Files
alknet/docs/architecture/call-protocol.md
glm-5.1 596c89ce24 refactor!: rebrand wraith to alknet
Rename all crates, CLI commands, constants, type names, doc comments,
and documentation from wraith to alknet. Includes wire-protocol changes:
ALPN wraith-ssh -> alknet-ssh, reserved destination prefix wraith- ->
alknet-, SSH auth username wraith -> alknet.
2026-06-05 10:04:32 +00:00

402 lines
17 KiB
Markdown

---
status: draft
last_updated: 2026-06-04
---
# Call Protocol
## What
A bidirectional, transport-agnostic call and event protocol that runs over
authenticated pipes. It supports request/response calls, streaming
subscriptions, and unidirectional events — all using the same wire format. The
protocol is defined as a spec + handler + registry; downstream consumers (NAPI,
Python, hub/spoke) register their own operations without modifying core.
## Why
The current control channel (ADR-018) is unidirectional (client → server) and
provides fire-and-forget event dispatch without request/response semantics.
The call protocol generalizes it to support bidirectional calls (ADR-024) and
downstream service registration (ADR-025), enabling the hub/spoke model where
spokes expose operations the hub invokes.
## Architecture
### Operation Paths
Operation names use slash-based paths aligned with URL routing conventions:
```
/{spoke}/{service}/{op}
```
- **spoke** — identity prefix of the node that exposes the operation. The hub
uses this segment to route calls to the correct connected node.
- **service** — the logical service namespace. Groups related operations
under one handler prefix.
- **op** — the specific operation within that service.
Examples:
| Path | Meaning |
|------|---------|
| `/dev1/fs/readFile` | Spoke `dev1`, service `fs`, operation `readFile` |
| `/dev1/bash/exec` | Spoke `dev1`, service `bash`, operation `exec` |
| `/hub/agent/chat` | Hub's own `agent` service, operation `chat` |
| `/hub/sessions/list` | Hub's own `sessions` service, operation `list` |
| `/browser-1/notify/alert` | Browser spoke `browser-1`, `notify` service |
This three-level routing mirrors iroh's ALPN dispatch: the first segment
routes to a connected node (like ALPN routes to a protocol handler), the
remaining path dispatches within that node's registry. See ADR-025 for the
handler/spec separation decision.
The `namespace` field on `OperationSpec` is derived from the path (`namespace`
= second path segment). It's a convenience accessor for ACL matching and
service grouping.
### Wire Format: EventEnvelope
Every message on the wire is a length-prefixed JSON `EventEnvelope`:
```rust
pub struct EventEnvelope {
pub r#type: String, // Event type (e.g., "call.requested", "call.responded")
pub id: String, // Correlation key (requestId, topic, or "" for broadcasts)
pub payload: Value, // JSON payload — schema depends on event type
}
// Frame: 4-byte big-endian length prefix + UTF-8 JSON body
```
This is the same format used by `@alkdev/pubsub` adapters. It is JSON because
it must be consumable from JavaScript, Python, and any language. The envelope
is transport-agnostic — it runs over SSH channels, WebTransport streams, iroh
bidirectional streams, WebSocket, or Worker postMessage.
Binary payloads (postcard, protobuf, etc.) are base64-encoded in the `payload`
field. The envelope itself stays JSON for cross-language compatibility.
### Call Protocol Events
Five event types carry request/response and subscription semantics:
| Event | Direction | Purpose |
|-------|-----------|---------|
| `call.requested` | Caller → Handler | Initiate a call or subscription |
| `call.responded` | Handler → Caller | Deliver a result (one for calls, many for subscriptions) |
| `call.completed` | Handler → Caller | Signal end of subscription stream |
| `call.aborted` | Either side | Cancel the call/subscription |
| `call.error` | Handler → Caller | Signal an error |
**`call.error` payload**:
```json
{
"code": "string",
"message": "string",
"retryable": false
}
```
**A call is just a subscribe that resolves after one event.** Both `call()` and
`subscribe()` send the same `call.requested` event. The difference is
consumption pattern:
- **`call()`**: Sends `call.requested`, resolves `Promise` on first `call.responded`
- **`subscribe()`**: Sends `call.requested`, yields each `call.responded` until `call.completed` or `call.aborted`
The `id` field carries the `requestId` for correlation.
