feat: highlight injected languages in the diff view

CHANGELOG.md

@@ -8,6 +8,12 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Added
 
+- The diff view highlights code embedded in another language with that language's parser, not only the file's own grammar
+
+## [0.2.0] — 2026-10-09
+
+### Added
+
 - `dw` in the diff or the panel shows a partly staged file's changes since staging, works as a toggle
 - `ds` in the panel or the diff previews the commit: only the staged changes, each diffed against `HEAD`, also works as a toggle
 - `s` and `u` act on a hunk the index holds only part of, and the diff winbar tallies staged and partial hunks

lua/differ/syntax/init.lua

@@ -13,8 +13,16 @@ local M = {}
 local ns = vim.api.nvim_create_namespace("differ.syntax")
 
 -- layering: syntax foreground sits *under* the diff line background (paint
--- uses 100) and word-level spans (200), so the diff state always reads on top
+-- uses 100) and word-level spans (200), so the diff state always reads on top.
+-- an injected layer sits one above its host, capped short of the diff layer, so
+-- embedded code wins where the host also captures the region it was lifted from
 local PRIORITY = 90
+local MAX_DEPTH_BUMP = 9
+
+-- ceiling for parsing injections. the primary tree is still parsed above it, so a
+-- generated file keeps its tag/keyword level highlighting without paying for
+-- injection discovery over the whole tree
+local INJECTION_MAX_BYTES = 512 * 1024
 
 -- resolve a treesitter language for `path`, or nil when there's no filetype, no
 -- mapped language, or the parser isn't installed, in which case the pass is
@@ -39,10 +47,88 @@ local function resolve_lang(path)
     return lang
 end
 
--- parse `text` and collect highlight captures in source coordinates. mirrors the
--- core highlighter: hl group is `@<capture>.<lang>`, captures named `_…` are
--- internal and skipped. multi-line captures are clipped to one entry per line;
--- injections (embedded languages) are deferred for v1, only the primary tree
+-- treesitter handles are typed `any`: there's no nvim type library on the
+-- typecheck path, so naming TSTree/TSNode/Query would only be an undefined alias
+---@class differ.SyntaxLayer
+---@field tree any         -- TSTree
+---@field lang string
+---@field query any        -- vim.treesitter.Query
+---@field priority integer
+
+-- flatten a language tree into paint layers, host before its injections so
+-- embedded code is applied over the region the host lifted it from. one injected
+-- language holds one layer per region, not one per language, so a file with three
+-- `<script>` blocks yields three trees under a single child
+---@param parser any
+---@return differ.SyntaxLayer[]
+local function layers(parser)
+    local out = {}
+    local function walk(ltree, depth)
+        local lang = ltree:lang()
+        local query = vim.treesitter.query.get(lang, "highlights")
+        if query then
+            for _, tree in pairs(ltree:trees()) do
+                out[#out + 1] = {
+                    tree = tree,
+                    lang = lang,
+                    query = query,
+                    priority = PRIORITY + math.min(depth, MAX_DEPTH_BUMP),
+                }
+            end
+        end
+        for _, child in pairs(ltree:children()) do
+            walk(child, depth + 1)
+        end
+    end
+    walk(parser, 0)
+    return out
+end
+
+-- one capture becomes one entry per line it covers: head to EOL, full middle
+-- lines, tail to ecol (an empty tail is skipped)
+---@param out differ.SyntaxCapture[]
+---@param node any
+---@param hl string
+---@param lines string[]
+---@param priority integer
+local function push_capture(out, node, hl, lines, priority)
+    local srow, scol, erow, ecol = node:range()
+    local function push(row, col_start, col_end)
+        out[#out + 1] = {
+            row = row,
+            col_start = col_start,
+            col_end = col_end,
+            hl = hl,
+            priority = priority,
+        }
+    end
+    if srow == erow then
+        push(srow, scol, ecol)
+        return
+    end
+    push(srow, scol, #(lines[srow + 1] or ""))
+    for r = srow + 1, erow - 1 do
+        push(r, 0, #(lines[r + 1] or ""))
+    end
+    if ecol > 0 then
+        push(erow, 0, ecol)
+    end
+end
+
+-- captures named `_…` are internal to the query; spell/nospell carry spell
+-- metadata rather than a highlight group, which is what the core highlighter
+-- reads them as
+---@param name string
+---@return boolean
+local function is_paintable(name)
+    return not vim.startswith(name, "_") and name ~= "spell" and name ~= "nospell"
+end
+
+-- parse `text` and collect highlight captures in source coordinates, injected
+-- languages included. mirrors the core highlighter: the hl group is
+-- `@<capture>.<lang>` of the tree the capture came from, so embedded typescript
+-- in an astro file reads as typescript. injected ranges come back in the outer
+-- string's coordinates, so no translation is needed here or in project
 ---@param text string
 ---@param lang string
 ---@return differ.SyntaxCapture[]
@@ -51,41 +137,45 @@ local function captures_for(text, lang)
     if not ok or not parser then
         return {}
     end
-    local query = vim.treesitter.query.get(lang, "highlights")
-    if not query then
-        return {}
-    end
-    local tree = parser:parse()[1]
-    if not tree then
+    if not pcall(parser.parse, parser, #text <= INJECTION_MAX_BYTES) then
         return {}
     end
 
