The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates shell commands that directly incorporate the user-provided "$TASK_PROMPT" variable. While the variable is wrapped in double quotes in the template (Step 4), the instructions do not specify sanitization or escaping of shell metacharacters. This potentially allows arbitrary command execution via subshell expansion (e.g., using backticks or "$()") if the agent executes the string in a shell. This risk is combined with the support for a "YOLO mode" (--yolo) which suppresses tool call confirmations.
[PROMPT_INJECTION]: The skill exposes the agent to indirect prompt injection by design. It is used to ingest and synthesize data from external web searches and perform large-scale document analysis.
Ingestion points: Data enters the agent context via Google search results and multi-source research synthesis as described in Step 1.
Boundary markers: The task prompt structure in Step 4 provides basic context framing but lacks explicit instructions to ignore or delimit embedded commands within the fetched external data.
Capability inventory: The skill possesses the capability to execute local binaries (the Gemini CLI) and modify the local filesystem through project scaffolding tasks (Step 4).
Sanitization: There are no mechanisms described for sanitizing, escaping, or filtering potential instructions embedded in external data sources.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation of the "@google/gemini-cli" package from the official npm registry. This is an external tool provided by a well-known technology organization.

gemini-dispatcher