### Bidirectional Calls and Routing
Both sides of a connection can initiate calls. The hub routes calls to spokes
using the first path segment:
```
Hub (server) Spoke: "dev1" (client)
│ │
│ call.requested │
│ name: "/dev1/fs/readFile" │
│ payload: { path: "/src/main.rs" } │
│──────────────────────────────────────────▶│
│ │
│ call.responded │
│ id: <requestId> │
│ payload: { content: "fn main()..." } │
│◀──────────────────────────────────────────│
│ │
│ Spoke exposes /dev1/fs/*, │
│ /dev1/bash/* to hub │
│ │
│◀─ call.requested ────────────────────────│
│ name: "/hub/agent/chat" │
│ payload: { provider: "anthropic", ... } │
│ │
│── call.responded ──────────────────────▶ │
│ id: <requestId> │
│ payload: { completion: "..." } │
```
The hub's registry includes:
- **Hub-local operations** (`/hub/*`) — handled directly
- **Remote operations** (`/{spoke}/*`) — forwarded to the spoke connection
When the hub routes `/dev1/fs/readFile` to spoke `dev1`, it strips the spoke
prefix and delivers the call to the spoke's local registry as `/fs/readFile`.
The spoke doesn't need to know its own alias.
### Hub/Spoke Architecture
```
┌─────────────────────────────────┐
│ Hub │
│ │
│ Hub-local services: │
│ /hub/agent/chat (LLM coord) │
│ /hub/agent/complete │
│ /hub/sessions/list │
│ /hub/sessions/history │
│ │
│ Spoke registry (discovered): │
│ /dev1/fs/* → dev1 connection │
│ /dev1/bash/* → dev1 connection │
│ /dev2/fs/* → dev2 connection │
│ /browser-1/notify/* → WT conn │
└──────┬───────┬───────┬──────────┘
│ │ │
┌─────────▼┐ ┌───▼────┐ ┌▼───────────┐
│ Dev Spoke│ │Dev Spk │ │Browser Spoke│
│ "dev1" │ │"dev2" │ │"browser-1" │
│ /fs/* │ │/fs/* │ │/notify/* │
│ /bash/* │ │/bash/* │ │ │
│ /search/*│ │ │ │ │
└───────────┘ └────────┘ └─────────────┘
```
When a spoke connects, it registers its operations with the hub:
```
spoke → hub: call.requested { name: "/hub/services/register", payload: {
spoke: "dev1",
operations: ["/fs/readFile", "/fs/writeFile", "/bash/exec", "/search/query"]
}}
```
The hub adds these to its routing table with the spoke prefix. Other spokes
and browser clients can then call `/dev1/fs/readFile` without knowing how
the hub routes it internally.
### Operation Registry
The operation registry maps paths to specs and handlers. **Specs and handlers
are separate** — downstream consumers register both (ADR-025).
