Skip to content

Guard pack

Every install includes 14 deterministic local rules. The hook inspects tool names and arguments locally in memory, then passes only matched rule IDs to the signing worker. It does not call a model or send arguments to a server.

Each rule defaults to STOP in Enforce on every tier, including Free. Free includes full Enforce on one machine without a key. Shadow sessions record WARN with the rule ID without denying the call. The effective mode, not whether you pay, determines whether a default match stops the call.

Scan limits, unreadable branch state and internal failures select shadow with a reason. These failures allow the tool call. Check the effective mode and reason before assuming enforcement is active.

The categories cover remote shell execution, broad recursive deletion, Git history, infrastructure changes, credential writes and secret arguments, system security, and package sources. The IDs and reasons below match the policy reference.

Category Rule ID Reason
Remote shell execution GP001 A downloaded response is being executed by a shell or Python.
Recursive deletion GP002 Recursive deletion targets a root, home, current directory or shallow wildcard.
Git history GP003 A force push targets the shared main or master branch.
Git history GP004 A hard reset targets a shared branch or a branch whose state is unknown.
Git history GP005 Git clean would remove untracked directories and ignored files.
Infrastructure GP006 Terraform would apply infrastructure changes without interactive approval.
Infrastructure GP007 Kubernetes would delete a namespace.
Infrastructure GP008 A cloud command would delete, terminate or disable a resource.
Credential writes GP009 A write targets a credential, private-key, wallet or environment file.
Secret arguments GP010 An argument contains a known secret-key format.
System security GP011 Permissions would allow every local user to write and execute the target.
System security GP012 A command would turn off or flush a local firewall.
System security GP013 A write targets host resolution or sudo authorization settings.
Package sources GP014 A package install uses a URL or Git source rather than a registry version.

Only a shared team policy or org policy can downgrade a rule. Set a built-in ID in the guardPack.rules map to warn or off. For example, the site’s shared policy example sets GP001 to warn and GP014 to off. The only accepted values are stop, warn and off; unknown rule IDs and fields are rejected.

An org policy is the root layer. An omitted org rule keeps its default STOP, and lower layers cannot weaken it. A lower layer can tighten an explicitly downgraded rule. A personal policy alone cannot disable the pack. Protect policy and hook configuration from agent writes when they are organizational controls.

See Org policy for teams for publishing and the policy layers.

False positives on 356 benign commands: 0.

Measured 2026-09-20 on Apple M2 Max (darwin), Node v22.17.1 arm64. Static command strings were scanned, not executed. Hard-reset examples use an explicit feature-branch context. This result covers the named corpus, not arbitrary commands.

Read the corpus measurement and the security page for the measurement’s scope.

These checks cover recognizable command forms, not every equivalent shell program. Main and master are the shared branches recognized for hard reset. If branch context cannot be established, guard_branch_unknown selects shadow. Encoded commands, aliases and programs that construct commands later can bypass pattern checks. A clean result is not proof of safety.

Rules cover tool calls the host sends to trusted hooks. Disabled hooks, unsupported host paths and tools outside that boundary are not governed. The guard pack is not a complete shell interpreter, a malware detector or proof that allowed code is safe.