The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external Lark/Feishu spreadsheets which are untrusted sources. This creates a surface for indirect prompt injection where malicious instructions embedded in cells could influence the agent's behavior during data analysis or processing.
Ingestion points: Data is read via the lark-cli tool as coordinated in scripts/lark_sheet_read_cli.py and processed in scripts/lark_profile_table.py and scripts/lark_detect_subtables.py.
Boundary markers: The skill uses annotated_csv with [row=N] prefixes to help the agent distinguish between data and structure, and provides extensive documentation on correctly identifying data ranges to prevent interpretation of empty or footer rows as data.
Capability inventory: The skill possesses significant capabilities including executing shell commands via subprocess.run (in scripts/lark_sheet_read_cli.py), writing to the file system (via +workbook-export), and making extensive modifications to external spreadsheets.
Sanitization: The skill includes logic in scripts/lark_detect_subtables.py to handle malformed CSV quoting, providing some resilience against data parsing errors.
[COMMAND_EXECUTION]: The helper scripts in the scripts/ directory use subprocess.run to invoke the lark-cli binary. Although the implementation correctly avoids using shell=True, it constructs and executes commands based on arguments that are derived from the agent's interaction context.

lark-sheets