The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to run shell scripts and commands on the user's local host machine via the call_remote and queue_task functions, allowing execution outside the restricted agent sandbox.
[REMOTE_CODE_EXECUTION]: Facilitates the execution of local scripts and a free-form coding agent (run_claude.sh) on the host system, effectively bridging the gap between the agent's environment and the user's real machine.
[EXTERNAL_DOWNLOADS]: Instructs the agent to perform git clone and git fetch operations from a community-managed GitHub repository (abhinaykrupa/cowork-to-code-bridge) for the purpose of code review and inspection.
[DATA_EXFILTRATION]: The bridge daemon runs with local user permissions and provides access to local files, environment variables, and cloud credentials. The output from these operations (stdout and stderr) is returned to the agent context, creating a potential path for sensitive data exfiltration.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes command output from the host system, which may contain untrusted data or malicious instructions capable of influencing the agent's behavior.
Ingestion points: Command output (stdout/stderr) from the host machine via call_remote and poll_task_result (SKILL.md).
Boundary markers: The skill documentation includes warnings to treat all command output as "sensitive untrusted data."
Capability inventory: Broad access to read/write files and execute arbitrary commands on the user's host machine.
Sanitization: Instructs the agent to validate output formats and manually redact credentials, tokens, and private paths.
[DYNAMIC_EXECUTION]: Provides a "Free-Form Local Agent Boundary" using run_claude.sh, which enables a local agent to execute dynamic, multi-stage coding tasks on the host machine with varying permission levels, including edit and full access.

cowork-to-code-bridge