Files
unshell/PROTOCOL.md
T
Michael Mikovsky 1af134104e Add TreeTest
2026-04-22 10:03:24 -06:00

34 KiB

UnShell Network Protocol Specification

Version: 0.2.0
Status: Draft — implementation in progress
Last updated: 2026-04-21


Overview

The UnShell protocol is a tree-addressed, message-passing protocol for command and control (C2) operations. It is designed around a homogeneous node model: every participant (payload, operator, router) is structurally identical from the protocol's perspective. Each node owns a set of paths in a global tree and responds to requests addressed to those paths.

  /agents/abc123/shell/exec    ← a path owned by payload node "abc123"
  /agents/abc123/files/read    ← another path on the same payload
  /operator/sess1              ← operator node's own registration path
  /router/nodes                ← router's built-in endpoint

A router is a dumb relay. It reads the destination path from a packet header and forwards the packet body to whichever node registered that path. It has no application logic. It does not interpret payloads. Think of it as a post office: it reads the address on the envelope and delivers the contents without opening them.


Design Goals

  1. Shallow protocol, deep functionality. The base protocol is minimal. Complexity comes from APIs stacked on top (RESTful paths, modules), not from the wire format.

  2. Two communication patterns. One-time events (request/response) and streams (bidirectional channels) — not one-size-fits-all.

  3. Transport independence. TCP is the first transport, but the protocol must not assume TCP. HTTPS, ICMP, and other transports will be added later. The protocol layer sits above the transport layer via a Transport trait.

  4. No explicit node types. Nodes are identified by registered paths, not by type. This allows flexible deployment (implant, operator, relay, tunnel endpoint).

  5. Forward compatibility. Adding new fields to message types must not break existing implementations. Use rkyv's archived format, which supports this.

  6. Detection-aware. The handshake is kept simple. For stealth, swap in an encrypted transport (HTTPS, custom obfs) without changing the protocol.


Fundamental Design

The UnShell protocol has two communication patterns:

  1. One-time events — Request → Response, reliable, stateless on router
  2. Streams — Open → Bidirectional data flow → Close, persistent, fastpath routing

This mirrors HTTP (request/response) and WebSockets/VPNs (persistent streams).

No Explicit Node Types

The protocol does not distinguish between payloads, operators, or routers. Nodes are identified by their registered paths, not their type.

Recommended path conventions (not required):

  • /agents/<node_id>/ — for implants
  • /operator/<session_id>/ — for CLI sessions
  • /router/ — for built-in router endpoints
  • /tunnel/<name>/ — for stream endpoints

The complexity comes from APIs stacked on top, not from the protocol itself. This is intentional — the protocol is shallow; the functionality is in the routes.

┌─────────────────┐    ┌─────────────────────────────────────────────┐
│   Implant Node  │    │             Router Node                     │
│                 │    │                                             │
│  - Connects to  │    │  - Accepts TCP from any node                │
│    router       │    │  - Routes by path prefix match              │
│  - Registers    │    │  - Routes by stream_id for fastpath         │
│    paths        │    │  - NO application logic beyond routing      │
│  - Hosts API    │    │  - Has /router/ endpoints                   │
└────────┬────────┘    └─────────────────────────────────────────────┘
          │ TCP
          │
┌────────▼────────┐
│  Operator Node  │
│  (ush-cli)      │
│                 │
│  - Connects to  │
│    router       │
│  - Registers    │
│    paths        │
│  - Interactive  │
│    REPL shell   │
└─────────────────┘

NodeType enum (DEPRECATED): Removed in v0.2.0. Nodes are identified by paths, not types. Existing implementations should ignore or omit this field.


Wire Format

Every transmission uses a two-part framed message:

┌──────────────────────────────────────────────────────────────────────┐
│  Part 1: Header                          │  Part 2: Payload          │
│                                          │                           │
│  [u32 big-endian length]                 │  [u32 big-endian length]  │
│  [rkyv-serialised FrameHeader bytes]     │  [rkyv payload bytes]    │
│                                          │                           │
│  Router reads this to determine routing  │  Router forwards opaque   │
└──────────────────────────────────────────┴───────────────────────────┘

Both length fields are big-endian u32, so the maximum frame size is ~4GB per part. In practice, packets should be much smaller.

