection_downgrade",
+          ],
+          injectionMode: "full",
+          exhausted: true,
+        },
+        estimatedTokens: 1000,
+      };
+    },
   }),
-  reduceContextOverflow: async (
-    msgs: Message[],
-    cfg: unknown,
-    state: unknown,
-  ) => {
-    if (mockReducerStepFn) return mockReducerStepFn(msgs, cfg, state);
-    return {
-      messages: msgs,
-      tier: "forced_compaction",
-      state: {
-        appliedTiers: [
-          "forced_compaction",
-          "tool_result_truncation",
-          "media_stubbing",
-          "injection_downgrade",
-        ],
-        injectionMode: "full",
-        exhausted: true,
-      },
-      estimatedTokens: 1000,
-    };
-  },
-}));
+);
 
 // Policy: default to fail_gracefully
 let mockOverflowAction: string = "fail_gracefully";
@@ -471,7 +488,7 @@ function makeCtx(
     toolExecutor,
   });
 
-  return {
+  const ctx = {
     conversationId: "test-conv",
     messages: [
       { role: "user", content: [{ type: "text", text: "Hello" }] },
@@ -582,6 +599,8 @@ function makeCtx(
 
     ...ctxOverrides,
   } as unknown as Conversation;
+  fakeContextWindowManagers.set(conversationId, ctx.contextWindowManager);
+  return ctx;
 }
 
 /**

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

@@ -28,6 +28,20 @@ let mockUiConfig: { userTimezone?: string; detectedTimezone?: string } = {};
 
 // ── Module mocks (must precede imports of the module under test) ─────
 
+// The real AgentLoop resolves the per-conversation ContextWindowManager from
+// the compaction store keyed by conversationId. These orchestrator tests build
+// fake conversations whose manager is a canned stub, so register each stub in a
+// map the mocked store reads from.
+const fakeContextWindowManagers = new Map<string, unknown>();
+mock.module("../plugins/defaults/compaction/manager-store.js", () => ({
+  createContextWindowManager: () => undefined,
+  getContextWindowManager: (conversationId: string) =>
+    fakeContextWindowManagers.get(conversationId),
+  disposeContextWindowManager: (conversationId: string) => {
+    fakeContextWindowManagers.delete(conversationId);
+  },
+}));
+
 mock.module("../util/logger.js", () => ({
   getLogger: () =>
     new Proxy({} as Record<string, unknown>, { get: () => () => {} }),
@@ -100,35 +114,38 @@ mock.module("../context/token-estimator.js", () => ({
 let mockReducerStepFn:
   | ((msgs: Message[], cfg: unknown, state: unknown) => unknown)
   | null = null;
-mock.module("../daemon/context-overflow-reducer.js", () => ({
-  createInitialReducerState: () => ({
-    appliedTiers: [],
-    injectionMode: "full" as const,
-    exhausted: false,
+mock.module(
+  "../plugins/defaults/compaction/context-overflow-reducer.js",
+  () => ({
+    createInitialReducerState: () => ({
+      appliedTiers: [],
+      injectionMode: "full" as const,
+      exhausted: false,
+    }),
+    reduceContextOverflow: async (
+      msgs: Message[],
+      cfg: unknown,
+      state: unknown,
+    ) => {
+      if (mockReducerStepFn) return mockReducerStepFn(msgs, cfg, state);
+      return {
+        messages: msgs,
+        tier: "forced_compaction",
+        state: {
+          appliedTiers: [
+            "forced_compaction",
+            "tool_result_truncation",
+            "media_stubbing",
+            "injection_downgrade",
+          ],
+          injectionMode: "full",
+          exhausted: true,
+        },
+        estimatedTokens: 1000,
+      };
+    },
   }),
-  reduceContextOverflow: async (
-    msgs: Message[],
-    cfg: unknown,
-    state: unknown,
-  ) => {
-    if (mockReducerStepFn) return mockReducerStepFn(msgs, cfg, state);
-    return {
-      messages: msgs,
-      tier: "forced_compaction",
-      state: {
-        appliedTiers: [
-          "forced_compaction",
-          "tool_result_truncation",
-          "media_stubbing",
-          "injection_downgrade",
-        ],
-        injectionMode: "full",
-        exhausted: true,
-      },
-      estimatedTokens: 1000,
-    };
-  },
-}));
+);
 
