feat: author ${VAR} references, gated on whether the client expands them (#76)

The blocker on this issue was whether BCC or the client does the expanding.
Answer, from Anthropic's docs: Claude Code expands ${VAR} and
${VAR:-default} itself, in command, args, env, url and headers, for both
project .mcp.json and user-scope ~/.claude.json. Claude Desktop has no
documented support.

So this is a per-client capability, not a global one, and BCC does NOT
expand on write: resolving a reference into the file would put the secret
back on disk -- the whole thing the user is avoiding -- and would defeat a
feature the client already implements correctly. BCC authors, validates and
warns; expand_env_refs exists to preview what the client will do.

Semantics mirror the documented ones exactly, including the unusual bit:
an unset variable with no default is left as literal ${VAR} text rather
than blanked, because that is what Claude Code passes through.

Gating uses the existing profile_targets_claude_desktop(), so a config that
is correct under Claude Code and broken under Desktop is reported against
whichever profile is actually loaded. The two warnings are worded
differently on purpose -- 'this client will never expand these' is a
different problem from 'this variable looks unset here'.

Two existing behaviours were backwards for this feature and are fixed:

- Secret masking hid placeholders. is_secret_key('API_KEY') is true, so
  ${API_KEY} rendered as dots -- making a reference indistinguishable from
  a stored credential, which is the one distinction that makes the feature
  worth adopting. should_mask_value() now skips references, in the table
  delegate, _redact_server_data and redact_args alike.
- args_secret_warning fired on placeholders. Moving a token into ${VAR} is
  the recommended fix for that warning; continuing to warn punished the
  fix. It now skips references while still flagging a real secret that
  follows one.

Real secrets are still masked everywhere they were before -- asserted, not
assumed.

Refs #76

bcc.py

@@ -19,6 +19,7 @@ from typing import ClassVar
 from PySide6.QtCore import QRect, QSettings, QSize, Qt, QThread, QTimer, QUrl, Signal
 from PySide6.QtGui import (
     QAction,
+    QActionGroup,
     QColor,
     QCursor,
     QDesktopServices,
@@ -26,6 +27,7 @@ from PySide6.QtGui import (
     QIcon,
     QKeySequence,
     QPainter,
+    QPalette,
     QPixmap,
 )
 from PySide6.QtWidgets import (
@@ -66,89 +68,122 @@ import bcc_core as core
 # thread during drag-and-drop import, so skip anything larger than this.
 MAX_DROP_IMPORT_BYTES = 5 * 1024 * 1024  # 5 MB
 
-# --- One-line rebrand: change this to recolor the whole app --------------- #
-ACCENT = "#f97316"  # warm orange
-ACCENT_DIM = "#c2570b"
-BG = "#1b1d23"
-PANEL = "#23262e"
-PANEL_2 = "#2b2f39"
-TEXT = "#e7e9ee"
-MUTED = "#9aa0ad"
-BORDER = "#3a3f4b"
-GOOD = "#4ade80"
-BAD = "#f87171"
-WARN = "#fbbf24"
+# --- Theming (issue #75) -------------------------------------------------- #
+# The palette lives in bcc_core (testable without a Qt app); these module-level
+# names are rebound by `apply_palette()` whenever the theme changes.
+#
+# Why globals rather than passing a palette around: ~20 inline
+# `setStyleSheet(f"color: {MUTED}")` calls are scattered through this file, and
+# an f-string resolves its names when it runs, not when it's compiled. Rebinding
+# the globals means every one of those call sites picks up the new colour on its
+# next render, with no change to the call sites themselves.
+PALETTE = core.DARK_PALETTE
+ACCENT = ACCENT_DIM = BG = PANEL = PANEL_2 = TEXT = MUTED = BORDER = ""
+GOOD = BAD = WARN = REMOTE = ON_ACCENT = DISABLED_BG = MONO_BG = SEL_TEXT = ""
+STATUS_COLORS: dict[str, str] = {}
+HEALTH_COLORS: dict[str, str] = {}
 
-STATUS_COLORS = {"ok": GOOD, "missing": BAD, "warn": WARN, "remote": "#60a5fa", "unknown": WARN}
-STATUS_GLYPH = {"ok": "●", "missing": "●", "warn": "▲", "remote": "◆", "unknown": "○"}
+STATUS_GLYPH = {
+    "ok": "\u25cf",
+    "missing": "\u25cf",
+    "warn": "\u25b2",
+    "remote": "\u25c6",
+    "unknown": "\u25cb",
+}
 
 # Health dot (spawn-test outcome, see core.HealthStatus) shown per row in the
 # server tables' "Health" column -- distinct from the PATH-dependency Status
 # column above.
-HEALTH_COLORS = {"ok": GOOD, "failed": BAD, "untested": MUTED}
-HEALTH_GLYPH = {"ok": "●", "failed": "●", "untested": "○"}
+HEALTH_GLYPH = {"ok": "\u25cf", "failed": "\u25cf", "untested": "\u25cb"}
 
