Benchmarking showed qwen2.5-coder commonly wraps its one-liner in a single pair
of backticks (`cmd`), which the triple-fence stripper missed. Strip a matching
leading+trailing backtick pair only, leaving inline command substitution intact.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Default the ai mode to qwen2.5-coder:1.5b instead of the hardcoded (and
uninstalled) westenfelder/NL2SH, which left ai mode broken. The model is now
configurable via SUPERMAN_AI_MODEL / --model, with markdown-fence stripping and
a coder-specific prompt wrapper so general coder models emit a single bare
one-liner. NL2SH still receives the raw query. Adds --compare to run multiple
models side by side (SUPERMAN_AI_COMPARE_MODELS) with timing for benchmarking.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Rename bash-helper.sh -> super-man.sh and update all docs/tests to the
super-man name and alias. In interactive mode, pressing Esc in the flag
browser now returns directly to the home menu, removing the intermediary
"Press Enter to search another command" prompt.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
leetcrypt