 // Policy: default to fail_gracefully
 let mockOverflowAction: string = "fail_gracefully";
@@ -520,17 +537,6 @@ mock.module("../memory/llm-request-log-store.js", () => ({
   setAgentLoopExitReasonOnLatestLog: setAgentLoopExitReasonOnLatestLogMock,
 }));
 
-let mockHasProactiveArtifactCompleted = true;
-let mockTryClaimProactiveArtifactTrigger = false;
-const runProactiveArtifactJobMock = mock(
-  async (_params: Record<string, unknown>) => {},
-);
-mock.module("../proactive-artifact/index.js", () => ({
-  hasProactiveArtifactCompleted: () => mockHasProactiveArtifactCompleted,
-  tryClaimProactiveArtifactTrigger: () => mockTryClaimProactiveArtifactTrigger,
-  runProactiveArtifactJob: runProactiveArtifactJobMock,
-}));
-
 // ── Imports (after mocks) ────────────────────────────────────────────
 
 import { AgentLoop } from "../agent/loop.js";
@@ -594,7 +600,7 @@ function makeCtx(
     toolExecutor,
   });
 
-  return {
+  const ctx = {
     conversationId: "test-conv",
     messages: [
       { role: "user", content: [{ type: "text", text: "Hello" }] },
@@ -710,6 +716,8 @@ function makeCtx(
 
     ...ctxOverrides,
   } as unknown as Conversation;
+  fakeContextWindowManagers.set(conversationId, ctx.contextWindowManager);
+  return ctx;
 }
 
 type CompactionResult = Parameters<typeof applyCompactionResult>[1];
@@ -765,9 +773,6 @@ beforeEach(() => {
     title: null,
   };
   mockMessageById = null;
-  mockHasProactiveArtifactCompleted = true;
-  mockTryClaimProactiveArtifactTrigger = false;
-  runProactiveArtifactJobMock.mockClear();
   setConversationHistoryStrippedAtMock.mockClear();
   setConversationHistoryStrippedAtMock.mockImplementation(() => {});
   applyRuntimeInjectionsMock.mockClear();
@@ -862,59 +867,6 @@ describe("session-agent-loop", () => {
     });
   });
 
-  describe("proactive artifact trigger", () => {
-    test("does not start proactive artifact jobs after foreground user turns", async () => {
-      mockConversationRow = {
-        ...mockConversationRow,
-        id: "test-conv",
-        conversationType: "standard",
-      };
-      mockMessageById = {
-        id: "user-msg-1",
-        conversationId: "test-conv",
-        createdAt: 1000,
-      };
-      mockHasProactiveArtifactCompleted = false;
-      mockTryClaimProactiveArtifactTrigger = true;
-
-      // A two-call agent turn: the model invokes `app_create`, then wraps up
-      // with a final text reply.
-      const ctx = makeCtx({
-        conversationId: "test-conv",
-        providerResponses: [
-          {
-            content: [
-              { type: "text", text: "I'll build that app." },
-              {
-                type: "tool_use",
-                id: "tool-1",
-                name: "app_create",
-                input: { name: "Flow" },
-              },
-            ],
-            model: "mock-model",
-            usage: { inputTokens: 10, outputTokens: 5 },
-            stopReason: "tool_use",
-          },
-          textResponse("Done."),
-        ],
-        toolExecutor: async () => ({ content: "{}", isError: false }),
-      });
-      await runAgentLoopImpl(
-        ctx,
-        "build a kanban app",
-        "user-msg-1",
-        () => {},
-        {
-          isUserMessage: true,
-        },
-      );
-      await new Promise((resolve) => setTimeout(resolve, 0));
-
-      expect(runProactiveArtifactJobMock).toHaveBeenCalledTimes(0);
-    });
-  });
-
   describe("disk pressure injection context", () => {
     // The loop sets `ctx.diskPressureCleanupModeActive` for the duration of the
     // turn (the disk-pressure-warning injector reads it via the per-conversation

