Problem: after resolving all conflicts, vim.diagnostic.enable(true)
restored diagnostics that were cached while markers were present,
showing errors like "unexpected token end" on clean code.
Solution: call vim.diagnostic.reset() before re-enabling to flush
stale results and let the LSP re-analyze the resolved buffer.
Problem: resolving the last conflict called M.detach(), which cleared
attached_buffers[bufnr]. The TextChanged callback then returned true,
permanently deleting the autocmd. Undo restored conflict markers but
nothing re-highlighted or re-suppressed diagnostics.
Solution: inline the cleanup in refresh() instead of calling detach().
Keep attached_buffers set so the autocmd survives. Re-suppress
diagnostics when conflicts reappear after undo.
Problem: virtual text showed generic "current"/"incoming" labels with
no indication of which branch each side came from.
Solution: extract the branch name from the marker line itself
(e.g. <<<<<<< HEAD, >>>>>>> feature) and display as
"HEAD (current)" / "feature (incoming)".
Problem: LuaLS reports missing-fields errors because the parser builds
ConflictRegion tables incrementally, but the variable is typed as
diffs.ConflictRegion? which expects all required fields at construction.
Solution: type the work-in-progress variable as table? and cast to
diffs.ConflictRegion on insertion into the results array.
Problem: lua-language-server reports duplicate @class definitions for
ConflictKeymaps and ConflictConfig (defined in both init.lua and
conflict.lua), and inject-field errors for the untyped parser table.
Solution: remove duplicate @class annotations from conflict.lua
(init.lua is the canonical source), and annotate the parser's current
variable as diffs.ConflictRegion? so LuaLS knows its shape.
Problem: when git hits a merge conflict, users stare at raw <<<<<<<
markers with broken treesitter and noisy LSP diagnostics. Existing
solutions (git-conflict.nvim) use their own highlighting rather than
integrating with diffs.nvim's color blending pipeline.
Solution: add conflict.lua module that detects <<<<<<</=======/>>>>>>>
markers (with diff3 ||||||| support), highlights ours/theirs/base
regions with blended DiffsConflict* highlight groups, provides
resolution keymaps (doo/dot/dob/don) and navigation (]x/[x),
suppresses diagnostics while markers are present, and auto-detaches
when all conflicts are resolved. Fires DiffsConflictResolved user
event on last resolution.
Problem: diffs:// buffers could trigger spurious LSP diagnostics,
opening multiple diffs from fugitive created redundant splits, and
there was no quick way to close diff windows.
Solution: disable diagnostics on diff buffers, reuse existing
diffs:// windows in the tabpage instead of creating new splits,
and add a buffer-local q keymap to close diff windows.
Problem: users had no config-level way to override computed highlight
groups and had to call nvim_set_hl externally.
Solution: add highlights.overrides table that maps group names to
highlight definitions. Overrides are applied after all computed groups
without default = true, so they always win over both computed defaults
and colorscheme definitions.
Problem: the character-level blend intensity was hardcoded to 0.6,
giving users no way to tune how strongly changed characters stand out
from the line-level background.
Solution: add highlights.blend_alpha config option (number, 0-1,
default 0.6) with type validation and range check.
Problem: DiffsDiff* highlight groups lacked default = true, making them
impossible for colorschemes to override, inconsistent with the fugitive
unified diff groups which already had it.
Solution: add default = true to all four DiffsDiffAdd, DiffsDiffDelete,
DiffsDiffChange, and DiffsDiffText nvim_set_hl calls.
Problem: the vimdoc config example showed max_lines = 500 under vim,
but the actual default in init.lua is 200.
Solution: change the example to match the real default.
Problem: DiffsAddNr/DiffsDeleteNr used raw diffAdded/diffRemoved
foreground and the word-level blended background, instead of matching
the line-level and character-level highlight groups.
Solution: set gutter bg to the line-level blend (DiffsAdd/DiffsDelete)
and fg to the character-level blend (DiffsAddText/DiffsDeleteText).
Problem: word-level diff highlights were too intense at 70% alpha, and
line number backgrounds used the line-level blend instead of matching
the word-level highlights.
Solution: reduce DiffsAddText/DiffsDeleteText blend alpha from 0.7 to
0.6 and use the same blended background for DiffsAddNr/DiffsDeleteNr.
Problem: the header context string (e.g. "function M.setup()") was
parsed in isolation by treesitter, which couldn't recognize "function"
as @keyword.function because the snippet is an incomplete definition
with no body or "end".
Solution: append the already-built new_code lines as trailing context
when parsing the header string, giving treesitter a complete function
definition. Filter captures to row 0 only so body-line captures don't
produce extmarks on the header line.
