driftshield

Find where an AI agent first went wrong.

DriftShield reconstructs a failed agent run as a decision graph, shows where reasoning first drifted and matches the failure against known community signatures. It reads Claude Code, Codex CLI, Claude Desktop, Codex Desktop, OpenClaw, CrewAI and LangChain transcripts.

pip install driftshield-sdk
driftshield analyze session.jsonl

The distribution is called driftshield-sdk on PyPI because the name driftshield was already too close to an unrelated project. The import package and the command are still driftshield.

The library surface is two calls:

from driftshield import analyse_run, submit

run = analyse_run(open("session.jsonl", "rb").read(), source="session.jsonl")
print(run.qualification_state, run.findings)

receipt = submit(run)  # community lane: redacted, no key needed

analyse_run works offline and touches nothing outside the process. submit is the only call that sends anything off the machine, and it only ever sends run.redacted_transcript.

Documentation, the self hosted web UI and the full CLI reference live in the repository: https://github.com/tborys/driftshield

Licensed under AGPL-3.0-or-later.