macOS alpha, 0.32.0-alpha.1

Scripts

Settings reference: behavior, defaults and safeguards for this page of desktop Settings.

Published unchanged from docs/settings/scripts.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1).

Technical reference for Settings / Scripts. The app shows a short summary; this page keeps the full detail.

Repository script reviews

  • Agent assessments are advice, not a safety guarantee or permission.
  • Unsafe advice never opens an execution-approval prompt by itself; the user must force the exact unsafe source first.
  • Approving a safe assessment clears only that notification. It does not run the script or grant execution permission.
  • Forcing overrides only the unsafe advice. Hard policy, the exact source and root, and normal execution approval still apply.
  • In Sandbox, a forced script runs in a new model-free guest that uses the same verified private stage and reviewed native network boundary, never the Local source. Isolation does not make the code safe.
  • Locally, code may change your files or contact external services without OS isolation.
  • One original attempt is retained per review. Inspect its actual tool or job outcome; it cannot be replayed.
  • Moguri lists indicator findings (pattern matches such as downloading and running code, installing packages at run time or decoding hidden content) under the full source. They are not a malware or safety verdict.
  • Contract: docs/architecture/content-review.md.

Script reviews switch

Let the agent approve scripts with download or decode findings is off by default. Off: when a script has download-and-execute, run-time install or decode-and-execute findings, the agent can't mark it safe; it must report it as unsafe and only you can force it. On: you accept the risk that the agent's own "safe" verdict runs such scripts; normal approvals and hard policy still apply. Changing it counts as a permission change, so open script reviews must be inspected again.

Scripts (previously "Foundational and approved scripts")

Built-in grep_logs, fetch_page, search_memory and memory_graph are exposed through the single internal MCP, without generated one-off scripts.

Customize / Tools tab (protocol 114, project-less from 116)

  • Lists the agent's built-in toolsets, the foundational scripts and your scripts with On/Off switches. Listing grants nothing; each chat still gets only what its environment, mode and policy allow.
  • The tab isn't tied to a project. Write a script submits your source for every chat, with the same review as a model proposal. Ask the agent to write a script opens a project-free chat with an editable request; the agent submits it with propose_script, which needs the optional evolution switch.
  • Every chat sees its project's scripts plus every-chat scripts; a project script with the same name replaces the every-chat one in that project.
  • Scripts are Node.js ES modules run in the chat's working folder (project root, or Home for a project-free chat) that read {"arguments": {...}, "projectRoot": "..."} from stdin and write at most 128 KiB to stdout. A leading // description: ... comment becomes the description agents see once approved.
  • Suggested tools lists repeated requests (for example markdown_into_pdf) found by a local phrase match. The optional repeated-task check in Automations runs it on an interval and adds new suggestions to Inbox.
  • Contract: docs/architecture/skills-tools-artifacts.md.

Warning

Approved custom scripts execute as Local host code, not in a fictional sandbox. Review the full source, enable it explicitly, and approve each call. Use only scripts you trust.

Optional evolution

The switch was previously labeled "Allow model-proposed skills, scripts and style knowledge for human review". Off by default. Nothing applies itself; approval installs disabled content and enabling and execution remain separate actions.

Proposal review

The heading was "Review exact proposed content". The consent checkbox was "I reviewed this exact source. Approval does not grant execution permissions." The SHA-256 digest shown binds the approval to the exact content reviewed.

Back to the user guide