@markjaquith/agency@3.5.2
Manage agentic work across repositories with durable workbases
AI Security Review
scanned 1mo ago · by lpm-firewall-aiThe automatic global Pi extension is a meaningful agent-extension lifecycle risk and warrants a warning. The inspected code is transparent and lacks a concrete malware or self-dependency chain, so blocking is not supported.
Decision evidence
public snapshot- Install automatically runs a Pi extension installer.
- The installer copies package code into the user's global Pi extension directory.
- The installed extension adds skill paths and changes the agent system prompt before a Pi agent starts.
- The package can launch subprocesses, including Git commands, with inherited environment variables.
- Git support and dynamic UI loading are visible source features, not proof of malware.
- No runtime dependency on this package's own name was found.
- The installer only copies its bundled extension and removes that same extension on uninstall.
- The extension is scoped to discovered Agency workbases before it injects context.
- No direct HTTP client, credential harvesting, payload download, eval, or shell-string execution was found.
- Network-capable Git operations are part of the repository-management feature.
Source & flagged code
10 flagged · loading sourcePackage defines install-time lifecycle scripts.
package.jsonView on unpkgInstall-time lifecycle script is not statically allowlisted and needs review.
package.jsonView on unpkgPackage source references dynamic require/import behavior.
src/utils/interactive-loader.tsView on unpkg · L3Findings
Affected versions and remediation
This report applies to @markjaquith/agency@3.5.2.
See all versions of @markjaquith/agency for other recorded verdicts.
- Review the evidence and your use of @markjaquith/agency@3.5.2 before allowing it. Restrict the permissions described in this report.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.