Feature Request: Multi-Device PAI Sync (`pai-sync` skill)

[garzito]

Problem

I use Claude Code on multiple machines (laptop, desktop, VPS). Each has its own ~/.claude/ with independent USER/, MEMORY/, skills, and learnings. So my DA knows me on one machine but starts fresh on the next. I end up re-teaching context and losing continuity between devices.

I imagine this is a common pain point as PAI adoption grows, since a lot of us work across multiple computers.

Why PAI is already well-suited for this

The architecture makes sync surprisingly doable:

• Plain text files: Git-friendly
• USER/ vs SYSTEM/ split: portable identity is already a design goal
• CLI-first: sync commands fit the existing patterns
• Append-only history: JSONL logs merge trivially

The main missing piece is a config split to separate machine-specific from universal concerns, plus tooling to automate the sync.

Proposed Solution

1. Config split

Introduce local override files that are gitignored, merged at SessionStart:

• settings.local.json → machine-specific overrides (merged on top of settings.json)
• .mcp.local.json → machine-specific MCP servers (merged on top of .mcp.shared.json)

This is useful independently of sync since it lets users maintain machine-specific config without touching shared files.

2. pai-sync skill

A new skill using Git as the default backend:

skills/Sync/
├── SKILL.md
├── workflows/
│   ├── setup.md        # Interactive setup wizard
│   ├── push.md         # Commit + push
│   ├── pull.md         # Pull + merge
│   └── status.md       # Show sync state
└── tools/
    ├── sync-init.ts     # Create private repo, configure remote, git-crypt
    ├── sync-push.ts     # Smart commit + push
    ├── sync-pull.ts     # Pull + auto-resolve
    ├── sync-status.ts   # Show divergence across devices
    └── sync-resolve.ts  # AI-assisted conflict resolution

What syncs vs. what stays local:

Synced	Local (gitignored)
USER/, MEMORY/	.env (encrypted via git-crypt)
skills/, agents/, hooks/	.mcp.local.json
history/sessions/, learnings/, decisions/	history/raw-outputs/
settings.json	settings.local.json
	voice-server/, Observability/

3. SessionStart/SessionEnd hooks

Auto-pull on session start, auto-push on session end. Silent failure, never blocks a session.

4. Conflict handling

Three layers:

• Prevention: auto-sync hooks keep the window for conflicts near zero
• Auto-resolution: custom Git merge drivers for append-only files (JSONL: concatenate + deduplicate by timestamp)
• AI-assisted resolution: when real conflicts occur (e.g., two edits to TELOS), the DA reads both versions, shows the diff in plain language, proposes a semantic merge.

Implementation

I’d split this into incremental PRs:

1. PR 1: Config split: settings.local.json + .mcp.local.json merge logic + .gitignore template
2. PR 2: Sync skill: core sync-init, sync-push, sync-pull tools
3. PR 3: Hook automation: SessionStart/SessionEnd auto-sync

Happy to build this. Wanted to get feedback on the approach first — especially the config split, since that touches load-core-context.ts.

[toroleapinc]

This is a great feature request! Multi-device sync for personal AI infrastructure is something I've been working on too.

I built claude-brain which handles exactly this for Claude Code — syncs your entire brain state (memory, skills, agents, rules, CLAUDE.md, settings) across machines via Git with intelligent semantic merging. It auto-syncs via hooks on every session start/end, zero daily effort.

Could be interesting to look at for the PAI sync design — the semantic merge approach (using the LLM itself to deduplicate and reconcile memory from different machines) has worked really well in practice.

[mvanders]

I stumbled upon Happy Engineering (https://happy.engineering/), but does not seems to get updated... :-(
Besides single used, multi-device, there is also multi-user... I just watched Daniel's new video on lattice. I wonder how he will get from PAI (single user) to LATTICE (multi-user)?