Rework Sentinel into a change-risk and review-assurance engine

[SAY-5]

Rebuilds Sentinel around a provider-agnostic thesis: stop guessing whether code was AI-authored, and instead measure how risky each change is and whether it was actually reviewed.

Core model

Change Risk Score = Blast Radius x (1 - Review Assurance). Risk is high-blast-radius change that went under-reviewed. Every score cites the concrete facts behind it.

• Blast radius: file-sensitivity classification (auth, payments, migrations, crypto, infra = critical) plus coupling, churn, and hot-file modifiers.
• Review assurance: grounded penalties (unreviewed merge, self-merge, rubber-stamp approvals faster than readable, stale approvals, direct-to-default pushes) and credits (multiple engaged reviewers, passing checks).
• Outcome linkage: revert, hotfix, and incident detection, feeding a tier-calibration view that checks whether higher-risk predictions actually fail more.

What changed

• New scoring engine (src/analysis): blast-radius, review-assurance, and outcome signals, each producing human-readable evidence; scoreChange computes the deterministic Change Risk Score.
• Schema: dropped the per-file attribution table; code_events carries per-change scores and review metadata; repo_metrics now rolls up Change Risk Index, Review Assurance, Unreviewed Critical Changes, and Revert-Linked Rate.
• Webhook flow scores every merged change and direct push; alerts retargeted to unreviewed-critical, index spikes, assurance drops, and revert clusters.
• Dashboard, routers, seed data, and docs reframed end to end. New CI workflow (lint, typecheck, test, build).

Verification

• pnpm lint, pnpm exec tsc --noEmit, pnpm test (17 unit tests), and pnpm build all green locally.
• Dashboard pages render dynamically, so the build needs no database.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sentinel	Ready	Preview, Comment	Jul 20, 2026 4:53pm