v1.1.1

1.1.1 (2026-10-15)

Bug Fixes

• a skill with huge dialog files can no longer crash classifier training (82d39dc)
• a skill with huge dialog files can no longer crash classifier training (c7af3f7)

v1.1.0

1.1.0 (2026-10-15)

Features

• collision_resolution "specific" (default) for template collisions (97906b2)
• make collision_resolution "specific" the default (59bc68a)
• optional collision_resolution "specific" for template collisions (62cde38)

v1.0.0

1.0.0 (2026-10-15)

Features

• answer OVOS readiness; first build skips the debounce (ce44dbc)
• answer OVOS readiness; first build skips the debounce (49b28e1)
• catch-all templates yield to a confident classifier (c7ec5ed)
• catch-all templates yield to a confident classifier (9bf1aca)
• exact: collision_priority, prefer core or user skills on template collisions (bb6f9b2)
• exact: collision_priority, prefer core or user skills on template collisions (290bb34)
• exact: skill_priority, an ordered list of skills that win collisions (0edcb05)
• experimental decider step, a local decision model in place of the LLM (168890e)
• experimental decider step, a local decision model in place of the LLM (c8bb9fe)
• float32 classifier features (the only model) (619861c)
• float32 classifier features (the only model) (930574d)
• high_risk "confirm" hands held requests to Doublecheck (a45857a)
• high_risk "confirm" hands held requests to Doublecheck (d35ef53)
• OVOS pipeline plugin with Hangup declines and shadow mode (6b4e1f0)
• router library lifted from the intent-eval harness (81a1a14)
• strip leading wake phrases; optional prefix step for trailing speech (e7c6e97)
• strip leading wake phrases; optional prefix step for trailing speech (ad7cdef)
• Switchboard as four OVOS pipeline steps (echo, exact, classifier, llm) (d0dbd0c)
• train the classifier in a child process (feba61f)
• train the classifier in a child process (f0c9e5d)

Bug Fixes

• a held high-risk intent is declined, not passed on (07fdce4)
• a held high-risk intent is declined, not passed on (adfc22e)
• confirm: hold through Hangup when Doublecheck isn't loaded (c91929d)
• decider: own inference code instead of a lint-excluded vendored file (d96feb2)
• decline "not for me" speech the LLM finds no tool for (88692d6)
• decline "not for me" speech the LLM finds no tool for (eeb315a)
• echo: Doublecheck's quote of the user is not echo (51c3bbd)
• hand OCP's own intents to OCP's pipeline step (21e3e00)
• hand OCP's own intents to OCP's pipeline step (952acb8)
• keep typed-slot templates when skipping malformed ones (173235f)
• live startup on ovos-core; cache the trained classifier (91d9449)
• live startup on ovos-core; cache the trained classifier (d70466b)
• llm: decline when the model answers with a tool it wasn't offered (8771268)
• llm: decline when the model answers with a tool it wasn't offered (fb416a1)
• llm: give unknown_tool declines the full details the router reads (9c87c5d)
• say "unmapped intent" when the classifier passes an unmapped top class (f4f412c)
• say "unmapped intent" when the classifier passes an unmapped top class (1bd1023)
• skip malformed templates instead of failing the whole build (de85cf7)
• skip malformed templates instead of failing the whole build (0a79ceb)
• template words end at word boundaries (a249387)
• template words end at word boundaries (b239240)

Performance Improvements

• decider: one packed row per question, 128 -> 80 ms per decision (77d8764)

Documentation

• confirm-flow design proposal (for review; nothing built) (f0585a5)
• confirm-flow design reflects what was built and decided (c0c1d5f)
• describe the evaluation without private household details (6249960)
• design for confirming high-risk actions (535e5ed)
• Doublecheck's question wording (658f5e7)
• dynamic schema proposal for routing arbitrary skill intents (74eb381)
• dynamic schema proposal for routing arbitrary skill intents (527230d)
• evaluation results, limits and configurations per situation (53f7f74)
• evaluation results, limits and configurations per situation (6bee852)
• evaluation: decider latency is ~130 ms with the packaged implementation (3da9d2a)
• evaluation: desktop numbers re-scored; decider and 1.7B on float32 (8b52f4c)
• evaluation: desktop numbers re-scored; decider and 1.7B on float32 (840fa16)
• evaluation: how it fails (doing nothing, not the wrong thing); decider config; recommendation (5f64d57)
• evaluation: measured overheard-speech rates; LLM on CPU not viable; MLX caveat (32d1144)
• fix list nesting in the confirm-flow design (ac5a8d5)
• known limitations; confirm is opt-in by design (f0e9e6f)
• known limitati...