-STYLESHEET = f"""
+
+def build_stylesheet(p: core.Palette) -> str:
+    """Render the global QSS for a palette."""
+    return f"""
 /* No font-family here on purpose: Qt already uses the native system UI font
    on every platform (San Francisco / Segoe UI / desktop default). Naming
    web-CSS aliases like -apple-system forces a costly font-alias scan. */
-* {{ font-size: 13px; color: {TEXT}; }}
-QMainWindow, QDialog {{ background: {BG}; }}
+* {{ font-size: 13px; color: {p.text}; }}
+QMainWindow, QDialog {{ background: {p.bg}; }}
 QLabel#h1 {{ font-size: 15px; font-weight: 600; }}
-QLabel#muted {{ color: {MUTED}; }}
-QFrame#card {{ background: {PANEL}; border: 1px solid {BORDER}; border-radius: 10px; }}
+QLabel#muted {{ color: {p.muted}; }}
+QFrame#card {{ background: {p.panel}; border: 1px solid {p.border}; border-radius: 10px; }}
 QLineEdit, QPlainTextEdit, QComboBox {{
-    background: {PANEL_2}; border: 1px solid {BORDER}; border-radius: 7px;
-    padding: 6px 8px; selection-background-color: {ACCENT}; selection-color: #1a1205;
+    background: {p.panel_2}; border: 1px solid {p.border}; border-radius: 7px;
+    padding: 6px 8px; selection-background-color: {p.accent}; selection-color: {p.on_accent};
 }}
-QLineEdit:focus, QPlainTextEdit:focus, QComboBox:focus {{ border: 1px solid {ACCENT}; }}
+QLineEdit:focus, QPlainTextEdit:focus, QComboBox:focus {{ border: 1px solid {p.accent}; }}
 QComboBox::drop-down {{ border: none; width: 22px; }}
-QComboBox QAbstractItemView {{ background: {PANEL_2}; border: 1px solid {BORDER};
-    selection-background-color: {ACCENT}; outline: none; }}
-QPushButton {{ background: {PANEL_2}; border: 1px solid {BORDER}; border-radius: 7px;
+QComboBox QAbstractItemView {{ background: {p.panel_2}; border: 1px solid {p.border};
+    selection-background-color: {p.accent}; outline: none; }}
+QPushButton {{ background: {p.panel_2}; border: 1px solid {p.border}; border-radius: 7px;
     padding: 7px 13px; }}
-QPushButton:hover {{ border: 1px solid {ACCENT}; }}
-QPushButton:disabled {{ color: {MUTED}; background: {PANEL}; }}
-QPushButton#primary {{ background: {ACCENT}; border: 1px solid {ACCENT}; color: #1a1205; font-weight: 600; }}
-QPushButton#primary:hover {{ background: {ACCENT_DIM}; }}
-QPushButton#primary:disabled {{ background: {PANEL}; color: {MUTED}; border: 1px solid {BORDER}; }}
-QPushButton#danger:hover {{ border: 1px solid {BAD}; color: {BAD}; }}
-QTableWidget {{ background: {PANEL}; border: 1px solid {BORDER}; border-radius: 10px;
+QPushButton:hover {{ border: 1px solid {p.accent}; }}
+QPushButton:disabled {{ color: {p.muted}; background: {p.panel}; }}
+QPushButton#primary {{ background: {p.accent}; border: 1px solid {p.accent}; color: {p.on_accent}; font-weight: 600; }}
+QPushButton#primary:hover {{ background: {p.accent_dim}; }}
+QPushButton#primary:disabled {{ background: {p.panel}; color: {p.muted}; border: 1px solid {p.border}; }}
+QPushButton#danger:hover {{ border: 1px solid {p.bad}; color: {p.bad}; }}
+QTableWidget {{ background: {p.panel}; border: 1px solid {p.border}; border-radius: 10px;
     gridline-color: transparent; outline: none; }}
 QTableWidget::item {{ padding: 6px 8px; border: none; }}
-QTableWidget::item:selected {{ background: {ACCENT}; color: #1a1205; }}
+QTableWidget::item:selected {{ background: {p.accent}; color: {p.on_accent}; }}
 /* Inline cell editors: the global QLineEdit padding/radius clips the text
    inside a table row, so give editors a compact, flat style instead. */
 QTableWidget QLineEdit {{
-    background: {PANEL_2}; color: {TEXT}; border: 1px solid {ACCENT};
+    background: {p.panel_2}; color: {p.text}; border: 1px solid {p.accent};
     border-radius: 3px; padding: 0px 4px; margin: 0px;
-    selection-background-color: {ACCENT_DIM}; selection-color: #ffffff;
+    selection-background-color: {p.accent_dim}; selection-color: {p.selection_text};
 }}
-QHeaderView::section {{ background: {PANEL}; color: {MUTED}; border: none;
-    border-bottom: 1px solid {BORDER}; padding: 8px; font-weight: 600; }}
+QHeaderView::section {{ background: {p.panel}; color: {p.muted}; border: none;
+    border-bottom: 1px solid {p.border}; padding: 8px; font-weight: 600; }}
 QScrollBar:vertical {{ background: transparent; width: 10px; margin: 2px; }}
-QScrollBar::handle:vertical {{ background: {BORDER}; border-radius: 5px; min-height: 24px; }}
+QScrollBar::handle:vertical {{ background: {p.border}; border-radius: 5px; min-height: 24px; }}
 QScrollBar::add-line, QScrollBar::sub-line {{ height: 0; }}
-QLabel#statusbar {{ color: {MUTED}; padding: 4px 2px; }}
-QLabel#warnBanner {{ color: #1a1205; background: {WARN}; border-radius: 8px; padding: 8px 10px; font-weight: 600; }}
-QFrame#noticeBanner {{ background: {PANEL_2}; border: 1px solid {ACCENT}; border-radius: 8px; }}
-QLabel#noticeText {{ color: {TEXT}; }}
-QPushButton#noticeClose {{ background: transparent; border: none; color: {MUTED}; font-size: 14px; padding: 2px; }}
-QPushButton#noticeClose:hover {{ color: {TEXT}; }}
-QLabel#section {{ color: {MUTED}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
-QLabel#sectionDisabled {{ color: {MUTED}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
-QLabel#placeholder {{ color: {MUTED}; padding: 12px; background: {PANEL_2}; border: 1px dashed {BORDER}; border-radius: 8px; }}
-QTableWidget#disabledTable {{ background: #202229; }}
-QTableWidget#disabledTable::item:selected {{ background: {ACCENT}; color: #1a1205; }}
+QLabel#statusbar {{ color: {p.muted}; padding: 4px 2px; }}
+QLabel#warnBanner {{ color: {p.on_accent}; background: {p.warn}; border-radius: 8px; padding: 8px 10px; font-weight: 600; }}
+QFrame#noticeBanner {{ background: {p.panel_2}; border: 1px solid {p.accent}; border-radius: 8px; }}
+QLabel#noticeText {{ color: {p.text}; }}
+QPushButton#noticeClose {{ background: transparent; border: none; color: {p.muted}; font-size: 14px; padding: 2px; }}
+QPushButton#noticeClose:hover {{ color: {p.text}; }}
+QLabel#section {{ color: {p.muted}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
+QLabel#sectionDisabled {{ color: {p.muted}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
+QLabel#placeholder {{ color: {p.muted}; padding: 12px; background: {p.panel_2}; border: 1px dashed {p.border}; border-radius: 8px; }}
+QTableWidget#disabledTable {{ background: {p.disabled_bg}; }}
+QTableWidget#disabledTable::item:selected {{ background: {p.accent}; color: {p.on_accent}; }}
 QPlainTextEdit#diag {{ font-family: "Menlo", "Cascadia Code", "Consolas", "DejaVu Sans Mono", monospace;
-    font-size: 12px; background: #16181d; border: 1px solid {BORDER}; border-radius: 8px; }}
+    font-size: 12px; background: {p.mono_bg}; border: 1px solid {p.border}; border-radius: 8px; }}
 QFrame#diagCard {{ background: transparent; border: none; }}
 QSplitter::handle {{ background: transparent; }}
-QSplitter::handle:hover {{ background: {BORDER}; border-radius: 4px; }}
-QSplitter::handle:pressed {{ background: {ACCENT}; border-radius: 4px; }}
+QSplitter::handle:hover {{ background: {p.border}; border-radius: 4px; }}
+QSplitter::handle:pressed {{ background: {p.accent}; border-radius: 4px; }}
 """
 
 
+def apply_palette(p: core.Palette) -> str:
+    """Rebind the module-level colour names to `p` and return its stylesheet."""
+    global PALETTE, ACCENT, ACCENT_DIM, BG, PANEL, PANEL_2, TEXT, MUTED, BORDER
+    global GOOD, BAD, WARN, REMOTE, ON_ACCENT, DISABLED_BG, MONO_BG, SEL_TEXT
+    global STATUS_COLORS, HEALTH_COLORS
+    PALETTE = p
+    ACCENT, ACCENT_DIM = p.accent, p.accent_dim
+    BG, PANEL, PANEL_2 = p.bg, p.panel, p.panel_2
+    TEXT, MUTED, BORDER = p.text, p.muted, p.border
+    GOOD, BAD, WARN, REMOTE = p.good, p.bad, p.warn, p.remote
+    ON_ACCENT, DISABLED_BG, MONO_BG, SEL_TEXT = (
+        p.on_accent,
+        p.disabled_bg,
+        p.mono_bg,
+        p.selection_text,
+    )
+    STATUS_COLORS = {"ok": GOOD, "missing": BAD, "warn": WARN, "remote": REMOTE, "unknown": WARN}
+    HEALTH_COLORS = {"ok": GOOD, "failed": BAD, "untested": MUTED}
+    return build_stylesheet(p)
+
+
+STYLESHEET = apply_palette(core.DARK_PALETTE)
+
+
 # --------------------------------------------------------------------------- #
 # Background reachability tester (keeps the UI responsive during the request)
 # --------------------------------------------------------------------------- #
@@ -233,7 +268,7 @@ class _SecretMaskDelegate(QStyledItemDelegate):
         if self.revealed or not option.text:
             return
         key_item = self._table.item(index.row(), 0)
-        if key_item and core.is_secret_key(key_item.text()):
+        if key_item and core.should_mask_value(key_item.text(), option.text):
             option.text = core.MASK
 
 
@@ -1643,6 +1678,22 @@ class MainWindow(QMainWindow):
 
     # --- menu bar ---------------------------------------------------------- #
     def _build_menu_bar(self):
+        view_menu = self.menuBar().addMenu("&View")
+        theme_menu = view_menu.addMenu("Theme")
+        self._theme_group = QActionGroup(self)
+        self._theme_group.setExclusive(True)
+        current = stored_theme_setting()
+        for setting, label in (
+            (core.THEME_SYSTEM, "Match system"),
+            (core.THEME_LIGHT, "Light"),
+            (core.THEME_DARK, "Dark"),
+        ):
+            act = QAction(label, self, checkable=True)
+            act.setChecked(setting == current)
+            act.triggered.connect(lambda _checked=False, s=setting: self._set_theme(s))
+            self._theme_group.addAction(act)
+            theme_menu.addAction(act)
+
         help_menu = self.menuBar().addMenu("&Help")
 
         # "Check for updates" used to exist only as a button inside the About
@@ -1697,6 +1748,26 @@ class MainWindow(QMainWindow):
             self.update_banner.hide()
             self.status.setText(f"You're up to date ({core.__version__}).")
 
