The Agent Skills Directory

[COMMAND_EXECUTION]: The script uses subprocess.run to interact with local system utilities such as cmux, ps, lsof, and git. These interactions are required for the skill to identify workspaces and manage process lifecycles.
[DYNAMIC_EXECUTION]: The unpark feature uses os.execvp to resume Claude sessions by executing a shell command constructed from stored metadata. The risk of command injection is mitigated through the use of shlex.quote and strict UUID validation for session identifiers.
[PERSISTENCE]: The skill maintains state through a ledger of parked workspaces stored in ~/.claude/parked/. It also recommends a symlink in ~/.local/bin/ to ensure the tool is available across shell sessions.
[INDIRECT_PROMPT_INJECTION]: The skill reads terminal screen content to identify unsent drafts, representing an indirect injection surface. Ingestion points: Terminal screen buffers are read via cmux read-screen in park.py. Boundary markers: It identifies input boundaries using terminal box-drawing characters and specific prompt glyphs. Capability inventory: The skill has the ability to terminate processes and execute shell commands. Sanitization: The script uses strict regular expressions for identifiers and shlex.quote for all command-line arguments.
[DATA_EXPOSURE]: The skill stores workspace information including directories and session IDs. It enforces 0700 directory permissions and 0600 file permissions to ensure that session data is only accessible to the current user.

park-workspace