Repository navigation
Releases: Alaneel/video-toolkit
Release list
Video Agent source preview r3 — bounded feedback
Source-only review preview · 2.2.0-rc.1
This preview adds a bounded feedback loop to Video Agent: render an unapproved preview, review actual output/source JPEGs, validate an evidence-linked model revision and re-render within the existing task budgets. The native UI and versioned JSON CLI expose the same engine.
Candidate adoption is explicit. Adopting a new candidate revokes previous approval; final export still needs human confirmation of the exact saved revision. Separate preview-upload consent is required. Unknown external calls are not automatically resent, while saved responses can be settled and resumed idempotently.
Local verification: 179 ordinary Rust tests, 49 distinct real-media tests (including hard-kill recovery), 78 packaging checks and 61 evaluation-harness checks passed; strict workspace Clippy and formatting passed. PR #14 is merged. Quality/platform CI and four-target builds passed; public package steps were skipped and artifact count is zero.
Post-merge verification on master fc313562314cfd45ee232437847a7972669fb020 also passed: master CI and master four-target builds. Distribution gates remain in place; this source preview has no binary attachments.
Actual-model evidence and limits
Two synthetic red/blue development runs made five real GPT-6 Luna review calls. Both exercised real render → revision → re-render, but neither reached model acceptance: one stopped on unchanged edit ranges; the other rejected the second response. Neither selected, adopted, approved or exported a candidate. The failures, usage and media identities are preserved in docs/smart-feedback-evidence.json.
The earlier user-authorized 20-second rabbit edit remains a separate playable demonstration. Sparse-image model self-assessment is not full-video understanding, independent human semantic evaluation or evidence of improvement over uniform sampling. Formal semantic evaluation remains pending.
Distribution
No Agent binaries are attached. Six unresolved original dependency-license-text gaps continue to block public binary distribution. FFmpeg/ffprobe/ffplay remain external. The first runtime-validated Agent target is macOS Apple Silicon; other platform builds do not establish desktop runtime acceptance. No Developer ID signing or notarization is claimed.
Source preview r2 and stable Toolkit v2.1.0 are unchanged. This prerelease does not replace the latest stable release.
Video Agent 2.2.0-rc.1 — source preview r2
Video Agent — experimental source preview r2
Safety correction: this replaces the withdrawn first Agent preview. Do not run Unix crash tests from earlier Agent snapshots. r2 uses explicit process-group signal operands, checks the owned child's process-group identity and keeps its PID reserved until cleanup. The defect was confined to crash-test code, not the normal video-processing path. Details. Existing Toolkit v2.1.0 releases do not contain these new Agent tests.
A native Rust desktop workflow for timestamped observations, model-proposed clips, human editing and verified MP4 export. Built on the existing Video Toolkit media engine.
What is here
- Two bounded vision tools: request interval observations or propose evidence-linked clips. The model cannot approve or export.
- A native review surface for source playback, trimming, rejection, reordering and exact-revision approval.
- Persisted task state, conservative execution accounting, cancellation and verified-stage/shard recovery.
- A real OpenAI Responses backend, now using the user-approved GPT-6 Luna default.
- Frozen public-material recipes, a uniform baseline and failure-preserving evaluation capture tools.
Actual validation and limits
On the public development excerpt, Luna made two follow-up observation requests and produced a structurally valid five-clip, 25-second draft after a retained citation rejection. One clip lacks the requested visible rabbit. At the r2 source-tag snapshot, user-directed editing and model-origin export were awaiting the author's decision; see the subsequent update below. Formal human semantic scores remain pending. Three earlier legacy-model proposals also remain visible in the results. A valid schema or playable video is not a relevance score.
Native manual runs separately demonstrate editing, cancellation/reopening and actual playable exports. The first runtime-validated Agent target is macOS Apple Silicon. Other Agent platforms, a continuous real-model demonstration and heldout semantic comparisons are not established by this preview.
The corrected source at 5e5422909ae2cd557fe214796e1e8a1afc3dd3e8 passed the Linux/macOS/Windows CI checks, including all 29 real-media tests on Linux and the three corrected crash/recovery cases. Four-target builds also passed; the license gate correctly left zero binary artifacts. The final r2 follow-up adds documentation only, with the same tested Rust/script source. These checks are not semantic-quality evaluation or cross-platform desktop-runtime acceptance.
Build locally
Install Rust 1.87+, FFmpeg and ffprobe; ffplay enables interval playback. From this source checkout:
cargo build --locked --release
./target/release/video-toolkit agentUse Manual workflow — no AI without credentials. Real model calls require a privately configured local key and explicit consent to share sampled JPEGs and bounded task text; do not put keys or runtime task directories in Git.
Distribution status
This is a source-only prerelease, not an installer release. Six unresolved dependency license-text questions still block public binaries. Packaging fails closed, with no public package/checksum uploads while target-specific distribution approval is missing. FFmpeg is not bundled. Existing v2.1.0 downloads are the legacy Toolkit, not this Agent.
The preview's public commit contains the final source tree on the existing public history. Operator-declared intermediate development snapshot hashes in evidence records remain local and are not public checkout links or signed build attestations.
Setup and boundaries · Machine-readable evidence · Evaluation status
Post-release development evidence update — 2026-10-18
The user explicitly instructed the development agent to delete the first clip and export the remaining four. The native UI saved revision 2, approved its exact digest and produced the actual 20-second demonstration MP4. An independent check confirmed unchanged source hashes, the four original remaining intervals, approval/output binding and full video/audio decode; it also played in VLC. The 20.021333-second container includes AAC padding. Attribution and exact modifications.
The demonstration/evidence update is on current master; the r2 tag is not moved. The actual GUI runtime was the declared local 42f1909 core snapshot, not a retroactive run of the later final r2 tree. The initial 25-second model draft and failed proposals remain preserved. This is one user-directed development workflow, not a blind semantic score, heldout result, continuous screen recording or new public installer release.
v2.1.0
Download the archive matching your operating system and CPU, plus SHA256SUMS.
Linux and macOS packages use .tar.gz; Windows packages use .zip.
Each archive contains the executable, version/commit metadata, and INSTALL.txt.
Rust is not required. Install FFmpeg and ffprobe separately and make both available
in PATH. Linux also requires GTK3 and a desktop session for the GUI.
macOS binaries are not Developer ID signed or notarized; launch from a terminal
after reviewing the installation notes. Do not disable system-wide security checks.
All four archives were verified and their extracted executables passed CLI/profile
smoke tests on their native runners. Linux also passed real FFmpeg processing tests.
What's Changed
- Add plugin system and batch processing features by @Alaneel in #1
- v2 by @Alaneel in #2
- Support webm to gif and multiple other formats by @Alaneel in #3
- Support webm to gif and multiple other formats by @Alaneel in #4
- Polish Video ToolKit for v0.2.0 by @Alaneel in #5
- Standardize Video Toolkit naming by @Alaneel in #8
- chore: restore SemVer for v2.1.0 by @Alaneel in #9
- ci: upgrade checkout action to v7 by @Alaneel in #10
- Fix output formats, audio selection, and batch reliability by @Alaneel in #11
- Run validated saved profiles in CLI and desktop batch by @Alaneel in #12
- Ship verified native release packages for four platforms by @Alaneel in #13
New Contributors
Full Changelog: v2.0.0...v2.1.0
