Skip to content

v0.26.3

Choose a tag to compare

@github-actions github-actions released this 27 Oct 03:37
· 350 commits to main since this release
62b5e98

Patch Changes

  • #185 1d0bceb Filler and repeat removal with the language model (DISFLUENCY_LLM) and the AI packaging of clips you cut by hand now save their work as they go, so a timeout or restart continues where it stopped. A timeout used to finish the filler pass with only part of the transcript checked and no warning; it now fails the step so Retry completes it. A long selection for "find the best moment" is searched in parts instead of one prompt.

  • #185 f3a94cc Shot cuts are saved the moment they are found and the on-screen text scan is its own resumable step, so a text-scan timeout no longer loses them. A failed shot-change decode now retries and, if it keeps failing, shows as a failed scenes step with its reason instead of being stored as "no cuts".

  • #185 bf26baf Transcription now saves each audio chunk as a step, so its progress, time left and timeouts behave like finding moments and a retry continues after the last finished chunk. With S3 storage the speech-to-text cache is shared across hosts. A "no progress" warning now says which stage is stuck, how many parts are done and when the last one finished, and it also fires when a model keeps ticking but finishes no part.

  • #185 389a4de Fixes several races that could leave a project stuck or out of date: re-analyze no longer overwrites a settings change made at the same time, Retry after a failed re-analyze picks clips again, deleting the last clip of a rendering project finishes it, and a rename during the download survives. Refining a busy project is refused instead of overwriting its status.

  • #185 e3d14fe New MCP tools for agents: review and correct transcript fixes (list_transcript_flags, review_transcript_flags, add_transcript_correction, add_transcript_cut), read and set project defaults, open the billing portal, and start a project from an uploaded file with create_project's new uploadId.

  • #185 3f6c629 "Not found" answers now say what is missing in plain words, and the MCP clip tools report what changed and whether a render was queued (update_clip, create_clip, delete_clip(s), rerender_clip(s); get_clip also shows the style and framing it will render with and where they come from). Editing a clip's length now has the same 3 second to 10 minute limits as creating one. The command line can create projects, retry, re-analyze, refine, delete, export and create or edit clips, rerender prepares clips that never got a title, and recover retries a failed project. The templates and library lists load a page at a time.

  • #185 86e9b35 Each Reelsaw process now writes its own log file (reelsaw.<worker id>.<date>.log), so scaled workers no longer delete each other's logs; reelsaw logs and issue reports read and merge them all. Stopping the app now ends open event streams and waits briefly for open requests before draining jobs, and STOP_GRACE_PERIOD in .env sets how long Docker waits (raise it with SHUTDOWN_GRACE_SECONDS). Analysis and refine steps get more time on long videos (up to 3 hours, twice the video length), and the database pool is configurable with DB_POOL_MAX, with connect and idle-transaction timeouts, so model calls cannot starve job heartbeats. Upgrade: the release notes now list the irreversible migrations; take a pg_dump before upgrading.

  • #185 59de27d A job that keeps crashing its worker now fails with "Processing was interrupted 4 times" and a Retry button instead of being re-queued forever, and a worker short of disk no longer hangs while handing its job to another worker. /api/health now checks the database, live-update connection, job queue and worker heartbeat and answers 503 when one is down, and the live-update connection keeps retrying after a failed reconnect.

  • #185 2d994c7 A plan's free first video now goes to your first video whose length is known, not to one that was still downloading or failed early. Uploads that became a project are no longer counted twice against storage, and the storage limit is enforced on every upload path, including library imports and parallel uploads. Links longer than your allowance are refused before they are downloaded, and are not retried. Re-analyze and refine runs now count toward the hourly re-render limit. An unknown Stripe price is reported to admins, and an account with an existing Stripe subscription is sent to the Billing Portal instead of a second Checkout.

  • #185 7bef402 Follow-up fixes from review: a lost job no longer burns a retry when it was still running, uploads without a size reserve disk for the largest allowed file, a deleted project's late job is no longer reported as a failure, and the analysis step no longer times out on work that follows a finished step. Changing source text or caption style now tells you which clips were still rendering and need a re-render, cancelling uploads keeps projects that were already being created, and API keys without a permission record are no longer administrators (older keys were granted it explicitly, visible in Settings).

  • #185 2bc2caa API keys now have permissions: admins choose whether a key may change instance settings over MCP, and OAuth connectors never can. Connected apps can be reviewed and revoked in Settings, the consent screen shows where approval returns to, and client registration is rate-limited. New TRUSTED_PROXY_HEADER and TRUSTED_PROXY_HOPS settings stop spoofed client addresses, server-side link fetches can no longer be redirected inside your network, CSV exports neutralise spreadsheet formulas, uploaded SVGs from S3 keep their sandbox, and sign-up no longer reveals whether an email has an account.

  • #185 92351f2 Long videos on slow local models no longer fail after 30 minutes: finding moments now saves each part of the transcript as it finishes, so Retry (and any automatic retry) continues where it stopped instead of starting over, and the project page shows "13 of 43 parts, about 1 h left". A timeout now says how far it got and what to do, and two new settings, LLM_CALL_TIMEOUT_SECONDS and STAGE_IDLE_TIMEOUT_MINUTES, set how long a model call and a stage without progress may take.

  • #185 bd0e4b8 A crash during audio extraction can no longer leave a truncated file that a retry transcribes silently. Clip retention now frees the disk it promised (clips and thumbnails, local and S3 scratch), a project whose source was removed by retention is fetched again by Retry (clips then re-render on demand, with a clear message until then), abandoned uploads are cleaned up after a day, and S3 storage checks write access at boot and handles concurrent source pulls safely.

  • #185 b44e634 Agents and the CLI can now hide or keep text burned into the source (set_project_source_text, reelsaw source-text) and change a project's caption style (update_project_style, reelsaw update-style), as the screens already could. Docs say that a retry on a long video resumes from the last finished part.

  • #185 2093120 Live updates on the project page and the clip editor now recover after a sign-in expiry or a network blip (with a small "reconnecting" note), and a long activity log stays complete after a reconnect. Uploads check the file size against the server's limit up front, can be cancelled one by one or all at once, and warn before you close the tab mid-upload.


Install or upgrade: reelsaw-0.26.3.zip below holds the compose file and installer pinned to this release; an existing install runs docker compose pull && docker compose up -d. Details in deploy/README.md. Full history: CHANGELOG.md.