diff --git a/apps/desktop/src/app/session/hooks/use-session-actions.test.tsx b/apps/desktop/src/app/session/hooks/use-session-actions.test.tsx
index ea5cc084ce..866194227b 100644
--- a/apps/desktop/src/app/session/hooks/use-session-actions.test.tsx
+++ b/apps/desktop/src/app/session/hooks/use-session-actions.test.tsx
@@ -6,7 +6,13 @@ import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 
 import { $terminalTakeover, setTerminalTakeover } from '@/app/right-sidebar/store'
 import { noteActiveTreeGroup, revealTreePane } from '@/components/pane-shell/tree/store'
-import { getAllSessionMessages, getLatestSessionMessages, getSession, type SessionInfo } from '@/hermes'
+import {
+  getAllSessionMessages,
+  getLatestSessionMessages,
+  getSession,
+  type SessionInfo,
+  type SessionResumeResponse
+} from '@/hermes'
 import { createClientSessionState } from '@/lib/chat-runtime'
 import { clearSessionDraft, stashSessionDraft, takeSessionDraft } from '@/store/composer'
 import { $activeGatewayProfile, $newChatProfile, ensureGatewayProfile } from '@/store/profile'
@@ -1736,13 +1742,65 @@ describe('resumeSession warm-cache mapping integrity', () => {
     // resume RPC ran, for the session that was actually requested.
     const resumeCalls = requestGateway.mock.calls.filter(([method]) => method === 'session.resume')
     expect(resumeCalls.length).toBe(1)
-    expect(resumeCalls[0][1]).toMatchObject({ session_id: 'stored-A' })
+    expect(resumeCalls[0][1]).toMatchObject({
+      defer_history: true,
+      session_id: 'stored-A'
+    })
+    expect(getLatestSessionMessages).toHaveBeenCalledWith('stored-A', undefined)
 
     // The corrupt mapping was purged so it can't mis-resolve again.
     expect(runtimeIdByStoredSessionIdRef.current.has('stored-A')).toBe(false)
     expect(sessionStateByRuntimeIdRef.current.has('rt-recycled')).toBe(false)
   })
 
+  it('paints the bounded latest transcript before deferred resume finishes', async () => {
+    const latestPage = Array.from({ length: 500 }, (_, index) => ({
+      content: `message-${index}`,
+      role: index % 2 === 0 ? ('user' as const) : ('assistant' as const),
+      timestamp: index + 1
+    }))
+
+    setSessions([storedSession({ id: 'stored-A', message_count: 50_000 })])
+    vi.mocked(getLatestSessionMessages).mockReset()
+    vi.mocked(getLatestSessionMessages).mockResolvedValue({
+      messages: latestPage,
+      pagination: { limit: 500, offset: 0, order: 'latest', returned: 500 },
+      session_id: 'stored-A'
+    })
+
+    const deferredResume = deferred<SessionResumeResponse>()
+
+    const requestGatewayMock = vi.fn((method: string, _params?: Record<string, unknown>) => {
+      if (method === 'session.resume') {
+        return deferredResume.promise
+      }
+
+      return Promise.resolve({})
+    })
+
+    const requestGateway = <T,>(method: string, params?: Record<string, unknown>): Promise<T> =>
+      requestGatewayMock(method, params) as Promise<T>
+
+    let resume: ((storedSessionId: string, replaceRoute?: boolean) => Promise<unknown>) | null = null
+    render(<ResumeHarness onReady={value => (resume = value)} requestGateway={requestGateway} />)
+    await waitFor(() => expect(resume).not.toBeNull())
+    const resumePromise = resume!('stored-A', true)
+
+    await waitFor(() => expect($messages.get()).toHaveLength(500))
+    expect(getLatestSessionMessages).toHaveBeenCalledTimes(1)
+    expect(getLatestSessionMessages).toHaveBeenCalledWith('stored-A', undefined)
+
+    deferredResume.resolve({
+      session_id: 'rt-A',
+      resumed: 'stored-A',
+      message_count: 500,
+      messages: [],
+      info: {}
+    })
+    await resumePromise
+    expect($messages.get()).toHaveLength(500)
+  })
+
   it('honours a warm cache entry whose stored id matches and refreshes its persisted transcript', async () => {
     // Correctly-wired mapping: 'rt-A' <-> 'stored-A'. The fast-path should trust
     // it and never reach session.resume. session.activate refreshes the live
diff --git a/apps/desktop/src/app/session/hooks/use-session-actions/index.ts b/apps/desktop/src/app/session/hooks/use-session-actions/index.ts
index 361a553fd0..57e93046b9 100644
--- a/apps/desktop/src/app/session/hooks/use-session-actions/index.ts
+++ b/apps/desktop/src/app/session/hooks/use-session-actions/index.ts
@@ -1030,6 +1030,7 @@ export function useSessionActions({
           session_id: storedSessionId,
           cols: 96,
           source: 'desktop',
+          defer_history: !watchWindow,
           // REST is the transcript authority for Desktop. Avoid duplicating a
           // potentially huge compression lineage in the WebSocket response.
           // Watch windows attach lazily (live mirror). Every other cold resume
@@ -1079,7 +1080,6 @@ export function useSessionActions({
           prefetchApplied = true
           prefetchedStoredSessionId = prefetchedResult.session_id || storedSessionId
         }
-
         const currentMessages = $messages.get()
 
