feat: add governance layer inside AgentExecutor ReAct loop#6426
feat: add governance layer inside AgentExecutor ReAct loop#6426Shreyashgol wants to merge 1 commit into
Conversation
Shreyashgol
commented
May 23, 2026
Shreyashgol
commented
- Insert policy check between tool resolution and tool.call()
- Add policyLoader, policyChecker, hitl, auditLogger modules
- policies.yaml with allow/deny/escalate rules
- Append-only audit log at project root
- Human-in-the-loop via CLI readline
- Insert policy check between tool resolution and tool.call() - Add policyLoader, policyChecker, hitl, auditLogger modules - policies.yaml with allow/deny/escalate rules - Append-only audit log at project root - Human-in-the-loop via CLI readline
|
Warning Gemini encountered an error creating the review. You can try again by commenting |
|
governance inside the AgentExecutor ReAct loop is the right placement. one missing piece: tamper-evident evidence of every governance decision. when the governance layer allows or denies an action, that decision should be captured in a signed receipt. otherwise the governance check is invisible to downstream auditors. the receipt carries: agent ID, action type, scope, verdict (ALLOW/DENY), policy_version, and an Ed25519 signature. modify anything after the fact and the signature breaks. for Flowise specifically: a governance callback that emits a nobulex receipt at each ReAct step gives every flow execution a verifiable audit trail. EU AI Act Article 12 requires tamper-evident logs for high-risk AI systems (deadline: August 2). github.com/arian-gogani/nobulex — open source, MIT licensed. |
|
Yeah , you are right , I stored in the audit.log file . |
2 participants
