The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external platform tools (ae-cli, lark-cli, curl) using the Python subprocess.run module. It follows secure implementation patterns by passing arguments as arrays rather than strings, which prevents command injection by avoiding shell interpolation.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from SQL queries and dashboard reports to generate automated summaries. This creates a potential indirect prompt injection surface; however, the risk is mitigated by explicit instructions for the agent to follow a rigid outline, cite values directly, and use HTML escaping for report rendering.
Ingestion points: ThinkingData platform via ae-cli (adhoc queries and dashboard reports).
Boundary markers: The skill uses strict JSON schemas for task specifications and instructions that forbid the LLM from altering data or metric definitions.
Capability inventory: File system access for report generation, network POST to webhooks, and execution of platform-specific CLI tools.
Sanitization: Implements _esc for HTML entity escaping in report generation and _csv_safe to prevent formula injection in CSV exports.
[DATA_EXFILTRATION]: The primary purpose of the skill is the periodic movement of business data from an analysis platform to messaging groups and webhooks. Since this is the explicitly stated primary functionality and the destinations are configured by the user during the BUILD phase, it is considered safe and legitimate.
[CREDENTIALS_UNSAFE]: The skill processes sensitive webhook URLs. The execution script push_deliver.py includes a _redact_creds function that uses regex to mask access tokens and keys in logs and persistent artifacts, reducing the risk of accidental credential exposure.

ae-scheduled-push