Shadow

Shadow

[](https://github.com/manav8498/Shadow/actions/workflows/ci.yml) [](#license) [](SPEC.md)

[](CHANGELOG.md)

[](rust-toolchain.toml) [](python/pyproject.toml)

Find the exact change that broke your AI agent.

A teammate edits a system prompt or swaps a model. Code review looks fine. Unit tests pass. You merge. A week later a customer reports the refund agent stopped asking for confirmation. Nobody saw it coming because the code looked harmless.

Shadow catches that bug class on the PR. shadow diagnose-pr answers, in one PR comment:

1. Did agent behavior change? Yes / no, with severity.
2. How many traces are affected? Out of N representative production-like cases.
3. Which exact candidate change caused it? Prompt edit, model swap, tool-schema rename, config delta — named with a file:line citation.
4. With what confidence? ATE + 95% bootstrap CI + E-value when run with --backend live; deterministic delta-attribution otherwise.
5. What fix should verify-fix confirm before merge?

shadow diagnose-pr \
  --traces           prod-traces/ \
  --candidate-traces candidate-traces/ \
  --baseline-config  baseline.yaml \
  --candidate-config candidate.yaml \
  --policy           shadow-policy.yaml \
  --pr-comment       comment.md

→ Try the [60-second runnable demo](examples/refund-causal-diagnosis/), or read [docs/features/causal-pr-diagnosis.md](docs/features/causal-pr-diagnosis.md) for the full flow.

See a real Shadow PR comment ↗  ·  launch video (84 s)  ·  MP4 / WebM

Why ordinary CI can't catch this

Unit tests assert what code returns, not how an agent behaves. The refund agent that drops "always confirm" is still parseable Python. The model swap from GPT-4o to a cheaper one still returns valid responses. The tool schema renamed order_id → id still typechecks both sides. Every guardrail your normal CI gives you is irrelevant — the regression is in the agent's decisions, not the code.

Shadow turns "how should this agent behave?" into a YAML contract — your CI tests every PR against it; your runtime enforces it against every tool call. Same rule, both places. When the contract breaks, Shadow names the line in the PR that broke it.

What Shadow does, in one screen

Given a baseline .agentlog and a candidate change, Shadow answers three questions on the PR:

1. What behavior changed? A nine-axis diff scores response meaning, tool calls, refusals, length, latency, cost, output format, and more — with a plain-English summary on top.
2. Why did it change? If the PR touched multiple things at once, regression attribution names the specific change that most likely explains each regression.
3. Is it safe to merge? A YAML policy declares rules the agent must follow (tool ordering, output shape, forbidden outputs). The same policy enforces at runtime.

The report lands in the PR comment. No dashboard, no separate login, no trace upload. Traces stay on your disk.

Install

Step 1. Check you have Python 3.11 or newer:

python3 --version    # 3.11.x or higher

Step 2. Install Shadow from PyPI:

pip install shadow-diff

That's it. shadow --help should now work, and shadow quickstart runs the demo. No clone, no Rust toolchain, no separate setup step — pre-built wheels ship for Linux x8664, macOS arm64 (Apple Silicon), and Windows x8664.

On other platforms (Intel Mac, ARM Linux, older glibc, Alpine, FreeBSD), pip falls back to the source distribution and builds the Rust core locally. Install Rust first:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
pip install shadow-diff

For byte-identical installs (CI repro, enterprise audit), see [docs/quickstart/locked-install.md](docs/quickstart/locked-install.md).

Optional extras

Shadow's core install is lean. Most users want one of:

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]'  # better paraphrase-robust diff

All optional extras (framework adapters, dashboard, signing, OTel, …)

| Extra | Pulls in | Use case | |---|---|---| | shadow-diff[anthropic] | anthropic | Live Anthropic client wrapper for shadow record | | shadow-diff[openai] | openai | Live OpenAI client wrapper for shadow record | | shadow-diff[embeddings] | sentence-transformers | Paraphrase-robust semantic-similarity axis. The default lexical TF-IDF path stays fast and dependency-free; the Embedder trait accepts any backend (this extra, ONNX runtime, HF Inference API, OpenAI embeddings, …). | | shadow-diff[otel] | opentelemetry-sdk | Export traces to any OTel-compatible backend | | shadow-diff[serve] | fastapi, uvicorn, websockets | shadow serve (live .shadow/ dashboard) and shadow dashboard --report report.json (single-report viewer) | | shadow-diff[mcp] | mcp | shadow mcp-serve Model Context Protocol server | | shadow-diff[multimodal] | Pillow, imagehash | Image / multimodal diff | | shadow-diff[sign] | sigstore | Sigstore keyless signing for ABOM certificates | | shadow-diff[langgraph] | langgraph | LangGraph agent adapter | | shadow-diff[crewai] | crewai | CrewAI agent adapter | | shadow-diff[ag2] | ag2 | AG2 / Autogen agent adapter | | shadow-diff[all] | everything above | One-shot install for trying the full feature surface | | shadow-diff[dev] | test/lint/type-check tooling | Contributing to Shadow itself |

