@agent-rewind/core

@agent-rewind/core

The portable core of Agent Rewind — reversible execution for AI agents, as pure logic over logs and traces (never the filesystem or clock, so it runs unchanged across backend tiers).

Provides:

• WorldBackend — the whole-workspace snapshot/restore interface, with a Tier-0 git implementation (createGitBackend): append-only snapshot chain, atomic restore that rolls back on a mid-flight failure, and secret/heavy-dir excludes.
• The effect barrier (createEffectLedger) — admit-once / refuse-and-record over (scope, effectKey) identity, so a spent external effect can't be re-fired across a rewind.
• The tamper-evident hash chain (createMemoryEvidenceLedger, computeEntryHash, verifyChain) — append-only, per-scope, SHA-256-linked, fail-closed verification. The action vocabulary and authority resolver are injected, so it carries no product-specific vocabulary.
• Canonical JSON (canonicalize), the engine (createEngine), a replay-savings sink, and an AgentRewind-style recovery/failure-memory model.

npm install @agent-rewind/core

Requires Node ≥ 20. Licence: FSL-1.1-ALv2. See the monorepo for architecture and the full product breakdown.