The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to suggest a command to the user that downloads a shell script from a remote GitHub repository and pipes it directly into the bash interpreter. This 'curl | bash' pattern is a significant security risk as it bypasses local verification and package management audits.
Evidence: curl -fsSL https://raw.githubusercontent.com/Rheinmir/setup/orca/harness/poc-vendor-neutral/bootstrap.sh | bash suggested in the 'Pre-check' section of SKILL.md.
[COMMAND_EXECUTION]: The skill utilizes multiple shell commands to inspect project state, check for file existence, and read logs. While functional for a security tour, these represent a broad capability for local command execution.
Evidence: Execution of grep -q llmwiki-validate .claude/settings.json, test -f harness/poc-vendor-neutral/bin/llmwiki-validate.py, and cat .claude/audit/*.jsonl as described in the workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes project audit logs (.jsonl) that record previous tool interactions. This creates an attack surface for indirect prompt injection if an external entity can influence the content of these logs.
Ingestion points: .claude/audit/*.jsonl referenced in the 'Kết màn' (Conclusion) and workflow sections of SKILL.md.
Boundary markers: None present; the skill treats log content as a source of truth for the tour summary.
Capability inventory: File writing/deletion (llmwiki/, wiki/index.md) and shell command execution (grep, test, cat, python3).
Sanitization: No sanitization or validation of the audit log content is mentioned before it is displayed or processed.

harness-tour