hermes - 💡(How to fix) Fix Harden approval detection for remote command substitution [1 pull requests]

Summary

The dangerous-command approval heuristics currently flag remote content executed through pipes and process substitution, such as curl ... | sh and bash <(curl ...). Equivalent shell forms using command substitution are not covered, for example:

• eval $(curl ...)
• eval \wget ...``
• source $(curl ...)
• . $(wget ...)

Per the project threat model, approval regexes are defense-in-depth heuristics rather than a load-bearing security boundary. This issue tracks improving coverage so suspicious remote-content execution variants consistently request approval.

Expected behavior

Commands that download remote content with curl/wget and immediately execute it via eval, source, or dot-source should be treated as dangerous by detect_dangerous_command().

Notes

This should be treated as command approval hardening, not as an OS isolation or sandbox boundary bypass.

Proposed fix

Add a dangerous-command pattern for eval/source/. followed by command substitution from curl or wget, plus regression tests for $() and backtick variants.