feat(llm): retry upstream 4xx errors and ban the provider that keeps failing

LiteLLM does not retry 4xx and OpenRouter does not fall back on a provider's 400 or after the stream starts, so one bad upstream (Relace's "incomplete tool call", InferenceNet's 250k context cap) killed the agent. Every 4xx is now replayed within the existing 5 retries, except guardrail refusals and the gateway's own 401/402/403. Before each replay the upstream that failed is banned for the model: right away on provider_unavailable or when OpenRouter named it in a non-2xx reply, and after STRIX_PROVIDER_4XX_BAN_AFTER (default 2) mid-stream 4xx errors.

RequestLoggingModel attaches the attempt's HttpReply to the exception it raises, so the run cycle can read the upstream provider with request_log.failed_reply(exc).

strix/config/models.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import contextlib
 import inspect
+import json
 import logging
 import os
 import time
@@ -42,7 +43,7 @@ from strix.config import codex
 from strix.config.loader import load_settings
 from strix.config.tool_call_ids import TurnCallIdRewriter, dedupe_input
 from strix.config.tool_call_limits import TurnToolCallLimiter
-from strix.llm import request_log
+from strix.llm import provider_bans, request_log
 
 
 if TYPE_CHECKING:
@@ -745,6 +746,11 @@ def _install_openrouter_stream_cost_capture() -> None:
 
     class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler):
         def chunk_parser(self, chunk: dict[str, Any]) -> Any:
+            # Before parsing: LiteLLM raises on an error chunk and drops its
+            # top-level ``provider``.
+            request_log.record_upstream_provider(
+                chunk.get("provider"), _openrouter_error_type(chunk.get("error"))
+            )
             stream = super().chunk_parser(chunk)
             usage = chunk.get("usage")
             response_id = chunk.get("id") or getattr(stream, "id", None)
@@ -763,12 +769,24 @@ def _install_openrouter_stream_cost_capture() -> None:
                 json_mode=json_mode,
             )
 
+        def get_error_class(self, error_message: str, status_code: int, headers: Any) -> Any:
+            # A non-2xx reply names the provider in ``error.metadata``.
+            with contextlib.suppress(Exception):
+                error = json.loads(error_message)["error"]
+                request_log.record_upstream_provider(
+                    error["metadata"].get("provider_name"),
+                    _openrouter_error_type(error),
+                    rejected=True,
+                )
+            return super().get_error_class(error_message, status_code, headers)
+
         def transform_response(self, *args: Any, **kwargs: Any) -> Any:
             # Non-streamed replies (LLM_DISABLE_STREAMING) skip the chunk parser.
             response = super().transform_response(*args, **kwargs)
             raw_response = kwargs.get("raw_response", args[1] if len(args) > 1 else None)
             with contextlib.suppress(Exception):
                 body = raw_response.json()  # type: ignore[union-attr]
+                request_log.record_upstream_provider(body.get("provider"))
                 if body.get("usage"):
                     record_openrouter_provider(body.get("provider"), body["usage"])
             return response
@@ -781,6 +799,12 @@ def _install_openrouter_stream_cost_capture() -> None:
             if agent_id and load_settings().llm.openrouter_sticky_sessions:
                 session_id = _OPENROUTER_SESSION_IDS.setdefault(agent_id, str(uuid.uuid4()))
                 body.setdefault("session_id", session_id)
+            model = kwargs.get("model", args[0] if args else "")
+            if banned := provider_bans.banned_providers(model):
+                # Copy: strix-pro's hook sets ``provider`` to a dict it caches.
+                provider = dict(body.get("provider") or {})
+                provider["ignore"] = sorted({*provider.get("ignore", ()), *banned})
+                body["provider"] = provider
             return body
 
     # LiteLLM's provider-config factory reads litellm.OpenrouterConfig at call
@@ -789,6 +813,14 @@ def _install_openrouter_stream_cost_capture() -> None:
     litellm.OpenrouterConfig = _StrixOpenrouterConfig  # type: ignore[misc]
 
