[Feature]: Read agent responses aloud on mobile with background TTS

[AdamCLarsen]

Description

Add a Read Aloud action to agent responses in the T3 Code mobile app.

A common workflow for me is receiving a notification that an agent has finished while I am away from my desk. I want to open the response, tap Read Aloud, put the phone back in my pocket, and listen while I continue doing something else.

This is different from screen-reader accessibility or playing an audio file generated by an agent. The feature would take the text of an existing agent response and synthesize speech from it on demand.

Proposed UX

Add Read Aloud alongside the existing actions such as Copy on completed agent responses.

Tapping it should start reading that response and provide basic play/pause/stop controls.

Playback should continue when:

• T3 Code is in the background
• the user switches to another app
• the phone is locked

Where supported, playback should also integrate with the normal lock-screen / system media controls.

TTS providers

Native iOS/Android text-to-speech would be a useful zero-configuration default.

It would also be useful to allow optional BYO API keys for higher-quality TTS providers such as:

• ElevenLabs
• Deepgram
• OpenAI

These providers offer substantially more natural voices than typical system TTS. Users should be able to select the provider and voice, with native TTS remaining available without any external account.

Smallest useful scope

For a first version:

1. Read Aloud action on a completed agent response.
2. Native system TTS.
3. Play/pause/stop.
4. Playback continues while the app is backgrounded or the phone is locked.

External TTS providers and voice selection could be added without changing the basic interaction.

Related work

Related, but not duplicates:

• [Feature]: Read AI responses aloud via built-in TTS (per-message "read aloud" action) #3964 requests per-message TTS/read-aloud for the desktop app.
• [Feature]: Render agent-generated audio files inline #4587 is about rendering and playing audio files already generated by an agent.
• Existing screen-reader accessibility work makes the UI navigable with assistive technology; this request is for an explicit on-demand listening mode in the mobile app.

The main mobile-specific requirement here is reliable background playback so an agent response can be listened to after opening it from a completion notification.

[Kin230k]

A useful way to keep the first version small is to define a transport-neutral playback session before adding provider choice. I would treat one completed response as an immutable source snapshot and derive speakable segments with stable offsets. Markdown syntax can be removed while preserving heading and list order; code blocks and long URLs could be skipped by default with a short spoken marker and an explicit option to read them. Otherwise technical responses become difficult to follow when TTS reads fences and tokens literally.

The minimum lifecycle matrix I would pin is foreground, background, and locked playback; play, pause, stop, and segment navigation; phone-call or audio-focus interruption; Bluetooth route change; and response regeneration. Resume should only restore the same response revision, never begin reading a newer regenerated answer without the user choosing it.

Native TTS is also a clean privacy boundary for v1. If external providers arrive later, network synthesis should be explicit at provider selection and playback, show what response text leaves the device, and keep credentials in platform-secure storage. That prevents the zero-configuration path from quietly changing its data boundary.

For acceptance, lock-screen metadata can identify the project and response without exposing response content; system media controls should stay synchronized; the completion notification should deep-link to the exact response; and screen-reader navigation and Read Aloud should remain independent so one does not hijack the other. That playback contract seems more valuable for the first slice than provider selection, which can sit behind it later.

[Porkstone]

Would love this especially when I am out walking I need to look where I am going not at the screen.
I would basically want to have a conversation with the agent.
Sometimes I need to get away from my screen and being able to talk to the agent about a new feature would be great.

[PrismaMind]

Adding the Android specifics from #18501, which was closed as a duplicate of this thread.

On Android, TextToSpeech.speak() accepts at most TextToSpeech.getMaxSpeechInputLength() characters. That platform method returns 4000. A longer call does not read the rest of the message. Please split the message into chunks at or under that limit, on a paragraph or sentence boundary when one is nearby. Speak the first chunk with QUEUE_FLUSH and the following chunks with QUEUE_ADD, so one press reads the whole message in order.

Android TTS has no pause call, only stop(). The same button should pause and resume:

• While speaking, a second press calls stop() and keeps the character offset.
• The next press continues from that offset, not from the start.
• If another app takes audio focus (AUDIOFOCUS_LOSS or AUDIOFOCUS_LOSS_TRANSIENT), treat that as the same pause and keep the offset. Coming back to T3 Code should not restart from the first character.
• Remember the offset per message id. Starting another message stops the current one and leaves the previous offset in place.

UtteranceProgressListener.onRangeStart (API 26+) reports the range inside the current chunk. The saved position is the chunk start plus that range start. Some engines never fire onRangeStart. In that case remember the start of the chunk that was playing, so resume still does not jump back to the first character.

Read the same string the copy button already uses (presentation.text for a user message, renderedText for an assistant message). The control in ThreadFeed.tsx can sit beside CopyTextButton.

[Silverstar187]

Disclosure: I built this, so take it with that in mind.

On the "have a conversation with the agent" part from @Porkstone: there is a workaround today that sits next to T3 Code, not inside it. voicehook (https://voicehook.ai/?utm_source=github&utm_campaign=d8-t3code).

You start a voice call in the browser, on your phone if you like. Tap "invite agent" and paste the link into the running Codex or Claude Code session. The agent joins over plain HTTPS/curl, nothing to install. Instructions for the agent: https://voicehook.ai/agent/SKILL.md?utm_source=github&utm_campaign=d8-t3code

The voice in the call answers from a status board and FAQ that the agent keeps up to date. So you can ask where it's at while it keeps working. About 10 free minutes every day, no login needed.

It is not Read Aloud inside the app, so it doesn't solve this request. A native version would still be nicer.