Save recordings before transcription by ferologics · Pull Request #1024 · cjpais/Handy

ferologics · 2026-03-12T22:16:31Z

Summary

This PR fixes the data-loss path where a recording could disappear if transcription failed.

Save recordings before transcription

persist the WAV immediately after recording stops
create/update history entries for pending, completed, and failed states
preserve failed/pending entries in History instead of dropping them on cleanup
store transcription_error
allow retrying transcription from History using the saved WAV
keep tray copy behavior limited to completed entries
simplify History status copy so the badge/player/actions carry most of the state

Testing

bun run build
cargo check -j 1
manual verification of failed/pending History UI and retry flow using saved WAVs

cjpais · 2026-03-12T23:20:49Z

Let's not do the ORT error fix here.

We should be adding an option to transcribe-rs to do chunking. This is code which should be broadly applicable not just for Handy

It will also keep this PR more focused

ferologics · 2026-03-12T23:41:07Z

Dropped the ORT split-retry commit from this PR to keep it focused on the durability/history fix. I kept local notes and a preserved branch for a follow-up chunking/chunked-transcription pass in transcribe-rs.

VirenMohindra

minor nits, i think we can reduce the blast radius if we add back the comments + would be easier to review

VirenMohindra · 2026-03-19T05:57:15Z

src/bindings.ts

 },
+async retryHistoryEntryTranscription(id: number) : Promise<Result<null, string>> {
+    try {
+    return { status: "ok", data: await TAURI_INVOKE("retry_history_entry_transcription", { id }) };


needs a rerun of codegen + bun run format to fix

VirenMohindra · 2026-03-19T05:59:36Z

src-tauri/src/actions.rs

-                                error!("Failed to run paste on main thread: {:?}", e);
+                            }
+
+                            if processed.final_text.is_empty() {


when processed.final_text.is_empty(), the code silently hides the overlay and does nothing else but complete_transcription() was already called with an empty string, leaving the entry as status=Completed, transcription_text="". In the UI this shows as a successful entry with no text and users can't distinguish "completed but empty" from "failed." we should either call fail_transcription() with a descriptive error or skip complete_transcription() entirely for empty results

VirenMohindra · 2026-03-19T06:02:43Z

src-tauri/src/managers/history.rs

-            crate::settings::RecordingRetentionPeriod::Days3 => now - (3 * 24 * 60 * 60), // 3 days in seconds
-            crate::settings::RecordingRetentionPeriod::Weeks2 => now - (2 * 7 * 24 * 60 * 60), // 2 weeks in seconds
-            crate::settings::RecordingRetentionPeriod::Months3 => now - (3 * 30 * 24 * 60 * 60), // 3 months in seconds (approximate)


nit: would it be possible to avoid blowing this diff up and avoid removing comments? would be easier to review

cjpais · 2026-03-19T06:05:25Z

I think I have some inflight changes for this PR

VirenMohindra

some misc UI fixes ordered in terms of prio~

the retry button shows on completed entries and should only appear when !hasTranscription
content area shows "Transcription failed" during retry and should instead swap to a loading state
delete btn stays enabled during retry, and will be a possible race condition if you delete while retranscribing

i think we should do a proper UI pass before this lands

VirenMohindra · 2026-03-19T08:11:12Z

src-tauri/src/actions.rs

+                } else {
+                    // Save WAV concurrently with transcription
+                    let sample_count = samples.len();
+                    let file_name = format!("handy-{}.wav", chrono::Utc::now().timestamp());


lets use a uuid here instead. timestamp is a second class citizen and two recordings within the same second will overwrite each other

or even chrono::Utc::now().timestamp_nanos_opt().unwrap_or_else(|| chrono::Utc::now().timestamp_micros()) but il lean towards uuids

VirenMohindra · 2026-03-19T08:12:17Z

src-tauri/src/actions.rs

+                                    .await;
+
+                            // Save to history if WAV was saved and transcription is non-empty
+                            if wav_saved && !transcription.is_empty() {


orphaned wav files when there is silent audio and no transcription and then cleanup_old_entries won't touch it

the error path below this correctly writes an empty DB row for retry and the empty-transcription path should do the same

VirenMohindra · 2026-03-19T08:12:58Z

src-tauri/src/actions.rs

+                            // Verify: 44-byte header + num_samples * 2 bytes (16-bit)
+                            let expected_size = 44 + (sample_count * 2) as u64;


seems a bit fragile, rereading the header with WavReader::open and comparing sample counts is prob more robust

