The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The start-panes.sh script launches sub-agents with flags that explicitly bypass security constraints. Claude is launched with --dangerously-skip-permissions, and Codex is executed with --dangerously-bypass-approvals-and-sandbox. These flags disable user permission prompts, human-in-the-loop approvals, and execution sandboxing for the agents performing the workflow tasks.\n- [COMMAND_EXECUTION]: The skill uses tmux-bridge to dynamically construct and send commands to multiple terminal panes. This orchestration allows the skill to programmatically interact with shell environments and other AI agents.\n- [INDIRECT_PROMPT_INJECTION]: The skill features an indirect injection surface by reading content from markdown files (e.g., spec.md, change.md, test.md) and using it to drive automated workflow decisions and command generation.\n
Ingestion points: Logic in SKILL.md reads and parses project documentation and status files to determine phase transitions.\n
Boundary markers: None identified; file content (like status values and review comments) is processed directly by the orchestrator.\n
Capability inventory: Uses tmux-bridge for typing commands, tmux for pane control, and CronCreate for periodic monitoring.\n
Sanitization: No evidence of sanitization or validation of the data extracted from files before it influences command execution.

eo-workflow