# Documentation Page **Page:** kit/agent/voice [Download Raw Markdown](./kit/agent/voice.md) --- # Talk to your Bot by voice `hoody agent bots voice connect` connects your microphone and speakers to a Bot. You speak, the Bot's speech-to-text model writes down what you said and posts it to the Bot as your message, and the Bot's reply is spoken back through its text-to-speech model. The transcript of both sides prints in the terminal as you go. Typed messages and approvals work in the same session. The command runs audio programs already on your computer (PulseAudio, PipeWire, ALSA, SoX or FFmpeg tools) and sends raw audio over the Bot's voice connection (`GET /api/v1/agent/bots/{id}/voice`). It installs no audio library of its own. One connection controls a Bot's voice at a time. That includes this command, an OpenAI Realtime client and a phone call. You can also talk to a Bot in the Hub ([hub.hoody.com](https://hub.hoody.com)). Open **Bots**, select the Bot and stay on **Activity**. Press **Talk** in the message box, then hold **Hold to talk** while you speak and release to send, or switch the mode to **Hands-free** (use headphones). The browser asks for your microphone the first time. ## Requirements 1. **A Bot.** The default Bot, `chief`, is created on first use. `hoody agent bots list` lists the others. 2. **Speech models for the Bot.** The Bot uses its own `speech.stt_model` and `speech.tts_model` when they are set, otherwise the agent settings `audio.stt_model` and `audio.tts_model`, otherwise the Hoody AI voice relay when `audio.relay_default` is on (the free [Hoody Voice](/foundation/hoody-ai/models/#hoody-voice) models when the agent runs in a Hoody computer). `hoody agent audio models list --capability stt` lists the speech-to-text models and `--capability tts` the text-to-speech models, with each model's availability and a reason when it is unavailable. 3. **An audio program** for recording and one for playback. See [Audio programs](#audio-programs). 4. **A login.** The command uses the CLI's login (`hoody login`). Without one, or when the agent refuses it, the command stops with exit code 1 before anything is sent. 5. **A selected computer.** The Bot runs in the agent of one computer. Pass `-c `, set `HOODY_COMPUTER`, or save a default with `hoody use -c `. Without one, a normal run exits with code 2 before contacting the agent; `--doctor` reports the missing computer and exits with code 1. 6. **An agent that supports voice checks.** The command asks the agent for its capabilities first. An agent without the `bot_voice_check` capability still takes voice: the command connects without the check and says so. Text-only replies (`--audio-output none`) need the capability; against an older agent the command refuses them, because that agent would still make and bill speech. Update the agent to get both. When the agent's version cannot be read at all (no answer, a timeout, a server error), the command makes no claim about its age: a run with speech connects and shows only the connection's own outcome, and a text-only run stops before connecting with "Could not reach the agent: …" and exit code 1. Check the first three before the first call: ```bash hoody agent bots voice connect --id chief --doctor ``` By default, `--doctor` tries every audio program for this system. `--backend` limits both directions to the chosen program; `--input-backend` or `--output-backend` limits only recording or only playback. It records a 100 ms sample from the microphone with each recording program (checked and discarded on your machine, never sent) and plays 100 ms of silence with each playback program. It shows which one it picks for recording and for playback, and asks the agent whether the Bot's speech settings are ready (`GET /api/v1/agent/bots/{id}/voice/check`). It prints one line per speech stage, then a verdict: ```text Voice check of Bot chief: ok speech-to-text hoody-ai/hoody-stt ok text-to-speech minimax/speech-2.8-turbo (voice English_expressive_narrator) Ready. One utterance can last up to 60 s. ``` In the Hub, open **Bots**, select the Bot, press **Bot settings** and choose **Voice**. Turn on **Choose the models and voice myself** to pick a **Listening model**, **Speaking model**, **Voice** and **Language**, then press **Save**. After saving, press **Check it works** to run the same readiness check. A stage that is not ready is marked `no` with the reason, and `--doctor` exits with code 1. With `--audio-output none`, text-to-speech is marked `off` and not checked. With `--input none`, speech-to-text is marked `off`: it is not used, and its result is left out of the verdict. The check follows the same path as a real call: which models the Bot would use, that each model can transcribe or speak, that the voice and language are ones the model offers, and that the agent holds a key for the provider or can use the Hoody AI voice relay. It runs no transcription, bills nothing and takes at most 6 seconds. A ready check is not a promise that the provider will answer: an exhausted speech budget or a provider outage shows up only during the session. ## Voice sessions ```bash hoody agent bots voice connect --id chief ``` The session starts in push-to-talk mode: | Key | Action | |---|---| | Space | Start speaking. Press again to send what you said. | | Esc | Stop the reply that is playing. A turn that answers only what you said by voice in this session is cancelled too; a turn that includes a typed message keeps running. | | `t` | Type a message or a command instead of speaking. | | `q` | End the session. In a `--once` run, quitting before the run completes ends with exit code 1 and prints "Quit before the one-shot run completed." with the command that shows the answer to a question already sent. | | Ctrl+C | Stop the reply and the Bot's turn, then end the session. See [Session end and exit codes](#session-end-and-exit-codes). | Only the audio between the two presses of Space is sent. When it connects, the command asks the agent to check the speech settings it needs (speech-to-text unless `--input none`, plus text-to-speech unless replies are text-only). A missing model or key refuses the connection with the agent's reason, before the session opens. See [Troubleshooting](#troubleshooting). Your words print after `you:`, the Bot's reply to you after `bot:`. A reply to another message (one from a chat app, or one the Bot started itself) prints after `bot (other):` and is never taken as the answer to you. An approval or a question from the Bot prints as a line that starts with `The Bot asks:`. ### Hands-free mode ```bash hoody agent bots voice connect --id chief --mode vad ``` The microphone stays open and the agent hears when you start and stop speaking. Use headphones: through speakers the Bot hears its own reply and treats it as you talking over it. `--duplex` applies to hands-free mode only: with `half` (the default) the microphone is muted while the Bot speaks; `full` keeps it open, so you can interrupt the reply by speaking. When standard input is not a terminal, the command uses hands-free mode and says so. ### Text-only replies ```bash hoody agent bots voice connect --id chief --audio-output none ``` You speak, the Bot answers in text only. The command connects with `speak=false`: the agent makes no speech, every reply arrives as text, and nothing is billed for speech synthesis. Speech-to-text still runs: with the free Hoody Voice model it counts against your server's voice limits, and another provider's model is billed by that provider. Only a speech-to-text model is needed. Approvals the Bot would announce aloud are printed instead. ## One question from a file ```bash hoody agent bots voice connect --id chief --in-file question.wav --once ``` The command sends the recording, prints the transcript and the Bot's reply, and exits. Exit code 0 means the answer arrived and its speech did not fail. Exit code 1 means the answer did not arrive or could not be confirmed, or the speech service did not speak it in full; the reason is printed. With text-only replies (`--audio-output none`) only the answer counts. Raw audio works the same way: ```bash arecord -f S16_LE -r 16000 -c 1 -t raw | hoody agent bots voice connect --id chief --raw-stdin --once ``` Audio on standard input that runs past the utterance limit is cut there, and the command says so. The file must be a WAV file of 16-bit PCM, mono, at 8000, 16000, 24000 or 48000 Hz. The command sends it at the rate in the file and never resamples. Any other format is refused with the FFmpeg command that converts it: ```bash ffmpeg -i question.mp3 -ac 1 -ar 16000 -c:a pcm_s16le converted.wav ``` One recording is one utterance, at most 60 seconds; a longer file is refused before the command connects. Some speech-to-text models take less. A recording under 60 seconds but over the model's limit is still uploaded, then rejected with "That was longer than the Bot takes at once; say it in shorter parts.", and the run ends with exit code 1. `--doctor` prints the limit ("One utterance can last up to 30 s."). `--once` is on by default for file and standard-input input. It needs a spoken question: with `--input none` it exits with code 2, since a typed message cannot be matched to its reply. In a `--once` run with the microphone, a typed message is not sent, because the question is the one you speak; approval and answer commands still go out. It waits up to `--timeout` seconds (180 by default) for the upload, the reply and the agent's playback accounting. ### How the answer is found The command waits for the transcript of your question, then takes the first reply the agent marks as answering a message sent on this connection. Replies to other messages are shown and skipped. When the agent cannot tell which message a reply answers, the command checks the Bot's log up to three times, at 0, 2 and 5 seconds after it starts, and gives up within 8 seconds of starting. The check starts once the reply is judged, which is after the transcript of your message arrives if the reply came first. Each time it reads the newest 100 entries and, if the log has more, the 100 before them, and looks for the reply's turn with your message in it: - The Bot's reply is in the log: it is the answer. With text-only replies (`--audio-output none`) the exit code is 0. When speech is wanted the exit code is 1, because a reply the agent could not tie to this connection is never spoken; its text is shown. - A failure entry names your message (an error, a provider failure or a limit): the turn failed. The command prints the reply text, if any, and "The Bot's turn failed:" with the failure. Exit code 1. A failure entry wins over a reply entry for the same message. - Neither is found: the command prints the reply text followed by "(could not confirm this reply answers your question)". Exit code 1. ### Connection loss with --once Before your question is sent in full, the command may reconnect and send it again from the start, since nothing reached the Bot yet; it prints "Sending your recording again from the start." After that point it never reconnects: in push-to-talk, once the end of the question was sent; in hands-free mode, once the first audio was sent. A connection lost after that point leaves the question possibly delivered, so the command prints that it may still be answered, names `hoody agent bots logs get --id chief` to read the answer, and exits with code 1. ## Files, pipes and transcripts | Goal | Flags | |---|---| | Record what the Bot said | `--out-file reply.wav` | | Read raw audio from standard input | `--raw-stdin --in-rate 16000` | | Write raw audio to standard output | `--raw-stdout --out-rate 24000` | | Transcript as JSON lines | `--transcript json` | | No transcript | `--transcript off` | | Typed messages only, no microphone | `--input none` | `--out-file` is opened before the session starts: a file that cannot be created or written ends the run with code 1 and starts nothing. When the program reading standard output closes it (a broken pipe), the session stops cleanly with code 1. Raw audio is 16-bit little-endian mono PCM with no header. The command refuses to write raw audio to a terminal or read it from one. `--raw-stdout` implies `--audio-output stdout`. `--raw-stdin` is the same as `--input stdin`; together they are accepted. With `--input mic`, `file` or `none` it exits with code 2: "Standard input can carry only one thing: --raw-stdin reads audio from it, but --input is mic." With `--input stdin` (or `--raw-stdin`) standard input carries only the recording. It is sent push-to-talk by itself, no keys are read, and `--mode vad` is refused. `--out-file` records the audio the agent sent. Audio that was sent and then cancelled (because you interrupted the reply) stays in the file; the command says so on standard error. `--transcript json` writes one JSON object per event, without audio. A `reply.final` whose turn failed after its text carries the stop code in `failure` (`content_filter`, `refusal` or `max_steps`). When audio goes to standard output, the JSON lines go to standard error. With file or standard-output audio, nothing is played, so the agent cannot know whether you heard the reply. Each reply ends as not confirmed (`spoken: "unknown"`) 2 seconds after the later of its expected playback end and the end of the audio output, and the command shows "Waiting for the agent to settle the reply…" meanwhile. That is expected, so the command prints no note about it, the same as with `--audio-output none`, and a one-shot run still exits with code 0. ## Audio programs The command picks the first audio program that passes its check, separately for recording and playback. `--backend`, `--input-backend` and `--output-backend` choose one by name; `--list-backends` shows which are installed, and `--doctor` shows which work. | System | Recording, in order of preference | Playback, in order of preference | |---|---|---| | Linux and WSL | `parec`, `arecord`, `ffmpeg`, `pw-record`, `sox` | `pacat`, `pw-play`, `aplay`, `sox`, `ffplay` | | macOS | `sox`, `ffmpeg` | `sox`, `ffplay` | | Windows | `ffmpeg` (DirectShow), `sox` (waveaudio) | `ffplay`, `sox` (waveaudio) | On Linux, every program in the table typically starts recording within 0.4 seconds and playing within 0.2 seconds, with the settings the command uses. The PipeWire tools of Ubuntu 22.04 (PipeWire 0.3.48) cannot stream raw audio, so the command skips them there and uses the next program. On Windows, the programs in the table were tested on Windows 11 24H2. The macOS programs have not been tested yet. Report what `--doctor` prints if one fails. Install one of them: ```bash sudo apt install pulseaudio-utils ``` `pipewire-bin`, `alsa-utils`, `sox` or `ffmpeg` work too. ```bash brew install sox ``` ```powershell winget install Gyan.FFmpeg ``` This installs both `ffmpeg` and `ffplay`. SoX for Windows also works. ### Devices With automatic backend selection, `--list-devices` lists devices for the first installed recording and playback programs, which may differ from those a session selects. Use `--input-backend` and `--output-backend` to list devices for specific programs. `--input-device` and `--output-device` pick one; without them the program's default device is used. On Windows, FFmpeg needs a device name: without `--input-device` the command uses