alknet/docs/architecture/decisions/017-call-protocol-client-and-adapter-contract.md

ADR-017: Call Protocol Client and Adapter Contract

Status

Accepted (amended 2026-06-26 — see "Amendments" below)

Context

The call protocol spec (ADR-012) defined the stream model as bidirectional — "both sides can initiate calls." But the spec only described the server side: CallAdapter implements ProtocolHandler, accepts incoming QUIC connections, and dispatches to the operation registry. The client side — who opens the connection, how calls are sent, how remote operations are discovered and imported — was left as OQ-15.

The need for the client side is concrete and immediate:

• Head/worker dispatch: a head node manages worker nodes (Vast.ai, RunPod, local Docker). The head needs to call operations on workers (exec, sync, status) and workers need to call back (report status, request work). The POC at /workspace/@alkdev/dispatch demonstrated this over SSH+axum; under the call protocol, it's cross-node composition.
• NAPI/Python adapters: Node.js and Python clients need to call operations on an alknet node. They speak the EventEnvelope wire format over a QUIC connection.
• Agent tool dispatch: an agent handler needs to call operations on remote nodes (tools, services) the same way it calls local operations — through OperationEnv::invoke(). The from_call adapter makes remote operations appear in the local registry.
• Cross-protocol interop: external systems (HTTP APIs, MCP servers) are imported via from_openapi and from_mcp. The reverse direction — exposing local operations to external systems — needs to_openapi and to_mcp.

The @alkdev/operations TypeScript package demonstrated the adapter patterns (from_openapi, from_mcp) and the buildEnv composition mechanism. The Rust implementation defines the canonical traits (ADR-013).

OQ-15 was constrained by ADR-014 (adapters take credential sources, not static tokens) and ADR-015 (adapter-registered operations are Internal by default). This ADR locks the remaining one-way door: the client/adapter contract architecture.

Decision

1. CallClient opens connections and shares the dispatch loop

CallClient opens a QUIC connection to a remote node with ALPN alknet/call. Once connected, the connection is symmetric — both sides can send and receive call.requested. The CallClient is not just a caller; it is also a callee. It has its own operation registry to dispatch incoming calls from the remote side.

pub struct CallClient {
    registry: Arc<OperationRegistry>,
    identity_provider: Arc<dyn IdentityProvider>,
}

impl CallClient {
    pub async fn connect(&self, addr: SocketAddr, credentials: CallCredentials) -> Result<CallConnection>;
}

The dispatch loop is shared between CallAdapter and CallClient. Once a connection is established (whether accepted by the adapter or opened by the client), the same logic applies: read EventEnvelope frames, dispatch to the operation registry, write responses, and send outgoing call.requested events for calls initiated on this side. The only difference is who opened the connection.

CallConnection provides:

• call(operation_id, input) -> ResponseEnvelope — send call.requested, await call.responded (one result)
• subscribe(operation_id, input) -> Stream<ResponseEnvelope> — send call.requested, yield each call.responded until call.completed or call.aborted
• abort(request_id) — send call.aborted, cascade to descendants (ADR-016)
• services_list() -> Vec<OperationSpec> — call services/list
• services_schema(name) -> OperationSpec — call services/schema

2. Connection direction is independent of call direction

Who opens the QUIC connection (who has the public IP, who uses a relay, who connects out reverse-runner style) is a connection-layer concern, not a protocol-layer concern. Once connected, both sides can call each other.

Topology	Who advertises	Who opens connection	Who can call whom
Public service	Server (public IP/domain)	Client	Both directions
P2P (iroh relay)	Both (relay-assisted)	Either	Both directions
Reverse (runner pattern)	Head (public IP)	Worker connects out	Both directions
Reverse (dispatch pattern)	Worker (public SSH port)	Head connects out	Both directions

The protocol does not distinguish "server" and "client" after connection establishment. The CallAdapter accepts connections; the CallClient opens connections. Both dispatch incoming and outgoing calls through the same mechanism.

