JEP-Agent SDK 2.1 — JEP Core 0.7

Record signed statements about agent calls and inspect their evidence. The current source implements JEP Core 0.7; historical 2.0.x releases implement the earlier JEP-04/JAC-01 format. See migration before upgrading.

JEP is an individual IETF Internet-Draft, not an IETF-endorsed standard. Valid signatures do not establish factual truth, authorization, legality, successful external execution, or payment readiness.

Install

pip install jep-agent-sdk
# Optional framework dependencies:
pip install 'jep-agent-sdk[langchain,openai]'

For source development, clone this repository and use pip install -e '.[dev]'. Imports use jep_agent; the command is jep-agent, independent of jep-sdk-py and jep-cli.

Signed trace

from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PrivateKey
from jep_agent import AuditChain, record

key = Ed25519PrivateKey.generate()
chain = AuditChain("agent:example", private_key=key)

@record(issuer=chain.issuer, chain=chain)
def task(value):
    return value * 2

assert task(21) == 42
assert chain.verify_chain(key.public_key())
chain.save("events.jsonl")

Recording without a key is allowed for local traces, but produces unsigned, unverified records. Use a securely persisted key in a real deployment; this example generates a temporary key.

Core and companion boundaries

The in-memory JEPVerifier acceptance store is for a single process. It is not a durable distributed acceptance service. An independently trusted public key must be supplied; kid alone does not prove actor identity. Reference resolution, actor binding, domain policy and external effects remain unchecked unless provided by a separate profile or application.

A function failure/cancellation produces a result statement, not a Core Termination event. Async tracing records completion only after the call finishes.

Inspect and verify

jep-agent verify events.jsonl --public-key public-key.pem
jep-agent export events.jsonl --output report.html
jep-agent web --port 8080

The CLI checks Core signatures and any local audit-chain links. It does not resolve arbitrary external references. The viewer and HTML export show recorded relationships and Signed (unverified) status; visual links do not establish causality or legal responsibility.

Framework adapters are experimental: the OpenAI adapter targets synchronous Chat Completions, and the LangChain auto patch targets historical AgentExecutor APIs. New signed integrations use the callable recording path. The separate Agents SDK middleware is a retired unsigned observation experiment; it is not the maintained Core integration path.

Development

pip install '.[dev,langchain,openai]' build
make lint
python -m pytest -q
python validate.py
python -m build --wheel
python scripts/check_coexistence.py dist/*.whl

CI uses fixed Core 0.7 J/D/T/V vectors and a commit-pinned independent Core validator. It also installs the built wheel beside the Python SDK and CLI in both orders and checks independent uninstalls.

• Architecture and research-helper boundaries
• API
• Migration and historical compatibility
• Implementation limits
• Canonical JEP Core repository

BSD-3-Clause. Author: Yuqiang Wang, HJS Foundation, signal@humanjudgment.org.