fix(agent): classify status-less streaming render errors as format_error

LM Studio / llama.cpp raise a bare status-less APIError when chat-template
Jinja rendering fails mid-stream; classify as format_error with fallback
instead of burning same-provider retries. Stage fires only when status_code
is None. Hand-applied from #64834 (supersedes #62673) onto the staged
classifier pipeline.

Fixes #62662

agent/error_classifier.py

@@ -944,11 +944,23 @@ def _by_status(c: _Ctx) -> Optional[Verdict]:
     return _STATUS_HANDLERS[status](c) if status in _STATUS_HANDLERS else default
 
 
+# LM Studio / llama.cpp raise a bare status-less ``APIError`` when chat-template Jinja rendering
+# fails mid-stream (#62662). Deterministic for the request, so fall back instead of retrying.
+_STREAM_RENDER_ERROR_PATTERNS = ("error rendering", "rendering prompt", "jinja template", "jinja render")
+
+
+def _streaming_render_error(c: _Ctx) -> Optional[Verdict]:
+    """Status-less template render failure → format_error with fallback (only when no HTTP status)."""
+    if c.status_code is None and any(p in c.msg for p in _STREAM_RENDER_ERROR_PATTERNS):
+        return _v(_R.format_error, retryable=False, should_fallback=True)
+    return None
+
+
 # Stage order: plugin hooks → the provider's own profile hook → provider-specific special cases →
-# HTTP status → MoA shapes → structured error code → message patterns → SSL → disconnect +
+# status-less stream render errors → HTTP status → MoA shapes → structured error code → message patterns → SSL → disconnect +
 # large session → transport types → unknown (retryable with backoff).
 _STAGES: Sequence[Callable[[_Ctx], Optional[Verdict]]] = (
-    _plugin_verdict, _profile_verdict, _provider_special_cases, _by_status, _moa_special_cases,
+    _plugin_verdict, _profile_verdict, _provider_special_cases, _streaming_render_error, _by_status, _moa_special_cases,
     _by_error_code, _by_message, _by_transport,
 )
 

tests/agent/test_error_classifier.py

@@ -2112,3 +2112,19 @@ class TestAuthErrorNamesOffRouteEndpoint:
         for base_url in ("", "https://api.anthropic.com/v1"):
             result = classify_api_error(e, provider="anthropic", model="claude", base_url=base_url)
             assert result.message == "API keys are not supported by this endpoint.", base_url
+
+
+class TestStreamingRenderFormatError:
+    """Status-less Jinja render failures (LM Studio / llama.cpp) fail over; see #62662."""
+
+    def test_error_rendering_no_status_is_format_error(self):
+        e = MockAPIError("Error rendering prompt with jinja template: ...")
+        result = classify_api_error(e, provider="lm-studio", model="x")
+        assert result.reason == FailoverReason.format_error
+        assert result.retryable is False
+        assert result.should_fallback is True
+
+    def test_render_message_with_status_uses_http_path(self):
+        e = MockAPIError("Error rendering prompt with jinja template: ...", status_code=500)
+        result = classify_api_error(e, provider="lm-studio", model="x")
+        assert result.reason != FailoverReason.format_error