The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands provided as session configuration (e.g., benchmark workloads and correctness checks). The run_shell_command function in scripts/common.py utilizes subprocess.run(shell=True) and defaults to the system shell (e.g., /bin/bash), which allows for the execution of arbitrary shell strings.\n- [DYNAMIC_EXECUTION]: The skill workflow involves the agent generating local shell scripts (autoresearch.sh and autoresearch.checks.sh) based on user instructions, which are then programmatically executed by scripts/run_experiment.py during the measurement cycle.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes output from external benchmark workloads, which may contain untrusted data that is later reflected in reports or session logs.\n
Ingestion points: The run_single_trial function in scripts/run_experiment.py captures the full stdout and stderr of executed commands to parse metrics and diagnostics.\n
Boundary markers: The skill does not employ explicit delimiters or boundary markers when reading command output, although it specifically filters for lines prefixed with METRIC .\n
Capability inventory: The skill can execute shell commands (subprocess.run), perform local file writes (artifact logging), and modify git configuration (.git/info/exclude).\n
Sanitization: The render_html_report function in scripts/common.py utilizes html.escape() when embedding data into the generated HTML reports, mitigating potential cross-site scripting risks.

autoresearch