@vellumai/assistant 0.8.4 → 0.8.6

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

@@ -214,6 +214,155 @@ describe("forkConversation", () => {
     ]);
   });
 
+  test("pinned fork through a (createdAt, id) cutoff matches the cursor slice for same-timestamp rows", () => {
+    // Regression for the memory-retrospective cutoff/fork divergence: the job
+    // picks its cutoff from `getMessagesAfter`, which orders by `(createdAt,
+    // id)`. `forkConversation` must slice on the same order so same-millisecond
+    // siblings aren't skipped forever or reprocessed. Insert rows whose
+    // insertion order is the reverse of their `(createdAt, id)` order to expose
+    // the divergence: a `createdAt`-only slice would pick the wrong prefix.
+    const source = createConversation("Same-timestamp cutoff thread");
+    const db = getDb();
+    const createdAt = Date.now();
+    // Insert d, c, b, a so SQLite's createdAt-only tie order (≈ rowid /
+    // insertion order) is the opposite of the (createdAt, id) cursor order
+    // (a, b, c, d). All are plain user rows so no display-turn extension fires.
+    for (const id of ["msg-d", "msg-c", "msg-b", "msg-a"]) {
+      db.run(
+        `INSERT INTO messages (id, conversation_id, role, content, created_at) VALUES ('${id}', '${source.id}', 'user', '${id}', ${createdAt})`,
+      );
+    }
+
+    // Cutoff = "msg-c": the cursor treats {msg-a, msg-b, msg-c} as processed
+    // and {msg-d} as still-after-the-cutoff. The fork must contain exactly the
+    // first three in `(createdAt, id)` order.
+    const fork = forkConversation({
+      conversationId: source.id,
+      throughMessageId: "msg-c",
+    });
+
+    expect(getMessages(fork.id).map((message) => message.content)).toEqual([
+      "msg-a",
+      "msg-b",
+      "msg-c",
+    ]);
+  });
+
+  test("advances fork boundary through consecutive assistant rows after the requested message", async () => {
+    // When the read-path merges consecutive assistant DB rows into a single
+    // display row, the client only addresses the anchor id. Forking through
+    // the anchor must still include the merged tail rows that follow.
+    const source = createConversation("Multi-row turn thread");
+    await addMessage(source.id, "user", "Message 1", undefined, {
+      skipIndexing: true,
+    });
+    const anchor = await addMessage(
+      source.id,
+      "assistant",
+      "Assistant text segment",
+      undefined,
+      { skipIndexing: true },
+    );
+    const toolRow = await addMessage(
+      source.id,
+      "assistant",
+      "Tool turn row",
+      undefined,
+      { skipIndexing: true },
+    );
+    const tailRow = await addMessage(
+      source.id,
+      "assistant",
+      "Final assistant segment",
+      undefined,
+      { skipIndexing: true },
+    );
+    await addMessage(source.id, "user", "Next user turn", undefined, {
+      skipIndexing: true,
+    });
+
+    const fork = forkConversation({
+      conversationId: source.id,
+      throughMessageId: anchor.id,
+    });
+
+    // Boundary advances past the entire consecutive-assistant cluster, so the
+    // full turn is preserved in the fork — not just the anchor row.
+    expect(getMessages(fork.id).map((message) => message.content)).toEqual([
+      "Message 1",
+      "Assistant text segment",
+      "Tool turn row",
+      "Final assistant segment",
+    ]);
+    expect(fork.forkParentMessageId).toBe(tailRow.id);
+    expect(toolRow.id).not.toBe(anchor.id);
+  });
+
+  test("advances fork boundary across tool-result-only user rows between assistant rows", async () => {
+    // Read-path collapse folds tool-result-only user rows into the
+    // surrounding assistant turn (`mergeToolResultsIntoAssistantMessages`
+    // suppresses them). The client only sees a single display turn anchored
+    // at the first assistant row, so forking through the anchor must include
+    // both halves of the assistant turn plus the suppressed user row in
+    // between — otherwise the fork loses tool_use ↔ tool_result pairing
+    // and produces an invalid LLM history.
+    const source = createConversation("Tool-result gap thread");
+    await addMessage(
+      source.id,
+      "user",
+      "Find the latest sales numbers",
+      undefined,
+      {
+        skipIndexing: true,
+      },
+    );
+    const anchor = await addMessage(
+      source.id,
+      "assistant",
+      JSON.stringify([
+        { type: "text", text: "Looking up the data." },
+        { type: "tool_use", id: "tool_1", name: "lookup", input: {} },
+      ]),
+      undefined,
+      { skipIndexing: true },
+    );
+    const toolResultUserRow = await addMessage(
+      source.id,
+      "user",
+      JSON.stringify([
+        { type: "tool_result", tool_use_id: "tool_1", content: "data" },
+      ]),
+      undefined,
+      { skipIndexing: true },
+    );
+    const tailAssistantRow = await addMessage(
+      source.id,
+      "assistant",
+      "Here are the numbers.",
+      undefined,
+      { skipIndexing: true },
+    );
+    await addMessage(source.id, "user", "Thanks", undefined, {
+      skipIndexing: true,
+    });
+
+    const fork = forkConversation({
+      conversationId: source.id,
+      throughMessageId: anchor.id,
+    });
+
+    // All three DB rows of the assistant display turn — including the
+    // suppressed tool-result user row in the middle — land in the fork.
+    const forkedContent = getMessages(fork.id).map((m) => m.content);
+    expect(forkedContent).toHaveLength(4);
+    expect(forkedContent[0]).toBe("Find the latest sales numbers");
+    expect(forkedContent[1]).toContain("tool_use");
+    expect(forkedContent[2]).toContain("tool_result");
+    expect(forkedContent[3]).toBe("Here are the numbers.");
+    expect(fork.forkParentMessageId).toBe(tailAssistantRow.id);
+    expect(toolResultUserRow.id).not.toBe(anchor.id);
+  });
+
   test("preserves compacted context when forking from the visible window", async () => {
     const source = createConversation("Compacted thread");
     await addMessage(source.id, "user", "Message 1", undefined, {
@@ -297,7 +446,7 @@ describe("forkConversation", () => {
     ]);
   });
 
