The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing shell commands via cmux and git to manage development workspaces and repository worktrees. It constructs strings to launch fresh claude instances with specific instructions.
[DYNAMIC_EXECUTION]: The skill uses shell utilities to decode the CMUX_AGENT_LAUNCH_ARGV_B64 environment variable. This allows it to dynamically determine and replicate the agent's permission flags in child sessions.
[INDIRECT_PROMPT_INJECTION]: The skill creates a potential attack surface by interpolating user-provided topics into shell-executed strings for worktree naming and terminal labeling. The skill provides clear sanitization rules to mitigate this risk, but the underlying mechanism involves executing code based on untrusted input.
Ingestion points: The $target argument and the conversation history used to generate handoff summaries.
Boundary markers: The skill packages context into separate markdown files and launches new agent sessions with explicit opening prompts.
Capability inventory: Extensive use of cmux subcommands, git commands, and standard POSIX shell tools for string manipulation and process management.
Sanitization: The agent is explicitly instructed to enforce a strict character set ([a-z0-9-]) for slugs and to avoid shell metacharacters in labels and prompts.

handoff-to-worktree