diff --git a/agent/model_metadata.py b/agent/model_metadata.py
index 4493eae5f1f..01a96a735cd 100644
--- a/agent/model_metadata.py
+++ b/agent/model_metadata.py
@@ -700,7 +700,10 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
                 return _model_metadata_cache
 
     try:
-        response = requests.get(OPENROUTER_MODELS_URL, timeout=10, verify=_resolve_requests_verify())
+        # Tuple (connect, read) — flat timeout=10 means urllib3 can block 10s per
+        # retry stage through proxies that 403 CONNECT, ballooning to minutes
+        # (#46620). 5s connect / 10s read fails fast on unreachable hosts.
+        response = requests.get(OPENROUTER_MODELS_URL, timeout=(5, 10), verify=_resolve_requests_verify())
         response.raise_for_status()
         data = response.json()
 
@@ -778,7 +781,7 @@ def fetch_endpoint_model_metadata(
                 response = requests.get(
                     server_url.rstrip("/") + "/api/v1/models",
                     headers=headers,
-                    timeout=10,
+                    timeout=(5, 10),
                     verify=_resolve_requests_verify(),
                 )
                 response.raise_for_status()
@@ -826,7 +829,7 @@ def fetch_endpoint_model_metadata(
     for candidate in candidates:
         url = candidate.rstrip("/") + "/models"
         try:
-            response = requests.get(url, headers=headers, timeout=10, verify=_resolve_requests_verify())
+            response = requests.get(url, headers=headers, timeout=(5, 10), verify=_resolve_requests_verify())
             response.raise_for_status()
             payload = response.json()
             cache: Dict[str, Dict[str, Any]] = {}
@@ -1406,7 +1409,7 @@ def _query_anthropic_context_length(model: str, base_url: str, api_key: str) ->
             "x-api-key": api_key,
             "anthropic-version": "2023-06-01",
         }
-        resp = requests.get(url, headers=headers, timeout=10, verify=_resolve_requests_verify())
+        resp = requests.get(url, headers=headers, timeout=(5, 10), verify=_resolve_requests_verify())
         if resp.status_code != 200:
             return None
         data = resp.json()
@@ -1473,7 +1476,7 @@ def _fetch_codex_oauth_context_lengths(access_token: str) -> Dict[str, int]:
         resp = requests.get(
             "https://chatgpt.com/backend-api/codex/models?client_version=1.0.0",
             headers={"Authorization": f"Bearer {access_token}"},
-            timeout=10,
+            timeout=(5, 10),
             verify=_resolve_requests_verify(),
         )
         if resp.status_code != 200:
diff --git a/model_tools.py b/model_tools.py
index 0618138aa9a..5c1338f7fe1 100644
--- a/model_tools.py
+++ b/model_tools.py
@@ -552,7 +552,13 @@ def _resolve_active_context_length() -> int:
         if not model_id:
             return 0
         from agent.model_metadata import get_model_context_length
-        return int(get_model_context_length(model_id) or 0)
+        # Honor explicit `model.context_length` in config.yaml — short-circuits
+        # the OpenRouter /models probe at get_model_context_length step 0, so
+        # non-OpenRouter providers don't pay the ~2-3s OpenRouter fetch at every
+        # CLI startup.  See issue #46620.
+        raw_ctx = model_cfg.get("context_length")
+        config_ctx = raw_ctx if isinstance(raw_ctx, int) and raw_ctx > 0 else None
+        return int(get_model_context_length(model_id, config_context_length=config_ctx) or 0)
     except Exception as e:
         logger.debug("Could not resolve active context length: %s", e)
         return 0