The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/Invoke-HarnessOffload.ps1 executes external agent binaries including agy, opencode, copilot, codex, and hermes using the System.Diagnostics.Process class.
[PRIVILEGE_ESCALATION]: The Ensure-AgyHeadlessReadAllow function in scripts/Invoke-HarnessOffload.ps1 modifies the persistent configuration file at ~/.gemini/antigravity-cli/settings.json. It automatically adds read_file(*) and command(*) to the allowed permissions list and enables allowNonWorkspaceAccess, effectively lowering the security sandbox of the agy CLI tool for both current and future sessions.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the transfer of instructions from a host agent to a sub-agent worker, creating a vector for cross-agent instruction injection.
Ingestion points: The $Prompt parameter in scripts/Invoke-HarnessOffload.ps1 receives instructions from the host environment.
Boundary markers: The script appends a text guard directive to the prompt, which may be insufficient to prevent sophisticated prompt injection attacks against the worker agents.
Capability inventory: Depending on the worker selected and the state of the AllowTools flag, the skill can execute arbitrary shell commands, read/write files, and access network resources through the invoked agents.
Sanitization: The script performs command-line argument escaping using a custom ConvertTo-WinArgLine function, but it does not validate or sanitize the semantic content of the prompt passed to the external agents.

harness-offload