diff --git a/.config/jp/tools/src/git/hunk.rs b/.config/jp/tools/src/git/hunk.rs
index 737326f6..a5240d52 100644
--- a/.config/jp/tools/src/git/hunk.rs
+++ b/.config/jp/tools/src/git/hunk.rs
@@ -61,6 +61,83 @@ pub fn diff_header(diff_stdout: &str) -> Option<&str> {
     diff_stdout.split_once("\n@@ ").map(|(header, _)| header)
 }
 
+/// Parsed counts and start positions extracted from a `@@ ... @@` header.
+#[derive(Debug, Clone, Copy)]
+pub struct HunkCounts {
+    pub old_start: usize,
+    pub old_count: usize,
+    pub new_count: usize,
+}
+
+/// Parse the `-OS,N` and `+Y,M` ranges from a hunk header line.
+///
+/// Accepts both forms with explicit count (`-5,2`) and the implicit
+/// single-line form (`-5`, equivalent to `-5,1`). Anything after the second
+/// `@@` is ignored.
+pub fn parse_hunk_counts(header_line: &str) -> Option<HunkCounts> {
+    let mut parts = header_line.split_whitespace();
+    parts.next()?; // leading `@@`
+    let old_part = parts.next()?.strip_prefix('-')?;
+    let new_part = parts.next()?.strip_prefix('+')?;
+
+    let (old_start, old_count) = parse_range(old_part)?;
+    let (_new_start, new_count) = parse_range(new_part)?;
+
+    Some(HunkCounts {
+        old_start,
+        old_count,
+        new_count,
+    })
+}
+
+fn parse_range(s: &str) -> Option<(usize, usize)> {
+    let mut it = s.split(',');
+    let start: usize = it.next()?.parse().ok()?;
+    let count: usize = it.next().map_or(Some(1), |c| c.parse().ok())?;
+    Some((start, count))
+}
+
+/// Rewrite a hunk's header to use a canonical `+Y` line number.
+///
+/// Hunks emitted by `git diff-files --unified=0` carry `+Y` values that
+/// reflect the cumulative line shift of every preceding unstaged hunk in
+/// the same file. When staging only a subset of those hunks, the `+Y` of
+/// each selected hunk must be recomputed: `git apply --cached
+/// --unidiff-zero` positions changes by `+Y`, so a stale offset places
+/// the patch at the wrong line.
+///
+/// `cumulative_offset` is the net `(additions - removals)` line change
+/// produced by preceding *selected* hunks in the assembled patch, so that
+/// the rewritten `+Y` reflects the source state at the point this hunk
+/// is applied.
+#[allow(clippy::cast_possible_wrap, clippy::cast_sign_loss)]
+pub fn rewrite_hunk_y(hunk: &str, cumulative_offset: isize) -> Option<(String, HunkCounts)> {
+    let (header_line, body) = hunk.split_once('\n')?;
+    let counts = parse_hunk_counts(header_line)?;
+
+    // Canonical `+Y` assuming the hunk applies at its old position to the
+    // source state implied by `cumulative_offset`.
+    let canonical_y = if counts.old_count > 0 {
+        counts.old_start as isize
+    } else {
+        counts.old_start as isize + 1
+    };
+    let new_y = (canonical_y + cumulative_offset).max(0) as usize;
+
+    // Preserve any trailing context after the second `@@`.
+    let trailing = header_line
+        .splitn(4, ' ')
+        .nth(3)
+        .map_or(String::new(), |t| format!(" {t}"));
+
+    let new_header = format!(
+        "@@ -{},{} +{},{} @@{}",
+        counts.old_start, counts.old_count, new_y, counts.new_count, trailing
+    );
+
+    Some((format!("{new_header}\n{body}"), counts))
+}
+
 #[cfg(test)]
 #[path = "hunk_tests.rs"]
 mod tests;
diff --git a/.config/jp/tools/src/git/stage_patch.rs b/.config/jp/tools/src/git/stage_patch.rs
index 2f70a70c..8d39cc1b 100644
--- a/.config/jp/tools/src/git/stage_patch.rs
+++ b/.config/jp/tools/src/git/stage_patch.rs
@@ -6,7 +6,7 @@ use serde_json::{Map, Value};
 