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

@@ -164,7 +164,7 @@ mock.module("../memory/retriever.js", () => ({
   injectMemoryRecallAsUserBlock: (msgs: Message[]) => msgs,
 }));
 
-mock.module("../context/window-manager.js", () => ({
+mock.module("../plugins/defaults/compaction/window-manager.js", () => ({
   ContextWindowManager: class {
     constructor() {}
     updateConfig() {}

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

@@ -319,7 +319,6 @@ describe("classifyConversationError", () => {
       expect(result.errorCategory).toBe("image_dimensions_too_large");
       expect(result.retryable).toBe(false);
       expect(result.userMessage).toContain("image");
-      expect(result.userMessage).toContain("8000");
     });
 
     it("matches the singular 'image dimension exceeds' phrasing as well", () => {
@@ -333,6 +332,21 @@ describe("classifyConversationError", () => {
       expect(result.retryable).toBe(false);
     });
 
+    it("classifies the base64 payload 'exceeds 5 MB maximum' variant as IMAGE_TOO_LARGE (JARVIS-1041)", () => {
+      // This is the exact shape that wedged the reporting user: an oversized
+      // image nested in a tool_result. It must route to IMAGE_TOO_LARGE so the
+      // recovery path fires, not to a retryable PROVIDER_API loop.
+      const err = new ProviderError(
+        'Anthropic API error (400): 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.28.content.1.tool_result.content.1.image.source.base64: image exceeds 5 MB maximum: 7465044 bytes > 5242880 bytes"},"request_id":"req_011CbbFqvxBmDQdhVlqxtAfW"}',
+        "anthropic",
+        400,
+      );
+      const result = classifyConversationError(err, baseCtx);
+      expect(result.code).toBe("IMAGE_TOO_LARGE");
+      expect(result.errorCategory).toBe("image_dimensions_too_large");
+      expect(result.retryable).toBe(false);
+    });
+
     it("does not steal generic 400s that happen to mention 'image'", () => {
       const err = new ProviderError(
         "invalid request: image source is missing",

package/src/__tests__/conversation-history-web-search.test.ts

@@ -807,7 +807,12 @@ describe("web_search_tool_result structural guard", () => {
     // Media token counting iterates tool_result.contentBlocks for nested
     // image/file blocks. web_search_tool_result has opaque content with no
     // contentBlocks property, so it cannot contain nested media.
-    "daemon/context-overflow-reducer.ts",
+    "plugins/defaults/compaction/context-overflow-reducer.ts",
+
+    // Downgrades permanently-unsendable images nested in tool_result.contentBlocks
+    // (e.g. an oversized browser screenshot). web_search_tool_result has opaque
+    // content with no contentBlocks property, so it cannot hold such an image.
+    "daemon/persist-unsendable-image.ts",
 
     // Final orphan-pair safety pass in the Slack transcript renderer.
     // Server-side block types (`server_tool_use`, `web_search_tool_result`)

package/src/__tests__/conversation-media-retry.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect, test } from "bun:test";
 
-import { createContextSummaryMessage } from "../context/window-manager.js";
 import { stripMediaPayloadsForRetry } from "../daemon/conversation-media-retry.js";
+import { createContextSummaryMessage } from "../plugins/defaults/compaction/window-manager.js";
 import type { ContentBlock, Message } from "../providers/types.js";
 