+    def _set_theme(self, setting: str):
+        """Persist the theme choice and repaint the running window."""
+        QSettings("BCC", "BetterClaudeConfig").setValue("ui/theme", setting)
+        app = QApplication.instance()
+        if app is None:  # pragma: no cover - only in a headless test harness
+            return
+        app.setStyleSheet(theme_stylesheet_for(app, setting))
+        # The global stylesheet covers most of the UI, but the inline
+        # setStyleSheet calls (status dots, warning labels, update banner) only
+        # pick up the new palette when their widget next renders -- so re-render
+        # them now rather than leaving dark-on-light text behind.
+        self._repaint_themed_widgets()
+
+    def _repaint_themed_widgets(self):
+        """Re-run the inline-styled bits after a palette change."""
+        self.status.setStyleSheet(f"color: {MUTED};")
+        idx = self._current_index()
+        self._refresh_tables(select_index=idx if idx >= 0 else -1)
+        self._update_status(saved=False)
+
     def _show_about(self):
         AboutDialog(self).exec()
 
@@ -2045,9 +2116,21 @@ class MainWindow(QMainWindow):
                 return
             self.full_config = cfg
             repaired = True
+        # extract_servers tolerates malformed entries rather than raising (#72),
+        # but keep it inside the guard: a load failure must leave the previously
+        # loaded profile intact instead of half-swapping the window's state.
+        try:
+            servers = core.extract_servers(self.full_config)
+        except Exception as exc:  # pragma: no cover - defence in depth
+            QMessageBox.critical(
+                self,
+                "Could not read config",
+                f"{profile.path}\n\nThe server list couldn't be read: {exc}",
+            )
+            return
         self._loaded_stat = core.config_fingerprint(profile.path)
         self.current_profile = profile
-        self.servers = core.extract_servers(self.full_config)
+        self.servers = servers
         self.dirty = False
         self.restart_btn.hide()
         self._undo_stack.clear()
@@ -2370,7 +2453,7 @@ class MainWindow(QMainWindow):
         entry = self.servers[idx]
         old_name = entry.name
         entry.name = self.editor.current_name()
-        entry.data = self.editor.dump_data()
+        entry.set_data(self.editor.dump_data())
         # The server stays in its section (enable state unchanged), so update
         # its existing row in place rather than re-rendering.
         # An edit invalidates any cached "Test all" result -- the server that
@@ -2464,7 +2547,7 @@ class MainWindow(QMainWindow):
                 QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
             )
             if ans == QMessageBox.StandardButton.Yes:
-                self.servers[existing[name]].data = data
+                self.servers[existing[name]].set_data(data)
                 return False, True
             name = core.resolve_name_collision(name, {s.name for s in self.servers})
         self.servers.append(core.ServerEntry(name, data, True))
@@ -2539,6 +2622,12 @@ class MainWindow(QMainWindow):
             self.save_btn.setEnabled(False)
             return False
         lint_warnings = core.lint_servers(self.servers)
+        # ${VAR} references are only meaningful if the target client expands
+        # them -- Claude Desktop doesn't, so the same config is fine in one
+        # profile and broken in another (#76). Report against the loaded one.
+        for entry in self.servers:
+            for warning in core.env_ref_warnings(entry.data, self.current_profile):
+                lint_warnings.append(f"'{entry.name}': {warning}")
         if lint_warnings:
             self.validation_lbl.setText(f"⚠ {lint_warnings[0]}")
             self.validation_lbl.setStyleSheet(f"color: {WARN};")
@@ -2584,6 +2673,11 @@ class MainWindow(QMainWindow):
                 except Exception as e:
                     QMessageBox.critical(self, "Reload failed", str(e))
                     return
+                # The reload above is the on-disk truth for everything the user
+                # didn't touch -- but it also wipes BCC-authored keys the user
+                # changed in this session (named sets), which apply_servers
+                # doesn't write. Carry them over before saving (#73).
+                contested = core.carry_owned_keys(self.full_config, fresh)
                 core.apply_servers(fresh, self.servers)
                 try:
                     backup = core.write_config(self.current_profile.path, fresh)
@@ -2596,8 +2690,13 @@ class MainWindow(QMainWindow):
                 self.dirty = False
                 self.save_btn.setEnabled(False)
                 bnote = f"  ·  backup: {backup.name}" if backup else "  ·  (new file)"
+                cnote = (
+                    f"  ·  kept your {', '.join(contested)} (the file on disk had a different copy)"
+                    if contested
+                    else ""
+                )
                 self.status.setText(
-                    f"Merged & saved {self.current_profile.path}{bnote}"
+                    f"Merged & saved {self.current_profile.path}{bnote}{cnote}"
                     f"  ·  Restart {self.current_profile.label} to apply."
                 )
                 self._offer_restart_button()
@@ -2755,6 +2854,33 @@ class MainWindow(QMainWindow):
         e.accept()
 
 
+def system_is_dark(app: QApplication) -> bool:
+    """Whether the desktop is currently using a dark appearance.
+
+    Read from the style's own window colour rather than per-platform APIs --
+    Qt has already resolved the OS appearance by the time it builds the
+    default palette, so this works the same on all three platforms.
+    """
+    try:
+        return app.palette().color(QPalette.ColorRole.Window).lightness() < 128
+    except Exception:  # pragma: no cover - defensive; never block startup on theming
+        return True
+
+
+def stored_theme_setting() -> str:
+    """The user's theme choice, defaulting to following the system."""
+    value = QSettings("BCC", "BetterClaudeConfig").value("ui/theme", core.THEME_SYSTEM)
+    return value if value in core.THEME_CHOICES else core.THEME_SYSTEM
+
+
+def theme_stylesheet_for(app: QApplication, setting: str | None = None) -> str:
+    """Resolve setting + OS appearance into a palette, apply it, return the QSS."""
+    if setting is None:
+        setting = stored_theme_setting()
+    theme = core.resolve_theme(setting, system_is_dark(app))
+    return apply_palette(core.palette_for(theme))
+
+
 def main():
     if sys.platform == "win32":
         # Without an explicit AppUserModelID, Windows taskbar groups the app
@@ -2773,7 +2899,7 @@ def main():
     icon = _app_icon()
     if not icon.isNull():
         app.setWindowIcon(icon)
-    app.setStyleSheet(STYLESHEET)
+    app.setStyleSheet(theme_stylesheet_for(app))
     win = MainWindow()
     win.show()
     sys.exit(app.exec())

bcc_core.py

@@ -15,6 +15,7 @@ from __future__ import annotations
 
 import base64
 import contextlib
+import copy
 import difflib
 import functools
 import glob
@@ -49,6 +50,12 @@ DISABLED_KEY = "_disabledMcpServers"
 # parks the rest under DISABLED_KEY.
 SETS_KEY = "_bccServerSets"
 
+# Top-level keys BCC itself authors. They live in the client's config file, but
+# BCC is their owner, so on a stale-file merge the in-memory copy wins over the
+# on-disk one (see `carry_owned_keys`). Any future BCC-authored key belongs
+# here -- forgetting to add one is exactly how #73 happened.
+BCC_OWNED_KEYS = (SETS_KEY,)
+
 BACKUP_DIRNAME = ".bcc_backups"
 MAX_BACKUPS = 15
 
@@ -211,16 +218,195 @@ class Profile:
         self.path = Path(self.path)
 