Problem: hl_eol requires a multiline extmark to extend the background
past end-of-line. The extmark used end_col = line_len on the same
row, so it was single-line and hl_eol was effectively a no-op.
Solution: replace end_col with end_row targeting the next line so
the extmark is multiline and hl_eol takes effect. Split
number_hl_group into a separate extmark to prevent it bleeding to
adjacent lines.
Problem: highlight_treesitter only iterated captures from the base
language tree (trees[1]:root()). When code contained injected
languages (e.g. VimL inside vim.cmd()), those captures were never
read, so injected code got no syntax highlighting in diffs.
Solution: pass true to parse() to trigger injection discovery, then
use parser_obj:for_each_tree() to iterate all trees including
injected ones. Each tree gets its own highlights query looked up by
ltree:lang(), and @spell/@nospell captures are filtered out.
Problem: users who want keybindings must wrap commands in closures.
There is no stable public API for key binding.
Solution: define <Plug> mappings in the plugin file and document them
in a new MAPPINGS section in the vimdoc.
Problem: highlights.context was a plain integer, inconsistent with the
table structure used by treesitter, vim, and intra sub-configs.
Solution: change to { enabled = true, lines = 25 } with full
vim.validate() coverage matching the existing pattern.
Problem: treesitter parses each diff hunk in isolation, so incomplete
syntax constructs at hunk boundaries (e.g., a function definition with
no body) produce ERROR nodes and drop captures.
Solution: read N lines from the on-disk file before/after each hunk and
prepend/append them as unmapped padding lines. The line_map guard in
highlight_treesitter skips extmarks for unmapped lines, so padding
provides syntax context without visual output. Controlled by
highlights.context (default 25, 0 to disable). Also applies to the vim
syntax fallback path via a leading_offset filter.
Problem: highlight_treesitter concatenated all hunk lines (context, -,
+) into a single string. Mixed old/new code produced invalid syntax
(e.g. two return statements), causing treesitter error recovery to drop
captures on lines after the syntax error.
Solution: split hunk lines into two versions — new (context + added)
and old (context + deleted) — each parsed independently. Use a line_map
to resolve treesitter row indices to buffer lines, with the old version
only mapping deleted lines to avoid duplicate extmarks on context.
Also fixes three related issues exposed by the improved TS coverage:
- Replace Normal extmark with DiffsClear (explicit fg from Normal.fg).
Normal in extmarks doesn't reliably override vim :syntax foreground.
- Reorder priority stack to DiffsClear(198) < syntax(199) < line
bg(200) < char bg(201). TS capture groups can carry colorscheme
backgrounds that would override diff line backgrounds at higher
priority.
- Gate DiffsClear on per-line coverage tracking. Only clear fugitive
syntax fg on lines where TS/vim actually produced captures, preventing
force-clearing on lines where error recovery drops captures.
Problem: parse_diffopt() passes linematch from diffopt to byte-level
vim.diff() in char_diff_pair, where each "line" is a single byte.
linematch is meaningless at this granularity.
Solution: copy diff_opts without linematch before passing to byte_diff
in char_diff_pair. linematch remains in effect for the line-level diff
in diff_group_native where it belongs.
Problem: vim.json.encode fails with "excessively sparse array" when
extmark row numbers are used as integer table keys, since they create
gaps in the array.
Solution: use tostring(row) as keys instead. Also add hl_eol to dumped
extmark fields for debugging line background extmarks.
Problem: running :e on a :Gdiff buffer cleared all content because
diffs:// buffers had no BufReadCmd handler. Neovim tried to read the
buffer name as a file path, found nothing on disk, and emptied the
buffer. This affected all three buffer creation paths (gdiff,
gdiff_file, gdiff_section).
Solution: register a BufReadCmd autocmd for diffs://* that parses the
URL and regenerates diff content from git. Change buffer options from
nofile/wipe to nowrite/delete (matching fugitive's approach) so
buffer-local autocmds and variables survive across unload/reload
cycles. Store old filepath as buffer variable for rename support.
'default' inherits algorithm and linematch from diffopt, 'vscode' uses
the FFI library. Removes the need for diffs.nvim to duplicate settings
that users already control globally.
line_hl_group bg occupies a separate rendering channel from hl_group in
Neovim's extmark system, causing character-level bg-only highlights to be
invisible regardless of priority. Switching to hl_group + hl_eol ensures
all backgrounds compete in the same channel.
Also reorders priorities (Normal 198 < line bg 199 < syntax 200 < char
bg 201), bumps char-level blend alpha from 0.4 to 0.7 for visibility,
and adds debug logging throughout the intra pipeline.
