From d6b54ae2b028653efbc8f93feaf71660b4468501 Mon Sep 17 00:00:00 2001 From: Irvallen Saragih Date: Sat, 17 Oct 2026 05:27:47 +0900 Subject: [PATCH 01/26] fix(openai)!: route gpt-transcribe to GPT audio transcription config (#44021) * fix(openai): route gpt-transcribe to GPT audio transcription config * test(openai): move gpt-transcribe tests under tests/unit * fix(openai): bill gpt-transcribe from provider usage seconds when duration is unknown Fall back to the provider-reported usage.seconds when neither a measured file duration nor a response duration is available, so json responses for undecodable uploads are no longer billed at zero. Move the routing test next to the other provider config tests and replace the structure test with a behavior test. * fix(openai): bill provider usage seconds only for gpt transcription models * fix(cost): decide provider-billed transcription seconds from the deployment model, not its pricing id * test(integration): cover gpt-transcribe default request, provider-billed spend, and health probe --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/cost_calculator.py | 18 ++ litellm/utils.py | 2 +- .../test_audio_transcriptions_wire.py | 174 +++++++++++++++++- .../management/test_model_health_check.py | 37 ++++ tests/unit/test_cost_calculator.py | 72 ++++++++ tests/unit/test_utils.py | 27 +++ 6 files changed, 328 insertions(+), 2 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 0e315e97432..05dc2900349 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -86,6 +86,9 @@ from litellm.llms.openai.cost_calculation import ( from litellm.llms.openai.cost_calculation import ( video_output_cost_per_second, ) +from litellm.llms.openai.transcriptions.gpt_transformation import ( + OpenAIGPTAudioTranscriptionConfig, +) from litellm.llms.perplexity.cost_calculator import ( cost_per_token as perplexity_cost_per_token, ) @@ -325,6 +328,15 @@ def _transcription_usage_has_token_details( return (prompt_tokens_val > 0) or (completion_tokens_val > 0) +def _bills_provider_transcription_seconds(model: str | None, custom_llm_provider: str | None) -> bool: + if model is None or custom_llm_provider != LlmProviders.OPENAI.value: + return False + transcription_config: Final = ProviderConfigManager.get_provider_audio_transcription_config( + model=model.removeprefix(f"{LlmProviders.OPENAI.value}/"), provider=LlmProviders.OPENAI + ) + return isinstance(transcription_config, OpenAIGPTAudioTranscriptionConfig) + + OCRPricingField = Literal["ocr_cost_per_page", "ocr_cost_per_credit", "annotation_cost_per_page"] @@ -1739,6 +1751,12 @@ def completion_cost( audio_transcription_file_duration = _hidden.get( "audio_transcription_duration", getattr(completion_response, "duration", 0.0), + ) or ( + completion_response.usage.seconds + if isinstance(completion_response, TranscriptionResponse) + and isinstance(completion_response.usage, TranscriptionUsageDurationObject) + and _bills_provider_transcription_seconds(requested_model, custom_llm_provider) + else 0.0 ) elif call_type in _RERANK_CALL_TYPES: if completion_response is not None and isinstance(completion_response, RerankResponse): diff --git a/litellm/utils.py b/litellm/utils.py index 8aa6eb23c68..5a933d8f2b8 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -9252,7 +9252,7 @@ class ProviderConfigManager: return XAIAudioTranscriptionConfig() elif litellm.LlmProviders.OPENAI == provider: - if "gpt-4o" in model: + if "gpt-4o" in model or (model.startswith("gpt-") and "transcribe" in model): return litellm.OpenAIGPTAudioTranscriptionConfig() else: return litellm.OpenAIWhisperAudioTranscriptionConfig() diff --git a/tests/integration/compatibility/test_audio_transcriptions_wire.py b/tests/integration/compatibility/test_audio_transcriptions_wire.py index 7845e3d2423..02a09bce5fd 100644 --- a/tests/integration/compatibility/test_audio_transcriptions_wire.py +++ b/tests/integration/compatibility/test_audio_transcriptions_wire.py @@ -1,15 +1,18 @@ +import io import json import math import mimetypes import uuid +import wave from email.message import Message from email.parser import BytesParser from email.policy import HTTP +from hashlib import sha256 from typing import Final, cast import httpx import pytest -from integration._support.client import Gateway, eventually +from integration._support.client import JSON_OBJECT, Gateway, eventually, object_value from integration._support.database import read_rows from integration._support.wire import Reply, Request, wire_server from openai import OpenAI @@ -671,3 +674,172 @@ def test_audio_transcription_bills_audio_tokens_at_the_input_rate_when_no_audio_ rel_tol=1e-9, ), response.text assert [(request.method, request.target) for request in wire.drain()] == [("POST", "/v1/audio/transcriptions")] + + +_GPT_TRANSCRIBE_VERBOSE_JSON_REFUSAL: Final = { + "error": { + "message": "response_format 'verbose_json' is not compatible with model 'gpt-transcribe'", + "type": "invalid_request_error", + "param": "response_format", + "code": "unsupported_value", + } +} +_UNDECODABLE_M4A_BYTES: Final = b"\x00\x00\x00\x20ftypM4A \x00\x00\x02\x00M4A isommp42\x00\x00\x00\x08free" +_DEPLOYMENT_COST_PER_SECOND: Final = 0.0003 + + +def _openai_gpt_transcribe_reply(request: Request, reported_seconds: float) -> Reply: + fields: Final = dict(_text_parts(_multipart_parts(request))) + if fields.get("response_format") == "verbose_json": + return Reply(status=400, body=json.dumps(_GPT_TRANSCRIBE_VERBOSE_JSON_REFUSAL).encode()) + return Reply( + body=json.dumps({"text": "hello world", "usage": {"type": "duration", "seconds": reported_seconds}}).encode() + ) + + +def _uploaded_files(request: Request) -> tuple[tuple[str, str, bytes], ...]: + return tuple((name, filename, payload) for name, filename, _, payload in _file_parts(_multipart_parts(request))) + + +def _silent_wav(seconds: float) -> bytes: + output: Final = io.BytesIO() + with wave.open(output, "wb") as wav: + wav.setnchannels(1) + wav.setsampwidth(2) + wav.setframerate(16000) + wav.writeframes(b"\x00\x00" * round(16000 * seconds)) + return output.getvalue() + + +def _advertised_cost_per_second(gateway: Gateway, model: str) -> float: + entries: Final = gateway.get("/v2/model/info", {"model": model})["data"] + assert isinstance(entries, list) and len(entries) == 1, entries + price: Final = object_value(object_value(entries[0])["model_info"])["input_cost_per_second"] + assert isinstance(price, float) and price > 0, entries + return price + + +def _response_cost(response: httpx.Response) -> float: + return float(response.headers["x-litellm-response-cost"]) if "x-litellm-response-cost" in response.headers else 0.0 + + +def _logged_spend(call_id: str) -> float: + rows: Final = eventually( + lambda: read_rows('SELECT spend FROM "LiteLLM_SpendLogs" WHERE request_id = %s', (call_id,)), + lambda values: len(values) == 1, + seconds=70, + ) + return float(str(rows[0]["spend"])) + + +def test_gpt_transcribe_without_response_format_reaches_openai_without_verbose_json(gateway: Gateway) -> None: + upload: Final = _UNDECODABLE_M4A_BYTES + uuid.uuid4().bytes + + with ( + wire_server(lambda request: _openai_gpt_transcribe_reply(request, 2)) as wire, + gateway.scenario() as scenario, + ): + model: Final = scenario.model( + model="openai/gpt-transcribe", api_base=f"{wire.url}/v1", api_key="synthetic-openai-key" + ) + with _sdk(gateway, "/v1", gateway.key) as sdk: + response: Final = sdk.audio.transcriptions.with_raw_response.create( + model=model, file=("a.m4a", upload, "audio/mp4") + ) + requests: Final = wire.drain() + + assert response.status_code == 200, response.text + assert JSON_OBJECT.validate_json(response.content)["text"] == "hello world", response.text + assert [(request.method, request.target) for request in requests] == [("POST", "/v1/audio/transcriptions")] + assert _text_parts(_multipart_parts(requests[0])) == (("model", "gpt-transcribe"),) + assert _uploaded_files(requests[0]) == (("file", "a.m4a", upload),) + + +@pytest.mark.parametrize( + "configured_cost_per_second", + (None, _DEPLOYMENT_COST_PER_SECOND), + ids=("cost_map_price", "deployment_price"), +) +def test_gpt_transcribe_unmeasurable_upload_bills_the_seconds_openai_reports( + gateway: Gateway, configured_cost_per_second: float | None +) -> None: + reported_seconds: Final = 2.0 + upload: Final = _UNDECODABLE_M4A_BYTES + uuid.uuid4().bytes + + with ( + wire_server(lambda request: _openai_gpt_transcribe_reply(request, reported_seconds)) as wire, + gateway.scenario() as scenario, + ): + model: Final = ( + scenario.model( + model="openai/gpt-transcribe", + api_base=f"{wire.url}/v1", + api_key="synthetic-openai-key", + input_cost_per_second=configured_cost_per_second, + ) + if configured_cost_per_second is not None + else scenario.model( + model="openai/gpt-transcribe", api_base=f"{wire.url}/v1", api_key="synthetic-openai-key" + ) + ) + key: Final = scenario.key(models=[model]) + cost_per_second: Final = ( + configured_cost_per_second + if configured_cost_per_second is not None + else _advertised_cost_per_second(gateway, model) + ) + response: Final = gateway.request_multipart( + "/v1/audio/transcriptions", + {"model": model, "response_format": "json"}, + {"file": ("a.m4a", upload, "audio/mp4")}, + key=key, + ) + expected_cost: Final = reported_seconds * cost_per_second + + assert response.status_code == 200, response.text + assert JSON_OBJECT.validate_json(response.content)["text"] == "hello world", response.text + assert math.isclose(_response_cost(response), expected_cost), (_response_cost(response), expected_cost) + assert math.isclose(_logged_spend(response.headers["x-litellm-call-id"]), expected_cost) + key_rows: Final = eventually( + lambda: read_rows( + 'SELECT spend FROM "LiteLLM_VerificationToken" WHERE token = %s', + (sha256(key.encode()).hexdigest(),), + ), + lambda rows: len(rows) == 1 and float(str(rows[0]["spend"])) > 0, + seconds=70, + ) + assert math.isclose(float(str(key_rows[0]["spend"])), expected_cost), key_rows + requests: Final = wire.drain() + + assert [(request.method, request.target) for request in requests] == [("POST", "/v1/audio/transcriptions")] + assert _uploaded_files(requests[0]) == (("file", "a.m4a", upload),) + + +def test_gpt_transcribe_measured_upload_bills_the_measured_length_over_reported_seconds(gateway: Gateway) -> None: + measured_seconds: Final = 1.0 + upload: Final = _silent_wav(measured_seconds) + + with ( + wire_server(lambda request: _openai_gpt_transcribe_reply(request, 7)) as wire, + gateway.scenario() as scenario, + ): + model: Final = scenario.model( + model="openai/gpt-transcribe", + api_base=f"{wire.url}/v1", + api_key="synthetic-openai-key", + input_cost_per_second=_DEPLOYMENT_COST_PER_SECOND, + ) + response: Final = gateway.request_multipart( + "/v1/audio/transcriptions", + {"model": model, "response_format": "json"}, + {"file": ("a.wav", upload, _WAV_CONTENT_TYPE)}, + ) + expected_cost: Final = measured_seconds * _DEPLOYMENT_COST_PER_SECOND + + assert response.status_code == 200, response.text + assert math.isclose(_response_cost(response), expected_cost), (_response_cost(response), expected_cost) + assert math.isclose(_logged_spend(response.headers["x-litellm-call-id"]), expected_cost) + requests: Final = wire.drain() + + assert [(request.method, request.target) for request in requests] == [("POST", "/v1/audio/transcriptions")] + assert _uploaded_files(requests[0]) == (("file", "a.wav", upload),) diff --git a/tests/integration/management/test_model_health_check.py b/tests/integration/management/test_model_health_check.py index 2b9ec658d87..9945d16af07 100644 --- a/tests/integration/management/test_model_health_check.py +++ b/tests/integration/management/test_model_health_check.py @@ -1352,3 +1352,40 @@ def test_evaluation_mode_health_check_of_databricks_ai_decide_resolves_the_mode_ }, ) ] + + +def _openai_gpt_transcribe_health_reply(request: Request) -> Reply: + if _multipart_text_fields(_multipart_parts(request)).get("response_format") == "verbose_json": + return Reply( + status=400, + body=json.dumps( + { + "error": { + "message": "response_format 'verbose_json' is not compatible with model 'gpt-transcribe'", + "type": "invalid_request_error", + "param": "response_format", + "code": "unsupported_value", + } + } + ).encode(), + ) + return Reply(body=json.dumps({"text": "healthy", "usage": {"type": "duration", "seconds": 1}}).encode()) + + +def test_health_check_of_gpt_transcribe_probes_without_verbose_json_and_reports_it_healthy(gateway: Gateway) -> None: + with wire_server(_openai_gpt_transcribe_health_reply) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model="openai/gpt-transcribe", + api_base=f"{wire.url}/v1", + api_key="synthetic-openai-key", + model_info={"mode": "audio_transcription"}, + ) + report: Final = _health_report(gateway, model) + requests: Final = wire.drain() + + assert (report["healthy_count"], report["unhealthy_count"]) == (1, 0), report + assert requests, report + assert { + (request.method, request.target, tuple(sorted(_multipart_text_fields(_multipart_parts(request)).items()))) + for request in requests + } == {("POST", "/v1/audio/transcriptions", (("model", "gpt-transcribe"),))} diff --git a/tests/unit/test_cost_calculator.py b/tests/unit/test_cost_calculator.py index cc2a1d62d52..c61ebb3737c 100644 --- a/tests/unit/test_cost_calculator.py +++ b/tests/unit/test_cost_calculator.py @@ -6426,6 +6426,78 @@ def test_whisper_azure(): assert round(cost, 5) == round(expected_cost, 5) +@pytest.mark.parametrize( + ("provider", "requested_model", "response_extras", "expected_seconds"), + [ + ("openai", "openai/gpt-fake-transcribe", {}, 2.0), + ("openai", "openai/gpt-fake-transcribe", {"_audio_transcription_duration": 1.35}, 1.35), + ("openai", "openai/gpt-fake-transcribe", {"duration": 1.34}, 1.34), + ("openai", "openai/whisper-fake", {}, 0.0), + ("azure", "gpt-fake-transcribe", {}, 0.0), + ], + ids=[ + "undecodable_upload_bills_provider_seconds", + "measured_duration_still_wins", + "provider_duration_still_wins", + "whisper_models_never_bill_provider_seconds", + "azure_never_bills_provider_seconds", + ], +) +def test_transcription_cost_falls_back_to_provider_billed_seconds( + monkeypatch: pytest.MonkeyPatch, + provider: str, + requested_model: str, + response_extras: Mapping[str, float], + expected_seconds: float, +) -> None: + model: Final = requested_model.removeprefix(f"{provider}/") + monkeypatch.setitem( + litellm.model_cost, + model, + {"input_cost_per_second": 0.01, "litellm_provider": provider, "mode": "audio_transcription"}, + ) + transcription: Final = convert_to_model_response_object( + response_object={ + "text": "What is the weather in Paris?", + "usage": {"type": "duration", "seconds": 2}, + **response_extras, + }, + model_response_object=TranscriptionResponse(), + hidden_params={"model": model, "custom_llm_provider": provider}, + response_type="audio_transcription", + ) + + cost: Final = completion_cost( + completion_response=transcription, model=requested_model, call_type="atranscription" + ) + + assert cost == pytest.approx(0.01 * expected_seconds) + + +def test_custom_priced_transcription_deployment_bills_provider_seconds(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setitem( + litellm.model_cost, + "priced-deployment", + {"input_cost_per_second": 0.000075, "litellm_provider": "openai", "mode": "audio_transcription"}, + ) + transcription: Final = convert_to_model_response_object( + response_object={"text": "What is the weather in Paris?", "usage": {"type": "duration", "seconds": 2}}, + model_response_object=TranscriptionResponse(), + hidden_params={"model": "gpt-fake-transcribe", "custom_llm_provider": "openai"}, + response_type="audio_transcription", + ) + + cost: Final = completion_cost( + completion_response=transcription, + model="openai/gpt-fake-transcribe", + call_type="atranscription", + custom_pricing=True, + router_model_id="priced-deployment", + ) + + assert cost == pytest.approx(0.000075 * 2) + + def test_gpt_image_2_azure_cost_tracking(): azure_image_generation_response: Final = { "created": 1758585600, diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py index 98025f3a79a..75cec231f16 100644 --- a/tests/unit/test_utils.py +++ b/tests/unit/test_utils.py @@ -9585,6 +9585,33 @@ def test_get_provider_audio_transcription_config_resolves_for_every_provider() - assert isinstance(configs[LlmProviders.OPENAI], litellm.OpenAIWhisperAudioTranscriptionConfig) +@pytest.mark.parametrize( + ("model", "optional_params", "expected_response_format"), + [ + ("whisper-1", {}, "verbose_json"), + ("my-transcribe-whisper", {}, "verbose_json"), + ("whisper-1", {"response_format": "srt"}, "srt"), + ("gpt-4o-mini-transcribe", {}, None), + ("gpt-transcribe", {}, None), + ("gpt-transcribe", {"response_format": "text"}, "text"), + ], +) +def test_openai_transcription_defaults_verbose_json_only_for_whisper_models( + model: str, optional_params: dict[str, str], expected_response_format: str | None +) -> None: + config: Final = ProviderConfigManager.get_provider_audio_transcription_config( + model=model, provider=LlmProviders.OPENAI + ) + assert config is not None + + request: Final = config.transform_audio_transcription_request( + model=model, audio_file=b"audio", optional_params=dict(optional_params), litellm_params={} + ) + + assert isinstance(request.data, dict) + assert request.data.get("response_format") == expected_response_format + + def test_get_valid_models_from_provider_cache_invalidation(monkeypatch): """ Test that get_valid_models returns the correct models for a given provider From c7b115b56a310c4a0de8ca38295bde680fe7b2a2 Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Fri, 16 Oct 2026 13:44:24 -0700 Subject: [PATCH 02/26] feat(ui): add the 1 hour cache write cost to the add model and model edit forms (#46830) * feat(ui): add the 1 hour cache write cost to the add model and model edit forms Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(e2e): cover both cache write tiers in the add model and edit forms Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(ui): hoist the expected pricing payload out of the inline assertion Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(ui): share the per-token conversion between both cache write tiers Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(ui): derive the touched pricing field type from the pricing field list Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * refactor(ui): collapse the duplicated pricing source keys into one map Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(ui): convert add model pricing once for all selected models and describe the 1 hour fallback Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * test(e2e): type the cache write tier deployment read back and fill the 1 hour field Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * style(e2e): keep the cache write tier spec on the dashboard prettier config Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/types/utils.py | 2 +- .../tests/modelsPage/cacheWriteTiers.spec.ts | 146 ++++++++++++++++++ .../modelsPage/clearCustomPricing.spec.ts | 2 +- .../test_model_management_endpoints.py | 37 +++++ tests/unit/types/test_router.py | 1 + .../src/components/ModelInfoEditForm.tsx | 33 ++-- .../add_model/advanced_settings.test.tsx | 11 ++ .../add_model/advanced_settings.tsx | 20 +++ .../handle_add_model_submit.test.tsx | 55 +++++++ .../add_model/handle_add_model_submit.tsx | 111 +++++++------ .../src/components/model_info_view.test.tsx | 43 +++++- .../src/components/model_info_view.tsx | 21 ++- ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 + 13 files changed, 409 insertions(+), 75 deletions(-) create mode 100644 tests/e2e/ui/tests/modelsPage/cacheWriteTiers.spec.ts diff --git a/litellm/types/utils.py b/litellm/types/utils.py index a97e35a9021..054a0a1c034 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -3900,6 +3900,7 @@ class MirroredPricingParams(LiteLLMBaseModel): output_cost_per_character: float | None = None cache_read_input_token_cost: float | None = None cache_creation_input_token_cost: float | None = None + cache_creation_input_token_cost_above_1hr: float | None = None tiered_pricing: list[dict[str, Any]] | None = None @@ -3927,7 +3928,6 @@ class CustomPricingLiteLLMParams(MirroredPricingParams): input_cost_per_token_priority: float | None = None input_cost_per_token_balanced: float | None = None input_cost_per_token_ultrafast: float | None = None - cache_creation_input_token_cost_above_1hr: float | None = None cache_creation_input_token_cost_above_100k_tokens: float | None = None cache_creation_input_token_cost_above_1hr_above_100k_tokens: float | None = None cache_creation_input_token_cost_above_200k_tokens: float | None = None diff --git a/tests/e2e/ui/tests/modelsPage/cacheWriteTiers.spec.ts b/tests/e2e/ui/tests/modelsPage/cacheWriteTiers.spec.ts new file mode 100644 index 00000000000..7d0f02209cf --- /dev/null +++ b/tests/e2e/ui/tests/modelsPage/cacheWriteTiers.spec.ts @@ -0,0 +1,146 @@ +import { test, expect, type Page as PlaywrightPage } from "@playwright/test"; +import { ADMIN_STORAGE_PATH } from "../../constants"; +import { Role, users } from "../../fixtures/users"; +import { Page } from "../../fixtures/pages"; +import { navigateToPage } from "../../helpers/navigation"; +import { captureRequestBody, readBack } from "../../helpers/roundTrip"; + +const FIVE_MINUTE_PER_MILLION = 3.75; +const ONE_HOUR_PER_MILLION = 7.5; +const perToken = (perMillion: number): number => perMillion / 1_000_000; +const ONE_HOUR_LABEL = /^Cache Write Cost \(1 hour, per 1M tokens\)/; + +type Deployment = { + model_name?: string; + model_info?: { id?: string }; + litellm_params: { + cache_creation_input_token_cost?: number; + cache_creation_input_token_cost_above_1hr?: number; + }; +}; + +const masterKey = users[Role.ProxyAdmin].password; +const auth = { Authorization: `Bearer ${masterKey}` }; + +async function deleteModel(page: PlaywrightPage, id: string) { + await page.request.post("/model/delete", { headers: auth, data: { id } }); +} + +async function readDeployment(page: PlaywrightPage, modelId: string): Promise { + const body = await readBack<{ data: Deployment[] }>(page, `/model/info?litellm_model_id=${modelId}`); + return