@vellumai/assistant 0.8.8 → 0.8.9-dev.202606091926.ebb2d62

package/src/__tests__/compaction-circuit.test.ts

@@ -0,0 +1,258 @@
+/**
+ * Tests for the `CompactionCircuit` class.
+ *
+ * The breaker logic lives as direct methods on the per-conversation
+ * `CompactionCircuit` state container (`agent/compaction-circuit.ts`). These
+ * tests assert the threshold (3 consecutive failures) and cooldown (1 hour)
+ * exactly match the user-visible behavior:
+ *   (a) counter increments on each failure outcome
+ *   (b) circuit opens after exactly 3 consecutive failures
+ *   (c) successful compaction resets counter and clears the circuit
+ *   (d) isOpen() reflects state and cooldown expiry
+ *   (e) circuit re-opens after cooldown expiry when 3 more failures
+ *       accumulate (guards the stale-timestamp regression)
+ *   (f) isOpen() is query-only and never mutates the counter
+ *   (g) open→closed transition emits `compaction_circuit_closed` exactly once
+ *   (h) closed→closed transition emits nothing
+ */
+
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+
+import {
+  COMPACTION_CIRCUIT_COOLDOWN_MS,
+  COMPACTION_CIRCUIT_FAILURE_THRESHOLD,
+  CompactionCircuit,
+} from "../agent/compaction-circuit.js";
+import type { CompactionCircuitEvent } from "../plugins/types.js";
+
+// ─── Fixtures ───────────────────────────────────────────────────────────────
+
+const CONVERSATION_ID = "conv-breaker-test";
+
+function collectEvents(): {
+  events: CompactionCircuitEvent[];
+  onEvent: (msg: CompactionCircuitEvent) => void;
+} {
+  const events: CompactionCircuitEvent[] = [];
+  return { events, onEvent: (msg) => events.push(msg) };
+}
+
+describe("CompactionCircuit", () => {
+  let originalDateNow: () => number;
+
+  beforeEach(() => {
+    originalDateNow = Date.now;
+  });
+
+  afterEach(() => {
+    Date.now = originalDateNow;
+  });
+
+  test("threshold and cooldown constants match the user-visible contract", () => {
+    expect(COMPACTION_CIRCUIT_FAILURE_THRESHOLD).toBe(3);
+    expect(COMPACTION_CIRCUIT_COOLDOWN_MS).toBe(60 * 60 * 1000);
+  });
+
+  test("(a) counter increments on each failure outcome", async () => {
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent, events } = collectEvents();
+
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(1);
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    expect(events).toHaveLength(0);
+
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(2);
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    expect(events).toHaveLength(0);
+  });
+
+  test("(b) circuit opens after exactly 3 consecutive failures", async () => {
+    const fixedNow = 1_700_000_000_000;
+    Date.now = () => fixedNow;
+
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent, events } = collectEvents();
+
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    // Two failures — circuit still closed.
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    expect(events).toHaveLength(0);
+
+    await circuit.recordOutcome(true, onEvent);
+    // Third failure — circuit trips and fires the event exactly once.
+    expect(circuit.consecutiveCompactionFailures).toBe(3);
+    expect(circuit.compactionCircuitOpenUntil).toBe(fixedNow + 60 * 60 * 1000);
+    expect(await circuit.isOpen()).toBe(true);
+    expect(events).toHaveLength(1);
+    expect(events[0]).toEqual({
+      type: "compaction_circuit_open",
+      conversationId: CONVERSATION_ID,
+      reason: "3_consecutive_failures",
+      openUntil: fixedNow + 60 * 60 * 1000,
+    });
+
+    // Further failures do not re-fire the event while the circuit is open.
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(4);
+    expect(events).toHaveLength(1);
+  });
+
+  test("(c) successful outcome resets counter and clears circuit", async () => {
+    const fixedNow = 1_700_000_000_000;
+    Date.now = () => fixedNow;
+
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent } = collectEvents();
+
+    // Trip the breaker.
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.compactionCircuitOpenUntil).not.toBeNull();
+
+    // Success resets state.
+    await circuit.recordOutcome(false, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(0);
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+  });
+
+  test("(d) isOpen() reflects state and expiry", async () => {
+    const fixedNow = 1_700_000_000_000;
+    Date.now = () => fixedNow;
+
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent } = collectEvents();
+
+    // Fresh state: closed.
+    expect(await circuit.isOpen()).toBe(false);
+
+    // Trip the breaker.
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+
+    // While open.
+    expect(await circuit.isOpen()).toBe(true);
+
+    // After cooldown expires the breaker reports closed again, even without
+    // an explicit reset — the open-until timestamp is the only source of
