SpeechEdit Lands With a Swarm of Features

A Greenfield Day

I started speechedit (cef5304e) and then landed most of its feature surface in one session. SpeechEdit is an editor for talking-head video built around the transcript: you edit words, and the video follows.

The Feature Swarm

The commits read like a feature list: breaths, lead/trail, and repeat detection in analysis (0f18e3c3); an FFmpeg 4.4 export graph with volume monitoring (b553c28c); scenes[] and keepTake so takes can be chosen rather than just cut (3caa18df); pluggable ASR providers with a dedicated transcribe path (1bf12216); interchange, captions/titles, and a diarization hook (dc75b90e); a word-snap timeline with ripple edits and music beds (a6417cee); auto-edit review groups and a demo-guard restore (0ce9a350); play-from-word, P/A/C modes, revertAt, and put-back chips (b23873c1); and an MCP/agent loop on the project command bus (5c76deed).

I merged each stream back as it landed—analysis engine, export preview, transcribe providers, word snap, scene intelligence, auto-edit, intelligence, and captions/interchange/diarization. Working as parallel streams and merging kept the day from becoming one giant unverifiable change.

Correctness Under the Hood

I fixed bugs that only matter if the editor is going to be trusted: keeping island-word cuts from wiping the reel and making file I/O real (ec61aa5a), making preview play the same export job FFmpeg renders (b8d10276), letting users cancel export mid-FFmpeg and keeping import from getting stuck (cc2f4431), and keeping B-roll and music picks inside the one-job-at-a-time gate (1cbc721e). Preview matching export is the single most important correctness property of a video editor—if the preview lies, nothing else matters.

I added local whisper.cpp integration (6b53c26b) and proxy handling for media assets (012a2e31).

Meanwhile on the Client Work

I enhanced venue management with workspace features and avatar support (80eb7ac2), hardened mail sanitization for oversized stylesheets (0efa7b3c), improved event ownership and calendar UX (57e9d595), and streamlined event application logic (1a708910). I also fixed the Lake Forest officer panel by serving Livewire JavaScript (2c02fee4).

The Command Bus

The through-line of SpeechEdit is a project command bus. Every feature—scenes, edits, agent commands—goes through one place. An agent and a human operate on the same commands, which is what makes the MCP integration coherent rather than bolted on.