The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill directs the agent to generate Java source code files containing @Approximate classes, which are then compiled and executed at runtime by the opentaint CLI tool. This dynamic compilation workflow is defined in SKILL.md and references/java.md as the primary method for verifying dataflow models.
[INDIRECT_PROMPT_INJECTION]: The skill relies on parsing output from external tools, which presents a surface for indirect prompt injection if the ingested data is maliciously crafted.
Ingestion points: The agent processes test-result.json files via the scripts/check-test-result.py helper script to evaluate test outcomes.
Boundary markers: No explicit delimiters or instructions to ignore embedded content are provided when the agent reads the output of the result parsing script.
Capability inventory: The agent has the ability to execute shell commands (e.g., opentaint, unzip, javap, uv), write code and configuration files to the local filesystem, and read arbitrary project sources.
Sanitization: The check-test-result.py script performs standard JSON parsing but does not sanitize or validate the string content of the results (such as class or method names) before printing them to the agent's context.
[COMMAND_EXECUTION]: The workflow involves the execution of various command-line utilities, including the opentaint analysis engine, unzip and javap for inspecting compiled library dependencies, and uv for executing local Python scripts. All command execution is intended for local project analysis and model verification.

create-dataflow-approximation