apply_patch verification fails on canonically-equivalent Unicode — seekSequence needs an NFC pass

[beks-m]

Description

apply_patch fails with Failed to find expected lines in <file> when the file on disk and the model's context lines are canonically-equivalent Unicode but different byte sequences (NFC vs NFD). None of the 4 passes in seekSequence (packages/opencode/src/patch/index.ts:460) handle this:

1. exact
2. rstrip
3. trim
4. normalizeUnicode + trim — but normalizeUnicode (index.ts:418) only maps smart quotes, dashes, ellipsis and NBSP; it never touches combining characters.

So a file containing decomposed й (и U+0438 + U+0306) can never be matched by a patch containing precomposed й (U+0439), even though the strings are canonically equivalent and render identically. Models essentially always emit NFC, so any NFD content on disk makes the file permanently unpatchable via apply_patch — the model re-reads, retries, fails again, and burns tokens in a loop.

We hit this in production on Cyrillic markdown files (tables with rows like | Район | Астана | ... |). Related-but-different reports: #2904 (auto-closed stale), #31422 (formatter drift, closed not-planned), #27282/#11687 (Windows). None cover canonical equivalence.

Proposed fix

A 5th pass in seekSequence, after the current pass 4:

// Pass 5: NFC-normalized match (canonical equivalence — no false positives)
const nfc = tryMatch(
  lines, pattern, startIndex,
  (a, b) => a.normalize("NFC").trim() === b.normalize("NFC").trim(),
  eof,
)
return nfc

This is safe: NFC normalization followed by exact equality only matches strings Unicode defines as the same text. Happy to send a PR with this + tests if you'd take it.

Plugins

None

OpenCode version

1.15.4 (gap still present at current HEAD of packages/opencode/src/patch/index.ts)

Steps to reproduce

Engine-level (deterministic, no model needed):

import { Patch } from "./packages/opencode/src/patch" // deriveNewContentsFromChunks

const fileContent = "# Сводка\nРайон: Астана\n".normalize("NFD") // й stored decomposed
const chunk = {
  old_lines: ["Район: Астана".normalize("NFC")], // models emit NFC
  new_lines: ["Район: Алматы"],
  is_end_of_file: false,
}
// -> throws: Failed to find expected lines

End-to-end:

1. node -e 'require("fs").writeFileSync("test.md", "# Сводка\nРайон: Астана\n".normalize("NFD"))'
2. Ask any model to change Астана to Алматы in test.md using apply_patch.
3. apply_patch verification failed: Failed to find expected lines in test.md — and every retry fails the same way, since re-reading doesn't change the canonical-equivalence mismatch.

Operating System

Linux (Debian-based container)

Terminal

headless (server)

[leizd]

Hi, I'd like to work on this.

I confirmed that the current dev branch still lacks canonical Unicode
normalization in seekSequence(). I'll add a focused NFC matching pass and
regression tests covering NFC/NFD-equivalent text, while ensuring genuinely
different strings do not match.