Two Communication Patterns

The protocol supports two distinct patterns:

1. One-time Events (Request/Response):

  • Client sends FrameType::Request with dst_path and request_id
  • Router routes by longest-prefix match on dst_path
  • Server responds with FrameType::Response with same request_id
  • Reliable, stateless, exactly-once semantics via request_id

2. Streams (Bidirectional Channels):

  • Client sends FrameType::StreamOpen with dst_path
  • Router assigns stream_id (u16), registers in stream table, responds
  • Subsequent frames use FrameType::StreamData or StreamClose with stream_id
  • Router uses fastpath: looks up stream_id → node directly, no path matching
  • Bidirectional: both sides can send StreamData frames
  • Clean close: either side sends StreamClose, router cleans up

This mirrors HTTP (request/response) and WebSockets/VPN tunnels (persistent streams).

Why two parts?

The router needs to know where to send a packet. With a single rkyv blob, the router would have to deserialise the entire packet just to read the destination path. With a separate header, the router deserialises only the small header (typically < 100 bytes) and forwards the payload bytes untouched. This is efficient and keeps the protocol transport-agnostic at the router level.

FrameHeader

/// The frame header that every frame starts with.
/// For events: router reads dst_path for routing.
/// For streams: router reads stream_id for fastpath routing.
#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct FrameHeader {
    /// Frame type: REQUEST, RESPONSE, STREAM_OPEN, STREAM_DATA, STREAM_CLOSE
    pub frame_type: FrameType,

    /// Destination path for REQUEST and STREAM_OPEN.
    /// Ignored for RESPONSE (uses src_path from request) and STREAM_DATA/CLOSE (uses stream_id).
    pub dst_path: Option<String>,

    /// Source path of the sender.
    /// Used by the destination to know where to send responses.
    pub src_path: String,

    /// Request ID for correlation (REQUEST/RESPONSE pairs).
    /// None for stream frames.
    pub request_id: Option<u64>,

    /// Stream ID for fastpath routing (STREAM_DATA, STREAM_CLOSE).
    /// None for REQUEST/RESPONSE.
    pub stream_id: Option<u16>,
}

/// Discriminates between the two communication patterns.
#[derive(Archive, Serialize, Deserialize, Debug, Clone, PartialEq)]
pub enum FrameType {
    /// One-time event: request from client.
    Request = 0x01,

    /// One-time event: response from server.
    Response = 0x02,

    /// Stream: open a persistent bidirectional channel.
    StreamOpen = 0x03,

    /// Stream: data over an established stream (fastpath).
    StreamData = 0x04,

    /// Stream: close an established stream.
    StreamClose = 0x05,

    /// Legacy: sent by a newly connected node to register itself.
    Handshake = 0x10,

    /// Legacy: router's response to handshake.
    HandshakeAck = 0x11,
}

Why String for paths instead of Vec<String>?

A single /-delimited string serialises smaller (one allocation, no Vec overhead) and is easier for the router to do prefix matching on. Components are split at application layer, not at the wire level.


Handshake Protocol

A minimal registration handshake to tell the router which paths this node owns.

Node                         Router
 │                              │
 │──── TCP connect ────────────>│
 │                              │
 │──── Handshake ──────────────>│  (FrameType::Handshake)
 │     registered_paths: [...]  │
 │                              │
 │<─── HandshakeAck ────────────│  (FrameType::HandshakeAck)
 │     accepted: true           │
 │     assigned_base_path: "..."│
 │                              │
 │  [now registered, can send   │
 │   and receive frames]        │

Design note: The handshake is kept simple to minimize detection surface. However, the pattern (length-prefixed frames after TCP connect) is detectable. For stealth, use an encrypted transport layer (see Transport section).

Handshake timeout: If the node does not receive a HandshakeAck within 5 seconds, it closes the connection and retries.

Router timeout: If the router does not receive a Handshake within 10 seconds of a TCP connect, it closes the connection.

