I’m looking for narrow technical feedback on a small schema/demo for blocked agent tool calls.
The core idea is to represent a denied tool call as a completed policy decision, not as a failed tool execution. For example: decision=block, outcome=not_executed, payload_executed=false, and side_effect_occurred=false.
The schema also separates policy-resolution evidence, such as match stage, matched policy key, match kind, priority, fallback reason, and candidate policy keys.
The demo cases cover a safe read-only tool, a blocked SQL mutation, ambiguous overlapping policy matches that fail closed, and an MCP-style file/network boundary case. It also avoids putting raw paths, env vars, request bodies, secrets, or raw args into default trace metadata.
Main question: can this schema represent real blocked tool-call / trace / audit cases you’ve seen in LangGraph/LangChain-style systems?
Release/demo:
https://github.com/MkaliezZ/dhms-engine/releases/tag/v0.1.0-agentfuse-evidence-schema