The Agent Skills Directory

[SAFE]: The skill implements extremely restrictive security profiles for its advisory mode, explicitly disabling high-risk capabilities such as command execution, file writing, and network access for the consulting model using the Claude CLI --restricted and --disallowedTools flags.
[COMMAND_EXECUTION]: Shell commands are handled safely; Python scripts use subprocess.Popen with list-based arguments to prevent shell injection, and bash scripts utilize a hardened runner (consult-claude-model.sh) that validates all inputs against strict regular expressions (e.g., advisor_validate_thread_name).
[DATA_EXFILTRATION]: The work handoff mechanism, which relays data between sessions, is secured through deterministic target resolution. Cloud transfers are restricted to the official claude.ai domain, and live local transfers utilize a temporary 'courier' process monitored by a local PreToolUse hook that verifies the exact message and destination before execution.
[INDIRECT_PROMPT_INJECTION]: The skill manages the risk of malicious project data by wrapping advisor consultations in a hardened system prompt that establishes clear safety boundaries and by ensuring the model processing the project content has no write or execution permissions.
[SAFE]: Persistence and state management follow security best practices. All thread bindings, receipts, and session metadata are stored in user-private directories using umask 077 and explicit chmod 600/700 calls, preventing cross-user access or tampering.

claude-advisor