 
+def _openrouter_error_type(error: object) -> object:
+    if isinstance(error, dict):
+        metadata = error.get("metadata")
+        if isinstance(metadata, dict):
+            return metadata.get("error_type")
+    return None
+
+
 OPENROUTER_ATTRIBUTION_HEADERS = {
     "HTTP-Referer": "https://strix.ai",
     "X-Title": "Strix",

strix/config/settings.py

@@ -72,6 +72,8 @@ class LlmSettings(BaseSettings):
         default=False,
         alias="STRIX_OPENROUTER_STICKY_SESSIONS",
     )
+    openrouter_provider_bans: bool = Field(default=True, alias="STRIX_OPENROUTER_PROVIDER_BANS")
+    openrouter_ban_after_4xx: int = Field(default=2, ge=1, alias="STRIX_OPENROUTER_BAN_AFTER_4XX")
     disable_streaming: bool = Field(
         default=False,
         alias="LLM_DISABLE_STREAMING",

strix/core/execution.py

@@ -6,6 +6,7 @@ import asyncio
 import contextlib
 import logging
 import uuid
+from collections import Counter
 from collections.abc import Callable
 from functools import cache
 from typing import TYPE_CHECKING, Any, cast
@@ -20,6 +21,7 @@ from openai import (
 )
 
 from strix.config import codex
+from strix.config.loader import load_settings
 from strix.core.hooks import (
     BudgetExceededError,
     BudgetPausedError,
@@ -33,7 +35,7 @@ from strix.core.sessions import (
     seed_initial_input,
     strip_all_images_from_session,
 )
-from strix.llm import request_log
+from strix.llm import provider_bans, request_log
 from strix.llm.compaction import is_context_overflow, maybe_compact
 
 
@@ -123,6 +125,7 @@ async def _compact_session(
 _MAX_TRANSIENT_MODEL_RETRIES = 5
 _TRANSIENT_MODEL_RETRY_BASE_DELAY_S = 2.0
 _TRANSIENT_MODEL_RETRY_MAX_DELAY_S = 90.0
+_provider_4xx_counts: Counter[tuple[str, str]] = Counter()
 
 
 def _model_error_status_code(exc: BaseException) -> int | None:
@@ -138,13 +141,41 @@ def _is_transient_model_error(exc: BaseException) -> bool:
     ):
         return True
     code = _model_error_status_code(exc)
+    if code in (401, 402, 403):
+        # Auth, credit and policy errors from the gateway itself (e.g. OpenRouter's
+        # "User not found" 401) name no upstream provider; retrying cannot help.
+        reply = request_log.failed_reply(exc)
+        return reply is not None and reply.upstream_provider is not None
     if code is not None:
         import litellm
 
-        return bool(litellm._should_retry(code))
+        return 400 <= code < 500 or bool(litellm._should_retry(code))
     return isinstance(exc, APIError)
 
 
+async def _maybe_ban_provider(exc: BaseException, run_config: RunConfig) -> None:
+    reply = request_log.failed_reply(exc)
+    if reply is None or reply.upstream_provider is None:
+        return
+    model = _run_config_model(run_config)
+    if model is None:
+        return
+    provider = reply.upstream_provider
+    code = _model_error_status_code(exc)
+    if reply.upstream_error_type == "provider_unavailable":
+        reason = "provider_unavailable"
+    elif reply.upstream_rejected:
+        reason = f"upstream_rejected_{code}"
+    elif code is not None and 400 <= code < 500:
+        _provider_4xx_counts[model, provider] += 1
+        if _provider_4xx_counts[model, provider] < load_settings().llm.openrouter_ban_after_4xx:
+            return
+        reason = "repeated_4xx"
+    else:
+        return
+    await asyncio.to_thread(provider_bans.ban_provider, model, provider, reason)
+
+
 def _transient_model_retry_delay(attempt: int) -> float:
     delay = _TRANSIENT_MODEL_RETRY_BASE_DELAY_S * float(2 ** (attempt - 1))
     return min(delay, _TRANSIENT_MODEL_RETRY_MAX_DELAY_S)
@@ -844,6 +875,7 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                     input_data = []
                     continue
             if model_retries < _MAX_TRANSIENT_MODEL_RETRIES and _is_transient_model_error(exc):
+                await _maybe_ban_provider(exc, run_config)
                 model_retries += 1
                 delay = _transient_model_retry_delay(model_retries)
                 logger.warning(