body.data[0]; +} + +async function selectProvider(page: PlaywrightPage, providerName: string) { + const providerDropdown = page.getByRole("combobox", { name: "Provider", exact: true }); + await providerDropdown.click(); + await providerDropdown.fill(providerName); + await page.getByRole("option").filter({ hasText: providerName }).first().click(); + await expect(providerDropdown).toHaveValue(providerName); +} + +test.describe("Cache write cost tiers", () => { + test.use({ storageState: ADMIN_STORAGE_PATH }); + + let createdModelId: string | null = null; + + test.afterEach(async ({ page }) => { + if (createdModelId) { + await deleteModel(page, createdModelId); + createdModelId = null; + } + }); + + test("the edit form saves the 1 hour rate next to the 5 minute rate and the detail view shows both", async ({ + page, + }) => { + const modelName = `e2e-cache-tiers-edit-${Date.now()}`; + const created = await page.request.post("/model/new", { + headers: auth, + data: { + model_name: modelName, + litellm_params: { + model: "anthropic/claude-sonnet-4-6", + api_key: "sk-e2e-not-used", + cache_creation_input_token_cost: perToken(FIVE_MINUTE_PER_MILLION), + }, + model_info: {}, + }, + }); + expect(created.ok(), `POST /model/new for ${modelName}`).toBe(true); + createdModelId = (await created.json()).model_info?.id; + expect(createdModelId, "model id from /model/new").toBeTruthy(); + + await page.goto("/ui"); + await page.getByText("Models + Endpoints").click(); + const modelIdCell = page.getByTestId(`model-id-${createdModelId}`); + await expect(modelIdCell).toBeVisible({ timeout: 15_000 }); + await modelIdCell.click(); + await expect(page.getByText("Back to Models").first()).toBeVisible({ timeout: 10_000 }); + + await page.getByRole("button", { name: "Edit Settings" }).click(); + const oneHour = page.getByLabel(ONE_HOUR_LABEL); + await oneHour.waitFor({ timeout: 15_000 }); + await oneHour.fill(String(ONE_HOUR_PER_MILLION)); + + const patchBody = await captureRequestBody( + page, + { method: "PATCH", urlIncludes: `/model/${createdModelId}/update` }, + async () => { + await page.getByRole("button", { name: "Save Changes" }).click(); + }, + ); + expect(patchBody.litellm_params.cache_creation_input_token_cost_above_1hr, "1 hour rate sent per token").toBe( + perToken(ONE_HOUR_PER_MILLION), + ); + await expect(page.getByText("Model settings updated successfully")).toBeVisible({ timeout: 10_000 }); + + const stored = await readDeployment(page, createdModelId!); + expect(stored?.litellm_params?.cache_creation_input_token_cost_above_1hr, "1 hour rate persisted").toBe( + perToken(ONE_HOUR_PER_MILLION), + ); + expect(stored?.litellm_params?.cache_creation_input_token_cost, "5 minute rate untouched").toBe( + perToken(FIVE_MINUTE_PER_MILLION), + ); + + await expect(page.getByText(ONE_HOUR_LABEL)).toBeVisible(); + await expect(page.getByText(ONE_HOUR_PER_MILLION.toFixed(4))).toBeVisible({ timeout: 10_000 }); + await expect(page.getByText(FIVE_MINUTE_PER_MILLION.toFixed(4))).toBeVisible(); + }); + + test("Add Model sends both cache write tiers per token", async ({ page }) => { + const publicName = `e2e-cache-tiers-add-${Date.now()}`; + await navigateToPage(page, Page.Models); + await page.getByRole("tab", { name: "Add Model" }).click(); + await selectProvider(page, "Anthropic"); + + await page.getByRole("combobox", { name: "Select models" }).click(); + await page.getByRole("option", { name: "Custom Model Name (Enter below)" }).click(); + await page.keyboard.press("Escape"); + await page.getByPlaceholder("Enter custom model name").fill(publicName); + await page.locator("#api_key").fill("sk-e2e-not-used"); + + await page.getByText("Advanced Settings").click(); + await page.getByRole("switch", { name: "Custom Pricing" }).click(); + await page.getByLabel(/^Input Cost \(per 1M tokens\)/).fill("3"); + await page.getByLabel(/^Output Cost \(per 1M tokens\)/).fill("15"); + await page.getByLabel(/^Cache Write Cost \(per 1M tokens\)/).fill(String(FIVE_MINUTE_PER_MILLION)); + await page.getByLabel(ONE_HOUR_LABEL).fill(String(ONE_HOUR_PER_MILLION)); + + const body = await captureRequestBody(page, { method: "POST", urlIncludes: "/model/new" }, async () => { + await page.getByTestId("add-model-btn").click(); + }); + await expect(page.getByText("created successfully")).toBeVisible({ timeout: 15_000 }); + const listed = await readBack<{ data: Deployment[] }>(page, "/v2/model/info"); + createdModelId = listed.data.find((row) => row.model_name === publicName)?.model_info?.id ?? null; + + expect(body.model_name).toBe(publicName); + expect(body.litellm_params.cache_creation_input_token_cost, "5 minute rate sent per token").toBe( + perToken(FIVE_MINUTE_PER_MILLION), + ); + expect(body.litellm_params.cache_creation_input_token_cost_above_1hr, "1 hour rate sent per token").toBe( + perToken(ONE_HOUR_PER_MILLION), + ); + expect(createdModelId, "the model landed in the database").toBeTruthy(); + }); +}); diff --git a/tests/e2e/ui/tests/modelsPage/clearCustomPricing.spec.ts b/tests/e2e/ui/tests/modelsPage/clearCustomPricing.spec.ts index c532641b238..34a93c7ce14 100644 --- a/tests/e2e/ui/tests/modelsPage/clearCustomPricing.spec.ts +++ b/tests/e2e/ui/tests/modelsPage/clearCustomPricing.spec.ts @@ -87,7 +87,7 @@ test.describe("Clear custom pricing on a deployment", () => { // Both cache fields share the same placeholder ("Defaults to Input Cost if blank"), // so disambiguate via their labels. const cacheReadCost = page.getByLabel(/Cache Read Cost/); - const cacheWriteCost = page.getByLabel(/Cache Write Cost/); + const cacheWriteCost = page.getByLabel("Cache Write Cost (per 1M tokens)", { exact: true }); await inputCost.waitFor({ timeout: 15_000 }); for (const field of [inputCost, outputCost, cacheReadCost, cacheWriteCost]) { await field.click({ clickCount: 3 }); diff --git a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py index beffe454b8d..ad75344afba 100644 --- a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py @@ -3758,6 +3758,43 @@ class TestUpdateDBModelClearPricing: assert "cache_creation_input_token_cost" not in params assert "cache_creation_input_token_cost" not in info + def test_clear_cache_write_1h_cost_removes_from_both_blobs(self): + """The dashboard clears the 1 hour cache-write tier with an explicit null, which only + takes effect for fields in SPECIAL_MODEL_INFO_PARAMS.""" + from litellm.proxy.management_endpoints.model_management_endpoints import ( + update_db_model, + ) + from litellm.types.router import ( + Deployment, + LiteLLM_Params, + ModelInfo, + updateLiteLLMParams, + ) + + db_model = Deployment( + model_name="anthropic/*", + litellm_params=LiteLLM_Params( + model="anthropic/*", + cache_creation_input_token_cost=0.00000375, + cache_creation_input_token_cost_above_1hr=0.000006, + ), + model_info=ModelInfo(id="dep-cache-write-1h-0"), + ) + + result = update_db_model( + db_model=db_model, + updated_patch=updateDeployment( + litellm_params=updateLiteLLMParams(cache_creation_input_token_cost_above_1hr=None) + ), + ) + + params = json.loads(result["litellm_params"]) + info = json.loads(result["model_info"]) + assert "cache_creation_input_token_cost_above_1hr" not in params + assert "cache_creation_input_token_cost_above_1hr" not in info + assert params["cache_creation_input_token_cost"] == 0.00000375 + assert info["cache_creation_input_token_cost"] == 0.00000375 + def test_clear_cache_read_preserves_other_pricing(self): """Clearing cache_read must not touch input/output cost overrides.""" from litellm.proxy.management_endpoints.model_management_endpoints import ( diff --git a/tests/unit/types/test_router.py b/tests/unit/types/test_router.py index 6188bb966ee..8092437fbd3 100644 --- a/tests/unit/types/test_router.py +++ b/tests/unit/types/test_router.py @@ -63,6 +63,7 @@ def test_custom_pricing_params_keeps_every_field_it_had(): "output_cost_per_character", "cache_read_input_token_cost", "cache_creation_input_token_cost", + "cache_creation_input_token_cost_above_1hr", "cost_per_second", "input_cost_per_second", "cache_read_input_token_cost_flex", diff --git a/ui/litellm-dashboard/src/components/ModelInfoEditForm.tsx b/ui/litellm-dashboard/src/components/ModelInfoEditForm.tsx index c7860f35100..125857b9a94 100644 --- a/ui/litellm-dashboard/src/components/ModelInfoEditForm.tsx +++ b/ui/litellm-dashboard/src/components/ModelInfoEditForm.tsx @@ -63,20 +63,21 @@ const PTU_EDIT_FIELDS: PtuEditField[] = [ { name: PTU_END_FIELD, label: "PTU Effective To (UTC)", input: "datetime" }, ]; -export type TouchedPricingField = "input_cost" | "output_cost" | "cache_read_cost" | "cache_write_cost"; - -const PRICING_FIELDS: readonly TouchedPricingField[] = [ +const PRICING_FIELDS = [ "input_cost", "output_cost", "cache_read_cost", "cache_write_cost", + "cache_write_cost_1h", ] as const; +export type TouchedPricingField = (typeof PRICING_FIELDS)[number]; -const COST_SOURCES: Record = { - input_cost: { param: "input_cost_per_token", info: "input_cost_per_token" }, - output_cost: { param: "output_cost_per_token", info: "output_cost_per_token" }, - cache_read_cost: { param: "cache_read_input_token_cost", info: "cache_read_input_token_cost" }, - cache_write_cost: { param: "cache_creation_input_token_cost", info: "cache_creation_input_token_cost" }, +const COST_SOURCES: Record = { + input_cost: "input_cost_per_token", + output_cost: "output_cost_per_token", + cache_read_cost: "cache_read_input_token_cost", + cache_write_cost: "cache_creation_input_token_cost", + cache_write_cost_1h: "cache_creation_input_token_cost_above_1hr", }; export interface ModelEditFormValues { @@ -94,6 +95,7 @@ export interface ModelEditFormValues { output_cost?: string | number | null; cache_read_cost?: string | number | null; cache_write_cost?: string | number | null; + cache_write_cost_1h?: string | number | null; ptu_count?: string | number | null; cost_per_ptu_per_hour?: string | number | null; ptu_effective_from?: Dayjs | null; @@ -131,6 +133,7 @@ const modelEditShape = { output_cost: scalar, cache_read_cost: scalar, cache_write_cost: scalar, + cache_write_cost_1h: scalar, ptu_count: scalar, cost_per_ptu_per_hour: scalar, ptu_effective_from: z.custom().nullish(), @@ -243,6 +246,10 @@ export const toModelEditFormValues = (localModelData: any, isWildcardModel: bool localModelData.litellm_params?.cache_creation_input_token_cost, localModelData.model_info?.cache_creation_input_token_cost, ), + cache_write_cost_1h: perMillionTokens( + localModelData.litellm_params?.cache_creation_input_token_cost_above_1hr, + localModelData.model_info?.cache_creation_input_token_cost_above_1hr, + ), cache_control: localModelData.litellm_params?.cache_control_injection_points ? true : false, cache_control_injection_points: localModelData.litellm_params?.cache_control_injection_points || [], model_access_group: Array.isArray(localModelData.model_info?.access_groups) @@ -271,8 +278,8 @@ export const toModelEditFormValues = (localModelData: any, isWildcardModel: bool }); const displayCost = (localModelData: any, field: TouchedPricingField): string => { - const { param, info } = COST_SOURCES[field]; - const rate = localModelData?.litellm_params?.[param] ?? localModelData?.model_info?.[info]; + const key = COST_SOURCES[field]; + const rate = localModelData?.litellm_params?.[key] ?? localModelData?.model_info?.[key]; return rate != null ? (Number(rate) * 1_000_000).toFixed(4) : "Not Set"; }; @@ -527,6 +534,12 @@ const ModelInfoEditForm: React.FC = ({ "Defaults to Input Cost if blank", "If left blank on save, defaults to Input Cost (backend falls back to input_cost_per_token).", )} + {pricingField( + "cache_write_cost_1h", + "Cache Write Cost (1 hour, per 1M tokens)", + "Defaults to the cost map 1 hour rate, else Cache Write Cost", + "Rate for prompt cache writes with a 1 hour TTL. If left blank on save, known models use their cost map 1 hour rate, otherwise Cache Write Cost.", + )} {textField("api_base", "API Base", "Enter API base", localModelData.litellm_params?.api_base)} {textField( diff --git a/ui/litellm-dashboard/src/components/add_model/advanced_settings.test.tsx b/ui/litellm-dashboard/src/components/add_model/advanced_settings.test.tsx index ceb9d1ca713..15a13219d18 100644 --- a/ui/litellm-dashboard/src/components/add_model/advanced_settings.test.tsx +++ b/ui/litellm-dashboard/src/components/add_model/advanced_settings.test.tsx @@ -52,6 +52,17 @@ describe("AdvancedSettings", () => { }); }); + it("offers both cache write tiers once custom pricing is on", async () => { + renderAdvancedSettings(); + act(() => { + fireEvent.click(screen.getByText("Advanced Settings")); + }); + fireEvent.click(await screen.findByRole("switch", { name: "Custom Pricing" })); + + expect(await screen.findByLabelText(/^Cache Write Cost \(per 1M tokens\)/)).toBeInTheDocument(); + expect(screen.getByLabelText(/^Cache Write Cost \(1 hour, per 1M tokens\)/)).toBeInTheDocument(); + }); + it("hides every PTU field when PTU cost attribution is disabled", async () => { renderAdvancedSettings(); act(() => { diff --git a/ui/litellm-dashboard/src/components/add_model/advanced_settings.tsx b/ui/litellm-dashboard/src/components/add_model/advanced_settings.tsx index eb12e49ca89..6025db8e2e2 100644 --- a/ui/litellm-dashboard/src/components/add_model/advanced_settings.tsx +++ b/ui/litellm-dashboard/src/components/add_model/advanced_settings.tsx @@ -49,6 +49,7 @@ const USAGE_COST_FIELDS = [ "output_cost_per_token", "cache_read_input_token_cost", "cache_creation_input_token_cost", + "cache_creation_input_token_cost_above_1hr", "input_cost_per_second", ]; @@ -393,6 +394,25 @@ const AdvancedSettings: React.FC = ({ /> )} + + {(control) => ( + + )} + ) : ( { expect(deployment.modelInfoObj.team_id).toBe("team-123"); }); + it("converts both cache write tiers from per-million to per-token", async () => { + const formValues = { + model_mappings: [{ public_name: "claude-sonnet", litellm_model: "anthropic/claude-sonnet-4-6" }], + model_name: "anthropic/claude-sonnet-4-6", + custom_pricing: true, + input_cost_per_token: "3", + output_cost_per_token: "15", + cache_creation_input_token_cost: "3.75", + cache_creation_input_token_cost_above_1hr: "6", + }; + + const [deployment] = (await prepareModelAddRequest({ ...formValues }, "token", null))!; + + expect(deployment.litellmParamsObj.cache_creation_input_token_cost).toBeCloseTo(0.00000375, 12); + expect(deployment.litellmParamsObj.cache_creation_input_token_cost_above_1hr).toBeCloseTo(0.000006, 12); + }); + + it("converts the pricing once when several models are added together", async () => { + const formValues = { + model_mappings: [ + { public_name: "claude-sonnet", litellm_model: "anthropic/claude-sonnet-4-6" }, + { public_name: "claude-opus", litellm_model: "anthropic/claude-opus-4-2" }, + ], + model_name: "anthropic/claude-sonnet-4-6", + custom_pricing: true, + input_cost_per_token: "3", + cache_creation_input_token_cost: "3.75", + cache_creation_input_token_cost_above_1hr: "6", + }; + + const deployments = (await prepareModelAddRequest({ ...formValues }, "token", null))!; + + expect(deployments).toHaveLength(2); + for (const deployment of deployments) { + expect(deployment.litellmParamsObj.input_cost_per_token).toBeCloseTo(0.000003, 12); + expect(deployment.litellmParamsObj.cache_creation_input_token_cost).toBeCloseTo(0.00000375, 12); + expect(deployment.litellmParamsObj.cache_creation_input_token_cost_above_1hr).toBeCloseTo(0.000006, 12); + } + }); + + it("leaves the 1 hour cache write tier unset when its field is blank", async () => { + const formValues = { + model_mappings: [{ public_name: "claude-sonnet", litellm_model: "anthropic/claude-sonnet-4-6" }], + model_name: "anthropic/claude-sonnet-4-6", + custom_pricing: true, + cache_creation_input_token_cost: "3.75", + cache_creation_input_token_cost_above_1hr: "", + }; + + const [deployment] = (await prepareModelAddRequest({ ...formValues }, "token", null))!; + + expect(deployment.litellmParamsObj).not.toHaveProperty("cache_creation_input_token_cost_above_1hr"); + expect(deployment.litellmParamsObj.cache_creation_input_token_cost).toBeCloseTo(0.00000375, 12); + }); + it("uses a lowercase fallback for unrecognized custom providers", async () => { const fallbackValues = { model_mappings: [ diff --git a/ui/litellm-dashboard/src/components/add_model/handle_add_model_submit.tsx b/ui/litellm-dashboard/src/components/add_model/handle_add_model_submit.tsx index 5f2747e0415..70b499b108a 100644 --- a/ui/litellm-dashboard/src/components/add_model/handle_add_model_submit.tsx +++ b/ui/litellm-dashboard/src/components/add_model/handle_add_model_submit.tsx @@ -24,6 +24,65 @@ export const prepareModelAddRequest = async (formValues: Record, ac formValues["model"] = wildcardModel; } + // Handle pricing conversion before processing other fields + // Use explicit checks to allow 0 (zero cost models for budget bypass) + if ( + formValues.input_cost_per_token !== undefined && + formValues.input_cost_per_token !== null && + formValues.input_cost_per_token !== "" + ) { + formValues.input_cost_per_token = Number(formValues.input_cost_per_token) / 1000000; + } + if ( + formValues.output_cost_per_token !== undefined && + formValues.output_cost_per_token !== null && + formValues.output_cost_per_token !== "" + ) { + formValues.output_cost_per_token = Number(formValues.output_cost_per_token) / 1000000; + } + + // Cache Read Cost: if blank, default to Input Cost (already token-unit converted above) + if ( + formValues.cache_read_input_token_cost !== undefined && + formValues.cache_read_input_token_cost !== null && + formValues.cache_read_input_token_cost !== "" + ) { + formValues.cache_read_input_token_cost = Number(formValues.cache_read_input_token_cost) / 1000000; + } else if ( + formValues.input_cost_per_token !== undefined && + formValues.input_cost_per_token !== null && + formValues.input_cost_per_token !== "" + ) { + formValues.cache_read_input_token_cost = Number(formValues.input_cost_per_token); + } else { + delete formValues.cache_read_input_token_cost; + } + + // Cache Write Cost: explicit value if provided, else leave unset so the + // backend keeps the model-level default (per-second pricing, model_prices + // entries, etc.). Sending 0 here would overwrite that default. + // The backend falls back to input_cost_per_token when this key is absent. + if ( + formValues.cache_creation_input_token_cost !== undefined && + formValues.cache_creation_input_token_cost !== null && + formValues.cache_creation_input_token_cost !== "" + ) { + formValues.cache_creation_input_token_cost = Number(formValues.cache_creation_input_token_cost) / 1000000; + } else { + delete formValues.cache_creation_input_token_cost; + } + if ( + formValues.cache_creation_input_token_cost_above_1hr !== undefined && + formValues.cache_creation_input_token_cost_above_1hr !== null && + formValues.cache_creation_input_token_cost_above_1hr !== "" + ) { + formValues.cache_creation_input_token_cost_above_1hr = + Number(formValues.cache_creation_input_token_cost_above_1hr) / 1000000; + } else { + delete formValues.cache_creation_input_token_cost_above_1hr; + } + // Keep input_cost_per_second as is, no conversion needed + // Create a deployment for each mapping const deployments = []; for (const mapping of modelMappings) { @@ -34,55 +93,6 @@ export const prepareModelAddRequest = async (formValues: Record, ac const modelName = mapping.public_name; litellmParamsObj["model"] = mapping.litellm_model; - // Handle pricing conversion before processing other fields - // Use explicit checks to allow 0 (zero cost models for budget bypass) - if ( - formValues.input_cost_per_token !== undefined && - formValues.input_cost_per_token !== null && - formValues.input_cost_per_token !== "" - ) { - formValues.input_cost_per_token = Number(formValues.input_cost_per_token) / 1000000; - } - if ( - formValues.output_cost_per_token !== undefined && - formValues.output_cost_per_token !== null && - formValues.output_cost_per_token !== "" - ) { - formValues.output_cost_per_token = Number(formValues.output_cost_per_token) / 1000000; - } - - // Cache Read Cost: if blank, default to Input Cost (already token-unit converted above) - if ( - formValues.cache_read_input_token_cost !== undefined && - formValues.cache_read_input_token_cost !== null && - formValues.cache_read_input_token_cost !== "" - ) { - formValues.cache_read_input_token_cost = Number(formValues.cache_read_input_token_cost) / 1000000; - } else if ( - formValues.input_cost_per_token !== undefined && - formValues.input_cost_per_token !== null && - formValues.input_cost_per_token !== "" - ) { - formValues.cache_read_input_token_cost = Number(formValues.input_cost_per_token); - } else { - delete formValues.cache_read_input_token_cost; - } - - // Cache Write Cost: explicit value if provided, else leave unset so the - // backend keeps the model-level default (per-second pricing, model_prices - // entries, etc.). Sending 0 here would overwrite that default. - // The backend falls back to input_cost_per_token when this key is absent. - if ( - formValues.cache_creation_input_token_cost !== undefined && - formValues.cache_creation_input_token_cost !== null && - formValues.cache_creation_input_token_cost !== "" - ) { - formValues.cache_creation_input_token_cost = Number(formValues.cache_creation_input_token_cost) / 1000000; - } else { - delete formValues.cache_creation_input_token_cost; - } - // Keep input_cost_per_second as is, no conversion needed - // Iterate through the key-value pairs in formValues litellmParamsObj["model"] = mapping.litellm_model; for (const [key, value] of Object.entries(formValues)) { @@ -155,7 +165,8 @@ export const prepareModelAddRequest = async (formValues: Record, ac key === "output_cost_per_token" || key === "input_cost_per_second" || key === "cache_read_input_token_cost" || - key === "cache_creation_input_token_cost" + key === "cache_creation_input_token_cost" || + key === "cache_creation_input_token_cost_above_1hr" ) { if (value !== undefined && value !== null && value !