 use super::{
     apply::apply_patch_to_index,
-    hunk::{diff_header, hunk_id, split_hunks},
+    hunk::{diff_header, hunk_id, rewrite_hunk_y, split_hunks},
 };
 use crate::util::{
     OneOrMany, ToolResult,
@@ -189,13 +189,30 @@ fn build_file_patch<R: ProcessRunner>(
 
     // Deduplicate requested IDs and emit hunks in file order to satisfy
     // `git apply`'s monotonic-line requirement for multi-hunk patches.
+    //
+    // Each selected hunk's `+Y` is recomputed: the working-tree diff bakes
+    // in the cumulative line shift of every preceding unstaged hunk, but a
+    // partial stage skips some of those shifts. We track the net effect of
+    // preceding *selected* hunks and rewrite each header accordingly so
+    // `git apply --cached --unidiff-zero` lands the change at the right
+    // line.
     let requested: std::collections::HashSet<&str> =
         requested_ids.iter().map(String::as_str).collect();
-    let selected: Vec<&str> = available_hunks
-        .iter()
-        .filter(|h| requested.contains(hunk_id(h).as_str()))
-        .map(String::as_str)
-        .collect();
+
+    let mut cumulative_offset: isize = 0;
+    let mut selected: Vec<String> = vec![];
+    for hunk in &available_hunks {
+        if !requested.contains(hunk_id(hunk).as_str()) {
+            continue;
+        }
+        let (rewritten, counts) = rewrite_hunk_y(hunk, cumulative_offset)
+            .ok_or_else(|| format!("Malformed hunk header in diff: {hunk}"))?;
+        #[allow(clippy::cast_possible_wrap)]
+        {
+            cumulative_offset += counts.new_count as isize - counts.old_count as isize;
+        }
+        selected.push(rewritten);
+    }
 
     // Ensure the patch ends with a newline. Hunks emitted by `split_hunks`
     // may have lost their trailing newline during splitting (the newline
diff --git a/.config/jp/tools/src/git/stage_patch_lines.rs b/.config/jp/tools/src/git/stage_patch_lines.rs
index ca2b37d1..91105de8 100644
--- a/.config/jp/tools/src/git/stage_patch_lines.rs
+++ b/.config/jp/tools/src/git/stage_patch_lines.rs
@@ -56,10 +56,13 @@ enum DiffLineKind {
 }
 
 /// Parsed hunk header coordinates.
+///
+/// Only `old_start` is tracked: `new_start` from the original full-diff hunk
+/// header is intentionally discarded, since `build_sub_hunk` recomputes the
+/// `+Y` line number from scratch (see the offset-correction comment there).
 #[derive(Debug)]
 struct HunkHeader {
     old_start: usize,
-    new_start: usize,
 }
 
 fn parse_hunk_header(header: &str) -> Result<HunkHeader, String> {
@@ -68,20 +71,11 @@ fn parse_hunk_header(header: &str) -> Result<HunkHeader, String> {
     let old_part = parts
         .get(1)
         .ok_or("Invalid hunk header: missing old range")?;
-    let new_part = parts
-        .get(2)
-        .ok_or("Invalid hunk header: missing new range")?;
 
     let old_range = old_part.trim_start_matches('-');
     let (old_start, _old_count) = parse_range(old_range)?;
 
-    let new_range = new_part.trim_start_matches('+');
-    let (new_start, _new_count) = parse_range(new_range)?;
-
-    Ok(HunkHeader {
-        old_start,
-        new_start,
-    })
+    Ok(HunkHeader { old_start })
 }
 
 fn parse_range(range: &str) -> Result<(usize, usize), String> {
@@ -239,13 +233,26 @@ fn build_sub_hunk(hunk: &str, selected: &[usize]) -> Result<String, String> {
         body.push('\n');
     }
 