 // ---------------------------------------------------------------------------

package/src/__tests__/conversation-process-app-control-preactivation.test.ts

@@ -185,6 +185,7 @@ function makeQueuedMessage(opts: {
   requestId: string;
   content?: string;
   turnInterfaceContext?: TurnInterfaceContext;
+  sourceActorPrincipalId?: string;
 }): QueuedMessage {
   return {
     content: opts.content ?? "follow up",
@@ -194,6 +195,10 @@ function makeQueuedMessage(opts: {
     metadata: {},
     sentAt: Date.now(),
     turnInterfaceContext: opts.turnInterfaceContext,
+    sourceActorPrincipalId: opts.sourceActorPrincipalId,
+    authContext: opts.sourceActorPrincipalId
+      ? ({ actorPrincipalId: opts.sourceActorPrincipalId } as never)
+      : undefined,
   };
 }
 
@@ -332,6 +337,7 @@ describe("drainQueue preactivation re-add for host-proxy interfaces", () => {
       makeQueuedMessage({
         requestId: "req-web-1",
         turnInterfaceContext: ifCtx,
+        sourceActorPrincipalId: "user-1",
       }),
     );
     const ctx = makeFakeContext({ queue, turnInterfaceContext: ifCtx });
@@ -346,6 +352,39 @@ describe("drainQueue preactivation re-add for host-proxy interfaces", () => {
     expect(ctx.preactivatedSkillIdCalls).toContain("computer-use");
   });
 
+  test("drainSingleMessage filters cross-client preactivation by the queued requester, not guardian owner", async () => {
+    mockCapabilityClients = [
+      { clientId: "owner-macos-client", actorPrincipalId: "owner-user" },
+    ];
+    const queue = new MessageQueue();
+    const ifCtx: TurnInterfaceContext = {
+      userMessageInterface: "web",
+      assistantMessageInterface: "web",
+    };
+    queue.push(
+      makeQueuedMessage({
+        requestId: "req-web-attacker",
+        turnInterfaceContext: ifCtx,
+        sourceActorPrincipalId: "trusted-contact-user",
+      }),
+    );
+    const ctx = makeFakeContext({ queue, turnInterfaceContext: ifCtx });
+    ctx.trustContext = {
+      trustClass: "trusted_contact",
+      guardianPrincipalId: "owner-user",
+      sourceChannel: "vellum",
+    };
+
+    await drainQueue(ctx);
+
+    expect(ctx.preactivatedSkillIdCalls).not.toContain("computer-use");
+    expect(ctx.preactivatedSkillIdCalls).not.toContain("app-control");
+    expect(ctx.currentTurnAuthContext?.actorPrincipalId).toBe(
+      "trusted-contact-user",
+    );
+    expect(ctx.currentTurnSourceActorPrincipalId).toBe("trusted-contact-user");
+  });
+
   test("drainSingleMessage does NOT re-add 'app-control' for web-sourced message when no capable client is connected", async () => {
     // mockCapabilityClients remains [] (reset by afterEach from prior test)
     const queue = new MessageQueue();
@@ -380,6 +419,7 @@ describe("drainQueue preactivation re-add for host-proxy interfaces", () => {
       makeQueuedMessage({
         requestId: "req-web-3",
         turnInterfaceContext: ifCtx,
+        sourceActorPrincipalId: "user-1",
       }),
     );
     const ctx = makeFakeContext({ queue, turnInterfaceContext: ifCtx });

package/src/__tests__/conversation-process-callsite.test.ts

@@ -249,7 +249,7 @@ mock.module("../agent/loop.js", () => ({
   },
 }));
 