         // Keep the local snapshot when resume would only reshuffle runtime
diff --git a/apps/desktop/src/types/hermes.ts b/apps/desktop/src/types/hermes.ts
index b6291e8455..c9c28eef8f 100644
--- a/apps/desktop/src/types/hermes.ts
+++ b/apps/desktop/src/types/hermes.ts
@@ -606,6 +606,7 @@ export interface SessionResumeResponse {
     attempt: number
     interrupted_at: number
   }
+  hydrating?: boolean
   inflight?: null | {
     assistant?: string
     /** Mid-turn redirect corrections, oldest first. The turn's original prompt
diff --git a/tests/test_tui_gateway_server.py b/tests/test_tui_gateway_server.py
index fa6ba36e7e..26fa685f18 100644
--- a/tests/test_tui_gateway_server.py
+++ b/tests/test_tui_gateway_server.py
@@ -3282,6 +3282,221 @@ def test_lazy_child_watch_resume_serves_candidate_inclusive_display(monkeypatch,
     assert texts == ["child prompt", "child substantive answer", "child terse reply"]
 
 
+def test_session_resume_deferred_history_acknowledges_and_reuses(monkeypatch):
+    history_started = threading.Event()
+    release_history = threading.Event()
+    build_started = threading.Event()
+    history_calls = []
+    auto_continue_calls = []
+    ancestor = {"role": "assistant", "content": "ancestor"}
+    loaded = {"role": "user", "content": "loaded"}
+
+    class FakeDB:
+        def get_session(self, target):
+            return {"id": target, "message_count": 1200}
+
+        def resolve_resume_session_id(self, target):
+            return target
+
+        def reopen_session(self, target):
+            assert target == "large-session"
+
+        def get_resume_conversations(self, target):
+            history_calls.append(("resume", target))
+            history_started.set()
+            assert release_history.wait(timeout=2.0)
+            return [loaded], [ancestor, loaded]
+
+        def get_ancestor_display_prefix(self, target):
+            history_calls.append(("prefix", target))
+            return [ancestor]
+
+    monkeypatch.setattr(server, "_get_db", lambda: FakeDB())
+    monkeypatch.setattr(server, "_enable_gateway_prompts", lambda: None)
+    monkeypatch.setattr(server, "_schedule_session_cap_enforcement", lambda: None)
+    monkeypatch.setattr(
+        server,
+        "_start_agent_build",
+        lambda _sid, _session: build_started.set(),
+    )
+    monkeypatch.setattr(
+        server,
+        "_maybe_schedule_auto_continue",
+        lambda sid, session, stored_id: auto_continue_calls.append(
+            (sid, session, stored_id)
+        ),
+    )
+
+    try:
+        first = server._methods["session.resume"](
+            "r1",
+            {
+                "session_id": "large-session",
+                "source": "desktop",
+                "defer_history": True,
+            },
+        )
+
+        assert first["result"]["hydrating"] is True
+        assert first["result"]["messages"] == []
+        assert first["result"]["message_count"] == 1200
+        assert history_started.wait(timeout=1.0)
+
+        second = server._methods["session.resume"](
+            "r2",
+            {"session_id": "large-session", "defer_history": True},
+        )
+        assert second["result"]["session_id"] == first["result"]["session_id"]
+        assert second["result"]["hydrating"] is True
+        assert second["result"]["messages"] == []
+
+        release_history.set()
+        sid = first["result"]["session_id"]
+        assert server._sessions[sid]["resume_history_ready"].wait(timeout=1.0)
+        assert build_started.wait(timeout=1.0)
+        assert history_calls == [
+            ("resume", "large-session"),
+            ("prefix", "large-session"),
+        ]
+        assert server._sessions[sid]["history"] == [loaded]
+        assert server._sessions[sid]["display_history_prefix"] == [ancestor]
+        assert server._sessions[sid]["resume_message_count"] == 2
+        assert auto_continue_calls == [(sid, server._sessions[sid], "large-session")]
+    finally:
+        release_history.set()
+        for sid, session in list(server._sessions.items()):
+            if session.get("session_key") == "large-session":
+                lease = session.get("active_session_lease")
+                if lease is not None:
+                    lease.release()
+                server._sessions.pop(sid, None)
+
+
+def test_session_resume_deferred_history_failure_can_retry(monkeypatch):
+    first_released = threading.Event()
