diff --git a/strix/config/models.py b/strix/config/models.py index b6e6b0f0..dceb6654 100644 --- a/strix/config/models.py +++ b/strix/config/models.py @@ -277,6 +277,51 @@ def _configure_litellm_compatibility() -> None: litellm.suppress_debug_info = True _register_litellm_cost_callback() + _install_openrouter_stream_cost_capture() + + +def _install_openrouter_stream_cost_capture() -> None: + """Preserve OpenRouter's per-stream cost, which LiteLLM drops when streaming. + + OpenRouter reports the real charge in ``usage.cost`` of the final stream + chunk, but LiteLLM rebuilds streamed responses from token-only fields and + discards it (its non-streamed path stashes the cost in hidden params; the + streaming path does not). Every scan streams, so without this the cost is + lost and Strix falls back to a cost-map estimate that is missing entirely + for new models (e.g. kimi-k3), reporting $0. Subclass the OpenRouter + streaming handler to record the cost keyed by response id so the cost + callback can recover the exact charge for the matching rebuilt response. + """ + import litellm + from litellm.llms.openrouter.chat.transformation import ( + OpenRouterChatCompletionStreamingHandler, + OpenrouterConfig, + ) + + from strix.report.state import remember_streamed_openrouter_cost + + class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler): + def chunk_parser(self, chunk: dict[str, Any]) -> Any: + stream = super().chunk_parser(chunk) + remember_streamed_openrouter_cost( + chunk.get("id") or getattr(stream, "id", None), chunk.get("usage") + ) + return stream + + class _StrixOpenrouterConfig(OpenrouterConfig): + def get_model_response_iterator( + self, streaming_response: Any, sync_stream: bool, json_mode: bool | None = False + ) -> Any: + return _StrixOpenRouterStreamingHandler( + streaming_response=streaming_response, + sync_stream=sync_stream, + json_mode=json_mode, + ) + + # LiteLLM's provider-config factory reads litellm.OpenrouterConfig at call + # time, so overriding the attribute is enough for the subclass to take + # effect. (type: ignore — mypy rejects reassigning a class attribute.) + litellm.OpenrouterConfig = _StrixOpenrouterConfig # type: ignore[misc] _OPENROUTER_ATTRIBUTION_HEADERS = { diff --git a/strix/report/state.py b/strix/report/state.py index 1475ccf3..501673ce 100644 --- a/strix/report/state.py +++ b/strix/report/state.py @@ -1,6 +1,8 @@ import json import logging import subprocess +import threading +from collections import OrderedDict from collections.abc import Callable from datetime import UTC, datetime from importlib.metadata import PackageNotFoundError, version @@ -507,6 +509,63 @@ class ReportState: self._sync_llm_usage_record() +# LiteLLM rebuilds streamed responses from token-only chunks and drops the +# provider-reported ``usage.cost`` that OpenRouter sends in its final stream +# chunk (unlike the non-streamed path, which stashes it in hidden params). Since +# every scan streams, that cost never reaches the callback below. The OpenRouter +# streaming handler (see strix.config.models) stashes the cost here keyed by the +# response id so the callback can recover the exact charge for the matching +# rebuilt response. +_STREAMED_OPENROUTER_COST_LIMIT = 4096 +_streamed_openrouter_costs: OrderedDict[str, float] = OrderedDict() +_streamed_openrouter_costs_lock = threading.Lock() + + +def openrouter_stream_cost(usage: Any) -> float | None: + """Total OpenRouter-reported cost from a raw stream ``usage`` block, or None. + + Non-BYOK responses bill everything to ``usage.cost``. BYOK responses put the + OpenRouter fee in ``usage.cost`` (often 0) and the provider charge in + ``usage.cost_details.upstream_inference_cost``, so BYOK totals sum the two. + """ + if not isinstance(usage, dict): + return None + total = 0.0 + cost = usage.get("cost") + if isinstance(cost, int | float) and cost > 0: + total += float(cost) + if bool(usage.get("is_byok")): + details = usage.get("cost_details") + upstream = details.get("upstream_inference_cost") if isinstance(details, dict) else None + if isinstance(upstream, int | float) and upstream > 0: + total += float(upstream) + return total if total > 0 else None + + +def remember_streamed_openrouter_cost(response_id: Any, usage: Any) -> None: + """Record an OpenRouter stream's reported cost so the cost callback can read it.""" + if not isinstance(response_id, str) or not response_id: + return + cost = openrouter_stream_cost(usage) + if cost is None: + return + with _streamed_openrouter_costs_lock: + _streamed_openrouter_costs[response_id] = cost + _streamed_openrouter_costs.move_to_end(response_id) + while len(_streamed_openrouter_costs) > _STREAMED_OPENROUTER_COST_LIMIT: + _streamed_openrouter_costs.popitem(last=False) + + +def _take_streamed_openrouter_cost(completion_response: Any) -> float | None: + response_id = getattr(completion_response, "id", None) + if response_id is None and isinstance(completion_response, dict): + response_id = cast("dict[str, Any]", completion_response).get("id") + if not isinstance(response_id, str) or not response_id: + return None + with _streamed_openrouter_costs_lock: + return _streamed_openrouter_costs.pop(response_id, None) + + def litellm_cost_callback( kwargs: Any, completion_response: Any, @@ -541,6 +600,11 @@ def litellm_cost_callback( if cost is None: cost = _usage_reported_cost(completion_response) + # Recover the exact OpenRouter cost the streaming handler stashed for this + # response — LiteLLM drops it from streamed usage, so nothing above sees it. + if cost is None: + cost = _take_streamed_openrouter_cost(completion_response) + if cost is None: cost = _estimate_response_cost(kwargs, completion_response) diff --git a/tests/test_cost_tracking.py b/tests/test_cost_tracking.py index 543d3fd5..a4039c52 100644 --- a/tests/test_cost_tracking.py +++ b/tests/test_cost_tracking.py @@ -8,8 +8,21 @@ from unittest.mock import MagicMock, patch import litellm import pytest -from strix.config.models import _configure_litellm_compatibility -from strix.report.state import litellm_cost_callback +import strix.report.state as state_module +from strix.config.models import ( + _configure_litellm_compatibility, + _install_openrouter_stream_cost_capture, +) +from strix.report.state import ( + litellm_cost_callback, + openrouter_stream_cost, + remember_streamed_openrouter_cost, +) + + +@pytest.fixture(autouse=True) +def _clear_streamed_costs() -> None: + state_module._streamed_openrouter_costs.clear() def test_streaming_logging_stays_enabled_for_cost_callback() -> None: @@ -151,3 +164,83 @@ def test_cost_callback_records_nothing_when_no_cost_available() -> None: litellm_cost_callback({"response_cost": None, "model": "x/y"}, response) report_state.record_observed_llm_cost.assert_not_called() + + +def test_openrouter_stream_cost_extracts_plain_and_byok_totals() -> None: + assert openrouter_stream_cost({"cost": 0.003168}) == pytest.approx(0.003168) + assert openrouter_stream_cost( + {"cost": 0.01, "is_byok": True, "cost_details": {"upstream_inference_cost": 0.2}} + ) == pytest.approx(0.21) + # Upstream cost is only added for BYOK responses. + assert openrouter_stream_cost( + {"cost": 0.05, "is_byok": False, "cost_details": {"upstream_inference_cost": 0.04}} + ) == pytest.approx(0.05) + assert openrouter_stream_cost({"prompt_tokens": 10}) is None + assert openrouter_stream_cost(None) is None + + +def test_cost_callback_recovers_streamed_openrouter_cost_by_response_id() -> None: + report_state = MagicMock() + remember_streamed_openrouter_cost("gen-abc", {"cost": 0.42}) + # LiteLLM strips cost from the rebuilt streamed usage; only the id survives. + response = SimpleNamespace(id="gen-abc", usage=SimpleNamespace(cost=None), _hidden_params={}) + + with ( + patch("strix.report.state.get_global_report_state", return_value=report_state), + patch("litellm.completion_cost", side_effect=ValueError("unknown model")), + ): + litellm_cost_callback({"response_cost": None, "model": "moonshotai/kimi-k3"}, response) + + report_state.record_observed_llm_cost.assert_called_once_with(0.42) + # The entry is consumed so a later response cannot double-count it. + assert "gen-abc" not in state_module._streamed_openrouter_costs + + +def test_streamed_openrouter_cost_prefers_provider_report_over_estimate() -> None: + report_state = MagicMock() + remember_streamed_openrouter_cost("gen-xyz", {"cost": 0.9}) + response = SimpleNamespace( + id="gen-xyz", + usage=SimpleNamespace(prompt_tokens=10, completion_tokens=5, total_tokens=15), + _hidden_params={}, + ) + + with ( + patch("strix.report.state.get_global_report_state", return_value=report_state), + patch("litellm.completion_cost", return_value=0.1) as estimate, + ): + litellm_cost_callback({"response_cost": None, "model": "moonshotai/kimi-k3"}, response) + + report_state.record_observed_llm_cost.assert_called_once_with(0.9) + estimate.assert_not_called() + + +def test_remember_streamed_openrouter_cost_evicts_oldest_over_limit() -> None: + limit = state_module._STREAMED_OPENROUTER_COST_LIMIT + for i in range(limit + 5): + remember_streamed_openrouter_cost(f"gen-{i}", {"cost": 0.001}) + + assert len(state_module._streamed_openrouter_costs) == limit + assert "gen-0" not in state_module._streamed_openrouter_costs + assert f"gen-{limit + 4}" in state_module._streamed_openrouter_costs + + +def test_openrouter_stream_handler_records_cost() -> None: + _install_openrouter_stream_cost_capture() + handler_cls = ( + litellm.OpenrouterConfig() + .get_model_response_iterator(streaming_response=iter([]), sync_stream=True) + .__class__ + ) + + chunk = { + "id": "gen-stream", + "created": 1, + "model": "moonshotai/kimi-k3", + "choices": [{"index": 0, "delta": {"content": None}}], + "usage": {"prompt_tokens": 89, "completion_tokens": 138, "cost": 0.0035055}, + } + handler = handler_cls(streaming_response=iter([]), sync_stream=True) + handler.chunk_parser(chunk) + + assert state_module._streamed_openrouter_costs["gen-stream"] == pytest.approx(0.0035055)