v0.10.0

0.10.0 (2026-09-29)

⚠ BREAKING CHANGES

• metal-agent: serve engines on loopback behind an authenticated ingress (#1945)

Features

• controller: front Metal InferenceServices with an authenticated relay (#1944) (37be11b)
• metal-agent: serve engines on loopback behind an authenticated ingress (#1945) (cf55427)
• metal-agent: tensorfold runtime for MLX speculative decoding (#1923) (8c9820c)

Bug Fixes

• api: let InferenceService select Metal runtimes and fall back to the agent flag (#1912) (19191b4)
• metal-agent: enforce a typed allowlist on extraArgs (#1941) (c09a4be)
• metal-agent: enforce allowed roots on local model paths (#1939) (bcdd27f)
• metal-agent: fail fast when llama-server exits during startup (#1917) (23f1fc6)
• metal-agent: fail fast when mlx-server exits during startup (#1929) (c55f535)
• metal-agent: fail fast when vllm-swift exits during startup (#1928) (46e4e96)
• metal-agent: load local-path GGUF sources in place instead of downloading (#1920) (6c045b4)
• metal-agent: never overwrite or delete a Service or EndpointSlice it does not own (#1932) (799194a)
• metal-agent: size file:// model sources in the memory estimate (#1930) (39b1ef6)
• metal-agent: spell mlock as --load-mode on llama.cpp 0.5.0+ (#1916) (f760639)
• metal: stop routing to a Metal endpoint nothing is serving (#1921) (75c792f)