the first device FFmpeg lists. FFplay cannot pick an output device; use `--output-backend sox` to choose one. ### Playback accounting The command reports to the agent how far each reply has played. The agent confirms a reply as fully spoken only when playback reaches the end. After an interruption, it tells the Bot which complete speech segments you heard. If the audio program stops taking audio, the report stops advancing and the reply ends as not confirmed. A reply you interrupt counts as partly spoken only when at least one of its speech segments had played to the end. If you interrupt during the first segment, the reply is reported as not spoken. A player that delays sound longer than the agent waits for playback reports cannot confirm replies; `--doctor` names it, and its replies count as not confirmed. When a reply played on the speaker ends as not confirmed, the command prints a note; the run still succeeds. ## Approvals and questions Speech never approves anything. When the Bot itself waits on an approval or a question, the command prints it with the commands that answer it. Press `t` and type one: | Command | Effect | |---|---| | `/approve` | Approve this once. | | `/deny ` | Refuse, with the reason. | | `/allow-session` | Approve for the rest of the Bot's session, when the approval offers it. | | `/answer ` | Answer the Bot's question. | `/allow-session` goes out only when the approval offers it. Otherwise nothing is sent, and the command names the commands the approval takes. If the agent answers `not_offered`, your answer was not used and the gate stays open. The reason after `/deny` can be at most 500 characters; a longer one is not sent. A question is printed one line per question, with the options the Bot offers; past 20 questions the rest is counted and left to the Bot's log. When the agent reports the outcome, the command prints it in words, with the agent's code in brackets. Some outcomes end the gate. The others leave it waiting, because your answer was not used or not confirmed: `/approve` and `/answer` still go to it, and nothing is sent again by itself. When the answer is not confirmed, read the Bot's log before you answer again: | Agent code | The gate | Line | |---|---|---| | `applied` | ends | `The Bot's gate ended: your answer was applied [applied].` | | `resolved` | ends | `The Bot's gate ended: it was settled (answered elsewhere, or withdrawn) [resolved].` | | `stale` | may still be waiting | `The Bot's gate may still be waiting: it had already changed or ended; your answer may not have been used, see the Bot's log [stale].` | | `stale_gate` | ends | `The Bot's gate ended: it had already changed, so your answer was not used [stale_gate].` | | `unavailable` | may still be waiting | `The Bot's gate may still be waiting: your answer is not confirmed (it may or may not have been applied); check the Bot's log before trying again [unavailable].` | | `no_pending_gate` | ends | `The Bot's gate ended: nothing was waiting for an answer [no_pending_gate].` | | `gate_cancelled` | ends | `The Bot's gate ended: it was cancelled [gate_cancelled].` | | `decision_unconfirmed` | may still be waiting | `The Bot's gate may still be waiting: it is not known whether your answer was applied; see the Bot's log [decision_unconfirmed].` | | `not_offered` | may still be waiting | `The Bot's gate may still be waiting: this approval does not offer /allow-session, so your answer was not used; use /approve or /deny [not_offered].` | | `gate_already_resolved` | ends | `The Bot's gate ended: your answer was refused [gate_already_resolved].` | | Any other code | may still be waiting | `The Bot's gate may still be waiting: your answer was refused [{code}].` | A delegate's approval or question (from a session the Bot opened) is announced, and spoken when the session speaks. It cannot be answered here. With `speech.speak_gates` set to `false` on the Bot, it is neither announced nor printed. Answer it in the agent's TUI, a chat app or over HTTP. ## Session end and exit codes The first Ctrl+C stops the microphone and the reply, asks the agent to cancel the Bot's turn, closes the connection and exits with code 130. When the turn cannot be cancelled (it carries other messages), the command prints that the Bot's turn keeps running, and the turn finishes on its own. A second Ctrl+C exits at once. On Linux and macOS, SIGTERM and SIGHUP do the same as the first Ctrl+C and exit with codes 143 and 129. Windows has no SIGTERM: use Ctrl+C, or Ctrl+Break, which exits with code 149. Something you were saying when you pressed Ctrl+C is not sent. | Exit code | Meaning | |---|---| | 0 | Finished normally. With `--once`: the request succeeded. A failure or no reply exits 1. | | 1 | Failed. With `--once`: the answer did not arrive or could not be confirmed, the Bot's turn failed after part of the reply, or the speech service did not speak it in full. The reason is printed. An internal error in the command also ends with 1, after `Internal error:` and its message. | | 2 | The flags conflict or are wrong. Nothing was sent. A missing required option (such as `--id`) or an unknown option is caught by the option parser first and exits 1 instead. | | 130 | Stopped with Ctrl+C. | | 143 | Stopped with SIGTERM. | | 129 | Stopped with SIGHUP, for example when the terminal closed. | | 149 | Stopped with Ctrl+Break on Windows. | When the Bot's turn fails after part of its reply was written (a content filter stopped it, the model refused to go on, or it reached its step limit), the text stays the reply and the command prints after it that the reply may be cut off or refused and that earlier parts of it may still stand. With `--once` the run then ends with exit code 1, even when the speech played in full. On Windows, a Ctrl+C in the first moment after start can end the command with Windows' own interrupt status (0xC000013A) instead of 130. ## Reconnects In a live session, when the connection drops, the command reconnects up to 5 times, waiting at least 1, 2, 4, 8 and 8 seconds. If a reconnect attempt fails and the agent asks for a longer wait (a `Retry-After` header or a retry time in its error), the command waits that long instead. It does not reconnect after a refusal, a protocol error or a normal close. `--no-reconnect` turns reconnecting off. `--once` follows [its own rule](#connection-loss-with---once). What happens to the sentence you were speaking depends on when the connection dropped: - Before you pressed Space to send it, the command prints "Your unfinished sentence was not sent." Say it again. - After you sent it and before its transcript arrived, the command prints "Your last sentence may or may not have reached the Bot; check before saying it again." The agent may still have it, and the Bot can still answer it. A typed message or gate answer entered while the connection is dropping or being reconnected does not go out, and it is not shown as sent: the command prints "Not sent: the connection to the agent is down." Send it again once the connection is back. When the agent ends the connection because it could not deliver to the command in time, no close code arrives: the connection just ends, and the command treats it as dropped and reconnects. The command never sends a sentence again by itself. `--takeover` applies to the first connection only: a reconnect never takes the voice over again. The agent closes a session after 10 minutes without anything from the client; the Bot's own speech does not count. The command keeps a session open by itself while it runs. ## Voice takeover When another connection controls the Bot's voice, the command is refused and says so. `--takeover` takes