-mock.module("../context/window-manager.js", () => ({
+mock.module("../plugins/defaults/compaction/window-manager.js", () => ({
   ContextWindowManager: class {
     constructor() {}
     updateConfig() {}

package/src/__tests__/conversation-provider-retry-repair.test.ts

@@ -100,56 +100,59 @@ mock.module("../context/token-estimator.js", () => ({
 // but these tests exercise the Conversation-level flow, not the reducer internals.
 // The reducer mock delegates to the compactFn to simulate a forced compaction
 // tier, matching the real reducer's behavior for Tier 1.
-mock.module("../daemon/context-overflow-reducer.js", () => ({
-  createInitialReducerState: () => ({
-    appliedTiers: [],
-    injectionMode: "full" as const,
-    exhausted: false,
-  }),
-  reduceContextOverflow: async (
-    msgs: Message[],
-    _cfg: unknown,
-    _state: unknown,
-    compactFn?: (
-      m: Message[],
-      s: AbortSignal | undefined,
-      o: Record<string, unknown>,
-    ) => Promise<{
-      compacted: boolean;
-      messages: Message[];
-      compactedPersistedMessages?: number;
-      summaryText?: string;
-      [k: string]: unknown;
-    }>,
-    signal?: AbortSignal,
-  ) => {
-    let resultMessages = msgs;
-    let compactionResult;
-    if (compactFn) {
-      const cr = await compactFn(msgs, signal, { force: true });
-      if (cr.compacted) {
-        resultMessages = cr.messages;
-        compactionResult = cr;
+mock.module(
+  "../plugins/defaults/compaction/context-overflow-reducer.js",
+  () => ({
+    createInitialReducerState: () => ({
+      appliedTiers: [],
+      injectionMode: "full" as const,
+      exhausted: false,
+    }),
+    reduceContextOverflow: async (
+      msgs: Message[],
+      _cfg: unknown,
+      _state: unknown,
+      compactFn?: (
+        m: Message[],
+        s: AbortSignal | undefined,
+        o: Record<string, unknown>,
+      ) => Promise<{
+        compacted: boolean;
+        messages: Message[];
+        compactedPersistedMessages?: number;
+        summaryText?: string;
+        [k: string]: unknown;
+      }>,
+      signal?: AbortSignal,
+    ) => {
+      let resultMessages = msgs;
+      let compactionResult;
+      if (compactFn) {
+        const cr = await compactFn(msgs, signal, { force: true });
+        if (cr.compacted) {
+          resultMessages = cr.messages;
+          compactionResult = cr;
+        }
       }
-    }
-    return {
-      messages: resultMessages,
-      tier: "forced_compaction",
-      state: {
-        appliedTiers: [
-          "forced_compaction",
-          "tool_result_truncation",
-          "media_stubbing",
-          "injection_downgrade",
-        ],
-        injectionMode: "full",
-        exhausted: true,
-      },
-      estimatedTokens: 1000,
-      compactionResult,
-    };
-  },
-}));
+      return {
+        messages: resultMessages,
+        tier: "forced_compaction",
+        state: {
+          appliedTiers: [
+            "forced_compaction",
+            "tool_result_truncation",
+            "media_stubbing",
+            "injection_downgrade",
+          ],
+          injectionMode: "full",
+          exhausted: true,
+        },
+        estimatedTokens: 1000,
+        compactionResult,
+      };
+    },
+  }),
+);
 
 mock.module("../daemon/context-overflow-policy.js", () => ({
   resolveOverflowAction: () => "fail_gracefully",
@@ -218,7 +221,7 @@ mock.module("../memory/retriever.js", () => ({
 let maybeCompactCalls: Array<{ force: boolean }> = [];
 let forceCompactionEnabled = false;
 
-mock.module("../context/window-manager.js", () => ({
+mock.module("../plugins/defaults/compaction/window-manager.js", () => ({
   ContextWindowManager: class {
     constructor() {}
     updateConfig() {}

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

@@ -262,7 +262,7 @@ mock.module("../memory/retriever.js", () => ({
   injectMemoryRecallAsUserBlock: (msgs: Message[]) => msgs,
 }));
 
-mock.module("../context/window-manager.js", () => ({
+mock.module("../plugins/defaults/compaction/window-manager.js", () => ({
   ContextWindowManager: class {
     constructor() {}
     updateConfig() {}