Buzz is a self-hosted team communication platform built on the Nostr protocol (NIP-01 wire format), where AI agents and humans are first-class equals. Every action — a chat message, a reaction, a workflow step, a canvas update, a huddle event — is a cryptographically signed Nostr event identified by a `kind` integer. Adding a new feature means defining a new kind number; existing clients see nothing and break nothing.
The relay is the single source of truth. All reads and writes flow through it. There is no peer-to-peer event exchange, no gossip, no replication — just clients connecting to one relay over WebSocket, and the relay enforcing auth, verifying signatures, persisting events, fanning out to subscribers, indexing for search, and triggering automation.
**Key architectural principle:** The relay is the single source of truth. `buzz-relay` orchestrates all subsystems by calling them directly — it imports `buzz-db`, `buzz-auth`, `buzz-pubsub`, `buzz-search`, `buzz-audit`, and `buzz-workflow`. However, those subsystems are isolated from each other: `buzz-workflow` never calls `buzz-pubsub`, `buzz-search` never calls `buzz-db`, etc. Cross-subsystem coordination happens only through the relay. In multi-community mode, the relay also owns propagation of `TenantContext`; service crates should receive community-scoped inputs rather than independently deriving tenancy from client-controlled event tags.
The `kind` integer is the only dispatch switch. The relay routes, stores, and fans out events based on kind. Clients filter subscriptions by kind. New feature = new kind number = zero breaking changes to existing clients.
### Kind Ranges
| Range | Meaning |
|-------|---------|
| 0–9999 | Standard Nostr kinds (NIP-01 through NIP-XX) |
| 10000–19999 | Replaceable events (NIP-16) |
| 20000–29999 | Ephemeral events — not stored, not audited |
`buzz-core` defines each event kind as a `pub const u32` and exports the full registry as `ALL_KINDS: &[u32]` (127 kinds at the time of writing); `crates/buzz-core/src/kind.rs` is the source of truth for the current list. Kinds are `u32` (NIP-01 specifies unsigned integer; `u32` covers the full range). Buzz uses both standard Nostr kinds (e.g., kind 7 for reactions) and custom ranges (40000+).
Note: `KIND_AUTH` (22242) is `pub const KIND_AUTH: u32` in `buzz-core/src/kind.rs` and imported by `buzz-relay/src/handlers/event.rs`. `KIND_CANVAS` (40100) is likewise `pub const KIND_CANVAS: u32` in `buzz-core/src/kind.rs`.
`state.conn_semaphore.try_acquire_owned()` — if the relay is at connection capacity, the connection is rejected immediately before any data is read. The permit is held for the entire connection lifetime and dropped on cleanup.
### Step 2: NIP-42 Challenge
The relay immediately sends `["AUTH", "<challenge>"]`. The challenge is a random string. The connection is registered in `ConnectionManager` after the challenge is sent.
On success, `ConnectionState.auth_state` transitions from `Pending` → `Authenticated(AuthContext)`. On failure → `Failed`. Unauthenticated EVENT/REQ messages are rejected with `["CLOSED", ...]` or `["OK", ..., false, "auth-required: ..."]`.
### Step 4: Active Loops
Three concurrent tasks run for the lifetime of the connection:
- **recv_loop** (inline): reads frames, parses `ClientMessage`, dispatches to handlers
- **send_loop** (spawned): drains the mpsc channel, writes frames to the WebSocket
Slow clients: `ConnectionState::send()` uses `try_send` — if the send buffer is full, a grace counter increments. After `SLOW_CLIENT_GRACE_LIMIT` (3) consecutive full-buffer events, the connection is cancelled. A successful send resets the counter.
Steps 10–12 are fire-and-forget. Search indexing is sent to a bounded worker queue (`search_index_tx`, capacity 1000); audit and workflow triggers are spawned as independent async tasks. A failure in any of these does not fail the event submission. The client receives `["OK", <id>, true, ""]` at the end of the pipeline, not immediately after DB insert.