+    // truth for the gate.
+    Date.now = () => fixedNow + 60 * 60 * 1000 + 1;
+    expect(await circuit.isOpen()).toBe(false);
+  });
+
+  test("(e) circuit re-opens after cooldown expiry when 3 more failures accumulate", async () => {
+    // Regression: opening the breaker a second time must not require
+    // `compactionCircuitOpenUntil === null`. Once a cooldown expires, the
+    // gate correctly reports "closed" but a stale past-timestamp stays on the
+    // state, so the next 3-strike window must still trip a new cooldown. Any
+    // expired timestamp is treated the same as null.
+    const t0 = 1_700_000_000_000;
+    Date.now = () => t0;
+
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent, events } = collectEvents();
+
+    // Trip the breaker the first time.
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.compactionCircuitOpenUntil).toBe(t0 + 60 * 60 * 1000);
+    expect(events).toHaveLength(1);
+
+    // Advance past the cooldown window. Reset the counter — in production this
+    // happens when a subsequent `maybeCompact` call succeeds after the
+    // cooldown elapses, but the bug manifests even when the counter is reset:
+    // the stale `compactionCircuitOpenUntil` is what breaks re-opening.
+    const t1 = t0 + 60 * 60 * 1000 + 1;
+    Date.now = () => t1;
+    expect(await circuit.isOpen()).toBe(false);
+    circuit.consecutiveCompactionFailures = 0;
+    // `compactionCircuitOpenUntil` is deliberately left as the old timestamp
+    // to reproduce the bug condition.
+    expect(circuit.compactionCircuitOpenUntil).toBe(t0 + 60 * 60 * 1000);
+
+    // Three more failures must trip a fresh cooldown even though the old
+    // timestamp is still set.
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(3);
+    expect(circuit.compactionCircuitOpenUntil).toBe(t1 + 60 * 60 * 1000);
+    expect(events).toHaveLength(2);
+    expect(events[1]).toEqual({
+      type: "compaction_circuit_open",
+      conversationId: CONVERSATION_ID,
+      reason: "3_consecutive_failures",
+      openUntil: t1 + 60 * 60 * 1000,
+    });
+  });
+
+  test("(f) isOpen() is query-only and never mutates the counter", async () => {
+    // `maybeCompact()` early-return paths skip `recordOutcome` entirely and
+    // only gate on `isOpen()`. The query must never touch the 3-strike
+    // counter so those early returns can't silently reset it.
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent } = collectEvents();
+
+    await circuit.recordOutcome(true, onEvent);
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(2);
+
+    // Query-only — must NOT touch the counter.
+    expect(await circuit.isOpen()).toBe(false);
+    expect(circuit.consecutiveCompactionFailures).toBe(2);
+
+    // A third real failure then trips the breaker as expected.
+    await circuit.recordOutcome(true, onEvent);
+    expect(circuit.consecutiveCompactionFailures).toBe(3);
+    expect(circuit.compactionCircuitOpenUntil).not.toBeNull();
+  });
+
+  test("(g) open→closed transition emits compaction_circuit_closed exactly once", async () => {
+    // Regression: the reset branch must notify the client on open→closed.
+    // Otherwise the Swift banner set from `compaction_circuit_open` stays
+    // visible until the original `openUntil` deadline (up to 1h),
+    // misrepresenting the live state.
+    const fixedNow = 1_700_000_000_000;
+    Date.now = () => fixedNow;
+
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent, events } = collectEvents();
+
+    // Force the circuit into the open state directly — the emitted-event
+    // transition logic is what we're testing, not the tripping path.
+    circuit.compactionCircuitOpenUntil = fixedNow + 60 * 60 * 1000;
+    circuit.consecutiveCompactionFailures = 3;
+
+    await circuit.recordOutcome(false, onEvent);
+
+    expect(circuit.consecutiveCompactionFailures).toBe(0);
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    expect(events).toHaveLength(1);
+    expect(events[0]).toEqual({
+      type: "compaction_circuit_closed",
+      conversationId: CONVERSATION_ID,
+    });
+  });
+
+  test("(h) successful outcome against an already-closed circuit emits no event", async () => {
+    // Emitting `compaction_circuit_closed` on every successful compaction
+    // would spam the client (the breaker is closed in the common case).
+    // Only the open→closed transition is meaningful.
+    const circuit = new CompactionCircuit(CONVERSATION_ID);
+    const { onEvent, events } = collectEvents();
+
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    await circuit.recordOutcome(false, onEvent);
+    expect(circuit.compactionCircuitOpenUntil).toBeNull();
+    expect(events).toHaveLength(0);
+
+    // A second successful outcome while still closed — still no event.
+    await circuit.recordOutcome(false, onEvent);
+    expect(events).toHaveLength(0);
+  });
+});