-    // Preserve the offset between old and new positions from the original
-    // hunk header. This is critical for pure insertions where new_start
-    // differs from old_start (e.g. @@ -3,0 +4,1 @@ inserts AFTER line 3).
-    #[allow(clippy::cast_possible_wrap, clippy::cast_sign_loss)]
-    let new_start = {
-        let offset = header.new_start as isize - header.old_start as isize;
-        (old_start as isize + offset).max(0) as usize
+    // Compute `new_start` from the sub-hunk's own contents, NOT by
+    // preserving the offset of the original full-diff hunk header.
+    //
+    // The header we received comes from `git diff-files --unified=0` against
+    // the entire working tree, so its `+Y` reflects the cumulative line
+    // shift of every unstaged hunk earlier in the file. Applying a sub-hunk
+    // in isolation against a clean index via `git apply --cached
+    // --unidiff-zero` uses `+Y` to position the change — carrying that
+    // cumulative offset over would land the patch at the wrong line
+    // whenever other unstaged hunks (not part of this staging) shifted
+    // line numbers.
+    let new_start = if found_removal {
+        // Replacement or pure removal: new content starts at the same
+        // position as the first removed line.
+        old_start
+    } else {
+        // Pure insertion: the first new line lands at `old_start + 1`. For
+        // `old_start = 0` (insertion at file start) this naturally yields
+        // `1`.
+        old_start.saturating_add(1)
     };
 
     let hunk_header = format!("@@ -{old_start},{removals} +{new_start},{additions} @@");
diff --git a/.config/jp/tools/src/git/stage_patch_lines_tests.rs b/.config/jp/tools/src/git/stage_patch_lines_tests.rs
index 70482297..61a601ad 100644
--- a/.config/jp/tools/src/git/stage_patch_lines_tests.rs
+++ b/.config/jp/tools/src/git/stage_patch_lines_tests.rs
@@ -7,28 +7,24 @@ use crate::util::runner::MockProcessRunner;
 fn parse_hunk_header_simple() {
     let h = parse_hunk_header("@@ -5 +5 @@").unwrap();
     assert_eq!(h.old_start, 5);
-    assert_eq!(h.new_start, 5);
 }
 
 #[test]
 fn parse_hunk_header_with_counts() {
     let h = parse_hunk_header("@@ -5,3 +5,2 @@").unwrap();
     assert_eq!(h.old_start, 5);
-    assert_eq!(h.new_start, 5);
 }
 
 #[test]
 fn parse_hunk_header_zero_count() {
     let h = parse_hunk_header("@@ -5,0 +5,3 @@").unwrap();
     assert_eq!(h.old_start, 5);
-    assert_eq!(h.new_start, 5);
 }
 
 #[test]
 fn parse_hunk_header_different_old_new_start() {
     let h = parse_hunk_header("@@ -3,0 +4,1 @@").unwrap();
     assert_eq!(h.old_start, 3);
-    assert_eq!(h.new_start, 4);
 }
 
 #[test]
diff --git a/.config/jp/tools/tests/git_tests.rs b/.config/jp/tools/tests/git_tests.rs
index ab8e6ad9..587ce09d 100644
--- a/.config/jp/tools/tests/git_tests.rs
+++ b/.config/jp/tools/tests/git_tests.rs
@@ -125,6 +125,16 @@ fn commit_then_modify(root: &Utf8Path, path: &str, original: &str, modified: &st
     fs::write(root.join(path), modified).unwrap();
 }
 
+/// 20-line fixture used by the offset-correction regression tests.
+fn lines_1_to_20() -> String {
+    let mut s = String::new();
+    for i in 1..=20 {
+        use std::fmt::Write as _;
+        writeln!(s, "line{i}").unwrap();
+    }
+    s
+}
+
 /// Extract content-addressed patch IDs from a `git_list_patches` XML response.
 fn extract_patch_ids(xml: &str) -> Vec<String> {
     xml.lines()
@@ -1341,6 +1351,103 @@ async fn sequential_staging_across_tools() {
     assert_eq!(staged_content(&root, "mix.rs"), "A\nb\nc\nd\nE\n");
 }
 
