Bug Description

The analyze command shows that comments (#) and flags (--context) are being parsed as top commands:

🔝 Top 10 Base Commands:
   ─────────────────────────────────────────
   git           1411 (  6.7%) ███
   aws           1288 (  6.1%) ███
   #             1170 (  5.6%) ██        <-- Comments!
   --context     1144 (  5.4%) ██        <-- Flags!
   touch         1094 (  5.2%) ██

Impact

1. Model wastes capacity learning comment lines
2. Model learns flag patterns without their commands
3. Suggestions may include fragments like --context alone

Root Cause

History parsing splits on whitespace and treats first token as command, but doesn't filter:

• Lines starting with # (comments)
• Multiline commands where continuation has flags first
• Malformed history entries

Expected Behavior

🔝 Top 10 Base Commands:
   ─────────────────────────────────────────
   git           1411 (  6.7%) ███
   aws           1288 (  6.1%) ███
   touch         1094 (  5.2%) ██
   cargo          709 (  3.4%) █
   ...

Suggested Fixes

1. Filter comments during training

fn parse_command(line: &str) -> Option<&str> {
    let trimmed = line.trim();
    if trimmed.is_empty() || trimmed.starts_with('#') {
        return None;
    }
    // ...
}

2. Skip lines starting with flags

fn is_valid_command_start(first_token: &str) -> bool {
    !first_token.starts_with('-') && 
    !first_token.starts_with('#') &&
    !first_token.is_empty()
}

3. Handle multiline commands

ZSH history may contain multiline commands with \ continuations. These should be joined before parsing.

Related

• Affects model quality (currently 14.3% Hit@1)
• Related to Security: Filter sensitive commands from suggestions #86 (security filtering)