the voice over and closes the other connection. Use it only when you know which connection that is, for example a session you left open on another computer. A refused check (missing model or key) never takes the voice from anyone. ## Flags | Flag | Meaning | |---|---| | `--id ` | Required. The Bot. | | `--mode ` | ptt (push-to-talk, default) or vad (hands-free); recordings always use ptt. | | `--duplex ` | half (default: the microphone is muted while the Bot speaks, hands-free only) or full. | | `--input ` | mic (default), file (--in-file), stdin (raw PCM16) or none (typed messages only). | | `--in-file ` | A 16-bit mono PCM WAV at 8000, 16000, 24000 or 48000 Hz to send as one utterance. | | `--audio-output ` | speaker (default), file (--out-file), stdout (raw PCM16) or none (text only). | | `--out-file ` | Write the Bot's voice to this WAV file. | | `--raw-stdin` | Read raw PCM16 mono at --in-rate from standard input (implies --input stdin). | | `--raw-stdout` | Write raw PCM16 mono at --out-rate to standard output (implies --audio-output stdout). | | `--in-rate ` | Rate of raw input and the microphone: 8000, 16000 (default), 24000 or 48000. | | `--out-rate ` | Rate of the Bot's voice: 8000, 16000, 24000 (default) or 48000. | | `--once` | One utterance and its reply, then exit (on for --in-file and stdin). | | `--transcript ` | text (default), json (one object per line) or off. | | `--voice ` | The voice the Bot speaks with (default: the Bot's). | | `--language ` | The language you speak (default: the Bot's, or detected). | | `--backend ` | Audio program for both directions (default auto; see --list-backends). | | `--input-backend ` | Audio program for recording (default auto). | | `--output-backend ` | Audio program for playback (default auto). | | `--input-device ` | Microphone (default: the system default; Windows: the first listed). | | `--output-device ` | Speaker (default: the system default). | | `--takeover` | Take the Bot's voice from another connection (first connect only). | | `--no-reconnect` | Do not reconnect when the connection drops. | | `--timeout ` | One-shot runs: give up after this long (default 180). | | `--doctor` | Check the audio programs and the Bot's speech settings, and print what would run. | | `--list-backends` | List the audio programs this OS can use. | | `--list-devices` | List devices for the first installed audio programs, or the explicitly selected backends. | `--output` stays the global output-format flag of the CLI; the audio destination is `--audio-output`. ## Troubleshooting Every message names what happened and one thing to do. When the agent refuses or a speech call fails, the command prints the agent's own message inside its own line, for example "The voice connection could not be opened (HTTP 503 speech_unavailable): …" for a refusal that may pass, or "The voice connection was refused (HTTP 400 bad_request): …" for one that will not, followed by the reason and a support code (`diag`). The code names one failure in the agent's log; include it when you ask for help. When a refusal carries no message, the command runs the speech check and adds what it finds. ### Speech settings not ready These come from `--doctor`, from a refused connection, or from a failed speech call during the session. `M`, `V`, `L` and `openai` stand for the model, voice, language and provider named in the message. | Message | What to do | |---|---| | No speech model is set. Set the Bot's speech.stt_model or speech.tts_model, the agent setting audio.stt_model or audio.tts_model, or set audio.relay_default back to true (the Hoody AI voices). | Choose models with `hoody agent audio models list`, then set one of the settings named. | | No API key is stored for the speech provider minimax; add one with PUT /api/v1/agent/providers/minimax/auth/api-key or pick another speech model. | Add the provider's key, or pick a model whose provider has one. | | Voice through the Hoody AI relay is not set up on this agent. Set a speech model of a provider with a key for this Bot. | Set the Bot's speech models to a provider with a key. | | The Hoody AI voice relay refused this agent's credential. Run the agent in a Hoody computer, where the voice relay needs no key, or set a speech model of a provider with a key for this Bot. | Run the agent in a Hoody computer, or set the Bot's speech models to a provider with a key. | | This agent has no Hoody AI credential for the voice relay. Run the agent in a Hoody computer, where the voice relay needs no key, or set a speech model of a provider with a key for this Bot. | Run the agent in a Hoody computer, or set the Bot's speech models to a provider with a key. | | The Hoody AI voice relay does not offer this speech model right now. Pick another speech model for this Bot. | Pick another model. | | The Hoody AI voice relay offers no speech models of this kind here. Set a speech model of a provider with a key for this Bot. | Use a provider with a key. | | This agent may not connect to the Hoody AI voice relay. Set a speech model of a provider with a key for this Bot. | Use a provider with a key. | | The Hoody AI voice relay could not be reached. Try again in a minute. | Try again. | | The Hoody AI voice relay gave an answer this agent cannot use. Try again in a minute. | Try again. | | This speech model cannot be served here. Pick another speech model for this Bot. | Pick another model. | | The Bot's speech model is not a known speech model. Pick one from GET /api/v1/agent/audio/models. | Pick one from `hoody agent audio models list`. | | This model does not transcribe speech; pick a model with the stt capability. | Use `--capability stt` to list the right models. | | This model does not synthesize speech; pick a model with the tts capability. | Use `--capability tts` to list the right models. | | The speech model is not available. Pick another speech model for this Bot. | Pick another model. | | Voice V is not one of the voices of M. Pick one from GET /api/v1/agent/audio/voices, or leave voice out. | List them with `hoody agent audio voices list --model M`, or drop `--voice`. | | The speech model M does not take language L. Pick another language, or leave language out. | Drop `--language`, or pick another. | | The speech settings could not be checked in time. Try again in a moment. | Try again in a moment. | | The speech settings could not be checked right now. Try again in a moment. | Try again in a moment. | `--voice` and `--language` take 1 to 64 printable characters, spaces included, so a voice id such as `"Chinese (Mandarin)_News_Anchor"` can be chosen; quote it in the shell. The Bot's speech always takes `--language auto`. Speech-to-text takes `auto` only when its model lists no languages or lists `auto`. When every failing check is a voice or language the model does not offer, the session is refused with HTTP 400 before it opens. If any other row fails too, the refusal is HTTP 503 `speech_unavailable`. A refused check takes nothing over. ### Refusal codes | Status and code | What happened | What to do | |---|---|---| | `409 voice_busy` | Another voice connection controls this Bot. | Close it, or rerun with `--takeover`. | | `409 realm_blocked` | The realm selection is blocked. | Select a realm, then try again. | | `409 bot_owned_elsewhere` | Another agent process on this machine runs this Bot, or its owner lock could not be taken. | Use the Bot there, or try again after the time the message gives (about 5 seconds). | | `404 not_found` | No Bot with this id in the selected realm. | Check the id with `hoody agent bots list`. | | `403 forbidden` | Hoody Kit programs are reached through their URLs only. | Run the command against the agent's URL, as `hoody agent` does by default. | | `400 bad_request` | A value the agent does not accept, named in the message. | Correct that flag. | | `503 service_unavailable` | The agent or the Bot cannot take the connection right now. | Try again after the time given. | | `429 rate_limited` | Too many requests. | Wait and try again. | ### Speech failures during a session These are the agent's messages. For a failed transcription, the command prints the agent's message as it is, followed by `(reason R, diag D)` when those are known. Only when the agent sends no message does it print `Your speech could not be handed to the Bot (). Say it again.