Skip to content

v0.3.0

Choose a tag to compare

@github-actions github-actions released this 30 Oct 12:43
· 359 commits to main since this release

Changelog

  • 15d57d5: build(deps): bump the github-actions group with 6 updates (@fallais)
  • 7f1c91b: build(deps): bump the github-actions group with 9 updates (#129) (@dependabot[bot])
  • 56ff84a: build(deps): bump the go-modules group (@fallais)
  • 9b79b5e: build(deps): bump the go-modules group across 1 directory with 10 updates (#134) (@dependabot[bot])
  • e6ca64c: build(deps): bump the go-modules group across 1 directory with 11 updates (#130) (@dependabot[bot])
  • f9f639a: build(deps): move to go1.27.2 and x/net v0.60.0 (@fallais)
  • b13ea71: build(deps): move to go1.27.3 (@fallais)
  • 9ae02b5: chore(deps): bump Go modules, actions, tools and base images (@fallais)
  • 65d49a5: ci(security): add a Semgrep CE static analysis job (@fallais)
  • 24a7739: ci(security): leave the example servers' plain HTTP out of the Semgrep scan (@fallais)
  • 2fbb2ed: docs(changelog): drop the emoji from the breaking-change entries (@fallais)
  • 91df64b: docs(changelog): record LLMWithBackend, the backend worker and the fixes that came with them (@fallais)
  • 9d5cefb: docs(changelog): record backchannels, user-input, raw text, capabilities and bot-output fixes (@fallais)
  • 6cf3584: docs(changelog): record the engine, turn, service, transport and adapter fixes (@fallais)
  • a38c955: docs(changelog): record the provider fixes and the LMNT removal (@fallais)
  • abfe69a: docs(changelog): record the send timeout, observer cleanup, setup cancel and TTS fixes (@fallais)
  • 6e7684c: docs(changelog): record the speech, transport, RTVI, adapter and audio fixes (@fallais)
  • c1cd44e: docs(changelog): record the thinking defaults, Perplexity Agent API and connection fixes (@fallais)
  • 55108c2: docs(changelog): release 0.3.0 (@fallais)
  • 8286b10: docs(context): say why the sequencer has a lock without naming another project (@fallais)
  • 027a81c: feat(aggregators)!: answer a user turn that ends with no transcript (@fallais)
  • e9fb69c: feat(aggregators): announce each user turn written to the conversation (@fallais)
  • b2635f4: feat(aggregators): announce the turn an eager reply answers ahead of it (@fallais)
  • 2c8ae8d: feat(aggregators): run the turn's inference through the turn controller (@fallais)
  • 12ca129: feat(aggregators): tell whether a synchronous function call is in flight (@fallais)
  • 9b8077c: feat(anthropic): opt every request into interleaved thinking (@fallais)
  • 9cd476e: feat(anthropic): stream thinking blocks as thought frames (@fallais)
  • 318eedd: feat(assemblyai): add streaming text-to-speech (@fallais)
  • 472e4b7: feat(cartesia): pick the accent a multilingual voice speaks with (@fallais)
  • c3065e9: feat(classifier): answer typed questions about a state (@fallais)
  • 90314cf: feat(context): pass over the tags only the model's text carries (@fallais)
  • 4a027b3: feat(eval)!: judge with a classifier and explain with an LLM (@fallais)
  • 3bedb07: feat(eval)!: put a tool call to the judge (@fallais)
  • 217a222: feat(eval): check an event's text against regular expressions (@fallais)
  • 45ad2f1: feat(eval): check the expression the bot's TTS performed (@fallais)
  • 55b5a8a: feat(eval): keep the suite's concurrency in flight across entries (@fallais)
  • 47e9ec9: feat(eval): let a scenario forbid text as well as require it (@fallais)
  • 9034d87: feat(eval): let one file hold several scenarios (@fallais)
  • 149412c: feat(eval): name and cap a manifest entry, and spread the suite over them (@fallais)
  • 5758da1: feat(eval): record what each expectation matched (@fallais)
  • f94e468: feat(eval): report the markers a model emits, and assert on them (@fallais)
  • 7f7a1cc: feat(expressive): let a tagger after the LLM write the expression tags (@fallais)
