The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to manipulate project state files. It uses jq for processing JSON data, consistently employing the --arg and --argjson flags to prevent injection when interpolating variables into queries. It also performs standard shell operations like file redirection and moving temporary files.
[COMMAND_EXECUTION]: Step 0.1 implements explicit input sanitization for command-line arguments. It truncates the arguments to 200 characters and strips potentially dangerous shell characters like backticks and angle brackets. It also validates arguments against a whitelist for the --scope and --mode parameters.
[DATA_EXPOSURE]: The skill reads from and writes to the ./reports/ directory to manage research session states, quarantine files, and reports. Access is limited to the expected working directory of the research orchestrator.
[PROMPT_INJECTION]: The skill orchestrates a sub-agent (falsification-analyst) by defining its behavior and constraints in a system prompt. The prompt includes specific instructions to the sub-agent to use only web tools and never commit via git, which serves as a safety constraint for the sub-task.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from research findings in state.json and from the output of the falsification-analyst agent. While these are points where untrusted data could enter the context, the skill mitigates risk by sanitizing the CLI input and using structured JSON processing with schema validation (schemas/*.jq) before updating the primary session state.

falsify