HandshakeMessage

#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct HandshakeMessage {
    /// The path prefixes this node owns. The router registers these.
    /// Example: ["/agents/abc123"]
    /// All sub-paths are implicitly owned by this prefix.
    pub registered_paths: Vec<String>,
}

HandshakeAck

#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct HandshakeAck {
    /// Whether the router accepted this node's registration.
    pub accepted: bool,

    /// The canonical base path assigned by the router (usually matches
    /// the first registered_path the node sent, but the router may adjust it).
    /// Empty string if rejected.
    pub assigned_base_path: String,

    /// Human-readable rejection reason if accepted == false.
    pub rejection_reason: Option<String>,
}

HandshakeAck

#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct HandshakeAck {
    /// Whether the router accepted this node's registration.
    pub accepted: bool,

    /// The canonical base path assigned by the router (usually matches
    /// the first registered_path the node sent, but the router may adjust it).
    /// Empty string if rejected.
    pub assigned_base_path: String,

    /// Human-readable rejection reason if accepted == false.
    pub rejection_reason: Option<String>,
}

Rejection reasons (v0.2):

  • "invalid_path" — a registered path is malformed or conflicts with a reserved prefix
  • "duplicate_path" — this path prefix is already registered by another node

Application Protocol: TreeRequest / TreeResponse

After the handshake, nodes communicate using TreeRequest / TreeResponse pairs.

A request travels: sender → router → destination node
A response travels: destination → router → original sender (using src_path from the request header as the destination path for the response)

TreeRequest

#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct TreeRequest {
    /// Unique ID for this request, generated by the sender.
    /// The responder echoes this back in TreeResponse.request_id.
    /// Enables correlation when multiple requests are in-flight.
    pub request_id: u64,

    /// The operation type.
    pub request_type: RequestType,

    /// Content-type string describing how to interpret `data`.
    /// Convention: "core/None", "core/Utf8String", "core/Bytes", etc.
    pub content_type: String,

    /// The operation payload. Interpretation depends on content_type.
    pub data: Vec<u8>,
}

#[derive(Archive, Serialize, Deserialize, Debug, Clone, PartialEq)]
pub enum RequestType {
    /// Read a value at this path.
    Read = 0,

    /// List available sub-paths and procedures at this path.
    GetProcedures = 1,

    /// Write a value to this path.
    Write = 2,

    /// Invoke a named procedure at this path.
    CallProcedure = 3,
}

TreeResponse

#[derive(Archive, Serialize, Deserialize, Debug, Clone)]
pub struct TreeResponse {
    /// Echoed from the corresponding TreeRequest.request_id.
    pub request_id: u64,

    /// Whether the operation succeeded or failed.
    pub status: ResponseStatus,

    /// Content-type of the response data.
    pub content_type: String,

    /// Response payload. Empty if status is an error with no data.
    pub data: Vec<u8>,
}

#[derive(Archive, Serialize, Deserialize, Debug, Clone, PartialEq)]
pub enum ResponseStatus {
    /// Operation completed successfully.
    Ok = 0,

    /// The requested path does not exist at the destination node.
    NoBranchError = 1,

    /// The requested operation is not supported at this path.
    UnsupportedOperation = 2,

    /// The destination node encountered an error executing the request.
    ExecutionError = 3,

    /// The request payload was malformed.
    ProtocolError = 4,
}

Content Type Convention

The content_type field in requests and responses follows a namespaced string convention, similar to MIME types but simpler:

Content type Meaning
"core/None" No data (empty payload)
"core/Utf8String" Raw UTF-8 string in data
"core/Bytes" Raw bytes (no specific interpretation)
"core/ProcedureList" Response to GetProcedures: rkyv-serialised Vec<ProcedureDescriptor>
"shell/Output" Shell command output (UTF-8 stdout + stderr)
"files/Bytes" Raw file contents

Custom module content types should use the module name as the namespace: "mymodule/MyType".


Path Routing

The router uses two routing methods:

1. Path-based Routing (Events)

For FrameType::Request and FrameType::StreamOpen, the router does longest-prefix match:

Registered paths:        Incoming dst_path:         Routes to:
/agents/abc123           /agents/abc123/shell/exec  → node "abc123"
/agents/xyz456           /agents/xyz456/files/read  → node "xyz456"
/router                  /router/nodes              → router's built-in handler

Rules:

  1. Split dst_path by /, find all nodes whose registered_paths is a prefix of dst_path.
  2. Choose the node with the longest matching prefix (most specific).
  3. If no match, return a TreeResponse { status: NoBranchError, ... } to the sender.
  4. If multiple nodes match with equal prefix length, route to most recently registered.