== "") { litellmParamsObj[key] = Number(value); diff --git a/ui/litellm-dashboard/src/components/model_info_view.test.tsx b/ui/litellm-dashboard/src/components/model_info_view.test.tsx index 79717988292..db3e9354350 100644 --- a/ui/litellm-dashboard/src/components/model_info_view.test.tsx +++ b/ui/litellm-dashboard/src/components/model_info_view.test.tsx @@ -1695,13 +1695,54 @@ describe("ModelInfoView", () => { const [cacheRead, cacheWrite] = screen.getAllByPlaceholderText("Defaults to Input Cost if blank"); await user.type(cacheRead, "5"); await user.type(cacheWrite, "9"); + await user.type(screen.getByPlaceholderText("Defaults to the cost map 1 hour rate, else Cache Write Cost"), "15"); const payload = await save(user); - expect(payload.litellm_params).toMatchObject({ + const expectedPricing = { output_cost_per_token: 0.000012, cache_read_input_token_cost: 0.000005, cache_creation_input_token_cost: 0.000009, + cache_creation_input_token_cost_above_1hr: 0.000015, + }; + expect(payload.litellm_params).toMatchObject(expectedPricing); + }); + + describe("1 hour cache write tier", () => { + const withCacheWriteTiers = { + ...defaultModelData, + litellm_params: { + ...defaultModelData.litellm_params, + cache_creation_input_token_cost: 0.00000375, + cache_creation_input_token_cost_above_1hr: 0.000006, + }, + }; + + beforeEach(() => { + mockUseModelsInfo.mockReturnValue({ data: { data: [withCacheWriteTiers] }, isLoading: false, error: null }); + mockModelInfoV1Call.mockResolvedValue({ data: [withCacheWriteTiers] }); + }); + + it("shows both configured cache write rates per million tokens", async () => { + render(, { wrapper }); + + expect(await screen.findByText("Cache Write Cost (1 hour, per 1M tokens)")).toBeInTheDocument(); + expect(screen.getByText("3.7500")).toBeInTheDocument(); + expect(screen.getByText("6.0000")).toBeInTheDocument(); + }); + + it("clearing the 1 hour rate sends an explicit null and keeps the 5 minute rate", async () => { + const user = userEvent.setup(); + await enterEditMode(user); + + const oneHour = screen.getByPlaceholderText("Defaults to the cost map 1 hour rate, else Cache Write Cost"); + expect(oneHour).toHaveValue(6); + await user.clear(oneHour); + + const payload = await save(user); + + expect(payload.litellm_params.cache_creation_input_token_cost_above_1hr).toBeNull(); + expect(payload.litellm_params.cache_creation_input_token_cost).toBe(0.00000375); }); }); diff --git a/ui/litellm-dashboard/src/components/model_info_view.tsx b/ui/litellm-dashboard/src/components/model_info_view.tsx index 21fad41b2e5..a8df9999f52 100644 --- a/ui/litellm-dashboard/src/components/model_info_view.tsx +++ b/ui/litellm-dashboard/src/components/model_info_view.tsx @@ -51,6 +51,9 @@ import { Tag } from "./tag_management/types"; import { getDisplayModelName } from "./view_model/model_name_display"; import { Dialog, DialogContent, DialogFooter, DialogHeader, DialogTitle } from "@/components/ui/dialog"; +const perTokenOrNull = (perMillion: string | number | null | undefined): number | null => + perMillion === undefined || perMillion === null || perMillion === "" ? null : Number(perMillion) / 1_000_000; + interface ModelInfoViewProps { modelId: string; onClose: () => void; @@ -327,19 +330,13 @@ export default function ModelInfoView({ } } - // Cache Write Cost: explicit value if provided, else explicit null so the - // backend removes the override and falls back to the model-level default. - // Sending 0 here would persist a zero rate even when the user intended to unset it. + // Cache write tiers: explicit value if provided, else explicit null so the backend removes the + // override. Sending 0 would persist a zero rate even when the user intended to unset it. if (isFieldTouched("cache_write_cost")) { - if ( - values.cache_write_cost !== undefined && - values.cache_write_cost !== null && - values.cache_write_cost !== "" - ) { - updatedLitellmParams.cache_creation_input_token_cost = Number(values.cache_write_cost) / 1_000_000; - } else { - updatedLitellmParams.cache_creation_input_token_cost = null; - } + updatedLitellmParams.cache_creation_input_token_cost = perTokenOrNull(values.cache_write_cost); + } + if (isFieldTouched("cache_write_cost_1h")) { + updatedLitellmParams.cache_creation_input_token_cost_above_1hr = perTokenOrNull(values.cache_write_cost_1h); } const storedCredentialName: string | null = localModelData?.litellm_params?.litellm_credential_name ?? null; diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index b77399c963f..bdf936d8bf3 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -49929,6 +49929,8 @@ export interface components { blocked?: boolean | null; /** Cache Creation Input Token Cost */ cache_creation_input_token_cost?: number | null; + /** Cache Creation Input Token Cost Above 1Hr */ + cache_creation_input_token_cost_above_1hr?: number | null; /** Cache Read Input Token Cost */ cache_read_input_token_cost?: number | null; /** Cost Per Ptu Per Hour */ From 60af2c601da29345ebd62df32819bff725934346 Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 16 Oct 2026 20:44:34 +0000 Subject: [PATCH 03/26] fix(proxy): emit spend queue size after updates and flushes (#46956) Co-authored-by: Mateo --- .../spend_update_queue.py | 3 + .../test_spend_update_queue.py | 69 +++++++++++++++++++ 2 files changed, 72 insertions(+) diff --git a/litellm/proxy/db/db_transaction_queue/spend_update_queue.py b/litellm/proxy/db/db_transaction_queue/spend_update_queue.py index 2b8535cb113..e88b9fc2bfb 100644 --- a/litellm/proxy/db/db_transaction_queue/spend_update_queue.py +++ b/litellm/proxy/db/db_transaction_queue/spend_update_queue.py @@ -29,6 +29,7 @@ class SpendUpdateQueue(BaseUpdateQueue): ) -> DBSpendUpdateTransactions: """Flush all updates from the queue and return all updates aggregated by entity type.""" updates: Final = await self.flush_all_updates_from_in_memory_queue() + await self._emit_new_item_added_to_queue_event(queue_size=self.update_queue.qsize()) if len(updates) > 0: verbose_proxy_logger.info( "Spend tracking - flushed %d spend update items from in-memory queue", @@ -48,6 +49,7 @@ class SpendUpdateQueue(BaseUpdateQueue): "Spend update queue is full. Aggregating all entries in queue to concatenate entries." ) await self.aggregate_queue_updates() + await self._emit_new_item_added_to_queue_event(queue_size=self.update_queue.qsize()) async def aggregate_queue_updates(self): """Concatenate all updates in the queue to reduce the size of in-memory queue""" @@ -55,6 +57,7 @@ class SpendUpdateQueue(BaseUpdateQueue): aggregated_updates: Final = self._get_aggregated_spend_update_queue_item(updates) for update in aggregated_updates: await self.update_queue.put(update) + await self._emit_new_item_added_to_queue_event(queue_size=self.update_queue.qsize()) def _get_aggregated_spend_update_queue_item( self, updates: list[SpendUpdateQueueItem] diff --git a/tests/unit/proxy/db/db_transaction_queue/test_spend_update_queue.py b/tests/unit/proxy/db/db_transaction_queue/test_spend_update_queue.py index 43f1a820885..9a4b02ba3ba 100644 --- a/tests/unit/proxy/db/db_transaction_queue/test_spend_update_queue.py +++ b/tests/unit/proxy/db/db_transaction_queue/test_spend_update_queue.py @@ -1,12 +1,81 @@ import asyncio import json +from typing import Final import pytest from fastapi.testclient import TestClient +from prometheus_client import REGISTRY +import litellm from litellm.constants import MAX_SIZE_IN_MEMORY_QUEUE from litellm.proxy._types import Litellm_EntityType, SpendUpdateQueueItem from litellm.proxy.db.db_transaction_queue.spend_update_queue import SpendUpdateQueue +from litellm.types.services import ServiceTypes + + +@pytest.mark.asyncio +@pytest.mark.parametrize("aggregation_threshold", (2, 3)) +async def test_spend_queue_gauge_tracks_enqueue_aggregation_and_drain( + monkeypatch: pytest.MonkeyPatch, aggregation_threshold: int +) -> None: + monkeypatch.setattr(litellm, "service_callback", ["prometheus_system"]) + queue: Final = SpendUpdateQueue() + queue.MAX_SIZE_IN_MEMORY_QUEUE = aggregation_threshold + response_cost: Final = 0.5 + update: Final = SpendUpdateQueueItem( + entity_type=Litellm_EntityType.USER, entity_id="queue-user", response_cost=response_cost + ) + + await queue.add_update(update) + await asyncio.sleep(0) + assert REGISTRY.get_sample_value( + "litellm_in_memory_spend_update_queue_size", + {"in_memory_spend_update_queue": str(ServiceTypes.IN_MEMORY_SPEND_UPDATE_QUEUE)}, + ) == queue.update_queue.qsize() == 1 + + await queue.add_update(update) + await asyncio.sleep(0) + assert REGISTRY.get_sample_value( + "litellm_in_memory_spend_update_queue_size", + {"in_memory_spend_update_queue": str(ServiceTypes.IN_MEMORY_SPEND_UPDATE_QUEUE)}, + ) == queue.update_queue.qsize() == (1 if aggregation_threshold == 2 else 2) + + await queue.aggregate_queue_updates() + await asyncio.sleep(0) + assert REGISTRY.get_sample_value( + "litellm_in_memory_spend_update_queue_size", + {"in_memory_spend_update_queue": str(ServiceTypes.IN_MEMORY_SPEND_UPDATE_QUEUE)}, + ) == queue.update_queue.qsize() == 1 + + transactions: Final = await queue.flush_and_get_aggregated_db_spend_update_transactions() + await asyncio.sleep(0) + assert transactions["user_list_transactions"] == {"queue-user": 2 * response_cost} + assert REGISTRY.get_sample_value( + "litellm_in_memory_spend_update_queue_size", + {"in_memory_spend_update_queue": str(ServiceTypes.IN_MEMORY_SPEND_UPDATE_QUEUE)}, + ) == queue.update_queue.qsize() == 0 + + +@pytest.mark.asyncio +async def test_spend_queue_gauge_tracks_partial_flush(monkeypatch: pytest.MonkeyPatch) -> None: + import litellm.proxy.db.db_transaction_queue.base_update_queue as base_queue_module + + monkeypatch.setattr(litellm, "service_callback", ["prometheus_system"]) + monkeypatch.setattr(base_queue_module, "MAX_IN_MEMORY_QUEUE_FLUSH_COUNT", 1) + queue: Final = SpendUpdateQueue() + update: Final = SpendUpdateQueueItem( + entity_type=Litellm_EntityType.USER, entity_id="queue-user", response_cost=0.5 + ) + await queue.add_update(update) + await queue.add_update(update) + + transactions: Final = await queue.flush_and_get_aggregated_db_spend_update_transactions() + await asyncio.sleep(0) + assert transactions["user_list_transactions"] == {"queue-user": update["response_cost"]} + assert REGISTRY.get_sample_value( + "litellm_in_memory_spend_update_queue_size", + {"in_memory_spend_update_queue": str(ServiceTypes.IN_MEMORY_SPEND_UPDATE_QUEUE)}, + ) == queue.update_queue.qsize() == 1 From 388251a5b863cf1da7b1829875a735f0a398fa9a Mon Sep 17 00:00:00 2001 From: AxelRay Date: Sat, 17 Oct 2026 04:00:37 +0700 Subject: [PATCH 04/26] fix(bedrock): drop dangerous-tool-use beta without safeguards (#44471) * fix(bedrock): drop dangerous beta without safeguards * fix(bedrock): drop the dangerous-tool-use beta only next to structured outputs, keep it on Mantle Bedrock rejects the pair with 'invalid beta flag' on Claude models before Opus 4.7, while requests without structured outputs and Bedrock Mantle accept the beta, so the drop now only applies to Invoke requests with output_config.format or a strict tool and no safeguards. * fix(anthropic): keep the AnthropicThinkingParam import that main now uses * fix(bedrock): drop the client's dangerous-tool-use beta whenever safeguards is absent Bedrock rejects the beta with 'invalid beta flag' on Sonnet 4.5, Opus 4.5, and Opus 4.6 even without structured outputs, so dropping it only next to structured outputs left those beta-only requests failing. The beta does nothing without safeguards, so Invoke now drops it in every such request. Mantle keeps it. The thinking display helper goes back to main's version. * fix(bedrock): drop the client's dangerous-tool-use beta on Invoke chat without safeguards * test(bedrock): wire cells for the dangerous-tool-use beta on Invoke messages, chat and responses * test(bedrock): type the new dangerous-tool-use beta tests * test(bedrock_mantle): type the dangerous-tool-use beta Mantle test --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../anthropic_claude3_transformation.py | 8 +- .../anthropic_claude3_transformation.py | 14 +- .../bedrock/messages/mantle_transformation.py | 3 + ...ock_invoke_dangerous_tool_use_beta_wire.py | 138 ++++++++++++++++++ ...ations_anthropic_claude3_transformation.py | 29 ++++ .../test_anthropic_claude3_transformation.py | 54 +++++++ ..._bedrock_mantle_messages_transformation.py | 9 ++ 7 files changed, 253 insertions(+), 2 deletions(-) create mode 100644 tests/integration/providers/test_bedrock_invoke_dangerous_tool_use_beta_wire.py diff --git a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py index 04747405375..fc6088ddc66 100644 --- a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py @@ -32,6 +32,7 @@ from litellm.llms.bedrock.common_utils import ( tools_without_eager_input_streaming, ) from litellm.types.llms.anthropic import ( + ANTHROPIC_BETA_HEADER_VALUES, ANTHROPIC_FINE_GRAINED_TOOL_STREAMING_BETA_HEADER, ANTHROPIC_TOOL_SEARCH_BETA_HEADER, AnthropicThinkingParam, @@ -270,7 +271,12 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): is_mid_conversation_output_config_used: Final = self.is_mid_conversation_output_config_used(messages) is_thinking_display_updates_used: Final = self.is_thinking_display_updates_used(thinking) - user_beta_set: Final = set(get_anthropic_beta_from_headers(headers)) + client_beta_set: Final = set(get_anthropic_beta_from_headers(headers)) + user_beta_set: Final = ( + client_beta_set + if "safeguards" in optional_params + else client_beta_set - {ANTHROPIC_BETA_HEADER_VALUES.DANGEROUS_TOOL_USE_2026_09_03.value} + ) beta_set: Final = set(user_beta_set) auto_betas: Final = self.get_anthropic_beta_list( model=model, diff --git a/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py index 88ae0d4d514..95e81fa3ee7 100644 --- a/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py @@ -90,6 +90,9 @@ class AmazonAnthropicClaudeMessagesConfig( def beta_headers_provider(self) -> str: return self.custom_llm_provider or "bedrock" + def drops_dangerous_tool_use_beta_without_safeguards(self) -> bool: + return True + BEDROCK_INVOKE_ALLOWED_TOP_LEVEL_FIELDS = frozenset(BedrockInvokeAnthropicMessagesRequest.__annotations__.keys()) def get_error_class( @@ -552,8 +555,17 @@ class AmazonAnthropicClaudeMessagesConfig( if anthropic_model_info.is_eager_input_streaming_used(tools): beta_set.add(ANTHROPIC_FINE_GRAINED_TOOL_STREAMING_BETA_HEADER) + dangerous_tool_use_beta: Final = ANTHROPIC_BETA_HEADER_VALUES.DANGEROUS_TOOL_USE_2026_09_03.value if anthropic_messages_optional_request_params.get("safeguards") is not None: - beta_set.add(ANTHROPIC_BETA_HEADER_VALUES.DANGEROUS_TOOL_USE_2026_09_03.value) + beta_set.add(dangerous_tool_use_beta) + elif dangerous_tool_use_beta in beta_set and self.drops_dangerous_tool_use_beta_without_safeguards(): + beta_set.discard(dangerous_tool_use_beta) + verbose_logger.debug( + "Bedrock Invoke: dropping %s from client headers because the request has no safeguards. Bedrock " + "answers it with 'invalid beta flag' next to structured outputs before Opus 4.7, and on Sonnet 4.5, " + "Opus 4.5, and Opus 4.6 even alone.", + dangerous_tool_use_beta, + ) self._filter_context_management_for_bedrock_invoke( anthropic_messages_request=anthropic_messages_request, diff --git a/litellm/llms/bedrock/messages/mantle_transformation.py b/litellm/llms/bedrock/messages/mantle_transformation.py index ca93cc3e617..1ad84761bc9 100644 --- a/litellm/llms/bedrock/messages/mantle_transformation.py +++ b/litellm/llms/bedrock/messages/mantle_transformation.py @@ -67,6 +67,9 @@ class AmazonMantleMessagesConfig(AmazonAnthropicClaudeMessagesConfig): def should_filter_anthropic_beta_headers(self) -> bool: return False + def drops_dangerous_tool_use_beta_without_safeguards(self) -> bool: + return False + def get_complete_url( self, api_base: str | None, diff --git a/tests/integration/providers/test_bedrock_invoke_dangerous_tool_use_beta_wire.py b/tests/integration/providers/test_bedrock_invoke_dangerous_tool_use_beta_wire.py new file mode 100644 index 00000000000..ca22ab4f23d --- /dev/null +++ b/tests/integration/providers/test_bedrock_invoke_dangerous_tool_use_beta_wire.py @@ -0,0 +1,138 @@ +import json +from collections.abc import Callable, Mapping +from typing import Final +from urllib.parse import unquote + +import pytest +from integration._support.client import Gateway +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue, TypeAdapter + +CLAUDE_ID: Final = "us.anthropic.claude-sonnet-4-5-20250929-v1:0" +MANTLE_CLAUDE_ID: Final = "anthropic.claude-haiku-4-5" +TOKEN: Final = "synthetic-bedrock-bearer" +DANGEROUS_TOOL_USE_BETA: Final = "dangerous-tool-use-2026-09-03" +KEPT_BETA: Final = "context-1m-2025-08-07" +SAFEGUARDS: Final[JsonValue] = [{"type": "dangerous_tool_use"}] +NO_CACHE: Final[JsonValue] = {"no-cache": True} +INVOKE_TARGET: Final = f"/model/{CLAUDE_ID}/invoke" +MANTLE_TARGET: Final = "/anthropic/v1/messages" +REPLY: Final = json.dumps( + { + "id": "msg_scripted", + "type": "message", + "role": "assistant", + "model": CLAUDE_ID, + "content": [{"type": "text", "text": "scripted reply"}], + "stop_reason": "end_turn", + "stop_sequence": None, + "usage": {"input_tokens": 3, "output_tokens": 2}, + } +).encode() +_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) +_BETAS: Final = TypeAdapter(tuple[str, ...]) + + +def _peer(request: Request) -> Reply: + if unquote(request.target) not in (INVOKE_TARGET, MANTLE_TARGET): + return Reply(status=404, body=json.dumps({"error": f"unscripted target {request.target}"}).encode()) + assert request.headers["authorization"] == f"Bearer {TOKEN}", request.headers + return Reply(body=REPLY) + + +def _invoke_body_betas(request: Request) -> frozenset[str]: + return frozenset(_BETAS.validate_python(_OBJECT.validate_json(request.body).get("anthropic_beta") or ())) + + +def _mantle_header_betas(request: Request) -> frozenset[str]: + return frozenset(beta.strip() for beta in request.headers.get("anthropic-beta", "").split(",") if beta.strip()) + + +def _delivered(gateway: Gateway, wire: Wire, path: str, body: Mapping[str, JsonValue], client_betas: str) -> Request: + response: Final = gateway.request("POST", path, body, headers={"anthropic-beta": client_betas}) + assert response.status_code == 200, response.text + (sent,) = wire.drain() + return sent + + +def _messages_body(model: str, **extra: JsonValue) -> dict[str, JsonValue]: + return { + "model": model, + "max_tokens": 16, + "messages": [{"role": "user", "content": "hi"}], + "cache": NO_CACHE, + **extra, + } + + +def _chat_body(model: str, **extra: JsonValue) -> dict[str, JsonValue]: + return {"model": model, "messages": [{"role": "user", "content": "hi"}], "cache": NO_CACHE, **extra} + + +def _responses_body(model: str, **extra: JsonValue) -> dict[str, JsonValue]: + return {"model": model, "input": "hi", "cache": NO_CACHE, **extra} + + +def test_invoke_messages_without_safeguards_sends_the_betas_of_a_request_that_never_had_the_beta( + gateway: Gateway, +) -> None: + with wire_server(_peer) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"bedrock/{CLAUDE_ID}", api_key=TOKEN, aws_region_name="us-east-1", api_base=wire.url + ) + with_beta: Final = _invoke_body_betas( + _delivered(gateway, wire, "/v1/messages", _messages_body(model), f"{KEPT_BETA},{DANGEROUS_TOOL_USE_BETA}") + ) + without_beta: Final = _invoke_body_betas( + _delivered(gateway, wire, "/v1/messages", _messages_body(model), KEPT_BETA) + ) + assert KEPT_BETA in without_beta, without_beta + assert with_beta == without_beta, (with_beta, without_beta) + + +def test_invoke_messages_with_safeguards_keeps_the_dangerous_beta(gateway: Gateway) -> None: + with wire_server(_peer) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"bedrock/{CLAUDE_ID}", api_key=TOKEN, aws_region_name="us-east-1", api_base=wire.url + ) + sent: Final = _invoke_body_betas( + _delivered( + gateway, + wire, + "/v1/messages", + _messages_body(model, safeguards=SAFEGUARDS), + f"{KEPT_BETA},{DANGEROUS_TOOL_USE_BETA}", + ) + ) + assert {KEPT_BETA, DANGEROUS_TOOL_USE_BETA} <= sent, sent + + +@pytest.mark.parametrize( + ("path", "build"), + [("/v1/chat/completions", _chat_body), ("/v1/responses", _responses_body)], + ids=["chat_completions", "responses"], +) +def test_invoke_chat_shaped_routes_without_safeguards_drop_only_the_dangerous_beta( + gateway: Gateway, path: str, build: Callable[[str], dict[str, JsonValue]] +) -> None: + with wire_server(_peer) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"bedrock/invoke/{CLAUDE_ID}", api_key=TOKEN, aws_region_name="us-east-1", api_base=wire.url + ) + with_beta: Final = _invoke_body_betas( + _delivered(gateway, wire, path, build(model), f"{KEPT_BETA},{DANGEROUS_TOOL_USE_BETA}") + ) + without_beta: Final = _invoke_body_betas(_delivered(gateway, wire, path, build(model), KEPT_BETA)) + assert KEPT_BETA in without_beta, without_beta + assert with_beta == without_beta, (with_beta, without_beta) + + +def test_mantle_messages_without_safeguards_keeps_the_dangerous_beta(gateway: Gateway) -> None: + with wire_server(_peer) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"bedrock_mantle/{MANTLE_CLAUDE_ID}", api_key=TOKEN, aws_region_name="us-east-1", api_base=wire.url + ) + sent: Final = _mantle_header_betas( + _delivered(gateway, wire, "/v1/messages", _messages_body(model), f"{KEPT_BETA},{DANGEROUS_TOOL_USE_BETA}") + ) + assert DANGEROUS_TOOL_USE_BETA in sent, sent diff --git a/tests/unit/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/unit/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py index 72857246d85..16f44f4f35f 100644 --- a/tests/unit/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py +++ b/tests/unit/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py @@ -1087,6 +1087,35 @@ def test_bedrock_chat_invoke_eager_input_streaming_beta_not_duplicated_with_clie assert result["anthropic_beta"] == [FINE_GRAINED_TOOL_STREAMING_BETA] +DANGEROUS_TOOL_USE_BETA: Final = "dangerous-tool-use-2026-09-03" + + +def _chat_invoke_betas(extra_optional_params: dict[str, object]) -> object: + config: Final = AmazonAnthropicClaudeConfig() + model: Final = "us.anthropic.claude-sonnet-4-5-20250929-v1:0" + optional_params: Final = config.map_openai_params( + non_default_params={"max_tokens": 64}, optional_params={}, model=model, drop_params=False + ) + return config.transform_request( + model=model, + messages=[{"role": "user", "content": "hi"}], + optional_params={**optional_params, **extra_optional_params}, + litellm_params={}, + headers={"anthropic-beta": f"claude-code-20250219,{DANGEROUS_TOOL_USE_BETA}"}, + ).get("anthropic_beta") + + +def test_bedrock_chat_invoke_drops_client_dangerous_tool_use_beta_without_safeguards() -> None: + assert _chat_invoke_betas({}) == ["claude-code-20250219"] + + +def test_bedrock_chat_invoke_keeps_client_dangerous_tool_use_beta_with_safeguards() -> None: + assert _chat_invoke_betas({"safeguards": [{"type": "dangerous_tool_use"}]}) == [ + "claude-code-20250219", + DANGEROUS_TOOL_USE_BETA, + ] + + def _mid_conversation_system_conversation() -> list[dict]: return [ { diff --git a/tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py b/tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py index dffe2c36e20..c91eb75dbd9 100644 --- a/tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py +++ b/tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py @@ -1677,6 +1677,60 @@ def test_bedrock_messages_does_not_add_dangerous_tool_use_beta_without_safeguard assert "dangerous-tool-use-2026-09-03" not in result.get("anthropic_beta", []) +_NAME_SCHEMA: Final = {"type": "object", "properties": {"name": {"type": "string"}}, "required": ["name"]} + + +@pytest.mark.parametrize( + "request_params", + [ + {}, + {"output_config": {"format": {"type": "json_schema", "schema": _NAME_SCHEMA}}}, + {"tools": [{"name": "get_name", "strict": True, "input_schema": _NAME_SCHEMA}]}, + ], + ids=["plain", "output_config_format", "strict_tool"], +) +def test_bedrock_messages_drops_client_dangerous_tool_use_beta_without_safeguards( + local_model_cost_map: None, local_beta_headers_config: None, request_params: dict[str, object] +) -> None: + from litellm.types.router import GenericLiteLLMParams + + result: Final = AmazonAnthropicClaudeMessagesConfig().transform_anthropic_messages_request( + model="us.anthropic.claude-haiku-4-5-20251001-v1:0", + messages=[{"role": "user", "content": [{"type": "text", "text": "Name a color."