+    build_started = threading.Event()
+    attempts = 0
+
+    class FakeDB:
+        def get_session(self, target):
+            return {"id": target, "message_count": 1}
+
+        def resolve_resume_session_id(self, target):
+            return target
+
+        def reopen_session(self, _target):
+            pass
+
+        def get_resume_conversations(self, _target):
+            nonlocal attempts
+            attempts += 1
+            if attempts == 1:
+                first_released.set()
+                raise RuntimeError("sqlite read failed")
+            loaded = [{"role": "user", "content": "retry loaded"}]
+            return loaded, loaded
+
+        def get_ancestor_display_prefix(self, _target):
+            return []
+
+    monkeypatch.setattr(server, "_get_db", lambda: FakeDB())
+    monkeypatch.setattr(
+        server,
+        "_claim_active_session_slot",
+        lambda *_args, **_kwargs: pytest.fail(
+            "resume must not claim a session slot before the first prompt"
+        ),
+    )
+    monkeypatch.setattr(server, "_enable_gateway_prompts", lambda: None)
+    monkeypatch.setattr(server, "_schedule_session_cap_enforcement", lambda: None)
+    monkeypatch.setattr(
+        server,
+        "_start_agent_build",
+        lambda _sid, _session: build_started.set(),
+    )
+
+    try:
+        first = server._methods["session.resume"](
+            "r1",
+            {"session_id": "retry-session", "defer_history": True},
+        )
+        first_sid = first["result"]["session_id"]
+        assert first_released.wait(timeout=1.0)
+        assert first_sid not in server._sessions
+
+        second = server._methods["session.resume"](
+            "r2",
+            {"session_id": "retry-session", "defer_history": True},
+        )
+        second_sid = second["result"]["session_id"]
+        assert second_sid != first_sid
+        assert server._sessions[second_sid]["resume_history_ready"].wait(timeout=1.0)
+        assert build_started.wait(timeout=1.0)
+    finally:
+        for sid, session in list(server._sessions.items()):
+            if session.get("session_key") == "retry-session":
+                lease = session.get("active_session_lease")
+                if lease is not None:
+                    lease.release()
+                server._sessions.pop(sid, None)
+
+
+def test_session_resume_deferred_history_close_cancels_build(monkeypatch):
+    history_started = threading.Event()
+    release_history = threading.Event()
+    build_started = threading.Event()
+
+    class FakeDB:
+        def get_session(self, target):
+            return {"id": target, "message_count": 1}
+
+        def resolve_resume_session_id(self, target):
+            return target
+
+        def reopen_session(self, _target):
+            pass
+
+        def get_resume_conversations(self, _target):
+            history_started.set()
+            assert release_history.wait(timeout=2.0)
+            loaded = [{"role": "user", "content": "late"}]
+            return loaded, loaded
+
+        def get_ancestor_display_prefix(self, _target):
+            return []
+
+    monkeypatch.setattr(server, "_get_db", lambda: FakeDB())
+    monkeypatch.setattr(server, "_enable_gateway_prompts", lambda: None)
+    monkeypatch.setattr(server, "_schedule_session_cap_enforcement", lambda: None)
+    monkeypatch.setattr(
+        server,
+        "_start_agent_build",
+        lambda _sid, _session: build_started.set(),
+    )
+
+    response = {}
+    try:
+        response = server._methods["session.resume"](
+            "r1",
+            {"session_id": "cancel-session", "defer_history": True},
+        )
+        sid = response["result"]["session_id"]
+        session = server._sessions[sid]
+        assert history_started.wait(timeout=1.0)
+
+        assert server._close_session_by_id(sid, end_reason="tui_close") is True
+        assert session["resume_history_ready"].is_set()
+        assert session["resume_history_error"] == "session resume cancelled"
+
+        release_history.set()
+        time.sleep(0.05)
+        assert not build_started.is_set()
+        assert sid not in server._sessions
+    finally:
+        release_history.set()
+        server._sessions.pop(response.get("result", {}).get("session_id", ""), None)
+
+
 def test_session_resume_follows_compression_tip(monkeypatch, tmp_path):
     """Resuming a rotated-out parent id must load the continuation's messages.
 
