JanusLangChainAgent (adapter depth 3) cannot be constructed under LangChain 1.x

[ish-codes-magic]

Summary

JanusLangChainAgent (depth 3 of the LangChain adapter) cannot be constructed
under LangChain 1.x. It imports AgentExecutor and create_tool_calling_agent
from langchain.agents unconditionally, and LangChain 1.0 removed both.

The langchain extra declares langchain>=0.3, which permits 1.x, and
uv sync --extra all --extra dev currently resolves to langchain 1.2.10. So
the adapter's turnkey entry point is broken in the project's own canonical
environment, not just in some unusual user setup.

Depths 1 and 2 are unaffected — see Scope below.

Environment

• janus-guard @ 2512be1e6 (main)
• langchain 1.2.10, langchain-core 1.2.16 (as resolved by uv sync --extra all --extra dev)
• Python 3.13, Windows 11

Steps to reproduce

from janus.adapters.langchain import JanusLangChainAgent

JanusLangChainAgent(model="openai/gpt-4o", tools=[], policy=None)

Actual behaviour

ImportError: cannot import name 'AgentExecutor' from 'langchain.agents'
(.../site-packages/langchain/agents/__init__.py)

The failure is at construction time, before any policy or tool wiring happens.

langchain.agents in 1.2.10 exports only:

['AgentState', 'create_agent', 'factory', 'middleware', 'structured_output']

Expected behaviour

Either the adapter works against the LangChain versions its own extra permits,
or the extra is narrowed so that a resolvable install is a working one.

Root cause

janus/adapters/langchain.py:261-264, inside JanusLangChainAgent.__init__:

from langchain.agents import (  # type: ignore[attr-defined]
    AgentExecutor,
    create_tool_calling_agent,
)

LangChain 1.0 replaced the AgentExecutor + create_tool_calling_agent pair
with a single LangGraph-backed create_agent. The import is unconditional and
runs on every construction, so there is no version in which depth 3 degrades
gracefully — it either imports or raises.

pyproject.toml:44-48 declares:

langchain = [
    "langchain>=0.3",
    "langchain-core>=0.3",
    "langchain-openai>=0.2",
]

The lower bound is correct for the 0.3 API; there is no upper bound excluding
the 1.x API that removed it.

Scope

Affected

• Depth 3, JanusLangChainAgent — cannot be constructed at all.
• The docstring example at janus/adapters/langchain.py:89 and the depth-3
  description at lines 20 and 205, which all reference AgentExecutor.

Not affected

• Depth 1, secure_langchain_tools() — imports only
  langchain_core.tools.StructuredTool (line 43).
• Depth 2, wrap_langchain_tools() — same, plus BaseTool.

So policy enforcement itself is fine on LangChain 1.x. It is only the turnkey
agent wrapper that breaks. No test covers depth 3 construction, which is why CI
is green.

Suggested fix

Option A — version-tolerant branch (recommended). Detect which generation
is installed and build the corresponding loop:

try:
    from langchain.agents import create_agent   # LangChain >= 1.0
except ImportError:
    create_agent = None

if create_agent is not None:
    self._executor = create_agent(model=llm, tools=self.lc_tools,
                                  system_prompt=system_prompt)
else:
    from langchain.agents import AgentExecutor, create_tool_calling_agent
    ...

The two generations also differ at invoke time and need matching branches:

	LangChain 0.3	LangChain 1.x
invoke	{"input": task}	{"messages": [{"role": "user", "content": task}]}
result	result["output"]	result["messages"][-1].content
tool trace	result["intermediate_steps"]	tool_calls on the message list

This keeps the declared langchain>=0.3 honest.

Option B — cap the extra at langchain>=0.3,<1.0. Smaller change, but it
locks users out of LangChain 1.x for a limitation that only affects depth 3, and
depths 1 and 2 work fine there.

Option C — deprecate depth 3 in favour of depth 1 plus a user-built loop,
and document the create_agent pattern.

Whichever is chosen, the chat_model= escape hatch does not currently help:
the AgentExecutor import happens regardless of whether a pre-built model was
passed.

Workaround

demos/demo_2 (PR #6) hits this and works around it by using
secure_langchain_tools() (depth 1) and building the loop itself with a
create_agent / AgentExecutor branch. See
demos/demo_2/agent.py::IssueExplainer for a working implementation of
Option A, including the invoke-shape differences above.

Note for whoever picks this up

The fix touches janus/adapters/, so per CLAUDE.md it needs the
enforcement-review checklist before commit. Worth adding a construction test
for all three depths against whichever LangChain generation CI installs —
the absence of one is what let this ship.