}]}], + anthropic_messages_optional_request_params={"max_tokens": 64, **request_params}, + litellm_params=GenericLiteLLMParams(), + headers={"anthropic-beta": "dangerous-tool-use-2026-09-03"}, + ) + + assert all(result.get(key) == value for key, value in request_params.items()) + assert "dangerous-tool-use-2026-09-03" not in result.get("anthropic_beta", []) + + +def test_bedrock_messages_keeps_client_dangerous_tool_use_beta_with_safeguards_and_structured_outputs( + local_model_cost_map: None, local_beta_headers_config: None +) -> None: + from litellm.types.router import GenericLiteLLMParams + + safeguards: Final = [{"type": "dangerous_tool_use", "classifier_context": {"v": 1, "permission_mode": "auto"}}] + strict_tool: Final = {"name": "get_name", "strict": True, "input_schema": _NAME_SCHEMA} + + result: Final = AmazonAnthropicClaudeMessagesConfig().transform_anthropic_messages_request( + model="us.anthropic.claude-sonnet-5", + messages=[{"role": "user", "content": [{"type": "text", "text": "Name a color."}]}], + anthropic_messages_optional_request_params={ + "max_tokens": 64, + "safeguards": safeguards, + "tools": [strict_tool], + }, + litellm_params=GenericLiteLLMParams(), + headers={"anthropic-beta": "dangerous-tool-use-2026-09-03"}, + ) + + assert result["safeguards"] == safeguards + assert result["tools"] == [strict_tool] + assert result["anthropic_beta"].count("dangerous-tool-use-2026-09-03") == 1 + + def test_bedrock_messages_stream_decoder_keeps_safeguard_results(): """Bedrock streams the classifier verdicts on message_start and on the final message_delta, exactly as api.anthropic.com does.""" decoder = AmazonAnthropicClaudeMessagesStreamDecoder(model="us.anthropic.claude-sonnet-5") diff --git a/tests/unit/llms/bedrock_mantle/test_bedrock_mantle_messages_transformation.py b/tests/unit/llms/bedrock_mantle/test_bedrock_mantle_messages_transformation.py index 923572c4f46..3236efbbfba 100644 --- a/tests/unit/llms/bedrock_mantle/test_bedrock_mantle_messages_transformation.py +++ b/tests/unit/llms/bedrock_mantle/test_bedrock_mantle_messages_transformation.py @@ -10,6 +10,7 @@ Bearer/SigV4 auth chain, and the wire request through the public entrypoint. """ import json +from typing import Final from unittest.mock import MagicMock import httpx @@ -453,6 +454,14 @@ class TestBetaHeadersOnTheWire: assert _sent_betas(route) == ["dangerous-tool-use-2026-09-03"] assert _sent_body(route)["safeguards"] == safeguards + @pytest.mark.asyncio + @respx.mock + async def test_client_dangerous_tool_use_beta_reaches_mantle_without_safeguards(self) -> None: + route: Final = await self._send(extra_headers={"anthropic-beta": "dangerous-tool-use-2026-09-03"}) + + assert _sent_betas(route) == ["dangerous-tool-use-2026-09-03"] + assert "safeguards" not in _sent_body(route) + @pytest.mark.asyncio @respx.mock async def test_betas_and_version_never_travel_in_the_body(self): From 78372b440d7c353aeb648926fbf3c69106cb8a4b Mon Sep 17 00:00:00 2001 From: "berriai-litellm-provider-info-sync[bot]" <328147090+berriai-litellm-provider-info-sync[bot]@users.noreply.github.com> Date: Fri, 16 Oct 2026 14:03:10 -0700 Subject: [PATCH 05/26] feat(cost-map): add together_ai reflection/Beam-1.0-501B-A23B (#46964) Price-Sync: litellm-providers Co-authored-by: berriai-litellm-provider-info-sync[bot] <328147090+berriai-litellm-provider-info-sync[bot]@users.noreply.github.com> --- litellm/model_prices_and_context_window_backup.json | 13 +++++++++++++ model_prices_and_context_window.json | 13 +++++++++++++ 2 files changed, 26 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 49d2ce90af4..15bf988c96b 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -75218,5 +75218,18 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true + }, + "together_ai/reflection/Beam-1.0-501B-A23B": { + "cache_read_input_token_cost": 4e-08, + "input_cost_per_token": 4e-07, + "litellm_provider": "together_ai", + "max_input_tokens": 1048576, + "max_tokens": 1048576, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_response_schema": true } } diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 49d2ce90af4..15bf988c96b 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -75218,5 +75218,18 @@ "supports_response_schema": true, "supports_tool_choice": true, "supports_vision": true + }, + "together_ai/reflection/Beam-1.0-501B-A23B": { + "cache_read_input_token_cost": 4e-08, + "input_cost_per_token": 4e-07, + "litellm_provider": "together_ai", + "max_input_tokens": 1048576, + "max_tokens": 1048576, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://api.together.ai/v1/models", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_response_schema": true } } From 5bcced0187d8091c4cb21553c6c18b7cabfa3f41 Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Fri, 16 Oct 2026 14:04:36 -0700 Subject: [PATCH 06/26] feat(ui): redesign the playground chat around a single composer and inspector (#46934) * feat(ui): add clean-slate playground chat prototype * feat(ui): replace the playground chat with the redesigned composer and inspector * fix(ui): close the playground chat gaps found in self-review Resend only complete text exchanges, so empty replies, MCP direct calls and transcriptions no longer leak into chat history. A stopped request can no longer append output, attach its gateway trace to a newer reply or clear the newer request's loading state. MCP selections are only pruned once both the server and toolset lists load. The MCP tool picker lists toolsets again and calls the server that backs the chosen tool, and a saved model the current key cannot use is cleared when the model list loads, matching the old chat Also vendors the scroll-fade-x utility the attachment primitive uses * fix(ui): resolve playground toolset tools per server and drop late session ids A toolset now lists only its own tools, each resolved on the server the toolset names, so the args form and the call always target the same tool. A saved toolset no longer asks for tools with its own id while the toolset list loads, and a failed tool list says so instead of claiming the server has no tools A stopped request can no longer overwrite the Responses session id, starting a new chat drops a request that is still running, embeddings replies are no longer resent as chat history, and a saved model missing from an already cached model list is cleared too. A failing toolset list falls back right away instead of retrying while the MCP pickers show loading * fix(ui): guard playground sends and label inspector stats Pressing Enter twice no longer sends the same message twice, an attachment the new endpoint cannot take is dropped when the endpoint changes, a stopped Responses request can no longer show its code output on a newer reply, and a reply that broke mid-stream is no longer resent as history The inspector's speed, cost and routing stats get the same icons as the reply footer, and the footer's metrics now line up with its copy and Inspect buttons * fix(ui): write playground key settings only through the encoded storage path Key source, virtual key and custom headers now go through their own write loop, separate from the plain settings, so a key value can never take the plain sessionStorage branch * fix(ui): restore playground BYOK connect and reconnect Picking a BYOK server for a direct tool call now shows a key button next to the server picker, so the key can be connected without leaving the MCP endpoint. A connected BYOK server gets a Replace key action in the MCP dialog, also shown when its tools fail to load, and saving a key reloads that server's tool list * revert: restore the single playground settings storage loop Reverts ca4e742115. The CodeQL finding covers the same encoded sessionStorage handling the old chat used, so it is left as is * test(ui): type the BYOK credential mock without deep instantiation * fix(ui): let users back out of replacing a BYOK key The replace-key form in the MCP dialog now has Cancel, and it resets when another server is picked or the dialog closes, so a working key is never forced to be overwritten * refactor(ui): name the MCP dialog row focus handler * fix(ui): accept exact custom models, tags and decimal temperatures in the playground A typed model or tag that is a prefix of a listed one could not be used, because Enter picked the partial match and the custom action only showed when nothing matched. The custom value is now a list option whenever no exact match exists. Temperature gets an editable input beside the slider so values finer than the 0.1 step can be sent again. * fix(ui): stop claiming no guardrails ran when the gateway sent no headers Direct MCP calls, A2A agents and any request that fails before headers arrive built their trace from an empty header set, so the inspector said no guardrails or policies ran. The gateway fields are now left unset in that case and the inspector says the endpoint does not report them. * fix(ui): leave policies unknown when the gateway headers are hidden The proxy exposes x-litellm-applied-guardrails to other origins but not x-litellm-applied-policies, so a dashboard on another origin read every request as having run no policies. Policies are now unknown unless the call id header is readable, and the inspector says so. --- tests/e2e/ui/helpers/playground.ts | 55 +- .../guardrails/presidioUserStory.spec.ts | 10 +- .../tests/internal-user/internalUser.spec.ts | 9 +- .../ui/tests/playground/playground.spec.ts | 23 +- .../e2e/ui/tests/team-admin/teamAdmin.spec.ts | 15 +- ui/litellm-dashboard/eslint-suppressions.json | 15 + .../components/chat/ByokConnect.tsx | 152 ++++ .../components/chat/Composer.test.tsx | 61 ++ .../playground/components/chat/Composer.tsx | 196 ++++++ .../components/chat/ComposerControls.tsx | 120 ++++ .../components/chat/Conversation.tsx | 388 +++++++++++ .../playground/components/chat/Inspector.tsx | 292 ++++++++ .../playground/components/chat/McpDialog.tsx | 391 +++++++++++ .../playground/components/chat/ModelName.tsx | 74 ++ .../components/chat/Pickers.test.tsx | 83 +++ .../playground/components/chat/Pickers.tsx | 193 +++++ .../chat/PlaygroundChat.integration.test.tsx | 658 ++++++++++++++++++ .../components/chat/PlaygroundChat.tsx | 489 +++++++++++++ .../components/chat/ReplyContent.tsx | 92 +++ .../components/chat/SettingsPopover.tsx | 395 +++++++++++ .../components/chat/composerFiles.ts | 52 ++ .../components/chat/dispatchRequest.test.ts | 58 ++ .../components/chat/dispatchRequest.ts | 295 ++++++++ .../components/chat/endpointCapabilities.ts | 104 +++ .../chat/playgroundSettings.test.ts | 149 ++++ .../components/chat/playgroundSettings.ts | 192 +++++ .../playground/components/chat/useIsWide.ts | 16 + .../components/chat/usePlaygroundCatalog.ts | 147 ++++ .../components/chat/usePlaygroundSend.ts | 263 +++++++ .../components/chat/usePlaygroundSettings.ts | 38 + .../llm_calls/anthropic_messages.tsx | 3 + .../playground/llm_calls/audio_speech.tsx | 3 + .../llm_calls/audio_transcriptions.tsx | 3 + .../playground/llm_calls/embeddings_api.tsx | 3 + .../playground/llm_calls/image_edits.tsx | 3 + .../playground/llm_calls/image_generation.tsx | 3 + .../playground/llm_calls/interactions_api.tsx | 3 + .../app/(dashboard)/playground/page.test.tsx | 4 +- .../src/app/(dashboard)/playground/page.tsx | 87 ++- ui/litellm-dashboard/src/app/globals.css | 196 ++++++ .../src/components/chat_ui/types.ts | 24 + .../components/llm_calls/chat_completion.tsx | 23 +- .../components/llm_calls/reply_trace.test.ts | 64 ++ .../src/components/llm_calls/reply_trace.ts | 49 ++ .../components/llm_calls/responses_api.tsx | 18 +- .../src/components/ui/accordion.tsx | 59 ++ .../src/components/ui/attachment.tsx | 185 +++++ .../src/components/ui/item.tsx | 179 +++++ .../src/components/ui/kbd.tsx | 20 + .../src/components/ui/marker.tsx | 67 ++ .../src/components/ui/message-scroller.tsx | 117 ++++ .../src/components/ui/message.tsx | 74 ++ .../src/components/ui/spinner.tsx | 16 + .../src/components/ui/toggle-group.tsx | 87 +++ .../src/components/ui/toggle.tsx | 37 + .../src/lib/http/observeResponseHeaders.ts | 9 + 56 files changed, 6245 insertions(+), 116 deletions(-) create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ByokConnect.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ComposerControls.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Conversation.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Inspector.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/McpDialog.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ModelName.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.integration.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ReplyContent.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/SettingsPopover.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/composerFiles.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/dispatchRequest.test.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/dispatchRequest.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/endpointCapabilities.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/playgroundSettings.test.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/playgroundSettings.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/useIsWide.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/usePlaygroundCatalog.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/usePlaygroundSend.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/usePlaygroundSettings.ts create mode 100644 ui/litellm-dashboard/src/components/llm_calls/reply_trace.test.ts create mode 100644 ui/litellm-dashboard/src/components/llm_calls/reply_trace.ts create mode 100644 ui/litellm-dashboard/src/components/ui/accordion.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/attachment.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/item.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/kbd.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/marker.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/message-scroller.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/message.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/spinner.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/toggle-group.tsx create mode 100644 ui/litellm-dashboard/src/components/ui/toggle.tsx create mode 100644 ui/litellm-dashboard/src/lib/http/observeResponseHeaders.ts diff --git a/tests/e2e/ui/helpers/playground.ts b/tests/e2e/ui/helpers/playground.ts index 866334e1c94..b46a2263072 100644 --- a/tests/e2e/ui/helpers/playground.ts +++ b/tests/e2e/ui/helpers/playground.ts @@ -2,42 +2,57 @@ import { expect, type Locator, type Page as PlaywrightPage } from "@playwright/t import { navigateToPage, dismissFeedbackPopup } from "./navigation"; import { Page } from "../fixtures/pages"; -/** Controls for the Test Key / Playground page, shared with the router-fallback specs. */ - -/** - * The configuration panel is rendered twice, docked and overlay, with one visible at a time. - * Every control is narrowed to the visible copy or it trips strict mode against its hidden twin. - */ export const onlyVisible = (locator: Locator): Locator => locator.filter({ visible: true }).first(); -/** The model combobox, addressed by the placeholder its search input shows before selection. */ -export const modelSelect = (page: PlaywrightPage): Locator => onlyVisible(page.getByPlaceholder("Select a Model")); +export const composer = (page: PlaywrightPage): Locator => + onlyVisible(page.getByRole("group").filter({ has: page.getByRole("textbox", { name: "Message" }) })); + +export const modelSelect = (page: PlaywrightPage): Locator => + onlyVisible(page.getByRole("combobox", { name: "Model", exact: true })); export const sendButton = (page: PlaywrightPage): Locator => - onlyVisible(page.getByRole("button", { name: "Send message" })); + composer(page).getByRole("button", { name: "Send message" }); -/** The Virtual Key Source dropdown, addressed by the accessible name on its trigger. */ -export const keySourceSelect = (page: PlaywrightPage): Locator => - onlyVisible(page.getByLabel("Virtual Key Source")); +export const keySourceSelect = (page: PlaywrightPage): Locator => onlyVisible(page.getByLabel("Send as")); export async function openPlayground(page: PlaywrightPage): Promise { await navigateToPage(page, Page.LlmPlayground); await dismissFeedbackPopup(page); - await expect(onlyVisible(page.getByText("Virtual Key Source"))).toBeVisible({ - timeout: 20_000, - }); + await expect(modelSelect(page)).toBeVisible({ timeout: 20_000 }); + const searchHint = page.getByRole("button", { name: "Dismiss search hint" }); + if (await searchHint.isVisible()) { + await searchHint.click(); + } +} + +export async function openRequestSettings(page: PlaywrightPage): Promise { + await composer(page).getByRole("button", { name: "Settings", exact: true }).click(); + await expect(keySourceSelect(page)).toBeVisible({ timeout: 10_000 }); +} + +export async function enterVirtualKey(page: PlaywrightPage, key: string): Promise { + await openRequestSettings(page); + await keySourceSelect(page).click(); + await onlyVisible(page.getByRole("option", { name: "Virtual key", exact: true })).click({ timeout: 15_000 }); + const keyInput = onlyVisible(page.getByLabel("Virtual key", { exact: true })); + await expect(keyInput).toBeVisible({ timeout: 10_000 }); + await keyInput.fill(key); + await page.keyboard.press("Escape"); +} + +export async function searchModels(page: PlaywrightPage, query: string): Promise { + await modelSelect(page).click(); + await onlyVisible(page.getByPlaceholder("Search models")).fill(query); } export async function selectModel(page: PlaywrightPage, model: string): Promise { - const select = modelSelect(page); - await select.click(); - // Virtualized: options outside the rendered window are absent from the DOM, so search first. - await select.fill(model); + await searchModels(page, model); await onlyVisible(page.getByRole("option", { name: model, exact: true })).click({ timeout: 15_000 }); + await expect(modelSelect(page)).toContainText(model); } export async function sendMessage(page: PlaywrightPage, message: string): Promise { - const input = onlyVisible(page.getByPlaceholder("Type your message", { exact: false })); + const input = composer(page).getByRole("textbox", { name: "Message" }); await expect(input).toBeVisible({ timeout: 15_000 }); await input.fill(message); await sendButton(page).click(); diff --git a/tests/e2e/ui/tests/guardrails/presidioUserStory.spec.ts b/tests/e2e/ui/tests/guardrails/presidioUserStory.spec.ts index b4dc7c4e8da..c57a9147ab0 100644 --- a/tests/e2e/ui/tests/guardrails/presidioUserStory.spec.ts +++ b/tests/e2e/ui/tests/guardrails/presidioUserStory.spec.ts @@ -3,7 +3,7 @@ import { ADMIN_STORAGE_PATH, MOCK_PRESIDIO_URL, PROPAGATION_TIMEOUT_MS } from ". import { navigateToPage, dismissFeedbackPopup } from "../../helpers/navigation"; import { Page } from "../../fixtures/pages"; import { CHAT_MODEL_A, masterKey, rootPath, uniqueSuffix, waitForSpendLog } from "../../helpers/traffic"; -import { openPlayground, selectModel, sendButton, onlyVisible } from "../../helpers/playground"; +import { composer, openPlayground, selectModel, sendButton, onlyVisible } from "../../helpers/playground"; const RAW_EMAIL = "jane.doe@example.com"; const RAW_PHONE = "555-867-5309"; @@ -123,14 +123,12 @@ test.describe("Presidio PII guardrail, end to end from the dashboard", () => { await openPlayground(page); await selectModel(page, CHAT_MODEL_A); - const guardrailSelect = onlyVisible(page.getByPlaceholder("Select guardrails")); - await expect(guardrailSelect).toBeVisible({ timeout: 20_000 }); - await guardrailSelect.click(); - await guardrailSelect.fill(guardrailName); + await composer(page).getByRole("combobox", { name: "Guardrails" }).click(); + await onlyVisible(page.getByPlaceholder("Search guardrails")).fill(guardrailName); await onlyVisible(page.getByRole("option", { name: guardrailName, exact: true })).click({ timeout: 20_000 }); await page.keyboard.press("Escape"); - const input = onlyVisible(page.getByPlaceholder("Type your message", { exact: false })); + const input = composer(page).getByRole("textbox", { name: "Message" }); await