The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute a provided bash script for both initialization and search operations.
Evidence: bash <CURRENT_SKILL_MD_PATH_DIR>/scripts/search.sh --ping is called at the start of sessions to verify connectivity.
[EXTERNAL_DOWNLOADS]: The script uses curl to communicate with an external Model Context Protocol (MCP) gateway (defaulting to mcp-gw.dingtalk.com) to retrieve search data.
Evidence: The mcp_post function in scripts/search.sh utilizes curl for network requests.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes results from external web searches, which could contain instructions designed to influence agent behavior.
Ingestion points: scripts/search.sh captures search results from the MCP gateway and returns them to the agent.
Boundary markers: Absent; search content is not wrapped in explicit delimiters or warnings.
Capability inventory: The skill package includes command execution capabilities via the search.sh script and network access via curl.
Sanitization: Absent; the script parses JSON responses but does not filter for malicious instruction patterns within the search text.
[DYNAMIC_EXECUTION]: The bash script utilizes python -c to dynamically parse JSON data from the user-provided configuration and search responses.
Evidence: The parse_mcp_url and cmd_search functions in scripts/search.sh invoke Python to handle JSON processing.

dingtalk-ai-web-search