diff --git a/strix/config/models.py b/strix/config/models.py
index a54edd4d..858d7dd7 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -5,6 +5,7 @@ from __future__ import annotations
 import os
 from typing import TYPE_CHECKING
 
+import httpx
 from agents import set_default_openai_api, set_default_openai_key, set_tracing_disabled
 from agents.models.multi_provider import MultiProvider
 from agents.retry import (
@@ -21,17 +22,19 @@ if TYPE_CHECKING:
     from strix.config.settings import Settings
 
 
-def request_timeout_extra_args(timeout_s: float | None) -> dict[str, float] | None:
-    """Per-request model timeout (connect + read/inactivity) as ``extra_args``.
+def request_timeout_extra_args(timeout_s: float | None) -> dict[str, httpx.Timeout] | None:
+    """Per-request model timeout as ``extra_args``, forwarded to the provider call.
 
-    Restores pre-v1 behavior: a stalled model stream trips this timeout and is
-    retried by ``DEFAULT_MODEL_RETRY`` instead of hanging the agent indefinitely.
-    The value is forwarded to the underlying ``responses.create`` /
-    ``chat.completions.create`` / ``litellm.acompletion`` call.
+    Uses ``read`` for inactivity (matching pre-v1's per-chunk ``wait_for``): a stalled
+    stream trips ``read`` and is retried by ``DEFAULT_MODEL_RETRY``, while a healthy
+    long stream that keeps emitting tokens never trips it. An explicit ``httpx.Timeout``
+    (rather than a scalar) keeps this a read-inactivity timeout instead of a
+    total-duration deadline on every httpx-based backend (``responses.create`` /
+    ``chat.completions.create`` / ``litellm.acompletion``).
     """
     if not timeout_s or timeout_s <= 0:
         return None
-    return {"timeout": float(timeout_s)}
+    return {"timeout": httpx.Timeout(timeout_s, connect=min(timeout_s, 30.0))}
 
 
 def _retry_statusless_provider_errors(context: RetryPolicyContext) -> bool:
diff --git a/tests/test_inputs.py b/tests/test_inputs.py
index 7c51c795..b14280c1 100644
--- a/tests/test_inputs.py
+++ b/tests/test_inputs.py
@@ -164,7 +164,8 @@ def test_make_model_settings_sets_request_timeout() -> None:
         request_timeout=300.0,
     )
 
-    assert settings.extra_args == {"timeout": 300.0}
+    assert settings.extra_args is not None
+    assert settings.extra_args["timeout"].read == 300.0
 
 
 def test_make_model_settings_omits_timeout_when_unset() -> None:
@@ -182,4 +183,5 @@ def test_make_model_settings_timeout_survives_reasoning_resolve() -> None:
         request_timeout=120.0,
     )
 
-    assert settings.extra_args == {"timeout": 120.0}
+    assert settings.extra_args is not None
+    assert settings.extra_args["timeout"].read == 120.0
diff --git a/tests/test_models.py b/tests/test_models.py
index cda1c6b6..df1231c4 100644
--- a/tests/test_models.py
+++ b/tests/test_models.py
@@ -17,8 +17,18 @@ def test_recommended_models_are_accepted(model_name: str) -> None:
 
 
 def test_request_timeout_extra_args_positive() -> None:
-    assert request_timeout_extra_args(300) == {"timeout": 300.0}
-    assert request_timeout_extra_args(120.5) == {"timeout": 120.5}
+    args = request_timeout_extra_args(300)
+    assert args is not None
+    timeout = args["timeout"]
+    # read (inactivity) carries the configured value; connect is capped so a dead
+    # endpoint fails fast rather than waiting the full read window.
+    assert timeout.read == 300.0
+    assert timeout.connect == 30.0
+
+    short = request_timeout_extra_args(10)
+    assert short is not None
+    assert short["timeout"].read == 10.0
+    assert short["timeout"].connect == 10.0
 
 
 @pytest.mark.parametrize("value", [None, 0, -1])