Module apply_patch 

apply_patch: apply a structured multi-file patch atomically.

The patch format is the V4A flavour popularised by Anthropic and used by OpenAI’s Codex CLI: no line numbers, context lines locate the change in the file. This is materially more robust than unified diff for LLM-authored edits — models reliably mis-count line numbers, but they reliably preserve a few lines of context.

*** Begin Patch
*** Update File: src/foo.rs
@@ optional anchor
 context line (unchanged, must match the file exactly)
-line to remove (must match the file exactly)
+line to add
 context line
*** Add File: new/file.rs
+line one
+line two
*** Delete File: doomed/file.rs
*** End Patch

Reliability strategy:

• Strict matching. Context+remove lines must match the live file exactly. No fuzzy matching.
• Single-match. If a hunk’s pattern appears 0 or >1 times in the file, the patch is rejected (the model can add an @@ anchor for disambiguation or include more context).
• Atomic. All hunks are applied in memory first; only after every file’s final state is computed do we touch the disk. A failure in hunk N rolls the whole patch back; nothing on disk changes.
• Sandboxed. Every path goes through resolve_within.
• Useful errors. The tool returns structured diagnostics (which file, which hunk, what was being matched) so the model can self-correct.

Structs

ApplyPatch