strix/llm/provider_bans.py

@@ -0,0 +1,74 @@
+"""Per-run bans on the OpenRouter upstream providers behind a model."""
+
+from __future__ import annotations
+
+import logging
+import threading
+
+import httpx
+
+from strix.config.loader import load_settings
+
+
+logger = logging.getLogger(__name__)
+
+_ENDPOINTS_URL = "https://openrouter.ai/api/v1/models/{model}/endpoints"
+
+_LOCK = threading.Lock()
+# OpenRouter reports display names ("InferenceNet"); provider.ignore takes slugs ("inference-net").
+_SLUGS: dict[str, dict[str, str]] = {}
+_BANNED: dict[str, set[str]] = {}
+
+
+def _model_id(model: str) -> str:
+    model = model.strip()
+    for prefix in ("litellm/", "openrouter/"):
+        model = model.removeprefix(prefix)
+    return model
+
+
+def _provider_slugs(model: str) -> dict[str, str]:
+    if model not in _SLUGS:
+        response = httpx.get(_ENDPOINTS_URL.format(model=model), timeout=5)
+        response.raise_for_status()
+        slugs: dict[str, str] = {}
+        for endpoint in response.json()["data"]["endpoints"]:
+            if not (tag := endpoint.get("tag")):
+                continue
+            # A tag is the provider slug plus an optional variant, e.g. "inference-net/fp4".
+            slug = str(tag).split("/")[0]
+            slugs[slug] = slug
+            slugs[str(endpoint.get("provider_name") or slug).lower()] = slug
+        _SLUGS[model] = slugs
+    return _SLUGS[model]
+
+
+def ban_provider(model: str, provider: str, reason: str) -> None:
+    """Ban for the rest of the run; ``provider`` may be a display name or slug."""
+    if not load_settings().llm.openrouter_provider_bans:
+        return
+    model = _model_id(model)
+    try:
+        slugs = _provider_slugs(model)
+        slug = slugs[provider.strip().lower()]
+    except Exception as exc:  # noqa: BLE001
+        logger.warning(
+            "provider_ban_skipped upstream=%s model=%s reason=%s: %r", provider, model, reason, exc
+        )
+        return
+    with _LOCK:
+        banned = _BANNED.setdefault(model, set())
+        if slug in banned:
+            return
+        banned.add(slug)
+        logger.warning(
+            "provider_banned upstream=%s slug=%s model=%s reason=%s", provider, slug, model, reason
+        )
+        if banned >= set(slugs.values()):
+            banned.clear()
+            logger.warning("provider_bans_reset model=%s: every provider was banned", model)
+
+
+def banned_providers(model: str) -> list[str]:
+    with _LOCK:
+        return sorted(_BANNED.get(_model_id(model), ()))

strix/llm/request_log.py

@@ -159,9 +159,43 @@ class HttpReply:
 
     status_code: int | None = None
     headers: dict[str, str] | None = None
+    upstream_provider: str | None = None
+    upstream_error_type: str | None = None
+    # The provider rejected the request (named in the gateway's error body).
+    upstream_rejected: bool = False
 
 
 _http_reply: ContextVar[HttpReply | None] = ContextVar("strix_llm_http_reply", default=None)
+_FAILED_REPLY_ATTR = "_strix_http_reply"
+
+
+def record_upstream_provider(
+    provider: object, error_type: object = None, *, rejected: bool = False
+) -> None:
+    """Remember which provider behind a gateway served the attempt in flight."""
+    reply = _http_reply.get()
+    if reply is None:
+        return
+    if isinstance(provider, str) and provider:
+        reply.upstream_provider = provider
+        reply.upstream_rejected = rejected
+    if isinstance(error_type, str) and error_type:
+        reply.upstream_error_type = error_type
+
+
+def _upstream_details(reply: HttpReply | None) -> list[tuple[str, object]]:
+    if reply is None:
+        return []
+    return [
+        ("upstream_provider", reply.upstream_provider),
+        ("upstream_error_type", reply.upstream_error_type),
+    ]
+
+
+def failed_reply(exc: BaseException) -> HttpReply | None:
+    """The reply of the attempt that raised ``exc``."""
+    reply = getattr(exc, _FAILED_REPLY_ATTR, None)
+    return reply if isinstance(reply, HttpReply) else None
 
 
 async def record_http_reply(response: Response) -> None:
@@ -772,6 +806,7 @@ def _litellm_success(
             ("request", _litellm_request_details(kwargs)),
             ("response", _litellm_response_details(response)),
             ("litellm", _litellm_hidden_details(hidden, slo)),
+            *_upstream_details(_http_reply.get()),
         ),
     )
 