3. from_call adapter imports remote operations

from_call does for call protocol endpoints what from_openapi does for HTTP APIs: discovers operations and registers them in the local registry with forwarding handlers.

pub async fn from_call(
    connection: &CallConnection,
    config: FromCallConfig,
) -> Vec<HandlerRegistration>

The adapter:

1. Calls services/list on the remote node → gets the list of External operations
2. Calls services/schema for each → gets the input/output JSON Schemas and declared error_schemas (ADR-023)
3. For each discovered operation, constructs a HandlerRegistration bundle:
   • The spec mirrors the remote operation's name, namespace, type, schemas (input, output, and error_schemas — ADR-023), and access control
   • The handler sends call.requested through the CallConnection and awaits call.responded (or streams for subscriptions)
   • provenance: FromCall, composition_authority: None, scoped_env: None (leaves — ADR-022)
4. The caller registers these bundles in their local registry (into the connection's overlay — ADR-024)

from_call-registered operations are Internal by default (ADR-015) — they are composition material, not directly callable from the wire. The handler that composes them is External.

The FromCallConfig includes:

• The credential source for the outbound connection (ADR-014) — TLS identity, auth token, or capability-provided credentials
• An optional namespace prefix (to avoid collisions when importing from multiple remote nodes)
• An optional operation filter (to import only specific operations)

4. to_openapi and to_mcp adapters export local operations

The reverse direction — exposing local operations to external systems:

• to_openapi: generates an OpenAPI spec from the local registry's External operations. External systems (HTTP clients, API gateways) can discover and call alknet operations through a standard HTTP interface.
• to_mcp: exposes local operations as MCP tools. MCP clients (editors, AI tools) can discover and call alknet operations through the MCP protocol.

These adapters are outbound bridges — they translate the call protocol's operation model into external protocol formats. They do not modify the local registry; they project it.

5. The adapter contract trait

The adapter patterns share a common shape: they produce HandlerRegistration bundles that register in the local registry. The trait:

#[async_trait]
pub trait OperationAdapter: Send + Sync {
    async fn import(&self) -> Vec<HandlerRegistration>;
}

The return type is Vec<HandlerRegistration> (not (OperationSpec, Handler) pairs) — ADR-022 changed the registration API to the bundle shape, and adapters must produce bundles. Adapter convenience methods construct bundles with composition_authority: None and scoped_env: None for the leaf ops they produce.

The trait is async because from_call requires async discovery (services/list + services/schema over a QUIC connection). A synchronous trait cannot accommodate from_call without a separate async pre-step that populates a cache. The sync adapters (from_openapi, from_mcp reading a static spec) trivially satisfy an async trait — their import() bodies contain no .await points. The async/sync question is decided: the trait is async.

Implementations:

• FromOpenAPI — imports from an OpenAPI spec (HTTP-backed handlers)
• FromMCP — imports from an MCP server (MCP-backed handlers)
• FromCall — imports from a remote call protocol endpoint (call-protocol-backed handlers)
• FromJsonSchema — imports from a JSON Schema definition (schema-only, no handler — used for validation or client generation)

The to_* adapters are outbound projections, not OperationAdapter implementations — they consume the registry, they don't produce entries for it.

The specific trait signatures (error types, configuration parameters) are two-way doors for implementation. The one-way doors are the architectural commitments: adapters produce HandlerRegistration bundles (ADR-022), the trait is async (required by from_call), and adapters live in alknet-call.

6. Cross-node call tree and abort cascade

When a from_call handler sends call.requested to a remote node, the call participates in the local call tree via parent_request_id. If the parent is aborted, the cascade (ADR-016) reaches the from_call handler, which sends call.aborted to the remote node. The remote node cascades to its own descendants. The abort crosses the node boundary transparently.

Head node                                    Worker node
  r1: /dispatch/run_training
    r1-a: worker/exec (from_call handler)
      → call.requested { id: r1-a } ────────→ receives, dispatches to exec
                                                r1-a-1: exec spawns child
  user aborts r1
    cascade to r1-a
      from_call handler sends:
        call.aborted { id: r1-a } ───────────→ receives, cascades to r1-a-1
                                                  aborts exec and children

7. Credential sources for connections

The CallClient needs credentials to authenticate to the remote node. These come from capabilities (ADR-014), not environment variables. The credential types:

• TLS identity: the local node's Ed25519 key (RFC 7250 raw key) or X.509 cert, derived from the vault at startup
• Auth token: an opaque token for call-protocol-level authentication, decrypted from the vault or derived from a shared secret
• Remote identity verification: the expected fingerprint or cert of the remote node, stored as a capability (not an env var or config file)

The from_call adapter receives these credentials at registration time, same as from_openapi receives HTTP credentials.