diff --git a/tests/tui_gateway/test_session_resume_db_ownership.py b/tests/tui_gateway/test_session_resume_db_ownership.py
index 309d1e3998..f31117b7cb 100644
--- a/tests/tui_gateway/test_session_resume_db_ownership.py
+++ b/tests/tui_gateway/test_session_resume_db_ownership.py
@@ -28,6 +28,7 @@ Pinned here, in both directions:
 
 from __future__ import annotations
 
+import threading
 import types
 
 import pytest
@@ -196,6 +197,43 @@ def test_resume_closes_profile_db_on_deferred_cold_resume(profile_dbs, monkeypat
     assert profile_dbs[0].closed == 1
 
 
+def test_resume_hands_profile_db_to_deferred_history_worker(profile_dbs, monkeypatch):
+    """Incremental hydration owns the profile handle until its read completes."""
+    history_started = threading.Event()
+    release_history = threading.Event()
+
+    class _BlockingDB(_RecordingDB):
+        def get_resume_conversations(self, _target):
+            history_started.set()
+            assert release_history.wait(timeout=2.0)
+            assert self.closed == 0
+            return ([], [])
+
+    def _factory(db_path=None, **kwargs):
+        db = _BlockingDB(db_path=db_path, **kwargs)
+        db.rows["s1"] = {"id": "s1", "cwd": "", "message_count": 0}
+        profile_dbs.append(db)
+        return db
+
+    monkeypatch.setattr("hermes_state.SessionDB", _factory)
+    monkeypatch.setattr(server, "_stored_session_runtime_overrides", lambda _found: {})
+
+    try:
+        resp = _resume(
+            session_id="s1", profile="work", defer_history=True
+        )
+        sid = resp["result"]["session_id"]
+        db = profile_dbs[0]
+        assert history_started.wait(timeout=1.0)
+        assert db.closed == 0
+
+        release_history.set()
+        assert server._sessions[sid]["resume_history_ready"].wait(timeout=1.0)
+        assert db.closed == 1
+    finally:
+        release_history.set()
+
+
 def test_resume_keeps_profile_db_open_after_ownership_transfer(profile_dbs, monkeypatch):
     """A COMPLETED resume transfers the handle to the agent — do not close it.
 
diff --git a/tui_gateway/methods_session.py b/tui_gateway/methods_session.py
index d7e95edcdc..7e357ab030 100644
--- a/tui_gateway/methods_session.py
+++ b/tui_gateway/methods_session.py
@@ -317,6 +317,7 @@ def _(rid, params: dict) -> dict:
     # local profile's state.db. None/own profile → the launch profile (unchanged).
     profile = (params.get("profile") or "").strip() or None
     profile_home = _profile_home(profile)
+    defer_history = is_truthy_value(params.get("defer_history", False))
     # Desktop hydrates persisted transcripts through the authenticated REST
     # route in parallel. Suppress the duplicate WebSocket transcript only when
     # the caller explicitly requests it; other clients keep upstream behavior.
@@ -424,6 +425,12 @@ def _(rid, params: dict) -> dict:
                 omit_messages=omit_messages,
             )
             payload["resumed"] = target
+            if defer_history:
+                payload["messages"] = []
+                payload["message_count"] = int(
+                    session.get("resume_message_count") or payload["message_count"]
+                )
+                payload["hydrating"] = bool(session.get("resume_hydrating"))
             # A lazy watch session never owns a run loop, so its payload's running
             # flag is always False — overlay the child-run registry so a reconnecting
             # watch window keeps its busy indicator while the child is still mid-run.
@@ -512,6 +519,64 @@ def _(rid, params: dict) -> dict:
                 },
             )
 
+        # Desktop can ask for a bounded acknowledgement and hydrate the display
+        # transcript through the paginated REST endpoint. Register the runtime now,
+        # then load model history and initialize optional providers in background.
+        # Repeated requests reuse the record through the live fast path above.
+        if defer_history and not is_truthy_value(params.get("eager_build", False)):
+            sid = uuid.uuid4().hex[:8]
+            source = _resolve_session_source(str(params.get("source") or "").strip() or None)
+            lease = None  # claimed lazily on the first turn (_ensure_active_session_slot)
+            _enable_gateway_prompts()
+            overrides = _stored_session_runtime_overrides(found) or {}
+            model_override = overrides.get("model_override") or {}
+            cwd = profile_resume_cwd or _default_session_cwd()
+            record = _deferred_session_record(
+                target,
+                cols=cols,
+                cwd=cwd,
+                history=[],
+                lease=lease,
+                source=source,
+                close_on_disconnect=is_truthy_value(params.get("close_on_disconnect", False)),
+                profile_home=profile_home,
+                model_override=overrides.get("model_override"),
+                resume_runtime_overrides=overrides or None,
+            )
+            record["resume_history_ready"] = threading.Event()
+            record["resume_hydrating"] = True
+            record["resume_message_count"] = int(found.get("message_count") or 0)
+            if (live := _claim_or_reuse_live(sid, target, record, lease)) is not None:
+                return _ok(rid, _reuse_live_payload(*live))
+
+            _schedule_resume_hydration(sid, target, db, close_db=owns_db)
+            # The hydration worker now owns a profile-scoped handle and closes it
+            # after the transcript read. The shared launch DB is process-owned.
+            if owns_db:
+                owns_db = False
+            _schedule_session_cap_enforcement()
+            return _ok(
+                rid,
+                {
+                    "session_id": sid,
+                    "resumed": target,
+                    "message_count": record["resume_message_count"],
+                    "messages": [],
+                    "hydrating": True,
+                    "info": _lazy_resume_info(
+                        cwd,
+                        model=model_override.get("model") or "",
+                        provider=overrides.get("provider_override") or "",
+                        profile=profile,
+                    ),
+                    "inflight": None,
+                    "running": False,
+                    "session_key": target,
+                    "started_at": record["created_at"],
+                    "status": "resuming",
+                },
+            )
+
         # Cold resume default: register the live session and read its stored
         # transcript, but build the agent OFF the response path. _make_agent can
         # block for seconds (MCP discovery, prompt/skill build, AIAgent
diff --git a/tui_gateway/server.py b/tui_gateway/server.py
index 267a6839a5..d8dfcaeda3 100644
--- a/tui_gateway/server.py
+++ b/tui_gateway/server.py
@@ -738,6 +738,10 @@ def _finalize_session(session: dict | None, end_reason: str = "tui_close") -> No
     if not session or session.get("_finalized"):
         return
     session["_finalized"] = True
+    history_ready = session.get("resume_history_ready")
+    if history_ready is not None and not history_ready.is_set():
+        session["resume_history_error"] = "session resume cancelled"
+        history_ready.set()
     _release_active_session_slot(session)
     stop_event = session.get("_notif_stop")
     if stop_event is not None:
@@ -2277,6 +2281,15 @@ def _start_agent_build(sid: str, session: dict) -> None:
         owns_db = False
         profile_home = current.get("profile_home")
         try:
+            history_ready = current.get("resume_history_ready")
+            if history_ready is not None:
+                if not history_ready.wait(timeout=300.0):
+                    raise TimeoutError("session history hydration timed out")
+                if history_error := current.get("resume_history_error"):
+                    raise RuntimeError(str(history_error))
+                with _sessions_lock:
+                    if _sessions.get(sid) is not current:
+                        return
             tokens = _set_session_context(key)
             # Build against the session's profile (global-remote): bind its
             # HERMES_HOME so config/skills/model resolve to it, and hand the
@@ -8352,6 +8365,75 @@ def _schedule_agent_build(sid: str, delay: float = 0.05) -> None:
     timer.start()
 