  • 976e253: feat(expressive): let the LLM hear how the user spoke (@fallais)
  • f19a1c3: feat(expressive): read expression tags and format them for a TTS service (@fallais)
  • a184683: feat(frames): add backchannels, for the user and for the bot (@fallais)
  • 3829479: feat(frames): decide interruptibility per frame (@fallais)
  • 8d5397b: feat(frames): report function calls made outside the pipeline (@fallais)
  • bb0fd0c: feat(gemini): push thoughts and keep thought signatures (@fallais)
  • debb59f: feat(groq)!: take the reasoning control Groq adds to the request (@fallais)
  • 136a16b: feat(inception)!: default to mercury-3 (@fallais)
  • c483465: feat(jobcontext): let a job group keep every event its workers report (@fallais)
  • d06fa72: feat(live): carry a thinking configuration into the session setup (@fallais)
  • b7282a3: feat(live): hold the bot's turn while the model is still working (@fallais)
  • 729be71: feat(live): say whether the model should wait for a function call (@fallais)
  • a32db80: feat(llm)!: let a one-shot inference enforce a reply schema (@fallais)
  • c63bdc7: feat(llm): cancel every function call in flight, and let a service decline a pairing role (@fallais)
  • 479175c: feat(llm): let a model with no thinking configured run at its provider's default (@fallais)
  • f1abbbe: feat(llm): let an output gate drop what an LLM service pushes (@fallais)
  • bb28584: feat(llm): teach the LLM the expression tags its TTS performs (@fallais)
  • d805ce4: feat(llmwithbackend)!: restructure the pairing instruction so a stop is delegated (@fallais)
  • 8a32f46: feat(llmwithbackend): delegate a turn early when a classifier predicts it (@fallais)
  • e62a966: feat(llmwithbackend): have the frontend acknowledge a delegation once (@fallais)
  • 16b2366: feat(llmwithbackend): pair a conversational frontend LLM with a backend that does the work (@fallais)
  • 9dae782: feat(llmwithbackend): shorten the frontend's acknowledgement of a delegation (@fallais)
  • d458ee9: feat(llmworker): add a backend worker a frontend delegates its heavy work to (@fallais)
  • 57d6c00: feat(lmnt)!: remove the LMNT TTS provider (@fallais)
  • cc47ff8: feat(observers): tell an observer about a frame once, on its first push (@fallais)
  • e4eea56: feat(openai): run gpt-6.1-terra with reasoning effort none by default (@fallais)
  • b3a511a: feat(perplexity): add the Agent API service and deprecate Sonar (@fallais)
  • 31051ed: feat(pipeline): hand metadata from outside a switcher to its inactive services (@fallais)
  • 40d7c39: feat(processor): cancel the frame in progress without dropping the queue (@fallais)
  • 3537f33: feat(providers): carry seven settings that had no way through (@fallais)
  • 6b2976f: feat(rtvi)!: report the bot's capabilities in bot-ready (@fallais)
  • f78753f: feat(rtvi)!: report what the user said or typed as user-input (@fallais)
  • c0e3324: feat(rtvi): add the raw text messages, and deprecate the old text messages (@fallais)
  • dfe4531: feat(rtvi): report a tool call's arguments without its result (@fallais)
  • 279a7b1: feat(rtvi): report the expression the bot performed, protocol 2.3.0 (@fallais)
  • 9136a3a: feat(rtvi): say on bot-output what kind of text it holds as output_type (@fallais)
  • d4ff9ed: feat(smallest): add the content filter, date order and number language (@fallais)
  • f0887e4: feat(text)!: find sentence boundaries with language rules in the TTS language (@fallais)
  • bdef8a1: feat(together)!: add the reasoning toggle and the reasoning effort (@fallais)
  • 77b00ac: feat(tts): let a TTSSpeakFrame say what kind of text it speaks (@fallais)
  • 7717766: feat(tts): perform expression on Cartesia, ElevenLabs, Fish and MiniMax (@fallais)
  • f6de43b: feat(tts): perform expression on Inworld, Soniox, Mistral and Azure (@fallais)
  • 23db982: feat(tts): perform expression tags in an expressive conversation (@fallais)
