openclaw-kirocli-coding-agent

Run Codex CLI, Claude Code, Kiro CLI, OpenCode, or Pi Coding Agent via background process for programmatic control.

This skill is coherent with its stated purpose, but take these precautions before installing or using it: - Avoid enabling insecure automation flags unless you understand the consequences. --trust-all-tools, --yolo, and similar flags disable confirmations/sandboxing and can let CLIs execute arbitrary actions in the workdir or host if combined with elevated modes. - Prefer running agents inside temporary or dedicated review workdirs (mktemp, git worktree) and never point them at system folders or OpenClaw's own codebase. - Be cautious installing CLIs via curl | bash; prefer vetted package managers or verified release binaries when possible and inspect the installer if you must use it. - When authenticating Kiro/other CLIs, scope credentials with least privilege (avoid long-lived full-access IAM credentials). Use identity providers or short-lived credentials when possible. - Monitor background sessions (list/poll/log) and kill them if they behave unexpectedly; restrict what files tools can access by setting workdir appropriately. If you want a stricter evaluation, provide examples of how OpenClaw will invoke these bash commands at runtime (the exact execution wrapper), and whether your OpenClaw environment enforces a sandbox/elevation policy—that would allow raising confidence above 'medium'.

Type: OpenClaw Skill Name: openclaw-kirocli-coding-agent Version: 1.5.0 The skill bundle provides instructions for the OpenClaw agent to orchestrate powerful AI coding tools (Kiro CLI, Codex, Claude Code) using high-risk configurations. Specifically, it directs the agent to use flags like `--yolo` and `--trust-all-tools` in SKILL.md, which explicitly bypass sandboxing and manual approval prompts for file writes and command execution. While these capabilities are aligned with the stated goal of automated coding, they grant the agent broad, unverified permissions over the host system. No evidence of intentional malice, data exfiltration, or persistence was found, but the instructions encourage a high-risk operational mode.