     local lines = vim.split(text, "\n", { plain = true })
     local out = {}
-    for id, node in query:iter_captures(tree:root(), text, 0, -1) do
-        local name = query.captures[id]
-        if not vim.startswith(name, "_") then
-            local hl = "@" .. name .. "." .. lang
-            local srow, scol, erow, ecol = node:range()
-            if srow == erow then
-                out[#out + 1] = { row = srow, col_start = scol, col_end = ecol, hl = hl }
-            else
-                -- head to EOL, full middle lines, tail to ecol (skip an empty tail)
-                out[#out + 1] =
-                    { row = srow, col_start = scol, col_end = #(lines[srow + 1] or ""), hl = hl }
-                for r = srow + 1, erow - 1 do
-                    out[#out + 1] =
-                        { row = r, col_start = 0, col_end = #(lines[r + 1] or ""), hl = hl }
-                end
-                if ecol > 0 then
-                    out[#out + 1] = { row = erow, col_start = 0, col_end = ecol, hl = hl }
-                end
+    for _, layer in ipairs(layers(parser)) do
+        for id, node in layer.query:iter_captures(layer.tree:root(), text, 0, -1) do
+            local name = layer.query.captures[id]
+            if is_paintable(name) then
+                push_capture(out, node, "@" .. name .. "." .. layer.lang, lines, layer.priority)
             end
         end
     end
     return out
 end
 
+-- a model's old_text/new_text are fixed once built, so a rerender (a layout or
+-- context toggle) reprojects rather than reparsing both sides. weak keys: an
+-- entry falls away with the model it was parsed from
+---@type table<differ.DiffModel, table<string, differ.SyntaxCapture[]>>
+local cache = setmetatable({}, { __mode = "k" })
+
+---@param model differ.DiffModel
+---@param lang string
+---@param side "old"|"new"
+---@return differ.SyntaxCapture[]
+local function side_captures(model, lang, side)
+    local entry = cache[model]
+    if not entry then
+        entry = {}
+        cache[model] = entry
+    end
+    if not entry[side] then
+        entry[side] = captures_for(side == "old" and model.old_text or model.new_text, lang)
+    end
+    return entry[side]
+end
+
 -- apply the syntax pass to one column's buffer: parse the real source(s) the
 -- column draws from (old, new, or both for a unified/stacked column), project the
 -- captures through the map, and paint them as extmarks. no-op when the language
@@ -105,13 +195,13 @@ function M.apply(bufnr, column, model)
     if column.side == "old" or column.side == "unified" then
         vim.list_extend(
             marks,
-            project.project(captures_for(model.old_text, lang), column.map.from_old)
+            project.project(side_captures(model, lang, "old"), column.map.from_old)
         )
     end
     if column.side == "new" or column.side == "unified" then
         vim.list_extend(
             marks,
-            project.project(captures_for(model.new_text, lang), column.map.from_new)
+            project.project(side_captures(model, lang, "new"), column.map.from_new)
         )
     end
 
@@ -121,7 +211,7 @@ function M.apply(bufnr, column, model)
         pcall(vim.api.nvim_buf_set_extmark, bufnr, ns, m.row, m.col_start, {
             end_col = m.col_end,
             hl_group = m.hl,
-            priority = PRIORITY,
+            priority = m.priority,
         })
     end
 end
@@ -131,7 +221,8 @@ end
 -- recorded buffer rows, shifted right by col_offset (the spine + marker prefix). same
 -- derived-buffer rule as apply: never parse the page itself. clears the namespace once,
 -- so one call per repaint covers every snippet; a path with no parser is skipped and
--- its hunk stays plain
+-- its hunk stays plain. a snippet is a fragment, so an injection whose opening
+-- delimiter was cut away simply doesn't fire and that line stays plain
 ---@param bufnr integer
 ---@param snippets { path: string, col_offset: integer, lines: { row: integer, text: string }[] }[]|nil
 function M.apply_snippets(bufnr, snippets)
@@ -155,7 +246,7 @@ function M.apply_snippets(bufnr, snippets)
                         {
                             end_col = c.col_end + s.col_offset,
                             hl_group = c.hl,
-                            priority = PRIORITY,
+                            priority = c.priority,
                         }
                     )
                 end

lua/differ/syntax/project.lua

@@ -12,12 +12,14 @@ local M = {}
 ---@field col_start integer -- byte col, 0-based inclusive
 ---@field col_end integer   -- byte col, 0-based exclusive
 ---@field hl string         -- resolved highlight group
+---@field priority integer  -- host layer, one higher per injection depth
 
 ---@class differ.SyntaxMark
 ---@field row integer       -- 0-based buffer row
 ---@field col_start integer
 ---@field col_end integer
 ---@field hl string
+---@field priority integer
 
