feat(voice): message listening versions persist and sync across clients

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.

[pandec]

Creating a listening version was one long HTTP request owned by the message row. The spinner lived in component state, so leaving the thread hid all progress, a disconnect aborted the synthesis, and other clients never learned a recording existed because the server wrote the projection table directly and emitted no event.

Speech synthesis is now an event-sourced job. A thread.message.speech.request command validates against the SQL projection under a shared per-message lock and emits a requested event that marks pending state on the message. A reactor synthesizes the audio (concurrent across messages, serialized per message, 3-minute timeout) and completes through thread.message.speech.complete; both projectors store the artifact and clear pending, and subscribed clients receive both transitions as live deltas. Recordings survive navigation, reconnects, and device switches, user recordings are event-owned and survive projection replay, and interrupted jobs are cleared on startup. The legacy HTTP endpoint joins the same jobs and keeps typed failure reasons, so old clients keep working. Web and mobile derive button state from the projection when the server advertises textToSpeech.persistentJobs and keep the old path against older servers.

Implemented by gpt-5.6-sol subagents, orchestrated and reviewed (gpt-5.6-sol + Opus 5 panels) by Claude Fable 5 in Claude Code.