@@ -821,6 +856,7 @@ def _litellm_failure(
             ("request", _litellm_request_details(kwargs)),
             ("error", error or None),
             ("litellm", _litellm_hidden_details(hidden, slo)),
+            *_upstream_details(_http_reply.get()),
         ),
     )
 
@@ -894,15 +930,18 @@ def _observe_sdk_shared_http_client() -> None:
 
 
 def _log_line_sink(event: LlmRequestEvent) -> None:
+    details = event.details or {}
     level = logging.DEBUG if event.outcome == "success" else logging.WARNING
     logger.log(
         level,
-        "llm_request route=%s provider=%s model=%s host=%s outcome=%s status=%s "
-        "request_id=%s response_id=%s stream=%s duration_ms=%d ttft_ms=%s "
+        "llm_request route=%s provider=%s upstream=%s upstream_error=%s model=%s host=%s "
+        "outcome=%s status=%s request_id=%s response_id=%s stream=%s duration_ms=%d ttft_ms=%s "
         "req_bytes=%s res_bytes=%s finish=%s "
         "in=%s out=%s cached=%s cost=%s agent=%s attempt=%d%s",
         event.route,
         event.provider or "-",
+        details.get("upstream_provider") or "-",
+        details.get("upstream_error_type") or "-",
         event.model,
         event.api_host or "-",
         event.outcome,
@@ -1070,6 +1109,7 @@ class RequestLoggingModel(Model):
                 details=merge_details(
                     ("request", request.details),
                     ("response", _openai_response_details(response, raw_response)),
+                    *_upstream_details(reply),
                 ),
             )
         status, request_id = _openai_error_fields(exc)
@@ -1089,6 +1129,7 @@ class RequestLoggingModel(Model):
             details=merge_details(
                 ("request", request.details),
                 ("error", _exception_details(exc)),
+                *_upstream_details(reply),
             ),
         )
 
@@ -1178,6 +1219,7 @@ class RequestLoggingModel(Model):
                 prompt=prompt,
             )
         except BaseException as exc:
+            setattr(exc, _FAILED_REPLY_ATTR, reply)
             if self._should_emit(exc):
                 emit(
                     self._event(
@@ -1272,6 +1314,7 @@ class RequestLoggingModel(Model):
                     raw_response = event.response
                 yield event
         except BaseException as exc:
+            setattr(exc, _FAILED_REPLY_ATTR, reply)
             if self._should_emit(exc):
                 emit(
                     self._event(

