The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive process environment information by reading /proc/$pid/environ and using ps eww. This is used to identify CLAUDE_CONFIG_DIR but potentially exposes other secrets such as API keys and tokens present in the environment.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill scans shell configuration files (~/.zshrc, ~/.bashrc) and agent-specific configuration directories (~/.claude*) to discover environment settings and account wrappers.
[INDIRECT_PROMPT_INJECTION]: The skill implements an indirect prompt injection surface by fetching and processing untrusted data from external issue trackers (GitHub/GitLab API) and monitoring output from other agent panes.
Ingestion points: Reads issue bodies and comments via gh api and glab api. Monitors tmux panes using tmux capture-pane.
Boundary markers: Utilizes signed messages (e.g., --- ORCHESTRATOR (round N) ---) and manual human-in-the-loop checkpoints to distinguish authorized instructions.
Capability inventory: Executes shell commands via tmux send-keys, manages git worktrees, and performs repository merges.
Sanitization: Recommends file-based API body editing to prevent shell interpretation of markdown content, but lacks formal sanitization of ingested issue text.
[DYNAMIC_EXECUTION]: The skill drives external agent sessions by dynamically injecting commands into tmux panes using tmux send-keys. This includes passing instructions and data between the orchestrator and worker instances.

orchestrate-tmux