Inspects a value, then strips Ecto bookkeeping noise from the resulting string — __meta__: #Ecto.Schema.Metadata<…> and #Ecto.Association.NotLoaded<…> — so a trace shows %Order{id: 12, status: "paid", …} instead of the full Ecto internals.

One value-shaped exception precedes the string cleaning: inspect/2 first delegates a top-level Ecto.Query value to CodeStory.QueryLabel, which renders it as a compact #Ecto.Query<Schema> label (at the summary detail levels only). Everything else follows the string-cleaning path below.

We clean the inspected string, not the value: removing __meta__ from the value pre-inspect turns the struct into a plain map (%{…, __struct__: Mod}), losing the %Mod{…} name. CodeStory has no Ecto dependency — detection is purely string-shaped.

Known limitations

Because this is string-cleaning, not value-cleaning:

• a struct with __meta__ as its only field isn't stripped (no adjacent ,) — real schemas always have an id/fields, so this never happens in practice;
• a NotLoaded association whose name isn't a bare identifier (\w+) isn't dropped — rare;
• a genuine string value that literally contains the full __meta__: #Ecto.Schema.Metadata<…> field shape would be edited — the unavoidable cost of the chosen approach; low likelihood (a bare #Ecto.Schema.Metadata<…> substring, with no __meta__: prefix, is untouched);
• under a compact :limit, stripping a NotLoaded/nested-struct field can show one fewer real field at that level (only the top-level __meta__ slot is compensated by the limit bump below).