feat: expressiveness mode, stateless Instructions, structured LLM output by theomonnom · Pull Request #5635 · livekit/agents

feat: expressiveness mode, stateless Instructions, structured LLM output

theomonnom

committed

fix: strip TTS markup from streamed transcripts when expressiveness is enabled

theomonnom

committed

feat: XML-aware sentence tokenizer + regression tests

theomonnom

committed

remove unused get_tags from _provider_format

theomonnom

committed

test: expand XML tokenizer coverage to 46 tests

theomonnom

committed

test: use full sentences inside wrapping tags to ensure merge is exercised

theomonnom

committed

fix: restore original pipeline instruction behavior, fix all CI test failures

theomonnom

committed

refactor: ExpressivenessOptions as TypedDict, stateless Instructions with render(), improved provider prompts

theomonnom

committed

feat: expression tag for bracket-based TTS providers (Inworld, ElevenLabs v3)

theomonnom

committed

feat: add Markup support to Inworld TTS plugin

theomonnom

committed

refactor: plugins override Markup inner class instead of manually setting self._markup

theomonnom

committed

refactor: audio_recognition property raises instead of returning None

theomonnom

committed

feat: comprehensive Inworld TTS 2 support

theomonnom

committed

refactor: trim provider prompts — keep syntax/values, cut redundant examples

theomonnom

committed

feat: better Inworld delivery examples showing full range of expressiveness

theomonnom

committed

example: switch drive-thru to Inworld TTS 2 with expressiveness

theomonnom

committed

feat: max_token_len on sentence tokenizer for TTS input limits

theomonnom

committed

fix: move markup conversion from token-level to sentence-level

theomonnom

committed

switch drive-thru to gpt-4.1-mini, don't truncate TTS debug logs

theomonnom

committed

example: soften drive-thru guardrails to not block expressiveness

theomonnom

committed

fix: richer Inworld delivery examples — short labels sound flat

theomonnom

committed

fix: XML regex was greedy — never detected self-closing tags with attributes

theomonnom

committed

test: regression tests for self-closing XML tag regex bug

theomonnom

committed

feat(tts): XML-aware sentence tokenizer, provider prompts, markup pipeline

theomonnom

and

Théo Monnom

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: expressiveness mode, stateless Instructions, structured LLM output#5635

feat: expressiveness mode, stateless Instructions, structured LLM output#5635
theomonnom wants to merge 24 commits intomainfrom
theo/expressiveness-mode

Commits on May 8, 2026

feat: expressiveness mode, stateless Instructions, structured LLM output

fix: strip TTS markup from streamed transcripts when expressiveness is enabled

feat: XML-aware sentence tokenizer + regression tests

remove unused get_tags from _provider_format

test: expand XML tokenizer coverage to 46 tests

test: use full sentences inside wrapping tags to ensure merge is exercised

fix: restore original pipeline instruction behavior, fix all CI test failures

refactor: ExpressivenessOptions as TypedDict, stateless Instructions with render(), improved provider prompts

feat: expression tag for bracket-based TTS providers (Inworld, ElevenLabs v3)

feat: add Markup support to Inworld TTS plugin

refactor: plugins override Markup inner class instead of manually setting self._markup

refactor: audio_recognition property raises instead of returning None

feat: comprehensive Inworld TTS 2 support

refactor: trim provider prompts — keep syntax/values, cut redundant examples

feat: better Inworld delivery examples showing full range of expressiveness

example: switch drive-thru to Inworld TTS 2 with expressiveness

feat: max_token_len on sentence tokenizer for TTS input limits

fix: move markup conversion from token-level to sentence-level

switch drive-thru to gpt-4.1-mini, don't truncate TTS debug logs

example: soften drive-thru guardrails to not block expressiveness

fix: richer Inworld delivery examples — short labels sound flat

fix: XML regex was greedy — never detected self-closing tags with attributes

test: regression tests for self-closing XML tag regex bug

feat(tts): XML-aware sentence tokenizer, provider prompts, markup pipeline