The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python and shell scripts (pre-optimize-check.sh, recover-streaming-source.py) and uses the Databricks CLI to interact with workspace metadata. These commands are necessary for its data-ops functions and follow safe implementation patterns.
[DYNAMIC_EXECUTION]: The skill uses a PreToolUse hook (streaming-guard-hook.py) to intercept shell commands and block destructive operations like DROP or VACUUM when active consumers are detected. Subprocess calls within the hook are handled securely with structured arguments rather than shell strings.
[INDIRECT_PROMPT_INJECTION]: Analysis of external data ingestion for diagnostic logic: 1. Ingestion points: Data is read from Databricks system tables via the CLI Statement Execution API. 2. Boundary markers: Data is consumed by local script logic for decision-making and not interpolated into agent prompts as instructions. 3. Capability inventory: The skill can block specific Bash commands and provide diagnostic advice. 4. Sanitization: Ingested data is parsed as structured JSON and validated before processing.
[PROMPT_INJECTION]: The eval-spec.yaml file contains adversarial test cases used for verifying the agent's safety protocols. These are standard defensive test patterns and do not represent a malicious attempt to override behavior during normal use.

databricks-streaming-guardian