2. Stream ID Fastpath

For FrameType::StreamData and FrameType::StreamClose, the router uses stream ID lookup:

Stream table (router):
stream_id: u16  →  node (connection handle)

Frame header:
stream_id: 42   →  Direct lookup → node "abc123"

Rules:

  1. Router maintains a HashMap<u16, Node> for active streams.
  2. StreamOpen returns a unique stream_id (assigned by router).
  3. All subsequent StreamData frames use this stream_id for O(1) lookup.
  4. StreamClose removes the entry from the stream table.
  5. If stream_id not found (already closed), frame is discarded with warning.

Router Built-in Endpoints

The router itself hosts a small set of endpoints at /router/:

Path RequestType Returns
/router/nodes GetProcedures List of all connected nodes with their paths and types
/router/ping Read "pong" (latency check)

Real-World Scenario Analysis

This section stress-tests the protocol against conditions you'll actually encounter on an engagement or in the wild.

Scenario 1: Flaky Network / Payload Reconnect

Situation: A payload is behind a NAT and its TCP connection to the router drops (firewall timeout, network hiccup, target rebooted).

What happens:

  1. Payload's recv() call returns TransportError::Disconnected (EOF) or TransportError::Io.
  2. Payload closes the TcpStream, waits 5 seconds, attempts reconnect.
  3. Router's node thread for this connection receives EOF, removes the NodeInfo entry from the registry, exits cleanly.
  4. Payload reconnects, sends a new HandshakeMessage with the same registered_paths.
  5. Router re-registers it. The operator runs list and sees the payload appear again.

Operator experience: The operator may see the payload disappear from list briefly during the reconnect window. Sessions associated with that payload become temporarily unresponsive. After reconnect they work again.

Stream impact: Any open streams are lost on disconnect. Client must re-establish with new StreamOpen after reconnect.


Scenario 2: Operator Disconnects Mid-Session

Situation: The operator closes the CLI (Ctrl+C, terminal crash) while a payload is still connected.

What happens:

  1. Router's operator node thread receives EOF. Removes /operator/sess1 from registry.
  2. Any in-flight TreeRequest from that operator that the payload hasn't responded to yet: the payload sends a TreeResponse back, router tries to route it to /operator/sess1, finds no registered node, discards the response and logs a warning.
  3. Payloads remain connected. The payload's modules keep running (persistence).

Operator experience: When the operator reconnects, it gets a new session ID (/operator/sess2). It runs list to see what payloads are still connected. Background operations on payloads that were running continue.

Key insight: The payload is the persistent state. The operator is ephemeral. This is the "background services without another process" design — payload modules keep running even when no operator is connected.


Scenario 3: Multiple Operators

Situation: Two operators connect simultaneously (e.g., red team lead and junior analyst).

What happens:

  1. Both connect, get unique session IDs: /operator/sess1 and /operator/sess2.
  2. Both can send requests to any payload path.
  3. Responses go back to the requesting operator's src_path.
  4. There is no access control in v1. Both operators have full access to all paths.

Collision scenario: Both operators call /agents/abc123/shell/exec "ls" at the same time. The payload processes requests sequentially (single-threaded recv loop). It sends two responses, each echoing the correct request_id. Each response routes to the operator that sent the matching request (via src_path in the request header).

Failure mode in v1: No locking on the payload side. If a Write and a Read to the same resource happen simultaneously, the result is whatever order the TCP stack delivers them. This is acceptable for v1 red team use where multiple operators are unlikely to stomp each other on the same target simultaneously.

Future: Add an optional exclusive-lock request type for sensitive operations.


Scenario 4: Large Data Transfer (File Exfiltration)

Situation: Operator requests a large file (100MB) from a target.

Problem with current design: The u32 length prefix allows up to 4GB per packet, but buffering 100MB in RAM on the payload before sending is problematic on constrained targets.

V1 approach: Accept this limitation. Files up to ~50MB should be fine in practice for most engagements. The TreeRequest.data field holds the serialised request; the TreeResponse.data field holds the file bytes. For v1, the payload reads the entire file into a Vec<u8> and sends it.

