AI Swarm Orchestration

Key things to check before installing or running this skill: - Audit spawn-batch.sh, integration-watcher.sh, notify-on-complete.sh and any file that runs 'git add/commit/push' or 'gh pr' to confirm whether auto‑endorse/auto‑merge behaviors exist and can be disabled. The SKILL.md says human approval is required, but some scripts/docs indicate automatic endorsement — get clarity from the author. - Inspect and remove any agent CLI flags that bypass permissions (examples in docs: --permission-mode bypassPermissions, --dangerously-bypass-approvals-and-sandbox, --dangerously-skip-permissions). Those flags defeat safety checks and let agents run non‑interactive privileged actions. - Decide where you will run the swarm: use a throwaway or test repository and an isolated machine/VM first. Do not run this against production repos until you are confident. - Supply only minimal credentials and tokens. Create a dedicated machine/service account or GitHub user with limited repo scope (avoid granting broad org/admin rights) and a Telegram token/channel that is intended for notifications only. - Provide a swarm.conf and explicitly set SWARM_NOTIFY_TARGET, OBSIDIAN_BASE, and any other paths so scripts don't guess locations. Consider making notify/push steps manual (disable auto-push or require an explicit command) and enable logging/audit of all commits prior to push. - Test the orchestration in dry-run mode if possible, and search the codebase for all commands that call external CLIs (openclaw, gh, tmux, git, claude/codex/gemini) to verify behaviour. If you want, I can list the exact files/lines that introduce the biggest risks (auto‑merge, bypass flags, implicit credential use) so you can review or modify them before use.