Agent Harness Protocol for LangChain + LangGraph

agent-harness-protocol-langgraph implements the Python Agent Harness Protocol (AHP) 0.2.3 port with a LangChain BaseChatModel running inside a LangGraph StateGraph. It is a separate distribution from the core agent-harness-protocol package, which contains no LangChain or LangGraph dependency.

Each AHP session compiles a START -> model -> END graph with its own checkpoint thread. Sequential tasks in one session share the conversation; different sessions are isolated. The adapter reports what this graph can actually guarantee and rejects the rest at admission instead of simulating it.

Installation

python -m pip install agent-harness-protocol-langgraph

Requirements: Python 3.11+, langchain-core>=1.2.8,<2, langgraph>=1.1,<2.

Minimal example

from agent_harness_protocol import Completed, SessionSpec, TaskRequest, TextInput
from agent_harness_protocol_langgraph import LangGraphHarness
from langchain_core.language_models import BaseChatModel


async def run(model: BaseChatModel, model_id: str) -> None:
    async with LangGraphHarness(model, model_id=model_id) as harness:
        session = await harness.create_session(SessionSpec(instructions="Answer briefly"))
        task = await session.start_task(TaskRequest(TextInput("Summarize the report")))

        async for event in task.events():
            print(event)

        outcome = await task.await_outcome()
        if isinstance(outcome, Completed):
            print(outcome.output)

SessionSpec.model is admitted only when it equals the model_id given at construction. Without a model_id, any explicit model request is rejected.

Support matrix

Only TextInput is accepted. Semantic events are TaskStarted, MessageCompleted, UsageChanged, and one terminal event. MessageDelta is not produced because the graph is not streamed yet.

Disposition declarations

Whether stored context is ephemeral and whether inputs stay hidden depend on facts the adapter cannot observe, so they are declared at construction and reflected identically in support, validate(), admission, and session.disposition.

from agent_harness_protocol import ContextRetentionDisposition, UserHistoryVisibility

harness = LangGraphHarness(
    model,
    checkpointer=my_saver,
    context_retention=ContextRetentionDisposition.MATERIALIZED,
    history_visibility=UserHistoryVisibility.VISIBLE,
)

Visibility defaults to UNKNOWN because provider request logs or LangSmith tracing can expose inputs outside the application. Declare HIDDEN only when that is known to be false for your deployment. A type name is not evidence: InMemorySaver(factory=...) can write to disk.

Cancellation semantics

Cancelling a LangChain coroutine does not by itself prove that a remote provider stopped working, so the outcome depends on cancellation_semantics:

Choose the second value only when coroutine termination really ends the native work, for example with a local in-process model.

The adapter delivers one cancellation request per task. While a graph run that outlived its settled task is still running, start_task() raises SessionBlockedError, so two runs never write to the same checkpoint thread. The session becomes usable again when that run ends.

A cancelled turn's input remains in the checkpoint and is part of the next task's context, without a matching model reply, because LangGraph records the input before the model node runs.

Cleanup

session.release() and harness.aclose() settle active tasks within the published CleanupBudget (default: 1 s per task, 3 s total). Cancelling the caller does not interrupt cleanup.

Releasing a session deletes its checkpoint thread within the remaining total budget. If deletion fails or times out, CheckpointCleanupError is raised; task outcomes are unchanged and the handle is still closed. aclose() settles every task before raising one aggregated CheckpointCleanupError. When a graph run outlives cleanup, its thread is deleted again after the run ends; that background deletion has no failure report.

Usage accounting

AIMessage.usage_metadata becomes AgentUsage. Missing metadata or a missing field is None (unknown), never zero. Session usage is committed once per task outcome and a field stays unknown once any task left it unknown.

Failures are reported as Failed(FailureKind.UNKNOWN) with the original exception as cause; the adapter does not classify errors from message text.

Verification

The tests replace only the chat model with a controllable BaseChatModel; the harness, compiled graph, and checkpointer are real, and every observation is read from what the model received.

Bound shared suites:

• RuntimeProfileConformanceTests
• RequirementsConformanceTests, with four disposition profiles
• CleanupBudgetConformanceTests, with confirmed, unconfirmed, and cancellation-ignoring models

Suites for interactions, approval scope, persistence, workspace, and execution constraints are not bound because this graph has no such route; the requirements suite verifies that those requirements are rejected. The full list, including suites not bound yet, is in the conformance guide.

From implementations/langgraph in a repository checkout:

python -m pip install -e "../..[test]" -e ".[test]"
python -m pytest

Design decisions are recorded in docs/langgraph-adapter-design.md.