  • 106cd1e: feat(tts): pronounce words from IPA in each provider's markup (@fallais)
  • 087fe22: feat(tts,wsservice): report token streaming and a reconnect in progress (@fallais)
  • 5bcbc76: feat(turns)!: decide on the LLM's turn-completion marker in the stop strategy (@fallais)
  • 6383470: feat(turns)!: decide user turns from every strategy's decision (@fallais)
  • e0e039f: feat(turns)!: report decisions from the start and stop strategies (@fallais)
  • e875545: feat(turns): add a user turn stop strategy that asks a classifier (@fallais)
  • 2d17e5a: feat(turns): add the user turn strategy modes, the merge modes and the response cancel frame (@fallais)
  • 0490360: feat(vad): let Silero analyzers share one model session (@Sud-Puth)
  • 56e35f7: feat(voicemail)!: decide with a classifier once the caller stops (@fallais)
  • 195670f: feat(xai): send the STT model and default to grok-voice-transcribe-2.0 (@fallais)
  • 8d86936: fix(adapter): copy native Anthropic messages and read stored thoughts (@fallais)
  • 1bdbd47: fix(adapter): copy native Gemini contents and apply kept thought signatures (@fallais)
  • edb2539: fix(adapter/gemini): wrap every non-object tool result and adapt tool schemas (@fallais)
  • b41d7c4: fix(aggregators): announce a stand-alone marker downstream instead of re-running the model (@fallais)
  • 5900b0d: fix(aggregators): hold an assistant-side append's run like a tool result's (@fallais)
  • 041c2f9: fix(aggregators): keep a tool result's run owed until something makes it (@fallais)
  • b8e74e1: fix(aggregators): mute and consume the eager transcription in the user aggregator (@fallais)
  • 89ccee2: fix(aggregators): recover a turn's voided answer like an empty user turn (@fallais)
  • c45a2e2: fix(aggregators): settle a tool call's result where the call was made (@fallais)
  • da1817e: fix(anthropic): cache the system instruction (@fallais)
  • 43e0dbe: fix(assemblyai): default the model to universal-3-6-pro (@fallais)
  • 9f1cc31: fix(assemblyai): reject a prompt off U3 Pro and map the missing languages (@fallais)
  • b757666: fix(assemblyai): treat universal-3-6-pro as a U3 Pro model (@fallais)
  • d3a2471: fix(bedrock): leave the interleaved-thinking beta off Bedrock requests (@fallais)
  • 77ea7c5: fix(context): attribute a closing tag to the word it is written against (@fallais)
  • 41cf8a1: fix(context): keep the text type of a sentence promoted from streamed text (@fallais)
  • b3ce2b0: fix(context): keep word tracking in step on markdown-heavy replies (@fallais)
  • b19055d: fix(context): pair force-completed text with the progress that matches it (@fallais)
  • fe033e6: fix(deepgram): report Deepgram's reason when it refuses a connection (@fallais)
  • eac406d: fix(deepseek)!: turn thinking off by default and send reasoning_content after tool calls (@fallais)
  • c8ea497: fix(events): let a handler clean up the object that raised its event (@fallais)
  • 87bd14b: fix(flows): keep a node's context and tools through an interruption (@fallais)
  • d7a92f1: fix(frames)!: make the resume and turn inference completed frames system frames (@fallais)
  • 80a82d0: fix(frames): let a late async tool result be delivered on its own (@fallais)
  • ed9a2df: fix(gladia): report rejected audio chunks and failed translations (@fallais)
  • f6be927: fix(google): default to the Gemini 3.8 models before the old ones retire (@fallais)
  • cc648b5: fix(google): follow a trailing model turn with a user turn (@fallais)
  • f4e72db: fix(google): sign a function call the model did not make (@fallais)
  • bec5fb1: fix(gradium): report the words held when the STT connection drops (@fallais)
  • d068a83: fix(gradium): resample STT input at sample rates Gradium does not accept (@fallais)
  • 37e91e5: fix(livekit): convert room audio to the pipeline's sample rates (@fallais)