 
+class _NoRaw:
+    """Sentinel for ServerEntry.raw.
+
+    `None` can't do this job: `{"mcpServers": {"foo": null}}` is legal JSON and
+    a real malformed-entry case, so None has to mean "the config said null",
+    not "there was nothing here".
+    """
+
+    __slots__ = ()
+
+    def __repr__(self) -> str:  # keeps ServerEntry reprs readable in test output
+        return "<no raw>"
+
+
+NO_RAW = _NoRaw()
+
+
 @dataclass
 class ServerEntry:
+    """One server definition.
+
+    `data` is always a dict so every consumer can treat it as one. When the
+    config held something that wasn't a JSON object for this server (a string,
+    a number, a list -- all legal JSON, all wrong here), `data` is empty and
+    the original value is preserved verbatim in `raw` so Save round-trips it
+    instead of silently deleting the user's line. `lint_servers` surfaces it.
+    `raw` defaults to the NO_RAW sentinel rather than None, because a config
+    value of literal `null` is itself a malformed entry worth preserving.
+
+    Assigning `data` means the user replaced the definition through the editor,
+    which retires `raw` -- use `set_data` so that can't be forgotten.
+    """
+
     name: str
     data: dict
     enabled: bool = True
+    raw: object = NO_RAW
 
     @property
     def kind(self) -> str:
         return "remote" if "url" in self.data and "command" not in self.data else "stdio"
 
+    @property
+    def malformed(self) -> bool:
+        """True when the config value for this server wasn't a JSON object."""
+        return self.raw is not NO_RAW
+
+    def set_data(self, data: dict) -> None:
+        """Replace the definition from the editor, clearing any malformed original."""
+        self.data = data
+        self.raw = NO_RAW
+
+    def config_value(self):
+        """What to write back to the config: the edited dict, or the untouched
+        malformed original when the user never edited it."""
+        return self.data if self.raw is NO_RAW else self.raw
+
+
+# --------------------------------------------------------------------------- #
+# Theming (issue #75)
+# --------------------------------------------------------------------------- #
+THEME_SYSTEM = "system"
+THEME_LIGHT = "light"
+THEME_DARK = "dark"
+THEME_CHOICES = (THEME_SYSTEM, THEME_LIGHT, THEME_DARK)
+
+
+@dataclass(frozen=True)
+class Palette:
+    """Every colour the UI draws with.
+
+    Deliberately exhaustive: the stylesheet used to inline a handful of
+    near-black literals (`#1a1205` for text on the accent, `#202229` for the
+    disabled table, `#16181d` for the diagnostics pane), which is fine while
+    there's one theme and invisible breakage the moment there are two. Each
+    gets a slot here so a light palette can't silently inherit a dark value.
+    """
+
+    name: str
+    accent: str
+    accent_dim: str
+    bg: str
+    panel: str
+    panel_2: str
+    text: str
+    muted: str
+    border: str
+    good: str
+    bad: str
+    warn: str
+    remote: str
+    on_accent: str  # text drawn on top of an accent fill
+    disabled_bg: str  # the parked-servers table
+    mono_bg: str  # diagnostics / log panes
+    selection_text: str
+
+
+# The shipping theme through v1.3.0. These values are carried over verbatim --
+# adding a light theme must not restyle the dark one.
+DARK_PALETTE = Palette(
+    name="dark",
+    accent="#f97316",
+    accent_dim="#c2570b",
+    bg="#1b1d23",
+    panel="#23262e",
+    panel_2="#2b2f39",
+    text="#e7e9ee",
+    muted="#9aa0ad",
+    border="#3a3f4b",
+    good="#4ade80",
+    bad="#f87171",
+    warn="#fbbf24",
+    remote="#60a5fa",
+    on_accent="#1a1205",
+    disabled_bg="#202229",
+    mono_bg="#16181d",
+    selection_text="#ffffff",
+)
+
+# The semantic colours are NOT the dark ones lightened. #4ade80 / #fbbf24 sit
+# around 1.7:1 against white -- illegible. These are darkened to clear 4.5:1,
+# which `test_light_palette_meets_contrast` enforces so nobody "tidies" them
+# back toward the dark hues later.
+LIGHT_PALETTE = Palette(
+    name="light",
+    accent="#c2410c",
+    accent_dim="#9a3412",
+    bg="#f6f7f9",
+    panel="#ffffff",
+    panel_2="#eef0f4",
+    text="#1b1d23",
+    muted="#5c6270",
+    border="#d3d7de",
+    good="#15803d",
+    bad="#b91c1c",
+    warn="#a16207",
+    remote="#1d4ed8",
+    on_accent="#ffffff",
+    disabled_bg="#e9ebef",
+    mono_bg="#f0f2f5",
+    selection_text="#ffffff",
+)
+
+PALETTES = {DARK_PALETTE.name: DARK_PALETTE, LIGHT_PALETTE.name: LIGHT_PALETTE}
+
+
+def resolve_theme(setting: str, system_is_dark: bool) -> str:
+    """Map a stored theme setting + the OS appearance onto a concrete palette name.
+
+    Anything unrecognised (a hand-edited QSettings value, a setting written by
+    a future version) falls back to following the system rather than to a
+    fixed theme -- the user's desktop is the better guess.
+    """
+    if setting == THEME_DARK:
+        return THEME_DARK
+    if setting == THEME_LIGHT:
+        return THEME_LIGHT
+    return THEME_DARK if system_is_dark else THEME_LIGHT
+
+
+def palette_for(theme: str) -> Palette:
+    """Concrete palette by name; unknown names fall back to dark (the historical look)."""
+    return PALETTES.get(theme, DARK_PALETTE)
+
+
+def _hex_to_rgb(value: str) -> tuple[int, int, int]:
+    v = value.lstrip("#")
+    if len(v) == 3:
+        v = "".join(ch * 2 for ch in v)
+    return int(v[0:2], 16), int(v[2:4], 16), int(v[4:6], 16)
+
+
+def relative_luminance(color: str) -> float:
+    """WCAG relative luminance for a #rrggbb colour."""
+
+    def chan(c: int) -> float:
+        srgb = c / 255.0
+        return srgb / 12.92 if srgb <= 0.04045 else ((srgb + 0.055) / 1.055) ** 2.4
+
+    r, g, b = (chan(c) for c in _hex_to_rgb(color))
+    return 0.2126 * r + 0.7152 * g + 0.0722 * b
+
+
+def contrast_ratio(fg: str, bg: str) -> float:
+    """WCAG contrast ratio between two #rrggbb colours (1.0 to 21.0)."""
+    a, b = relative_luminance(fg), relative_luminance(bg)
+    lighter, darker = max(a, b), min(a, b)
+    return (lighter + 0.05) / (darker + 0.05)
+
 
 # --------------------------------------------------------------------------- #
 # Discovery
@@ -485,13 +671,30 @@ def repair_config_file(path: str | os.PathLike) -> tuple[dict, list[str], str]:
     return obj, notes, pretty
 
 
+def _server_entry(name: str, data, enabled: bool) -> ServerEntry:
+    """Build a ServerEntry, tolerating a value that isn't a JSON object.
+
+    A hand-edited config can legally hold `{"mcpServers": {"foo": "oops"}}` --
+    valid JSON, wrong shape. Calling dict() on that raises, which used to take
+    the whole load down before the linter ever got a look at it (#72). Keep the
+    original instead and let the linter report it.
+    """
+    if isinstance(data, dict):
+        return ServerEntry(name=name, data=dict(data), enabled=enabled)
+    return ServerEntry(name=name, data={}, enabled=enabled, raw=data)
+
+
 def extract_servers(cfg: dict) -> list[ServerEntry]:
-    """Pull enabled (`mcpServers`) and disabled (`_disabledMcpServers`) servers."""
+    """Pull enabled (`mcpServers`) and disabled (`_disabledMcpServers`) servers.
+
+    Never raises on a structurally-odd config -- malformed entries come back as
+    empty-data entries carrying their original value (see `_server_entry`).
+    """
     out: list[ServerEntry] = []
     for name, data in (cfg.get("mcpServers") or {}).items():
-        out.append(ServerEntry(name=name, data=dict(data), enabled=True))
+        out.append(_server_entry(name, data, True))
     for name, data in (cfg.get(DISABLED_KEY) or {}).items():
-        out.append(ServerEntry(name=name, data=dict(data), enabled=False))
+        out.append(_server_entry(name, data, False))
     return out
 
