The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends running code directly from a personal, unverified GitHub repository (louchi1984-coder/deepseek-claude-code-worker-mcp) using the npx command. This bypasses standard package security audits and allows the execution of remote code without version pinning or integrity checks.
[COMMAND_EXECUTION]: Several tools provided by the skill, including deepseek_start_implementation and deepseek_implement_in_workspace, accept a validation_commands parameter. This parameter allows for the execution of arbitrary shell commands in the project directory, which can be exploited to run malicious scripts or modify the system.
[CREDENTIALS_UNSAFE]: The documentation specifies that the skill looks for sensitive API keys in specific local file paths such as ~/.codex/secrets/deepseek_api_key. Standardizing access to these paths creates a target for credential harvesting scripts.
[PROMPT_INJECTION]: The task parameter in the implementation tools serves as a significant surface for indirect prompt injection. Instructions provided within this string are passed directly to a worker model that has file system and shell access.
Ingestion points: The task argument in deepseek_start_implementation and deepseek_implement_in_workspace tools.
Boundary markers: The skill uses directory-based boundaries (allowed_dirs, forbidden_paths) but does not implement sanitization or escaping for the natural language task instruction.
Capability inventory: The skill allows for file modification, creation, and arbitrary bash command execution via the background worker process.
Sanitization: No input sanitization is documented for the task string; the skill relies on a pre-execution hook within the externally downloaded code.
[EXTERNAL_DOWNLOADS]: The skill automates the installation of the @anthropic-ai/claude-code CLI package during setup. This is a reference to a well-known service.

deepseek-claude-code-worker-mcp