+#[tokio::test]
+async fn stage_patch_addition_with_unrelated_unstaged_removal() {
+    if !has_git() {
+        return;
+    }
+
+    let (_dir, root) = init_repo();
+    let original = lines_1_to_20();
+    let modified: String = {
+        let mut lines: Vec<String> = original.lines().map(String::from).collect();
+        for _ in 0..9 {
+            lines.remove(2);
+        }
+        lines.insert(9, "INSERTED".to_string());
+        lines.join("\n") + "\n"
+    };
+
+    fs::write(root.join("sp.rs"), &original).unwrap();
+    git(&root, &["add", "sp.rs"]);
+    git(&root, &["commit", "-m", "init sp"]);
+    fs::write(root.join("sp.rs"), &modified).unwrap();
+
+    let ids = patch_ids(&root, "sp.rs").await;
+    assert_eq!(ids.len(), 2, "expected 2 hunks, got {}", ids.len());
+
+    // Stage only the addition hunk (the second one).
+    run_ok(
+        ctx(&root),
+        tool_with_answers(
+            "git_stage_patch",
+            &json!({"patches": [{"path": "sp.rs", "ids": [&ids[1]]}]}),
+            &json!({"stage_changes": true}),
+        ),
+    )
+    .await;
+
+    let expected: String = {
+        let mut lines: Vec<String> = original.lines().map(String::from).collect();
+        lines.insert(18, "INSERTED".to_string());
+        lines.join("\n") + "\n"
+    };
+    assert_eq!(staged_content(&root, "sp.rs"), expected);
+}
+
+#[tokio::test]
+async fn stage_patch_lines_addition_with_unrelated_unstaged_removal() {
+    if !has_git() {
+        return;
+    }
+
+    let (_dir, root) = init_repo();
+
+    // Original file: 20 lines.
+    let original = lines_1_to_20();
+
+    // Working tree: remove lines 3-11 (9 lines) AND insert "INSERTED" after
+    // old line 18. Two distant unstaged hunks in a single file.
+    let modified: String = {
+        let mut lines: Vec<String> = original.lines().map(String::from).collect();
+        for _ in 0..9 {
+            lines.remove(2);
+        }
+        // After removals: indices 0..=10 hold line1, line2, line12..line20.
+        // Old line 18 is now at index 8. Insert "INSERTED" right after it.
+        lines.insert(9, "INSERTED".to_string());
+        lines.join("\n") + "\n"
+    };
+
+    fs::write(root.join("two_hunk.rs"), &original).unwrap();
+    git(&root, &["add", "two_hunk.rs"]);
+    git(&root, &["commit", "-m", "init two_hunk"]);
+    fs::write(root.join("two_hunk.rs"), &modified).unwrap();
+
+    // Expect two distinct hunks in the listing.
+    let ids = patch_ids(&root, "two_hunk.rs").await;
+    assert_eq!(ids.len(), 2, "expected 2 hunks, got {}", ids.len());
+
+    // Stage just the insertion hunk's only line.
+    run_ok(
+        ctx(&root),
+        tool(
+            "git_stage_patch_lines",
+            &json!({"path": "two_hunk.rs", "patch_id": &ids[1], "lines": [0]}),
+        ),
+    )
+    .await;
+
+    // The index should match HEAD plus exactly one inserted line after
+    // old line 18 — the unrelated removal must remain unstaged.
+    let expected: String = {
+        let mut lines: Vec<String> = original.lines().map(String::from).collect();
+        lines.insert(18, "INSERTED".to_string());
+        lines.join("\n") + "\n"
+    };
+    assert_eq!(staged_content(&root, "two_hunk.rs"), expected);
+}
+
 #[tokio::test]
 async fn stage_patch_multiple_files_single_call() {
     if !has_git() {