expect(input).toBeVisible({ timeout: 15_000 }); await input.fill(prompt); diff --git a/tests/e2e/ui/tests/internal-user/internalUser.spec.ts b/tests/e2e/ui/tests/internal-user/internalUser.spec.ts index f392c5104da..d34bc034453 100644 --- a/tests/e2e/ui/tests/internal-user/internalUser.spec.ts +++ b/tests/e2e/ui/tests/internal-user/internalUser.spec.ts @@ -9,7 +9,7 @@ import { import { Page } from "../../fixtures/pages"; import { navigateToPage, clickTeamId } from "../../helpers/navigation"; import { CHAT_MODEL_A, MOCK_RESPONSE_TEXT, masterKey } from "../../helpers/traffic"; -import { keySourceSelect, onlyVisible, openPlayground, selectModel, sendMessage } from "../../helpers/playground"; +import { openPlayground, selectModel, sendMessage, enterVirtualKey } from "../../helpers/playground"; test.describe("Internal User", () => { test.use({ storageState: INTERNAL_USER_STORAGE_PATH }); @@ -73,12 +73,7 @@ test.describe("Internal User", () => { try { await openPlayground(page); - await keySourceSelect(page).click(); - await onlyVisible(page.getByRole("option", { name: "Virtual Key" })).click({ timeout: 15_000 }); - - const keyInput = onlyVisible(page.getByPlaceholder("Enter custom Virtual Key")); - await expect(keyInput).toBeVisible({ timeout: 10_000 }); - await keyInput.fill(apiKey); + await enterVirtualKey(page, apiKey); await selectModel(page, CHAT_MODEL_A); await sendMessage(page, `internal user team key ping ${keyName}`); diff --git a/tests/e2e/ui/tests/playground/playground.spec.ts b/tests/e2e/ui/tests/playground/playground.spec.ts index 9a9ecd539b3..205cc57ae46 100644 --- a/tests/e2e/ui/tests/playground/playground.spec.ts +++ b/tests/e2e/ui/tests/playground/playground.spec.ts @@ -1,7 +1,14 @@ import { test, expect } from "@playwright/test"; import { ADMIN_STORAGE_PATH } from "../../constants"; import { CHAT_MODEL_A, CHAT_MODEL_B, MOCK_RESPONSE_TEXT, createVirtualKey } from "../../helpers/traffic"; -import { keySourceSelect, onlyVisible, openPlayground, selectModel, sendMessage } from "../../helpers/playground"; +import { + keySourceSelect, + openPlayground, + openRequestSettings, + selectModel, + sendMessage, + enterVirtualKey, +} from "../../helpers/playground"; /** * The one flow that exercises the dashboard's own LLM call path rather than an admin CRUD endpoint, @@ -14,8 +21,9 @@ test.describe("Playground", () => { test(`chats with ${model} using the current UI session`, async ({ page }) => { await openPlayground(page); - // "Current UI Session" is the default: the logged-in admin's key, nothing pasted. - await expect(keySourceSelect(page)).toContainText("Current UI Session"); + await openRequestSettings(page); + await expect(keySourceSelect(page)).toContainText("Current UI session"); + await page.keyboard.press("Escape"); await selectModel(page, model); const prompt = `playground ping for ${model}`; @@ -34,14 +42,7 @@ test.describe("Playground", () => { await openPlayground(page); - // Switch the source to "Virtual Key" and paste the key we just minted. - await keySourceSelect(page).click(); - await onlyVisible(page.getByRole("option", { name: "Virtual Key" })).click({ timeout: 15_000 }); - - const keyInput = onlyVisible(page.getByPlaceholder("Enter custom Virtual Key")); - await expect(keyInput).toBeVisible({ timeout: 10_000 }); - await keyInput.fill(key); - + await enterVirtualKey(page, key); await selectModel(page, CHAT_MODEL_A); await sendMessage(page, "playground ping via virtual key"); diff --git a/tests/e2e/ui/tests/team-admin/teamAdmin.spec.ts b/tests/e2e/ui/tests/team-admin/teamAdmin.spec.ts index f3c031f0172..ff41778b2c9 100644 --- a/tests/e2e/ui/tests/team-admin/teamAdmin.spec.ts +++ b/tests/e2e/ui/tests/team-admin/teamAdmin.spec.ts @@ -10,7 +10,7 @@ import { Page } from "../../fixtures/pages"; import { navigateToPage, dismissFeedbackPopup, clickTeamId } from "../../helpers/navigation"; import { captureRequestBody, readBack } from "../../helpers/roundTrip"; import { CHAT_MODEL_A, masterKey } from "../../helpers/traffic"; -import { keySourceSelect, modelSelect, onlyVisible, openPlayground } from "../../helpers/playground"; +import { onlyVisible, openPlayground, searchModels, enterVirtualKey } from "../../helpers/playground"; /** * Every identifier a roster is addressable by. Which of user_id / user_email is populated depends on @@ -228,21 +228,14 @@ test.describe("Team Admin", () => { .toBe(true); await openPlayground(page); - await keySourceSelect(page).click(); - await onlyVisible(page.getByRole("option", { name: "Virtual Key" })).click({ timeout: 15_000 }); + await enterVirtualKey(page, teamKey); - const keyInput = onlyVisible(page.getByPlaceholder("Enter custom Virtual Key")); - await expect(keyInput).toBeVisible({ timeout: 10_000 }); - await keyInput.fill(teamKey); - - const select = modelSelect(page); - await select.click(); - await select.fill(teamModelName); + await searchModels(page, teamModelName); await expect(onlyVisible(page.getByRole("option", { name: teamModelName }))).toBeVisible({ timeout: 15_000, }); - await select.fill(CHAT_MODEL_A); + await onlyVisible(page.getByPlaceholder("Search models")).fill(CHAT_MODEL_A); await expect(onlyVisible(page.getByRole("option", { name: CHAT_MODEL_A }))).toBeVisible({ timeout: 15_000, }); diff --git a/ui/litellm-dashboard/eslint-suppressions.json b/ui/litellm-dashboard/eslint-suppressions.json index b1b6cef7320..9eead41a102 100644 --- a/ui/litellm-dashboard/eslint-suppressions.json +++ b/ui/litellm-dashboard/eslint-suppressions.json @@ -2090,11 +2090,26 @@ "count": 1 } }, + "src/components/ui/attachment.tsx": { + "local/no-ad-hoc-z-index": { + "count": 2 + } + }, + "src/components/ui/message-scroller.tsx": { + "local/no-noop-hover-variant": { + "count": 1 + } + }, "src/components/ui/slider.tsx": { "no-nested-ternary": { "count": 1 } }, + "src/components/ui/toggle-group.tsx": { + "local/no-ad-hoc-z-index": { + "count": 2 + } + }, "src/components/update_model_credentials_modal.tsx": { "local/filename-pascal-case": { "count": 1 diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ByokConnect.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ByokConnect.tsx new file mode 100644 index 00000000000..f94c43b720a --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ByokConnect.tsx @@ -0,0 +1,152 @@ +"use client"; + +import { useId, useState } from "react"; +import { Check, ExternalLink, KeyRound } from "lucide-react"; +import { PasswordInput } from "@/components/shared/PasswordInput"; +import { Button } from "@/components/ui/button"; +import { InputGroupButton } from "@/components/ui/input-group"; +import { Popover, PopoverContent, PopoverTrigger } from "@/components/ui/popover"; +import { Field, FieldDescription, FieldGroup, FieldLabel } from "@/components/ui/field"; +import { Spinner } from "@/components/ui/spinner"; +import { Switch } from "@/components/ui/switch"; +import type { MCPServer } from "@/components/mcp_tools/types"; +import { fetchClient } from "@/lib/http/api"; +import { ApiError } from "@/lib/http/client"; +import { toast } from "@/lib/toast"; + +const saveErrorMessage = (error: unknown): string => { + if (error instanceof ApiError) { + const detail = (error.body as { detail?: { error?: string } } | null)?.detail?.error; + if (detail) return detail; + } + return error instanceof Error && error.message ? error.message : "Failed to connect"; +}; + +export function ByokConnect({ + server, + name, + onConnected, + onCancel, +}: { + server: MCPServer; + name: string; + onConnected: () => void; + onCancel?: () => void; +}) { + const [credential, setCredential] = useState(""); + const [save, setSave] = useState(true); + const [isSaving, setIsSaving] = useState(false); + const keyId = useId(); + const saveId = useId(); + + const connect = async () => { + if (!credential.trim()) { + toast.error("Enter your API key"); + return; + } + setIsSaving(true); + try { + await fetchClient.POST("/v1/mcp/server/{server_id}/user-credential", { + params: { path: { server_id: server.server_id } }, + body: { credential: credential.trim(), save }, + }); + toast.success(`Connected to ${name}`); + setCredential(""); + onConnected(); + } catch (error) { + toast.error(saveErrorMessage(error)); + } finally { + setIsSaving(false); + } + }; + + return ( + +
+

+ + {server.has_user_credential ? `Replace your ${name} API key` : `${name} needs your API key`} +

+

+ {server.has_user_credential + ? `The new key replaces the one the gateway uses for ${name}.` + : `The gateway calls ${name} with your own key. Connect it to list and use its tools.`} +

+
+ {server.byok_description && server.byok_description.length > 0 && ( +
    + {server.byok_description.map((item) => ( +
  • + + {item} +
    • + ))} +
+ )} + + {name} API key + setCredential(event.target.value)} + placeholder="Paste your API key" + /> + {server.byok_api_key_help_url && ( + + + Where to find this key + + + + )} + + + + Save the key for next time + + + +
+ + {onCancel && ( + + )} +
+ + ); +} + +export function ByokKeyButton({ + server, + name, + onConnected, +}: { + server: MCPServer; + name: string; + onConnected: () => void; +}) { + const [open, setOpen] = useState(false); + return ( + + }> + + {server.has_user_credential ? "Replace key" : "Connect key"} + + + { + setOpen(false); + onConnected(); + }} + /> + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.test.tsx new file mode 100644 index 00000000000..c62b839e5b3 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.test.tsx @@ -0,0 +1,61 @@ +import { createRef } from "react"; +import { createPortal } from "react-dom"; +import { render, screen } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { describe, expect, it, vi } from "vitest"; +import { EndpointType } from "@/components/chat_ui/mode_endpoint_mapping"; +import type { MCPToolArgumentsFormRef } from "@/components/mcp_tools/MCPToolArgumentsForm"; +import { Composer, type ChecklistProps, type DraftState } from "./Composer"; +import { capabilitiesFor } from "./endpointCapabilities"; + +const checklist: ChecklistProps = { groups: [], value: [], onValueChange: vi.fn(), isLoading: false }; + +const draft: DraftState = { + text: "", + setText: vi.fn(), + files: [], + attach: vi.fn(), + removeFile: vi.fn(), + isLoading: false, + send: vi.fn().mockResolvedValue(undefined), + cancel: vi.fn(), +}; + +const renderComposer = () => + render( + } + draft={draft} + directTool={null} + toolArgsForm={createRef()} + mcpControl={null} + vectorStores={checklist} + guardrails={checklist} + codeInterpreter={{ enabled: false, onToggle: vi.fn() }} + settingsControl={createPortal(, document.body)} + sendDisabled + />, + ); + +describe("Composer", () => { + it("keeps focus in a popover field opened from the composer toolbar", async () => { + const user = userEvent.setup(); + renderComposer(); + + await user.click(screen.getByLabelText("Request tags")); + + expect(screen.getByLabelText("Request tags")).toHaveFocus(); + }); + + it("focuses the message box when empty toolbar space is clicked", async () => { + const user = userEvent.setup(); + const { container } = renderComposer(); + const toolbar = container.querySelectorAll('[data-slot="input-group-addon"]'); + + await user.click(toolbar[toolbar.length - 1]); + + expect(screen.getByLabelText("Message")).toHaveFocus(); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.tsx new file mode 100644 index 00000000000..a9fbb777393 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Composer.tsx @@ -0,0 +1,196 @@ +"use client"; + +import { useRef, type DragEvent, type KeyboardEvent, type MouseEvent, type ReactNode, type RefObject } from "react"; +import { ArrowUp, Code2, Database, Paperclip, Shield, Square } from "lucide-react"; +import { EndpointType } from "@/components/chat_ui/mode_endpoint_mapping"; +import MCPToolArgumentsForm, { type MCPToolArgumentsFormRef } from "@/components/mcp_tools/MCPToolArgumentsForm"; +import type { MCPTool } from "@/components/mcp_tools/types"; +import { InputGroup, InputGroupAddon, InputGroupButton, InputGroupTextarea } from "@/components/ui/input-group"; +import { Toggle } from "@/components/ui/toggle"; +import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"; +import { DraftAttachments } from "./ComposerControls"; +import type { DraftFile } from "./composerFiles"; +import { ACCEPT_FOR, ATTACH_LABEL, type Capabilities } from "./endpointCapabilities"; +import { ChecklistMenu, type OptionGroup } from "./Pickers"; + +export type ChecklistProps = Readonly<{ + groups: readonly OptionGroup[]; + value: readonly string[]; + onValueChange: (value: readonly string[]) => void; + isLoading: boolean; +}>; + +export type DraftState = Readonly<{ + text: string; + setText: (text: string) => void; + files: readonly DraftFile[]; + attach: (files: readonly File[]) => void; + removeFile: (index: number) => void; + isLoading: boolean; + send: () => Promise; + cancel: () => void; +}>; + +function AttachButton({ + capabilities, + attach, +}: { + capabilities: Capabilities; + attach: (files: readonly File[]) => void; +}) { + const input = useRef(null); + if (!capabilities.attach) return null; + const label = ATTACH_LABEL[capabilities.attach]; + return ( + <> + + input.current?.click()} />} + > + + + {label} + + { + attach(Array.from(event.target.files ?? [])); + event.target.value = ""; + }} + /> + + ); +} + +function SendButton({ draft, disabled }: { draft: DraftState; disabled: boolean }) { + if (draft.isLoading) + return ( + + + + ); + return ( + void draft.send()} + > + + + ); +} + +export function Composer({ + capabilities, + endpoint, + targetRow, + draft, + directTool, + toolArgsForm, + mcpControl, + vectorStores, + guardrails, + codeInterpreter, + settingsControl, + sendDisabled, +}: { + capabilities: Capabilities; + endpoint: EndpointType; + targetRow: ReactNode; + draft: DraftState; + directTool: MCPTool | null; + toolArgsForm: RefObject; + mcpControl: ReactNode; + vectorStores: ChecklistProps; + guardrails: ChecklistProps; + codeInterpreter: Readonly<{ enabled: boolean; onToggle: (enabled: boolean) => void }>; + settingsControl: ReactNode; + sendDisabled: boolean; +}) { + const isDirectCall = endpoint === EndpointType.MCP; + const isRealtime = endpoint === EndpointType.REALTIME; + const message = useRef(null); + const focusMessage = (event: MouseEvent) => { + const target = event.target as HTMLElement; + if (!event.currentTarget.contains(target) || target.closest("button, input, textarea")) return; + message.current?.focus(); + }; + const handleDrop = (event: DragEvent) => { + if (!capabilities.attach) return; + event.preventDefault(); + draft.attach(Array.from(event.dataTransfer.files)); + }; + const handleKeyDown = (event: KeyboardEvent) => { + if (event.key !== "Enter" || event.shiftKey || event.nativeEvent.isComposing) return; + event.preventDefault(); + if (!sendDisabled) void draft.send(); + }; + const directCallBody = ( +
+ {directTool ? ( + + ) : ( +

Pick a server and a tool above to fill in its arguments

+ )} +
+ ); + const textBody = ( + draft.setText(event.target.value)} + onKeyDown={handleKeyDown} + placeholder={capabilities.placeholder} + rows={2} + aria-label="Message" + className="max-h-60" + /> + ); + return ( + capabilities.attach && event.preventDefault()} onDrop={handleDrop}> + + {targetRow} + + {draft.files.length > 0 && ( + + + + )} + {!isRealtime && (isDirectCall ? directCallBody : textBody)} + + {capabilities.mcp && mcpControl} + {capabilities.vectorStores && ( + } emptyText="No vector stores yet" {...vectorStores} /> + )} + {capabilities.guardrails && ( + } emptyText="No guardrails yet" {...guardrails} /> + )} + {capabilities.codeInterpreter && ( + + + Code interpreter + + )} + + {settingsControl} + + {!isRealtime && } + + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ComposerControls.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ComposerControls.tsx new file mode 100644 index 00000000000..e3e52be3b7b --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ComposerControls.tsx @@ -0,0 +1,120 @@ +"use client"; + +import { ChevronDown, FileText, Music, X } from "lucide-react"; +import { EndpointType } from "@/components/chat_ui/mode_endpoint_mapping"; +import { + Attachment, + AttachmentAction, + AttachmentActions, + AttachmentContent, + AttachmentDescription, + AttachmentGroup, + AttachmentMedia, + AttachmentTitle, +} from "@/components/ui/attachment"; +import { + DropdownMenu, + DropdownMenuContent, + DropdownMenuLabel, + DropdownMenuRadioGroup, + DropdownMenuRadioItem, + DropdownMenuTrigger, +} from "@/components/ui/dropdown-menu"; +import { InputGroupButton } from "@/components/ui/input-group"; +import { ENDPOINT_OPTIONS, OPEN_AI_VOICE_SELECT_OPTIONS, type OpenAIVoice } from "../chat_ui/chatConstants"; +import { formatBytes, type DraftFile } from "./composerFiles"; + +const shortEndpoint = (label: string): string => label.replace(/^\/v1(beta)?\//, ""); + +export function EndpointMenu({ + value, + onValueChange, +}: { + value: EndpointType; + onValueChange: (value: EndpointType) => void; +}) { + const current = ENDPOINT_OPTIONS.find((option) => option.value === value); + return ( + + }> + {shortEndpoint(current?.label ?? value)} + + + + onValueChange(next as EndpointType)}> + Endpoint + {ENDPOINT_OPTIONS.map((option) => ( + + {option.label} + + ))} + + + + ); +} + +export function VoiceMenu({ + value, + onValueChange, +}: { + value: OpenAIVoice; + onValueChange: (value: OpenAIVoice) => void; +}) { + const current = OPEN_AI_VOICE_SELECT_OPTIONS.find((option) => option.value === value); + return ( + + }> + Voice: {current?.label.split(" - ")[0] ?? value} + + + + onValueChange(next as OpenAIVoice)}> + Voice + {OPEN_AI_VOICE_SELECT_OPTIONS.map((option) => ( + + {option.label} + + ))} + + + + ); +} + +function FileMedia({ draft }: { draft: DraftFile }) { + if (draft.previewUrl) + return ( + + + + ); + return {draft.file.type.startsWith("audio/") ? : }; +} + +export function DraftAttachments({ + files, + onRemove, +}: { + files: readonly DraftFile[]; + onRemove: (index: number) => void; +}) { + return ( + + {files.map((draft, index) => ( + + + + {draft.file.name} + {formatBytes(draft.file.size)} + + + onRemove(index)}> + + + + + ))} + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Conversation.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Conversation.tsx new file mode 100644 index 00000000000..31eb395408a --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Conversation.tsx @@ -0,0 +1,388 @@ +"use client"; + +import type { ReactNode } from "react"; +import { + ArrowDownToLine, + ArrowUpFromLine, + Brain, + Clock, + Coins, + DatabaseZap, + FileText, + Music, + PlayCircle, + Repeat, + ScanSearch, + Timer, + Wrench, +} from "lucide-react"; +import { EndpointType } from "@/components/chat_ui/mode_endpoint_mapping"; +import MCPEventsDisplay from "@/components/chat_ui/MCPEventsDisplay"; +import ReasoningContent from "@/components/chat_ui/ReasoningContent"; +import type { MessageAttachment, MessageType } from "@/components/chat_ui/types"; +import type { CodeInterpreterResult } from "@/components/llm_calls/code_interpreter_handler"; +import type { MCPEvent } from "@/components/mcp_tools/types"; +import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert"; +import { + Attachment, + AttachmentContent, + AttachmentDescription, + AttachmentGroup, + AttachmentMedia, + AttachmentTitle, +} from "@/components/ui/attachment"; +import { Badge } from "@/components/ui/badge"; +import { Bubble, BubbleContent } from "@/components/ui/Bubble"; +import { Button } from "@/components/ui/button"; +import { Empty, EmptyContent, EmptyDescription, EmptyHeader, EmptyMedia, EmptyTitle } from "@/components/ui/empty"; +import { Marker, MarkerContent, MarkerIcon } from "@/components/ui/marker"; +import { Message, MessageContent, MessageFooter, MessageHeader } from "@/components/ui/message"; +import { + MessageScroller, + MessageScrollerButton, + MessageScrollerContent, + MessageScrollerItem, + MessageScrollerProvider, + MessageScrollerViewport, +} from "@/components/ui/message-scroller"; +import { Spinner } from "@/components/ui/spinner"; +import { Toggle } from "@/components/ui/toggle"; +import A2AMetrics from "../chat_ui/A2AMetrics"; +import CodeInterpreterOutput from "../chat_ui/CodeInterpreterOutput"; +import { SearchResultsDisplay } from "../chat_ui/SearchResultsDisplay"; +import { formatBytes } from "./composerFiles"; +import { CopyIconButton, FallbackBadge, LongValue } from "./ModelName"; +import { ReplyBody } from "./ReplyContent"; + +export const formatSeconds = (ms: number): string => `${(ms / 1000).toFixed(2)}s`; + +export const formatCost = (cost: number): string => + `$${cost < 0.0001 && cost > 0 ? cost.toPrecision(2) : cost.toFixed(4)}`; + +const MCP_EVENT_ENDPOINTS: readonly EndpointType[] = [EndpointType.CHAT, EndpointType.RESPONSES]; + +const textOf = (message: MessageType): string => (typeof message.content === "string" ? message.content : ""); + +const fellBack = (message: MessageType): boolean => (message.trace?.attemptedFallbacks ?? 0) > 0; + +export const servedModel = (message: MessageType): string | undefined => + message.trace?.modelGroup ?? (message.isError ? undefined : message.model ?? message.trace?.requestedModel); + +const replyModel = (message: MessageType): Readonly<{ value: string; detail: string }> | null => { + const served = servedModel(message); + if (served) return { value: served, detail: fellBack(message) ? "Served by, after a fallback" : "Served by" }; + const requested = message.trace?.requestedModel ?? message.model; + return requested ? { value: requested, detail: "Requested model" } : null; +}; + +function Metric({ icon, value, label }: { icon: ReactNode; value: string; label?: string }) { + return ( + + {icon} + {value} + {label} + + ); +} + +function ReplyMetrics({ message }: { message: MessageType }) { + const usage = message.usage; + const cost = usage?.cost ?? message.trace?.cost; + const cached = usage?.servedFromResponseCache || message.trace?.cacheHit; + return ( +
