Skip to content

v1.33.0

Latest

Choose a tag to compare

@perminder-klair perminder-klair released this 18 Nov 14:35
a81c2b2

1.33.0 (2026-11-18)

Features

  • add Cheaper Inference as an OpenAI-compatible LLM provider (#2097) (3d2c92a)
  • llm: reasoning levels per leg (#2356) (76c344a), closes #2296
  • llm: request parameters and tool-call overrides for OpenAI-compatible legs (#2355) (ee27d73), closes #2296
  • search: add Parallel as a keyless web-search provider (#2278) (f449e8a)
  • search: SpecStream as a web-search backend (#2346) (8fd4285)
  • tagging: energy from the audio for LLM-tagged tracks: fill an empty one, overrule a clearly wrong one (both opt-in) (#2315) (207f92d)
  • transitions: cut a fade ending at its audible end, read off a loudness curve (#2332) (3c54c38)

Bug Fixes

  • analysis: Sound: line vector follow-ups from #2240 (#2354) (857e622), closes #2313
  • analyzer: a tail whose curve meter failed settles instead of re-reading every pass (#2360) (5e56a6f)
  • analyzer: drop diffq; a quantised DEMUCS_MODEL loads htdemucs (#2243) (62e7fdd), closes #2198
  • analyzer: fall back to cpu when the GPU can't run a kernel, and publish a cu124 analyzer build (#2311) (77ce5e0), closes #2280
  • cli: stop compiled binaries loading the working directory's .env (#2207) (07ba00b)
  • compose: run the controller and web under tini (init: true) (#2344) (50efadd)
  • liquidsoap: drop a coast request resolved from a replaced auto.m3u, and warn on an off-coast play (#2364) (2cd1c11), closes #2213
  • llm: Default reasoning on manual-thinking Claude thinks on a fixed 1,024-token budget (#2363) (d3bc501), closes #2357
  • stations: creating a second station no longer takes the AIO off the air (#2350) (dc68bac)
  • stems: a usage journal keeps the stored cache size true across a walk (#2351) (9d7b9ab), closes #2317
  • stems: a writer that dies mid-change stops the stored cache size being trusted (#2362) (c9c4523), closes #2352
  • transitions: the stem blend reads the outgoing tail through the projection (#2359) (e0676be)
  • tts: rescue each persona in its own voice when its engine fails (#2089) (9ad51b2)
  • ui: make Last.fm tags work without scrobbling obvious (#2316) (c382ac8)

Refactors

  • library: name the audio-energy pass for what it does (#2361) (06a49c4)