  • 2246b0a: fix(llm): rebind a toolset handler when its tool carries a new one (@fallais)
  • 52a5a12: fix(llm): send each model its lowest documented thinking setting (@fallais)
  • 776c80d: fix(llm): send the service's built-in tools when the conversation has none (@fallais)
  • 751fd6c: fix(llm): start a response's tool calls once the response has ended (@fallais)
  • 45987ed: fix(llm): teach the turn-completion protocol in prose (@fallais)
  • 2205774: fix(llm): track a function call from dispatch so an interruption cannot miss it (@fallais)
  • 48c8510: fix(llmworker): keep an idle check for the frontend it was asked for (@fallais)
  • 02ad3d0: fix(loudness)!: measure volume without allocating, to the reference meter (@fallais)
  • 7ef8d25: fix(mcp): drop a dead session, and tell the model why a call failed (@fallais)
  • f8f0f72: fix(metrics): label a service by its kind, not its instance (@fallais)
  • 239c86f: fix(novasonic): default to Nova 2 Sonic now that the first model is retired (@fallais)
  • 8ff036d: fix(nvidia): default the LLM to nemotron-3-super-120b-a12b (@fallais)
  • dafc7e8: fix(nvidia): default the LLM to nemotron-3.5-super-vl-120b-a12b (@fallais)
  • 523bc92: fix(observers): clean up an observer when it is removed or the pipeline stops (@fallais)
  • cb58068: fix(openai): lower reasoning on gpt-5 and leave models that reject none at their default (@fallais)
  • f0ec43d: fix(openai/realtime): resample input audio to the session's 24 kHz (@fallais)
  • 494d9c9: fix(pipeline): cancel a worker's children when its pipeline finishes (@fallais)
  • efbbcc3: fix(pipeline): drain a flush through a parallel pipeline (@fallais)
  • 62d8326: fix(pipeline): leave RTVI off a bridged worker by default (@fallais)
  • a69a406: fix(pipeline): let a cancel stop a pipeline that is still setting up (@fallais)
  • 46ae092: fix(pipeline): let an observer remove itself from its own callback (@fallais)
  • 0596a63: fix(processor): never run two process goroutines on one processor (@fallais)
  • 32339c2: fix(processor): report a panic while handling or pushing a frame instead of crashing (@fallais)
  • 3b887f9: fix(processor): stop processing the frames queued behind an interrupted one (@fallais)
  • b89828a: fix(responses): leave gpt-6-astra at its own reasoning default (@fallais)
  • 1586cbe: fix(rtvi): declare protocol 2.2.0 (@fallais)
  • 5bed27b: fix(rtvi): keep the words and progress of a skipped segment from the client (@fallais)
  • d84ad6d: fix(rtvi): leave a function call running on an intermediate result (@fallais)
  • 1ae272c: fix(rtvi): report spoken text under the segment it was spoken in (@fallais)
  • 0c2a980: fix(rtvi): send bot-output for text that is not spoken, and drop cut-off text (@fallais)
  • a53bbf8: fix(soniox): authenticate with a Bearer header on the handshake (@fallais)
  • 73fbad5: fix(speechmatics): send Filipino as the Tagalog language pack (@fallais)
  • 403f909: fix(stt): cut segments on the VAD, and transcribe them in order (@fallais)
  • 2415305: fix(stt): drop the TTFB measurement when speech starts or on interruption (@fallais)
  • 00980e8: fix(stt): finalize a segment's transcript only while it is the latest (@fallais)
  • 78ed7ef: fix(stt): honor STTMuteFrame and stop sending audio while muted (@fallais)
  • 72991f3: fix(stt): keep the keepalive running after a failed send (@fallais)
  • ce08f4c: fix(stt): retry a connection Flux or Cartesia Turns never confirms (@fallais)
  • 71f8e5e: fix(stt): retry a streaming session that fails to open (@fallais)
  • 8729101: fix(stt): transcribe user audio and attribute transcripts to the user (@fallais)
  • da32577: fix(telnyx): send the pipeline's application messages to the client (@fallais)
  • a03f217: fix(text): find a tag of any pair, and a pair with one delimiter (@fallais)
  • 8fd1640: fix(text): keep Latin abbreviations whole in Japanese sentences (@fallais)
