Format, normalization, and small utilities

A grab-bag of support modules that keep the core small. Most callers never touch these directly — they exist so the parser, patcher, and host renderers share one source of truth.

format — the single source of truth for the syntax

Every sigil and separator the language uses is a constant here, so the parser, the tokenizer, the prompt, and any renderer can never drift apart:

Hashing

compute_file_hash(text) is the content tag function: it normalizes the text (trim trailing [ \t\r] from every line so CRLF endings and display-trimmed lines don’t invalidate a tag), runs xxh32, and returns the low 16 bits as 4 uppercase hex chars ("{:04X}"). Any read of byte-identical content mints the same tag — that’s what makes tags verifiable.

Formatting helpers

• format_hashline_header(path, hash) → ¶path#hash — the header a tool emits after reading or editing, and the anchor for follow-up edits.
• format_replace_header(start, end) / format_delete_header(start, end) / format_insert_header(&cursor) — build op headers programmatically (delete 5 for a single line, delete 5..7 for a range).
• format_numbered_line(n, line) → 5| text — one displayed line.
• format_numbered_lines(text, start_line) → a whole file with N| -style prefixes, hashline mode.
• describe_anchor_examples(prefix) → "160", "42", "7" — example anchors for error messages.
• hl_file_hash_re_raw() / hl_line_body_sep_re_raw() — regex fragments for embedding in larger patterns.

normalize — text-shape round-trip

The patcher canonicalizes file text to LF before applying edits and restores the original shape on write-back:

• strip_bom(content) → BomResult { bom, text } — strips a UTF-8 BOM, returning both halves.
• detect_line_ending(content) → LineEnding::Crlf | Lf — first line ending wins; defaults to Lf.
• normalize_to_lf(text) — every \r\n/\r → \n.
• restore_line_endings(text, ending) — LF text back to CRLF if needed.

prefixes — stripping display prefixes from payloads

Payloads authored against read/search output may carry display prefixes (line numbers 5| , diff +, or >>>/>> continuation markers). strip_new_line_prefixes (opportunistic) and strip_hashline_prefixes (strict — only when every content line is prefixed) recover the raw text before tokenizing; hashline_parse_text is the single-call wrapper. Truncation notices like [Showing lines 1-10 of 42 … Use :L5 are also detected and dropped.

diff and diff_preview — showing what changed

• structured_patch(old, new, context) → a Patch of Hunks — a minimal structured diff used by recovery and previews.
• apply_patch(target, &patch) → Option<String> — applies a structured patch, None when it doesn’t fit cleanly.
• Patch::to_unified_diff(old_path, new_path) — render as classic unified diff.
• build_compact_diff_preview(old, new, &options) → CompactDiffPreview { preview, added_lines, removed_lines } — a bounded preview that truncates unchanged runs (max_unchanged_run, default 2) so a huge edit still produces a reviewable snippet.

stream — chunked formatting

stream_hash_lines(source, &StreamOptions) → Vec<String> — formats a file with hashline line prefixes in bounded chunks (default 200 lines / 64 KiB per chunk, configurable via StreamOptions), so a renderer can stream a large file without holding one giant string.

messages — the exact words

All error and warning strings are centralized as named constants (EMPTY_REPLACE, DELETE_TAKES_NO_BODY, BARE_BODY_AUTO_PIPED_WARNING, the RECOVERY_* banners, HEADTAIL_DRIFT_WARNING, …) plus two builders: missing_snapshot_tag_message(path) and block_unresolved_message(line). The full text of each appears inline in the syntax reference, apply, patcher, and recovery. MISMATCH_CONTEXT (2) sets the context window in mismatch diagnostics. The envelope markers *** Begin Patch / *** End Patch / *** Abort are also defined here. (A few constants — e.g. REPLACE_PAIR_COALESCED_WARNING — are defined but not yet emitted by any code path; treat the constants as the source of truth for wording.)