` instead. For a reply the speech service could not speak, it prints `The reply was not spoken: .` (or `only partly spoken`), with no reason or diag id; `--transcript json` carries the full error, with its `diag_id`. | Message | What happened | What to do | |---|---|---| | The speech provider refused the stored credential. Check the provider's API key. | The provider rejected the key. | Replace the provider's key. | | The speech provider is rate limiting requests. Try again shortly. | Rate limit at the provider. | Wait, then speak again. | | The speech provider did not answer in time. Try again. | Timeout. | Speak again. | | The speech provider failed to answer. Try again; if it keeps failing, pick another speech model. | Provider error. | Speak again, or change the model. | | The transcript does not match the amount of speech in the audio; it was not used. Say it again. | The transcript looked wrong and was dropped. | Say it again. | | The speech quota is used up. Wait until the budget renews, or raise it. / The daily speech limit is reached; it resets at the next UTC midnight. | A daily speech limit for the Bot, its realm or the agent would be exceeded. A transcription failure keeps the spoken question from reaching the Bot; a synthesis failure can leave the reply unspoken or partly spoken. | Wait until the next UTC midnight, or raise the applicable limit. Read the reply in the transcript if it was not fully spoken. | | Speech is busy. Try again in a moment. | The agent had no room for the call. | Speak again in a moment. | | Speech usage could not be recorded. Try again in a moment. | The agent could not record the call's usage, so it did not make the call. | Speak again in a moment. | | The Bot's session is not available right now; nothing was billed. Say it again. | The Bot could not take the message. | Say it again. | | Your message arrived after the Bot was reset. Say it again. | The Bot was reset while you spoke. | Say it again. | | The Bot was deleted before your message reached it. | The Bot is gone. | Use another Bot. | | Your speech could not be handed to the Bot. Say it again. | The message did not reach the Bot. | Say it again. | | The speech could not be transcribed. Say it again. | Transcription failed without a known cause. | Say it again. | | The speech provider stopped early. The rest is on screen. | Only part of the reply was spoken. | Read the rest in the transcript. | | The reply could not be spoken. It is on screen. | The reply was not spoken. | Read it in the transcript. | ### Messages from the command #### Flag errors | Message | When it is printed | Exit code | |---|---|---| | `--mode must be one of ptt, vad (got "talk").` | A flag got a value it does not take. | 2 | | `--in-rate must be 8000, 16000, 24000 or 48000 (got "44100").` | --in-rate or --out-rate is not a rate the voice route takes. | 2 | | `--in-file and --raw-stdin both choose where the question comes from. Give only one of them.` | --in-file together with --raw-stdin: both choose where the question comes from. | 2 | | `--out-file and --raw-stdout both choose where the reply audio goes. Give only one of them.` | --out-file together with --raw-stdout: both choose where the reply audio goes. | 2 | | `--timeout must be a positive number of seconds (got "0").` | --timeout is not a positive number of seconds. | 2 | | `--mode vad does not work with --input file: a recording is always sent push-to-talk. Leave --mode out.` | --mode vad together with --input file or --input stdin. | 2 | | `--in-file reads a recording, but --input is mic. Use --input file, or leave --input out.` | --in-file together with an --input other than file. | 2 | | `--input file needs --in-file .` | --input file without --in-file. | 2 | | `--audio-output file needs --out-file .` | --audio-output file without --out-file. | 2 | | `--out-file writes a recording, but --audio-output is speaker. Use --audio-output file, or leave --audio-output out.` | --out-file together with an --audio-output other than file. | 2 | | `Standard input can carry only one thing: --raw-stdin reads audio from it, but --input is mic.` | --raw-stdin together with --input mic, file or none: standard input would carry the audio and also the keys, typed messages or a second recording. (--raw-stdin with --input stdin is fine: they are the same input.) | 2 | | `Raw audio would be written to the terminal. Redirect standard output to a file or a program, or pick another --audio-output.` | Raw audio would be written to a terminal (--audio-output stdout or --raw-stdout without a redirect). | 2 | | `Raw audio is read from standard input, but it is the terminal. Pipe audio in, or use --input mic.` | --input stdin or --raw-stdin while standard input is the terminal. | 2 | | `--once asks one spoken question, but --input is none. Use --input mic or --input file, or leave --once out.` | --once together with --input none: a typed message cannot be matched to its reply. | 2 | | ``No computer selected. Pass -c , set HOODY_COMPUTER, or save a default with `hoody use -c `.`` | No computer is selected. The Bot runs in the agent of one computer, so nothing was sent. | 2 for a normal run; 1 for `--doctor` | | `Unknown audio backend "jack" for recording. Known here: parec, arecord, ffmpeg-pulse, ffmpeg-alsa, pw-record, sox. See --list-backends.` | A backend flag names no known audio backend. | 2 | #### Recordings | Message | When it is printed | Exit code | |---|---|---| | `Cannot read question.wav: no such file. Check the path.` | The --in-file recording cannot be opened or read. | 1 | | `question.mp3 is not a WAV file; the voice route takes 16-bit mono PCM WAV at 8000, 16000, 24000 or 48000 Hz. Convert it with: ffmpeg -i question.mp3 -ac 1 -ar 16000 -c:a pcm_s16le converted.wav` | The --in-file recording is not 16-bit mono PCM WAV at 8, 16, 24 or 48 kHz. | 1 | | `question.wav holds no audio. Record something first.` | The recording holds no audio. | 1 | | `question.wav is 75.2 s long; one utterance can be at most 60 s. Cut it shorter.` | The recording is longer than one utterance can be (60 s). Nothing is sent. | 1 | | `The audio on standard input ran past 60 s; only the first 60 s were sent.` | Audio from standard input ran past the utterance limit; the rest was not sent. | continues | #### Audio programs on this computer | Message | When it is printed | Exit code | |---|---|---| | `No program to record from the microphone works here. Install pulseaudio-utils (parec, pacat), pipewire-bin (pw-record, pw-play; Debian and Ubuntu), alsa-utils, sox or ffmpeg. Run with --doctor to see what was tried.` | No audio program for recording was found, or none passed its check. | 1 | | `No program to play the Bot's voice works here. Install pulseaudio-utils (parec, pacat), pipewire-bin (pw-record, pw-play; Debian and Ubuntu), alsa-utils, sox or ffmpeg. Or use --audio-output none for text only. Run with --doctor to see what was tried.