 
@@ -583,8 +786,8 @@ def apply_servers(cfg: dict, servers: list[ServerEntry]) -> dict:
     Write the server list back into `cfg` in place, preserving every other key
     and the position of `mcpServers`. Returns the same dict for convenience.
     """
-    enabled = {s.name: s.data for s in servers if s.enabled}
-    disabled = {s.name: s.data for s in servers if not s.enabled}
+    enabled = {s.name: s.config_value() for s in servers if s.enabled}
+    disabled = {s.name: s.config_value() for s in servers if not s.enabled}
 
     cfg["mcpServers"] = enabled  # replaces value if key existed; appends otherwise
     if disabled:
@@ -597,6 +800,34 @@ def apply_servers(cfg: dict, servers: list[ServerEntry]) -> dict:
 # --------------------------------------------------------------------------- #
 # Write (atomic, with rotating backups)
 # --------------------------------------------------------------------------- #
+def carry_owned_keys(local_cfg: dict, fresh_cfg: dict) -> list[str]:
+    """Carry BCC-authored top-level keys from `local_cfg` onto `fresh_cfg`.
+
+    Used by the stale-file "Merge & save" path, which reloads the file from
+    disk and re-applies the user's server edits. That reload used to drop
+    anything BCC owns but `apply_servers` doesn't write -- named server sets
+    vanished without a word (#73). BCC owns these keys, so the in-memory copy
+    wins; mutates `fresh_cfg` in place.
+
+    Returns the keys where the on-disk copy differed and was overwritten, so
+    the caller can tell the user something was actually contested rather than
+    merely carried across.
+
+    Deliberately one-directional: a key absent locally is left alone on disk.
+    We can't tell "user deleted their last set" from "user never had sets and
+    another machine just added some", and silently deleting someone else's
+    data is the worse of the two failures.
+    """
+    conflicts: list[str] = []
+    for key in BCC_OWNED_KEYS:
+        if key not in local_cfg:
+            continue
+        if key in fresh_cfg and fresh_cfg[key] != local_cfg[key]:
+            conflicts.append(key)
+        fresh_cfg[key] = copy.deepcopy(local_cfg[key])
+    return conflicts
+
+
 def _make_backup(path: Path) -> Path:
     bdir = path.parent / BACKUP_DIRNAME
     bdir.mkdir(exist_ok=True)
@@ -662,13 +893,28 @@ def backup_label(backup_path: Path | str) -> str:
     return f"{ts[:4]}-{ts[4:6]}-{ts[6:8]}  {ts[9:11]}:{ts[11:13]}:{ts[13:]}"
 
 
+def should_mask_value(key: str, value) -> bool:
+    """Whether an env/header value should be masked for display.
+
+    A ${VAR} reference is NOT a secret -- it's a pointer to one, and it's the
+    thing we want users to adopt. Masking it to dots would make a reference
+    indistinguishable from a stored credential, hiding exactly the distinction
+    that makes the feature worth using (#76).
+    """
+    if not is_secret_key(key):
+        return False
+    return not is_env_ref(value) if isinstance(value, str) else True
+
+
 def _redact_server_data(data: dict) -> dict:
     """Return a copy of a server definition with secrets masked for display."""
     out = dict(data)
     if "args" in out:
         out["args"] = redact_args(list(out["args"] or []))
     if "env" in out:
-        out["env"] = {k: (MASK if is_secret_key(k) else v) for k, v in (out["env"] or {}).items()}
+        out["env"] = {
+            k: (MASK if should_mask_value(k, v) else v) for k, v in (out["env"] or {}).items()
+        }
     return out
 
 
@@ -1219,6 +1465,155 @@ MASK = "••••••••"
 _EMBEDDED_CRED_RE = re.compile(r"://[^:@/\s]+:[^:@/\s]+@")
 
 
+# --------------------------------------------------------------------------- #
+# Environment-variable references (issue #76)
+#
+# Claude Code expands ${VAR} and ${VAR:-default} itself, in command, args, env,
+# url and headers. So BCC does NOT expand these on write -- resolving them into
+# the file would put the secret back on disk, which is the whole thing the user
+# is avoiding, and would defeat a feature the client already implements. BCC
+# authors, validates and warns.
+#
+# Claude Desktop has no documented support, so the same text there is passed to
+# the server literally. That makes this a per-client capability, not a global
+# one -- see client_expands_env_refs().
+# --------------------------------------------------------------------------- #
+# ${NAME} or ${NAME:-default}. Names follow the shell convention (letter or
+# underscore first) so a bare "${}" or "${1}" isn't mistaken for a reference.
+_ENV_REF_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)(?::-([^}]*))?\}")
+
+# The five fields Claude Code documents as expansion sites.
+ENV_REF_FIELDS = ("command", "args", "env", "url", "headers")
+
+
+class EnvRef(NamedTuple):
+    """One ${VAR} / ${VAR:-default} occurrence found in a server definition."""
+
+    name: str
+    default: str | None
+    field: str  # which of ENV_REF_FIELDS it was found in
+
+    @property
+    def has_default(self) -> bool:
+        return self.default is not None
+
+
+def find_env_refs(text: str, field: str = "") -> list[EnvRef]:
+    """Every ${VAR} / ${VAR:-default} reference in a single string."""
+    if not isinstance(text, str):
+        return []
+    return [EnvRef(m.group(1), m.group(2), field) for m in _ENV_REF_RE.finditer(text)]
+
+
+def is_env_ref(value: str) -> bool:
+    """True when the value contains at least one ${VAR} reference.
+
+    Used to keep placeholders OUT of secret masking: `${API_KEY}` under a
+    secret-looking key is a reference, not a secret, and masking it to dots
+    would hide the one distinction the user needs to see.
+    """
+    return bool(find_env_refs(value))
+
+
+def server_env_refs(data: dict) -> list[EnvRef]:
+    """Every env reference in a server definition, tagged with its field.
+
+    Only inspects the fields Claude Code actually expands; a ${VAR} written
+    into some other key is not a reference and shouldn't be reported as one.
+    """
+    out: list[EnvRef] = []
+    if not isinstance(data, dict):
+        return out
+    for field in ENV_REF_FIELDS:
+        value = data.get(field)
+        if isinstance(value, str):
+            out.extend(find_env_refs(value, field))
+        elif isinstance(value, list):
+            for item in value:
+                out.extend(find_env_refs(item, field))
+        elif isinstance(value, dict):
+            for v in value.values():
+                out.extend(find_env_refs(v, field))
+    return out
+
+
+def expand_env_refs(text: str, environ: dict | None = None) -> str:
+    """Expand ${VAR} / ${VAR:-default} the way Claude Code documents it.
+
+    Provided for previewing what the client will do -- BCC never writes the
+    expanded form back to the config. Unset with no default is left as the
+    literal ${VAR} text, matching Claude Code: the config still loads and the
+    unexpanded text is passed through.
+    """
+    if not isinstance(text, str):
+        return text
+    env = os.environ if environ is None else environ
+
+    def repl(m: re.Match) -> str:
+        name, default = m.group(1), m.group(2)
+        if name in env:
+            return env[name]
+        return default if default is not None else m.group(0)
+
+    return _ENV_REF_RE.sub(repl, text)
+
+
+def unresolved_env_refs(data: dict, environ: dict | None = None) -> list[EnvRef]:
+    """References that would not resolve: variable unset AND no default.
+
+    Best-effort by nature -- BCC's environment isn't necessarily the client's,
+    so this warns rather than blocks, and the warning text says so.
+    """
+    env = os.environ if environ is None else environ
+    return [r for r in server_env_refs(data) if not r.has_default and r.name not in env]
+
+
+def client_expands_env_refs(profile: Profile) -> bool:
+    """Whether the client behind `profile` expands ${VAR} itself.
+
+    Claude Code does, in command/args/env/url/headers, for both project
+    `.mcp.json` and user-scope `~/.claude.json`. Claude Desktop has no
+    documented support, so a reference there reaches the server as literal
+    text -- which surfaces as a confusing auth failure rather than an obvious
+    config error, hence the warning.
+    """
+    return not profile_targets_claude_desktop(profile)
+
+
+def env_ref_warnings(
+    data: dict, profile: Profile | None = None, environ: dict | None = None
+) -> list[str]:
+    """Advisory warnings about env references in one server definition.
+
+    Two distinct problems, deliberately worded differently:
+      - the target client won't expand them at all (Claude Desktop)
+      - the client will expand them, but a variable looks unset here
+    """
+    refs = server_env_refs(data)
+    if not refs:
+        return []
+
+    if profile is not None and not client_expands_env_refs(profile):
+        names = ", ".join(sorted({f"${{{r.name}}}" for r in refs}))
+        return [
+            f"{names} will NOT be expanded by Claude Desktop -- it has no "
+            f"documented support for variable references, so the server "
+            f"receives the literal text. Use a real value here, or move this "
+            f"server to a Claude Code config."
+        ]
+
+    missing = unresolved_env_refs(data, environ)
+    if not missing:
+        return []
+    names = ", ".join(sorted({r.name for r in missing}))
+    return [
+        f"{names} is not set in this environment and has no ':-default'. "
+        f"Claude Code will pass the reference through unexpanded. "
+        f"(Checked against BCC's environment, which may differ from the "
+        f"client's.)"
+    ]
+
+
 def is_secret_key(name: str) -> bool:
     """Does this env-var / header / flag name look like it holds a secret?"""
     return bool(_SECRET_KEY_RE.search(name or ""))
@@ -1235,17 +1630,22 @@ def redact_args(args: list[str]) -> list[str]:
       --api-key=abc123      ->  --api-key=••••••••   (inline flag=value)
       ghp_abc123            ->  ••••••••             (well-known token prefix)
     Everything else passes through untouched.
+
+    ${VAR} references are left visible: they name a secret rather than being
+    one, and hiding them would obscure the difference between "this config
+    leaks a token" and "this config points at one" (#76).
     """
     out: list[str] = []
     mask_next = False
     for a in args:
         s = str(a)
         if mask_next:
-            out.append(MASK)
             mask_next = False
+            out.append(s if is_env_ref(s) else MASK)
             continue
         if s.startswith("-") and "=" in s and is_secret_key(s.split("=", 1)[0]):
-            out.append(s.split("=", 1)[0] + "=" + MASK)
+            flag, value = s.split("=", 1)
+            out.append(f"{flag}={value}" if is_env_ref(value) else f"{flag}={MASK}")
             continue
         if s.startswith("-") and is_secret_key(s):
             out.append(s)
@@ -1272,6 +1672,11 @@ def args_secret_warning(data: dict) -> str | None:
     args = [str(a) for a in (data.get("args") or [])]
     mask_next = False
     for a in args:
+        # A ${VAR} reference is the recommended fix for this very warning --
+        # continuing to warn after the user adopts it punishes the fix (#76).
+        if is_env_ref(a):
+            mask_next = False
+            continue
         if mask_next:
             mask_next = False
             if not a.startswith("-"):
@@ -1413,9 +1818,21 @@ def lint_server(name: str, data: dict) -> list[str]:
 