package/src/__tests__/compaction-direct.test.ts

@@ -0,0 +1,146 @@
+/**
+ * Tests for the default compaction plugin (`defaultCompact`).
+ *
+ * The agent loop calls {@link defaultCompact} directly with a
+ * {@link CompactionContext} rather than routing through a middleware pipeline.
+ * These tests assert that the default implementation resolves the
+ * conversation's {@link ContextWindowManager} from the compaction store and
+ * forwards the request's conversational options verbatim. The orchestrator
+ * integration path (conversation-agent-loop) is exercised by
+ * `conversation-agent-loop-overflow.test.ts`.
+ */
+
+import { describe, expect, mock, test } from "bun:test";
+
+// `defaultCompact` resolves the manager from the per-conversation compaction
+// store. Register each test's canned 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);
+  },
+}));
+
+const { defaultCompact } =
+  await import("../plugins/defaults/compaction/compact.js");
+
+type ContextWindowResultShape = {
+  compacted: boolean;
+  summaryText: string;
+  messages: unknown[];
+  previousEstimatedInputTokens: number;
+  estimatedInputTokens: number;
+  maxInputTokens: number;
+  thresholdTokens: number;
+  compactedMessages: number;
+  compactedPersistedMessages: number;
+  summaryCalls: number;
+  summaryInputTokens: number;
+  summaryOutputTokens: number;
+  summaryModel: string;
+  reason?: string;
+};
+
+function makeResult(
+  overrides: Partial<ContextWindowResultShape> = {},
+): ContextWindowResultShape {
+  return {
+    compacted: true,
+    summaryText: "default-summary",
+    messages: [],
+    previousEstimatedInputTokens: 1000,
+    estimatedInputTokens: 100,
+    maxInputTokens: 100000,
+    thresholdTokens: 80000,
+    compactedMessages: 3,
+    compactedPersistedMessages: 3,
+    summaryCalls: 1,
+    summaryInputTokens: 500,
+    summaryOutputTokens: 120,
+    summaryModel: "default-model",
+    ...overrides,
+  };
+}
+
+/** Register a stub manager under a conversation id and return its recorder. */
+function registerManager(
+  conversationId: string,
+  result: ContextWindowResultShape,
+) {
+  const observed: {
+    messages: unknown;
+    signal: unknown;
+    options: unknown;
+  }[] = [];
+  fakeContextWindowManagers.set(conversationId, {
+    maybeCompact: async (
+      messages: unknown,
+      signal: unknown,
+      options: unknown,
+    ) => {
+      observed.push({ messages, signal, options });
+      return result;
+    },
+  });
+  return observed;
+}
+
+describe("defaultCompact", () => {
+  test("delegates to the manager and returns its result unchanged", async () => {
+    // GIVEN a manager whose maybeCompact records its arguments
+    const expected = makeResult({
+      summaryText: "manager-summary",
+      compactedMessages: 7,
+    });
+    const observed = registerManager("conv-delegates", expected);
+    const messages = [{ role: "user", content: "hi" }] as never;
+    const signal = new AbortController().signal;
+
+    // WHEN defaultCompact runs with those messages and a signal
+    const result = (await defaultCompact({
+      conversationId: "conv-delegates",
+      messages,
+      signal,
+    })) as unknown as ContextWindowResultShape;
+
+    // THEN the manager saw the same messages and signal
+    expect(observed).toHaveLength(1);
+    expect(observed[0]!.messages).toBe(messages);
+    expect(observed[0]!.signal).toBe(signal);
+
+    // AND the returned result is the manager's object, unmodified
+    expect(result).toBe(expected);
+    expect(result.summaryText).toBe("manager-summary");
+    expect(result.compactedMessages).toBe(7);
+  });
+
+  test("forwards the request's compaction options to the manager", async () => {
+    // GIVEN a manager and a fully-populated compaction context
+    const observed = registerManager("conv-options", makeResult());
+
+    // WHEN defaultCompact runs with force/profile/trust options
+    await defaultCompact({
+      conversationId: "conv-options",
+      messages: [] as never,
+      force: true,
+      overrideProfile: "fast-profile",
+      precomputedEstimate: 1234,
+      minKeepRecentUserTurns: 0,
+      actorTrustClass: "guardian",
+    });
+
+    // THEN the manager received exactly those options, and the conversation id,
+    // messages, and signal are not leaked into the options bag
+    expect(observed).toHaveLength(1);
+    expect(observed[0]!.options).toEqual({
+      force: true,
+      overrideProfile: "fast-profile",
+      precomputedEstimate: 1234,
+      minKeepRecentUserTurns: 0,
+      actorTrustClass: "guardian",
+    });
+  });
+});