Combine extras with comma-separated form:

pip install 'shadow-diff[anthropic,openai,embeddings,otel]'

Telemetry: off by default, opt-in only

Shadow ships an opt-in usage-telemetry hook. When enabled, each event includes: CLI command / event name, SDK version, Python version, OS, CPU architecture, and an anonymous install ID (random UUID4 generated once, stored at ~/.shadow/install_id, never tied to identity). No traces, no prompts, no user data. Telemetry is off by default — the first-run prompt asks before enabling. CI environments are detected and skipped automatically. Hard-disable with SHADOW_TELEMETRY=off in your shell. Source: [python/src/shadow/_telemetry.py](python/src/shadow/_telemetry.py).

Shadow never uploads .agentlog content, prompt text, response text, or tool arguments. The only fields collected when telemetry is enabled are listed in the module docstring.

Try it in sixty seconds

git clone https://github.com/manav8498/Shadow.git
cd Shadow/examples/refund-causal-diagnosis
./demo.sh

That runs shadow diagnose-pr end-to-end on a refund-agent scenario where the candidate config drops a "always confirm before refunding" instruction. Shadow names the prompt change as the dominant cause and tells you the fix:

## Shadow verdict: STOP

This PR violates a critical policy and must not merge as-is.
This PR changes agent behavior on **3** / **3** production-like traces.

### Dominant cause
`prompt.system` appears to be the main cause.
- Axis: `trajectory`
- ATE: `+0.60`
- 95% CI: `[0.60, 0.60]`
- E-value: `6.7`

### Why it matters
3 traces violate the `confirm-before-refund` policy rule.

### Suggested fix
Review the prompt change at `prompts/candidate.md` — restore the
instruction or constraint it removed.

For the underlying nine-axis behavior diff (without the causal layer), shadow demo runs the same fixtures through shadow diff and prints a severity table — useful when you want raw signal numbers:

shadow demo

That output looks like this (abbreviated):

signal             baseline  candidate    change     severity
─────────────────────────────────────────────────────────────
response meaning      1.000      0.435    -0.565     severe
tool calls            0.000      0.000    +0.000     none
refusals              0.000      0.333    +0.333     severe
response length      26.000     52.000   +26.000     minor
response time        98.000    412.000  +314.000     severe
output format         1.000      0.000    -1.000     severe

top divergences:
  #1  turn 0 — tool set changed: removed `search_files(query)`,
                                  added `search_files(limit,query)`
  #2  turn 2 — stop_reason changed: `end_turn` → `content_filter`

recommendations:
  error   Refusal rate is up severely. Check for stricter system instructions.
  error   Review tool-schema change at turn 0: call shape diverged.
  warning Review response text at turn 1: semantic content shifted.

Three things to read top-to-bottom: the severity column tells you which signals moved, the top-divergences list names the specific changes, and the recommendations tell you what to check first. A reviewer doesn't need to know any Shadow vocabulary — the recommendation lines speak plain English.

Use shadow quickstart when you want a writable copy of the demo files (agent.py, configs, fixtures) to edit and re-run:

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

Daily workflow — Shadow as pytest for agent behavior

The four commands you'll actually run all day:

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

shadow inspect opens a .agentlog file in your terminal — turn, role, tokens, latency, cost, redactions, first divergence. What pytest -v is to test runs, this is to recorded traces. Pass two paths to compare side-by-side; the first divergent turn is highlighted in red.

shadow scan walks your .agentlog files looking for OpenAI / Anthropic / GitHub / AWS keys, JWTs, PEM private keys, emails, phone numbers, and Luhn-valid credit cards. Exit code is non-zero on any hit, so it composes into CI before shadow gate-pr. Add company-specific patterns via --patterns ci/extra-secrets.txt.

shadow baseline is a frozen-trace workflow modeled on Insta and Jest snapshots:

shadow baseline create baseline_traces/    # pin the hash into shadow.yaml
shadow baseline verify                     # exit non-zero on drift
shadow baseline update --force             # re-pin after a deliberate regen
shadow baseline approve candidate_traces/ --force  # promote candidate to baseline

The hash lives in shadow.yaml so PRs that change the baseline show up in git diff as a single line. Reviewers see "the baseline hash changed in this PR" — a deliberate signal that the agent's expected behavior was reset.

shadow gate-pr prints a 1-screen failure summary that reads like a pytest assertion failure:

✗ Shadow gate-pr: STOP (exit 2) — 3/3 traces affected
  Cause: prompts/refund.md:17 appears to be the main cause.
  - removed: 4. Always confirm the refund amount before issuing the refund.
  trajectory: ATE=+0.60  95% CI=[0.50, 0.70]  E=6.7
  Policy: 3 new violation(s) of confirm-before-refund
  Fix: Restore the prompt instruction at prompts/refund.md:17
  Verify: shadow verify-fix --report .shadow/diagnose-pr/report.json

Pass -v / --verbose to also dump the full JSON report. The PR-comment markdown is always written via --pr-comment regardless of verbosity. Exit code: 0 ship, 1 hold/probe, 2 stop, 3 internal error.

