Independent convergence: we run a 40-agent human+AI fleet on a signed, file-based bus — interop questions

Hey Buzz team 👋

For the past months I've been running a personal/SMB "fleet": ~40 registered agents (Claude Cowork/Code, Codex-family, Cursor) working alongside me and two human collaborators across 4 machines (2 Macs, a Windows/WSL2 box, an Ubuntu server). Reading your README was uncanny — we independently converged on most of your core bets, just in duct-tape form:

- **Agents as members, not bots** — every agent has a registry card (YAML), a callsign, and its own ed25519 keypair
- **One shared log** — a file-based message bus (`~/.claude-agents`) synced via Syncthing, with a git committer as the signed audit archive
- **Scoped by identity, not permission flags** — agents run scoped to areas of a shared "brain" monorepo, each area with its own steward agent
- **Externalized memory** — agents write ledgers/state to the repo so sessions are disposable
- **Human approval gates** — irreversible ops (deploys, deletes, purchases) always stop for a signed GO

Where you're ahead of us: full-text search over everything, branch-as-room, one desktop UI over the whole thing. Where our setup differs: we're **offline-first** — devices drop off (laptop on a train) and reconcile on reconnect; no always-on relay required.

Questions from someone who wants to pilot Buzz as the UI/search layer over an existing fleet:

1. **Bridge pattern** — for mirroring an external bus into a relay: one bridge identity, or one nsec per agent posting via `buzz-cli`? Any idempotency/dedup guidance for mirrored events?
2. **Offline story** — any plans for a client-side outbox in `buzz-cli` (write while relay unreachable, flush+reconcile later)? This is our only structural blocker.
3. **Persona packs** — is the `buzz-persona` schema stable enough to target? Our agent identity cards are YAML; compiling them into packs looks feasible.
4. **ACP + Claude Code** — daily-driver ready, or still experimental?

Happy to share more detail on our setup if useful — feels like we're building the same thing from opposite ends. 🐝

No files, tests, or concrete entry points are named. Start with the README and trace the relay and buzz-cli behavior related to bridge identities, deduplication, offline outboxes, persona packs, and ACP; done would require maintainers to define one scoped interoperability change, since the issue currently presents questions rather than an implementation target.

Written by the indexing model from the issue text.