Leiter's distillation pipeline had three weaknesses that compounded:

1. The pre-processing stripped too much context — tool_use blocks were
   silently dropped, so the distilling agent couldn't see what tools
   were used when the user corrected something. A user saying "no, don't
   do that" after an [assistant tool]: Edit(src/main.rs) line is far
   more informative than just seeing the correction in isolation. This
   behavior was introduced earlier when we added log pre-processing
   to cut down on token use use - but that was primarly from tool output,
   which we are still filtering out.
2. The writing guidelines didn't help the agent decide whether to
   record an observation. Every correction got recorded equally, leading
   to one-off debugging decisions polluting the soul alongside genuine
   lasting preferences.
3. The user never saw what distillation changed. The sub-agent would
   silently modify the soul and the user had no visibility into what
   was learned, making it hard to catch bad entries or confirm good ones.

Changes:

• Richer pre-processing: assistant messages with tool_use blocks now
  emit [assistant tool]: Name(param) summary lines. Key parameter is
  chosen heuristically (file_path > command > pattern). Tool results
  remain dropped — the tool name from the assistant side is sufficient.
• Recording judgment guidelines: new subsection in SOUL_WRITING_GUIDELINES
  teaching the distiller to prefer patterns over one-offs, use tool
  context to understand corrections, skip ephemeral decisions, and
  capture implicit positive signals.
• Distillation feedback: sub-agent now ends with a summary of what it
  changed, and the main agent relays it to the user after mark-distilled.
  Works identically for manual and auto-distill (foreground vs background
  sub-agent).

changelog: include