Loop 7 - MCP / AI Agent Guardrail Pro Foundation

[sairintechnologycom]

Loop 7 of the PkgSafe Loop Engineering roadmap.

Goal: strengthen MCP guardrails for AI coding agents while keeping PkgSafe npm-first and local-first.

Scope:

• validate package install
• validate install command
• explain package risk
• safe alternative suggestion
• agent audit log

Validation target:

• gofmt -w .
• go test ./...
• go test -race ./...
• go vet ./...
• make build
• MCP stdio tests
• evidence recorded in evidence/loops/loop-07-mcp-ai-agent-guardrail.md

Do not add SaaS, billing, SSO, hosted services, or behavior analysis changes. Use behavior_analysis wording, not sandbox, for heuristic behavior.

[sairintechnologycom]

Loop 7 evidence recorded in evidence/loops/loop-07-mcp-ai-agent-guardrail.md.

Summary:

• Added agent_instruction to validate_package_install and validate_install_command responses.
• Agent actions are now explicit: proceed, ask_human, and never_install.
• Updated MCP tool descriptions so agents see that BLOCK means never install and WARN means ask a human.
• Added requested_by to validate_install_command, defaulting to ai_agent.
• Added local MCP audit events for package and install-command decisions using the existing audit logger.
• Added tests for BLOCK/WARN agent instructions, command instructions, MCP audit entries, and malformed stdio JSON-RPC behavior.

Validation passed: gofmt, go test ./internal/mcp, go test ./..., go test -race ./..., go vet ./..., make build, make package, MCP stdio samples, and wording/secrets audit.

Known caveat: this branch is stacked on uncommitted Loop 1-6 work.