 
+def _schedule_resume_hydration(
+    sid: str, stored_id: str, db, *, close_db: bool = False
+) -> None:
+    """Load a cold resume's transcript off the JSON-RPC response path."""
+
+    def _run() -> None:
+        session = _sessions.get(sid)
+        try:
+            if session is None:
+                return
+            _emit(
+                "session.resume_progress",
+                sid,
+                {"phase": "history", "status": "loading"},
+            )
+            db.reopen_session(stored_id)
+            raw_history, display_history = db.get_resume_conversations(stored_id)
+            prefix = db.get_ancestor_display_prefix(stored_id)
+            history = sanitize_replay_history(raw_history)
+
+            if _sessions.get(sid) is not session:
+                return
+            with session["history_lock"]:
+                session["history"] = history
+                session["display_history_prefix"] = prefix
+                session["resume_hydrating"] = False
+                session["resume_message_count"] = len(display_history)
+            session["resume_history_ready"].set()
+            _emit(
+                "session.resume_progress",
+                sid,
+                {
+                    "message_count": len(display_history),
+                    "phase": "history",
+                    "status": "complete",
+                },
+            )
+            _maybe_schedule_auto_continue(sid, session, stored_id)
+            _start_agent_build(sid, session)
+        except Exception as exc:
+            if _sessions.get(sid) is not session:
+                return
+            message = f"resume failed: {exc}"
+            session["resume_hydrating"] = False
+            session["resume_history_error"] = message
+            session["agent_error"] = message
+            session["resume_history_ready"].set()
+            session["agent_ready"].set()
+            _emit(
+                "session.resume_progress",
+                sid,
+                {"message": message, "phase": "history", "status": "failed"},
+            )
+            _emit("error", sid, {"message": message})
+            with _sessions_lock:
+                discarded = _sessions.pop(sid, None) if _sessions.get(sid) is session else None
+            lease = (discarded or {}).get("active_session_lease")
+            if lease is not None:
+                lease.release()
+        finally:
+            if close_db and hasattr(db, "close"):
+                try:
+                    db.close()
+                except Exception:
+                    logger.debug("failed to close resume db for %s", sid, exc_info=True)
+
+    threading.Thread(target=_run, daemon=True).start()
+
+
 def _session_pending_kind(sid: str) -> str:
     for rid, (owner_sid, _ev) in list(_pending.items()):
         if owner_sid != sid: