Video Evidence Workbench v0.3.0

Video Evidence Workbench v0.3.0

Analyze videos through a globally installed Codex Skill, resume long breakdowns
in small batches, and preserve the evidence needed to review the result.

Global Codex installation

Download video-evidence-workbench-skill-0.3.0.zip and
video_analysis_mvp-0.3.0-py3-none-any.whl from this release, extract the Skill
archive, then run with Python 3.11+:

python3 -I video-evidence-workbench/scripts/install.py \
  --wheel /absolute/video_analysis_mvp-0.3.0-py3-none-any.whl \
  --extras api --extras export --extras pdf

The installer places the Skill in $CODEX_HOME/skills (default
~/.codex/skills) and binds an isolated runtime under
~/.local/share/video-evidence-workbench/runtimes. Existing versions are backed
up outside Skill discovery. It keeps the caller's working directory and does
not change global PATH or another project's virtualenv. Pip may download the
declared package dependencies during this explicit installation; the wheel
itself must be supplied locally. ffmpeg/ffprobe remain prerequisites. PDF
rendering additionally needs the documented preinstalled Node/browser/font
runtime; the installer does not download those tools or ASR models.

Before installing dependencies, the installer checks its managed pip and upgrades
older versions to the tested 26.2.1 baseline. This addresses the vulnerable pip
bundled by some Python installers; system pip is not changed. See the
pip advisory.

On the next Codex turn, invoke:

$video-evidence-workbench Analyze this video and produce a complete, traceable breakdown.

Workflow changes

• codex next returns a bounded batch, ordered supporting frames and neighboring
  shots; codex submit checkpoints results and commits only the complete scope.
• Repeated prepare preserves valid applied work. Intentional reanalysis uses
  prepare --refresh; a failed final commit can be retried with submit --finish.
• New requests support up to 1024 eligible shots, with at most 32 per batch.
  V1 requests and receipts remain compatible; direct apply retains its 1 MiB cap.
• Video-only sources remain video-only. The workflow records no audio stream,
  leaves WAV absent and does not infer measured silence.
• Full-shot creative Markdown retains specific evidence while consolidating
  repeated model templates and shared constraints.

Correctness and review

Single detected cuts are retained; short CFR extraction includes tested 25 fps
single-frame/two-frame and 100 fps cases. Caught commit errors restore the prior
shot/receipt state. Protected human rows and later human corrections coexist
with unchanged model bindings.

See the optimization review
and the attached release-candidate receipt for exact local verification. GitHub
Actions on the release's commit/tag records remote platform checks separately.
The release API and release verification determine the platform's actual
immutability and asset provenance state.

This remains pre-1.0. External provider accuracy, general ASR/semantic accuracy,
exact VFR PTS, Windows and automatic SIGKILL/power-loss recovery are not claimed.
Model submissions remain proposals; no automated check asserts human review.

Video Evidence Workbench v0.2.2

Video Evidence Workbench v0.2.2

v0.2.2 is a release-integrity correction to v0.2.1. It does not change video
analysis, review, Finalize, migration, Codex handoff or client-export behavior.

Integrity correction

• Candidate CI binds the exact named tagged Git tree, product digest, current
  release status and four separately hashed review-evidence files.
• v0.2.1 was created while GitHub release immutability was disabled. After the
  policy was enabled and its note was updated, the API reported
  immutable: true and release verification succeeded. This timeline is
  reported without inferring why observed platform behavior differs from the
  general GitHub documentation.
• Repository release immutability was enabled before v0.2.2 was created. The
  GitHub release API is authoritative for whether the published release, tag and
  attached assets are platform-locked.

Verification boundary

The release remains pre-1.0. Structural and integrity checks do not certify
shot interpretation, copyright ownership, live OpenAI/MiniMax/BridgeDeck
behavior, Windows, ASR accuracy, semantic VO/music/SFX identity, adoption or
stars.

Video Evidence Workbench v0.2.1

Integrity timeline (2026-09-04): v0.2.1 was created while repository release immutability was disabled. After the policy was enabled and this note was updated, the GitHub API currently reports immutable:true and gh release verify succeeds for the tag and attached receipt. No destructive mutation probe was performed.

Video Evidence Workbench v0.2.1

v0.2.1 is a maintenance release for the public Video Evidence Workbench. It
contains the post-v0.2.0 release-governance fixes and the separately verified
GitHub Actions v7 upgrades; it does not broaden the product's factual accuracy
or live-provider claims.

What changed

• Ordinary pull requests keep deterministic UI, screenshot, source and packaged
  asset checks without being compared with the historical v0.2.0 product digest.
• v* tag builds verify the named tagged Git tree, release status, mature
  candidate receipt and four separately bound review-evidence files.
• npm audits still fail on confirmed high/critical findings, malformed reports
  and unknown process errors. A verified advisory-endpoint outage is surfaced as
  unavailable, never as a clean audit or a vulnerability finding.
• checkout, setup-python, setup-node and upload-artifact use their reviewed v7
  releases on GitHub-hosted runners.
• Dependabot groups only frontend minor/patch updates. Frontend major versions
  must be proposed and visually validated independently.

Compatibility and migration

No workspace schema migration is required from v0.2.0. Existing projects remain
subject to the same readiness, human-review, Finalize and explicit-export gates.
The GitHub Actions v7 upgrades require runner 2.327.1 or newer; the public CI
uses GitHub-hosted runners and exercises Linux and macOS on Python 3.11/3.12.

Verification boundary

The exact tag CI and attached mature-candidate receipt are authoritative for this
release. A successful structural gate does not certify shot interpretation,
copyright ownership, live OpenAI/MiniMax/BridgeDeck behavior, Windows support,
ASR accuracy or semantic VO/music/SFX identity accuracy.

Video Evidence Workbench v0.2.0

Video Evidence Workbench v0.2.0

v0.2.0 turns the original local video-analysis prototype into a reviewable,
evidence-first workbench for short videos, brand films and media research.

What users can do

• analyze a local video into timecoded shots, frames, audio and rhythm evidence;
• review every shot and unresolved VO/music/SFX event in the browser;
• explicitly Finalize a digest-bound evidence package;
• generate professional XLSX and PDF client files only when requested;
• use the same evidence contract from Codex, CLI or the local API;
• inspect provenance, readiness, migration and artifact-lifecycle receipts.

Upgrade notes

Existing pre-readiness-v3 projects should first be inspected without writes:

analyze-video --workspace ./analysis-projects migrate PROJECT_ID

Use --apply only after reviewing the migration report. Migration invalidates
legacy publication state and requires the normal explicit Finalize step; it does
not rerun media analysis or generate client files.

Verification

The pre-publication candidate passed 519 Python tests with 28 optional-runtime
skips and 393 subtests. The explicit render matrix passed 71/71 tests, including
a real XLSX/PDF transaction and LibreOffice workbook opening. Frontend build,
same-origin integration, responsive E2E, clean wheel installation, migration,
synthetic benchmarks, Ruff, Bandit and dependency audits also passed locally.

GitHub Actions checks on the tagged revision are the remote verification source.
Local evidence and the independent review are preserved in
docs/evidence/mature-candidate-receipt.json
and docs/cold-review.md.

Boundaries

• Readiness means structural evidence completeness, not factual certification.
• Live OpenAI, MiniMax and BridgeDeck behavior is optional and not proven by the
  deterministic test suite.
• Windows is not currently supported.
• URL ingest is for a trusted local operator and is not an SSRF sandbox.
• yt-dlp>=2026.7.4 is required to exclude
  CVE-2026-55404 / GHSA-6v4j-43gg-vj32.
• No GitHub-star, adoption or accuracy outcome is guaranteed.

Remote release verification

• Merge commit: 672e67c7c9dd3f610091007eada529df5873044a
• Main CI: 8/8 jobs passed on attempt 2.
• The only first-attempt main failure was an investigated npm Advisory API 503 Service Unavailable; the exact same merged tree passed on the failed-job retry.
• PR review and history: #1

This is a pre-1.0 GitHub pre-release. It is not a Windows, hosted-service, live-provider, or adoption guarantee.