shadow-diff 3.2.0

Behavior contracts for AI agents — tested in your PR, enforced at runtime.

shadow-diff

Find the exact change that broke your AI agent.

Shadow is a CI-native regression-forensics tool for LLM agents. One command on the PR — shadow diagnose-pr — answers:

1. Did agent behavior change?
2. How many traces are affected?
3. Which exact prompt / model / tool / config change caused it?
4. With what confidence (ATE + bootstrap CI + E-value when run with --backend live)?
5. What fix should verify-fix confirm before merge?

The PyPI distribution is shadow-diff. The Python import path is shadow. The CLI is shadow.

Install

pip install shadow-diff

Requires Python 3.11+. Pre-built wheels ship for Linux x86_64, macOS arm64, and Windows x86_64; other platforms build from source (Rust required).

Optional extras:

pip install 'shadow-diff[anthropic]'   # if your agent uses Claude
pip install 'shadow-diff[openai]'      # if your agent uses GPT
pip install 'shadow-diff[embeddings]'  # paraphrase-robust semantic diff
pip install 'shadow-diff[all]'         # everything

60-second tour

shadow demo                  # nine-axis diff on bundled fixtures, no API key
shadow quickstart            # writable copy of a runnable scenario

Then run diff against the writable scenario:

cd shadow-quickstart
shadow diff fixtures/baseline.agentlog fixtures/candidate.agentlog

For the full diagnose-pr flow against your own agent, see docs/features/causal-pr-diagnosis.md and the runnable refund-causal-diagnosis demo.

Record your own agent

from shadow.sdk import Session

with Session(output_path="trace.agentlog"):
    # Your existing Anthropic / OpenAI code, unchanged.
    client.messages.create(model="claude-sonnet-4-6", messages=[...])

Shadow auto-instruments the Anthropic and OpenAI SDKs and writes content-addressed .agentlog files. Secrets are redacted by default. Or skip the code change entirely:

shadow record -o trace.agentlog -- python your_agent.py

Daily workflow — Shadow as pytest for agent behavior

shadow inspect trace.agentlog                  # debug a single trace
shadow scan baseline_traces/                   # block secret leaks
shadow baseline create baseline_traces/        # pin the gold standard
shadow gate-pr ...                             # gate every PR

Full docs

The canonical README, the .agentlog spec, runnable examples, and the comparison against adjacent agent-eval and runtime-governance tools all live at https://github.com/manav8498/Shadow.

License

Apache-2.0. See LICENSE-APACHE in this distribution. The .agentlog spec is independently published under Apache-2.0.