Get a Shadow comment on every PR (≈ 10 minutes)

The end-to-end setup, from a fresh repo to seeing your first Shadow comment land on a real PR. Skip any step you've already done.

1. Install Shadow (one line):

pip install shadow-diff

2. Record baseline + candidate traces. Wrap the place where your agent runs in a Session block. Shadow auto-instruments OpenAI, Anthropic, LiteLLM, LangChain ChatOpenAI, and Vercel AI SDK (the ai package) calls and writes content-addressed .agentlog files:

# scripts/record_traces.py
from shadow.sdk import Session

with Session(output_path="baseline.agentlog"):
    run_my_agent()   # whatever your existing entry-point is

Run it once on baseline (main), once on the candidate (your PR branch), and commit both .agentlog directories. Traces stay on your disk and inside your repo — nothing is uploaded.

3. Drop in the diagnose-pr GitHub Action. Scaffold a workflow that runs shadow diagnose-pr on every PR — names the exact change that broke the agent, posts the verdict + suggested fix as a comment, and gates the merge with verdict-mapped exit codes (ship=0, hold/probe=1, stop=2):

shadow init --github-action
git add .github/workflows/shadow-diagnose-pr.yml && git commit -m "ci: shadow diagnose-pr"

The generated workflow uses the [shadow-diagnose-pr](.github/actions/shadow-diagnose-pr/action.yml) composite action with the recorded backend by default (offline, no API spend). Switch to --backend live (and add an OPENAI_API_KEY secret) when you want real intervention-based ATE + bootstrap CI + E-value on the dominant cause.

For a hand-rolled workflow, the action is a one-liner:

- uses: manav8498/Shadow/.github/actions/shadow-diagnose-pr@main
  with:
    baseline-traces: fixtures/baseline_traces
    candidate-traces: fixtures/candidate_traces
    baseline-config:  configs/baseline.yaml
    candidate-config: configs/candidate.yaml
    policy:           configs/shadow-policy.yaml
    backend:          recorded

Forked PRs run with a read-only GITHUB_TOKEN, so the action falls back to the workflow summary instead of attempting to post a comment that would fail. The marker-dedup means a re-run on the same PR edits the existing comment rather than stacking new ones.

4. Open a PR. The Shadow comment lands automatically. It leads with the verdict, names the dominant cause (or lists likely candidates when attribution can't crown one), explains the policy violation in plain English, and ends with the shadow verify-fix command to confirm a fix before merge. See [docs/sample-pr-comment.md](docs/sample-pr-comment.md) for an example, or [examples/refund-causal-diagnosis/](examples/refund-causal-diagnosis/) for a runnable scenario you can cargo-cult into your own repo.

That's it. The earlier path of shadow init --github-action --legacy-diff (raw nine-axis shadow diff without causal attribution) still ships for repos that prefer the older flow; everything below is optional depth.

Writing behavior rules

The diff tells you what changed. A policy tells you what is not allowed to change. Write one YAML file that declares the agent's behavioral contract:

# shadow-policy.yaml
rules:
  - id: confirm-before-refund
    kind: must_call_before
    params: { first: confirm_refund_amount, then: issue_refund }
    severity: error

  - id: never-leak-ssn
    kind: forbidden_text
    params: { text: "SSN:" }
    severity: error

  - id: finish-cleanly
    kind: required_stop_reason
    params: { allowed: [end_turn, tool_use] }
    severity: error

  - id: cost-ceiling
    kind: max_total_tokens
    params: { limit: 100000 }

Run:

shadow diff baseline.agentlog candidate.agentlog --policy shadow-policy.yaml

The candidate trace is checked against every rule. Violations that are new in the candidate are flagged as regressions. Violations that existed in the baseline and are now cleared are flagged as fixes. Twelve rule kinds ship today: must_call_before, must_call_once, no_call, max_turns, required_stop_reason, max_total_tokens, must_include_text, forbidden_text, must_match_json_schema, must_remain_consistent, must_followup, must_be_grounded (cheap lexical grounding gate, not NLI-backed faithfulness — see [docs/features/policy.md](docs/featur

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: manav8498
• Source: manav8498/Shadow
• License: Apache-2.0
• Homepage: https://manav8498.github.io/Shadow/

Install and usage instructions live in the source repository linked above.