Consequences

Positive:

• Cross-node composition works the same as local composition. A handler calls env.invoke("worker", "exec", ...) and doesn't know (or care) whether worker/exec is a local operation or a from_call-imported remote operation. The composition is transparent.
• The head/worker pattern (dispatch, runners) is a connection topology, not a protocol feature. Workers can connect to heads (runner pattern) or heads can connect to workers (dispatch pattern) — the protocol handles both.
• from_call is the same pattern as from_openapi and from_mcp: discover, register, forward. The adapter contract is unified.
• to_openapi and to_mcp enable interop with non-alknet systems without those systems needing to speak EventEnvelope.
• The abort cascade (ADR-016) crosses node boundaries transparently. No consumer needs to implement cross-node abort propagation.
• The NAPI and Python adapters can use CallClient directly to call remote operations — they don't need a separate client implementation.

Negative:

• CallClient has its own operation registry (for dispatching incoming calls from the remote side). This is a second registry instance, not the global one — it needs to be populated with the operations this node wants to expose to that specific remote peer. The specific mechanism (sharing the global registry, a peer-scoped subset, or a separate registry) is a two-way door.
• from_call-registered operations have a latency cost: each invocation sends a call.requested over QUIC and awaits a call.responded. This is inherent to remote calls and not specific to the adapter pattern. Caching or batching strategies are consumer concerns.
• The to_* adapters need to translate the call protocol's operation model (JSON Schema, EventEnvelope, subscribe/stream) into external formats (OpenAPI paths, MCP tools). Some semantics don't map cleanly (e.g., subscriptions in OpenAPI, bidirectional calls in MCP). The adapters handle these with best-effort mappings and document the gaps.
• Published to_* specs are compatibility contracts. The "best-effort" mapping label is internal framing. Once a generated spec is published and external clients build against it, the mapping semantics (e.g., subscriptions → SSE long-poll) become a de facto contract. Changing the mapping later breaks every client. to_* mapping choices are two-way before first publication but one-way after. Version the generated specs (e.g., OpenAPI spec version tied to the registry's External operation set version) and emit a spec version marker so consumers can detect mapping changes. This is the "published artifact is a contract" blind spot in ADR-009's framework: it classifies doors by reversal cost in the codebase, not by compatibility cost for external consumers.
• Sharing the global registry with a CallClient exposes local capabilities to the remote peer. Each HandlerRegistration carries Capabilities with secret material. If the CallClient shares the global registry, a remote peer calling an External operation triggers dispatch that populates OperationContext.capabilities from the local registration bundle — meaning the local node's API keys and signing keys are used for the remote peer's call. A peer-scoped subset must filter by capability remote-safety (is this operation's capability safe to expose to this peer?), not just operation name. The registry-mechanism choice (share global vs subset vs separate) is two-way mechanically but has a security dimension post-ADR-022: the "share global" option is a capability-exposure decision, not just a dispatch decision.
• The CallConnection abstraction adds a layer between the handler and the raw QUIC stream. This is necessary for the from_call handler to be transparent — it shouldn't know about QUIC streams, only about call/request semantics.

Assumptions

1. The connection is symmetric after establishment. Both sides can send and receive call.requested. If a future use case requires one-directional connections (e.g., a fire-and-forget notification where the receiver can't call back), the model needs extension. The assumption is that bidirectional is the correct default.

2. services/list and services/schema are the discovery mechanism for from_call. The remote node exposes its External operations through these built-in operations. If a remote node doesn't support service discovery (e.g., a minimal worker that only accepts specific calls), from_call needs an alternative discovery mechanism (static config, manual spec). The assumption is that nodes participating in cross-node composition support service discovery.

3. The from_call handler is transparent to composition. A handler that calls env.invoke("worker", "exec", ...) doesn't know it's a remote call. If the remote node is unreachable or the connection drops, the handler gets a call.error (same as a local handler error). The assumption is that remote call failures are handled the same as local handler failures.

4. from_call-registered operations mirror the remote spec. The imported OperationSpec has the same name, namespace, type, schemas (input, output, and error_schemas per ADR-023), and access control as the remote operation. If the remote operation changes (new schema, renamed), the imported spec is stale until re-import. The assumption is that re-import happens on reconnection or is triggered explicitly. Hot-swapping imported specs is a two-way door.

5. The to_* adapters are projections, not live bridges. to_openapi generates a spec; it doesn't proxy HTTP requests. An external HTTP client calling the generated OpenAPI endpoints needs an HTTP handler (alknet-http) that translates HTTP requests into call protocol operations. The assumption is that to_* generates specs/tools, and a separate HTTP/MCP handler bridges the actual traffic.

References

• ADR-005: irpc as call protocol foundation
• ADR-012: Call protocol stream model (bidirectional streams)
• ADR-013: Rust as canonical implementation language (adapter traits in Rust)
• ADR-014: Secret material flow (credential sources, not static tokens)
• ADR-015: Privilege model (adapter ops are Internal by default)
• ADR-016: Abort cascade (cross-node abort propagation)
• ADR-028: Peer-Scoped Registry Filtering for CallClient Inbound Dispatch (resolves the §1 Consequences security dimension flagged as a two-way door)
• OQ-15: Call protocol client and adapter contract (resolved by this ADR)
• OQ-25..28: Two-way-door remainders from the call-completion gap analysis (DC-1 shape, DC-4 error type, DC-2 re-import trigger, DC-3 namespace collision — see open-questions.md)
• call-protocol.md
• operation-registry.md
• client-and-adapters.md — the spec that operationally fills the gap this ADR left to implementation
• docs/research/alknet-call-completion/gap-analysis.md — DC-1..4, the decisions that needed resolution before implementation
• TypeScript @alkdev/operations — from_openapi, from_mcp, buildEnv prior art
• POC at /workspace/@alkdev/dispatch — head/worker dispatch over SSH+axum

Amendments (2026-06-26)

This ADR left four decisions as two-way doors (§1 Consequences flagged DC-1's security dimension; §5 noted trait signatures are two-way doors; Assumption 4 noted re-import hot-swap is a two-way door; §3 mentioned the namespace prefix). The call-completion gap analysis (docs/research/alknet-call-completion/gap-analysis.md DC-1..4) resolved them. The resolutions:

DC-1 — CallClient registry scope: resolved by ADR-028, superseded by ADR-029

The §1 Consequences security dimension was originally resolved by ADR-028 (default-deny remote_safe: bool + trusted_peer opt-in). ADR-028 is now superseded by ADR-029 (2026-06-27): the flat-namespace single-peer model ADR-028 built on cannot express the head→N-workers pattern, and the remote_safe/trusted_peer gate duplicates the existing AccessControl/Identity machinery while reintroducing the blanket-bypass anti-pattern ADR-015 killed. ADR-029 replaces the flat overlay with peer-keyed overlays + PeerRef routing, and retires remote_safe/ trusted_peer in favor of AccessControl::check(peer_identity) — the existing authorization path that was already in the dispatch path. The peer- scoping question this section flagged is now answered structurally (peer-keyed overlays), not by a parallel boolean gate.

DC-4 — OperationAdapter trait error type: resolved

§5 showed async fn import(&self) -> Vec<HandlerRegistration> with no error type. The trait returns Result<Vec<HandlerRegistration>, AdapterError> where AdapterError is a crate-level enum. The presence of the error type is recorded in client-and-adapters.md; the exact variants are the two-way-door remainder, tracked as OQ-26.

DC-2 — from_call re-import on reconnection: default set

Assumption 4 noted re-import "happens on reconnection or is triggered explicitly." The v1 default is auto-re-import on connection establishment. The overlay is per-connection (Layer 2, ADR-024), so re-import is naturally scoped; a stale overlay dies with the connection. Explicit re-import via a future CallConnection::refresh() is additive. Two-way door; recorded in client-and-adapters.md; tracked as OQ-27.

DC-3 — from_call namespace collision: default set

§3's FromCallConfig namespace prefix is optional, default no prefix, collision = error. A node importing from two remotes that both expose the same unprefixed op name should fail loudly. The operator adds prefixes when importing from multiple sources. Two-way door; recorded in client-and-adapters.md; tracked as OQ-28.

Operational spec

The gap this ADR left to implementation — the CallClient API, the from_call/from_jsonschema flows, the trait signature, the adapter location map, the no-env-vars invariant, and the exchange-of-operations pattern — is specified in client-and-adapters.md. That document is the operational complement to this ADR; this ADR remains the architectural authority.