The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill monitors terminal output using cmux read-screen to track task status. This terminal content is treated as authoritative evidence but originates from potentially untrusted data or agent-controlled processes.
Ingestion points: Terminal output retrieved via cmux read-screen --workspace <ref> (referenced in SKILL.md and references/lifecycle.md).
Boundary markers: None present to distinguish between tool output and potential embedded instructions.
Capability inventory: The skill calls cmux, git, start-issue, and direnv (in scripts/start-in-cmux).
Sanitization: No sanitization is performed on the screen output before it is processed by the agent.
[DYNAMIC_EXECUTION]: The scripts/start-in-cmux script dynamically assembles a shell command string from the issue identifier and agent name provided by the user.
Evidence: The script uses a loop to build command_text and passes it to cmux new-workspace --command. While it uses shell-safe quoting (printf %q) to ensure parameters are handled correctly, it still constitutes runtime generation of executable commands.
[COMMAND_EXECUTION]: The skill relies on executing external CLI tools (git, cmux, start-issue, direnv) to manage the repository and workspace lifecycle.
Evidence: scripts/start-in-cmux verifies the presence of these commands and executes them to initialize worktrees and terminal environments.

start-issue-cmux