v0.3.8

What's New

Free Association — LLM-powered step-by-step sequencer

A new sequencer where you pick the first clip and an LLM proposes each next clip based on its metadata, providing a rationale for every transition.

• Two-column dialog: stacked interaction pages on the left, a running rationale log on the right
• Accept / Reject / Stop flow — reject to get a different proposal for the same position, stop to end early with the clips accepted so far
• Metadata-grounded rationales reference actual clip fields (shot type, colors, objects, motion, transcript, etc.) rather than abstract poetry
• Tiered prompt strategy keeps per-step cost bounded (~800 tokens) regardless of total clip count — local cosine-similarity shortlisting absorbs scaling
• Per-clip rationales persist on the sequence through project save/reload

Generate Embeddings — user-triggerable analysis operation

DINOv2 visual embeddings are now a first-class analysis checkbox in the Analyze tab alongside Extract Colors, Classify Shots, and others.

• Populates clip.embedding (768-dim DINOv2 vectors) directly, rather than only as a side effect of running similarity-based sequencers
• Used by Free Association, Human Centipede, Staccato, and Reference Guide
• Shows in the cost gate with time estimates like any other operation
• Dialog disables the checkbox when every clip in scope already has embeddings

Agent-native improvements

All 10 recommendations from the agent-native audit shipped, bringing the chat agent closer to parity with user capabilities across tabs, clip operations, and sequencing.

Fixes

• UI: Removed redundant analysis category labels; added cloud icons to cloud-capable operations; quit→restart no longer leaks state
• LLM: Caught ValueError from tiktoken when registry data is missing; patched litellm encoding fallback for frozen (PyInstaller) builds
• Linux/AppImage CI: Resolved version in recipe before build, pinned packaging<22 for appimage-builder compatibility, imported rotated Ubuntu GPG key, added GPG key_url to AppImageBuilder apt sources

Upgrading

The Free Association sequencer and Generate Embeddings operation use the same model settings as Storyteller and Exquisite Corpus (now labeled LLM Sequencer — Storyteller / Exquisite Corpus / Free Association in Preferences). Embeddings extraction downloads ~450 MB of DINOv2 weights on first run.

v0.3.8-rc2

Changes since v0.3.8-rc1

• fix(ui): Remove analysis category labels, add cloud icons, fix quit restart
• fix(llm): Catch ValueError from tiktoken with missing registry data
• fix(llm): Patch litellm encoding fallback for frozen builds

v0.3.8-rc1

Merge pull request #86 from dvschultz/feat/semantic-clip-search

feat(search): add semantic clip search with enhanced filtering and similarity

v0.3.7

What's New

Category Pill Bar for Sequence Tab

• 6 category filters (All, Arrange, Find, Connect, Audio, Text) above the algorithm card grid
• Algorithms can appear in multiple categories via tags
• Grid dynamically reflows when switching categories
• Selected category persists across sessions

Semantic Clip Search & Enhanced Filtering

• 4 new ClipBrowser filters: gaze direction, object detection, brightness range, description search
• "Find Similar" visual similarity search using DINOv2 embeddings
• Expanded agent filter_clips tool with 16 new parameters (gaze, brightness, volume, OCR, tags, notes, cinematography, similarity)
• Filter controls disable with tooltips when analysis data is missing

v0.3.6-rc2

feat(ui): add bug report dialog to Help menu

New "Report Bug..." item in Help menu opens a dialog with:
- Text area for describing the issue
- Auto-collected system info (app version, OS, Python version)
- Three submission options:
  1. GitHub: opens pre-filled issue with system info + last 50 log lines
  2. Gmail: opens compose window, opens Finder to log file for attachment
  3. Email: opens mailto link, opens Finder to log file for attachment

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

v0.3.6-rc1

fix(ci): exclude on-demand ML packages from PyInstaller collection

torch bundled unsigned binaries (protoc, torch_shm_manager) into the
app, causing Apple notarization to reject the archive with "critical
validation errors". Exclude all on-demand ML packages from PyInstaller
since they're installed at runtime, not bundled.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

v0.3.3-rc7

fix(macos): retry ffmpeg runtime downloads

v0.3.3-rc6

fix(macos): retry ffmpeg runtime downloads

v0.3.3-rc5

fix(macos): retry ffmpeg runtime downloads

v0.3.3-rc4

fix(macos): retry ffmpeg runtime downloads