tests/test_cost_tracking.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 import uuid
 from types import SimpleNamespace
 from typing import TYPE_CHECKING, Any
@@ -389,3 +390,100 @@ def test_openrouter_request_carries_agent_session_id() -> None:
             assert body()["session_id"] == session_id
     finally:
         request_log.reset_call_context(token)
+
+
+def test_openrouter_request_ignores_banned_providers() -> None:
+    _install_openrouter_stream_cost_capture()
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="moonshotai/kimi-k3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    messages: list[AllMessageValues] = [{"role": "user", "content": "hi"}]
+    pro_preferences = {"order": ["novita", "relace"], "allow_fallbacks": True, "ignore": ["a"]}
+    with (
+        patch(
+            "litellm.llms.openrouter.chat.transformation.OpenrouterConfig.transform_request",
+            return_value={"provider": pro_preferences},
+        ),
+        patch("strix.llm.provider_bans.banned_providers", return_value=["relace"]),
+    ):
+        body = config.transform_request(
+            model="moonshotai/kimi-k3",
+            messages=messages,
+            optional_params={},
+            litellm_params={},
+            headers={},
+        )
+
+    assert body["provider"] == {
+        "order": ["novita", "relace"],
+        "allow_fallbacks": True,
+        "ignore": ["a", "relace"],
+    }
+    assert pro_preferences["ignore"] == ["a"]
+
+
+def _openrouter_config() -> Any:
+    _install_openrouter_stream_cost_capture()
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="z-ai/glm-5.3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    return config
+
+
+def test_openrouter_records_upstream_of_stream_and_mid_stream_error() -> None:
+    handler = _openrouter_config().get_model_response_iterator(
+        streaming_response=iter([]), sync_stream=True
+    )
+    reply = request_log.HttpReply()
+    token = request_log._http_reply.set(reply)
+    try:
+        handler.chunk_parser(
+            {
+                "id": "gen-a",
+                "created": 1,
+                "model": "z-ai/glm-5.3",
+                "provider": "Relace",
+                "choices": [{"index": 0, "delta": {"content": "x"}}],
+            }
+        )
+        assert reply.upstream_provider == "Relace"
+        assert reply.upstream_error_type is None
+        with pytest.raises(Exception, match="incomplete tool call"):
+            handler.chunk_parser(
+                {
+                    "id": "gen-a",
+                    "provider": "Relace",
+                    "error": {
+                        "code": 400,
+                        "message": "Generation stopped with an incomplete tool call.",
+                        "metadata": {"error_type": "invalid_request"},
+                    },
+                }
+            )
+    finally:
+        request_log._http_reply.reset(token)
+    assert reply.upstream_provider == "Relace"
+    assert reply.upstream_error_type == "invalid_request"
+    assert reply.upstream_rejected is False
+
+
+def test_openrouter_records_upstream_of_pre_stream_error() -> None:
+    body = {
+        "error": {
+            "message": "Provider returned error",
+            "code": 400,
+            "metadata": {"raw": "...", "provider_name": "InferenceNet"},
+        }
+    }
+    reply = request_log.HttpReply()
+    token = request_log._http_reply.set(reply)
+    try:
+        _openrouter_config().get_error_class(json.dumps(body), 400, {})
+        _openrouter_config().get_error_class("<html>502 Bad Gateway</html>", 502, {})
+    finally:
+        request_log._http_reply.reset(token)
+    assert reply.upstream_provider == "InferenceNet"
+    assert reply.upstream_error_type is None
+    assert reply.upstream_rejected is True

tests/test_execution_transient_retry.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
-from typing import Any, cast
+from collections import Counter
+from typing import Any
 
 import httpx
 import pytest
@@ -18,6 +19,7 @@ from openai import (
 from strix.config import codex
 from strix.core import execution
 from strix.core.agents import AgentCoordinator
+from strix.llm import provider_bans, request_log
 
 
 def _request() -> httpx.Request:
@@ -79,15 +81,74 @@ def test_content_guardrail_is_not_retried() -> None:
     assert execution._is_transient_model_error(guardrail) is False
 
 
-def test_client_errors_are_not_transient() -> None:
+def _upstream_error(
+    status: int, provider: str, *, error_type: str | None = None, rejected: bool = False
+) -> APIStatusError:
+    exc = _status_error(status)
+    reply = request_log.HttpReply(
+        upstream_provider=provider, upstream_error_type=error_type, upstream_rejected=rejected
+    )
+    setattr(exc, request_log._FAILED_REPLY_ATTR, reply)
+    return exc
+
+
+def test_client_errors_are_transient() -> None:
     bad_request = BadRequestError(
         "bad", response=httpx.Response(400, request=_request()), body=None
     )
-    assert execution._is_transient_model_error(bad_request) is False
-    assert execution._is_transient_model_error(_status_error(404)) is False
+    assert execution._is_transient_model_error(bad_request) is True
+    assert execution._is_transient_model_error(_status_error(404)) is True
     assert execution._is_transient_model_error(ValueError("nope")) is False
 
