feat: collapse generated files marked in .gitattributes

[ceffo]

Stacked on #524. GitHub can't base a cross-fork PR on a branch that doesn't exist here, so the first commit in this PR is #524 — please review that one first, and I'll rebase this to drop it once #524 lands. The second commit (feat: collapse generated files…) is the actual content here.

A review of hand-written code gets buried when a regenerated protobuf or API client lands in the same diff. This recognizes the marker the forges already define and lets the reviewer opt into collapsing those files, and into dropping them from review progress.

Both behaviors are off-by-default in the sense that matters: with no config, tuicr behaves exactly as it does today and pays no detection cost at all — no repository is opened and no attribute is read.

[generated]
collapse = false   # hide the diff body; Space expands one file
count    = true    # count toward the reviewed/total indicator

I want to flag up front that this is unrequested — there's no issue asking for it. I also noticed #230 asked for a built-in generated-file list and was resolved with the user-controlled .tuicrignore instead, which reads as a preference for user-owned declarative config over shipped opinion. This is deliberately in that spirit: the file list comes from the user's .gitattributes, tuicr ships no patterns and no heuristics. If you'd still rather not carry it, #524 stands on its own and I'm happy to have just that one.

Detection

.gitattributes only — linguist-generated (GitHub's Linguist) or gitlab-generated — resolved through libgit2's get_attr. That buys nested .gitattributes, .git/info/attributes, core.attributesFile, and [attr] macros behaving exactly as git check-attr does, for free.

I considered a header regex (// Code generated by …) and rejected it: the marker sits at line 1, outside every hunk, so it needs full file content that PR/MR mode doesn't have.

One trap worth calling out, because it's the kind of thing that silently hides code. GitHub documents linguist-generated=true and GitLab documents bare gitlab-generated. Those land in different AttrValue variants (String("true") vs True), and get_attr returns a sentinel string for a set-but-valueless attribute — git2's own docs warn against interpreting it directly. A naive is_some() check would read -gitlab-generated as an opt-in and hide a file the user explicitly asked to see. The whole variant matrix is tested.

Two rules, both chosen to fail toward showing the diff, because a needlessly shown diff is an annoyance and a needlessly hidden one is unreviewed code:

• An explicit opt-out on either attribute beats an opt-in on the other, so a directory-wide rule can be excepted per file.
• A value neither forge documents counts as not generated.

Why count isn't gated by collapse

Expanding a generated file to peek at it must not add it to the denominator, or review progress would regress as a side effect of looking. Both halves of the fraction drop the file — otherwise marking one reviewed would report 2/1.

count = false alone is also useful without collapsing: label the files, exclude them from progress, still show them.

Per-frame cost

is_file_collapsed runs once per file per frame, so the checks are ordered so the default configuration costs exactly what it does today: the reviewed lookup short-circuits first, and collapse_generated is a bare bool that's false unless the user opted in.

	lookups per file per frame
default	is_file_reviewed — 1, unchanged
opted in	+ generated_files (1 per unreviewed file), + expanded_generated (only for files that are generated)

file_header_prefix_text gains one lookup when opted in, for the [generated] label. Cancelling that out means threading the already-known reviewed state into it — worth doing, but it'd also change the reviewed path, so I'd rather send it separately than bundle it here. Multi-file view already pays two lookups there on main.

Where detection is hooked, and why not at the load sites

Detection runs from rebuild_annotations, not from each self.diff_files = … assignment. There are fifteen of those across pr.rs, commits.rs, and diff_load.rs, and every one of them rebuilds annotations afterwards — so one hook can't drift, whereas a missed assignment site would silently desynchronize the annotations from what's drawn, which is exactly the class of bug #524 exists to prevent.

Probed paths are memoized, and that memo doubles as the staleness test: a repeat call over an unchanged file set costs one hash lookup per file and no libgit2 work. Reload (:e) discards it, so editing .gitattributes mid-session takes effect without a restart.

State lives on App keyed by display path, never by index — reloads, watch ticks, and commit-selection changes replace diff_files wholesale and shift every index. There's a test for that specifically. I didn't add a DiffFile field: it'd touch all 31 struct literals for no benefit, since the expanded-state set has to live on App regardless.

UI

• Space in the diff panel expands the generated file under the cursor. It was unbound there, and already means "expand what's under the cursor" in the file list and the commit selector. Off a collapsed generated file it falls through to the shared handler exactly as before.
• [generated] after the file header rule, shown whether or not the body is collapsed — once expanded, it's the only thing left saying why it was hidden.
• Counter: Files · 0/2 · 2 generated.
• Tree: generated filenames are dimmed. Deliberately not a glyph — ▶ already means "collapsed directory" in that panel, and ▣/▢ stays about reviewed state alone.
• :set generated / :set nogenerated / :set generated! / :generated, mirroring the existing set commits triple. These work even with collapse = false, so the opt-in default stays tolerable.

The decoration is gated on the same condition as detection, not on "was anything ever detected", so :generated twice from a default start lands back exactly where it began. (I got this wrong first time round — the detected set is deliberately retained when the feature is switched off so re-enabling is free, which made the labels and counter survive a toggle-off. There's a round-trip test for it now.)

Limitations, stated plainly

• Needs a git repo. Mercurial and non-colocated jujutsu report nothing as generated rather than failing.
• In PR/MR mode the attributes come from your local checkout, not the PR head — the same tradeoff .tuicrignore already accepts. It reuses the existing local-checkout resolution, which is only set when the checkout matches the PR's target repo, so a foreign checkout can't mis-mark files.
• Collapse is inert in single-file and pristine --all-files views; the counter and label still apply.
• r on an expanded generated file marks it reviewed and re-collapses it, same as any file. Intentional.

Testing

cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test — clean, 1229 passing (30 new). Rebased on 5675284.

Dogfooded per CONTRIBUTING beyond just reviewing the diff: drove the actual TUI against a scratch repo with *.pb.go linguist-generated=true and confirmed collapse, Space expand and re-collapse, :set generated! in both directions, the counter, side-by-side mode, the toggle round-trip returning to the exact starting screen, an opt-out picked up mid-session via :e, and that a config with no [generated] section renders identically to before.

Follow-ups I'd rather not bundle

• The file_header_prefix_text double lookup above.
• reload_diff_files calls file_render_height unguarded, so in single-file view a collapsed current file makes reload clamp the cursor to the file top. Pre-existing for reviewed files; this PR makes it reachable for generated ones. Benign, and fixing it changes reviewed-file behavior, so it doesn't belong here.

🤖 Generated with Claude Code