Repository navigation
Reelsmith v0.2.2
Brief safety and missing-input hints
- A rejected
write_briefrequest no longer overwrites the previous workingbrief.toml. The new brief is validated in a separate file before replacement. - Missing clip and motion-asset names now include nearby existing-name suggestions, reducing the need for another folder-listing call after a typo.
- Suggestions exclude extension-case combinations that would leave the input unresolved on case-sensitive filesystems. This is conservative on case-insensitive filesystems, where some usable hints may be omitted.
Validation and limits
The brief-preservation fix was checked through real MCP calls: a negative-duration request leaves the old brief unchanged; valid updates, missing-input reports and concurrent successful writes still work. A fresh 360p procedural smoke render passed duration, dimensions, loudness and black-frame QA, and its sampled contact sheet was inspected.
Concurrent writes remain last-successful-writer-wins. Replacing a symlink destination replaces the link itself rather than writing through it. Abrupt process termination may leave a temporary file. No new performance gain is claimed; the audio-overlap experiment is not included.
The release workflow builds four platform archives and smoke-checks the MCP surface. The downloaded Linux x86_64 archive matched its published SHA256 and passed the 0.2.2 / 14-tool MCP smoke, rejected-edit SHA256 preservation, typo-hint checks, and a fresh uncached 360p render (1.3 seconds, -13.95 LUFS, -8.96 dBTP, no detected black frames). Sampled contact-sheet pixels were inspected. On the older local Linux runtime it printed a non-fatal weak GLIBC_2.39 version warning; these checks still completed, but this is not a general Linux compatibility guarantee. Windows, macOS and Linux ARM64 are not claimed as manually runtime-tested.
Known issue found after publication
A hand-edited config with procedural music and an empty music.progression = [] can panic during audio generation and leave that MCP session's render job stuck as running. Avoid empty procedural progressions; if already stuck, restart the MCP server process. The empty-progression trigger is fixed in v0.2.3 (PR #53), verified with the published Linux binary in the same-server invalid-then-valid flow. Upgrade to v0.2.3 for this fix. This issue predates the brief-preservation change.