The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on shell commands to manage agent sessions and repository state.
Uses tmux send-keys to dispatch instructions and control worker agent sessions across multiple terminal windows.
Employs git worktree for isolating the file system environment of each parallel agent.
Utilizes platform CLIs (gh and glab) to interact with issue trackers and pull request systems.
Inspects process environments (/proc/$pid/environ or ps eww) to discover CLAUDE_CONFIG_DIR for account quota management.
[PROMPT_INJECTION]: The skill defines a workflow where agents ingest and process untrusted data from external sources.
Ingestion points: Issues and comments are fetched from GitHub or GitLab via gh issue view and glab api (documented in worker.md and orchestrator.md).
Boundary markers: The skill mandates the use of orchestrator signatures (e.g., --- ORCHESTRATOR (round N) ---) to distinguish legitimate instructions from UI-generated suggestions or external content (provenance.md).
Capability inventory: Agents have access to git, tmux, and tracker CLIs, allowing for repository modification and further command execution.
Sanitization: The skill recommends using files for pull request body edits to prevent shell injection during command-line interaction (trackers.md).

agent-swarm