ADR-0022: Preserve encoding and line endings; cursors are frontend-local

Status: Accepted
Date: 2026-06-25
Deciders: jka
Decision: D22

Context

An editor used on real repositories across platforms must not rewrite files it did not mean to change. Silent CRLF conversion is a classic source of diff noise. Separately, the protocol needs a rule for cursor and selection state now that the document lives in the core (ADR-0003).

Decision

Encoding: UTF-8 by default. On open, the file's encoding is detected; on save, it is preserved. A byte-order mark, if present, is kept.
Line endings: the file's convention (LF or CRLF) is detected and preserved. New files use LF. A file with mixed endings keeps its dominant style; rio never normalises silently.
Cursor, selection and viewport belong to the frontend. The protocol carries edits as ranges (ADR-0012), never "where the cursor is".

Detection is bounded: UTF-8 with or without BOM, validated per RFC 3629, with a lossless ISO-8859-1 byte fallback; LF and CRLF. UTF-16, bare CR, and lazy loading of large files are out of scope. The detected encoding, BOM and line ending travel as buffer metadata so a save reproduces the on-disk form.

Consequences

rio can be trusted with existing files on every platform.
Two frontends, or two views of one buffer, move their cursors independently without round trips.
Features that need the cursor (find, undo grouping) must either receive the position from the frontend as a parameter or avoid needing it. ADR-0036 and ADR-0090 were designed around this.