-  test("inherits cleanedAt when forking past the clean event", async () => {
+  test("inherits historyStrippedAt when forking past the clean event", async () => {
     const source = createConversation("Clean thread");
     await addMessage(source.id, "user", "Message 1", undefined, {
       skipIndexing: true,
@@ -310,10 +459,10 @@ describe("forkConversation", () => {
       { skipIndexing: true },
     );
 
-    const cleanedAt = preClean.createdAt + 1;
+    const historyStrippedAt = preClean.createdAt + 1;
     getDb()
       .update(conversations)
-      .set({ cleanedAt })
+      .set({ historyStrippedAt })
       .where(eq(conversations.id, source.id))
       .run();
 
@@ -324,17 +473,17 @@ describe("forkConversation", () => {
       undefined,
       { skipIndexing: true },
     );
-    expect(postClean.createdAt).toBeGreaterThanOrEqual(cleanedAt);
+    expect(postClean.createdAt).toBeGreaterThanOrEqual(historyStrippedAt);
 
     const fork = forkConversation({
       conversationId: source.id,
       throughMessageId: postClean.id,
     });
 
-    expect(fork.cleanedAt).toBe(cleanedAt);
+    expect(fork.historyStrippedAt).toBe(historyStrippedAt);
   });
 
-  test("does not inherit cleanedAt when forking before the clean event", async () => {
+  test("does not inherit historyStrippedAt when forking before the clean event", async () => {
     const source = createConversation("Clean thread");
     await addMessage(source.id, "user", "Message 1", undefined, {
       skipIndexing: true,
@@ -347,10 +496,10 @@ describe("forkConversation", () => {
       { skipIndexing: true },
     );
 