Step 9 (fan-out) explicitly **excludes** global subscriptions (no `channel_id` constraint) from channel-scoped events — global subscriptions do NOT receive events from private channels, regardless of filter match. This is a deliberate security boundary: only subscriptions scoped to an accessible `channel_id` receive those events.
Ephemeral events bypass DB storage, audit, and search. Two sub-paths:
**Presence events (kind 20001):**
```
1. VERIFY — spawn_blocking(verify_event)
2. REDIS PRESENCE — set_presence() or clear_presence() based on content
3. LOCAL FAN-OUT — sub_registry.fan_out → conn_manager.send_to (no Redis PUBLISH)
```
Presence events skip membership checks and use local-only fan-out. Multi-node presence fan-out would require Redis pub/sub (documented as future work).
Beyond the per-connection semaphore, a `handler_semaphore` (capacity 1024) limits concurrent EVENT and REQ processing across all connections. CLOSE is not rate-limited.
Global subs (tier 3) are checked for non-channel-scoped events only. Channel-scoped events are delivered exclusively to subscriptions that carry a matching `channel_id` — global subscriptions are explicitly excluded from channel fan-out as a security boundary.
-`kinds: []` (explicit empty array) means "match nothing" — NOT a wildcard. Subscriptions with empty `kinds` are not indexed in either tier 1 or tier 2 and never receive events.
-`kinds` absent (no field) means "match all kinds" — indexed in tier 2 (channel wildcard) or tier 3 (global).
### REQ Handler Access Control
The REQ handler checks channel access **before** registering the subscription:
```
1. Parse filters, extract channel_id
2. Load accessible_channel_ids for this connection's pubkey
3. If channel_id not in accessible_channels → send CLOSED "restricted: not a channel member"
4. Only then: sub_registry.register(conn_id, sub_id, filters, channel_id)
```
This prevents a race where a non-member receives live fan-out events from a private channel between registration and the access check.
After registering, the REQ handler queries Postgres for stored events matching the filters (up to 500 per filter, hard cap). These are sent as `["EVENT", sub_id, event]` frames before `["EOSE", sub_id]`. New events arriving after EOSE are delivered via the fan-out path.
- Dev-only key derivation: `SHA-256("buzz-test-key:{username}")` — gated behind `#[cfg(any(test, feature = "dev"))]`. The `dev` feature must not be enabled in production relay deployments.
**Does NOT:** implement `RateLimiter` beyond a test stub (`AlwaysAllowRateLimiter`, gated behind `#[cfg(any(test, feature = "test-utils"))]`). No Redis-backed rate limiter exists anywhere in the codebase — rate limiting is not currently enforced. `RateLimitConfig` defines 4 tiers (human, agent-standard, agent-elevated, agent-platform) as a design target.
Manages Redis pub/sub fan-out, presence tracking, and typing indicators. In multi-community mode all tenant-visible keys are prefixed or otherwise partitioned by community (`buzz:{community}:...`) so channel fan-out, presence, typing, and cache invalidation cannot cross hosts.
The subscriber uses a **dedicated**`redis::aio::PubSub` connection — not from the pool. This is intentional: pool connections cannot hold `PSUBSCRIBE` state.
**Current state:** The subscriber loop is spawned in `buzz-relay/src/main.rs` and populates the broadcast channel. A consumer task subscribes via `pubsub.subscribe_local()`, calls `sub_registry.fan_out()` on each received event, and delivers matches to local WebSocket connections via `conn_manager.send_to()`. Multi-node fan-out is now wired end-to-end. Local-echo deduplication is implemented via `AppState.local_event_ids` — events published by the local relay instance are tracked and skipped when received via the Redis round-trip.
**Reconnection:** exponential backoff 1s → 30s (`backoff_secs * 2`). Backoff resets to 1s only after a clean stream end, not on each reconnect attempt.
**Presence:**`SET buzz:presence:{pubkey_hex} {status} EX 180` — 180-second TTL (3× the 60-second heartbeat interval). Single missed heartbeat does not cause presence flap.