```rust
pub struct OperationSpec {
pub name: String, // e.g., "/fs/readFile", "/agent/chat"
pub namespace: String, // e.g., "fs", "agent"
pub op_type: OperationType, // Query, Mutation, Subscription
pub input_schema: Value, // JSON Schema for input
pub output_schema: Value, // JSON Schema for output
pub access_control: AccessControl, // Required scopes/resources
}
pub enum OperationType {
Query, // Read-only, idempotent (e.g., "/fs/readFile", "/search/query")
Mutation, // Side effects (e.g., "/bash/exec", "/sessions/create")
Subscription, // Streaming (e.g., "/events/subscribe")
}
pub struct AccessControl {
pub required_scopes: Vec<String>, // AND-checked
pub required_scopes_any: Option<Vec<String>>, // OR-checked
pub resource_type: Option<String>, // e.g., "service"
pub resource_action: Option<String>, // e.g., "read"
}
```
**Registration is separated from implementation:**
```rust
// Core registers discovery operations
registry.register(OperationSpec { name: "/services/list", ... }, list_services_handler);
registry.register(OperationSpec { name: "/services/schema", ... }, schema_handler);
// A dev env spoke registers its tools
registry.register(OperationSpec { name: "/fs/readFile", ... }, fs_read_handler);
registry.register(OperationSpec { name: "/bash/exec", ... }, bash_exec_handler);
// A browser client registers notification UDFs
registry.register(OperationSpec { name: "/notify/alert", ... }, notify_handler);
```
Core-provided operations use short paths without a spoke prefix
(`/services/list`, `/services/schema`). They live on whatever node the
caller is connected to. Spoke-prefixed operations (`/dev1/fs/readFile`)
are routed by the hub.
### ACL Per Operation Path
Access control maps to path prefixes using standard URL-like matching:
| Pattern | Matches | Purpose |
|---------|---------|---------|
| `/dev1/*` | All operations on spoke `dev1` | Full access to a spoke |
| `/*/fs/*` | `fs` service on any spoke | Read file access across dev envs |
| `/*/bash/*` | `bash` service on any spoke | Shell access (higher risk) |
| `/hub/agent/*` | Hub LLM agent | LLM calls |
| `/hub/sessions/*` | Hub session management | Session history |
| `/browser-1/notify/alert` | Specific operation on specific spoke | One UI notification |
Higher-risk operations (shell, filesystem write) can require tighter scopes
than read-only operations. The ACL evaluates against the caller's
`Identity.scopes` and `Identity.resources` from the auth layer (see auth.md).
### Service Discovery
The `/services/list` and `/services/schema` operations expose what a node
offers. Read-only — no admin operations:
| Operation | Type | Description |
|-----------|------|-------------|
| `/services/list` | Query | List registered operation paths + metadata |
| `/services/schema` | Query | Get `OperationSpec` for a specific operation |
These tell the caller: "here's what you can call." They are not a control
panel. Access control is enforced at the operation level.
### PendingRequestMap
Manages in-flight calls and subscriptions. Correlates `call.responded` events
back to the original `call.requested`:
```rust
pub struct PendingRequestMap {
pending: HashMap<String, PendingEntry>,
}
enum PendingEntry {
Call {
tx: oneshot::Sender<Result<Value>>,
timeout: Instant,
},
Subscribe {
tx: mpsc::Sender<Result<Value>>,
timeout: Option<Instant>,
},
}
```
When a `call.responded` event arrives:
- If `PendingEntry::Call` → resolve the oneshot, delete entry
- If `PendingEntry::Subscribe` → push to the mpsc channel, keep entry alive
When `call.completed` arrives on a subscription → close the mpsc channel, delete
entry. When `call.aborted` arrives → cancel/drop whichever side initiated it. A
`call.aborted` for an unknown `requestId` is silently discarded — no error
response is generated.
Timeouts prevent dangling entries. A background task sweeps expired entries
periodically.
### Protocol Adapter Layer
The call protocol is transport-agnostic by design. It maps to any transport
that carries `EventEnvelope` frames:
| Transport | Channel mechanism | Direction |
|-----------|-------------------|-----------|
| SSH | Reserved `direct_tcpip` destination (ADR-018) | Bidirectional over SSH channel |
| WebTransport | Bidirectional stream after CONNECT | Bidirectional over WT stream |
| iroh QUIC | Bidirectional `open_bi()` / `accept_bi()` | Bidirectional over QUIC stream |
| WebSocket | Single WS connection | Bidirectional over WS frames |
| Worker | `postMessage` | Bidirectional over structured clone |
The framing is always: 4-byte BE length prefix + JSON. The envelope shape is
the same regardless of transport.