+ {message.timeToFirstToken !== undefined && ( + } value={formatSeconds(message.timeToFirstToken)} label="to first token" /> + )} + {message.totalLatency !== undefined && ( + } value={formatSeconds(message.totalLatency)} label="total" /> + )} + {usage?.promptTokens !== undefined && ( + } value={String(usage.promptTokens)} label="in" /> + )} + {usage?.completionTokens !== undefined && ( + } value={String(usage.completionTokens)} label="out" /> + )} + {usage?.reasoningTokens ? ( + } value={String(usage.reasoningTokens)} label="reasoning" /> + ) : null} + {usage?.cacheReadTokens ? ( + } value={String(usage.cacheReadTokens)} label="cached" /> + ) : null} + {cost !== undefined && } value={formatCost(cost)} />} + {message.toolName && } value={message.toolName} />} + {cached && Cached response} +
+ ); +} + +function AttachmentMediaFor({ attachment }: { attachment: MessageAttachment }) { + if (attachment.previewUrl) + return ( + + + + ); + return {attachment.kind === "audio" ? : }; +} + +function UserTurn({ message }: { message: MessageType }) { + const text = textOf(message); + return ( + + + {message.attachments && message.attachments.length > 0 && ( + + {message.attachments.map((attachment, index) => ( + + + + {attachment.name} + {formatBytes(attachment.size)} + + + ))} + + )} + {message.mcpToolCall && ( + + + + + Called {message.mcpToolCall.tool} + + 
+                {JSON.stringify(message.mcpToolCall.arguments, null, 2)}
+
+ + + )} + {text && ( + + {text} + + )} + + + ); +} + +export type LastReplyExtras = Readonly<{ + endpoint: EndpointType; + mcpEvents: readonly MCPEvent[]; + codeInterpreter: CodeInterpreterResult | null; + accessToken: string; +}>; + +function AssistantReply({ + message, + isLast, + extras, + inspected, + onInspect, +}: { + message: MessageType; + isLast: boolean; + extras: LastReplyExtras; + inspected: boolean; + onInspect: () => void; +}) { + const model = replyModel(message); + const showMcpEvents = isLast && extras.mcpEvents.length > 0 && MCP_EVENT_ENDPOINTS.includes(extras.endpoint); + const showCode = isLast && extras.codeInterpreter && extras.endpoint === EndpointType.RESPONSES; + return ( + + + {model && ( + + + {fellBack(message) && } + + )} + {message.reasoningContent && } + {showMcpEvents && } + {message.searchResults && } + {showCode && extras.codeInterpreter && ( + + )} + {message.isError ? ( + + The request failed + {textOf(message)} + + ) : ( + + )} + +
+ {message.a2aMetadata ? ( + + ) : ( + + )} +
+ + {!message.isImage && !message.isAudio && } + {message.trace && ( + + + Inspect + + )} + + + + + ); +} + +function WaitingTurn({ target }: { target: string | undefined }) { + return ( + + +
+ + {target ? `Waiting for ${target}` : "Waiting for a reply"} +
+ + + ); +} + +function SwitchMarker({ model }: { model: string }) { + return ( + + + + + Switched to {model} + + ); +} + +function EmptyConversation({ + suggestions, + onPick, +}: { + suggestions: readonly string[]; + onPick: (text: string) => void; +}) { + return ( + + + + + + Send a test request + + Pick a model and anything the request should go through, then inspect what the gateway did with it + + + {suggestions.length > 0 && ( + + {suggestions.map((suggestion) => ( + + ))} + + )} + + ); +} + +const switchedModel = (messages: readonly MessageType[], index: number): string | null => { + const message = messages[index]; + if (message?.role !== "user" || !message.model) return null; + const previous = messages.slice(0, index).findLast((candidate) => candidate.role === "user"); + return previous?.model && previous.model !== message.model ? message.model : null; +}; + +export function Conversation({ + messages, + isLoading, + waitingFor, + extras, + inspectedIndex, + onInspect, + suggestions, + onPickSuggestion, +}: { + messages: readonly MessageType[]; + isLoading: boolean; + waitingFor: string | undefined; + extras: LastReplyExtras; + inspectedIndex: number | null; + onInspect: (index: number) => void; + suggestions: readonly string[]; + onPickSuggestion: (text: string) => void; +}) { + if (messages.length === 0) return ; + const lastIndex = messages.length - 1; + const waiting = isLoading && messages[lastIndex]?.role === "user"; + return ( + + + + + {messages.flatMap((message, index) => { + const switched = switchedModel(messages, index); + const turn = ( + + {message.role === "user" ? ( + + ) : ( + onInspect(index)} + /> + )} + + ); + return switched + ? [ + + + , + turn, + ] + : [turn]; + })} + {waiting && ( + + + + )} + + + + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Inspector.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Inspector.tsx new file mode 100644 index 00000000000..3a9919ce38f --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Inspector.tsx @@ -0,0 +1,292 @@ +"use client"; + +import { useState, type ReactNode } from "react"; +import { + ArrowDownToLine, + ArrowUpFromLine, + Brain, + Clock, + Coins, + DatabaseZap, + Gauge, + RotateCw, + Server, + Shuffle, + Timer, +} from "lucide-react"; +import type { MessageType } from "@/components/chat_ui/types"; +import { Accordion, AccordionContent, AccordionItem, AccordionTrigger } from "@/components/ui/accordion"; +import { Badge } from "@/components/ui/badge"; +import { Empty, EmptyDescription, EmptyHeader, EmptyTitle } from "@/components/ui/empty"; +import { FieldDescription } from "@/components/ui/field"; +import { Item, ItemActions, ItemContent, ItemDescription, ItemGroup, ItemTitle } from "@/components/ui/item"; +import { ScrollArea, ScrollBar } from "@/components/ui/scroll-area"; +import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs"; +import { ToggleGroup, ToggleGroupItem } from "@/components/ui/toggle-group"; +import { formatCost, formatSeconds, servedModel } from "./Conversation"; +import { CopyIconButton, FallbackBadge, LongValue } from "./ModelName"; + +export type CodeSdk = "openai" | "azure"; + +function TraceRow({ title, description, children }: { title: string; description?: string; children?: ReactNode }) { + return ( + + + {title} + {description && {description}} + + {children && {children}} + + ); +} + +type StatValue = Readonly<{ label: string; icon: ReactNode; value: string | undefined }>; + +function StatGrid({ stats }: { stats: readonly StatValue[] }) { + const shown = stats.filter((stat): stat is StatValue & { value: string } => stat.value !== undefined); + if (shown.length === 0) return null; + return ( + + {shown.map((stat) => ( + + + {stat.icon} + {stat.label} + + {stat.value} + + ))} + + ); +} + +function TraceSection({ value, title, children }: { value: string; title: string; children: ReactNode }) { + return ( + + {title} + + {children} + + + ); +} + +function Muted({ children }: { children: ReactNode }) { + return

{children}

; +} + +const count = (value: number | undefined): string | undefined => (value === undefined ? undefined : String(value)); +const millis = (value: number | undefined): string | undefined => + value === undefined ? undefined : `${Math.round(value)}ms`; + +function Routing({ message }: { message: MessageType }) { + const trace = message.trace; + const served = servedModel(message); + const fellBack = (trace?.attemptedFallbacks ?? 0) > 0; + const reported = Boolean(trace?.callId || trace?.modelGroup || trace?.providerModel); + return ( + + {trace?.requestedModel && ( + + + + )} + {served && ( + + {fellBack && } + + + )} + {trace?.providerModel && ( + + + + )} + {trace?.apiBase && ( + + + + )} + {trace?.callId && ( + + {trace.callId} + + + )} + {trace?.responseId && ( + + {trace.responseId} + + + )} + , value: count(trace?.attemptedFallbacks) }, + { label: "Retries", icon: , value: count(trace?.attemptedRetries) }, + { label: "Overhead", icon: , value: millis(trace?.overheadMs) }, + ]} + /> + {!reported && ( + + The gateway did not share routing details with this page. Open the dashboard from the proxy to see them + + )} + + ); +} + +const noneRan = (policies: readonly string[] | undefined): string | null => { + if (policies === undefined) return "No guardrails ran on this request"; + return policies.length === 0 ? "No guardrails or policies ran on this request" : null; +}; + +function Guardrails({ message }: { message: MessageType }) { + const guardrails = message.trace?.appliedGuardrails; + const policies = message.trace?.appliedPolicies; + if (guardrails === undefined) { + return ( + + This endpoint does not report which guardrails or policies ran + + ); + } + const emptyNote = guardrails.length === 0 ? noneRan(policies) : null; + return ( + + {guardrails.map((name) => ( + + Ran + + ))} + {policies?.map((name) => ( + + Ran + + ))} + {emptyNote && {emptyNote}} + {policies === undefined && Policies are only reported when the dashboard is opened from the proxy} + + ); +} + +function SpeedAndCost({ message }: { message: MessageType }) { + const usage = message.usage; + const cost = usage?.cost ?? message.trace?.cost; + return ( + + , + value: message.timeToFirstToken === undefined ? undefined : formatSeconds(message.timeToFirstToken), + }, + { + label: "Total", + icon: , + value: message.totalLatency === undefined ? undefined : formatSeconds(message.totalLatency), + }, + { label: "Cost", icon: , value: cost === undefined ? undefined : formatCost(cost) }, + { label: "Tokens in", icon: , value: count(usage?.promptTokens) }, + { label: "Tokens out", icon: , value: count(usage?.completionTokens) }, + { label: "Cached", icon: , value: count(usage?.cacheReadTokens) }, + { label: "Reasoning", icon: , value: count(usage?.reasoningTokens) }, + { label: "Gateway time", icon: , value: millis(message.trace?.responseMs) }, + ]} + /> + + ); +} + +function NothingYet() { + return ( + + + Nothing to inspect yet + + Each reply shows here with the deployment that served it, fallbacks, guardrails, timing and cost + + + + ); +} + +function Trace({ message }: { message: MessageType | null }) { + if (!message) return ; + return ( + + + + + + ); +} + +function CodeTab({ code }: { code: (sdk: CodeSdk) => string }) { + const [sdk, setSdk] = useState("openai"); + const snippet = code(sdk); + return ( +
+
+ value[0] && setSdk(value[0] as CodeSdk)} + variant="outline" + size="sm" + spacing={0} + aria-label="SDK" + > + + OpenAI SDK + + + Azure SDK + + + + + +
+ + Send this request from your own app. Replace YOUR_LITELLM_API_KEY with a virtual key + + + 
{snippet}
+ + +
+ ); +} + +export function InspectorPanel({ message, code }: { message: MessageType | null; code: (sdk: CodeSdk) => string }) { + return ( + +
+ + Trace + Code + +
+ + + + + + + + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/McpDialog.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/McpDialog.tsx new file mode 100644 index 00000000000..e5e0828825d --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/McpDialog.tsx @@ -0,0 +1,391 @@ +"use client"; + +import { useId, useState, type ReactNode } from "react"; +import { Blocks, KeyRound, Search } from "lucide-react"; +import { Badge } from "@/components/ui/badge"; +import { Button } from "@/components/ui/button"; +import { Checkbox } from "@/components/ui/checkbox"; +import { + Dialog, + DialogClose, + DialogContent, + DialogDescription, + DialogFooter, + DialogHeader, + DialogTitle, + DialogTrigger, +} from "@/components/ui/dialog"; +import { Empty, EmptyContent, EmptyDescription, EmptyHeader, EmptyTitle } from "@/components/ui/empty"; +import { Field, FieldLabel, FieldLegend, FieldSet } from "@/components/ui/field"; +import { InputGroup, InputGroupAddon, InputGroupButton, InputGroupInput } from "@/components/ui/input-group"; +import { ScrollArea } from "@/components/ui/scroll-area"; +import { Spinner } from "@/components/ui/spinner"; +import type { MCPServer, MCPToolset } from "@/components/mcp_tools/types"; +import { uiHref } from "@/utils/uiHref"; +import { ByokConnect } from "./ByokConnect"; +import { + ALL_MCP_SERVERS, + TOOLSET_PREFIX, + allowedTools, + toggleMcpItem, + toggleToolLimit, + withoutStaleLimits, + type ToolLimits, +} from "./playgroundSettings"; +import { useServerTools } from "./usePlaygroundCatalog"; + +type Row = Readonly<{ value: string; label: string }>; + +export const serverName = (server: MCPServer): string => server.alias || server.server_name || server.server_id; + +const matches = (value: string, query: string): boolean => value.toLowerCase().includes(query.trim().toLowerCase()); + +const needsKey = (server: MCPServer | undefined): boolean => Boolean(server?.is_byok && !server.has_user_credential); + +function SearchField({ + value, + onChange, + placeholder, + autoFocus, +}: { + value: string; + onChange: (value: string) => void; + placeholder: string; + autoFocus: boolean; +}) { + return ( + + + + + onChange(event.target.value)} + placeholder={placeholder} + aria-label={placeholder} + autoFocus={autoFocus} + /> + + ); +} + +function MasterRow({ + row, + checked, + onCheckedChange, + active, + onSelect, + summary, +}: { + row: Row; + checked: boolean; + onCheckedChange: (checked: boolean) => void; + active: boolean; + onSelect: () => void; + summary?: string; +}) { + return ( +
+ + +
+ ); +} + +function ToolRow({ + label, + checked, + disabled, + onCheckedChange, +}: { + label: string; + checked: boolean; + disabled: boolean; + onCheckedChange: (checked: boolean) => void; +}) { + const id = useId(); + return ( + + + + {label} + + + ); +} + +function Note({ title, children, action }: { title: string; children: ReactNode; action?: ReactNode }) { + return ( + + + {title} + {children} + + {action && {action}} + + ); +} + +function RowGroup({ title, rows, render }: { title: string; rows: readonly Row[]; render: (row: Row) => ReactNode }) { + if (rows.length === 0) return null; + return ( +
+ + {title} + + {rows.map(render)} +
+ ); +} + +export function McpDialog({ + requestKey, + servers, + toolsets, + isLoading, + selection, + toolLimits, + onChange, + onCredentialSaved, + defaultOpen = false, +}: { + requestKey: string | null; + servers: readonly MCPServer[]; + toolsets: readonly MCPToolset[]; + isLoading: boolean; + selection: readonly string[]; + toolLimits: ToolLimits; + onChange: (selection: readonly string[], limits: ToolLimits) => void; + onCredentialSaved: () => void; + defaultOpen?: boolean; +}) { + const [focused, setFocused] = useState(selection.find((item) => item !== ALL_MCP_SERVERS) ?? ALL_MCP_SERVERS); + const [query, setQuery] = useState(""); + const [toolQuery, setToolQuery] = useState(""); + const [replacingKey, setReplacingKey] = useState(null); + const serverById = new Map(servers.map((server) => [server.server_id, server])); + const selectedServers = selection.filter((item) => serverById.has(item)); + const toolsByServer = useServerTools( + requestKey, + [...new Set([...selectedServers, ...(serverById.has(focused) ? [focused] : [])])].filter( + (id) => !needsKey(serverById.get(id)), + ), + ); + const toolNames = (serverId: string): readonly string[] => + (toolsByServer[serverId]?.tools ?? []).map((tool) => tool.name); + + const serverRows: Row[] = servers.map((server) => ({ value: server.server_id, label: serverName(server) })); + const toolsetRows: Row[] = toolsets.map((toolset) => ({ + value: `${TOOLSET_PREFIX}${toolset.toolset_id}`, + label: toolset.toolset_name, + })); + const visible = (rows: readonly Row[]) => rows.filter((row) => matches(row.label, query)); + + const toggle = (item: string, on: boolean) => { + const next = toggleMcpItem(selection, item, on); + onChange(next, withoutStaleLimits(toolLimits, next)); + if (on) setFocused(item); + }; + + const summary = (row: Row): string | undefined => { + if (!selection.includes(row.value)) return undefined; + const server = serverById.get(row.value); + if (!server) return undefined; + if (needsKey(server)) return "Needs your key"; + const tools = toolNames(row.value); + if (tools.length === 0) return undefined; + const allowed = allowedTools(toolLimits, row.value, tools).length; + return allowed === tools.length ? `All ${tools.length} tools` : `${allowed} of ${tools.length} tools`; + }; + + const focus = (item: string) => { + setFocused(item); + setToolQuery(""); + setReplacingKey(null); + }; + + const masterRow = (row: Row) => ( + toggle(row.value, on)} + active={row.value === focused} + onSelect={() => focus(row.value)} + summary={summary(row)} + /> + ); + + const detail = (): ReactNode => { + if (focused === ALL_MCP_SERVERS) + return Every server this key can reach, with all of its tools; + const toolset = toolsets.find((item) => `${TOOLSET_PREFIX}${item.toolset_id}` === focused); + if (toolset) + return ( + +
+

+ {toolset.description || "A toolset's tools are fixed by whoever created it"} +

+
+ {toolset.tools.map((tool) => ( + + {tool.tool_name} + + ))} +
+
+ + ); + const server = serverById.get(focused); + if (!server) return Pick a server or toolset on the left; + const name = serverName(server); + if (needsKey(server) || replacingKey === focused) + return ( + { + setReplacingKey(null); + onCredentialSaved(); + }} + onCancel={needsKey(server) ? undefined : () => setReplacingKey(null)} + /> + ); + const replaceKey = server.is_byok ? ( + + ) : undefined; + if (!selection.includes(focused)) + return ( + + Check this server to choose which of its tools the model can call + + ); + const state = toolsByServer[focused]; + if (!state || state.isLoading) + return ( +
+ +
+ ); + if (state.failed) + return ( + + Couldn't load this server's tools + + ); + const tools = toolNames(focused); + if (tools.length === 0) return This server has no tools; + const allowed = allowedTools(toolLimits, focused, tools); + return ( +
+
+ + + {replaceKey} +
+ +
+ {tools + .filter((tool) => matches(tool, toolQuery)) + .map((tool) => ( + + onChange(selection, toggleToolLimit(toolLimits, { server: focused, tools }, tool, on)) + } + /> + ))} +
+ +
+ ); + }; + + const nothingAvailable = !isLoading && servers.length === 0 && toolsets.length === 0; + + return ( + !open && setReplacingKey(null)}> + }> + + MCP + {selection.length > 0 && ( + + {selection.length} + + )} + + + + MCP servers + + Pick the servers the model can use, then limit their tools. A server allows every tool until you uncheck + some. + + + {nothingAvailable ? ( +
+ + Add one on the MCP servers page, or send as a key that can reach one + +
+ ) : ( +
+
+
+ +
+ + {isLoading ? ( +
+ +
+ ) : ( +
+ + +
+ )} + +
+
{detail()}
+
+ )} + + }>Done + + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ModelName.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ModelName.tsx new file mode 100644 index 00000000000..bbea909711a --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ModelName.tsx @@ -0,0 +1,74 @@ +"use client"; + +import { Copy, Split } from "lucide-react"; +import { Badge } from "@/components/ui/badge"; +import { Button } from "@/components/ui/button"; +import { HoverCard, HoverCardContent, HoverCardTrigger } from "@/components/ui/hover-card"; +import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"; +import { cn } from "@/lib/cva.config"; +import { toast } from "@/lib/toast"; + +export function copyText(value: string, label: string): void { + void navigator.clipboard.writeText(value).then( + () => toast.success(`${label} copied`), + () => toast.error("Unable to copy to clipboard"), + ); +} + +export function CopyIconButton({ value, label }: { value: string; label: string }) { + return ( + + copyText(value, label)} /> + } + > + + + Copy {label} + + ); +} + +export function LongValue({ + value, + detail, + copyLabel, + className, + align = "start", +}: { + value: string; + detail: string; + copyLabel: string; + className?: string; + align?: "start" | "end"; +}) { + return ( + + } + > + {value} + + +
+
+ {detail} + {value} +
+ +