**Hash chain:** each entry stores `prev_hash` (hash of the previous entry). In multi-community mode audit heads/chains are per-community; operator metrics may aggregate, but tenant-readable audit verification walks one community chain. `verify_chain()` walks entries and recomputes hashes to detect tampering. Genesis entry uses `GENESIS_HASH` (64 zeros).
Parses, validates, and executes channel-scoped workflow definitions. In multi-community mode workflow definitions, runs, approvals, webhook routes, and schedules inherit the host-derived community and evaluate triggers only against events in that community.
Note: Both `TriggerDef` and `ActionDef` use serde internally-tagged enums. Triggers use `on:` as the tag field; actions use `action:` as the tag field. Fields are flattened into the parent struct, not nested.
**Approval gates:**`request_approval` action returns `StepResult::Suspended` with a generated UUID token, but the engine does not yet persist the token or resume execution — runs that hit an approval gate are marked as failed (🚧 WF-08). `execute_from_step()` exists for future resumption support.
**Does NOT:** recursively resolve templates (single-pass only). Does NOT queue workflow runs when at capacity — returns `CapacityExceeded` immediately.
Real-time voice lives inside `buzz-relay` (`src/audio/`), not a separate crate. A WebSocket endpoint (`wss://.../huddle/{channel_id}/audio`) authenticates each participant with a NIP-42 challenge, checks channel membership, admits them to an in-memory room, and forwards opaque Opus frames between peers. No external SFU.
**Frame protocol (v2):** 8-byte big-endian header (sequence `u16`, 48 kHz timestamp `u32`, level dBov `i8`, flags `u8`) followed by an opaque Opus payload. Invalid `level_dbov` values are clamped rather than dropped — losing a metric beats losing audio.
**Room state:** an admission guard synchronizes joins against the room's ended flag; soft cap 25 peers (hard cap 255 via `u8` peer index). Per-peer audio uses a bounded channel (drop-on-full); the control channel is separate and never drops join/leave.
**Lifecycle events:** the relay emits Nostr events for participant joined / left and huddle ended; the desktop client emits huddle started and guidelines. When the last peer leaves, the room ends and the channel archives atomically.
`buzz-acp` spawns AI agent subprocesses (1–32, default 1), connects to the relay via WebSocket with NIP-42 auth, discovers channels via REST API, and queues `@mention` events per channel. At most one prompt is in-flight per channel. Queued events are batched into a single prompt sent via `session/prompt` over ACP.
The `buzz-admin` binary is shipped in the relay Docker image (`/usr/local/bin/buzz-admin`) and is the recommended way to manage relay membership in production. Use `./run.sh add-member`, `./run.sh remove-member`, and `./run.sh list-members` in Docker Compose deployments.
| `events` | All stored Nostr events; monthly range-partitioned by `PARTITION BY RANGE` on `created_at`; multi-community mode keys every tenant-visible event by `community_id` |
| `channels` | Channel records (type, visibility, canvas, topic); `community_id` is immutable after creation in multi-community mode |
| 3 | **No dedicated typing REST endpoint** | Typing indicators (kind 20002) are delivered via both local fan-out and Redis pub/sub (cross-node). There is no REST endpoint to query current typers — `/api/presence` returns online/away status only, not typing state. |
| 4 | **Huddle recording/tracks not built** | Voice, room lifecycle, and join/leave/end events are wired (see Huddle Audio above). Recording and per-track publishing have reserved kinds but no producer yet. |
| 5 | **Approval gates not wired end-to-end** | The executor returns `StepResult::Suspended` and the relay has grant/deny API endpoints with DB CRUD, but the engine intercepts before creating `WaitingApproval` rows — runs that hit an approval gate are marked as Failed (🚧 WF-08). |
| 6 | **Workflow actions partially stubbed** | The `send_dm` and `set_channel_topic` workflow actions are in the schema but return `NotImplemented` — a run that reaches one fails at execution (🚧 WF-07). |
