The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/lane_snapshot.py uses subprocess.run() to interact with the tmux CLI. These calls are well-defined, use argument lists to avoid shell injection, and are limited to retrieving environment information like pane lists and terminal output for the current lane manager.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from tmux terminal output and agent log files, creating a potential surface for indirect prompt injection. 1. Ingestion points: Pane content is captured via capture_pane and log files are parsed in extract_log_hints (scripts/lane_snapshot.py). 2. Boundary markers: The snapshot output is structured with clear text headers such as 'panes:' and 'codex log hints:'. 3. Capability inventory: The skill is limited to read-only tmux operations and printing summary data to stdout. 4. Sanitization: The skill includes a redact() function that filters sensitive tokens (GitHub, OpenAI, AWS) and private IP addresses, and content is compacted to prevent large-scale instruction injection.
[SAFE]: The skill demonstrates protective design by implementing a comprehensive redaction engine for secrets and PII. It does not perform network operations, and its file system access is restricted to the specific logs required for its stated purpose of agent lane orchestration.

tmux-agent-lane-orchestrator