copass-langchain

LangChain tool adapters for Copass. Python mirror of @copass/langchain. Pulls discover / interpret / search into any LangChain agent, and (optionally) wires them into LangGraph's ReAct prebuilt.

Install

pip install copass-langchain
# or, with the create_copass_agent convenience wrapper:
pip install copass-langchain[agent]

Depends on copass-core + copass-config + langchain-core. langgraph is optional — only needed for create_copass_agent.

Drop-in tools

from copass_core import ApiKeyAuth, CopassClient
from copass_langchain import copass_tools

client = CopassClient(auth=ApiKeyAuth(key="olk_..."))

tools = copass_tools(client=client, sandbox_id="sb_...")
# Pass `tools.all()` (a list of three StructuredTool instances) to your
# agent framework, or pull them individually:
# tools.discover, tools.interpret, tools.search

Tool descriptions and parameter descriptions come from copass-config — identical to the strings used across every other Copass adapter.

Window-aware retrieval (when a Context Window exists)

copass-core v0.1 does not yet ship a ContextWindow class (deferred to v0.2). Until then, copass_tools and CopassWindowCallback accept any object satisfying ContextWindowLike:

class ContextWindowLike(Protocol):
    def get_turns(self) -> list[ChatMessage]: ...
    def add_turn(self, turn: ChatMessage) -> Awaitable[None]: ...

When copass-core ships ContextWindow, it will satisfy this protocol and window-aware retrieval will light up without any consumer changes.

Full agent in one call

from copass_langchain import create_copass_agent
from langchain_anthropic import ChatAnthropic

agent = create_copass_agent(
    client=client,
    sandbox_id="sb_...",
    llm=ChatAnthropic(model="claude-opus-4-7"),
    # window=my_window,  # optional — enables window-aware retrieval + auto-mirroring
)

result = await agent.ainvoke({
    "messages": [("user", "why is checkout flaky?")]
})

Status

• copass_tools — shipped.
• CopassWindowCallback — shipped (generic on ContextWindowLike).
• create_copass_agent — shipped (lazy-imports langgraph).
• First-class ContextWindow integration — lands when copass-core v0.2 ships the primitive.

Conversation metadata

CopassWindowCallback delegates to ContextWindow.add_turn, so any speaker / participants configured on the underlying window flow through automatically — no LangChain-specific wiring needed:

window = await client.context_window.create(
    sandbox_id=sandbox_id,
    participants=["User", "agent:support-bot"],   # roster set once
)
callback = CopassWindowCallback(window=window)

When the LangChain callback emits a turn, the window's constructor-time roster + role-derived speaker ride on the envelope. Set ChatMessage.name on a turn for richer per-turn speaker provenance — see copass-core for the full envelope surface.

License

MIT.