+ + + ); +} + +export function FallbackBadge() { + return ( + + + Fallback + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.test.tsx new file mode 100644 index 00000000000..92258ed1428 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.test.tsx @@ -0,0 +1,83 @@ +import { render, screen } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { describe, expect, it, vi } from "vitest"; +import { ChecklistMenu, TargetPicker, type OptionGroup } from "./Pickers"; + +const GROUPS: readonly OptionGroup[] = [ + { + label: "openai", + options: [ + { value: "gpt-fast", label: "gpt-fast" }, + { value: "gpt-smart", label: "gpt-smart" }, + ], + }, +]; + +const renderTargetPicker = (onValueChange = vi.fn()) => { + render( + , + ); + return onValueChange; +}; + +describe("TargetPicker", () => { + it("selects the first match when Enter is pressed after searching", async () => { + const user = userEvent.setup(); + const onValueChange = renderTargetPicker(); + + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.type(await screen.findByPlaceholderText("Search models"), "smart{Enter}"); + + expect(onValueChange).toHaveBeenCalledWith("gpt-smart"); + }); + + it("uses a typed model name that matches nothing when Enter is pressed", async () => { + const user = userEvent.setup(); + const onValueChange = renderTargetPicker(); + + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.type(await screen.findByPlaceholderText("Search models"), "my-custom-model{Enter}"); + + expect(onValueChange).toHaveBeenCalledWith("my-custom-model"); + }); +}); + +describe("ChecklistMenu", () => { + it("toggles the first match when Enter is pressed after searching", async () => { + const user = userEvent.setup(); + const onValueChange = vi.fn(); + render( + , + ); + + await user.click(screen.getByRole("combobox", { name: "Guardrails" })); + await user.type(await screen.findByPlaceholderText("Search guardrails"), "tox{Enter}"); + + expect(onValueChange).toHaveBeenLastCalledWith(["pii-mask", "toxicity"]); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.tsx new file mode 100644 index 00000000000..28ce6fe2f4a --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/Pickers.tsx @@ -0,0 +1,193 @@ +"use client"; + +import { useState, type ReactNode } from "react"; +import { Plus } from "lucide-react"; +import { Badge } from "@/components/ui/badge"; +import { + Combobox, + ComboboxCollection, + ComboboxContent, + ComboboxEmpty, + ComboboxGroup, + ComboboxInput, + ComboboxItem, + ComboboxLabel, + ComboboxList, + ComboboxTrigger, +} from "@/components/ui/combobox"; +import { InputGroupButton } from "@/components/ui/input-group"; +import { Spinner } from "@/components/ui/spinner"; +import { cn } from "@/lib/cva.config"; +import type { PickerOption } from "./usePlaygroundCatalog"; + +export type OptionGroup = Readonly<{ label: string; options: readonly PickerOption[] }>; + +type GroupItems = Readonly<{ value: string; items: readonly string[] }>; + +const toItems = (groups: readonly OptionGroup[]): GroupItems[] => + groups + .filter((group) => group.options.length > 0) + .map((group) => ({ value: group.label, items: group.options.map((option) => option.value) })); + +const isListed = (index: ReadonlyMap, name: string): boolean => + index.has(name) || [...index.values()].some((option) => option.label === name); + +const optionIndex = (groups: readonly OptionGroup[]): ReadonlyMap => + new Map(groups.flatMap((group) => group.options.map((option) => [option.value, option] as const))); + +function GroupedOptions({ index, mono = false }: { index: ReadonlyMap; mono?: boolean }) { + return ( + + {(group: GroupItems) => ( + + {group.value} + + {(value: string) => { + const option = index.get(value); + if (!option) return ; + return ( + + + {option.label} + {option.description && ( + {option.description} + )} + + + ); + }} + + + )} + + ); +} + +function CustomItem({ value }: { value: string }) { + return ( + + + + Use {value} + + + ); +} + +function CountBadge({ count }: { count: number }) { + if (count === 0) return null; + return ( + + {count} + + ); +} + +export function ChecklistMenu({ + label, + icon, + groups, + value, + onValueChange, + isLoading, + emptyText, + defaultOpen = false, +}: { + label: string; + icon: ReactNode; + groups: readonly OptionGroup[]; + value: readonly string[]; + onValueChange: (value: readonly string[]) => void; + isLoading: boolean; + emptyText: string; + defaultOpen?: boolean; +}) { + const index = optionIndex(groups); + const items = toItems(groups); + return ( + onValueChange(next)} + itemToStringLabel={(item: string) => index.get(item)?.label ?? item} + defaultOpen={defaultOpen} + autoHighlight + > + }> + {icon} + {label} + + + + + + {isLoading ? : items.length === 0 && emptyText} + {!isLoading && items.length > 0 && "Nothing matches"} + + + + + ); +} + +export function TargetPicker({ + label, + placeholder, + groups, + value, + onValueChange, + isLoading, + emptyText, + allowCustom = false, + defaultOpen = false, +}: { + label: string; + placeholder: string; + groups: readonly OptionGroup[]; + value: string | null; + onValueChange: (value: string) => void; + isLoading: boolean; + emptyText: string; + allowCustom?: boolean; + defaultOpen?: boolean; +}) { + const [query, setQuery] = useState(""); + const index = optionIndex(groups); + const custom = query.trim(); + const offerCustom = allowCustom && custom !== "" && !isListed(index, custom); + const items = [...toItems(groups), ...(offerCustom ? [{ value: "Custom name", items: [custom] }] : [])]; + const unsetText = isLoading ? `Loading ${label.toLowerCase()}s` : placeholder; + return ( + next && onValueChange(next)} + inputValue={query} + onInputValueChange={setQuery} + onOpenChange={(open) => !open && setQuery("")} + itemToStringLabel={(item: string) => index.get(item)?.label ?? item} + defaultOpen={defaultOpen} + autoHighlight + > + }> + {isLoading && !value ? : null} + + {value ? index.get(value)?.label ?? value : unsetText} + + + + + + {items.length === 0 && !custom ? emptyText : `No matching ${label.toLowerCase()}`} + + + + + ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.integration.test.tsx new file mode 100644 index 00000000000..c80f47045c2 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.integration.test.tsx @@ -0,0 +1,658 @@ +import { act, fireEvent, screen, waitFor } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"; +import { renderWithProviders as render, testQueryClient } from "@/../tests/test-utils"; +import { makeOpenAIChatCompletionRequest } from "@/components/llm_calls/chat_completion"; +import { fetchAvailableModels } from "@/components/llm_calls/fetch_models"; +import { makeOpenAIResponsesRequest } from "@/components/llm_calls/responses_api"; +import { callMCPTool, fetchMCPServers, fetchMCPToolsets, listMCPTools, tagListCall } from "@/components/networking"; +import { setSecureItem } from "@/utils/secureStorage"; +import { fetchClient } from "@/lib/http/api"; +import { PlaygroundChat } from "./PlaygroundChat"; + +vi.mock("@/app/(dashboard)/hooks/useCan", () => ({ default: () => true })); + +vi.mock("@/lib/http/api", async (importOriginal) => { + const actual = await importOriginal(); + return { ...actual, fetchClient: { ...actual.fetchClient, POST: vi.fn() } }; +}); + +vi.mock("@/components/llm_calls/fetch_models", () => ({ fetchAvailableModels: vi.fn() })); + +vi.mock("@/components/llm_calls/chat_completion", () => ({ makeOpenAIChatCompletionRequest: vi.fn() })); + +vi.mock("@/components/llm_calls/responses_api", () => ({ makeOpenAIResponsesRequest: vi.fn() })); + +vi.mock("@/components/networking", () => ({ + tagListCall: vi.fn().mockResolvedValue({}), + vectorStoreListCall: vi.fn().mockResolvedValue({ data: [] }), + getGuardrailsList: vi.fn().mockResolvedValue({ guardrails: [{ guardrail_name: "pii-mask" }] }), + getPoliciesList: vi.fn().mockResolvedValue({ policies: [] }), + fetchMCPServers: vi.fn().mockResolvedValue([]), + fetchMCPToolsets: vi.fn().mockResolvedValue([]), + listMCPTools: vi.fn().mockResolvedValue({ tools: [] }), + callMCPTool: vi.fn(), + getGatewayBaseUrl: () => "http://gateway.test", + getGlobalLitellmHeaderName: () => "Authorization", +})); + +const chatRequest = vi.mocked(makeOpenAIChatCompletionRequest); + +const responsesRequest = vi.mocked(makeOpenAIResponsesRequest); + +const credentialPost = fetchClient.POST as unknown as ReturnType; + +type ChatArgs = Parameters; + +type ResponsesArgs = Parameters; + +const sentMessages = (call: number) => (chatRequest.mock.calls[call] as ChatArgs)[0]; + +const replyWith = + (text: string) => + async (...args: ChatArgs) => { + args[1](text, args[2]); + }; + +const preload = (values: Readonly>) => + Object.entries(values).forEach(([key, value]) => sessionStorage.setItem(key, value)); + +const deferred = () => { + const control: { release: () => void } = { release: () => undefined }; + const promise = new Promise((resolve) => { + control.release = resolve; + }); + return { promise, release: () => control.release() }; +}; + +const mcpTool = (name: string, required: readonly string[] = []) => ({ + name, + inputSchema: { + type: "object", + properties: Object.fromEntries(required.map((field) => [field, { type: "string" }])), + required: [...required], + }, +}); + +const settled = () => waitFor(() => expect(testQueryClient.isFetching()).toBe(0)); + +const renderChat = () => + render(); + +const send = async (user: ReturnType, text: string) => { + await user.type(screen.getByRole("textbox", { name: "Message" }), text); + await user.click(screen.getByRole("button", { name: "Send message" })); +}; + +describe("PlaygroundChat", () => { + beforeEach(() => { + vi.clearAllMocks(); + sessionStorage.clear(); + testQueryClient.clear(); + Element.prototype.scrollIntoView = vi.fn(); + chatRequest.mockReset(); + responsesRequest.mockReset(); + vi.mocked(callMCPTool).mockReset(); + credentialPost.mockReset().mockResolvedValue({ data: {}, response: new Response() }); + vi.mocked(fetchMCPServers).mockReset().mockResolvedValue([]); + vi.mocked(fetchMCPToolsets).mockReset().mockResolvedValue([]); + vi.mocked(listMCPTools).mockReset().mockResolvedValue({ tools: [] }); + vi.mocked(fetchAvailableModels).mockResolvedValue([ + { model_group: "gpt-fast", mode: "chat", providers: ["openai"] }, + ]); + }); + + afterEach(() => { + vi.unstubAllGlobals(); + }); + + it("sends the chosen model and guardrails with the session key", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + renderChat(); + + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.click(await screen.findByRole("option", { name: "gpt-fast" })); + await user.click(screen.getByRole("combobox", { name: "Guardrails" })); + await user.click(await screen.findByRole("option", { name: "pii-mask" })); + await user.keyboard("{Escape}"); + await send(user, "hello"); + + expect(await screen.findByText("Hi there")).toBeInTheDocument(); + const [messages, , model, apiKey, , , , , , , , guardrails] = chatRequest.mock.calls[0] as ChatArgs; + expect(messages).toEqual([{ role: "user", content: "hello" }]); + expect(model).toBe("gpt-fast"); + expect(apiKey).toBe("session-token"); + expect(guardrails).toEqual(["pii-mask"]); + }); + + it("keeps answered turns in the history but never resends a failed one", async () => { + const user = userEvent.setup(); + chatRequest + .mockImplementationOnce(replyWith("Hi there")) + .mockRejectedValueOnce(new Error("rate limited")) + .mockImplementationOnce(replyWith("Back again")); + renderChat(); + + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.click(await screen.findByRole("option", { name: "gpt-fast" })); + await send(user, "hello"); + expect(await screen.findByText("Hi there")).toBeInTheDocument(); + await send(user, "this one fails"); + expect(await screen.findByText("The request failed")).toBeInTheDocument(); + await send(user, "try again"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(3)); + expect(sentMessages(2)).toEqual([ + { role: "user", content: "hello" }, + { role: "assistant", content: "Hi there" }, + { role: "user", content: "try again" }, + ]); + }); + + it("sends every saved request setting with the chat request", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + vi.mocked(fetchMCPServers).mockResolvedValue([{ server_id: "srv-1", alias: "docs" }]); + preload({ + selectedModel: "gpt-fast", + selectedTags: JSON.stringify(["team-a"]), + selectedVectorStores: JSON.stringify(["vs-1"]), + selectedPolicies: JSON.stringify(["strict"]), + selectedMCPServers: JSON.stringify(["srv-1"]), + streamingEnabled: "false", + }); + setSecureItem("customHeaders", JSON.stringify([["x-team", "blue"]])); + renderChat(); + + await waitFor(() => expect(fetchMCPServers).toHaveBeenCalled()); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + const args = chatRequest.mock.calls[0] as ChatArgs; + expect(args[2]).toBe("gpt-fast"); + expect(args[4]).toEqual(["team-a"]); + expect(args[10]).toEqual(["vs-1"]); + expect(args[12]).toEqual(["strict"]); + expect(args[13]).toEqual(["srv-1"]); + expect(args[25]).toBe(false); + expect(args[26]).toMatchObject({ "x-team": "blue" }); + }); + + it("sends a custom model and tag that are prefixes of listed ones", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + vi.mocked(tagListCall).mockResolvedValueOnce({ + production: { name: "production", models: [], created_at: "", updated_at: "" }, + }); + renderChat(); + + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.type(await screen.findByPlaceholderText("Search models"), "gpt-f"); + await user.click(await screen.findByRole("option", { name: "Use gpt-f" })); + await user.click(screen.getByRole("button", { name: "Settings" })); + await user.type(await screen.findByLabelText("Tags"), "prod"); + await user.click(await screen.findByRole("option", { name: "Add prod" })); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + const args = chatRequest.mock.calls[0] as ChatArgs; + expect(args[2]).toBe("gpt-f"); + expect(args[4]).toEqual(["prod"]); + }); + + it("sends a temperature typed with two decimals", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + await user.click(screen.getByRole("button", { name: "Settings" })); + await user.click(await screen.findByRole("button", { name: "Advanced" })); + await user.type(screen.getByLabelText("Temperature value"), "0.05"); + await user.keyboard("{Escape}"); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + expect((chatRequest.mock.calls[0] as ChatArgs)[16]).toBe(0.05); + }); + + it("keeps the saved MCP selection when the server list cannot be fetched", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + vi.mocked(fetchMCPServers).mockRejectedValue(new Error("forbidden")); + preload({ selectedModel: "gpt-fast", selectedMCPServers: JSON.stringify(["srv-1"]) }); + renderChat(); + + await waitFor(() => expect(fetchMCPServers).toHaveBeenCalled()); + await settled(); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + expect((chatRequest.mock.calls[0] as ChatArgs)[13]).toEqual(["srv-1"]); + }); + + it("drops a saved model that the current key cannot use", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + preload({ selectedModel: "retired-model" }); + renderChat(); + + await waitFor(() => expect(fetchAvailableModels).toHaveBeenCalled()); + await waitFor(() => expect(screen.getByRole("combobox", { name: "Model" })).toHaveTextContent("Select a model")); + await user.click(screen.getByRole("combobox", { name: "Model" })); + await user.click(await screen.findByRole("option", { name: "gpt-fast" })); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + expect((chatRequest.mock.calls[0] as ChatArgs)[2]).toBe("gpt-fast"); + }); + + it("calls a toolset tool on the server that backs it", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPToolsets).mockResolvedValue([ + { toolset_id: "ts-1", toolset_name: "Research", tools: [{ server_id: "srv-2", tool_name: "search" }] }, + ]); + vi.mocked(listMCPTools).mockResolvedValue({ + tools: [{ name: "search", inputSchema: { type: "object", properties: {} } }], + }); + vi.mocked(callMCPTool).mockResolvedValue({ content: [{ type: "text", text: "found it" }] }); + preload({ endpointType: "mcp", selectedMCPServers: JSON.stringify(["toolset:ts-1"]) }); + renderChat(); + + await user.click(await screen.findByRole("combobox", { name: "Tool" })); + await user.click(await screen.findByRole("option", { name: /search/ })); + await user.click(screen.getByRole("button", { name: "Send message" })); + + await waitFor(() => expect(callMCPTool).toHaveBeenCalledTimes(1)); + expect(listMCPTools).toHaveBeenCalledWith("session-token", "srv-2"); + expect(listMCPTools).not.toHaveBeenCalledWith("session-token", "toolset:ts-1"); + expect(vi.mocked(callMCPTool).mock.calls[0]?.slice(1, 3)).toEqual(["srv-2", "search"]); + }); + + it("offers toolsets as targets and lists only the toolset's own tools", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPToolsets).mockResolvedValue([ + { + toolset_id: "ts-1", + toolset_name: "Research", + tools: [ + { server_id: "srv-a", tool_name: "search" }, + { server_id: "srv-b", tool_name: "fetch" }, + { server_id: "srv-b", tool_name: "search" }, + ], + }, + ]); + vi.mocked(listMCPTools).mockImplementation(async (_key: string, server: string) => ({ + tools: + server === "srv-a" ? [mcpTool("search"), mcpTool("fetch", ["a_arg"])] : [mcpTool("search"), mcpTool("fetch")], + })); + vi.mocked(callMCPTool).mockResolvedValue({ content: [{ type: "text", text: "fetched" }] }); + preload({ endpointType: "mcp" }); + renderChat(); + + await user.click(await screen.findByRole("combobox", { name: "Server" })); + await user.click(await screen.findByRole("option", { name: /Research/ })); + await settled(); + await user.click(screen.getByRole("combobox", { name: "Tool" })); + await waitFor(() => + expect(screen.getAllByRole("option").map((option) => option.textContent)).toEqual(["search", "fetch"]), + ); + await user.click(screen.getByRole("option", { name: "fetch" })); + await user.click(screen.getByRole("button", { name: "Send message" })); + + await waitFor(() => expect(callMCPTool).toHaveBeenCalledTimes(1)); + expect(vi.mocked(callMCPTool).mock.calls[0]?.slice(1, 4)).toEqual(["srv-b", "fetch", {}]); + }); + + it("says the tools could not be loaded when the server's tool list fails", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPServers).mockResolvedValue([{ server_id: "srv-1", alias: "docs" }]); + vi.mocked(listMCPTools).mockResolvedValue({ error: "server unreachable" }); + preload({ endpointType: "mcp", selectedMCPServers: JSON.stringify(["srv-1"]) }); + renderChat(); + + await settled(); + await user.click(screen.getByRole("combobox", { name: "Tool" })); + + expect(await screen.findByText("Unable to load tools for this server")).toBeInTheDocument(); + }); + + it("keeps the saved MCP selection when the toolset list cannot be fetched", async () => { + const user = userEvent.setup(); + chatRequest.mockImplementation(replyWith("Hi there")); + vi.mocked(fetchMCPServers).mockResolvedValue([{ server_id: "srv-1", alias: "docs" }]); + vi.mocked(fetchMCPToolsets).mockRejectedValue(new Error("database not connected")); + preload({ selectedModel: "gpt-fast", selectedMCPServers: JSON.stringify(["srv-1", "srv-gone"]) }); + renderChat(); + + await waitFor(() => expect(fetchMCPToolsets).toHaveBeenCalled()); + await settled(); + await send(user, "hello"); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(1)); + expect((chatRequest.mock.calls[0] as ChatArgs)[13]).toEqual(["srv-1", "srv-gone"]); + }); + + it("drops a saved model that is missing from an already loaded model list", async () => { + const first = renderChat(); + await waitFor(() => expect(fetchAvailableModels).toHaveBeenCalled()); + await settled(); + first.unmount(); + sessionStorage.setItem("selectedModel", "retired-model"); + + renderChat(); + + await waitFor(() => expect(screen.getByRole("combobox", { name: "Model" })).toHaveTextContent("Select a model")); + expect(fetchAvailableModels).toHaveBeenCalledTimes(1); + }); + + it("drops a request that is still running when a new chat starts", async () => { + const user = userEvent.setup(); + const first = deferred(); + chatRequest.mockImplementationOnce(async (...args: ChatArgs) => { + await first.promise; + args[1]("late reply", args[2]); + }); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + await send(user, "first"); + expect(await screen.findByRole("button", { name: "Stop" })).toBeInTheDocument(); + await user.click(screen.getByRole("button", { name: "New chat" })); + await act(async () => { + first.release(); + await chatRequest.mock.results[0]?.value; + }); + + expect(chatRequest).toHaveBeenCalledTimes(1); + expect(screen.queryByRole("button", { name: "Stop" })).not.toBeInTheDocument(); + expect(screen.queryByText("late reply")).not.toBeInTheDocument(); + }); + + it("keeps a stopped request's response id out of the next request", async () => { + const user = userEvent.setup(); + const first = deferred(); + responsesRequest + .mockImplementationOnce(async (...args: ResponsesArgs) => { + await first.promise; + args[15]?.("resp-stopped"); + }) + .mockImplementationOnce(async (...args: ResponsesArgs) => { + args[15]?.("resp-fresh"); + args[1]("assistant", "fresh reply", args[2]); + }) + .mockImplementationOnce(async (...args: ResponsesArgs) => { + args[1]("assistant", "third reply", args[2]); + }); + preload({ endpointType: "responses", selectedModel: "gpt-fast", useApiSessionManagement: "true" }); + renderChat(); + + await send(user, "first"); + await user.click(await screen.findByRole("button", { name: "Stop" })); + await send(user, "second"); + expect(await screen.findByText("fresh reply")).toBeInTheDocument(); + await act(async () => { + first.release(); + await responsesRequest.mock.results[0]?.value; + }); + await send(user, "third"); + + await waitFor(() => expect(responsesRequest).toHaveBeenCalledTimes(3)); + expect((responsesRequest.mock.calls[2] as ResponsesArgs)[14]).toBe("resp-fresh"); + }); + + it("ignores a stopped request that finishes after a newer one started", async () => { + const user = userEvent.setup(); + const first = deferred(); + const second = deferred(); + chatRequest + .mockImplementationOnce(async (...args: ChatArgs) => { + await first.promise; + args[1]("late reply", args[2]); + }) + .mockImplementationOnce(async (...args: ChatArgs) => { + await second.promise; + args[1]("fresh reply", args[2]); + }); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + await send(user, "first"); + await user.click(await screen.findByRole("button", { name: "Stop" })); + await send(user, "second"); + first.release(); + + await waitFor(() => expect(chatRequest).toHaveBeenCalledTimes(2)); + expect(screen.getByRole("button", { name: "Stop" })).toBeInTheDocument(); + expect(screen.queryByText("late reply")).not.toBeInTheDocument(); + second.release(); + expect(await screen.findByText("fresh reply")).toBeInTheDocument(); + expect(screen.queryByText("late reply")).not.toBeInTheDocument(); + }); + + it("only says no guardrails ran when the gateway reported its headers", async () => { + const user = userEvent.setup(); + vi.stubGlobal("matchMedia", (query: string) => ({ + matches: true, + media: query, + addEventListener: () => undefined, + removeEventListener: () => undefined, + })); + chatRequest + .mockImplementationOnce(replyWith("unreported reply")) + .mockImplementationOnce(async (...args: ChatArgs) => { + args[27]?.(new Headers({ "x-litellm-call-id": "call-reported" })); + args[1]("reported reply", args[2]); + }); + preload({ selectedModel: "gpt-fast", selectedGuardrails: JSON.stringify(["pii-mask"]) }); + renderChat(); + + await send(user, "first"); + expect(await screen.findByText("unreported reply")).toBeInTheDocument(); + expect(screen.getByText("This endpoint does not report which guardrails or policies ran")).toBeInTheDocument(); + expect(screen.queryByText("No guardrails or policies ran on this request")).not.toBeInTheDocument(); + + await send(user, "second"); + expect(await screen.findByText("call-reported")).toBeInTheDocument(); + expect(screen.getByText("No guardrails or policies ran on this request")).toBeInTheDocument(); + }); + + it("does not claim policies were skipped when the gateway's own headers are hidden", async () => { + const user = userEvent.setup(); + vi.stubGlobal("matchMedia", (query: string) => ({ + matches: true, + media: query, + addEventListener: () => undefined, + removeEventListener: () => undefined, + })); + chatRequest.mockImplementationOnce(async (...args: ChatArgs) => { + args[27]?.(new Headers({ "content-type": "application/json" })); + args[1]("cross-origin reply", args[2]); + }); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + await send(user, "hello"); + expect(await screen.findByText("cross-origin reply")).toBeInTheDocument(); + expect( + screen.getByText("Policies are only reported when the dashboard is opened from the proxy"), + ).toBeInTheDocument(); + expect(screen.queryByText(/policies ran on this request/)).not.toBeInTheDocument(); + }); + + it("keeps a stopped request's gateway details off the newer reply", async () => { + const user = userEvent.setup(); + vi.stubGlobal("matchMedia", (query: string) => ({ + matches: true, + media: query, + addEventListener: () => undefined, + removeEventListener: () => undefined, + })); + const first = deferred(); + chatRequest + .mockImplementationOnce(async (...args: ChatArgs) => { + args[27]?.