package/src/__tests__/compaction-events.test.ts

@@ -12,8 +12,8 @@ import { describe, expect, mock, test } from "bun:test";
 
 import { CompactionCircuit } from "../agent/compaction-circuit.js";
 import type { AgentEvent } from "../agent/loop.js";
-import type { ContextWindowResult } from "../context/window-manager.js";
 import type { ServerMessage } from "../daemon/message-protocol.js";
+import type { ContextWindowResult } from "../plugins/defaults/compaction/window-manager.js";
 import type { Message, ProviderResponse } from "../providers/types.js";
 
 // ---------------------------------------------------------------------------
@@ -187,11 +187,11 @@ let mockCompactResult: ContextWindowResult = {
   summaryText: "",
 };
 
-mock.module("../context/window-manager.js", () => ({
+mock.module("../plugins/defaults/compaction/window-manager.js", () => ({
   ContextWindowManager: class {
     nonPersistedPrefixCount = 0;
-    summaryIsInjected = false;
     constructor() {}
+    updateConfig() {}
     shouldCompact() {
       return { needed: false, estimatedTokens: 0 };
     }
@@ -228,10 +228,10 @@ mock.module("../agent/loop.js", () => ({
     getActiveModel() {
       return undefined;
     }
-    async run(
-      _messages: Message[],
-      _onEvent: (event: AgentEvent) => void,
-    ): Promise<Message[]> {
+    async run(_options: {
+      messages: Message[];
+      onEvent: (event: AgentEvent) => void;
+    }): Promise<Message[]> {
       return [];
     }
   },

package/src/__tests__/compaction.benchmark.test.ts

@@ -12,7 +12,7 @@ import { describe, expect, mock, test } from "bun:test";
 