-    const cleanedAt = preClean.createdAt + 1;
+    const historyStrippedAt = preClean.createdAt + 1;
     getDb()
       .update(conversations)
-      .set({ cleanedAt })
+      .set({ historyStrippedAt })
       .where(eq(conversations.id, source.id))
       .run();
 
@@ -363,10 +512,10 @@ describe("forkConversation", () => {
       throughMessageId: preClean.id,
     });
 
-    expect(fork.cleanedAt).toBeNull();
+    expect(fork.historyStrippedAt).toBeNull();
   });
 
-  test("inherits cleanedAt on a full-history fork", async () => {
+  test("inherits historyStrippedAt on a full-history fork", async () => {
     const source = createConversation("Clean thread");
     await addMessage(source.id, "user", "Message 1", undefined, {
       skipIndexing: true,
@@ -379,19 +528,19 @@ describe("forkConversation", () => {
       { skipIndexing: true },
     );
 
-    const cleanedAt = last.createdAt - 1;
+    const historyStrippedAt = last.createdAt - 1;
     getDb()
       .update(conversations)
-      .set({ cleanedAt })
+      .set({ historyStrippedAt })
       .where(eq(conversations.id, source.id))
       .run();
 
     const fork = forkConversation({ conversationId: source.id });
 
-    expect(fork.cleanedAt).toBe(cleanedAt);
+    expect(fork.historyStrippedAt).toBe(historyStrippedAt);
   });
 
-  test("leaves cleanedAt null when the source has no clean event", async () => {
+  test("leaves historyStrippedAt null when the source has no clean event", async () => {
     const source = createConversation("Unclean thread");
     await addMessage(source.id, "user", "Message 1", undefined, {
       skipIndexing: true,
@@ -402,7 +551,82 @@ describe("forkConversation", () => {
 
     const fork = forkConversation({ conversationId: source.id });
 
-    expect(fork.cleanedAt).toBeNull();
+    expect(fork.historyStrippedAt).toBeNull();
+  });
+
+  test("fork from a pre-compaction message preserves historical injection metadata", async () => {
+    const source = createConversation("Compacted thread");
+    const m1 = await addMessage(
+      source.id,
+      "user",
+      "Historical question",
+      {
+        pkbContextBlock: "<knowledge_base>\nstale\n</knowledge_base>",
+        nowScratchpadBlock:
+          "<NOW.md Always keep this up to date>\nstale\n</NOW.md>",
+      },
+      { skipIndexing: true },
+    );
+    const reply1 = await addMessage(
+      source.id,
+      "assistant",
+      "Reply 1",
+      undefined,
+      {
+        skipIndexing: true,
+      },
+    );
+    const tail = await addMessage(source.id, "user", "Tail turn", undefined, {
+      skipIndexing: true,
+    });
+    // Pin strictly-increasing timestamps so the pinned fork boundary is
+    // unambiguous. `addMessage` stamps `Date.now()`, and these three rows can
+    // land in the same millisecond; under the `(createdAt, id)` tie-break the
+    // pinned fork uses, that would let `m1`'s slice reorder relative to its
+    // siblings. Distinct timestamps keep this test focused on its intent —
+    // pre-compaction metadata + compaction-state inheritance.
+    const db = getDb();
+    const base = Date.now();
+    db.run(`UPDATE messages SET created_at = ${base} WHERE id = '${m1.id}'`);
+    db.run(
+      `UPDATE messages SET created_at = ${base + 1} WHERE id = '${reply1.id}'`,
+    );
+    db.run(
+      `UPDATE messages SET created_at = ${base + 2} WHERE id = '${tail.id}'`,
+    );
+    const compactedAt = base + 3;
+    getDb()
+      .update(conversations)
+      .set({
+        contextSummary: "summary",
+        contextCompactedMessageCount: 2,
+        contextCompactedAt: compactedAt,
+        historyStrippedAt: compactedAt,
+      })
+      .where(eq(conversations.id, source.id))
+      .run();
+
+    const fork = forkConversation({
+      conversationId: source.id,
+      throughMessageId: m1.id,
+    });
+
+    expect(fork.historyStrippedAt).toBeNull();
+    expect(fork.contextSummary).toBeNull();
+    expect(fork.contextCompactedMessageCount).toBe(0);
+
+    const forkedMessages = getMessages(fork.id);
+    expect(forkedMessages).toHaveLength(1);
+    const meta = parseMetadata(forkedMessages[0].metadata) as Record<
+      string,
+      unknown
+    >;
+    expect(meta.pkbContextBlock).toBe(
+      "<knowledge_base>\nstale\n</knowledge_base>",
+    );
+    expect(meta.nowScratchpadBlock).toBe(
+      "<NOW.md Always keep this up to date>\nstale\n</NOW.md>",
+    );
   });
 
   test("rejects forks when the source conversation has no persisted messages", () => {

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

@@ -45,11 +45,12 @@ mock.module("../config/loader.js", () => ({
 }));
 
 import { addMessage, createConversation } from "../memory/conversation-crud.js";
-import { getDb, resetDb } from "../memory/db-connection.js";
+import { getDb } from "../memory/db-connection.js";
 import { initializeDb } from "../memory/db-init.js";
 import { getPolicy } from "../runtime/auth/route-policy.js";
 import { mintToken } from "../runtime/auth/token-service.js";
 import { RuntimeHttpServer } from "../runtime/http-server.js";
+import { resetDbForTesting } from "./db-test-helpers.js";
 
 
 initializeDb();
@@ -98,7 +99,7 @@ describe("POST /v1/conversations/fork", () => {
 
   afterAll(async () => {
     await server?.stop();
-    resetDb();
+    resetDbForTesting();
   });
 
   test("returns the same conversation summary shape as GET /v1/conversations/:id", async () => {

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

@@ -50,6 +50,7 @@ mock.module("../memory/conversation-crud.js", () => ({
   },
   relinkAttachments: () => 0,
   deleteLastExchange: () => 0,
+  reserveMessage: mock(async () => ({ id: "msg-reserve" })),
 }));
 
 mock.module("../memory/conversation-queries.js", () => ({

package/src/__tests__/conversation-inference-profile-list.test.ts

@@ -42,9 +42,10 @@ import {
   createConversation,
   setConversationInferenceProfile,
 } from "../memory/conversation-crud.js";
-import { getDb, resetDb } from "../memory/db-connection.js";
+import { getDb } from "../memory/db-connection.js";
 import { initializeDb } from "../memory/db-init.js";
 import { RuntimeHttpServer } from "../runtime/http-server.js";
+import { resetDbForTesting } from "./db-test-helpers.js";
 
 initializeDb();
 
@@ -65,7 +66,7 @@ describe("conversation HTTP responses include inferenceProfile", () => {
 
   afterAll(async () => {
     await server?.stop();
-    resetDb();
+    resetDbForTesting();
   });
 
   test("GET /v1/conversations includes inferenceProfile for pinned conversations and omits it when unset", async () => {

package/src/__tests__/conversation-inference-profile-route.test.ts

@@ -30,11 +30,12 @@ import {
   createConversation,
   getConversation,
 } from "../memory/conversation-crud.js";
-import { getDb, resetDb } from "../memory/db-connection.js";
+import { getDb } from "../memory/db-connection.js";
 import { initializeDb } from "../memory/db-init.js";
 import { assistantEventHub } from "../runtime/assistant-event-hub.js";
 import { ROUTES } from "../runtime/routes/conversation-management-routes.js";
 import { BadRequestError, NotFoundError } from "../runtime/routes/errors.js";
+import { resetDbForTesting } from "./db-test-helpers.js";
 
 initializeDb();
 
@@ -57,7 +58,7 @@ describe("PUT /v1/conversations/:id/inference-profile", () => {
   });
 
   afterAll(() => {
-    resetDb();
+    resetDbForTesting();
     mock.restore();
   });
 

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

@@ -102,6 +102,7 @@ mock.module("../memory/conversation-crud.js", () => ({
   },
   setConversationOriginChannelIfUnset: () => {},
   setConversationOriginInterfaceIfUnset: () => {},
+  reserveMessage: mock(async () => ({ id: "msg-reserve" })),
 }));
 
 mock.module("../memory/conversation-queries.js", () => ({
@@ -379,7 +380,7 @@ describe("loadFromDb metadata injection rehydration", () => {
         metadata: JSON.stringify({
           memoryInjectedBlock: "mem payload",
           memoryV2StaticBlock:
-            "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+            "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
           pkbSystemReminderBlock:
             "<system_reminder>\npkb payload\n</system_reminder>",
         }),
@@ -406,7 +407,7 @@ describe("loadFromDb metadata injection rehydration", () => {
       { type: "text", text: "<memory>\nmem payload\n</memory>" },
       {
         type: "text",
-        text: "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+        text: "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
       },
       {
         type: "text",
@@ -416,6 +417,47 @@ describe("loadFromDb metadata injection rehydration", () => {
     ]);
   });
 
+  test("legacy <memory>-wrapped memoryV2StaticBlock rehydrates verbatim", async () => {
+    // `meta.memoryV2StaticBlock` may carry either `<info>…</info>` or
+    // legacy `<memory>…</memory>` wrappers depending on when the row was
+    // persisted. The rehydrate path replays the stored text verbatim,
+    // so both wrappers must round-trip unchanged.
+    mockConversation = defaultConv();
+    mockDbMessages = [
+      {
+        id: "m1",
+        role: "user",
+        content: JSON.stringify([{ type: "text", text: "First turn" }]),
+        metadata: JSON.stringify({
+          memoryV2StaticBlock:
+            "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+        }),
+      },
+      {
+        id: "m2",
+        role: "assistant",
+        content: JSON.stringify([{ type: "text", text: "Reply" }]),
+      },
+      {
+        id: "m3",
+        role: "user",
+        content: JSON.stringify([{ type: "text", text: "Tail turn" }]),
+      },
+    ];
+
+    const conversation = makeConversation();
+    await conversation.loadFromDb();
+    const messages = conversation.getMessages();
+
+    expect(messages[0].content).toEqual([
+      {
+        type: "text",
+        text: "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+      },
+      { type: "text", text: "First turn" },
+    ]);
+  });
+
   test("tail user row skips memoryV2StaticBlock", async () => {
     mockConversation = defaultConv();
     mockDbMessages = [
@@ -434,7 +476,7 @@ describe("loadFromDb metadata injection rehydration", () => {
         role: "user",
         content: JSON.stringify([{ type: "text", text: "Tail" }]),
         metadata: JSON.stringify({
-          memoryV2StaticBlock: "<memory>\n## Essentials\n\nleak\n</memory>",
+          memoryV2StaticBlock: "<info>\n## Essentials\n\nleak\n</info>",
         }),
       },
     ];
@@ -471,7 +513,7 @@ describe("loadFromDb metadata injection rehydration", () => {
           // survive the row-level filter for non-guardian views.
           provenanceTrustClass: "trusted_contact",
           memoryV2StaticBlock:
-            "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+            "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
         }),
       },
       {
@@ -500,7 +542,7 @@ describe("loadFromDb metadata injection rehydration", () => {
     expect(messages[0].content).toEqual([
       {
         type: "text",
-        text: "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+        text: "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
       },
       { type: "text", text: "First" },
     ]);
@@ -510,7 +552,7 @@ describe("loadFromDb metadata injection rehydration", () => {
     // Injection-time layout (per `applyRuntimeInjections` after-memory-
     // prefix splicing in ascending injector order: pkb-context 30,
     // pkb-reminder 35, memory-v2-static 38, now-md 40):
-    //   [<memory __injected>, <memory>v2static</memory>, <NOW.md>,
+    //   [<memory>dynamic</memory>, <info>v2static</info>, <NOW.md>,
     //    <system_reminder>, <knowledge_base>, ...original]
     // Rehydration must reproduce this exactly so Anthropic's prefix cache
     // matches msg[0] across daemon restarts.
@@ -523,7 +565,7 @@ describe("loadFromDb metadata injection rehydration", () => {
         metadata: JSON.stringify({
           memoryInjectedBlock: "mem payload",
           memoryV2StaticBlock:
-            "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+            "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
           nowScratchpadBlock: "<NOW.md>\nnow body\n</NOW.md>",
           pkbSystemReminderBlock:
             "<system_reminder>\npkb reminder body\n</system_reminder>",
@@ -551,7 +593,7 @@ describe("loadFromDb metadata injection rehydration", () => {
       { type: "text", text: "<memory>\nmem payload\n</memory>" },
       {
         type: "text",
-        text: "<memory>\n## Essentials\n\nAlice prefers VS Code.\n</memory>",
+        text: "<info>\n## Essentials\n\nAlice prefers VS Code.\n</info>",
       },
       { type: "text", text: "<NOW.md>\nnow body\n</NOW.md>" },
       {
@@ -575,7 +617,7 @@ describe("loadFromDb metadata injection rehydration", () => {
         metadata: JSON.stringify({
           provenanceTrustClass: "trusted_contact",
           memoryV2StaticBlock:
-            "<memory>\n## Essentials\n\nprivate memory\n</memory>",
+            "<info>\n## Essentials\n\nprivate memory\n</info>",
         }),
       },
       {
@@ -621,7 +663,7 @@ describe("loadFromDb metadata injection rehydration", () => {
         metadata: JSON.stringify({
           provenanceTrustClass: "trusted_contact",
           memoryV2StaticBlock:
-            "<memory>\n## Essentials\n\nprivate memory\n</memory>",
+            "<info>\n## Essentials\n\nprivate memory\n</info>",
         }),
       },
       {
@@ -649,7 +691,7 @@ describe("loadFromDb metadata injection rehydration", () => {
     expect(conversation.getMessages()[0].content).toEqual([
       {
         type: "text",
-        text: "<memory>\n## Essentials\n\nprivate memory\n</memory>",
+        text: "<info>\n## Essentials\n\nprivate memory\n</info>",
       },
       { type: "text", text: "First" },
     ]);

