v3.5.0: Foundation Complete

Foundation Complete

The culmination of v3.x development — all core infrastructure is in place for the editorial workflow.

Highlights

This release represents the complete v3.5 foundation: a production-ready system for processing PBS Wisconsin video transcripts into SEO-optimized metadata. The system combines automated LLM processing with a human-in-the-loop editing workflow powered by Cardigan, our Mister Rogers-inspired AI copy editor.

Everything in v3.5.0

• 4-phase processing pipeline (Analyst → Formatter → SEO → Manager) with multi-model routing
• React web dashboard with 8 pages, real-time WebSocket updates, WCAG 2.1 AA accessibility
• Remote ingest watcher — automated transcript/screengrab discovery from PBS media server
• Embedded chat — talk to Cardigan from the job detail view
• MCP server for Claude Desktop (9 tools, 6 prompts)
• Per-phase retry tracking with UI visibility
• GitHub Actions CI pipeline with code formatting enforcement

Infrastructure

• 9 API routers, 15 services
• SQLite database with 7 migrations
• APScheduler for ingest scheduling
• OpenRouter multi-model LLM routing
• Airtable SST integration (read-only)

What's next

See the roadmap for v3.6.0 (Stability & Identity) through v4.0.0 (Deployment).

v3.3.0: UX Refinement

UX Refinement

Sprint 13.x: Workflow improvements that make transcript discovery and screengrab management less overwhelming.

What shipped

• Promoted "Ready for Work" — moved from Settings sub-tab to second item in main navigation
• Smart filtering — debounced search, configurable date range, auto-hide files with existing jobs
• Contextual screengrab panel — screengrabs appear on JobDetail only when matching Media ID exists (replaced global panel)
• Simplified Settings — removed redundant ingest panels, added direct links to relevant pages
• Bundle optimization — reduced frontend bundle by 13KB through refactoring

Technical details

• Extended /api/ingest/available with search, days, and exclude_with_jobs parameters
• Added /api/ingest/screengrabs/for-media-id/{media_id} endpoint
• New components: ReadyForWork page, ScreengrabSlideout
• Frontend build: 472KB (down from 485KB)
• Ingest tests: 52/52 passing

v3.2.0: Chat Prototype

Chat Prototype

Sprint 12.1: Embedded chat with Cardigan directly from the job detail view.

What shipped

• Chat endpoint — POST /api/chat/message with full project context injection
• Slide-out chat panel — accessible from JobDetail page, chat while viewing job results
• Context builder — automatically loads transcript, brainstorming, and revision history into chat context
• Editor personality — Cardigan's Mister Rogers-inspired voice in chat responses
• Cost tracking — per-message token and cost tracking via chat_cost service
• DB persistence — chat history stored in SQLite for session continuity

Technical details

• New router: chat_prototype
• New services: chat_context, chat_cost
• REST-based (intentional — validates UX pattern before WebSocket investment)
• Context size: ~60K chars (transcript truncated at 10K)

v3.1.0: Remote Ingest Watcher

Remote Ingest Watcher

Sprint 11.1: Automated discovery of transcripts and screengrabs from PBS Wisconsin's media ingest server.

What shipped

• Ingest scanner service — scheduled scanning of remote media server for SRT transcripts and JPG screengrabs
• Smart matching — correlates files to Airtable SST records that have passed QC
• "Ready for Work" page — top-level dashboard page for transcript discovery with search and date filtering
• Screengrab auto-attachment — automatically links screengrabs to Airtable records (additive-only, with audit logging)
• Ingest configuration UI — Settings page controls for scan schedule and server URL
• APScheduler integration — configurable scan scheduling (default: daily)

Technical details

• New router: ingest (8 endpoints for scan, config, available files, screengrab management)
• New services: ingest_scanner, ingest_config, ingest_scheduler, screengrab_attacher
• 61 new tests (40 unit + 12 API + 9 attacher safety tests)
• Database migrations for ingest tables and available_files tracking

v3.0.0: Core Architecture

Core Architecture

The foundational rewrite from v2.0's CLI-based system to a modern API-first architecture.

What shipped

• FastAPI backend with SQLite database as single source of truth
• React web dashboard with real-time WebSocket job monitoring
• 4-phase processing pipeline (Analyst → Formatter → SEO → Manager)
• Multi-model LLM routing via OpenRouter with cost optimization
• Bulk transcript upload with drag-and-drop
• WebSocket live updates for job status changes
• WCAG 2.1 Level AA accessibility features (keyboard nav, screen reader support, reduced motion, high contrast)
• MCP server for Claude Desktop integration ("Cardigan" copy-editor persona)
• Airtable SST integration (read-only) for metadata lookup

Technical details

• 5 API routers: queue, jobs, config, upload, websocket
• 8 services: database, llm, worker, airtable, logging, utils, completeness
• 8 dashboard pages: Home, Queue, JobDetail, Projects, Settings, System, Help
• Model routing: Gemini 2.0 Flash (analysis/SEO), GPT-4o-mini (formatting)