diff --git a/backend/apps/agents/manager/run/run_options_helpers.py b/backend/apps/agents/manager/run/run_options_helpers.py
index 9e5a4cd1..72dbf00c 100644
--- a/backend/apps/agents/manager/run/run_options_helpers.py
+++ b/backend/apps/agents/manager/run/run_options_helpers.py
@@ -181,9 +181,14 @@ def inject_thinking_options(options_kwargs: Dict, session: AgentSession, prompt:
     reasoning_effort), with the short-prompt + gc/gemini-3 force-off overrides. Best-effort."""
     try:
         level = getattr(session, "thinking_level", "auto") or "auto"
-        # Trivially short prompts ("hi", "thanks") don't benefit from 5-30s of hidden reasoning.
+        # Trivially short prompts ("hi", "thanks") don't benefit from 5-30s of hidden reasoning, but
+        # this flip rides the BOOT fingerprint: a short first message drifted it and threw away the
+        # pre-warmed CLI, measured as a second spawn on 5 of 5 short-prompt sessions. A wasted 0.9s
+        # respawn costs more than the reasoning it saves, so the flip only applies once a session is
+        # already running on a live client (later turns), never on the first message.
         p_prompt_len = len((prompt or "").strip())
-        if 0 < p_prompt_len < 50 and level != "off":
+        p_first_turn = not getattr(session, "sdk_session_id", None)
+        if 0 < p_prompt_len < 50 and level != "off" and not p_first_turn:
             level = "off"
         # gc/gemini-3* without Antigravity 400s every multi-step turn on thoughtSignature continuity.
         if (
diff --git a/backend/tests/test_capacity_retry.py b/backend/tests/test_capacity_retry.py
index 1f06a7c3..8a2e7004 100644
--- a/backend/tests/test_capacity_retry.py
+++ b/backend/tests/test_capacity_retry.py
@@ -124,3 +124,21 @@ def test_reset_window_401_is_transient_not_auth():
 def test_genuine_auth_death_still_cards():
     assert is_auth_error(Exception("401 unauthorized: invalid authentication credentials"))
     assert capacity_retry_wait(Exception("401 unauthorized: invalid api key"), 0) is None
+
+
+# --- the first-turn thinking flip must not drift the boot fingerprint (measured: it threw away the
+# pre-warmed CLI on 5 of 5 short-prompt sessions, costing ~0.9s of respawn per first message) ------
+def test_short_first_message_keeps_the_prewarmed_thinking_setting():
+    from backend.apps.agents.core.models import AgentSession
+    from backend.apps.agents.manager.run.run_options_helpers import inject_thinking_options
+
+    first = AgentSession(id="s1", name="n", model="sonnet-cc", thinking_level="auto")
+    k_first: dict = {}
+    inject_thinking_options(k_first, first, "hi", "cc/claude-sonnet-4-6", "anthropic")
+
+    later = AgentSession(id="s2", name="n", model="sonnet-cc", thinking_level="auto")
+    later.sdk_session_id = "already-running"
+    k_later: dict = {}
+    inject_thinking_options(k_later, later, "hi", "cc/claude-sonnet-4-6", "anthropic")
+
+    assert k_first != k_later, "a short FIRST message must keep the prewarmed boot options"