package/src/__tests__/conversation-list-source.test.ts

@@ -44,9 +44,10 @@ mock.module("../config/loader.js", () => ({
 }));
 
 import { createConversation } from "../memory/conversation-crud.js";
-import { getDb, resetDb } from "../memory/db-connection.js";
+import { getDb } from "../memory/db-connection.js";
 import { initializeDb } from "../memory/db-init.js";
 import { RuntimeHttpServer } from "../runtime/http-server.js";
+import { resetDbForTesting } from "./db-test-helpers.js";
 
 initializeDb();
 
@@ -68,7 +69,7 @@ describe("GET /v1/conversations includes source discriminator", () => {
 
   afterAll(async () => {
     await server?.stop();
-    resetDb();
+    resetDbForTesting();
   });
 
   test("returns source for every listed conversation", async () => {

package/src/__tests__/conversation-load-history-repair.test.ts

@@ -104,6 +104,7 @@ mock.module("../memory/conversation-crud.js", () => ({
   },
   setConversationOriginChannelIfUnset: () => {},
   setConversationOriginInterfaceIfUnset: () => {},
+  reserveMessage: mock(async () => ({ id: "msg-reserve" })),
 }));
 
 mock.module("../memory/conversation-queries.js", () => ({
@@ -535,7 +536,7 @@ describe("loadFromDb history repair", () => {
       trustClass: "guardian",
       sourceChannel: "telegram",
     });
-    await conversation.persistUserMessage("Guardian follow-up", []);
+    await conversation.persistUserMessage({ content: "Guardian follow-up" });
     const messagesAfterPersist = conversation.getMessages();
 
     expect(messagesAfterPersist).toHaveLength(5);