The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/swain-test.sh allows for the execution of arbitrary shell commands defined in project configuration.
Evidence: The run_integration function in scripts/swain-test.sh executes the value of .integration.command from .agents/testing.json using bash -c. This allow a malicious repository to execute arbitrary code when the agent runs tests.
[COMMAND_EXECUTION]: The skill automatically executes test commands based on detected project files (e.g., package.json, Cargo.toml).
Evidence: The detect_test_command function triggers build-tool commands like npm test, which can execute arbitrary code defined in the project's dependency manifests or test scripts.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection from test outputs.
Ingestion points: Test results and error messages from scripts/swain-test.sh (captured via _RUN_OUTPUT) are emitted directly to the agent's context.
Boundary markers: No delimiters or isolation markers are used to wrap the untrusted test output.
Capability inventory: The agent has access to the shell and filesystem through this skill.
Sanitization: None; external content is interpolated directly into the tool response.

swain-test