LangChain Human In The Loop (CodeVF)

Minimal LangChain tool integration for routing review, debugging, and verification requests to CodeVF using the official Python SDK.

Quick Start

1. Install

pip install langchain-human-in-the-loop

Dependencies (codevf-sdk, langchain-core, pydantic) are installed automatically.

2. Set your API key

export CODEVF_API_KEY="cvf_live_..."

3. Run a direct review (no agent needed)

from langchain_human_in_the_loop import HumanInTheLoop

hitl = HumanInTheLoop(
    api_key="CODEVF_API_KEY",
    project_id=123,
    max_credits=50,
    mode="fast",
    tag_id=1,  # Optional: Engineer=1.7x, Vibe Coder=1.5x, General Purpose=1.0x
    timeout=300,
)
result = hitl.invoke(
    "Review this function for security issues and suggest fixes.",
    attachments=[
        {
            "file_name": "app.py",
            "mime_type": "text/x-python",
            "content": "def login(user, pwd): ...",
        }
    ],
)

print(result)

Install page:

https://pypi.org/project/langchain-human-in-the-loop/

Install From Source

pip install -e .

Usage

Attachment Format

Each attachment accepts either content (plain text) or base64 (for binary files). The SDK enforces limits and supported file types.

Engineer Expertise (tag_id)

Use an optional tag_id to select expertise pricing from GET /tags:

• Engineer -> 1.7x
• Vibe Coder -> 1.5x
• General Purpose -> 1.0x (default if omitted)

Invalid tag_id values return 400 from the API.

You can set tag_id at HumanInTheLoop(...) construction time or per call:

result = hitl.invoke("Review this.", tag_id=2)

Example: Multiple Files

from langchain_human_in_the_loop import HumanInTheLoop

hitl = HumanInTheLoop(project_id=123, max_credits=50)

result = hitl.invoke(
    "Find concurrency bugs and propose fixes.",
    attachments=[
        {
            "file_name": "worker.py",
            "mime_type": "text/x-python",
            "content": "def run(queue): ...",
        },
        {
            "file_name": "README.md",
            "mime_type": "text/plain",
            "content": "# Worker\n\nHow it is supposed to work...",
        },
    ],
)

print(result)

Example: Binary Attachment (Base64)

import base64
from langchain_human_in_the_loop import HumanInTheLoop

with open("diagram.png", "rb") as f:
    encoded = base64.b64encode(f.read()).decode("ascii")

hitl = HumanInTheLoop(project_id=123, max_credits=50)
result = hitl.invoke(
    "Review the architecture diagram for missing components.",
    attachments=[
        {
            "file_name": "diagram.png",
            "mime_type": "image/png",
            "base64": encoded,
        }
    ],
)

print(result)

LangChain Tool Helper

You can expose CodeVF as a LangChain structured tool via:

from langchain_human_in_the_loop import HumanInTheLoop, HumanInTheLoopInput

hitl = HumanInTheLoop(project_id=123)
codevf_tool = hitl.as_langchain_tool()

# Optional: use the tool input schema for structured calls
schema = HumanInTheLoopInput.schema()

See examples/human_in_the_loop.py for a complete runnable snippet.
If you want to test CodeVF with LangChain, use examples/codevf_tool_agent.py.

Research / University Samples

• examples/reproducibility_audit.py reviews an ML experiment for reproducibility gaps.
• examples/data_privacy_review.py checks a student-data pipeline for privacy and policy risks.
• examples/paper_review.py reviews methods/results for missing baselines and reporting issues.
• examples/irb_protocol_review.py reviews a study protocol, consent, and survey for IRB gaps.
• examples/bias_fairness_audit.py audits an admissions model for sampling bias and fairness risk.
• examples/replication_package_review.py checks a replication package for release readiness.

Testing

pip install -e .[dev]
pytest