 
 def lint_servers(servers: list[ServerEntry]) -> list[str]:
-    """Concatenate lint_server warnings across every entry, in order."""
+    """Concatenate lint_server warnings across every entry, in order.
+
+    Entries whose config value wasn't a JSON object at all are reported here
+    rather than in lint_server, which takes an already-dict `data` (#72).
+    """
     out: list[str] = []
     for s in servers:
+        if s.malformed:
+            nm = s.name.strip() or "(unnamed)"
+            out.append(
+                f"'{nm}': server definition is not an object "
+                f"(found {type(s.raw).__name__}) -- it is preserved as-is; "
+                f"edit it to replace it with a proper definition"
+            )
+            continue
         out.extend(lint_server(s.name, s.data))
     return out
 

tests/test_core.py

@@ -1,6 +1,7 @@
 """Pytest port of the original test_core.py script (same 23 behaviours, now
 proper test functions with tmp_path/monkeypatch fixtures)."""
 
+import dataclasses
 import json
 import os
 import re
@@ -2372,6 +2373,261 @@ def test_config_has_unfilled_placeholders_checks_env_too():
     assert c.config_has_unfilled_placeholders(cfg) is True
 
 
+# --------------------------------------------------------------------------- #
+# #72 -- a server value that isn't a JSON object must not take the load down
+# --------------------------------------------------------------------------- #
+@pytest.mark.parametrize("bad", ["not-a-dict", 123, ["a", "b"], None, True, 1.5])
+def test_extract_servers_survives_non_dict_server_value(bad):
+    entries = c.extract_servers({"mcpServers": {"foo": bad}})
+    assert len(entries) == 1
+    assert entries[0].name == "foo"
+    assert entries[0].data == {}
+    assert entries[0].malformed is True
+    assert entries[0].raw == bad
+
+
+def test_extract_servers_marks_only_the_bad_entry():
+    cfg = {"mcpServers": {"good": {"command": "npx"}, "bad": "oops"}}
+    by_name = {e.name: e for e in c.extract_servers(cfg)}
+    assert by_name["good"].malformed is False
+    assert by_name["good"].data == {"command": "npx"}
+    assert by_name["bad"].malformed is True
+
+
+def test_extract_servers_handles_malformed_disabled_entry():
+    entries = c.extract_servers({c.DISABLED_KEY: {"parked": ["nope"]}})
+    assert entries[0].enabled is False
+    assert entries[0].malformed is True
+
+
+def test_malformed_entry_round_trips_through_save_unchanged():
+    """The cardinal rule: never silently delete what the user had on disk."""
+    cfg = {"mcpServers": {"good": {"command": "npx"}, "bad": "oops"}}
+    servers = c.extract_servers(cfg)
+    out = c.apply_servers(dict(cfg), servers)
+    assert out["mcpServers"]["bad"] == "oops"
+    assert out["mcpServers"]["good"] == {"command": "npx"}
+
+
+def test_editing_a_malformed_entry_retires_the_raw_value():
+    entry = c.extract_servers({"mcpServers": {"bad": "oops"}})[0]
+    entry.set_data({"command": "npx"})
+    assert entry.malformed is False
+    assert entry.config_value() == {"command": "npx"}
+    assert c.apply_servers({}, [entry])["mcpServers"]["bad"] == {"command": "npx"}
+
+
+def test_lint_reports_the_malformed_entry_by_name():
+    servers = c.extract_servers({"mcpServers": {"bad": "oops"}})
+    warnings = c.lint_servers(servers)
+    assert len(warnings) == 1
+    assert "'bad'" in warnings[0]
+    assert "not an object" in warnings[0]
+    assert "str" in warnings[0]
+
+
+def test_lint_still_reports_normal_warnings_alongside_malformed():
+    cfg = {"mcpServers": {"bad": "oops", "sloppy": {"command": "npx", "args": "one two"}}}
+    warnings = c.lint_servers(c.extract_servers(cfg))
+    assert any("not an object" in w for w in warnings)
+    assert any("'args' should be a list" in w for w in warnings)
+
+
+# --------------------------------------------------------------------------- #
+# #73 -- the stale-file merge must not discard BCC-authored keys
+# --------------------------------------------------------------------------- #
+def test_carry_owned_keys_moves_sets_onto_the_reloaded_config():
+    local = {"mcpServers": {}, c.SETS_KEY: {"work": ["a", "b"]}}
+    fresh = {"mcpServers": {"external": {"command": "npx"}}}
+    contested = c.carry_owned_keys(local, fresh)
+    assert contested == []
+    assert fresh[c.SETS_KEY] == {"work": ["a", "b"]}
+    assert fresh["mcpServers"] == {"external": {"command": "npx"}}
+
+
+def test_carry_owned_keys_reports_a_genuine_conflict():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {c.SETS_KEY: {"work": ["a", "b"]}}
+    assert c.carry_owned_keys(local, fresh) == [c.SETS_KEY]
+    assert fresh[c.SETS_KEY] == {"work": ["a"]}  # local wins: BCC owns the key
+
+
+def test_carry_owned_keys_is_quiet_when_both_sides_agree():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {c.SETS_KEY: {"work": ["a"]}}
+    assert c.carry_owned_keys(local, fresh) == []
+
+
+def test_carry_owned_keys_leaves_disk_alone_when_absent_locally():
+    """Can't distinguish 'deleted my last set' from 'never had sets'; keep theirs."""
+    fresh = {c.SETS_KEY: {"remote": ["a"]}}
+    assert c.carry_owned_keys({}, fresh) == []
+    assert fresh[c.SETS_KEY] == {"remote": ["a"]}
+
+
+def test_carry_owned_keys_deep_copies_so_later_edits_do_not_leak():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {}
+    c.carry_owned_keys(local, fresh)
+    local[c.SETS_KEY]["work"].append("b")
+    assert fresh[c.SETS_KEY] == {"work": ["a"]}
+
+
+def test_merge_flow_preserves_sets_and_external_servers(tmp_path):
+    """End-to-end shape of the Merge & save path that lost sets in #73."""
+    path = tmp_path / "claude.json"
+    path.write_text(json.dumps({"mcpServers": {"old": {"command": "old"}}}))
+
+    # BCC loads, user saves a named set and edits servers in memory.
+    local = c.load_config(path)
+    servers = c.extract_servers(local)
+    c.save_server_set(local, "work", servers)
+
+    # Something else rewrites the file underneath us.
+    path.write_text(json.dumps({"mcpServers": {"external": {"command": "new"}}, "other": 1}))
+
+    # Merge & save: reload disk, carry BCC keys, re-apply the user's servers.
+    fresh = c.load_config(path)