(new Headers({ "x-litellm-call-id": "call-stopped" })); + await first.promise; + }) + .mockImplementationOnce(async (...args: ChatArgs) => { + args[27]?.(new Headers({ "x-litellm-call-id": "call-fresh" })); + args[1]("fresh reply", args[2]); + }); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + await send(user, "first"); + await user.click(await screen.findByRole("button", { name: "Stop" })); + await send(user, "second"); + expect(await screen.findByText("call-fresh")).toBeInTheDocument(); + await act(async () => { + first.release(); + await chatRequest.mock.results[0]?.value; + }); + + expect(screen.queryByRole("button", { name: "Stop" })).not.toBeInTheDocument(); + expect(screen.getByText("call-fresh")).toBeInTheDocument(); + expect(screen.queryByText("call-stopped")).not.toBeInTheDocument(); + }); + + it("sends once when Enter is pressed twice before the first request starts", async () => { + const user = userEvent.setup(); + const first = deferred(); + chatRequest.mockImplementation(async (...args: ChatArgs) => { + await first.promise; + args[1]("Hi there", args[2]); + }); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + const message = screen.getByRole("textbox", { name: "Message" }); + await user.type(message, "hello"); + fireEvent.keyDown(message, { key: "Enter" }); + fireEvent.keyDown(message, { key: "Enter" }); + + await waitFor(() => expect(chatRequest).toHaveBeenCalled()); + await act(async () => { + first.release(); + await chatRequest.mock.results[0]?.value; + }); + expect(chatRequest).toHaveBeenCalledTimes(1); + expect(screen.getAllByText("hello")).toHaveLength(1); + }); + + it("drops a chat attachment when the endpoint cannot use it", async () => { + const user = userEvent.setup(); + preload({ selectedModel: "gpt-fast" }); + renderChat(); + + fireEvent.drop(screen.getByRole("textbox", { name: "Message" }), { + dataTransfer: { files: [new File(["%PDF"], "notes.pdf", { type: "application/pdf" })] }, + }); + expect(await screen.findByRole("button", { name: "Remove notes.pdf" })).toBeInTheDocument(); + + await user.click(screen.getByRole("button", { name: "Endpoint" })); + await user.click(await screen.findByRole("menuitemradio", { name: "/v1/audio/transcriptions" })); + + await waitFor(() => + expect(screen.getByRole("button", { name: "Endpoint" })).toHaveTextContent("audio/transcriptions"), + ); + expect(screen.queryByRole("button", { name: "Remove notes.pdf" })).not.toBeInTheDocument(); + }); + + it("keeps a stopped request's code output off the newer reply", async () => { + const user = userEvent.setup(); + const first = deferred(); + responsesRequest + .mockImplementationOnce(async (...args: ResponsesArgs) => { + await first.promise; + args[18]?.({ code: "print('stopped run')", containerId: "c-1", annotations: [] }); + }) + .mockImplementationOnce(async (...args: ResponsesArgs) => { + args[1]("assistant", "fresh reply", args[2]); + }); + preload({ endpointType: "responses", selectedModel: "gpt-fast" }); + renderChat(); + + await send(user, "first"); + await user.click(await screen.findByRole("button", { name: "Stop" })); + await send(user, "second"); + expect(await screen.findByText("fresh reply")).toBeInTheDocument(); + await act(async () => { + first.release(); + await responsesRequest.mock.results[0]?.value; + }); + + expect(screen.queryByText("Python Code Executed")).not.toBeInTheDocument(); + }); + + it("lets a user connect a key for a BYOK server picked for a direct tool call", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPServers).mockResolvedValue([ + { server_id: "srv-byok", alias: "Linear", is_byok: true, has_user_credential: false }, + ]); + preload({ endpointType: "mcp", selectedMCPServers: JSON.stringify(["srv-byok"]) }); + renderChat(); + + await user.click(await screen.findByRole("button", { name: "Connect key" })); + await user.type(await screen.findByLabelText("Linear API key"), "lin-key-1"); + await user.click(screen.getByRole("button", { name: "Connect" })); + + await waitFor(() => expect(credentialPost).toHaveBeenCalledTimes(1)); + expect(credentialPost.mock.calls[0]?.[1]).toMatchObject({ + params: { path: { server_id: "srv-byok" } }, + body: { credential: "lin-key-1" }, + }); + await waitFor(() => expect(fetchMCPServers).toHaveBeenCalledTimes(2)); + }); + + it("lets a user replace the key of a connected BYOK server whose tools fail to load", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPServers).mockResolvedValue([ + { server_id: "srv-byok", alias: "Linear", is_byok: true, has_user_credential: true }, + ]); + vi.mocked(listMCPTools).mockResolvedValue({ error: "401 from upstream" }); + preload({ selectedModel: "gpt-fast", selectedMCPServers: JSON.stringify(["srv-byok"]) }); + renderChat(); + + await settled(); + await user.click(screen.getByRole("button", { name: "MCP" })); + await user.click(await screen.findByRole("button", { name: "Replace key" })); + await user.type(await screen.findByLabelText("Linear API key"), "lin-key-2"); + await user.click(screen.getByRole("button", { name: "Connect" })); + + await waitFor(() => expect(credentialPost).toHaveBeenCalledTimes(1)); + expect(credentialPost.mock.calls[0]?.[1]).toMatchObject({ body: { credential: "lin-key-2" } }); + await waitFor(() => + expect(vi.mocked(listMCPTools).mock.calls.filter(([, server]) => server === "srv-byok").length).toBeGreaterThan( + 1, + ), + ); + }); + + it("leaves replace-key mode on cancel or when the MCP dialog closes", async () => { + const user = userEvent.setup(); + vi.mocked(fetchMCPServers).mockResolvedValue([ + { server_id: "srv-byok", alias: "Linear", is_byok: true, has_user_credential: true }, + ]); + vi.mocked(listMCPTools).mockResolvedValue({ error: "401 from upstream" }); + preload({ selectedModel: "gpt-fast", selectedMCPServers: JSON.stringify(["srv-byok"]) }); + renderChat(); + + await settled(); + await user.click(screen.getByRole("button", { name: "MCP" })); + await user.click(await screen.findByRole("button", { name: "Replace key" })); + expect(await screen.findByLabelText("Linear API key")).toBeInTheDocument(); + await user.click(screen.getByRole("button", { name: "Cancel" })); + expect(screen.queryByLabelText("Linear API key")).not.toBeInTheDocument(); + + await user.click(screen.getByRole("button", { name: "Replace key" })); + expect(await screen.findByLabelText("Linear API key")).toBeInTheDocument(); + await user.keyboard("{Escape}"); + await waitFor(() => expect(screen.queryByLabelText("Linear API key")).not.toBeInTheDocument()); + await user.click(screen.getByRole("button", { name: "MCP" })); + + expect(await screen.findByRole("button", { name: "Replace key" })).toBeInTheDocument(); + expect(screen.queryByLabelText("Linear API key")).not.toBeInTheDocument(); + expect(credentialPost).not.toHaveBeenCalled(); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.tsx new file mode 100644 index 00000000000..5b2de80315d --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/PlaygroundChat.tsx @@ -0,0 +1,489 @@ +"use client"; + +import { useRef, useState } from "react"; +import { ChevronDown, PanelRight, RotateCcw, SquarePen } from "lucide-react"; +import useCan from "@/app/(dashboard)/hooks/useCan"; +import { generateCodeSnippet } from "@/components/chat_ui/CodeSnippets"; +import { EndpointType, getEndpointType } from "@/components/chat_ui/mode_endpoint_mapping"; +import type { MessageType } from "@/components/chat_ui/types"; +import type { ModelGroup } from "@/components/llm_calls/fetch_models"; +import { customHeadersFromPairs } from "@/components/llm_calls/request_headers"; +import type { MCPToolArgumentsFormRef } from "@/components/mcp_tools/MCPToolArgumentsForm"; +import type { MCPServer, MCPTool, MCPToolset } from "@/components/mcp_tools/types"; +import { Button } from "@/components/ui/button"; +import { ButtonGroup } from "@/components/ui/button-group"; +import { + DropdownMenu, + DropdownMenuContent, + DropdownMenuItem, + DropdownMenuTrigger, +} from "@/components/ui/dropdown-menu"; +import { Sheet, SheetContent, SheetTitle } from "@/components/ui/sheet"; +import { toast } from "@/lib/toast"; +import { useChatHistory } from "../../hooks/useChatHistory"; +import { useCodeInterpreter } from "../../hooks/useCodeInterpreter"; +import { filterModelsForEndpoint, isModelCompatibleWithEndpoint } from "../chat_ui/EndpointUtils"; +import RealtimePlayground from "../chat_ui/RealtimePlayground"; +import { Composer } from "./Composer"; +import { EndpointMenu, VoiceMenu } from "./ComposerControls"; +import { Conversation, formatCost } from "./Conversation"; +import { capabilitiesFor, CODE_INTERPRETER_SUGGESTIONS, SUGGESTIONS } from "./endpointCapabilities"; +import { InspectorPanel, type CodeSdk } from "./Inspector"; +import { ByokKeyButton } from "./ByokConnect"; +import { McpDialog, serverName } from "./McpDialog"; +import { TargetPicker, type OptionGroup } from "./Pickers"; +import { TOOLSET_PREFIX, withEndpoint, withReachableMcp } from "./playgroundSettings"; +import { SettingsPopover } from "./SettingsPopover"; +import { + useGuardrailOptions, + useMcpCatalog, + usePlaygroundAgents, + usePlaygroundModels, + usePolicyOptions, + useServerTools, + useTagOptions, + useVectorStoreOptions, + type PickerOption, + type ServerTools, +} from "./usePlaygroundCatalog"; +import { usePlaygroundSend, type SendDeps } from "./usePlaygroundSend"; +import { usePlaygroundSettings } from "./usePlaygroundSettings"; +import { useIsWide } from "./useIsWide"; + +export type ProxySettings = Readonly<{ PROXY_BASE_URL?: string; LITELLM_UI_API_DOC_BASE_URL?: string | null }>; + +const POLICY_PREFIX = "policy:"; + +const providerOf = (model: ModelGroup): string => model.providers?.[0] ?? "other"; + +const modelGroups = (models: readonly ModelGroup[]): readonly OptionGroup[] => + [...new Set(models.map(providerOf))].sort().map((provider) => ({ + label: provider, + options: models + .filter((model) => providerOf(model) === provider) + .map((model) => ({ value: model.model_group, label: model.model_group })), + })); + +const modelEmptyText = (failed: boolean, hasKey: boolean): string => { + if (failed) return "Unable to load models for this key"; + if (!hasKey) return "Enter a virtual key to load models"; + return "No models available for this endpoint"; +}; + +const reachableMcp = (servers: readonly MCPServer[], toolsets: readonly MCPToolset[]): ReadonlySet => + new Set([ + ...servers.map((server) => server.server_id), + ...toolsets.map((toolset) => `${TOOLSET_PREFIX}${toolset.toolset_id}`), + ]); + +const toolsetFor = (selection: string | null, toolsets: readonly MCPToolset[]): MCPToolset | null => + toolsets.find((toolset) => `${TOOLSET_PREFIX}${toolset.toolset_id}` === selection) ?? null; + +const backingServers = (selection: string | null, toolset: MCPToolset | null): readonly string[] => { + if (toolset) return [...new Set(toolset.tools.map((tool) => tool.server_id))]; + return selection && !selection.startsWith(TOOLSET_PREFIX) ? [selection] : []; +}; + +type DirectTarget = Readonly<{ server: string; tool: MCPTool }>; + +const toolsetTargets = (toolset: MCPToolset, loaded: Readonly>): DirectTarget[] => + toolset.tools + .flatMap((entry) => { + const tool = loaded[entry.server_id]?.tools.find((candidate) => candidate.name === entry.tool_name); + return tool ? [{ server: entry.server_id, tool }] : []; + }) + .filter((target, index, targets) => targets.findIndex((other) => other.tool.name === target.tool.name) === index); + +const directTargets = ( + servers: readonly string[], + toolset: MCPToolset | null, + loaded: Readonly>, +): readonly DirectTarget[] => { + if (toolset) return toolsetTargets(toolset, loaded); + return servers.flatMap((server) => (loaded[server]?.tools ?? []).map((tool) => ({ server, tool }))); +}; + +const prefixed = (options: readonly PickerOption[]): readonly PickerOption[] => + options.map((option) => ({ ...option, value: `${POLICY_PREFIX}${option.value}` })); + +const lastTracedReply = (messages: readonly MessageType[]): number => + messages.findLastIndex((message) => message.role === "assistant" && message.trace !== undefined); + +const totalCost = (messages: readonly MessageType[]): number => + messages.reduce((sum, message) => sum + (message.usage?.cost ?? message.trace?.cost ?? 0), 0); + +const conversationSummary = (messages: readonly MessageType[]): string => { + const requests = messages.filter((message) => message.role === "user").length; + if (requests === 0) return "New chat"; + const label = requests === 1 ? "1 request" : `${requests} requests`; + const cost = totalCost(messages); + return cost > 0 ? `${label}, ${formatCost(cost)}` : label; +}; + +export function PlaygroundChat({ + accessToken, + disabledPersonalKeyCreation, + proxySettings, +}: { + accessToken: string | null; + disabledPersonalKeyCreation: boolean; + proxySettings: ProxySettings | undefined; +}) { + const { settings: stored, update, reset } = usePlaygroundSettings(disabledPersonalKeyCreation); + const canViewPolicies = useCan("viewPolicies"); + const apiKey = (stored.keySource === "session" ? accessToken : stored.virtualKey.trim()) || null; + const mcp = useMcpCatalog(apiKey); + const settings = withReachableMcp(stored, mcp.isReady ? reachableMcp(mcp.servers, mcp.toolsets) : null); + const capabilities = capabilitiesFor(settings.endpoint); + const chat = useChatHistory({ simplified: false }); + const codeInterpreter = useCodeInterpreter(); + const toolArgsForm = useRef(null); + const [inspected, setInspected] = useState(null); + const [sheetOpen, setSheetOpen] = useState(false); + const isWide = useIsWide(); + + const models = usePlaygroundModels(apiKey); + const [checkedModels, setCheckedModels] = useState(undefined); + if (models.data && models.data !== checkedModels) { + setCheckedModels(models.data); + const loaded = models.data; + update((current) => + current.model && !loaded.some((option) => option.model_group === current.model) + ? { ...current, model: null } + : current, + ); + } + const agents = usePlaygroundAgents(apiKey, settings.proxyBaseUrl, settings.endpoint === EndpointType.A2A_AGENTS); + const guardrails = useGuardrailOptions(accessToken); + const policies = usePolicyOptions(accessToken, canViewPolicies); + const vectorStores = useVectorStoreOptions(accessToken); + const tags = useTagOptions(accessToken); + + const directSelection = settings.endpoint === EndpointType.MCP ? settings.mcpSelection[0] ?? null : null; + const directToolset = toolsetFor(directSelection, mcp.toolsets); + const directServers = backingServers(directSelection, directToolset); + const directServerTools = useServerTools(apiKey, directServers); + const directByok = directToolset ? undefined : mcp.servers.find((server) => server.server_id === directSelection); + const targets = directTargets(directServers, directToolset, directServerTools); + const directTarget = targets.find((target) => target.tool.name === settings.directTool) ?? null; + const directTool = directTarget?.tool ?? null; + const toolsFailed = directServers.some((server) => directServerTools[server]?.failed); + const customHeaders = customHeadersFromPairs(settings.headerPairs); + + const sendDeps: SendDeps = { + settings, + capabilities, + apiKey, + signedIn: Boolean(accessToken), + customHeaders, + mcpServers: mcp.servers, + mcpToolsets: mcp.toolsets, + directCall: () => (directTarget ? { serverId: directTarget.server, tool: directTarget.tool.name } : null), + toolArgsForm, + chat, + codeInterpreter, + }; + const draft = usePlaygroundSend(sendDeps); + + const messages = chat.chatHistory; + const inspectedIndex = inspected ?? lastTracedReply(messages); + const inspectedMessage = inspectedIndex >= 0 ? messages[inspectedIndex] ?? null : null; + const modelList = models.data ?? []; + + const selectModel = (model: string) => { + const info = modelList.find((option) => option.model_group === model); + const switchTo = + info?.mode && !isModelCompatibleWithEndpoint(info, settings.endpoint) ? getEndpointType(info.mode) : null; + update((current) => ({ ...(switchTo ? withEndpoint(current, switchTo) : current), model })); + }; + + const inspect = (index: number) => { + setInspected(index); + if (!isWide) setSheetOpen(true); + }; + + const clearChat = () => { + chat.clearChatHistory(); + draft.clearDraft(); + codeInterpreter.clearResult(); + setInspected(null); + }; + + const newChat = () => { + clearChat(); + toast.success("Chat history cleared."); + }; + + const resetAll = () => { + clearChat(); + reset(); + toast.success("Chat and settings reset"); + }; + + const code = (sdk: CodeSdk): string => { + const snippet: Parameters[0] = { + apiKeySource: "custom", + accessToken: null, + apiKey: "", + inputMessage: draft.text, + chatHistory: messages, + selectedTags: [...settings.tags], + selectedVectorStores: [...settings.vectorStores], + selectedGuardrails: [...settings.guardrails], + selectedPolicies: [...settings.policies], + selectedMCPServers: [...settings.mcpSelection], + mcpServers: [...mcp.servers], + mcpServerToolRestrictions: Object.fromEntries( + Object.entries(settings.toolLimits).map(([server, tools]) => [server, [...tools]]), + ), + endpointType: settings.endpoint, + selectedModel: settings.model ?? undefined, + selectedSdk: sdk, + selectedVoice: settings.voice, + proxySettings, + customHeaders, + }; + return generateCodeSnippet(snippet); + }; + + const targetRow = ( + <> + {capabilities.target === "model" && ( + + )} + {capabilities.target === "agent" && ( + ({ + value: agent.agent_name, + label: agent.agent_name || agent.agent_id, + description: agent.agent_card_params?.description, + })), + }, + ]} + value={settings.agent} + onValueChange={(agent) => update({ agent })} + isLoading={agents.isLoading} + emptyText="No agents available for this key" + /> + )} + {capabilities.target === "mcpTool" && ( + <> + ({ value: server.server_id, label: serverName(server) })), + }, + { + label: "Toolsets", + options: mcp.toolsets.map((toolset) => ({ + value: `${TOOLSET_PREFIX}${toolset.toolset_id}`, + label: toolset.toolset_name, + description: toolset.description, + })), + }, + ]} + value={directSelection} + onValueChange={(server) => update({ mcpSelection: [server], directTool: null })} + isLoading={mcp.isLoading} + emptyText="No MCP servers available for this key" + /> + {directByok?.is_byok && ( + + )} + {directSelection && ( + ({ + value: tool.name, + label: tool.name, + description: tool.description, + })), + }, + ]} + value={settings.directTool} + onValueChange={(tool) => update({ directTool: tool })} + isLoading={directServers.some((server) => directServerTools[server]?.isLoading)} + emptyText={toolsFailed ? "Unable to load tools for this server" : "This server has no tools"} + /> + )} + + )} + update((current) => withEndpoint(current, endpoint))} + /> + {capabilities.voice && update({ voice })} />} + + ); + + const guardrailGroups: readonly OptionGroup[] = [ + { label: "Guardrails", options: guardrails.data ?? [] }, + ...(canViewPolicies ? [{ label: "Policies", options: prefixed(policies.data ?? []) }] : []), + ]; + + const composer = ( + update({ mcpSelection, toolLimits })} + onCredentialSaved={mcp.refreshCredentials} + /> + } + vectorStores={{ + groups: [{ label: "Vector stores", options: vectorStores.data ?? [] }], + value: settings.vectorStores, + onValueChange: (value) => update({ vectorStores: value }), + isLoading: vectorStores.isLoading, + }} + guardrails={{ + groups: guardrailGroups, + value: [...settings.guardrails, ...settings.policies.map((policy) => `${POLICY_PREFIX}${policy}`)], + onValueChange: (value) => + update({ + guardrails: value.filter((item) => !item.startsWith(POLICY_PREFIX)), + policies: value + .filter((item) => item.startsWith(POLICY_PREFIX)) + .map((item) => item.slice(POLICY_PREFIX.length)), + }), + isLoading: guardrails.isLoading || policies.isLoading, + }} + codeInterpreter={{ enabled: codeInterpreter.enabled, onToggle: codeInterpreter.setEnabled }} + settingsControl={ + + } + sendDisabled={draft.isLoading} + /> + ); + + const suggestions = + settings.endpoint === EndpointType.RESPONSES && codeInterpreter.enabled + ? CODE_INTERPRETER_SUGGESTIONS + : SUGGESTIONS[settings.endpoint] ?? []; + + const inspector = ; + + return ( +
+
+
+ {conversationSummary(messages)} + + {!isWide && ( + + )} + + + + } + > + + + + + + New chat, keep settings + + + + Reset chat and settings + + + + + +
+ {settings.endpoint === EndpointType.REALTIME ? ( +
+ 0 ? [...settings.guardrails] : undefined} + /> +
+ ) : ( + + )} +
{composer}
+
+ {isWide ? ( + + ) : ( + + + Inspector + {inspector} + + + )} +
+ ); +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ReplyContent.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ReplyContent.tsx new file mode 100644 index 00000000000..5733ede7168 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat/ReplyContent.tsx @@ -0,0 +1,92 @@ +"use client"; + +import ReactMarkdown from "react-markdown"; +import { Prism as SyntaxHighlighter } from "react-syntax-highlighter"; +import { coy } from "react-syntax-highlighter/dist/esm/styles/prism"; +import remarkGfm from "remark-gfm"; +import type { MessageType } from "@/components/chat_ui/types"; +import { ScrollArea, ScrollBar } from "@/components/ui/scroll-area"; +import { Table, TableBody, TableCell, TableHead, TableHeader, TableRow } from "@/components/ui/table"; +import { useSyntaxTheme } from "@/hooks/useSyntaxTheme"; +import AudioRenderer from "../chat_ui/AudioRenderer"; + +const textOf = (message: MessageType): string => (typeof message.content === "string" ? message.content : ""); + +function CodeBlock({ language, code }: { language: string; code: string }) { + const theme = useSyntaxTheme(coy); + return ( + + {code} + + ); +} + +export function Markdown({ children }: { children: string }) { + return ( +

, + ul: ({ node: _node, ...props }) =>