` | No audio program for playback was found, or none passed its check. | 1 | | `pw-record does not work here: pw-record is not installed. Run with --doctor, or leave the backend flag out to pick one automatically.` | The backend named with --backend / --input-backend / --output-backend failed its check. | 1 | | `parec stopped (exit 1): Connection refused. Run with --doctor to check the audio setup.` | The recording or playback program stopped on its own during the session. | 1 | | `arecord could not be started (ENOENT). Run with --doctor to check the audio setup.` | The recording program that passed the audio check could not be started for the session. | 1 | | `ffmpeg finds no microphone. Plug one in, or name one with --input-device; --list-devices shows them.` | Windows: ffmpeg lists no DirectShow microphone. | 1 | | `ffplay cannot pick an output device; --output-device is ignored. Use --output-backend sox to choose one.` | An --output-device was given for ffplay, which cannot pick a device. | continues | | `Audio on macOS is untested. If it fails, run with --doctor and report what it shows.` | Audio programs on macOS (built from their documentation, not yet tested on a Mac). | continues | | `ffplay plays with more delay than the agent waits for; replies through it are recorded as "unknown" instead of heard.` | --doctor: the playback program delays sound longer than the agent waits for playback reports. | continues | | `The audio player is not taking audio; the agent is told playback paused.` | The playback program stopped taking audio for more than half a second. | continues | | `Cannot write reply.wav: permission denied. Check the path and that its folder exists.` | The --out-file file cannot be created or written. Nothing is started. | 1 | | `Standard output was closed; stopping.` | Standard output was closed by the program reading it (a broken pipe): the session stops cleanly. | 1 | | `Standard output was not read in time; the output may be incomplete.` | The program reading the output stopped reading before the command ended. The output may be incomplete. | 1 | #### Connection and session start | Message | When it is printed | Exit code | |---|---|---| | `The voice connection was refused (HTTP 409 voice_busy): another voice connection controls this Bot. Close the other connection, or rerun with --takeover.` | HTTP 409 voice_busy: another connection holds this Bot's voice. The command never takes over by itself. | 1 | | `The voice connection was refused (HTTP 400 bad_request): Voice "nova2" is not one of the voices of "minimax/speech-2.8-turbo". Pick one from GET /api/v1/agent/audio/voices, or leave voice out. (reason voice_not_found)` | The agent refused the voice connection for good (HTTP 4xx except 408, 425, 429 and 409 voice_busy): a bad flag value, an unknown Bot, a voice or language the speech model does not offer. | 1 | | `The voice connection could not be opened (HTTP 503 speech_unavailable): The Hoody AI voice relay could not be reached. Try again in a minute. (reason relay_listing_unreachable, diag 77a0c3e15b9d2f48)` | The voice connection could not be opened for a reason that may pass (HTTP 5xx, 408, 425, 429): speech not ready, the agent busy or restarting. | 1 | | `` The CLI is not logged in, or its login was refused. Run `hoody login`, then try again. `` | The CLI is not logged in, or its saved login was refused. Nothing was sent. | 1 | | `The agent accepted the connection but did not start the voice session in time. Try again; if it repeats, update the agent.` | The agent accepted the voice connection but did not start the session in time. | 1 | | `Could not reach the agent: connection timed out. Check the address and the network.` | The agent could not be reached at all (network, address, TLS). Also before a text-only connect (--audio-output none) whose GET /version failed without showing an old agent: the connect could not turn speech off, so it does not start. | 1 | | `Text-only voice (--audio-output none) needs a newer agent. Update the agent, or use voice output.` | --audio-output none against an agent shown to lack text-only voice (GET /version answered without it, or 404/405): it would still synthesize and bill speech. Checked before connecting, and again on the session it opens. | 1 | | `The agent opened the session hands-free=true instead of push-to-talk. Update the agent.` | The agent opened the session in another talk mode than asked. | 1 | #### During a session | Message | When it is printed | Exit code | |---|---|---| | `Nothing was heard. Speak closer to the microphone, or check --input-device.` | No speech was heard in the recording. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `That was longer than the Bot takes at once; say it in shorter parts.` | The utterance was longer than the Bot takes at once. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `Part of your audio was lost on the way; say it again.` | Part of the audio was lost on the way to the agent. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `The Bot had no room for that; say it again in a moment.` | The Bot had no room for another message. With --once the run ends when the utterance itself is refused; a busy notice alone does not end it. | 1 with --once; a live session goes on | | `The Bot did not take that (stt_suspect). Say it again.` | The agent did not take the utterance, for the reason shown. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `The speech provider did not answer in time. Try again. (reason deadline, diag 3f9a1c0b7e2d4a65)` | The speech could not be turned into a message. The line is the agent's own message, saying what happened and what to do next, with the reason and a diagnostic id when known. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `Your speech could not be handed to the Bot (submit). Say it again.` | The speech could not be turned into a message, and the agent gave no reason. With --once the run ends; a live session goes on. | 1 with --once; a live session goes on | | `The reply was only partly spoken: the speech service failed.` | Speech was wanted, but the reply was not spoken, or only in part. With --once the run ends with 1; a live session goes on. | 1 with --once; a live session goes on | | `The Bot's turn ended early (a content filter stopped it): this reply may be cut off or refused, and earlier parts of it may still stand.` | The Bot's turn failed after the text shown, so the reply may be cut off or refused; earlier parts of it may still stand. With --once the run ends with 1; a live session goes on. | 1 with --once; a live session goes on | | `The agent reported: The Hoody AI voice relay could not be reached. Try again in a minute. (reason relay_listing_unreachable, diag 77a0c3e15b9d2f48)` | The agent reported an error. The session continues unless the agent closes it. | continues | | `Unknown command /aprove. Use /approve, /deny , /allow-session or /answer .` | A typed line starts with / but is not a gate command. | continues | | `Not sent: type the answer after /answer.` | /answer was typed without an answer. Nothing is sent; the question stays open. | continues | | `What the Bot waits on changed while you typed, so nothing was sent. Check the request above and type your answer again.` | You typed an approval or answer command, but what the Bot waits on changed while you typed. Nothing was sent. | continues | | `The Bot is not waiting on anything.` | An approval or answer command was typed while the Bot waits on nothing. | continues | | `The Bot waits on a question; use /answer .` | An approval command was typed for a question, or /answer for an approval. | continues | | `Not sent: the connection to the agent is down. Send it again once it is back.