Future (chunked streaming): Add PacketType::Stream and PacketType::StreamEnd to support chunked transfers. The router passes stream packets through without buffering. The operator reassembles chunks. This requires a stream ID in the header to demultiplex concurrent streams.


Scenario 5: AV / EDR Detection via Network Traffic

Situation: The payload is on a monitored network. The router is a VPS. Plain TCP connections from the target to an unknown IP may trigger alerts.

V1 limitation: Plaintext TCP. Easy to detect.

Transport abstraction payoff: The Transport trait makes this the router's and payload's responsibility, not the protocol's. To switch to HTTPS:

  1. Implement HttpsTransport: Transport for the payload.
  2. Have the payload connect to a domain name (baked at compile time) on port 443.
  3. The router terminates TLS and speaks the same framing protocol underneath.
  4. From the network's perspective: an HTTPS connection to what looks like a CDN.

Nothing in the protocol spec changes. Only the Transport implementation swaps.


Scenario 6: Router Crash / Restart

Situation: The router process crashes or is restarted (e.g., VPS reboot).

What happens:

  1. All node TCP connections drop simultaneously.
  2. All nodes (payloads and operators) receive Disconnected errors.
  3. All nodes enter reconnect loops.
  4. Once the router restarts and starts accepting connections, nodes reconnect and re-register in whatever order their reconnect loops fire.
  5. The router comes back to a clean state (no session persistence across restarts in v1).

Failure mode: In-flight requests at the time of crash are lost. The operator may see commands that appear to hang. The operator should use a timeout on requests.

V1 mitigation: Request timeout is on the operator's TODO list. For now, the operator can detect a crash by the payload disappearing from list.

Future: The router could persist its node registry to disk and recover after restart.


Scenario 7: Malformed Packet / Bad Actor

Situation: Something sends a malformed packet to the router (fuzzer, compromised node, network corruption).

Defense layers:

  1. Length prefix: If the announced frame length is > a max limit (e.g., 64MB), the router closes the connection with TransportError::FrameTooLarge. No allocation.
  2. rkyv deserialisation: If the header bytes don't decode to a valid PacketHeader, rkyv::access returns an error. The router closes the connection.
  3. Unknown dst_path: Routes to no node, sends back NoBranchError.
  4. No authentication in v1: Any node can send to any path. This is acceptable for v1 where the router address is only known to the operator. Authentication (shared secret or challenge-response) is a v2 concern.

Scenario 8: Pivot / Multi-Hop (Future)

Situation: A payload on an internal network can only reach another internal host, not the external router. A "pivot" payload acts as a relay.

How the tree model enables this:

  1. Pivot payload registers at /agents/pivot1/ on the external router.
  2. Pivot payload also acts as a local router for sub-agents.
  3. Sub-agents connect to the pivot payload's local listener and register.
  4. The pivot payload's /agents/pivot1/agents/ prefix forwards packets to sub-agents.
  5. From the external operator's perspective: /agents/pivot1/agents/sub1/shell/exec is just a deeper path. The routing is recursive.

Protocol requirement to enable this: Add NodeType::Router to the enum. A pivot payload registers as a Router node, not a Payload node. The external router knows to forward any path with /agents/pivot1/ prefix to the pivot connection, and the pivot routes further from there.

This does not require protocol changes to v1. Only the NodeType enum needs the Router variant added back.


Transport Trait

All transports implement this interface:

/// A bidirectional framed transport.
///
/// Implementations are responsible for framing: the two-part header+payload format
/// described in the wire format spec. Each `send` call transmits exactly one
/// logical frame (header + payload). Each `recv` call receives exactly one.
///
/// Implementations MUST use `read_exact`-style loops (not single `read` calls)
/// because TCP is a stream protocol and may deliver partial frames.
///
/// # Example (TCP)
///
/// ```rust
/// impl Transport for TcpTransport {
///     fn send(&mut self, header: &FrameHeader, payload: &[u8]) -> Result<(), TransportError> {
///         // 1. Serialise header to rkyv bytes
///         // 2. Write [u32 header_len][header bytes][u32 payload_len][payload bytes]
///         // 3. Use write_all() to ensure complete write
///     }
///     fn recv(&mut self) -> Result<(FrameHeader, Vec<u8>), TransportError> {
///         // 1. read_exact 4 bytes → header length
///         // 2. read_exact N bytes → header bytes
///         // 3. Deserialise header
///         // 4. read_exact 4 bytes → payload length
///         // 5. read_exact M bytes → payload bytes
///         // 6. Return (header, payload)
///     }
/// }
/// ```
pub trait Transport: Send {
    /// Send a frame (header + payload) over this transport.
    /// Blocks until all bytes are written.
    fn send(&mut self, header: &FrameHeader, payload: &[u8]) -> Result<(), TransportError>;