 import { DEFAULT_CONFIG } from "../config/defaults.js";
 import { estimatePromptTokens } from "../context/token-estimator.js";
-import { ContextWindowManager } from "../context/window-manager.js";
+import { ContextWindowManager } from "../plugins/defaults/compaction/window-manager.js";
 import type { Message, Provider } from "../providers/types.js";
 
 mock.module("../util/logger.js", () => ({

package/src/__tests__/compactor-web-search-strip.test.ts

@@ -0,0 +1,213 @@
+import { describe, expect, mock, test } from "bun:test";
+
+function makeLoggerStub(): Record<string, unknown> {
+  const stub: Record<string, unknown> = {};
+  for (const m of [
+    "info",
+    "warn",
+    "error",
+    "debug",
+    "trace",
+    "fatal",
+    "silent",
+    "child",
+  ]) {
+    stub[m] = m === "child" ? () => makeLoggerStub() : () => {};
+  }
+  return stub;
+}
+
+mock.module("../util/logger.js", () => ({
+  getLogger: () => makeLoggerStub(),
+}));
+
+mock.module("../memory/conversation-crud.js", () => ({
+  getMessages: () => [],
+  reserveMessage: mock(async () => ({ id: "msg-reserve" })),
+}));
+
+mock.module("../memory/attachments-store.js", () => ({
+  getAttachmentMetadataForMessage: () => [],
+  getAttachmentContent: () => null,
+}));
+
+import {
+  runAssistantDrivenCompaction,
+  runEmergencyCompaction,
+} from "../context/compactor.js";
+import type { Message, Provider } from "../providers/types.js";
+
+const TAIL_TIMESTAMP =
+  "2026-05-21 (Thursday) 10:00:00 -05:00 (America/Chicago)";
+
+const compactionResponse = `
+<compaction_result>
+<summary>
+Earlier turns summarized here.
+</summary>
+
+<key_state>
+- Nothing critical pending.
+</key_state>
+
+<tail_start timestamp="${TAIL_TIMESTAMP}" preview="tail anchor message" />
+</compaction_result>
+`;
+
+const ENCRYPTED_TOKEN = "expired_encrypted_token_abc123";
+
+const userText = (text: string): Message => ({
+  role: "user",
+  content: [{ type: "text", text }],
+});
+
+const userTextWithTurnContext = (text: string, timestamp: string): Message => ({
+  role: "user",
+  content: [
+    {
+      type: "text",
+      text: `<turn_context>\ncurrent_time: ${timestamp}\n</turn_context>\n${text}`,
+    },
+  ],
+});
+
+const assistantWithWebSearch = (): Message => ({
+  role: "assistant",
+  content: [
+    {
+      type: "server_tool_use",
+      id: "stu_1",
+      name: "web_search",
+      input: { query: "cats" },
+    },
+    {
+      type: "web_search_tool_result",
+      tool_use_id: "stu_1",
+      content: [
+        {
+          type: "web_search_result",
+          url: "https://cats.example",
+          title: "All About Cats",
+          encrypted_content: ENCRYPTED_TOKEN,
+        },
+      ],
+    },
+  ],
+});
+
+const assistantToolUse = (): Message => ({
+  role: "assistant",
+  content: [
+    { type: "tool_use", id: "tu_1", name: "bash", input: { cmd: "ls" } },
+  ],
+});
+
+const toolResult = (): Message => ({
+  role: "user",
+  content: [{ type: "tool_result", tool_use_id: "tu_1", content: "ok" }],
+});
+
+function serializeBlocks(messages: Message[]): string {
+  return JSON.stringify(messages);
+}
+
+// Records the exact message list the provider is asked to summarize.
+function recordingProvider(sink: { sent: Message[] }): Provider {
+  return {
+    name: "mock-provider",
+    sendMessage: async (msgs: Message[]) => {
+      sink.sent = msgs;
+      return {
+        content: [{ type: "text", text: compactionResponse }],
+        model: "mock-model",
+        usage: { inputTokens: 100, outputTokens: 50 },
+        stopReason: "end_turn",
+      };
+    },
+  };
+}
+
+// Expired Anthropic web-search tokens must never reach the summarization LLM:
