The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute shell commands and browser automation scripts defined in an external test plan (test.md). This is the primary intended function of the test runner, and the skill provides mechanisms for capturing and auditing all command output via logs.
[REMOTE_CODE_EXECUTION]: The skill supports cross-AI evaluation using the agent-delegate.sh utility. This allows the test driver to delegate execution to other models while maintaining a strict sandbox (workspace-write) and verifying the resulting artifacts via sha256 checksums and file size checks.
[PROMPT_INJECTION]: The skill processes test cases from test.md which influence subagent behavior, representing a potential indirect prompt injection surface.
Ingestion points: Test cases are parsed from .specs/{feature}/test.md as described in SKILL.md Step 1.
Boundary markers: Commands are executed literally as defined in the input file; framing is provided by the evaluator-prompt.md instructions.
Capability inventory: The execution environment includes bash command execution, browser automation (Playwright), and file system writes for evidence collection.
Sanitization: No input sanitization is performed on the test commands, but security risks are mitigated by the 'Evidence over Self-Report' principle and the 'Golden Rules' which explicitly forbid accessing real user data and require synthetic fixtures.

spec-evaluate