 -- project source captures onto buffer rows via `from_map` (source lnum 1-based ->
 -- buffer lnum 1-based). returns extmark specs in buffer coordinates
@@ -34,6 +36,7 @@ function M.project(captures, from_map)
                 col_start = c.col_start,
                 col_end = c.col_end,
                 hl = c.hl,
+                priority = c.priority,
             }
         end
     end

test/nvim/syntax_spec.lua

@@ -16,19 +16,40 @@ local function model(old, new, path)
     })
 end
 
--- syntax extmarks for a buffer: { {row, col, end_col, hl}, ... }
+-- syntax extmarks for a buffer: { {row, col, end_col, hl, priority}, ... }
 local function syntax_marks(bufnr)
     local out = {}
     for _, m in ipairs(vim.api.nvim_buf_get_extmarks(bufnr, ns, 0, -1, { details = true })) do
-        out[#out + 1] = { row = m[2], col = m[3], end_col = m[4].end_col, hl = m[4].hl_group }
+        out[#out + 1] = {
+            row = m[2],
+            col = m[3],
+            end_col = m[4].end_col,
+            hl = m[4].hl_group,
+            priority = m[4].priority,
+        }
     end
     return out
 end
 
+-- the first mark with `hl` on `row` (optionally starting at `col`), or nil
+local function find(marks, row, hl, col)
+    for _, m in ipairs(marks) do
+        if m.row == row and m.hl == hl and (col == nil or m.col == col) then
+            return m
+        end
+    end
+    return nil
+end
+
 -- is there a mark with `hl` on `row` (optionally starting at `col`)?
 local function has(marks, row, hl, col)
+    return find(marks, row, hl, col) ~= nil
+end
+
+-- is there a mark with `hl` anywhere in the buffer?
+local function has_hl(marks, hl)
     for _, m in ipairs(marks) do
-        if m.row == row and m.hl == hl and (col == nil or m.col == col) then
+        if m.hl == hl then
             return true
         end
     end
@@ -89,6 +110,50 @@ describe("syntax pass (split)", function()
     end)
 end)
 
+-- injected languages: a fenced lua block in markdown. both parsers ship with nvim, so
+-- this holds wherever the suite runs; astro/vue/svelte are the same shape, with the bulk
+-- of the file living in injections rather than in the host tree
+describe("syntax pass (injections)", function()
+    -- stacked, whole-file context: "# t"(0), ""(1), "```lua"(2), old(3), new(4), "```"(5)
+    local function md_view()
+        return view("# t\n\n```lua\nlocal x = 1\n```\n", "# t\n\n```lua\nlocal y = 2\n```\n", {
+            path = "x.md",
+        })
+    end
+
+    it("highlights the embedded language, not just the host", function()
+        local v = md_view()
+        v:open()
+        local marks = syntax_marks(v.columns[1].bufnr)
+        assert.is_true(has(marks, 3, "@keyword.lua", 0)) -- `local` on the old line
+        assert.is_true(has(marks, 4, "@keyword.lua", 0)) -- and on the new line
+        assert.is_true(has(marks, 0, "@markup.heading.1.markdown")) -- host tree still paints
+        v:close()
+    end)
+
+    it("layers an injected capture over the host's on the same row", function()
+        local v = md_view()
+        v:open()
+        local marks = syntax_marks(v.columns[1].bufnr)
+        -- markdown captures the whole fence as raw block, lua captures inside it
+        local host = find(marks, 3, "@markup.raw.block.markdown")
+        local injected = find(marks, 3, "@keyword.lua")
+        assert.is_not_nil(host)
+        assert.is_not_nil(injected)
+        assert.is_true(injected.priority > host.priority)
+        v:close()
+    end)
+
+    it("skips spell captures, which carry metadata rather than a highlight", function()
+        local v = md_view()
+        v:open()
+        local marks = syntax_marks(v.columns[1].bufnr)
+        assert.is_false(has_hl(marks, "@spell.markdown"))
+        assert.is_false(has_hl(marks, "@nospell.markdown"))
+        v:close()
+    end)
+end)
+
 -- apply_snippets projects treesitter captures onto arbitrary buffer rows (the overview's
 -- boxed hunk lines), parsed from the stripped source text and shifted right by col_offset
 -- to clear the box spine + inset. same namespace/helpers as the View syntax pass
@@ -122,6 +187,25 @@ describe("apply_snippets (overview hunk syntax)", function()
         assert.is_true(has(marks, 1, "@function.call.lua")) -- keep() on the next snippet row
     end)
 
+    it("projects injected captures too, shifted by col_offset", function()
+        local offset = 3
+        local pad = string.rep(" ", offset)
+        local src = { "```lua", "local x = 1", "```" }
+        local b = scratch({ pad .. src[1], pad .. src[2], pad .. src[3] })
+        syntax.apply_snippets(b, {
+            {
+                path = "x.md",
+                col_offset = offset,
+                lines = {
+                    { text = src[1], row = 0 },
+                    { text = src[2], row = 1 },
+                    { text = src[3], row = 2 },
+                },
+            },
+        })
+        assert.is_true(has(syntax_marks(b), 1, "@keyword.lua", offset))
+    end)
+
     it("is a no-op for a path with no treesitter language", function()
         local b = scratch({ "local x = 1" })
         syntax.apply_snippets(b, {