+    c.carry_owned_keys(local, fresh)
+    c.apply_servers(fresh, servers)
+    c.write_config(path, fresh)
+
+    saved = c.load_config(path)
+    assert saved[c.SETS_KEY] == {"work": ["old"]}  # the set survived
+    assert saved["other"] == 1  # unrelated external key preserved
+    assert "old" in saved["mcpServers"]  # user's servers re-applied
+
+
+def test_null_server_value_is_malformed_not_mistaken_for_absent():
+    """`{"mcpServers": {"foo": null}}` is legal JSON and a real malformed case,
+    so None must not double as the 'nothing here' sentinel."""
+    entry = c.extract_servers({"mcpServers": {"foo": None}})[0]
+    assert entry.malformed is True
+    assert entry.raw is None
+    assert c.apply_servers({}, [entry])["mcpServers"]["foo"] is None
+
+
+def test_a_normal_entry_is_not_malformed():
+    entry = c.extract_servers({"mcpServers": {"foo": {"command": "npx"}}})[0]
+    assert entry.malformed is False
+    assert entry.raw is c.NO_RAW
+
+
+# --------------------------------------------------------------------------- #
+# #75 -- theming
+# --------------------------------------------------------------------------- #
+@pytest.mark.parametrize(
+    "setting,system_dark,expected",
+    [
+        (c.THEME_DARK, False, "dark"),
+        (c.THEME_DARK, True, "dark"),
+        (c.THEME_LIGHT, False, "light"),
+        (c.THEME_LIGHT, True, "light"),
+        (c.THEME_SYSTEM, True, "dark"),
+        (c.THEME_SYSTEM, False, "light"),
+    ],
+)
+def test_resolve_theme_covers_every_setting_and_appearance(setting, system_dark, expected):
+    assert c.resolve_theme(setting, system_dark) == expected
+
+
+@pytest.mark.parametrize("junk", ["", "solarized", None, "DARK", 3])
+def test_resolve_theme_falls_back_to_following_the_system(junk):
+    """A hand-edited or future QSettings value should follow the desktop,
+    not pin a fixed theme."""
+    assert c.resolve_theme(junk, True) == "dark"
+    assert c.resolve_theme(junk, False) == "light"
+
+
+def test_palette_for_known_names():
+    assert c.palette_for("dark") is c.DARK_PALETTE
+    assert c.palette_for("light") is c.LIGHT_PALETTE
+
+
+def test_palette_for_unknown_name_falls_back_to_dark():
+    assert c.palette_for("chartreuse") is c.DARK_PALETTE
+
+
+def test_dark_palette_is_unchanged_from_the_shipped_look():
+    """v1.3.0 shipped these exact colours; adding a light theme must not
+    quietly restyle the dark one."""
+    p = c.DARK_PALETTE
+    assert (p.accent, p.bg, p.panel, p.panel_2) == ("#f97316", "#1b1d23", "#23262e", "#2b2f39")
+    assert (p.text, p.muted, p.border) == ("#e7e9ee", "#9aa0ad", "#3a3f4b")
+    assert (p.good, p.bad, p.warn, p.remote) == ("#4ade80", "#f87171", "#fbbf24", "#60a5fa")
+    assert (p.on_accent, p.disabled_bg, p.mono_bg) == ("#1a1205", "#202229", "#16181d")
+
+
+def test_both_palettes_define_every_slot():
+    """A missing slot should fail here rather than render a broken window."""
+    for pal in (c.DARK_PALETTE, c.LIGHT_PALETTE):
+        for f in dataclasses.fields(c.Palette):
+            value = getattr(pal, f.name)
+            assert value, f"{pal.name}.{f.name} is empty"
+            if f.name != "name":
+                assert re.fullmatch(r"#[0-9a-fA-F]{6}", value), f"{pal.name}.{f.name}={value!r}"
+
+
+@pytest.mark.parametrize("pal_name", ["dark", "light"])
+@pytest.mark.parametrize("slot", ["text", "muted", "good", "bad", "warn", "remote", "accent"])
+def test_palette_meets_contrast_on_panel(pal_name, slot):
+    """Every colour drawn as text/glyph must clear WCAG AA (4.5:1) against the
+    surface it sits on. The light palette's semantic colours are NOT the dark
+    ones lightened -- #4ade80 sits near 1.7:1 on white -- so this guards
+    against someone 'harmonising' them back toward the dark hues."""
+    pal = c.palette_for(pal_name)
+    assert c.contrast_ratio(getattr(pal, slot), pal.panel) >= 4.5
+
+
+@pytest.mark.parametrize("pal_name", ["dark", "light"])
+def test_on_accent_is_legible_against_the_accent_fill(pal_name):
+    """Primary buttons and selected rows draw on_accent on top of accent."""
+    pal = c.palette_for(pal_name)
+    assert c.contrast_ratio(pal.on_accent, pal.accent) >= 4.5
+
+
+def test_contrast_ratio_endpoints():
+    assert c.contrast_ratio("#000000", "#ffffff") == pytest.approx(21.0, abs=0.01)
+    assert c.contrast_ratio("#123456", "#123456") == pytest.approx(1.0, abs=0.001)
+    assert c.contrast_ratio("#ffffff", "#000000") == pytest.approx(21.0, abs=0.01)
+
+
+def test_relative_luminance_extremes():
+    assert c.relative_luminance("#000000") == pytest.approx(0.0)
+    assert c.relative_luminance("#ffffff") == pytest.approx(1.0)
+
+
+def test_stylesheet_builder_has_no_hardcoded_colours():
+    """Every colour in the QSS must come from the palette.
+
+    Three near-black literals used to be inlined here (#1a1205, #202229,
+    #16181d). Harmless with one theme; with two, they silently render dark
+    chrome on a light window. Reads the source rather than importing bcc,
+    which needs PySide6.
+    """
+    src = (Path(__file__).resolve().parent.parent / "bcc.py").read_text(encoding="utf-8")
+    start = src.index("def build_stylesheet")
+    body = src[start : src.index("def apply_palette")]
+    assert re.findall(r"#[0-9a-fA-F]{6}", body) == []
+
+
+def test_every_palette_slot_is_consumed():
+    """A slot added to Palette but never wired up is dead weight.
+
+    Checks for `p.<slot>` anywhere in bcc.py, which covers both the QSS and
+    apply_palette's global bindings -- not every slot belongs in the
+    stylesheet (`good` and `remote` feed the inline status dots via
+    STATUS_COLORS/HEALTH_COLORS, never the QSS). This won't catch a slot bound
+    to a global that nothing then uses; it does catch the common mistake of
+    extending the dataclass and forgetting to plumb it through.
+    """
+    src = (Path(__file__).resolve().parent.parent / "bcc.py").read_text(encoding="utf-8")
+    for f in dataclasses.fields(c.Palette):
+        if f.name == "name":
+            continue
+        assert f"p.{f.name}" in src, f"palette slot {f.name!r} is never consumed"
+
+
 # --------------------------------------------------------------------------- #
 # #78/#79 -- update notice: when to show it, and what it says
 # --------------------------------------------------------------------------- #
@@ -2424,3 +2680,138 @@ def test_update_notice_renders_both_versions_the_same_way():
     assert "1.3.0" in n["text"] and "1.2.0" in n["text"]
     # the machine-readable field keeps the real tag
     assert n["version"] == "v1.3.0"
+
+
+# --------------------------------------------------------------------------- #
+# #76 -- ${VAR} references. Semantics mirror Claude Code's documented
+# behaviour: ${VAR} and ${VAR:-default}, expanded in command/args/env/url/
+# headers, and an unset variable with no default left as literal text.
+# --------------------------------------------------------------------------- #
+def test_find_env_refs_plain_and_defaulted():
+    refs = c.find_env_refs("${A} and ${B:-fallback}")
+    assert [(r.name, r.default) for r in refs] == [("A", None), ("B", "fallback")]
+
+
+@pytest.mark.parametrize("text", ["${}", "${1BAD}", "$NOTBRACED", "{NOPE}", "plain", "$${X"])
+def test_find_env_refs_ignores_non_references(text):
+    assert c.find_env_refs(text) == []
+
+
+def test_find_env_refs_allows_an_empty_default():
+    """`${VAR:-}` is a documented way to say 'blank if unset'."""
+    refs = c.find_env_refs("${A:-}")
+    assert refs[0].default == ""
+    assert refs[0].has_default is True
+
+
+def test_server_env_refs_covers_all_five_documented_fields():
+    data = {
+        "command": "${BIN}",
+        "args": ["--x", "${ARG}"],
+        "env": {"K": "${ENVV}"},
+        "url": "${URL}/mcp",
+        "headers": {"Authorization": "Bearer ${HDR}"},
+    }
+    found = {(r.name, r.field) for r in c.server_env_refs(data)}
+    assert found == {
+        ("BIN", "command"),
+        ("ARG", "args"),
+        ("ENVV", "env"),
+        ("URL", "url"),
+        ("HDR", "headers"),
+    }
+
+
+def test_server_env_refs_ignores_unexpanded_fields():
+    """Claude Code expands five fields; a ${VAR} elsewhere isn't a reference."""
+    assert c.server_env_refs({"description": "${NOPE}", "timeout": "${ALSO_NO}"}) == []
+
+
+def test_expand_env_refs_matches_documented_semantics():
+    env = {"SET": "value"}
+    assert c.expand_env_refs("${SET}", env) == "value"
+    assert c.expand_env_refs("${MISSING:-dflt}", env) == "dflt"
+    assert c.expand_env_refs("${SET:-dflt}", env) == "value"
+    # unset with no default: left as literal text, exactly as Claude Code does
+    assert c.expand_env_refs("${MISSING}", env) == "${MISSING}"
+
+
+def test_expand_env_refs_handles_several_in_one_string():
+    assert c.expand_env_refs("${A}/${B:-two}/${C}", {"A": "one"}) == "one/two/${C}"
+
+
+def test_unresolved_env_refs_only_flags_unset_without_default():
+    data = {"env": {"A": "${SET}", "B": "${UNSET}", "C": "${OTHER:-has_default}"}}
+    assert [r.name for r in c.unresolved_env_refs(data, {"SET": "x"})] == ["UNSET"]
+
+
+# --- the two interactions that were backwards for this feature ------------
+def test_placeholder_under_a_secret_key_is_not_masked():
+    """A ${VAR} names a secret rather than being one. Masking it would make a
+    reference indistinguishable from a stored credential."""
+    assert c.should_mask_value("API_KEY", "${API_KEY}") is False
+    assert c.should_mask_value("API_KEY", "ghp_realsecret") is True
+    assert c.should_mask_value("NOT_SECRET", "${API_KEY}") is False
+
+
+def test_redacted_display_keeps_placeholders_but_masks_real_secrets():
+    out = c._redact_server_data({"env": {"API_KEY": "${API_KEY}", "TOKEN": "ghp_real"}})
+    assert out["env"]["API_KEY"] == "${API_KEY}"
+    assert out["env"]["TOKEN"] == c.MASK
+
+
+def test_redact_args_keeps_placeholders_visible():
+    assert c.redact_args(["--token", "${GH_TOKEN}"]) == ["--token", "${GH_TOKEN}"]
+    assert c.redact_args(["--api-key=${K}"]) == ["--api-key=${K}"]
+    # real secrets still masked
+    assert c.redact_args(["--token", "ghp_real"]) == ["--token", c.MASK]
+    assert c.redact_args(["--api-key=sk-real"]) == [f"--api-key={c.MASK}"]
+
+
+def test_args_secret_warning_is_silenced_by_a_placeholder():
+    """Moving a token into ${VAR} is the recommended fix for this warning --
+    still warning afterwards would punish the fix."""
+    assert c.args_secret_warning({"args": ["--token", "ghp_real"]}) is not None
+    assert c.args_secret_warning({"args": ["--token", "${GH_TOKEN}"]}) is None
+
+
+def test_args_secret_warning_still_fires_on_the_arg_after_a_placeholder():
+    """A placeholder must clear the pending-flag state, not blanket-suppress."""
+    assert c.args_secret_warning({"args": ["${SAFE}", "--token", "ghp_real"]}) is not None
+
+
+# --- per-client gating ----------------------------------------------------
+def _profile(path):
+    return c.Profile(label="p", path=Path(path), config_exists=True)
+
+
+def test_claude_code_profiles_expand_references():
+    assert c.client_expands_env_refs(_profile(Path.home() / ".claude.json")) is True
+    assert c.client_expands_env_refs(_profile("/repo/.mcp.json")) is True
+
+
+def test_claude_desktop_profile_does_not_expand_references():
+    desktop = _profile(c.app_support_base() / "Claude" / c.CONFIG_FILENAME)
+    assert c.client_expands_env_refs(desktop) is False
+
+
+def test_desktop_profile_warns_that_references_are_literal():
+    data = {"env": {"API_KEY": "${API_KEY}"}}
+    desktop = _profile(c.app_support_base() / "Claude" / c.CONFIG_FILENAME)
+    warnings = c.env_ref_warnings(data, desktop, {"API_KEY": "set"})
+    assert len(warnings) == 1
+    assert "NOT be expanded" in warnings[0]
+    assert "${API_KEY}" in warnings[0]
+
+
+def test_claude_code_profile_warns_only_about_unset_variables():
+    code = _profile(Path.home() / ".claude.json")
+    data = {"env": {"A": "${UNSET_ONE}"}}
+    assert c.env_ref_warnings(data, code, {}) != []
+    assert c.env_ref_warnings(data, code, {"UNSET_ONE": "x"}) == []
+    # a default means it always resolves
+    assert c.env_ref_warnings({"env": {"A": "${X:-d}"}}, code, {}) == []
+
+
+def test_no_references_means_no_warnings():
+    assert c.env_ref_warnings({"command": "npx", "args": ["-y", "pkg"]}, None) == []