    /// Receive one frame from this transport.
    /// Blocks until a complete header+payload pair is received.
    fn recv(&mut self) -> Result<(FrameHeader, Vec<u8>), TransportError>;
}

#[derive(Debug, thiserror::Error)]
pub enum TransportError {
    #[error("I/O error: {0}")]
    Io(#[from] std::io::Error),

    #[error("frame header too large: {0} bytes (max {1})")]
    HeaderTooLarge(usize, usize),

    #[error("frame payload too large: {0} bytes (max {1})")]
    PayloadTooLarge(usize, usize),

    #[error("connection closed cleanly")]
    Disconnected,

    #[error("rkyv deserialisation failed")]
    DeserialiseError,
}

Alternative Transports

The protocol is transport-agnostic. Implementations can swap transports without changing protocol logic:

Transport Use Case
TcpTransport Default, straightforward
TlsTransport Encrypted channel (looks like HTTPS)
HttpTransport Tunnel over HTTP (looks like web traffic)
DnsTransport Tunnel over DNS queries
IcmpTransport Tunnel over ICMP (looks like ping)

For stealth, use a transport that blends with legitimate traffic. The protocol logic remains the same — only the transport layer changes.

Reconnect Policy

Payloads: On Disconnected or Io(_) from recv() or send():

  1. Close the transport.
  2. Wait 5 seconds.
  3. Attempt to create a new transport connection.
  4. If connect fails, wait 5 more seconds, retry. No maximum retry limit.
  5. On connect success, run the handshake again.

Operator CLI: On disconnect, print a message and exit. The operator restarts the CLI manually. (In a future version, the CLI could auto-reconnect and restore session.)


Frame Size Limits

Limit Value Reason
Max header length 64 KB Headers should never be this large; anything bigger is a bug or attack
Max payload length 64 MB Sufficient for most file transfers; larger files need chunked streaming (future)
Handshake timeout 10 s (router) Prevent resource exhaustion from hanging connections
Handshake ack timeout 5 s (node) Keep reconnect loops responsive

Version Compatibility

rkyv's archived format allows adding new fields (with #[rkyv(default)] for missing fields when reading older messages). This means:

  • New fields can be added to any message type without breaking existing implementations.
  • Removing or renaming fields IS a breaking change.
  • The FrameType enum should only gain variants, never lose them.

When breaking changes are necessary, bump the protocol version (future: add a version field to the framing format).


Implementation Checklist

  • src/protocol/mod.rs — re-exports all protocol types
  • src/protocol/types.rs — FrameHeader, FrameType, TreeRequest, TreeResponse, HandshakeMessage, HandshakeAck
  • src/protocol/content_types.rs — content type constants
  • src/transport/mod.rs — Transport trait, TransportError (add PayloadTooLarge variant)
  • src/transport/tcp.rs — TcpTransport implementing Transport
  • src/tree/mod.rs — Tree, Endpoint trait
  • ush-router/ — router binary with stream fastpath routing
  • ush-payload/ — payload binary with transport layer
  • ush-cli/ — operator REPL binary
  • Unit tests for framing round-trips, tree routing correctness
  • Integration test: two nodes through a real router
  • Stream test: open stream, send data both directions, close stream
  • Alternative transport: TlsTransport (stealth mode)

Leaf System Architecture

Terminology

Term Definition
Tree The network of endpoints connected through the UnShell protocol
Endpoint A node connected to the tree (payload, operator, router)
Leaf A data object or service hosted on an endpoint

Design Goals

  1. Rich leaves, simple protocol — The protocol stays shallow. Complexity lives in leaves.
  2. Self-contained — Each leaf is an object with config, state, RPC, and streams.
  3. Composable — Leaves can be composed; a TTY leaf might wrap a process leaf.

Leaf Structure

Every leaf has three aspects:

Leaf {
  config:  Map<String, LeafValue>    // Stored configuration
  state:   LeafState                 // Running, Stopped, Error
  rpc:     Map<Name, Handler>        // Synchronous calls
  streams: Map<Name, StreamHandle>   // Bidirectional data flows
}

Configuration

Leaves expose configurable parameters as key-value pairs:

Type Example Use
Int rows: 24, cols: 80 Dimensions, limits
Bool echo: true, raw: false Mode flags
String shell: "/bin/bash", env: "TERM=xterm" Commands, env vars
Bytes (reserved for large config) Certificates, keys

RPC (Remote Procedure Call)

Synchronous request/response operations:

Request                              Response
------                               --------
start()    →                     →  { ok: true, state: Running }
reset()    →                     →  { ok: true, state: Running }
halt()     →                     →  { ok: true, state: Stopped }
resize(80, 24) →                 →  { ok: true }
config.get("rows") →              →  { value: 24 }
config.set("cols", 120) →        →  { ok: true }

Streams

Bidirectional data channels for long-lived connections:

Client                                                        Leaf
  │                                                             │
  ├───── StreamOpen(path="/tty/0/input") ────────────────────>│
  │<──── StreamOpenAck(stream_id=42) ──────────────────────────│
  │                                                             │
  ├───── StreamData(stream_id=42, data="ls -la\n") ──────────>│
  ├───── StreamData(stream_id=42, data="echo $TERM\n") ──────>│
  │<──── StreamData(stream_id=42, data="total 12\n") ─────────│
  │<──── StreamData(stream_id=42, data="drwxr-xr-x  2 user user 4096 Apr 21 10:30 .\n") │
  │<──── StreamData(stream_id=42, data="xterm-256color\n") ──│
  │                                                             │
  ├───── StreamData(stream_id=42, data="\x03") ───────────────>│ (Ctrl+C)
  │                                                             │
  ├───── StreamClose(stream_id=42) ──────────────────────────>│

Reference Implementation: TTY Leaf

Configuration:

struct TtyConfig {
    rows: u16,           // Terminal rows (default: 24)
    cols: u16,           // Terminal columns (default: 80)
    pixel_width: u16,   // Pixel width (default: 0)
    pixel_height: u16,  // Pixel height (default: 0)
    shell: String,      // Shell to spawn (default: "/bin/sh")
    env: Vec<(String, String)>,  // Environment variables
}

RPC Methods:

Method Description Returns
start() Spawn PTY and begin session { state: "Running", pid: u32 }
reset() Kill and respawn process { state: "Running", pid: u32 }
halt() Kill the process { state: "Stopped" }
resize(rows, cols) Update PTY size { ok: true }
config.get(key) Get config value { value: LeafValue }
config.set(key, value) Set config value { ok: true }
state() Get current state { state: LeafState, pid: Option<u32> }

Stream Bindings:

Stream Direction Description
input Client → TTY Send keystrokes to terminal
output TTY → Client Receive terminal output
both Bidirectional Combined input+output over single stream

Leaf Discovery

Endpoints expose available leaves via the GetProcedures mechanism:

REQUEST dst: "/agents/abc123/"
  request_type: GetProcedures
  content_type: "core/Utf8String"
  data: ""

RESPONSE
  status: Ok
  content_type: "core/ProcedureList"
  data: rkyv([...]) of ProcedureDescriptor:
    - path: "/tty/0"
      name: "tty/0"
      description: "PTY shell session 0"
      methods: ["start", "reset", "halt", "resize", "state", "config.get", "config.set"]
      streams: ["input", "output", "both"]
    - path: "/files"
      name: "files"
      description: "File system access"
      methods: ["read", "write", "list"]
      streams: []

Future Leaf Types

Leaf Config RPC Streams
TTY rows, cols, shell start, halt, resize input, output
Process cmd, args, env spawn, kill, wait stdout, stderr
TCP Tunnel lport, rhost, rport open, close, stats tunnel
FileSystem root_path read, write, list (none)
DNS domain, record_type query (none)