+// both compaction provider calls funnel through the same request builder, so
+// each is guarded here.
+describe("compaction summary calls — historical web-search sanitization", () => {
+  test("runAssistantDrivenCompaction strips expired encrypted_content from the request while leaving durable history untouched", async () => {
+    // GIVEN a history whose older turn contains a native web_search_tool_result
+    // block carrying an opaque, expired `encrypted_content` token
+    const messages: Message[] = [
+      userText("old user turn"),
+      assistantWithWebSearch(),
+      userTextWithTurnContext("tail anchor message", TAIL_TIMESTAMP),
+      assistantWithWebSearch(),
+    ];
+
+    // AND a snapshot of the caller's array so we can assert it is not mutated
+    const inputSnapshot = serializeBlocks(messages);
+    const sink = { sent: [] as Message[] };
+
+    // WHEN compaction runs over that history
+    const result = await runAssistantDrivenCompaction({
+      conversationId: "conv-web-search",
+      messages,
+      provider: recordingProvider(sink),
+      systemPrompt: "system",
+      compaction: { enabled: true, autoThreshold: 0.7 },
+      maxInputTokens: 1000,
+      previousEstimatedInputTokens: 900,
+    });
+
+    // THEN the summary request the provider received carries no expired token
+    // and no native web_search blocks, but keeps the search content as text
+    const sentSerialized = serializeBlocks(sink.sent);
+    expect(sentSerialized).not.toContain(ENCRYPTED_TOKEN);
+    expect(sentSerialized).not.toContain("web_search_tool_result");
+    expect(sentSerialized).not.toContain("server_tool_use");
+    expect(sentSerialized).toContain("All About Cats");
+    expect(sentSerialized).toContain("https://cats.example");
+
+    // AND the caller's durable history array is left byte-for-byte untouched,
+    // so persisted messages keep the original rich blocks
+    expect(serializeBlocks(messages)).toBe(inputSnapshot);
+    expect(result.compacted).toBe(true);
+  });
+
+  test("runEmergencyCompaction strips expired encrypted_content from the prefix request", async () => {
+    // GIVEN a mid-turn history whose prefix (before the last tool pair) holds a
+    // web_search_tool_result block with an expired token
+    const messages: Message[] = [
+      userText("old user turn"),
+      assistantWithWebSearch(),
+      userText("keep going"),
+      assistantToolUse(),
+      toolResult(),
+    ];
+
+    const inputSnapshot = serializeBlocks(messages);
+    const sink = { sent: [] as Message[] };
+
+    // WHEN emergency compaction splits at the last tool pair and summarizes the
+    // prefix
+    const result = await runEmergencyCompaction({
+      conversationId: "conv-web-search-emergency",
+      messages,
+      provider: recordingProvider(sink),
+      systemPrompt: "system",
+      compaction: { enabled: true, autoThreshold: 0.7 },
+      // Large budget so the prefix is not front-truncated — this test exercises
+      // the web-search sanitizer, not the truncation fallback.
+      maxInputTokens: 100000,
+      previousEstimatedInputTokens: 90000,
+    });
+
+    // THEN the prefix sent to the provider carries no expired token
+    const sentSerialized = serializeBlocks(sink.sent);
+    expect(sentSerialized).not.toContain(ENCRYPTED_TOKEN);
+    expect(sentSerialized).not.toContain("web_search_tool_result");
+    expect(sentSerialized).not.toContain("server_tool_use");
+    expect(sentSerialized).toContain("All About Cats");
+
+    // AND the caller's durable history is untouched
+    expect(serializeBlocks(messages)).toBe(inputSnapshot);
+    expect(result.compacted).toBe(true);
+  });
+});