HuggingFace repos can legitimately contain 0-byte files (config stubs,
etc.). The models_are_downloaded check treated these as missing, causing
the download status to never report completion and trapping the user in
the download dialog with no way to exit.

Remove the file size check so that file existence alone is sufficient.
The empty-directory check is preserved since that genuinely indicates a
failed or incomplete download.

Signed-off-by: RyanOnTheInside <[email protected]>

* Output app logs from e2e as a github artifact

Signed-off-by: Max Holland <[email protected]>

* switch to prod daydream api

Signed-off-by: Max Holland <[email protected]>

---------

Signed-off-by: Max Holland <[email protected]>

* Optional logging of OSC messages

Signed-off-by: Thom Shutt <[email protected]>

* fix: use server-confirmed state for OSC logging toggle

Instead of optimistically updating local state (which was not reverted on
error), use the status response returned by PUT /api/v1/osc/settings to
set the UI state. Also deduplicate the OscStatus interface into a shared
OscStatusResponse type in api.ts.

Co-Authored-By: Claude Opus 4.6 <[email protected]>
Signed-off-by: Rafał Leszko <[email protected]>

* fix: rename OSC log toggle label from "Unknown only" to "Errors only"

Co-Authored-By: Claude Opus 4.6 <[email protected]>
Signed-off-by: Rafał Leszko <[email protected]>

---------

Signed-off-by: Thom Shutt <[email protected]>
Signed-off-by: Rafał Leszko <[email protected]>
Co-authored-by: Rafał Leszko <[email protected]>
Co-authored-by: Claude Opus 4.6 <[email protected]>

* fix: batch-level FPS tracking for smooth frame delivery

Replace per-frame delta FPS tracking with batch-level throughput
measurement (sum(frames) / sum(intervals)) in PipelineProcessor.

The old approach measured inter-frame deltas, but in batched pipelines
(e.g., 12 frames emitted at once), near-zero intra-batch deltas mixed
with large inter-batch gaps caused the FPS estimate to oscillate
permanently. This produced unstable frame pacing downstream.

The new approach records one sample per pipeline call as a
(num_frames, interval) tuple, then computes FPS over a sliding window
of 10 batches. This gives a stable, accurate estimate regardless of
batch size.

Benchmark results (10s runs, real PipelineProcessor with stub pipelines):

Scenario            | Metric         |     Old |     New
--------------------|----------------|---------|--------
LongLive (batch=12) | Jitter         |  84.1%  |   0.3%
                    | Stalls (>1.5x) |   7.9%  |   0.0%
                    | FPS error      |  34.3%  |   0.2%
                    | FPS stdev      |   7.23  |   0.01
SDv2 (batch=4)      | Jitter         |  29.1%  |   5.8%
                    | Stalls (>1.5x) |   4.6%  |   0.0%
                    | FPS error      |  13.8%  |   1.0%
                    | FPS stdev      |   5.24  |   0.69
Single-frame        | Jitter         |   2.3%  |   4.5%
                    | FPS error      |   2.7%  |   2.7%
                    | FPS stdev      |   1.89  |   1.35
Pause/resume        | Jitter         | 112.9%  |   0.3%
                    | Stalls (>1.5x) |   7.3%  |   0.0%
                    | FPS error      |  53.8%  |   0.2%
                    | FPS stdev      |  12.10  |   0.01

Single-frame results are equivalent, confirming no regression for
non-batching pipelines.

Signed-off-by: RyanOnTheInside <[email protected]>

* Reduce OUTPUT_QUEUE_MAX_SIZE_FACTOR to 2

Signed-off-by: Rafal Leszko <[email protected]>

---------

Signed-off-by: RyanOnTheInside <[email protected]>
Signed-off-by: Rafal Leszko <[email protected]>
Co-authored-by: Rafal Leszko <[email protected]>

When this PR cleanup action ran it was silently failing if the app was not found

Signed-off-by: Max Holland <[email protected]>

* feat: MIDI support

Signed-off-by: gioelecerati <[email protected]>

* midi mappable plugins

Signed-off-by: gioelecerati <[email protected]>

* midi ui fixes

Signed-off-by: gioelecerati <[email protected]>

* es lint midi fixes

Signed-off-by: gioelecerati <[email protected]>

* midi ui formatting

Signed-off-by: gioelecerati <[email protected]>

* feat: ui fixes for the midi input

Signed-off-by: gioelecerati <[email protected]>

* feat: add more midi inputs and remove superfluos cache management button

Signed-off-by: gioelecerati <[email protected]>

* feat: add play button to the midi controller

Signed-off-by: gioelecerati <[email protected]>

* formatting

Signed-off-by: gioelecerati <[email protected]>

* fix bug with rerenders & remove unused fun

Signed-off-by: gioelecerati <[email protected]>

* formatting

Signed-off-by: gioelecerati <[email protected]>

* feat: add MIDI prompt blend controls

Replace the old prompt slot workflow with direct prompt solo and blend controls so MIDI mappings can switch and rebalance prompts live while streaming.