 
+def test_gateway_auth_credit_and_policy_errors_are_not_transient() -> None:
+    for status in (401, 402, 403):
+        assert execution._is_transient_model_error(_status_error(status)) is False
+    assert execution._is_transient_model_error(_upstream_error(401, "Relace")) is True
+
+
+@pytest.fixture
+def bans(monkeypatch: pytest.MonkeyPatch) -> list[tuple[str, str, str]]:
+    calls: list[tuple[str, str, str]] = []
+
+    def _ban(model: str, provider: str, reason: str) -> None:
+        calls.append((model, provider, reason))
+
+    monkeypatch.setattr(provider_bans, "ban_provider", _ban)
+    monkeypatch.setattr(execution, "_provider_4xx_counts", Counter())
+    return calls
+
+
+_OPENROUTER_RUN_CONFIG = RunConfig(model="openrouter/z-ai/glm-5.3")
+
+
+@pytest.mark.asyncio
+async def test_bans_upstream_right_away_on_provider_unavailable_or_rejection(
+    bans: list[tuple[str, str, str]],
+) -> None:
+    unavailable = _upstream_error(502, "Relace", error_type="provider_unavailable")
+    await execution._maybe_ban_provider(unavailable, _OPENROUTER_RUN_CONFIG)
+    rejected = _upstream_error(400, "InferenceNet", rejected=True)
+    await execution._maybe_ban_provider(rejected, _OPENROUTER_RUN_CONFIG)
+    assert bans == [
+        ("openrouter/z-ai/glm-5.3", "Relace", "provider_unavailable"),
+        ("openrouter/z-ai/glm-5.3", "InferenceNet", "upstream_rejected_400"),
+    ]
+
+
+@pytest.mark.asyncio
+async def test_bans_upstream_after_repeated_mid_stream_4xx(
+    bans: list[tuple[str, str, str]],
+) -> None:
+    for _ in range(2):
+        await execution._maybe_ban_provider(_upstream_error(400, "Relace"), _OPENROUTER_RUN_CONFIG)
+        await execution._maybe_ban_provider(
+            _upstream_error(504, "Together"), _OPENROUTER_RUN_CONFIG
+        )
+    await execution._maybe_ban_provider(_status_error(400), _OPENROUTER_RUN_CONFIG)
+    assert bans == [("openrouter/z-ai/glm-5.3", "Relace", "repeated_4xx")]
+
+
 class _FakeStream:
     def __init__(self, exc: BaseException | None = None) -> None:
         self._exc = exc
@@ -128,7 +189,7 @@ async def _run_once(
         coordinator,
         "root",
         input_data="task",
-        run_config=cast("RunConfig", object()),
+        run_config=RunConfig(),
         context={},
         max_turns=5,
         session=None,
@@ -166,9 +227,6 @@ async def test_run_cycle_gives_up_after_max_retries(
 async def test_run_cycle_does_not_retry_permanent_error(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    bad_request = BadRequestError(
-        "bad", response=httpx.Response(400, request=_request()), body=None
-    )
-    streams = [_FakeStream(exc=bad_request), _FakeStream()]
-    with pytest.raises(BadRequestError):
+    streams = [_FakeStream(exc=_status_error(401)), _FakeStream()]
+    with pytest.raises(APIStatusError):
         await _run_once(monkeypatch, streams)

tests/test_llm_request_log.py

@@ -394,9 +394,30 @@ def test_log_line_sink_formats_without_content(caplog: pytest.LogCaptureFixture)
     assert "request_id=req_line01" in line
     assert "status=400" in line
     assert "provider=anthropic" in line
+    assert "upstream=- upstream_error=-" in line
     assert "SECRET PROMPT" not in line
 
 
+def test_log_line_names_upstream_recorded_during_attempt(
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    exc = _anthropic_error(400, ANTHROPIC_BLOCK_BODY, {})
+    token = request_log._http_reply.set(request_log.HttpReply())
+    try:
+        request_log.record_upstream_provider("Together", "provider_unavailable")
+        event = request_log.event_from_litellm(
+            _anthropic_kwargs(exc), None, None, None, outcome="error"
+        )
+    finally:
+        request_log._http_reply.reset(token)
+    assert event.details is not None
+    assert event.details["upstream_provider"] == "Together"
+    with caplog.at_level(logging.DEBUG, logger="strix.llm.request_log"):
+        request_log._log_line_sink(event)
+    line = caplog.records[-1].getMessage()
+    assert "upstream=Together upstream_error=provider_unavailable" in line
+
+
 # --------------------------------------------------------------------------- #
 # sizes, timing, finish reason, free-form headers and details                  #
 # --------------------------------------------------------------------------- #
@@ -1214,6 +1235,23 @@ async def test_litellm_route_wrapper_only_logs_abandoned_attempts(
     assert event.request_bytes is not None
 