### Relationship to @alkdev/pubsub and @alkdev/operations
The call protocol in core is a Rust reimplementation of the same protocol
defined in `@alkdev/operations`. The TypeScript implementation provides:
- `PendingRequestMap` — request/response correlation
- `CallHandler` — bridges pubsub events to operation registry
- `OperationSpec`, `AccessControl`, `Identity` — type definitions
The Rust implementation mirrors these types and behaviors. TypeScript consumers
continue using `@alkdev/operations` over `@alkdev/pubsub` adapters (including
the `event-target-alknet` adapter). Rust consumers use core's registry directly.
Both speak the same wire protocol and can interoperate.
The key principle: **the same `EventEnvelope` can flow from a Rust handler
through core, out over SSH channel, into a JavaScript pubsub adapter, and
be dispatched through `@alkdev/operations`'s call handler** — with zero
translation at the wire level.
### Agent Service Pattern
The hub commonly runs an agent service that coordinates between LLM providers
and tool calls. This service is just another set of registered operations —
no special treatment:
- `/hub/agent/chat` — send a message, get a completion. Routes to the
appropriate LLM provider based on available spokes and configuration.
- `/hub/agent/complete` — streaming completion. Yields tokens as they arrive.
- `/hub/sessions/list` — list session histories (backed by Honker or other
durable storage).
- `/hub/sessions/history` — retrieve a specific session's message history.
The agent service uses the same call protocol to invoke tools on spokes:
`/dev1/fs/readFile` for file access, `/dev1/bash/exec` for shell commands. It
stores session state via whatever mechanism the hub deployment provides — core
doesn't mandate Honker or any specific storage.
## Constraints
- The call protocol does not depend on Honker, SQLite, or any database. The
`PendingRequestMap` is in-memory. Durable session storage is a consumer concern.
- Operation specs use JSON Schema. Complex sub-structures (postcard, protobuf)
can be carried as base64-encoded blobs in the `payload`, but the envelope
itself is always JSON.
- Service discovery (`/services/list`, `/services/schema`) is read-only. No
admin operations are exposed through the call protocol itself.
- Batch is not a protocol primitive. Multiple `call.requested` events with
correlated `requestId`s provide equivalent semantics.
- The spoke prefix in the operation path is a routing mechanism, not a security
boundary. ACL is enforced at the `AccessControl` level, not by path prefix
alone. A spoke that exposes `/dev1/bash/exec` can restrict access via
`required_scopes` — not every authenticated identity should have shell access.
## Open Questions
- **OQ-20**: How does the hub track which spokes expose which operations when
spokes connect and disconnect? Registration on connect and cleanup on
disconnect, or heartbeat-based discovery? See
[open-questions.md](open-questions.md).
- **OQ-22**: Should the call protocol support streaming inputs (client streaming
in gRPC terms), or is client→server always a single request payload with
streaming only server→client? See [open-questions.md](open-questions.md).
## Design Decisions
| ADR | Decision | Summary |
|-----|----------|---------|
| [018](decisions/018-control-channel-for-pubsub.md) | Control channel for pubsub | Reserved destination for event bus |
| [024](decisions/024-bidirectional-call-protocol.md) | Bidirectional call protocol | Generalizes ADR-018, both sides can call |
| [025](decisions/025-handler-spec-separation.md) | Handler/spec separation | Downstream registers operations without modifying core |
## References
- [auth.md](auth.md) — Identity and `IdentityProvider` trait
- [napi-and-pubsub.md](napi-and-pubsub.md) — NAPI wrapper and pubsub adapter
- [server.md](server.md) — Channel handling and control channel routing
- [transport.md](transport.md) — Transport abstraction
- [configuration.md](../research/configuration.md) — ForwardingPolicy, service metadata
- `@alkdev/pubsub` — TypeScript event target adapters and `EventEnvelope`
- `@alkdev/operations` — TypeScript call protocol, `OperationSpec`, registry
- `@alkdev/storage``peer_credentials` table, ACL graph, `Identity`
- [irpc](/workspace/irpc) — iroh streaming RPC (postcard-only, Rust-to-Rust)
- [iroh](/workspace/iroh) — P2P QUIC transport