Earlier versions made the server's proxies of HEIF frames from their embedded
previews. Those carry the frame's capture time, so the proxy pass kept them
and the sequence stayed marked as having all its proxies; renders and
analysis then read thumbnail-sized files and no client or render node was
asked for a real one. The proxy pass now treats a proxy whose long edge falls
short of the policy's for the frame (its own size when smaller, a tenth of
slack, frames of unknown size not judged) as stale, and the server refuses
such a proxy when one is sent (422 proxy_too_small).
PUT /api/v1/frames/{id}/proxy also took a proxy of a frame the server
decodes itself, replacing its own proxy with any JPEG of the right capture
time; it now answers 422 proxy_not_wanted for those and for frames missing or
damaged. Its follow-up jobs are queued even when the sender disconnects once
the proxy is filed.
A render node cleared the server's Undecodable flag on every frame, so a
damaged frame the server's own render skips was downloaded and failed the
node's job. Only a format the server cannot decode and the node can loses the
flag now; other undecodable originals are neither fetched nor rendered unless
read from their proxy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
The server decodes with managed decoders only, so HEIF originals had no
proxies, thumbnails from their small previews and no pixels for analysis or
managed renders, and nothing could send it proxies (ENG-05). Manifests now
carry the archive's proxy spec and, per frame, its flags and whether the
server wants its proxy. PUT /api/v1/frames/{id}/proxy (Edit or RenderNode)
stores a JPEG of at most 64 MiB that decodes and whose Exif capture time is
the frame's, then queues thumbnails and the proxy pass. An upload from the
desktop makes and sends the wanted proxies after the originals and counts
them as local changes; a render node sends those it can make from the
originals it fetched once the job is done. The server never makes such a
proxy from a preview, drops one left by a renumbered frame, takes thumbnails
from a proxy, and analyses and renders those frames from their proxies.
Render nodes fetched every frame of the manifest: frames missing, corrupt or
excluded on the server failed the download, and originals not hashed yet
failed the SHA-256 check (ENG-07). Nodes now skip those frames, fetch unhashed
originals into the job folder with a size check, judge decodability with their
own decoders, and use the proxy of a frame they cannot decode.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
The heartbeat check slept 300 ms to get past the claim that finds the queue
empty and then gave heartbeats 400 ms to move the node's last sign of life: a
slow runner could fail it, or pass it on the claim's own touch. It now counts
the claims the server answered, reads the last beat after the third one, waits
for a later beat with a deadline and checks that no claim came since.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Originals are resumed from their partial file because their SHA-256 checks
the result. Proxies have no hash on the server and can be rebuilt between
two attempts, so appending to what a dropped attempt left could mix two
versions; a retry now fetches the proxy whole.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
A node announced one size limit, the largest of all its encoders (managed
MJPEG at 65535 px), so it claimed H.264 jobs larger than its H.264 encoder
takes. Each offered encoder now carries its own limit and the server matches
codec and size per encoder; the node-wide limit is the largest offered one,
for encoders that announce none. A hand-written batch with null or blank
entries no longer throws when a job is posted or examined.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
A node registering again gives back the job its previous run abandoned, so
a restart within the heartbeat timeout no longer leaves it running forever.
Frame downloads retry dropped connections from the partial file; a 404 at
registration waits for the next poll; render remarks reach the job log; a
render the server engine started stays with it. POST /jobs splits a batch
into one job per preset, and a node fails a batch it cannot send back
instead of reporting it done. Renders honour Settings > Rendering: hardware
encoders only where the preference allows them, also for what a node offers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC