API reference
Agentdock construction, streaming, checkpoint reads, and run options.
new Agentdock(graph, options?)
Adapts a compiled LangGraph graph to the Agentdock event contract. Input and context types are inferred from the graph where LangGraph exposes them.
import { Agentdock } from "@agentdock-ai/agentdock";
const runtime = new Agentdock(graph);| Option | Description |
|---|---|
recursionLimit | Optional positive safe integer override. Otherwise LangGraph resolves its configured limit. |
interruptFormat | Agentdock.OPAQUE (default) or Agentdock.HITL for native LangChain approval presentation. The equivalent strings remain supported. |
validateResume | Optional application preflight hook (resumeValue, pendingInterrupts, currentContext). May be asynchronous; runs before graph invocation or SSE headers. |
onError | Optional server diagnostic callback (error, { threadId, runId, stage }). Stage is graph, mapper, or checkpoint; client errors remain sanitized. |
Streaming methods
| Method | Description |
|---|---|
stream(run) | Return the transport-free async event iterable. |
pipe(response, run) | Write SSE events to a Node-compatible response with backpressure and disconnect cancellation. |
toResponse(run) | Return a promise for a Web Response with a cancelable event stream. |
A run has an authorized threadId and one of these forms:
- Start:
{ threadId, input }. - Resume a dynamic interrupt:
{ threadId, resume }. - Continue a static breakpoint:
{ threadId, continue: true }.
All forms accept optional context, signal, and config. config forwards
LangGraph options such as callbacks, tags, metadata, store, and additional
configurable values. Agentdock overwrites configurable.thread_id with the
authorized threadId. Context, signal, stream modes, and recursion are controlled
by serving; configure a recursion override in the constructor.
Resume values are forwarded to LangGraph unchanged. For multiple pending interrupts, use a map keyed by their native IDs. See interrupt and resume.
Checkpoint reads
| Method | Description |
|---|---|
getMessages(threadId, { channel?, config? }?) | Read a checkpoint message channel; defaults to messages. Returns unknown[] or null when there is no checkpoint. Normalize provider messages in your application. |
getResumeState(threadId, config?) | Project pending native interrupts or static breakpoints into a reducer seed, or return null when no work is pending. Invalid projections throw. |
The resume seed has runId: null and preserves pending native IDs and complete
payloads. Load conversation history separately. Authorize checkpoint reads as
well as writes and continuations.
No Agentdock graph state schema is required. Agentdock reads native checkpoint
control state and never writes event bookkeeping into graph state. Each invocation
gets a fresh runId with sequence starting at 1.
Contracts and validation
Import event types, reducers, and validation utilities from
@agentdock-ai/contracts. The core package exports Agentdock, serving types,
and validateToolApprovalResume. This helper checks the LangChain HITL response
envelope and action count; application and native middleware policies still own
permissions.
Ownership boundaries
Your application owns routes, request parsing, authentication, authorization, thread identity, context, side effects, provider setup, and saver lifecycle. LangChain and LangGraph own agent execution, tools, models, retries, checkpoints, interrupts, and resume. Agentdock owns event mapping and serving mechanics.