Missing EOF marker patches still fail when the hunk header has a section heading

[mwolson]

• Aube: 2.6.1, Linux x64 (also 2.2.17, 2.3.0 and 2.6.0)
• Pnpm: 12.8.1
• Repro: https://github.com/mwolson/tmp-aube-issues/tree/main/pnpm-patch-missing-eof-marker-hunk-heading

The missing EOF marker fix from #1515 still fails with error applying hunk #1
when the hunk header has git's section heading after the second @@, such as
@@ -2,4 +2,4 @@ module.exports = 1;. A bare @@ -2,4 +2,4 @@ header works,
and so does the headed hunk when the patch includes
\ No newline at end of file. Pnpm applies all three.

apply_with_eof_context re-renders the parsed patch with diffy's Display
before appending the marker. Diffy 0.5.2 writes a heading back with an extra
space and an extra blank line, so @@ -2,5 +2,5 @@ p becomes:

@@ -2,5 +2,5 @@  p

 q

The blank line re-parses as an empty context line, so the retried hunk can no
longer match. The same round trip fails even for a file that ends in a
newline, so the retry cannot succeed for any headed hunk. The original #1514
repro passed because its hunk starts at line 1, where git writes no heading.

Git writes a heading for most hunks that do not start at line 1, so real
pnpm patch-commit output hits this often. I found it with the
react-native-keyboard-controller@1.21.13 patch in pingdotgg/t3code, which has
53 headed hunks. With the headings stripped, aube 2.6.1 installs it and
preserves the unterminated last lines.

The repro uses a five-line local package with no registry dependencies. Each
install gets fresh isolated storage. It checks a bare-header control, a headed
patch with the canonical marker, and the headed patch without it, with exact
output-byte assertions.

Could the retry append the marker to the original section text instead of the
re-rendered patch, or drop the heading before re-rendering?

[jdx]

Thanks for the detailed report and repro. You diagnosed it correctly: the EOF-marker retry re-rendered the patch through diffy's Display, which writes a hunk's section heading back with an extra space and a blank line. The blank line re-parsed as an empty context line, so any headed hunk failed on retry.

I opened #1678. The retry now writes the @@ headers itself without the heading, then appends the marker. Matching is still exact, so content mismatches are still rejected and explicit newline changes behave as before. It includes a regression test for a headed hunk that doesn't start at line 1 and has no marker. I haven't run your repro repository or the t3code patch against it yet, so a check on your side once it's released would be welcome.

AI-assisted — Tool: Claude Code; model: anthropic/claude-sonnet-5-5; version: unavailable.

[mwolson]

Retested with the release, aube 2.7.0 linux-x64 (2026-10-07), including the t3code patch you asked about:

• Repo: https://github.com/mwolson/tmp-aube-issues
• Case: pnpm-patch-missing-eof-marker-hunk-heading

PNPM_BIN=/path/to/real/pnpm ./repro.sh

It now exits 0: aube applies the plain missing-marker, headed canonical, and headed missing-marker variants, with native pnpm 12.8.1 as the control.

For t3code, I installed pingdotgg/t3code at 0fcd5f9061, which still has the original react-native-keyboard-controller@1.21.13.patch (53 hunks with section headings and no \ No newline at end of file markers). aube install --frozen-lockfile applies the patch and leaves pnpm-lock.yaml unchanged. Current t3code main, now on keyboard-controller 1.22.6, installs as well.

Fix shipped in aube 2.7.0 via #1678. Marking your reply as the answer.