promptfoo Files

0.123.1 (2026-09-18)

Features

• bedrock: refresh auth across HTTP adapters (#10123) (9d3f063)
• providers: add Gemini 3.8 and Vertex Live (#10956) (1417f22)
• providers: add GPT-Live voice sessions (#10891) (316a334)
• providers: add OpenAI Agents API (#10892) (ad3bee3)
• providers: add portable HTTP config schemas (#10968) (fcd0604)
• providers: add portable MCP config schemas (#10965) (f5ccb3c)
• providers: support Ollama 0.34 features and refresh model docs (#10964) (ac01a4f)
• providers: surface Ollama thinking output and finish reason (#10954) (27d4c22)
• providers: update Gemini tools and media support (#10172) (627bdd0)

Bug Fixes

• assertions: bound embedded tool call parsing (#10877) (78202a3)
• assertions: bound tool-call parser state (#10887) (15290b4)
• assertions: default RAG assertion thresholds to 0.5 (#10202) (4831b60), closes #9910 #9848
• assertions: invert score along with pass for search-rubric (#10103) (26e8f2f)
• assertions: never invert grader failures on not-classifier and not-search-rubric (#10904) (e049702)
• assertions: reject malformed webhook results (#10955) (fc57736)
• assertions: reject non-string search rubrics (#10934) (7bbb79e)
• auth: harden custom header requests (#10524) (5325084)
• cli: handle large values during result export (#10884) (be798f4)
• cli: preserve in-memory results for extensions (#10885) (60483d9)
• code-scan: extractValidLineRanges off-by-one on trailing newline (#10107) (2cbabdb)
• config: preserve discovery in paths containing quotes (#10549) (d9b3927)
• db: preserve shared blobs after persistence errors (#10897) (3b353fb)
• deps: update dependency @anthropic-ai/sdk to v0.123.0 (#10901) (14f7c06)
• deps: update dependency @anthropic-ai/sdk to v0.124.0 (#10940) (29a15d1)
• deps: update engine.io to v6.6.10 (#10915) (e54a968)
• deps: update opentelemetry (#10886) (280bc5e)
• deps: update type definitions (#10924) (01c6397)
• drain database writes and preserve provider options (#10911) (32bfa9e)
• eval: keep concurrent derived metrics consistent (#10873) (2a3a3bd)
• eval: preserve audio grading results and order (#10822) (395d21e)
• eval: preserve cloud UUID config semantics (#7732) (8544fed)
• fetch: decode URL credentials and match Authorization case-insensitively (#10909) (b8aa6ff)
• fetch: treat credit_balance_exhausted as a hard quota error (#10881) (7f21bb2)
• honor SDK options and load AVIF/TIFF images (#10902) (15bba21)
• integrations: honor --env-file and config env for Helicone key and Langfuse base URL (#10942) (8bd9f37)
• integrations: read Langfuse env at fetch time (#10937) (033080c)
• integrations: share Langfuse prompt fetches and surface real SDK load errors (#10943) (379ecd0)
• matchers: tag context-faithfulness grader failures (#9907) (77bc3ba)
• matchers: tag RAG grader failures so inverse assertions cannot pass (#10494) (b3d21d2)
• mcp: accept fine-tuned model identifiers (#10847) (9442129)
• opencode: capture skill calls from full session history (#10011) (290973d)
• prompts: CSV rows collapse to duplicate labels/IDs when basePrompt.label is set (#10102) (25963c3)
• providers: align ElevenLabs requests and audio inputs (#10736) (de12341)
• providers: apply normalized Envoy base URL after user config spread (#10642) (1b6e793)
• providers: clear ElevenLabs request and stream timers (#10919) (23d245b)
• providers: forward Snowflake provider configuration (#10878) (7d0bacd)
• providers: handle billing quotas and retryable throttles (#10866) (16fdeb6)
• providers: honor Foundry overrides and bound fetch reads (#10948) (6b153ad)
• providers: honor Nova Sonic inference configuration (#10310) (aa2abb3)
• providers: honor prompt Gemini tool policy over passthrough (#10890) (2fd7add)
• providers: isolate Agents endpoint credentials (#10951) (f385d96)
• providers: migrate Ollama embeddings to /api/embed and honor config (#10958) (6c33d1f)
• providers: preserve default MCP enablement (#10661) (8a8a08b)
• providers: preserve HTTP transform errors (#10454) (a7e4ffc)
• providers: preserve patterned OpenCode skill history (#10971) (ba0eeb4)
• providers: preserve streamed Ollama tool calls (#10946) (7867eb2)
• providers: route OpenClaw default agent (#9046) (139267a)
• providers: stop clobbering resolved TogetherAI model params (#10846) (783d71f)
• providers: stop leaking local config into request bodies (#10845) (329b748)
• providers: surface Ollama HTTP errors instead of empty output (#10953) (c66b468)
• redteam: align redteam-ollama example with its README (#10959) (ccb4b5a)
• redteam: align tool discovery generation and grading (#10621) (d672144)
• redteam: ground agent skill plans in application evidence (#10903) (3c5429b)
• redteam: keep vulnerabilities when scoring fails (#10931) (37608b6)
• redteam: preserve strategy grading context (#10969) (08e91da)
• validators: default session source to client when provider.config is undefined (#10086) (62dbbc0)
• webui: handle malformed login redirects (#10460) (ebb1a6d)