This crate contains code derived from OpenAI Codex‘s codex-rs/apply-patch, licensed under Apache-2.0. Original notice: OpenAI Codex, Copyright 2025 OpenAI. See the repository NOTICE, the repository’s Apache-2.0 LICENSE, and the pinned upstream license.
parser.rs adapts parse_one_hunk and parse_update_file_chunk from parser.rs at d45513ce5ad99d37638f355601f075d55860001e. This is the batch parser immediately before the executor filesystem migration. The newer streaming parser is intentionally not imported.seek_sequence.rs derives from seek_sequence.rs at 624ccf794703e2d84e748fc3ef547d6191a8c0a4.update.rs extracts compute_replacements's PreserveLineEndings behavior from file_update.rs at 624ccf794703e2d84e748fc3ef547d6191a8c0a4.text_file.rs derives from text_file.rs at 624ccf794703e2d84e748fc3ef547d6191a8c0a4.lib.rs / parser.rs derives from UpdateFileChunk at that same current revision. Tests and public adapters are Maka additions.Filesystem access, PathUri resolution, CLI and shell invocation parsing, streaming, networking, diff display, and upstream integration tests are excluded. No returned path is authorized by this crate. Callers must apply their capability checks, per-operation identity checks, pinned file operations, and durable mutation outcome recording.
The parser preserves upstream batch chunk rules, including optional first @@, blank context lines, EOF markers, and fuzzy whitespace / Unicode matching. Maka rejects Move, environment directives, shell wrappers, empty envelopes, empty Add operations and NUL paths. Repeated paths and aliases such as x and ./x retain source order, matching Maka's sequential batch contract. Callers must capture/prepare each operation against the state at that point, not assume distinct spellings identify independent files or prepare every update from the same initial snapshot.
Update preparation always uses upstream‘s line-ending preservation mode. Context keeps its original bytes and endings, new lines use the first existing ending (LF for empty sources), and an unterminated final line gains an ending. Structured OpenAI create diffs remove the single newline forced by envelope Add parsing, matching the SDK’s join-without-forced-newline behavior.
Preparation retains complete source, patch and result strings and parsed lines in memory. Multiple fuzzy passes and Unicode normalization retain upstream's nonlinear worst case. Opaque chunk fields and an overlap/range check protect reconstruction indices.
The integration corpus covers exact and fuzzy matches, context preservation, multiple chunks, EOF, mixed CRLF/LF/CR endings, unterminated sources, Add/Delete, Move rejection, ordered repeated paths, malformed syntax, structured-diff operation injection, large content and ordered operation batches. The extraction is not a full current Codex parser: multi-environment and streaming grammar are explicitly outside the contract.
An external probe also compared eight LF/exact/fuzzy/EOF cases against the unmodified, pinned d45513ce upstream crate and verified that upstream's preview left its snapshot files unchanged. All eight outputs agreed; mixed line endings are intentionally tested against the newer preservation contract instead.