fix(jj): enable unchanged-context expansion

[samuela]

Problem

Git-backed reviews can load complete file contents and expand context omitted from a patch. The Jujutsu adapter returned only patch text, so the same gaps appeared as static ··· N unchanged lines ··· rows in hunk show and hunk diff.

Approach

• Resolve single-revision JJ reviews to immutable commit IDs before generating the patch, so lazy source reads cannot race a moving @ or bookmark.
• Load complete files with a bounded, path-safe jj file show --ignore-working-copy reader and expose it through the existing VCS extension contract.
• Preserve rename, addition, deletion, source-cache, and custom-executable behavior.
• Keep aggregate multi-revision patches reviewable without attaching a guessed source pair.
• For merges, expose the exact new side used by gap expansion without pretending that any one parent is JJ's virtual merged-parent baseline.

The resulting JJ review uses the existing expandable row and z interaction:

Before: ··· 6 unchanged lines ···
After:  ▾ 6 unchanged lines

And here's some screenshot of it in action:

Validation

• bun test src/extensions/default/vcs/jujutsu/commands.test.ts src/extensions/default/vcs/jujutsu/source.test.ts src/extensions/default/vcs/jujutsu/index.test.ts — 24 passed
• bun run test — 1627 passed, 3 skipped
• bun run typecheck
• bun run format:check
• bun run lint
• bun run deps:check
• git diff --check
• Manually verified hunk show mq in a pure JJ workspace: the previously static gap expanded and collapsed with z.

Tested on NixOS/Linux with JJ 0.41.0. macOS and Windows were not tested manually.

[vercel]

@samuela is attempting to deploy a commit to the Modem Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

PR author is not in the allowed authors list.

[benvinegar]

Awesome thank you!

[samuela]

@benvinegar thanks for the review, i believe i have addressed all your comments, lmkwyt

[benvinegar]

Thanks — the three earlier threads are addressed. I found two additional user-config interactions that still make the new “exact source” path unsafe, plus the Windows CI failure:

1. commands.ts:23,228-259: use an alias-insensitive commit-ID template. JJ template aliases can shadow the bare commit_id keyword. With a valid template-aliases.commit_id that emits an older full commit ID, hunk show @ pins the patch, source reader, and cache key to that older commit. I reproduced the adapter showing an old add-file patch instead of the current @ diff. self.commit_id() bypassed the alias in the same repository.

2. source.ts:45-47,74-86: neutralize fileset and file_show template customization. root-file:<value> is itself shadowable through fileset-aliases; I reproduced root-file:"a.txt" selecting both a.txt and b.txt. Separately, jj file show defaults to user-configured templates.file_show, which can prepend metadata to the source. Together, the reader returned PREFIX:a.txt\none\nPREFIX:b.txt\nbee\n as one file's attested source. A bare JSON-quoted path from repoRoot plus an explicit empty -T template (-T '""') avoided both customizations in my local JJ 0.41.0 test. Please add regression coverage for both configurations.

3. source.test.ts:66: make the integration filename Windows-valid. The double quote in -note [exact] "file".txt is illegal on Windows, and the Windows compatibility job is currently red with ENOENT at the following writeFileSync. The quote case can stay in the mocked argv test; the real filesystem fixture should use only cross-platform-valid metacharacters.

The focused source tests, typecheck, lint, dependency check, and tracked-file format check otherwise pass locally. I also exercised the real branch TUI under tmux against local JJ repositories: z expanded the omitted source correctly, and a nested-cwd show @ -- file.txt selected only sub/file.txt and expanded its source.

This comment was generated by Pi using OpenAI GPT-5.6-sol

[samuela]

done in 9d999ab

[benvinegar]

The requested fixes are addressed, including the Windows fileset regression, and the full CI matrix is green.

This comment was generated by Pi using OpenAI GPT-5.6-sol