Skip to content

developer.md

Latest commit

 

History

History
96 lines (73 loc) · 4.12 KB

developer.md

File metadata and controls

96 lines (73 loc) · 4.12 KB
Raw

For developers

You're here because you need a working audit trail for an AI decision function — and you don't want to stand up a SaaS, an API key, or a cloud account to get one. Fine. Five commands, no accounts, and you're running.

90 seconds to your first trace

pipx install 'sentinel-kernel[pdf]'
sentinel quickstart
python hello_sentinel.py
sentinel audit-gap

That's a real trace store (SQLite at ./.sentinel/traces.db), a real hash-only decision record per call, and a real readiness score you can re-run every time you change something.

Wrap your own agent

from sentinel import Sentinel

sentinel = Sentinel()        # privacy-by-default, SQLite, zero config

@sentinel.trace
def approve(request: dict) -> dict:
    # your existing logic — unchanged
    return {"decision": "approved", "amount": request["amount"]}

approve({"amount": 5_000, "requester": "alice"})

One line on top of whatever function makes the decision. Sync or async. LLM, ML classifier, rule engine, robotic control — Sentinel doesn't care what decides. It records that the decision happened, what policy ran, and what hashed inputs produced what hashed output.

When you want raw payloads

Default is hash-only (GDPR Art. 25 by design). When you have legal basis and controlled access, opt in explicitly:

sentinel = Sentinel(store_inputs=True, store_outputs=True)

See the honest reasoning: docs/provability.md and docs/explainability-art-22.md.

Framework integrations

Framework Import Extra
LangChain from sentinel.integrations.langchain import SentinelCallbackHandler [langchain]
CrewAI from sentinel.integrations.crewai import SentinelCrewCallback [crewai]
AutoGen from sentinel.integrations.autogen import SentinelAutoGenHook [autogen]
Haystack from sentinel.integrations.haystack import SentinelHaystackCallback [haystack]
OpenTelemetry from sentinel.integrations.otel import OTelExporter [otel]
LangFuse from sentinel.integrations.langfuse import LangFuseEnricher [langfuse]
Prometheus from sentinel.integrations.prometheus import PrometheusExporter [prometheus]
FastAPI from sentinel.integrations.fastapi import SentinelMiddleware [fastapi]
Django from sentinel.integrations.django import SentinelMiddleware [django]
Jupyter from sentinel.integrations.jupyter import SentinelWidget [jupyter]

Detail: docs/integration-guide.md. Every class name above is a real, imported symbol in v3.4 — each table row is copy-paste safe.

Debug / day-to-day commands

sentinel status                    # decision activity + readiness, one screen
sentinel demo --no-kill-switch     # narrative defence-logistics walk
sentinel scan                      # dependency + CI/CD sovereignty
sentinel export --output out.ndjson
sentinel verify --all              # recompute hashes for every stored trace

Breaking change — v3.2.0 privacy default

v3.1.0 defaulted to store_inputs=True. v3.2.0 defaults to False. If you upgrade without changing config, your new traces will hold hashes only — your existing traces are unaffected. Full migration note: docs/migration-v3.2.md.

Next

  • Your sentinel audit-gap score < 80 %? Run sentinel fix kill-switch and sentinel fix retention --days 2555. Score moves. That's the point.
  • Ready for an auditor? sentinel evidence-pack --output audit.pdf.
  • Need deployment help? Open a pilot enquiry on GitHub.