 
+@pytest.mark.asyncio
+async def test_failed_attempt_reply_rides_on_its_exception() -> None:
+    exc = _openai_status_error(400, None)
+    model = request_log.RequestLoggingModel(
+        _Inner(exc=exc),
+        model_name="openrouter/z-ai/glm-5.3",
+        provider="openrouter",
+        base_url=None,
+        route="litellm",
+        abandoned_only=True,
+    )
+    with pytest.raises(APIStatusError):
+        await model.get_response(*_CALL_ARGS, **_CALL_KWARGS)
+    assert isinstance(request_log.failed_reply(exc), request_log.HttpReply)
+    assert request_log.failed_reply(ValueError()) is None
+
+
 @pytest.mark.asyncio
 async def test_stream_closed_early_is_logged_as_abandoned(captured: list[LlmRequestEvent]) -> None:
     created = ResponseCreatedEvent(

tests/test_provider_bans.py

@@ -0,0 +1,75 @@
+"""Tests for per-run OpenRouter provider bans."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+
+from strix.llm import provider_bans
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+
+MODEL = "openrouter/z-ai/glm-5.3"
+ENDPOINTS = [
+    {"provider_name": "Relace", "tag": "relace"},
+    {"provider_name": "InferenceNet", "tag": "inference-net/fp4"},
+    {"provider_name": "Google", "tag": "google-vertex/us-east5"},
+    {"provider_name": "Google", "tag": "google-vertex"},
+]
+
+
+@pytest.fixture(autouse=True)
+def endpoints() -> Iterator[MagicMock]:
+    for state in (provider_bans._SLUGS, provider_bans._BANNED):
+        state.clear()
+    response = httpx.Response(
+        200,
+        json={"data": {"endpoints": ENDPOINTS}},
+        request=httpx.Request("GET", "https://openrouter.ai"),
+    )
+    with patch("strix.llm.provider_bans.httpx.get", return_value=response) as get:
+        yield get
+
+
+def test_ban_maps_provider_names_to_slugs(endpoints: MagicMock) -> None:
+    provider_bans.ban_provider(MODEL, "InferenceNet", "context_length")
+    provider_bans.ban_provider("z-ai/glm-5.3", "google", "provider_unavailable")
+    provider_bans.ban_provider(MODEL, "InferenceNet", "context_length")
+
+    assert provider_bans.banned_providers(MODEL) == ["google-vertex", "inference-net"]
+    assert provider_bans.banned_providers("openrouter/other/model") == []
+    endpoints.assert_called_once_with(
+        "https://openrouter.ai/api/v1/models/z-ai/glm-5.3/endpoints", timeout=5
+    )
+
+
+def test_banning_every_provider_unbans_them_all() -> None:
+    provider_bans.ban_provider(MODEL, "Relace", "incomplete_tool_call")
+    provider_bans.ban_provider(MODEL, "InferenceNet", "context_length")
+    assert provider_bans.banned_providers(MODEL) == ["inference-net", "relace"]
+    provider_bans.ban_provider(MODEL, "Google", "provider_unavailable")
+
+    assert provider_bans.banned_providers(MODEL) == []
+
+
+def test_ban_is_skipped_for_an_unknown_provider_or_failed_lookup(endpoints: MagicMock) -> None:
+    provider_bans.ban_provider(MODEL, "Nobody", "provider_unavailable")
+    endpoints.side_effect = httpx.ConnectError("offline")
+    provider_bans.ban_provider("openrouter/other/model", "Relace", "x")
+
+    assert provider_bans.banned_providers(MODEL) == []
+
+
+def test_bans_can_be_turned_off(endpoints: MagicMock) -> None:
+    with patch("strix.llm.provider_bans.load_settings") as settings:
+        settings.return_value.llm.openrouter_provider_bans = False
+        provider_bans.ban_provider(MODEL, "Relace", "incomplete_tool_call")
+
+    assert provider_bans.banned_providers(MODEL) == []
+    endpoints.assert_not_called()