The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from project assets, connection metadata, and job logs which could contain malicious instructions. This data is then used to inform the generation of executable flow definitions.
Ingestion points: Metadata is retrieved via discover_connection_data and inspect_project_asset (e.g., column names), data content is read via read_data_preview, and execution logs are processed via get_job_run_logs (including error messages).
Boundary markers: The instructions do not prescribe the use of delimiters or specific warnings to ignore embedded instructions when interpolating this external content into prompts or code templates.
Capability inventory: The skill has extensive capabilities including creating and updating flows (create_pyflow, update_datastage_flow), executing runs (create_job_run), and deleting assets (delete_asset) across various scripts such as author.md, run.md, and recover.md.
Sanitization: No explicit sanitization or validation of external input is mandated before it is utilized in the dynamic generation of flow logic.
[DYNAMIC_EXECUTION]: The skill's primary function involves the runtime generation and execution of domain-specific code to manage data flows.
Evidence: The skill assembles pyflow DSL and DataStage SDK Python code dynamically based on user requirements and discovered schemas.
File context: This behavior is central to the state machine logic described in references/author.md and references/registry.md.

di-agent-flow-lifecycle