  • 912ab95: fix(text): keep an ellipsis whole when looking ahead for a boundary (@fallais)
  • c3c0d32: fix(text): keep sentence boundaries out of tags that have closed (@fallais)
  • dda2f44: fix(together): default the LLM to zai-org/GLM-5.2 (@fallais)
  • 7f0e5cc: fix(together): escape the TTS voice and send the language (@fallais)
  • 6cfc450: fix(transport): convert WebRTC audio to the pipeline's sample rates (@fallais)
  • c25a9f2: fix(transport): keep the audio loop running through a barge-in (@fallais)
  • 4f05f43: fix(transport): never write an urgent keypress over a chunk the audio loop is writing (@fallais)
  • 29fc662: fix(transport): route every filter control frame to the input filter (@fallais)
  • f523bdd: fix(tts): always push the response end behind the turn's audio (@fallais)
  • 78ab37d: fix(tts): call the interruption hook while each context is still open (@fallais)
  • d1f9c2a: fix(tts): close the open contexts when a provider replaces its connection (@fallais)
  • 755825f: fix(tts): hold frames until the open audio contexts end, not the bot speaking (@fallais)
  • bbb1352: fix(tts): keep TTS audio frames aligned to whole samples (@fallais)
  • ffa48b7: fix(tts): keep skipped text behind the sentence before it, and trim streamed sentences (@fallais)
  • 90d225e: fix(tts): keep uninterruptible speech through an interruption (@fallais)
  • 68159c5: fix(tts): leave text a transform reduces to whitespace unspoken (@fallais)
  • 05170f1: fix(tts): promote the pending sentence at the end of a streamed turn (@fallais)
  • 660b2b3: fix(tts): reopen the turn's audio context for audio arriving after its timeout (@fallais)
  • b98a313: fix(tts): report interrupted contexts in creation order (@fallais)
  • 6072ee5: fix(tts): reset the text filters when a reply ends (@fallais)
  • 1081099: fix(turns)!: release the end-of-turn analyzer when its stop strategy is cleaned up (@fallais)
  • db39cca: fix(turns)!: report turn metrics only for a prediction the model made (@fallais)
  • 89c3bf9: fix(vad): update Silero to v6.2 and reset it only during non-speech (@fallais)
  • 6ae4ec3: fix(workers): reach the hooks a worker built on another worker overrides (@fallais)
  • b5fba1a: fix(wsserver): bound the auto hang-up request and make it before the run ends (@fallais)
  • 59377f3: fix(wsservice): stop reconnecting once the session is canceled (@fallais)
  • 8e225a1: fix(wsutil): bound each WebSocket send to a peer that stops reading (@fallais)
  • aa0b314: fix(wsutil): keep API keys out of a failed dial's error (@fallais)
  • cdab9fc: fix(xai): accept 11025 Hz in the realtime sample rate (@fallais)
  • a726b26: fix(xai): report chunk finals as interims so each utterance is transcribed once (@fallais)
  • a6a181b: fix: handle user audio wherever input audio is handled (@fallais)
  • 6a7e075: perf(openai): stop waiting for an interrupted Responses WebSocket reply (@fallais)
  • bd1feed: refactor(frames)!: call what kind of text a frame holds its TextType (@fallais)
  • 60eca61: refactor(llmwithbackend)!: split the package into prompts, strategies and connector (@fallais)
  • 6d06e84: test(aggregators): let the turn's interruption go by in the turn-strategy tests (@fallais)
  • 8e07ae3: test(aggregators): wait for the user turn frame itself before ending the run (@fallais)
  • 835c92b: test(jev): a broken TLS connection is a classifier error (@fallais)
  • ffb037a: test(live): speak on the session that declared the tools (@fallais)
  • 6db8718: test(openai): wait for the user transcript before checking the realtime events (@fallais)
  • f3d1b20: test(pipeline): stop the unusable-policy tests racing the end of the run (@fallais)
  • bf0bc05: test(rtvi): let the answer reach the aggregator before typing over it (@fallais)
  • e80103d: test(transport): wait for a stream's first write before timing its end (@fallais)