v0.2.1 — Protocol Stability & Persistence Fix

Overview

This release represents a critical stability and persistence overhaul of the BioVerify AI Agent architecture. It transitions the protocol from ephemeral, memory-based state to a durable, multi-agent persistence layer using Postgres (Neon), while resolving race conditions within serverless environments.

Key Changes

1. Durable Agent Persistence (LangGraph + Postgres)

• Infrastructure: Migrated from MemorySaver to PostgresSaver using a singleton connection pool for Neon.
• State Recovery: Agents can now durably "pause" for Human-In-The-Loop (HITL) interactions and resume across different execution contexts without state loss.
• Schema Setup: Added db:setup-agents script to automate checkpointer table initialization in the agent database.

2. Multi-Agent Namespacing (Thread ID Architecture)

• Collision Prevention: Refactored getThreadId to use a namespaced format: AGENT_TYPE::PUB_ID::ROOT_CID. This prevents state pollution between Submission and Review graphs.
• Parsing Logic: Updated EIP-712 signature verification to utilize the new namespaced thread format, ensuring validators correctly map signatures to specific agent types.

3. Vercel & Serverless Reliability

• Process Suspension Fix: Applied await to AI agent invocations within the CQRS command handlers. This prevents Vercel from suspending the serverless function before the agent completes its automated audit or on-chain settlement.
• Error Handling: Implemented comprehensive try/catch blocks in agent entry points with Telegram alerting for real-time crash reporting.

4. Security & Debugging

• Authorization Guards: Added explicit PRODUCTION_AUTH_CHECK logging and validation in resume nodes to ensure only assigned peer/senior reviewers can advance the state machine.
• EIP-712 Alignment: Synchronized the server-side crypto utilities with the new thread ID delimiter (::) to ensure human reviewer signatures validate against the correct publication metadata.

---

Technical Breakdown

Package	Change Description
@packages/agents	Transitioned submissionGraph and reviewersGraph to use pgCheckpointer.
@packages/cqrs	Updated processContractEvent to await agent results and log execution status.
@packages/utils-server	Updated verifyRewiewEip712 to parse namespaced threadId.
@packages/utils	Unified AgentType enum and getThreadId logic.
@packages/db	Updated workspace scripts for centralized database management.

---

Verification Results

The protocol's multi-gate state machine has been manually verified on Base Sepolia staging across all logical paths. Note: In all scenarios, participant rewards, stake slashes, and reputation scores were confirmed to be correctly computed and updated on-chain.

Gate 1: Submission Agent (Automated Audit)

• Scenario: Early Slash
  • AI detects plagiarism or integrity failure -> earlySlashPublication triggered.
  • Result: Status set to 1 (EARLY_SLASHED). No reviewers assigned.

Gate 2: Review Agent (Human-In-The-Loop)

• Scenario: Consensus PASS
  • Submission agent passes (IN_REVIEW) -> Human reviewers agree on validity.
  • Result: Status set to 4 (PUBLISHED).
• Scenario: Consensus FAIL
  • Submission agent passes (IN_REVIEW) -> Human reviewers agree on invalidity.
  • Result: Status set to 3 (SLASHED).
• Scenario: Escalation (Conflict Resolution)
  • Submission agent passes (IN_REVIEW) -> Human reviewers disagree -> Senior Reviewer tie-break.
  • Result: AI aligns on Senior verdict; status updated to 3 or 4 accordingly.