Llama Stack 1.5.0

Minor release off the new release-1.5.x branch. 108 commits since v1.4.0.

Breaking changes

• Serve conversations and prompts with responses, or when explicitly listed (#6561)
• An explicit apis: list is now authoritative, even when empty (#6562)
• Raise the FastAPI floor to 0.137.2 and drop legacy route collection (#6629)
• Allow an empty choices list in the OpenAI legacy completions usage chunk (#6541)

New features

Inference and providers

• Add a Text-Embeddings-Inference inference provider (#6657)
• Add native /v1/messages passthrough for DeepSeek (#6677) and Fireworks AI (#6676)
• Support thinking for reasoning models on the Anthropic translation path (#6663)
• Refresh the Fireworks provider and drop the unused SDK pin (#6556)
• Support rerank models in llama-cpp-server (#6620), and classify its models via models.dev to match vLLM (#6619)
• Normalize the inference provider-data key by provider id (#6468)

Tools and search

• Add an Exa web search provider (#6643)
• Add a Serply web search provider (#6636)
• Support file_search ranking_options.hybrid_search in responses (#6576)
• Support the mcp 2.x client SDK (#6684)

Storage and files

• Add PostgreSQL connection URI support (#6690)
• Store S3 objects under an optional key prefix (#6639)

Fixes

45 fixes landed in this release across the server, providers, responses, and storage layers. See the full changelog for the complete list.

Full changelog: v1.4.0...v1.5.0