` | A typed message or gate answer could not be sent because the connection was dropping or being reconnected. | continues | | `Not sent: with --once the question is the one you speak. Say it, or run without --once to type messages.` | A message was typed during a --once run with the microphone. Only the question you speak is sent; approval and answer commands still go out. | continues | | `This approval does not offer /allow-session; use /approve or /deny .` | /allow-session was typed for an approval that does not offer it. Nothing is sent; the line names the commands it takes. | continues | | `That reason is over 500 characters; send a shorter one.` | The reason after /deny is longer than the Bot takes. Nothing is sent. | continues | | `That message is over 64 KiB; send a shorter one.` | A typed message is over 64 KiB. | continues | | `Not sent: this run asks only the recorded question. Approval and answer commands still go out.` | A message was typed during a run that sends a recording. Only the recorded question is sent; approval and answer commands still go out. | continues | #### Connection loss | Message | When it is printed | Exit code | |---|---|---| | `Another connection took over this Bot's voice.` | Another connection took over this Bot's voice with --takeover. | 1 | | `The agent did not accept the voice protocol. Update the CLI or the agent.` | The agent closed the session because it did not accept the voice protocol. | 1 | | `The agent closed the session over something the CLI sent: protocol error. Update the CLI; if it persists, report it.` | The agent closed the session over something the command sent. | 1 | | `A message was over 64 KiB and the agent closed the session. Send shorter messages.` | A message was over 64 KiB, so the agent closed the session. | 1 | | `Connection lost (the connection fell behind); reconnecting in 2 s (attempt 2 of 5).` | The connection dropped; the command reconnects. The reason in brackets appears only when it is known. | continues | | `Could not reconnect after 5 attempts. Run the command again.` | Every reconnect attempt failed. | 1 | | `The connection was lost. Run the command again.` | The connection dropped and is not reopened (--no-reconnect, or the agent ended it in a way that cannot be retried). With --once the connection is reopened only before your question was sent. | 1 | | `Your unfinished sentence was not sent.` | The connection dropped while you were talking (before Space sent it). | continues | | `` The connection was lost after your question was sent; it may still be answered. See `hoody agent bots logs get --id chief`. `` | A one-shot run lost its connection after its question was sent (push-to-talk released, or hands-free audio started); it never resends then. | 1 | | `Your last sentence may or may not have reached the Bot; check before saying it again.` | The connection dropped after a sentence was sent but before its transcript came back. | continues | #### One-shot results | Message | When it is printed | Exit code | |---|---|---| | `` No reply within 180 s. Raise --timeout, or see `hoody agent bots logs get --id chief`. `` | --once: no reply within --timeout. | 1 | | `(could not confirm this reply answers your question)` | A one-shot run got a reply the agent could not attribute, and the Bot's log did not confirm it answers your question in up to three attempts within 8 s of the check starting (which waits for your question's transcript if the reply came first). | 1 | | `The Bot's turn failed: [provider] The model provider refused the request (HTTP 429). Try again later.` | A one-shot run: the Bot's log says the turn that took your question failed. | 1 | | `` The Bot answered with no text. See `hoody agent bots logs get --id chief` for what happened. `` | --once: the reply ended with no text. | 1 | | `` Quit before the one-shot run completed. If your question was sent, see `hoody agent bots logs get --id chief` for its answer. `` | You quit a one-shot run before it completed. The reply may already be on screen; a question already sent may still be answered. | 1 | #### Internal errors | Message | When it is printed | Exit code | |---|---|---| | `Internal error: WebSocket is not connected. Run the command again; if it repeats, report it.` | An unexpected error inside the command (a bug): the run stops with the normal teardown (players and microphone stopped) instead of a stack trace. | 1 | ## Other lines the command prints Status lines, labels and the output of `--doctor`, `--list-backends` and `--list-devices`. Examples use sample values. ### Flag notes | Line | When it is printed | |---|---| | `--takeover applies to the first connection only; a reconnect never takes over again.` | --takeover was given to a live session that may reconnect: it applies to the first connection only. | ### Audio output notes | Line | When it is printed | |---|---| | `reply.wav records what the agent sent. Speech that was stopped (barge-in, Esc) stays in it.` | --audio-output file: what the file holds. | | `Audio goes to standard output; the transcript goes to standard error.` | Audio goes to standard output; transcript lines go to standard error. | ### Voice check notice | Line | When it is printed | |---|---| | `The voice check says text-to-speech is not ready: No API key is stored for the speech provider openai; add one with PUT /api/v1/agent/providers/openai/auth/api-key or pick another speech model. (reason no_credentials, diag 9b2e47c01d5f8a36)` | Under a failed connection that came without the agent's reason, and in --doctor: a speech stage the voice check says is not ready (model, key, voice or language). | | `The voice check did not answer (HTTP 502).` | --doctor: the agent's voice check could not be run. | | `This agent cannot check the Bot's speech settings before connecting; connecting without the check. Update the agent for clearer errors.` | The agent answered GET /version without the voice check, or answered it 404/405 (it predates the route): the command connects without checking the speech settings first. Never shown when the agent could not be reached or did not answer in time; the connect error says that. | ### Session start and prompts Four status lines appear only when the transcript goes to a terminal (standard output, or standard error when the audio goes to standard output): Listening, Sending, Waiting for the Bot, and Waiting for the agent to settle the reply. The connection lines Connected, Reconnected and Stopping always print, on standard error. With --transcript json all of them arrive as events. | Line | When it is printed | |---|---| | `Try again in 30 s.` | Appended to a connection failure the agent says may pass after a while. | | `Connected to Bot bot-7f3a.` | The voice session is open. | | `Space: talk, Space again: send. Esc: stop the reply. t: type a message. q: quit. Ctrl+C: stop now.` | Push-to-talk session on a terminal. | | `Hands-free: use headphones, or the Bot will hear itself. t: type a message. q: quit.` | Hands-free (--mode vad) session. | | `Space: talk, Space again: send your question. Esc: stop the reply. t: answer the Bot's own approval or question. q: quit. Ctrl+C: stop now.` | Push-to-talk with --once: you ask one question by voice; t only answers the Bot's own approval or question. | | `Hands-free: use headphones, or the Bot will hear itself. t: answer the Bot's own approval or question. q: quit.` | Hands-free with --once: you ask one question by voice; t only answers the Bot's own approval or question. | | `Typed messages only. t: type a message. Esc: stop the reply. q: quit.` | --input none: typed messages only. | | `The keyboard is not a terminal, so the microphone stays open (hands-free).