The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to suppress standard AI safety behaviors regarding user oversight and confirmation.
references/interaction-wizard.md and references/autonomous-loop-protocol.md mandate that the agent must "NEVER pause to ask the user anything during the loop -- not for clarification, not for confirmation, not for permission."
This behavior is justified in-context as allowing the user to "sleep" while the agent works, but it effectively removes human-in-the-loop control for high-risk actions.
[COMMAND_EXECUTION]: The skill's architecture relies on executing arbitrary shell commands provided by the user or inferred from the repository structure.
Multiple Python scripts, including scripts/autoresearch_init_run.py, scripts/autoresearch_record_iteration.py, and scripts/autoresearch_health_check.py, use subprocess.run to execute git operations and user-defined "verify" and "guard" commands.
The references/exec-workflow.md specifically encourages users to run the agent with the --dangerously-bypass-approvals-and-sandbox flag in CI/CD environments.
[REMOTE_CODE_EXECUTION]: The skill supports a "background" mode that spawns detached long-running processes.
scripts/autoresearch_runtime_ops.py uses subprocess.Popen with start_new_session=True to launch a detached runtime controller that manages its own lifecycle and nested Codex sessions.
[EXTERNAL_DOWNLOADS]: The skill includes mechanisms for fetching data and code from external sources.
references/web-search-protocol.md defines a protocol for autonomous web searches when the agent is "stuck," treating search results as inputs for new code-generation hypotheses.
The documentation in README.md and docs/INSTALL.md provides instructions for downloading the skill via git clone from a non-vendor GitHub repository.

codex-autoresearch