MotionMix Convergence Directive
MotionMix is no longer at the "invent the architecture" stage. The major runtime pieces already exist: the iPhone/iPad capture app, `multicam-server`, the Live Director macOS app, the Convex production ledger, the still-upload path, and the low-latency hero-feed/WebRTC seam. The current job is convergence. That means turning multiple partially completed coding sessions into one coordinated push with strict lane ownership, shared situational awareness, and no duplicate implementation. The product goal for this conve
Full Public Reader
MotionMix Convergence Directive
What Is Happening
MotionMix is no longer at the "invent the architecture" stage. The major runtime pieces already exist: the iPhone/iPad capture app, `multicam-server`, the Live Director macOS app, the Convex production ledger, the still-upload path, and the low-latency hero-feed/WebRTC seam. The current job is convergence. That means turning multiple partially completed coding sessions into one coordinated push with strict lane ownership, shared situational awareness, and no duplicate implementation. The product goal for this convergence pass is not "train the whole music brain today." The goal is to make the AI photoshoot lane trustworthy, operable, and logged well enough that it can be used as a real creative workflow.
Product Priority Order
1. Stabilize the AI photoshoot runtime.
2. Add operator supervision for Ghost mode.
3. Complete Convex production-session/event writes from the live photoshoot path.
4. Start the offline training pipeline as a separate branch, without touching the live runtime.
What Already Exists And Must Not Be Rebuilt
- Photoshoot Mode shell exists in Live Director.
- Coverage map exists in Live Director.
- Convex production ledger exists.
- WebRTC hero-feed signaling/publisher/consumer seam exists.
- Per-device flip controls already exist.
- Smarter iPhone lens switching has already been patched.
- Real still-upload path exists.
- Optional Mac webcam feeds are already hidden by default in the photoshoot workflow.
Shared Ground Truth
Runtime/control surfaces:
- `MotionMixApp` on iPhone/iPad:
- `[home]/Desktop/MotionMixApp/MotionMixApp`
- `multicam-server`:
- `[home]/Desktop/MotionMix/multicam-server/src/main.rs`
- `MotionMixLiveDirector`:
- `[home]/Desktop/MotionMixLiveDirector`
- Convex production backend:
- `[home]/Desktop/Comp-Core/apps/convex-memory/convex`
Supporting docs:
- `[home]/Desktop/MotionMix/AI-PHOTOSHOOT-TODAY-RUNBOOK.md`
- `[home]/Desktop/MotionMix/AI-PHOTOSHOOT-MOTION-MUSIC-PLAN.md`
- `[home]/Desktop/MotionMix/CONVEX-PRODUCTION-IMPLEMENTATION-HANDOFF.md`
- `[home]/Desktop/MotionMix/CANONICAL-STATE-CONTRACT.md`
- `[home]/Desktop/MotionMix/HANDOFF-AUDIO-PREP.md`
Agent Map
Codex / Orchestrator
Owns runtime convergence, placement-metadata correctness, rebuild/redeploy/relaunch, and final integration truth.
See:
- `[home]/Desktop/MotionMix/agent-handoffs/CODEX-ORCHESTRATOR-LANE.md`
Claude Agent 1
Owns Ghost-mode operator supervision in Live Director.
See:
- `[home]/Desktop/MotionMix/agent-handoffs/CLAUDE-AGENT-1-GHOST-UI.md`
Claude Agent 2
Owns Convex production-session/event writes for the AI photoshoot lane.
See:
- `[home]/Desktop/MotionMix/agent-handoffs/CLAUDE-AGENT-2-PRODUCTION-LOGGING.md`
OpenCode
Owns the offline training/pairing branch only, starting with dataset alignment and builder work.
See:
- `[home]/Desktop/MotionMix/agent-handoffs/OPENCODE-TRAINING-PAIRING.md`
Anti-Overlap Rules
These are global and apply to every agent:
- Do not rebuild Photoshoot Mode shell.
- Do not rebuild Coverage Map.
- Do not redesign WebRTC hero-feed transport.
- Do not redesign the Convex production schema from scratch.
- Do not duplicate lens-control or per-device flip work unless fixing a regression inside your exact lane.
- Do not let offline training work reach into live runtime code.
Convergence Milestone
The convergence pass is complete only when all of this is true:
1. Live Director shows a real Ghost proposal banner with accept/veto flow.
2. Photoshoot session start/stop and live events write into Convex production tables.
3. The offline training branch can produce a reproducible paired dataset from real local audio/pose data.
4. The runtime can be rebuilt and redeployed without regressing:
- per-device flip
- iPhone `0.5x`
- still capture path
- live mobile device registry
- coverage/placement visibility
Required Reporting Format
Every agent must end with an RTD Verification block covering:
1. Structure
2. Compilation
3. Integration
4. Content
5. User Journey
6. Deployment
Every agent must also state:
- what they changed
- what they verified
- what they intentionally did not touch
- residual risk
Promotion Decision
Attach run IDs, datasets, metrics, and reproduction commands.
Source Anchor
MotionMix/agent-handoffs/00-CONVERGENCE-DIRECTIVE.md
Detected Structure
Method · Evaluation · Code Anchors · Architecture