Integrate cactus-based local STT pipeline and shared LLM types#4003
Integrate cactus-based local STT pipeline and shared LLM types#4003
Conversation
✅ Deploy Preview for hyprnote-storybook canceled.
|
✅ Deploy Preview for hyprnote canceled.
|
Introduce new cactus/cactus-sys/transcribe-cactus crates, wire plugins to the new server path, and retire the older VAD-specific integration points. Co-authored-by: Cursor <cursoragent@cursor.com>
Track vendor/cactus by gitlink so checkout and sync follow .gitmodules rather than vendoring full source into the monorepo. Co-authored-by: Cursor <cursoragent@cursor.com>
Align the vendored cactus dependency to the latest 1.7 release and clear local submodule drift so builds use a reproducible upstream tag. Co-authored-by: Cursor <cursoragent@cursor.com>
yujonglee
force-pushed
the
feat/cactus-stt-integration
branch
from
25c6501 to
162e667
Compare
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
| lfs: true | ||
| fetch-depth: 0 | ||
| fetch-tags: true | ||
| submodules: recursive |
There was a problem hiding this comment.
ARM64 Linux runner builds x86_64 binaries
High Severity
The Linux build job now runs on depot-ubuntu-24.04-arm-8 (ARM64 runner) but still builds for x86_64-unknown-linux-gnu target. Cross-compilation from ARM64 to x86_64 on Linux typically requires additional toolchain setup that isn't present in the workflow, causing build failures or producing incorrect binaries.
| /> | ||
| <button | ||
| onClick={handleUseCactus} | ||
| disabled={active || !modelPath} |
There was a problem hiding this comment.
Cactus button remains disabled without model path
Low Severity
The "Use Cactus" button is disabled when !modelPath, but the Rust backend has fallback logic that tries multiple paths including /tmp/cactus-model and models_dir/whisper-small. Users with valid fallback paths can't activate cactus because the UI incorrectly requires the setting to be populated, despite the backend working without it.
| .map_err(|e| format!("failed to flush temp file: {}", e))?; | ||
|
|
||
| let model = | ||
| hypr_cactus::Model::new(model_path).map_err(|e| format!("failed to load model: {}", e))?; |
There was a problem hiding this comment.
Batch transcription loads model on every request
Medium Severity
transcribe_batch calls hypr_cactus::Model::new(model_path) on every HTTP batch request, loading the model from disk each time. Model initialization is typically an expensive operation (seconds), making the batch endpoint impractically slow for any repeated use. The streaming (WebSocket) path loads the model once per connection in run_transcriber, but the batch path has no such caching. The TranscribeService already holds model_path; a shared, lazily-initialized model instance would avoid the repeated load cost.
Additional Locations (1)
| "{:04}-{:02}-{:02}T{:02}:{:02}:{:02}.{:03}Z", | ||
| year, month, day, hours, minutes, seconds, millis | ||
| ) | ||
| } |
There was a problem hiding this comment.
Custom calendar arithmetic reimplements standard library functionality
Low Severity
format_timestamp_now is a 55-line manual implementation of epoch-to-ISO-8601 conversion, including leap-year calculation and month-day table iteration. The chrono crate is already in the workspace dependency tree and would replace this entire function with a single expression like Utc::now().to_rfc3339_opts(SecondsFormat::Millis, true). Hand-rolled calendar math is hard to review and maintain.
| .collect::<String>(); | ||
|
|
||
| assert_eq!(restored, text); | ||
| } |
There was a problem hiding this comment.
Flaky test asserts wrong value on successful parse
Medium Severity
The second block of test_summary asserts restored == text where text includes raw <think>...</think> tags. But when the parser successfully parses a think block, parse_think_block strips the tags and trims the content. The restored string then omits the tags and surrounding whitespace, making the assertion fail. The test only passes by accident (~80% of runs) when random chunk sizes split the <think> tag, causing the parser to fail and preserve raw tags as TextDelta.
| permissions: | ||
| contents: write | ||
| runs-on: depot-ubuntu-24.04-8 | ||
| runs-on: depot-ubuntu-24.04-arm-8 |
There was a problem hiding this comment.
Linux builds switched from x86_64 to ARM
High Severity
The Linux runner changed from depot-ubuntu-24.04-8 (x86_64) to depot-ubuntu-24.04-arm-8 (ARM) in CD, and similarly in CI. Since the Linux build job has no explicit --target cross-compilation flag, it produces binaries matching the host architecture. This means release artifacts will now be ARM-only, breaking the app for all x86_64 Linux users.
Additional Locations (1)
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.
| } | ||
| return true; | ||
| }); | ||
| const models = configuredProviders?.[providerId]?.models ?? []; |
There was a problem hiding this comment.
Removed filter allows selecting non-downloaded models
Medium Severity
The previous code filtered out Quantized* models that weren't yet downloaded, preventing users from selecting unavailable models. That filter was entirely removed, so all models now appear in the dropdown regardless of download status. Users can select a Quantized model that hasn't been downloaded, which will fail at runtime when the server tries to load a nonexistent model file.
1 participant
Summary
cactus-sys,cactus, andtranscribe-cactuscrates to support cactus-backed local STT with both streaming and batch service layers.llm-typescrate and wire dependent crates/plugins to consume the new interfaces.Test plan
cargo checkMade with Cursor