Signed-off-by: gioelecerati <[email protected]>

* feat: replace inline MIDI learning badge with a toast

Signed-off-by: gioelecerati <[email protected]>

* revert: restore cache management UI to match main

Reverts the cache management changes introduced alongside MIDI support,
restoring the Manage Cache toggle and Reset Cache button to their
original behavior.

Made-with: Cursor
Signed-off-by: gioelecerati <[email protected]>

* fix build

Signed-off-by: gioelecerati <[email protected]>

* build fix, manage cache & forbid bypass of disabled params

Signed-off-by: gioelecerati <[email protected]>

* address relevant CR comments

Signed-off-by: gioelecerati <[email protected]>

* formatting

Signed-off-by: gioelecerati <[email protected]>

* address minor CR comments

Signed-off-by: gioelecerati <[email protected]>

---------

Signed-off-by: gioelecerati <[email protected]>
Co-authored-by: James Dawson <[email protected]>

…630)

* Fix vace_context_scale not sent in initial params without ref images

vace_context_scale was only included in initial stream parameters and
load params when reference images were present, because it was gated
inside getVaceParams() which returns {} without ref images. When using
VACE with input video (no ref images), the scale was never sent at
stream start, so the backend always defaulted to 1.0 regardless of
the UI slider value.

Decouple vace_context_scale from the ref-images check so it is always
sent when the pipeline supports VACE.

Signed-off-by: RyanOnTheInside <[email protected]>

* Propagate VACE params (vace_context_scale, ref_images) for Krea pipeline

The Krea branch in _load_pipeline_implementation inlined its VACE setup
to use the 14B checkpoint instead of the default 1.3B, but this skipped
the vace_context_scale and ref_images extraction that _configure_vace
provides. Add the missing parameter extraction and tests.

Signed-off-by: RyanOnTheInside <[email protected]>

---------

Signed-off-by: RyanOnTheInside <[email protected]>

- Add retry logic for transient VP8 decode errors in cloud WebRTC input loops
- Fix PLI keyframe request for aiortc >= 1.14.0 (pass media_ssrc)
- Poll for remote streams before sending PLI instead of blind retries
- Log first frame receipt to verify PLI/keyframe roundtrip

Signed-off-by: emranemran <[email protected]>

Fix WebRTC VP8 decode errors and PLI keyframe requests

* Add native Workflow export to daydream.live

Signed-off-by: Thom Shutt <[email protected]>

* Handle URL opening better when running in browser

Signed-off-by: Thom Shutt <[email protected]>

---------

Signed-off-by: Thom Shutt <[email protected]>

When the fal WebSocket closes (e.g. MAX_DURATION_EXCEEDED timer), the
WebRTC peer connection (UDP) between the client and the scope backend
continues running independently — the pipeline keeps processing frames
and the video keeps streaming.

Add a DELETE /api/v1/webrtc/offer/{session_id} endpoint to the backend,
and call it from fal_app.py's finally block to explicitly tear down the
WebRTC session when the signaling WebSocket closes.

Signed-off-by: emranemran <[email protected]>

* Cache pipeline schemas and plugin list responses

The /api/v1/pipelines/schemas and /api/v1/plugins endpoints were
recomputing everything on every request (schema generation, distribution
scanning, update check subprocesses). Since these don't change until a
plugin install/uninstall (which restarts the server), cache the responses
after first computation.

Co-Authored-By: Claude Opus 4.6 <[email protected]>
Signed-off-by: Rafal Leszko <[email protected]>

* Fix test: clear plugin/schema caches between tests

The cached plugin list response from a previous test was being returned
instead of the mock, causing test_returns_plugin_list to fail.

Co-Authored-By: Claude Opus 4.6 <[email protected]>
Signed-off-by: Rafal Leszko <[email protected]>

---------

Signed-off-by: Rafal Leszko <[email protected]>
Co-authored-by: Claude Opus 4.6 <[email protected]>

Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: leszko <[email protected]>