diff --git a/apps/server/src/cli/thread.test.ts b/apps/server/src/cli/thread.test.ts
index c2e4bc5c8168c..e09f3be00312b 100644
--- a/apps/server/src/cli/thread.test.ts
+++ b/apps/server/src/cli/thread.test.ts
@@ -2,7 +2,7 @@ import { assert, it } from "@effect/vitest";
 import { RunId, ThreadId } from "@t3tools/contracts";
 import * as Effect from "effect/Effect";
 
-import { type ThreadTools, waitForRun } from "./thread.ts";
+import { readLatestMessages, type ThreadTools, waitForRun } from "./thread.ts";
 
 const threadId = ThreadId.make("thread-1");
 const runId = RunId.make("run-1");
@@ -73,3 +73,90 @@ it.effect("--wait reports its run's result when a later run is the one waiting",
     });
   }),
 );
+
+/** Serves t3_thread_read pages the way the server does: positions count every item. */
+const timelineTools = (types: ReadonlyArray<string>) => {
+  const items = types.map((type, position) => ({ position, itemId: `item-${position}`, type }));
+  const served: Array<number> = [];
+  const call = ((_tool: string, input: unknown) => {
+    const args = input as { readonly afterPosition?: number; readonly limit: number };
+    // Scanning one match past the page keeps reads of huge fixtures linear.
+    const matching: typeof items = [];
+    for (let position = (args.afterPosition ?? -1) + 1; position < items.length; position++) {
+      const item = items[position]!;
+      if (["user_message", "assistant_message", "proposed_plan"].includes(item.type)) {
+        matching.push(item);
+        if (matching.length > args.limit) break;
+      }
+    }
+    const page = matching.slice(0, args.limit);
+    served.push(page.length);
+    return Effect.succeed({
+      thread: { itemCount: items.length },
+      items: page,
+      nextPosition: page.at(-1)?.position ?? null,
+      hasMore: matching.length > page.length,
+    });
+  }) as ThreadTools["call"];
+  const latest = (count: number) =>
+    items
+      .filter((item) => item.type === "user_message" || item.type === "assistant_message")
+      .slice(-count)
+      .map((item) => item.itemId);
+  return { tools: { call } satisfies ThreadTools, served, latest };
+};
+
+const tail = (tools: number) => [...Array<string>(tools).fill("tool_call"), "assistant_message"];
+
+it.effect("show reads the end of a long thread, not its whole history", () =>
+  Effect.gen(function* () {
+    const { tools, served, latest } = timelineTools([
+      ...Array.from({ length: 1000 }, (_, index) =>
+        index % 2 === 0 ? "user_message" : "assistant_message",
+      ),
+      ...Array.from({ length: 10 }, () => tail(30)).flat(),
+      "proposed_plan",
+    ]);
+
+    const { messages } = yield* readLatestMessages(tools, threadId, 10);
+
+    assert.deepStrictEqual(
+      messages.map((item) => item.itemId),
+      latest(10),
+    );
+    assert.isBelow(
+      served.reduce((sum, count) => sum + count, 0),
+      100,
+    );
+  }),
+);
+
+it.effect("show widens back to the first item when the thread has fewer messages", () =>
+  Effect.gen(function* () {
+    const { tools, latest } = timelineTools([
+      "user_message",
+      ...Array.from({ length: 4 }, () => tail(150)).flat(),
+    ]);
+
+    const { messages } = yield* readLatestMessages(tools, threadId, 10);
+
+    assert.deepStrictEqual(
+      messages.map((item) => item.itemId),
+      latest(10),
+    );
+    assert.lengthOf(messages, 5);
+  }),
+);
+
+it.effect("show returns a --messages count larger than one call can spread", () =>
+  Effect.gen(function* () {
+    const { tools, latest } = timelineTools(Array<string>(200_000).fill("assistant_message"));
+
+    const { messages } = yield* readLatestMessages(tools, threadId, 200_000);
+
+    assert.deepStrictEqual(
+      messages.map((item) => item.itemId),
+      latest(200_000),
+    );
+  }),
+);
diff --git a/apps/server/src/cli/thread.ts b/apps/server/src/cli/thread.ts
index 989028afccb9d..6aa28a2c18810 100644
--- a/apps/server/src/cli/thread.ts
+++ b/apps/server/src/cli/thread.ts
@@ -332,6 +332,63 @@ const readTimeline = Effect.fn("cli.thread.readTimeline")(function* (
   return { thread: page.thread, items };
 });
 
+/**
+ * The latest `count` user and assistant messages, oldest first. Positions
+ * count every timeline item, so the read starts near the end and widens
+ * backward only while tool and plan items leave too few messages.
+ */
+export const readLatestMessages = Effect.fn("cli.thread.readLatestMessages")(function* (
+  tools: ThreadTools,
+  threadId: ThreadId,
+  count: number,
+) {
+  let { thread } = yield* tools.call(
+    "t3_thread_read",
+    { threadId, limit: 1, runLimit: 1 },
+    OrchestratorMcpThreadReadResult,
+  );
+  let messages: Array<OrchestratorMcpThreadTimelineItem> = [];
+  // The newest window has no upper bound, so items added since the first
+  // read are still shown.
+  let before: number | undefined;
+  let start = thread.itemCount;
+  let span = Math.max(count, READ_PAGE_LIMIT);
+  do {
+    const end = before;
+    start = Math.max(0, start - span);
+    const window: Array<OrchestratorMcpThreadTimelineItem> = [];
+    let afterPosition = start - 1;
+    let page: OrchestratorMcpThreadReadResult;
+    do {
+      page = yield* tools.call(
+        "t3_thread_read",
+        {
+          threadId,
+          view: "messages",
+          limit: READ_PAGE_LIMIT,
+          runLimit: 1,
+          ...(afterPosition < 0 ? {} : { afterPosition }),
+        },
+        OrchestratorMcpThreadReadResult,
+      );
+      window.push(
+        ...page.items.filter(
+          (item) =>
+            (item.type === "user_message" || item.type === "assistant_message") &&
+            (end === undefined || item.position < end),
+        ),
+      );
+      afterPosition = page.nextPosition ?? Number.POSITIVE_INFINITY;
+    } while (page.hasMore && afterPosition < (end ?? Number.POSITIVE_INFINITY) - 1);
+    thread = page.thread;
+    // A window can hold the whole --messages count, too many to spread as arguments.
+    messages = window.concat(messages);
+    before = start;
+    span *= 2;
+  } while (messages.length < count && start > 0);
+  return { thread, messages: messages.slice(Math.max(0, messages.length - count)) };
+});
+
 /** Item text past the read's per-item cap. */
 const readFullText = Effect.fn("cli.thread.readFullText")(function* (
   tools: ThreadTools,
@@ -661,11 +718,11 @@ const threadShowCommand = Command.make("show", {
       yield* withThreadTools(
         flags,
         Effect.fn("cli.thread.show")(function* (tools) {
-          const { thread, items } = yield* readTimeline(tools, threadId, "messages");
-          const conversation = items.filter(
-            (item) => item.type === "user_message" || item.type === "assistant_message",
+          const { thread, messages: latest } = yield* readLatestMessages(
+            tools,
+            threadId,
+            flags.messages,
           );
-          const latest = conversation.slice(Math.max(0, conversation.length - flags.messages));
           const messages = yield* Effect.forEach(latest, (item) =>
             readFullText(tools, threadId, item).pipe(
               Effect.map((text) => ({