` | Microphone input while standard input is not a terminal: there is no key to talk with. | | `Listening… (Space to send)` | Push-to-talk: recording. Shown only when the transcript goes to a terminal (standard output, or standard error when the audio goes to standard output). | | `Sending 4.2 s of audio…` | An --in-file recording is being sent. Shown only when the transcript goes to a terminal (standard output, or standard error when the audio goes to standard output). | | `Waiting for the Bot…` | The utterance reached the Bot; its reply is awaited. Shown only when the transcript goes to a terminal (standard output, or standard error when the audio goes to standard output). | | `Waiting for the agent to settle the reply…` | In a `--once` run, the utterance reached the Bot and audio output goes to a file or standard output. The agent waits for playback accounting to settle: without a playback confirmation, the reply settles 2 seconds after the later of its expected playback end and the end of the audio output. Shown only when the transcript goes to a terminal (standard output, or standard error when the audio goes to standard output). With `--audio-output none`, the command shows the normal reply-wait status instead. | | `type>` | Typing a message after t. | ### Transcript labels | Line | When it is printed | |---|---| | `you:` | Your transcribed or typed words. | | `bot:` | The Bot's reply to you. | | `bot (other):` | A reply to another message (a wake, or a chat app), shown but never taken as the answer to you. | ### Session notices | Line | When it is printed | |---|---| | `The agent did not hear back that the reply was played, so it is recorded as not confirmed.` | The speaker played the reply, but the agent did not hear back in time that it was played (a slow player). It is recorded as not confirmed; the run still succeeds. | | `` The reply was cut to fit; the whole text is in `hoody agent bots logs get --id chief`. `` | The reply text was too long for one message and was cut. | | `The Bot changed its reply while speaking; the final text is below.` | The final reply text differs from what was spoken. | | `Speech stopped. The Bot's turn keeps running.` | You pressed Esc: the speech stopped and the Bot's turn goes on. | | `Stopped.` | You pressed Esc and the reply stopped. | | `Nothing to stop.` | You pressed Esc while nothing was playing. | | `The Bot asks: Delete the old build cache? Press t, then type /approve, /deny or /allow-session.` | The Bot waits on an approval or a question. | | `The Bot asks: Delete the old build cache? Press Esc to clear the line you are typing, then press t and type /approve, /deny or /allow-session.` | The Bot waits on an approval or a question while you are typing a line. Press Esc to clear that line first; anything typed now is added to it. | | `The Bot's gate ended: your answer was applied [applied].` | The Bot's own approval or question was settled; the line says how. | | `The Bot's gate may still be waiting: your answer is not confirmed (it may or may not have been applied); check the Bot's log before trying again [unavailable].` | Your approval or answer to the Bot's own gate was not confirmed or not used; the gate may still be waiting, and the line says why. | | `Question 1: Which branch? (options: main, dev)` | One of the questions the Bot asks, with the options it offers. | | `(+3 more questions; see the Bot's log)` | The Bot asks more questions than are shown. | | `The Bot asks: Delete the old build cache? This run cannot read typed commands, and speech never answers it: answer it where you chat with the Bot.` | The Bot waits on an approval or a question, and this run cannot read typed commands. Answer it where you chat with the Bot; speech never answers it. | | `A delegate waits on a gate (session ses-41c2). Answer it where that session runs.` | A delegate waits on an approval or a question; answer it where that delegate's session runs. | | `Speech never decides a gate; type the command.` | Shown with a pending own gate: saying "yes" is a message, not a decision. | ### Connection notices | Line | When it is printed | |---|---| | `The agent closed the voice session (idle for 10 minutes).` | The agent closed the voice session: after 10 minutes without activity, because the Bot was deleted or suspended, or normally; the reason in brackets says which. A live session exits 0; with --once the run exits 1, since it ended without its answer. | | `(the connection fell behind)` | The connection could not keep up with the agent. | | `(the agent hit an internal error)` | The agent hit an internal error and closed the connection. | | `(the agent ran out of memory for voice)` | The agent ran out of memory for voice and closed the connection. | | `(the agent could not send to this connection)` | The agent could not send to this connection and closed it. | | `Reconnected.` | A reconnect succeeded. | | `Sending your recording again from the start.` | A one-shot run reconnected before its question was sent: the recording goes again from the start. | ### Interrupt notices | Line | When it is printed | |---|---| | `Stopping… (Ctrl+C again to quit at once)` | First Ctrl+C (or SIGINT/SIGTERM): stopping cleanly. | | `` The Bot's turn keeps running; its reply will be in `hoody agent bots logs get --id chief`. `` | After Ctrl+C the agent says the Bot's turn goes on. | ### Doctor and listing output | Line | When it is printed | |---|---| | `Audio check for linux (x64):` | --doctor: first line. | | `ok parec 16.1` | --doctor: a backend passed. | | `no pw-record not installed` | --doctor: a backend failed. | | `parec --raw --format=s16le --rate=16000 --channels=1 --latency-msec=20` | --doctor: the command line a backend runs. | | `Recording: parec` | --doctor: the backend picked. | | `Playback: none works. Install pulseaudio-utils (parec, pacat), pipewire-bin (pw-record, pw-play; Debian and Ubuntu), alsa-utils, sox or ffmpeg.` | --doctor: nothing works for a direction. | | `Voice check of Bot chief:` | --doctor: the agent part, after the audio programs. | | `ok text-to-speech minimax/speech-2.8-turbo (voice English_expressive_narrator)` | --doctor: a speech stage the voice check says is ready, with its model (and voice or language when set). | | `no text-to-speech minimax/speech-2.8-turbo: No API key is stored for the speech provider minimax. (reason no_credentials, diag 9b2e47c01d5f8a36)` | --doctor: a speech stage the voice check says is not ready, and why. | | `off text-to-speech not used with --audio-output none` | --doctor with --audio-output none (text-to-speech, not checked) or --input none (speech-to-text, still checked by the agent but left out of the result): the stage is not used. | | `Ready. One utterance can last up to 60 s.` | --doctor: the voice check says the Bot's speech settings are ready for these flags. | | `Not ready: fix what is marked "no" above, then run --doctor again.` | --doctor: the voice check says something is not ready. Ends the run with exit code 1. | | `This agent has no voice check. Update the agent to check speech settings and to use --audio-output none.` | --doctor: the agent predates the voice check (and text-only voice). | | `Recording:` | --list-backends: one direction. | | `parec installed` | --list-backends: the program is installed (not yet probed; --doctor probes). | | `pw-record not installed (pw-record)` | --list-backends: the program is not installed. | | `pw-record cannot list devices; use its default, or name one with --input-device.` | --list-devices: the backend cannot list devices. | | `Recording devices (parec):` | --list-devices: a backend's devices. | ## What's next