The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructions explicitly direct the agent to run commands "unsandboxed" and to "request escalated permissions." This is a direct attempt to circumvent security boundaries and platform-enforced sandbox restrictions.
[COMMAND_EXECUTION]: The skill constructs and executes shell commands targeting a local companion script (agy-companion.mjs) using the node runtime.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a proxy for user input, passing research prompts verbatim to a secondary agent without sanitization.
Ingestion points: User research topics provided via CLI arguments, prompt files, or stdin in SKILL.md.
Boundary markers: None; instructions explicitly mandate passing content verbatim to the delegated tool.
Capability inventory: Execution of node scripts with escalated permissions and network access (localhost ports).
Sanitization: Explicitly absent; the agent is instructed to keep all instructions (including potentially side-effectful ones) in the prompt verbatim.
[CREDENTIALS_UNSAFE]: The skill documentation notes that the process requires access to an "OAuth token file," indicating the execution environment handles sensitive credentials that could be exposed if the unsandboxed execution is compromised.

researcher