VirenMohindra · 2026-03-19T08:14:12Z

src-tauri/src/commands/history.rs

+
+    transcription_manager.initiate_model_load();
+
+    match transcription_manager.transcribe(samples) {


lets wrap this with tauri::async_runtime::spawn_blocking if possible

this can run for 5 to 30s depending on model and hardware. calling it directly inside an async fn would block the thread for the entire time

VirenMohindra · 2026-03-19T08:14:59Z

src-tauri/src/commands/history.rs

+            history_manager
+                .update_transcription(
+                    id,
+                    transcription,
+                    processed.post_processed_text,
+                    processed.post_process_prompt,
+                )


UI has no way to distinguish "retry succeeded with silence" from the original failed state and the user will retry again with the same result indefinitely

we should prob return an error when transcription.is_empty() so the fe can show a meaningful state

cjpais · 2026-03-19T09:42:04Z

I think retry should be for everything, mainly if you want to re-transcribe with another model or you picked the wrong language (happens to me all the time)

Otherwise other 2 I agree

VirenMohindra · 2026-03-19T16:54:57Z

also please add screenshots to the PR

- Save recordings before transcription to prevent data loss on crash (cjpais#1024) - Linux Wayland titlebar: disable decorations for custom titlebar (cjpais#1042) - Add change_secondary_selected_language_setting command (cjpais#1106 gap) - Fix CLI --auto-submit key type being ignored (cjpais#1099 bug)

cjpais · 2026-03-21T05:00:10Z

@VirenMohindra let me know what you think of the current set of changes. Not sure if the original author will jump in, but I think things are doing okay right now. Would love a second opinion

cjpais · 2026-03-21T05:18:22Z

github-actions · 2026-03-21T06:40:17Z

🧪 Test Build Ready

Build artifacts for PR #1024 are available for testing.

Download artifacts from workflow run

Artifacts expire after 30 days.

VirenMohindra

LGTM, the ui feedback is addressed, thanks. the spawn_blocking and filename collision are worth fixing before this lands because both are easy changes and the former can stall the runtime for 30s+

the header sample count check and empty transcription handling can be follow-ups if needed

ferologics · 2026-03-27T19:05:55Z

thx for getting this across the finish line @cjpais & @VirenMohindra !!
now we just need to recover from failures better :P

ferologics force-pushed the main branch 2 times, most recently from 97eaaac to 47f44a8 Compare March 12, 2026 22:53

ferologics marked this pull request as draft March 12, 2026 23:01

ferologics force-pushed the main branch from 47f44a8 to 35857a2 Compare March 12, 2026 23:40

ferologics changed the title ~~Save recordings before transcription and retry Parakeet ORT failures~~ Save recordings before transcription Mar 12, 2026

ferologics marked this pull request as ready for review March 12, 2026 23:50

save recordings before transcription

13cf43e

ferologics force-pushed the main branch from 35857a2 to 13cf43e Compare March 13, 2026 16:05

cjpais added 2 commits March 19, 2026 08:48

Merge branch 'main' into ferologics/main

68eab05

impart cj preferences

af15874

VirenMohindra reviewed Mar 19, 2026

View reviewed changes

cjpais added 6 commits March 19, 2026 15:40

cleanup

8090e51

re-transcribe anything

4588561

Merge branch 'main' into ferologics/main

869818d

format

9fb42a9

fix tests

fb2879e

format

ce20b1e

VirenMohindra reviewed Mar 19, 2026

View reviewed changes

cjpais added 3 commits March 21, 2026 09:40

ui cleanup

b078f8f

format

93dbdf3

Merge branch 'main' into ferologics/main

e75c8af

cjpais mentioned this pull request Mar 21, 2026

improve history performance #1107

Merged

cjpais added 2 commits March 21, 2026 12:21

Merge branch 'main' into ferologics/main

f4f11f4

cleanup

8c6d5c8

bunch of fixes, mainly for vm

b717b5a

put comments back

00b80d0

cjpais added 2 commits March 21, 2026 14:25

update transcription failure text

d37368a

alert -> toast

781bd4a

VirenMohindra approved these changes Mar 21, 2026

View reviewed changes

spawn_blocking

4dc64ae

cjpais merged commit 17277cf into cjpais:main Mar 22, 2026
4 checks passed

		// Verify: 44-byte header + num_samples * 2 bytes (16-bit)
		let expected_size = 44 + (sample_count * 2) as u64;


		transcription_manager.initiate_model_load();

		match transcription_manager.transcribe(samples) {

Uh oh!

Conversation

ferologics commented Mar 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary