From 85314ed1622cbdcbd6621ebf97307606f4baf89d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 8 Jun 2026 23:10:35 -0700
Subject: [PATCH] Studio frontend: reduce and tighten code comments (#6099)

Trim and tighten code comments across studio/frontend TS/JS. Comment-only: every changed file verified code-identical to main via the TypeScript printer signature comparison.
---
 studio/frontend/src/app/provider.tsx          |  18 +-
 studio/frontend/src/app/routes/__root.tsx     |   4 +-
 studio/frontend/src/app/routes/settings.tsx   |   8 +-
 .../frontend/src/components/app-sidebar.tsx   |  51 ++-
 .../components/assistant-ui/attachment.tsx    |   4 +-
 .../components/assistant-ui/code-plugin.ts    |   4 +-
 .../components/assistant-ui/code-themes.ts    |   9 +-
 .../src/components/assistant-ui/image.tsx     |   5 +-
 .../components/assistant-ui/markdown-text.tsx |  18 +-
 .../assistant-ui/message-timing.tsx           |   7 +-
 .../model-selector/folder-browser.tsx         |  41 +--
 .../assistant-ui/model-selector/pickers.tsx   |  78 ++---
 .../src/components/assistant-ui/reasoning.tsx |   6 +-
 .../src/components/assistant-ui/sources.tsx   |  24 +-
 .../assistant-ui/think-aria-label.ts          |   7 +-
 .../src/components/assistant-ui/thread.tsx    | 216 ++++++------
 .../assistant-ui/tool-call-spinner.tsx        |   6 +-
 .../components/assistant-ui/tool-group.tsx    |   4 +-
 .../assistant-ui/tool-ui-code-execution.tsx   |  14 +-
 .../assistant-ui/tool-ui-image-generation.tsx |  28 +-
 .../assistant-ui/tool-ui-python.tsx           |   2 +-
 .../assistant-ui/tool-ui-render-html.tsx      |   9 +-
 .../assistant-ui/tool-ui-web-search.tsx       |   5 +-
 .../use-intent-aware-autoscroll.tsx           | 276 ++++++---------
 .../src/components/shutdown-dialog.tsx        |   9 +-
 .../src/components/tauri/window-titlebar.tsx  |   2 +-
 studio/frontend/src/config/env.ts             |  10 +-
 studio/frontend/src/features/auth/api.ts      |  12 +-
 .../features/auth/components/auth-form.tsx    |  28 +-
 studio/frontend/src/features/auth/session.ts  |  13 +-
 .../src/features/auth/tauri-auto-auth.ts      |   6 +-
 .../src/features/chat/api-provider-logo.tsx   |   9 +-
 .../src/features/chat/api/chat-adapter.ts     | 327 ++++++++----------
 .../src/features/chat/api/chat-api.ts         |  46 ++-
 .../features/chat/api/openai-containers.ts    |  14 +-
 .../src/features/chat/api/providers-api.ts    |   7 +-
 .../chat/artifacts/artifact-surface.tsx       |   7 +-
 .../features/chat/artifacts/html-frame.tsx    |   5 +-
 .../src/features/chat/artifacts/types.ts      |   4 +-
 .../features/chat/chat-mcp-servers-dialog.tsx |   7 +-
 .../frontend/src/features/chat/chat-page.tsx  | 168 ++++-----
 .../features/chat/chat-providers-dialog.tsx   |  12 +-
 .../src/features/chat/chat-settings-sheet.tsx |  44 +--
 .../chat/components/context-usage-bar.tsx     |   2 +-
 .../chat/components/model-load-status.tsx     |  21 +-
 .../components/openai-code-exec-section.tsx   | 209 +++++------
 studio/frontend/src/features/chat/db.ts       |   5 +-
 .../src/features/chat/external-providers.ts   |  63 ++--
 .../chat/hooks/use-chat-model-runtime.ts      |  65 ++--
 .../chat/hooks/use-chat-search-index.ts       |   7 +-
 .../chat/hooks/use-chat-sidebar-items.ts      |   5 +-
 .../chat/hooks/use-pill-activation-order.ts   |   5 +-
 .../features/chat/hooks/use-transfer-stats.ts |  20 +-
 .../src/features/chat/lib/friendly-names.ts   |  37 +-
 .../src/features/chat/mcp-composer-button.tsx |  15 +-
 .../features/chat/provider-capabilities.ts    | 224 +++++-------
 .../src/features/chat/runtime-provider.tsx    |  45 ++-
 .../src/features/chat/shared-composer.tsx     | 154 ++++-----
 .../chat/stores/chat-runtime-store.ts         | 132 ++++---
 .../frontend/src/features/chat/tour/steps.tsx |   4 +-
 studio/frontend/src/features/chat/types.ts    |  31 +-
 .../frontend/src/features/chat/types/api.ts   |  75 ++--
 .../src/features/chat/types/runtime.ts        |   4 +-
 .../chat/utils/chat-history-storage.ts        |  62 ++--
 .../chat/utils/chat-thread-tombstones.ts      |  10 +-
 .../chat/utils/delete-thread-message.ts       |  17 +-
 .../features/chat/utils/format-transfer.ts    |  20 +-
 .../chat/utils/image-input-support.ts         |  13 +-
 .../chat/utils/parse-assistant-content.ts     |   8 +-
 .../src/features/chat/utils/transfer-stats.ts |  35 +-
 .../src/features/export/api/export-api.ts     |   6 +-
 .../export/components/export-dialog.tsx       |  84 ++---
 .../src/features/export/export-page.tsx       |  13 +-
 .../components/steps/hyperparameters-step.tsx |   8 +-
 .../components/steps/model-selection-step.tsx |   2 +-
 .../components/steps/model-type-step.tsx      |   2 +-
 .../src/features/profile/utils/jwt-subject.ts |   4 +-
 .../profile/utils/resize-image-file.ts        |   4 +-
 .../components/executions/executions-view.tsx |   8 +-
 .../components/inline/inline-model.tsx        |   6 +-
 .../models/local-recipe-model-selector.tsx    |   4 +-
 .../dialogs/models/model-config-dialog.tsx    |   6 +-
 .../hooks/use-recipe-executions.ts            |  21 +-
 .../recipe-studio/recipe-studio-page.tsx      |  44 ++-
 .../recipe-studio/stores/recipe-studio.ts     |   7 +-
 .../recipe-studio/utils/graph/fit-view.ts     |  15 +-
 .../utils/graph/recipe-graph-connection.ts    |   8 +-
 .../features/recipe-studio/utils/layout.ts    |  24 +-
 .../utils/payload/builders-model.ts           |   9 +-
 .../utils/payload/builders-sampler.ts         |   2 +-
 .../settings/components/create-key-form.tsx   |   5 +-
 .../src/features/settings/settings-dialog.tsx |  15 +-
 .../settings/stores/settings-dialog-store.ts  |   8 +-
 .../features/settings/stores/theme-store.ts   |  16 +-
 .../features/settings/tabs/api-keys-tab.tsx   |   5 +-
 .../features/settings/tabs/general-tab.tsx    |  19 +-
 .../studio/historical-training-view.tsx       |   4 +-
 .../src/features/studio/history-card-grid.tsx |  14 +-
 .../dataset-preview-dialog-mapping.tsx        |   4 +-
 .../sections/dataset-preview-dialog.tsx       |  14 +-
 .../studio/sections/model-section.tsx         |   8 +-
 .../studio/sections/progress-section.tsx      |   2 +-
 .../studio/sections/training-section.tsx      |  11 +-
 .../src/features/studio/studio-page.tsx       |   6 +-
 .../studio/training-start-overlay.tsx         |  28 +-
 .../features/tour/lib/confetti-fireworks.ts   |   7 +-
 .../src/features/training/api/models-api.ts   |  10 +-
 .../training/hooks/use-training-actions.ts    |   4 +-
 .../hooks/use-training-history-sidebar.ts     |  12 +-
 .../hooks/use-training-unload-guard.ts        |   7 +-
 .../src/features/training/lib/yaml-config.ts  |   8 +-
 .../training/stores/training-config-store.ts  |  44 +--
 .../src/hooks/use-collapse-scroll-lock.ts     |  10 +-
 studio/frontend/src/hooks/use-gpu-info.ts     |   6 +-
 .../frontend/src/hooks/use-gpu-utilization.ts |   6 +-
 .../frontend/src/hooks/use-hardware-info.ts   |   8 +-
 .../src/hooks/use-hf-dataset-search.ts        |   4 +-
 .../src/hooks/use-hf-dataset-splits.ts        |   5 +-
 .../frontend/src/hooks/use-hf-model-search.ts |  51 ++-
 .../src/hooks/use-hf-token-validation.ts      |   5 +-
 .../src/hooks/use-recommended-model-vram.ts   |   4 +-
 .../frontend/src/hooks/use-tauri-backend.ts   |  29 +-
 studio/frontend/src/lib/copy-to-clipboard.ts  |   8 +-
 .../frontend/src/lib/format-fastapi-error.ts  |  15 +-
 studio/frontend/src/lib/hf-cache.ts           |  11 +-
 studio/frontend/src/lib/latex.ts              |  65 ++--
 studio/frontend/src/lib/open-link.ts          |  11 +-
 studio/frontend/src/lib/vram.ts               |  61 +---
 studio/frontend/src/main.tsx                  |   4 +-
 studio/frontend/src/utils/index.ts            |   1 -
 130 files changed, 1553 insertions(+), 2141 deletions(-)

diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 6f3c7618be..4bd348effc 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -48,11 +48,10 @@ async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void>
 
   const win = getCurrentWindow();
   // Decide first-launch vs restore from the on-disk state file BEFORE touching the
-  // window. Probing the window itself after restoreStateCurrent is unreliable:
-  // on GTK, set_size against a hidden window is deferred until show(), so
-  // innerSize() reads a stale value and any baseline fallback would overwrite the
-  // queued restore. On macOS the same probe works, hence the inconsistency
-  // between previous iterations of this code.
+  // window. Probing the window after restoreStateCurrent is unreliable: on GTK,
+  // set_size on a hidden window is deferred until show(), so innerSize() reads a
+  // stale value and a baseline fallback would overwrite the queued restore. On
+  // macOS the same probe works, hence the inconsistency between prior iterations.
   const hasSavedState = await invoke<boolean>("has_saved_window_state");
   if (!isCurrent()) return;
 
@@ -60,9 +59,8 @@ async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void>
   if (!isCurrent()) return;
 
   if (hasSavedState) {
-    // Subsequent launch: the plugin handles size, position, and maximized,
-    // with built-in off-screen protection (monitor-intersection check) for
-    // positions saved on a now-disconnected display.
+    // Subsequent launch: plugin restores size/position/maximized, with built-in
+    // off-screen protection for positions saved on a now-disconnected display.
     await restoreStateCurrent(
       StateFlags.SIZE | StateFlags.POSITION | StateFlags.MAXIMIZED,
     );
@@ -87,8 +85,8 @@ async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void>
   if (!isCurrent()) return;
   await win.show();
   if (!isCurrent()) return;
-  // Apply constraints after restore/show. Setting constraints before plugin restore
-  // can emit a Resized event and overwrite the plugin's cached saved size.
+  // Apply constraints after restore/show: doing so before plugin restore can emit
+  // a Resized event and overwrite the plugin's cached saved size.
   await win.setSizeConstraints({ minWidth: MIN_WINDOW_WIDTH, minHeight: MIN_WINDOW_HEIGHT });
 }
 
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index da74a9e8a7..200200d190 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -55,8 +55,8 @@ function isChatOnlyAllowed(pathname: string): boolean {
 
 export const Route = createRootRoute({
   beforeLoad: async ({ location }) => {
-    // Ensure platform info is fetched before checking chat-only guard.
-    // fetchDeviceType caches after first call, so subsequent navigations are instant.
+    // Fetch platform info before the chat-only guard. fetchDeviceType caches,
+    // so later navigations are instant.
     await fetchDeviceType();
     const chatOnly = usePlatformStore.getState().isChatOnly();
     if (chatOnly && !isChatOnlyAllowed(location.pathname)) {
diff --git a/studio/frontend/src/app/routes/settings.tsx b/studio/frontend/src/app/routes/settings.tsx
index fa97a450f7..84650bdb8b 100644
--- a/studio/frontend/src/app/routes/settings.tsx
+++ b/studio/frontend/src/app/routes/settings.tsx
@@ -7,10 +7,10 @@ import { useSettingsDialogStore } from "@/features/settings";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-// /settings is a deep link to the modal. Open it, then redirect home.
-// Tab title is driven by useSettingsDialogStore in __root.tsx since the
-// redirect means /settings never stays matched; staticData is just a
-// safety net if beforeLoad ever stops throwing.
+// /settings deep-links the modal: open it, then redirect home. Tab title is
+// driven by useSettingsDialogStore in __root.tsx since the redirect means
+// /settings never stays matched; staticData is a safety net if beforeLoad
+// ever stops throwing.
 export const Route = createRoute({
   getParentRoute: () => rootRoute,
   path: "/settings",
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 941c2e8a74..8d56d90f98 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -138,11 +138,8 @@ function getTourId(pathname: string): string | null {
   return null;
 }
 
-// Hugeicons' TestTube01Icon ships with two interior bubbles (paths #4
-// and #5 of the 5-path definition). Slicing to the first three paths
-// keeps the test-tube outline + horizontal cap + liquid line, dropping
-// the bubbles. The original export stays untouched, and HugeiconsIcon
-// renders this trimmed array exactly the same way.
+// TestTube01Icon's last 2 paths are interior bubbles; slice to the first
+// 3 (outline + cap + liquid line) to drop them. Original export untouched.
 const TestTubeOutlineIcon = TestTube01Icon.slice(
   0,
   3,
@@ -261,11 +258,11 @@ export function AppSidebar() {
 
   const scrollRef = useRef<HTMLDivElement | null>(null);
   const [scrolled, setScrolled] = useState(false);
-  // Bottom fade hides at the very bottom (and for short, non-scrolling lists)
-  // so the last row isn't washed out - Gemini-style.
+  // Bottom fade hides at the very bottom / for short lists so the last row
+  // isn't washed out (Gemini-style).
   const [canScrollDown, setCanScrollDown] = useState(false);
-  // Driven only from onScroll + a content-change effect below. Deliberately NO
-  // ResizeObserver: its callback-driven setState created a render loop (React
+  // Driven only from onScroll + a content-change effect below. No
+  // ResizeObserver: its callback-driven setState caused a render loop (React
   // #185). Both setters bail out when unchanged, so neither path can loop.
   const syncScrollState = (el: HTMLDivElement) => {
     const nextScrolled = el.scrollTop > 0;
@@ -311,10 +308,9 @@ export function AppSidebar() {
   const selectedHistoryRunId = useTrainingRuntimeStore((s) => s.selectedHistoryRunId);
   const setSelectedHistoryRunId = useTrainingRuntimeStore((s) => s.setSelectedHistoryRunId);
 
-  // Recompute the bottom-fade state on mount and whenever the list height can
-  // change (items load, sections collapse/expand, route switches the visible
-  // list) - onScroll never fires for short, non-scrolling lists. Guarded
-  // setState below means this can't loop even if a dep is a fresh reference.
+  // Recompute bottom-fade on mount and whenever list height can change
+  // (items load, sections toggle, route switch) - onScroll never fires for
+  // short, non-scrolling lists. Guarded setState below can't loop.
   useEffect(() => {
     const el = scrollRef.current;
     if (!el) return;
@@ -761,9 +757,8 @@ export function AppSidebar() {
               label={t("shell.navigation.search")}
               active={false}
               onClick={() => {
-                // Search is read-only over chat history and never runs
-                // inference, so it stays available while training (unlike
-                // New chat, which is gated on `chatDisabled`).
+                // Search is read-only and never runs inference, so it stays
+                // available while training (unlike New chat, gated on chatDisabled).
                 useChatSearchStore.getState().open();
                 closeMobileIfOpen();
               }}
@@ -796,8 +791,7 @@ export function AppSidebar() {
                   closeMobileIfOpen();
                 }}
               />
-              {/* Train has its own labelled section when expanded; surface it as
-                  a plain icon here only while the sidebar is collapsed. */}
+              {/* Train has a labelled section when expanded; plain icon here only when collapsed. */}
               <NavItem
                 icon={TestTubeOutlineIcon}
                 label={t("shell.navigation.train")}
@@ -899,12 +893,10 @@ export function AppSidebar() {
               <SidebarGroupContent className="px-2">
                 <SidebarMenu>
                   {runItems.map((run) => {
-                    // An explicit sidebar selection wins. Otherwise highlight
-                    // the active job only while the "Current Run" tab is the
-                    // view - that covers a live run (it auto-switches there) and
-                    // a just-finished/errored run you're still viewing, while
-                    // keeping the Configure tab unhighlighted even though
-                    // `activeJobId` stays pinned to the last job.
+                    // Explicit selection wins. Otherwise highlight the active
+                    // job only while the "Current Run" tab is the view, keeping
+                    // the Configure tab unhighlighted even though activeJobId
+                    // stays pinned to the last job.
                     const isActiveRun =
                       selectedHistoryRunId != null
                         ? run.id === selectedHistoryRunId
@@ -988,10 +980,9 @@ export function AppSidebar() {
       </SidebarContent>
 
       <SidebarFooter className="relative group-data-[collapsible=icon]:px-0">
-        {/* Fade above the profile box, shown only while there's more list below
-            the fold; at the very bottom (or for short lists) it fades out so the
-            last row shows fully (Gemini-style). `right-2` keeps it clear of the
-            8px scrollbar gutter so the scrollbar isn't faded out. */}
+        {/* Fade above the profile box, shown only when there's more list below
+            the fold; at the bottom (or short lists) it fades so the last row
+            shows fully (Gemini-style). right-2 keeps it clear of the 8px scrollbar gutter. */}
         <div
           aria-hidden="true"
           className={cn(
@@ -1082,8 +1073,8 @@ export function AppSidebar() {
                 </DropdownMenuItem>
                 <DropdownMenuItem
                   onSelect={async () => {
-                    // Best-effort server-side revocation; ignore network errors
-                    // so the local clear path still runs and the user lands on /login.
+                    // Best-effort server revocation; ignore network errors so
+                    // the local clear still runs and the user lands on /login.
                     try {
                       await logout();
                     } catch {
diff --git a/studio/frontend/src/components/assistant-ui/attachment.tsx b/studio/frontend/src/components/assistant-ui/attachment.tsx
index a96dd7a4e1..0a59d75964 100644
--- a/studio/frontend/src/components/assistant-ui/attachment.tsx
+++ b/studio/frontend/src/components/assistant-ui/attachment.tsx
@@ -164,8 +164,8 @@ const AttachmentUI: FC = () => {
         throw new Error(`Unknown attachment type: ${type as string}`);
     }
   });
-  // Include filename in accessible name so screen readers distinguish
-  // same-typed attachments. Sighted users get it via the tooltip.
+  // Filename in accessible name lets screen readers distinguish same-typed
+  // attachments. Sighted users get it via the tooltip.
   const accessibleName = name
     ? `${typeLabel} attachment: ${name}`
     : `${typeLabel} attachment`;
diff --git a/studio/frontend/src/components/assistant-ui/code-plugin.ts b/studio/frontend/src/components/assistant-ui/code-plugin.ts
index 5df7ac4f95..1e70871c06 100644
--- a/studio/frontend/src/components/assistant-ui/code-plugin.ts
+++ b/studio/frontend/src/components/assistant-ui/code-plugin.ts
@@ -10,8 +10,8 @@ import {
 } from "@streamdown/code";
 import type { BundledLanguage } from "shiki";
 
-// Fence tags LLMs/users commonly write that shiki doesn't expose as aliases.
-// Keys are lower-cased input; values are canonical shiki language ids.
+// Common fence tags shiki doesn't expose as aliases.
+// Keys: lower-cased input; values: canonical shiki language ids.
 const LANGUAGE_ALIAS_OVERRIDES: Record<string, BundledLanguage> = {
   objectivec: "objective-c",
   "obj-c": "objective-c",
diff --git a/studio/frontend/src/components/assistant-ui/code-themes.ts b/studio/frontend/src/components/assistant-ui/code-themes.ts
index 2557b45ef0..67db02b9b4 100644
--- a/studio/frontend/src/components/assistant-ui/code-themes.ts
+++ b/studio/frontend/src/components/assistant-ui/code-themes.ts
@@ -5,11 +5,10 @@ import oneDarkPro from "@shikijs/themes/one-dark-pro";
 import oneLight from "@shikijs/themes/one-light";
 import type { ThemeRegistrationAny } from "shiki";
 
-// Canonical Atom One Dark / One Light themes, shipped by `@shikijs/themes`.
-// We only override the background so the code block blends into the app's
-// `--code-block` surface instead of painting its own. Every token color and
-// scope mapping is left intact — that's what gives consistent multi-language
-// highlighting (including Objective-C, Go, Rust, etc.) out of the box.
+// Canonical Atom One Dark / One Light themes from `@shikijs/themes`. Only the
+// background is overridden so the code block blends into the app's `--code-block`
+// surface; all token colors/scopes are kept intact for consistent multi-language
+// highlighting out of the box.
 const withTransparentBg = (theme: ThemeRegistrationAny): ThemeRegistrationAny => ({
   ...theme,
   bg: "transparent",
diff --git a/studio/frontend/src/components/assistant-ui/image.tsx b/studio/frontend/src/components/assistant-ui/image.tsx
index 3cd7ecaaf5..0850a8230a 100644
--- a/studio/frontend/src/components/assistant-ui/image.tsx
+++ b/studio/frontend/src/components/assistant-ui/image.tsx
@@ -378,10 +378,7 @@ function ImageContentFilterError({
 
 export type ImageActionsProps = {
   part: ImageMessagePart;
-  /**
-   * Wire to your own generation call to show a regenerate button. The button
-   * renders only when this is set and the part carries a `prompt`.
-   */
+  /** Shows a regenerate button (only when set and the part has a `prompt`). */
   onRegenerate?: () => void | Promise<void>;
   className?: string;
 };
diff --git a/studio/frontend/src/components/assistant-ui/markdown-text.tsx b/studio/frontend/src/components/assistant-ui/markdown-text.tsx
index 107672d938..fc547da0b1 100644
--- a/studio/frontend/src/components/assistant-ui/markdown-text.tsx
+++ b/studio/frontend/src/components/assistant-ui/markdown-text.tsx
@@ -157,8 +157,7 @@ const UNSAFE_SVG_RE =
 
 function sanitizeSvg(source: string): string | null {
   if (UNSAFE_SVG_RE.test(source)) return null;
-  // Strip XML declaration (<?xml ...?>) -- not needed for data URI
-  // rendering and can cause issues with some renderers.
+  // Strip XML declaration: unneeded for data URIs and breaks some renderers.
   return source.replace(/^\s*<\?xml[^?]*\?>\s*/i, "");
 }
 
@@ -381,11 +380,10 @@ function StreamdownBlock(props: BlockProps) {
 }
 const AUDIO_PLAYER_RE = /<audio-player\s+src="([^"]+)"\s*\/>/;
 
-// Coalesce markdown re-parses to one per animation frame while streaming: the
-// runtime notifies on every token (hundreds/sec) and the monitor can't paint
-// that fast. When not streaming we return live text rather than the throttled
-// state, so the final text never lags and a reused instance (parts are keyed by
-// index) shows a completed message's text immediately instead of a stale frame.
+// Coalesce markdown re-parses to one per frame while streaming: tokens arrive
+// hundreds/sec, faster than the monitor can paint. When not streaming we return
+// live text (not the throttled state) so final text never lags and a reused
+// instance (parts keyed by index) shows completed text instead of a stale frame.
 function useRafCoalescedText(text: string, isStreaming: boolean): string {
   const [displayed, setDisplayed] = useState(text);
   const pendingRef = useRef(text);
@@ -408,9 +406,9 @@ function useRafCoalescedText(text: string, isStreaming: boolean): string {
     }
   }, [text, isStreaming]);
 
-  // Unmount cleanup. Cancel the in-flight rAF and null the handle so a
-  // StrictMode remount isn't gated out by a stale id. Kept separate from the
-  // scheduling effect so it doesn't cancel mid-stream and defeat the throttle.
+  // Unmount cleanup: cancel the in-flight rAF and null the handle so a
+  // StrictMode remount isn't gated by a stale id. Separate from the scheduling
+  // effect so it doesn't cancel mid-stream and defeat the throttle.
   useEffect(() => {
     return () => {
       if (rafRef.current !== null) {
diff --git a/studio/frontend/src/components/assistant-ui/message-timing.tsx b/studio/frontend/src/components/assistant-ui/message-timing.tsx
index 31f742bc5e..447fd763e3 100644
--- a/studio/frontend/src/components/assistant-ui/message-timing.tsx
+++ b/studio/frontend/src/components/assistant-ui/message-timing.tsx
@@ -52,10 +52,9 @@ export const MessageTiming: FC<{
   // Anthropic-only cache-write count.
   const cacheWrites = custom?.contextUsage?.cacheWriteTokens ?? 0;
 
-  // Guard unphysical tok/s: llama.cpp emits predicted_ms=0 on no-op
-  // turns, blowing the rate up to Infinity. Require >=1 token AND a
-  // non-zero decode window AND a finite rate. Fast cached single-token
-  // responses (sub-10ms) are legitimate and must stay visible.
+  // Guard unphysical tok/s: llama.cpp emits predicted_ms=0 on no-op turns,
+  // blowing the rate up to Infinity. Require >=1 token, a non-zero decode
+  // window, and a finite rate. Fast cached sub-10ms responses are legit.
   const hasPredicted =
     (st?.predicted_n ?? 0) >= 1 && (st?.predicted_ms ?? 0) > 0;
   const predictedRate =
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx b/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
index 42bd1716a1..3cb80ca2e1 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
@@ -32,12 +32,9 @@ export interface FolderBrowserProps {
 
 function splitBreadcrumb(path: string): { label: string; value: string }[] {
   if (!path) return [];
-  // Distinguish path styles BEFORE normalizing separators. On POSIX
-  // backslashes are valid filename characters, so we cannot blindly
-  // rewrite ``\`` -> ``/`` -- doing so would mangle directory names
-  // like ``my\backup`` into ``my/backup`` and produce breadcrumb
-  // values that 404 on the server. Only Windows-style absolute paths
-  // (drive letter, or UNC ``\\server\share``) get the conversion.
+  // Detect path style BEFORE normalizing: on POSIX, `\` is a valid filename
+  // char, so blindly rewriting `\` -> `/` mangles names like `my\backup` into
+  // 404ing breadcrumbs. Only Windows-style paths (drive letter, or UNC) convert.
   const isWindowsDrive = /^[A-Za-z]:[\\/]/.test(path) || /^[A-Za-z]:$/.test(path);
   const isUnc = /^\\\\/.test(path);
   const isWindows = isWindowsDrive || isUnc;
@@ -57,11 +54,9 @@ function splitBreadcrumb(path: string): { label: string; value: string }[] {
     return parts;
   }
 
-  // Windows-ish drive path (C:, D:): first segment is the drive. Use
-  // ``C:/`` (drive-absolute) as the crumb value so clicking the drive
-  // root navigates to the root of the drive rather than the
-  // drive-relative current working directory on that drive (``C:``
-  // alone resolves to ``CWD-on-C``, not ``C:\``).
+  // Windows drive path (C:, D:): first segment is the drive. Use `C:/` as the
+  // crumb value so clicking the drive root navigates to the drive root, not the
+  // drive-relative CWD (`C:` alone resolves to CWD-on-C, not `C:\`).
   if (/^[A-Za-z]:$/.test(segments[0])) {
     const driveRoot = `${segments[0]}/`;
     let cur = driveRoot;
@@ -102,8 +97,8 @@ export function FolderBrowser({
       abortRef.current = ctrl;
       setLoading(true);
       setError(null);
-      // Forward the signal so cancelled navigation actually cancels the
-      // backend enumeration instead of just discarding the response.
+      // Forward the signal so cancelled navigation aborts the backend
+      // enumeration, not just the response.
       browseFolders(target, hidden, ctrl.signal)
         .then((res) => {
           if (ctrl.signal.aborted) return;
@@ -112,17 +107,13 @@ export function FolderBrowser({
         })
         .catch((err) => {
           if (ctrl.signal.aborted) return;
-          // Surface the error, but if the very first request (typically
-          // a typo'd or denylisted ``initialPath``) fails AND the
-          // browser is empty (no ``data`` to render against), fall
-          // back to the user's HOME so the modal is navigable instead
-          // of an irrecoverable dead end.
+          // Surface the error; if the first request (e.g. a bad initialPath)
+          // fails, fall back to HOME so the modal stays navigable.
           const message = err instanceof Error ? err.message : String(err);
           setError(message);
           if (opts?.fallbackOnError && target !== undefined) {
             // Re-issue without a target -> backend defaults to HOME.
-            // Don't recurse if HOME itself fails (paranoia: shouldn't
-            // happen since the sandbox allowlist always includes HOME).
+            // Don't recurse if HOME itself fails (allowlist always has HOME).
             queueMicrotask(() => navigate(undefined, hidden));
           }
         })
@@ -133,15 +124,13 @@ export function FolderBrowser({
     [],
   );
 
-  // Fetch when the dialog opens.  Only re-run when the dialog transitions
-  // closed -> open; subsequent navigation is driven by `navigate()` so we
-  // don't want `path` in the dependency list here.
+  // Fetch only on closed -> open; later navigation is driven by `navigate()`,
+  // so `path` is deliberately kept out of the dependency list.
   // eslint-disable-next-line react-hooks/exhaustive-deps
   useEffect(() => {
     if (!open) return;
-    // ``fallbackOnError``: if the user-supplied ``initialPath`` is bad
-    // (typo, denylisted, deleted) we recover into HOME instead of
-    // showing an empty modal with no breadcrumbs/entries.
+    // fallbackOnError: recover into HOME if initialPath is bad, rather than
+    // showing an empty modal.
     navigate(initialPath, showHidden, { fallbackOnError: true });
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [open]);
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 2d715bf8de..516aa409fc 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -62,7 +62,7 @@ function dedupe(values: string[]): string[] {
   return [...new Set(values.filter(Boolean))];
 }
 
-/** Normalize a string for fuzzy search: lowercase, strip separators. */
+/** Lowercase and strip separators for fuzzy search. */
 function normalizeForSearch(s: string): string {
   return s.toLowerCase().replace(/[\s\-_\.]/g, "");
 }
@@ -296,14 +296,14 @@ function GgufVariantExpander({
     [gpuGb, gpuBudgetGb, totalBudgetGb],
   );
 
-  // If the backend-recommended variant is OOM, pick the largest fitting
-  // variant instead; if all are OOM, recommend the smallest one.
+  // If the recommended variant is OOM, pick the largest fitting one;
+  // if all are OOM, recommend the smallest.
   const effectiveRecommended = useMemo(() => {
     if (!variants || !gpuGb || gpuGb <= 0) return defaultVariant;
     const defaultV = variants.find((v) => v.quant === defaultVariant);
     if (defaultV && getGgufFit(defaultV.size_bytes) !== "oom")
       return defaultVariant;
-    // Default is OOM -- pick largest non-OOM variant (best quality that fits)
+    // Largest non-OOM variant (best quality that fits)
     const fitting = variants.filter((v) => getGgufFit(v.size_bytes) !== "oom");
     if (fitting.length > 0) {
       fitting.sort((a, b) => b.size_bytes - a.size_bytes);
@@ -460,8 +460,8 @@ function isGgufRepo(id: string, hintedIsGguf?: boolean): boolean {
 
 /** Extract param count label from model name (e.g. "Qwen3-0.6B" -> "0.6B"). */
 function extractParamLabel(id: string): string | undefined {
-  // Match patterns like "0.6B", "1B", "4B", "3.5B", "70B", "1.5B" etc.
   const name = id.split("/").pop() ?? id;
+  // Match patterns like "0.6B", "1B", "4B", "3.5B", "70B", "1.5B" etc.
   const match = name.match(/(?:^|[-_])(\d+(?:\.\d+)?)[Bb](?:[-_]|$)/);
   return match ? `${match[1]}B` : undefined;
 }
@@ -504,11 +504,10 @@ export function HubModelPicker({
   const { results, isLoading, isLoadingMore, fetchMore } =
     useHfModelSearch(debouncedQuery);
 
-  // Sets of lowercased repo ids that the store or HF search have
-  // confirmed are GGUF. Absence means "no hint" and lets hasGgufSuffix
-  // take over as fallback, rather than conflating unknown with known-
-  // not-GGUF. Keys are lowercased so that store IDs and HF search IDs
-  // that differ only by casing still match the same hint.
+  // Lowercased repo ids confirmed GGUF by the store or HF search.
+  // Absence means "no hint" -> hasGgufSuffix is the fallback (don't
+  // conflate unknown with known-not-GGUF). Lowercased so store and HF
+  // IDs differing only by casing match the same hint.
   const modelGgufIds = useMemo(() => {
     const ids = new Set<string>();
     for (const model of models) {
@@ -538,8 +537,8 @@ export function HubModelPicker({
   const [customFoldersCollapsed, setCustomFoldersCollapsed] = useState(false);
   const [recommendedCollapsed, setRecommendedCollapsed] = useState(false);
 
-  // Cached (already downloaded) repos -- use module-level cache so
-  // re-mounting the popover does not flash an empty "Downloaded" section.
+  // Cached (downloaded) repos -- module-level cache avoids flashing an
+  // empty "Downloaded" section when the popover re-mounts.
   const [cachedGguf, setCachedGguf] =
     useState<CachedGgufRepo[]>(_cachedGgufCache);
   const [cachedModels, setCachedModels] =
@@ -548,8 +547,7 @@ export function HubModelPicker({
     _cachedGgufCache.length > 0 || _cachedModelsCache.length > 0;
   const [cachedReady, setCachedReady] = useState(alreadyCached);
 
-  // LM Studio local models -- module-level cache so re-mounting the
-  // popover does not flash an empty section (same pattern as GGUF/models).
+  // LM Studio local models -- module-level cache, same pattern as above.
   const [lmStudioModels, setLmStudioModels] =
     useState<LocalModelInfo[]>(_lmStudioCache);
   const [customFolderModels, setCustomFolderModels] =
@@ -589,24 +587,20 @@ export function HubModelPicker({
   }, []);
 
   const handleAddFolder = useCallback(async (overridePath?: string) => {
-    // Accept an explicit path so the folder browser can submit the
-    // chosen path in the same tick it calls `setFolderInput`; reading
-    // `folderInput` alone would race the state update.
+    // Explicit path lets the folder browser submit in the same tick it
+    // calls `setFolderInput`; reading `folderInput` would race the update.
     const raw = overridePath !== undefined ? overridePath : folderInput;
     const trimmed = raw.trim();
     if (!trimmed || folderLoading) return;
     setFolderError(null);
     setFolderLoading(true);
-    // True when the request originated from the folder browser's
-    // ``onSelect`` (one-click "Use this folder"). In that flow the
-    // typed-input panel is closed, so the inline ``folderError``
-    // paragraph is invisible. Surface failures via toast instead so
-    // the action doesn't appear to silently no-op when the backend
-    // rejects (denylisted path, sandbox 403, etc.).
+    // From the folder browser's one-click "Use this folder": the typed-
+    // input panel is closed, so the inline folderError is invisible.
+    // Surface failures (denylisted path, sandbox 403, etc.) via toast.
     const fromBrowser = overridePath !== undefined;
     try {
       const created = await addScanFolder(trimmed);
-      // Backend returns existing row for duplicates, so deduplicate
+      // Backend returns the existing row for duplicates, so dedupe.
       const next = _scanFoldersCache.some((f) => f.id === created.id || f.path === created.path)
         ? _scanFoldersCache
         : [..._scanFoldersCache, created];
@@ -632,7 +626,7 @@ export function HubModelPicker({
   const handleRemoveFolder = useCallback(async (id: number) => {
     try {
       await removeScanFolder(id);
-      // Optimistic update so the folder disappears immediately
+      // Optimistic: drop it immediately.
       const next = _scanFoldersCache.filter((f) => f.id !== id);
       _scanFoldersCache = next;
       setScanFolders(next);
@@ -662,18 +656,17 @@ export function HubModelPicker({
   }, [refreshLocalModelsList]);
 
   useEffect(() => {
-    // Always refresh LM Studio + custom folder models (not gated by alreadyCached)
+    // Always refresh LM Studio + custom folder models (not gated by alreadyCached).
     refreshLocalModelsList();
     refreshScanFolders();
     listRecommendedFolders()
       .then(setRecommendedFolders)
       .catch(() => {});
 
-    // Always refetch cached GGUF/model lists. The module-level caches give
-    // an instant render with stale data (no spinner flash), but newly
-    // downloaded repos won't appear unless we re-hit the backend on every
-    // mount.  Initial state already has cachedReady=alreadyCached, so the
-    // background refresh is invisible when we already had data.
+    // Always refetch cached GGUF/model lists. The module-level caches render
+    // instantly with stale data (no spinner flash), but newly downloaded
+    // repos need a fresh backend hit. cachedReady=alreadyCached initially,
+    // so the background refresh is invisible when we already had data.
     let done = 0;
     const check = () => {
       if (++done >= 2) setCachedReady(true);
@@ -694,8 +687,8 @@ export function HubModelPicker({
       .finally(check);
   }, [refreshLocalModelsList, refreshScanFolders]);
 
-  // Deduplicate: don't show downloaded models in the recommended list.
-  // Compare case-insensitively since HF cache lowercases repo IDs.
+  // Hide downloaded models from the recommended list. Case-insensitive
+  // since the HF cache lowercases repo IDs.
   const downloadedSet = useMemo(() => {
     const s = new Set<string>();
     for (const c of cachedGguf) s.add(c.repo_id.toLowerCase());
@@ -756,10 +749,9 @@ export function HubModelPicker({
     return recommendedIds.filter((id) => normalizeForSearch(id).includes(q));
   }, [showHfSection, debouncedQuery, recommendedIds]);
 
-  // Fetch VRAM info for visible models, plus any models surfaced by a search
-  // query so that filtered recommended models also show VRAM badges.
-  // Skip GGUF repos: they have no safetensors metadata and the render layer
-  // already shows a static "GGUF" badge instead of VRAM data.
+  // VRAM info for visible models plus any surfaced by a search query, so
+  // filtered recommended models also show VRAM badges. Skip GGUF repos:
+  // no safetensors metadata, and the render layer shows a "GGUF" badge.
   const idsForVram = useMemo(() => {
     const ids = showHfSection
       ? [...new Set([...visibleRecommendedIds, ...filteredRecommendedIds])]
@@ -851,9 +843,8 @@ export function HubModelPicker({
   );
 
   // Sentinel + IntersectionObserver for recommended infinite scroll.
-  // We disconnect after each fire so the observer doesn't loop while
-  // React re-renders; the effect re-creates it on the next page.
-  // Uses a callback ref for the sentinel so we detect mount/unmount reliably.
+  // Disconnect after each fire so it doesn't loop during re-render; the
+  // effect re-creates it next page. Callback ref detects mount/unmount.
   const [recommendedSentinel, setRecommendedSentinel] =
     useState<HTMLDivElement | null>(null);
   const recommendedSentinelRef = useCallback((node: HTMLDivElement | null) => {
@@ -872,7 +863,7 @@ export function HubModelPicker({
       },
       { threshold: 0, root },
     );
-    // Small delay so the browser finishes layout after the previous page render
+    // Small delay so layout settles after the previous page render.
     const timer = setTimeout(() => obs.observe(recommendedSentinel), 100);
     return () => {
       clearTimeout(timer);
@@ -1184,9 +1175,8 @@ export function HubModelPicker({
                 onSelect={(picked) => {
                   setFolderInput(picked);
                   setFolderError(null);
-                  // One-click UX: the "Use this folder" button submits
-                  // the scan folder directly. Pass the path explicitly
-                  // because `folderInput` state hasn't flushed yet.
+                  // Pass the path explicitly: `folderInput` state hasn't
+                  // flushed yet when "Use this folder" submits.
                   void handleAddFolder(picked);
                 }}
               />
diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index 633c3bc3e9..3eaa21a19e 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -358,18 +358,18 @@ const ReasoningGroupImpl: ReasoningGroupComponent = ({
     }
   }, [isReasoningStreaming]);
 
-  // Reset dismissed flag when a new stream starts
+  // Reset dismissed flag on new stream.
   useEffect(() => {
     if (isReasoningStreaming) {
       setDismissedWhileStreaming(false);
     }
   }, [isReasoningStreaming]);
 
-  // Derived: open during streaming (unless dismissed), or if user manually opened after
+  // Open while streaming (unless dismissed), or once manually opened.
   const isOpen = (isReasoningStreaming && !dismissedWhileStreaming) || manualOpen;
   const variant = isOpen ? "outline" : "ghost";
 
-  // Allow closing during streaming (matches ChatGPT)
+  // Allow closing during streaming (matches ChatGPT).
   const handleOpenChange = useCallback(
     (open: boolean) => {
       if (isReasoningStreaming) {
diff --git a/studio/frontend/src/components/assistant-ui/sources.tsx b/studio/frontend/src/components/assistant-ui/sources.tsx
index 5593369bf9..7e5b52f7d0 100644
--- a/studio/frontend/src/components/assistant-ui/sources.tsx
+++ b/studio/frontend/src/components/assistant-ui/sources.tsx
@@ -128,9 +128,8 @@ function Source({
 
 interface SourceData {
   /**
-   * Stable per-citation key. Two Anthropic document citations into
-   * different spans of the same source share a ``url``, so React keys
-   * on ``id`` to keep each footnote distinct.
+   * Stable per-citation key. Two Anthropic citations into different spans of
+   * the same source share a `url`, so React keys on `id` to keep them distinct.
    */
   id: string;
   url: string;
@@ -185,7 +184,6 @@ const SourcesGroup: FC = () => {
   const [visibleCount, setVisibleCount] = useState<number | null>(null);
   const [expanded, setExpanded] = useState(false);
 
-  // Extract source parts from the message
   const sources: SourceData[] = [];
   if (message.content) {
     for (const part of message.content) {
@@ -220,7 +218,6 @@ const SourcesGroup: FC = () => {
     const children = Array.from(container.children) as HTMLElement[];
     if (children.length === 0) return;
 
-    // Find the top of the first child as baseline
     const firstTop = children[0].offsetTop;
     let rowCount = 1;
     let prevTop = firstTop;
@@ -263,17 +260,12 @@ const SourcesGroup: FC = () => {
 
   return (
     <div className="relative mt-2 mb-3">
-      {/* Hidden measurement container. Renders all badges off-screen so we
-          can read each child's offsetTop and decide how many fit in two
-          rows. Wrapped in an absolute, h-0, overflow-hidden box so the
-          measurement pills do NOT contribute to the viewport's scrollable
-          overflow region. Without this clip, every hidden source row
-          adds ~30px to scrollHeight, producing a phantom empty scroll
-          area below the message: visible to users as unbounded blank
-          space below the assistant action bar. The inner div still
-          flex-wraps its children for measurement; offsetTop reads
-          correctly because the wrapper is positioned (absolute) and the
-          children's offsetTop is measured relative to it. */}
+      {/* Hidden measurement container: renders all badges off-screen to read
+          each child's offsetTop and decide how many fit in two rows. The
+          absolute/h-0/overflow-hidden wrapper clips the pills so they don't add
+          to scrollHeight (~30px per row) and create a phantom empty scroll area
+          below the message. offsetTop reads correctly because the wrapper is
+          positioned (absolute) and the flex-wrapped children measure against it. */}
       <div
         aria-hidden
         className="absolute pointer-events-none overflow-hidden h-0 w-full left-0 top-0"
diff --git a/studio/frontend/src/components/assistant-ui/think-aria-label.ts b/studio/frontend/src/components/assistant-ui/think-aria-label.ts
index 9a7eb31a86..193a5a5ad1 100644
--- a/studio/frontend/src/components/assistant-ui/think-aria-label.ts
+++ b/studio/frontend/src/components/assistant-ui/think-aria-label.ts
@@ -1,8 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Shared aria-labels for the Think pill (toggle + effort dropdown) so both
-// controls use the same pre-load / unsupported-model wording.
+// Shared aria-labels for the Think pill toggle + effort dropdown, so both
+// use the same pre-load / unsupported-model wording.
 
 export interface ThinkToggleState {
   reasoningLockedOn: boolean;
@@ -27,8 +27,7 @@ export interface ThinkEffortState {
   reasoningEffort: string;
 }
 
-// Locked-on isn't special-cased: the effort dropdown stays interactive
-// (users can still pick an effort level).
+// Locked-on isn't special-cased: the effort dropdown stays interactive.
 export function thinkEffortAriaLabel(state: ThinkEffortState): string {
   if (!state.modelLoaded) return "Thinking (model not loaded)";
   if (state.reasoningDisabled)
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 6b94bfff6c..302c107e26 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -124,20 +124,18 @@ import {
   useState,
 } from "react";
 
-// True while a file is dragged anywhere over the chat page (not just the
-// composer), so the composer can show its "Drop files here" affordance.
+// True while a file is dragged anywhere over the chat page, so the composer
+// can show its "Drop files here" affordance.
 const PageDragContext = createContext(false);
 
-// Gap (px) between the last message and the floating composer. The bottom
-// spacer tracks composer height plus this gap so the chat can always be
-// scrolled fully above the composer.
+// Gap (px) between last message and floating composer; bottom spacer tracks
+// composer height plus this gap so chat can scroll fully above the composer.
 const COMPOSER_SCROLL_GAP_PX = 24;
 // The scroll-to-bottom footer sits 10px below the spacer top.
 const FOOTER_GAP_BELOW_SPACER_PX = 10;
-// Composer shrinks this soon after a run start (send clears the chips)
-// apply immediately: the run-start pin owns the bottom, so the clamp is
-// the intended glide. Covers instant responses where isRunning is
-// already false by the time the dock resize is observed.
+// Window after a run start during which composer shrinks apply immediately:
+// the run-start pin owns the bottom, so the clamp is the intended glide.
+// Covers instant responses where isRunning is already false by resize time.
 const RUN_SHRINK_WINDOW_MS = 1000;
 
 export const Thread: FC<{
@@ -145,10 +143,9 @@ export const Thread: FC<{
   hideWelcome?: boolean;
   targetThreadId?: string;
 }> = ({ hideComposer, hideWelcome, targetThreadId }) => {
-  // Intent-aware autoscroll: replaces assistant-ui's built-in autoscroll
-  // to prevent the streaming-mutation race that makes the viewport snap
-  // back to the bottom while the user is scrolling up (see the hook for
-  // the full explanation).
+  // Intent-aware autoscroll replaces assistant-ui's built-in autoscroll to
+  // prevent the streaming-mutation race that snaps the viewport back to the
+  // bottom while the user scrolls up (see the hook for the full explanation).
   const { ref: viewportRef, context: autoScrollContext } =
     useIntentAwareAutoScroll();
 
@@ -167,10 +164,9 @@ export const Thread: FC<{
       ? null
       : composerHeight + COMPOSER_SCROLL_GAP_PX - FOOTER_GAP_BELOW_SPACER_PX;
 
-  // The viewport element is owned by the autoscroll hook; mirror it
-  // locally for the spacer clamp math below. State, not a ref: the keyed
-  // provider below remounts the viewport on thread switches, and the
-  // scroll listener effect must re-attach to the new element.
+  // Viewport element is owned by the autoscroll hook; mirror it locally for
+  // the spacer clamp math. State, not a ref: the keyed provider remounts the
+  // viewport on thread switches and the scroll listener must re-attach.
   const [viewportEl, setViewportEl] = useState<HTMLElement | null>(null);
   const composedViewportRef = useCallback(
     (node: HTMLElement | null) => {
@@ -180,13 +176,13 @@ export const Thread: FC<{
     [viewportRef],
   );
 
-  // Bottom spacer sizing. Invariant: the chat never moves on its own when
-  // the composer resizes.
-  // - Grow (attachment added, multiline input): grow the spacer at once.
-  //   Growth below the scroll position is invisible and only adds room.
+  // Bottom spacer sizing. Invariant: chat never moves on its own on composer
+  // resize.
+  // - Grow (attachment added, multiline): grow at once; growth below the
+  //   scroll position is invisible and only adds room.
   // - Shrink (attachment removed): shrinking scrollHeight near the bottom
-  //   would clamp scrollTop and yank the chat down. Defer the shrink until
-  //   it is invisible (user scrolled up) or a bottom-pinning moment.
+  //   clamps scrollTop and yanks the chat down. Defer until invisible (user
+  //   scrolled up) or a bottom-pinning moment.
   // Applied imperatively so a remounted spacer can be sized from refs even
   // when composerHeight did not change (e.g. thread switch).
   const spacerElRef = useRef<HTMLDivElement | null>(null);
@@ -214,8 +210,8 @@ export const Thread: FC<{
   const spacerRef = useCallback(
     (node: HTMLDivElement | null) => {
       spacerElRef.current = node;
-      // Fresh mounts (thread switch, first message) start at the desired
-      // size; deferral state from a previous mount is moot.
+      // Fresh mounts (thread switch, first message) start at desired size;
+      // deferral state from a previous mount is moot.
       const desired = desiredSpacerPxRef.current;
       if (node && desired != null) {
         applySpacerPx(desired);
@@ -249,7 +245,7 @@ export const Thread: FC<{
         aui.thread().getState().isRunning ||
         performance.now() - runStartAtRef.current < RUN_SHRINK_WINDOW_MS;
       // At the bottom the shrink only drops blank spacer, so apply it now
-      // instead of stranding dead space until the next pin.
+      // rather than strand dead space until the next pin.
       if (
         runOwnsBottom ||
         distance >= applied - desired ||
@@ -260,20 +256,20 @@ export const Thread: FC<{
       // else: deferred; released on scroll or a bottom-pinning event.
     }
     if (prev != null && composerHeight > prev) {
-      // The chat is now above the new bottom. Detach as if the user had
-      // scrolled up so no later signal re-pins and shoves the chat up.
-      // Scrolling back down re-attaches; explicit pins still work.
-      // Mid-run growth comes from tool-status rows, not the user, and
-      // detaching then would break streaming autoscroll, so skip it.
+      // Chat is now above the new bottom. Detach as if the user scrolled up
+      // so no later signal re-pins and shoves the chat up (scrolling back
+      // down re-attaches; explicit pins still work). Skip mid-run: that
+      // growth is tool-status rows, not the user, and detaching would break
+      // streaming autoscroll.
       if (!aui.thread().getState().isRunning) {
         autoScrollContext.detachFromBottom();
       }
     }
   }, [composerHeight, hideComposer, autoScrollContext, aui, applySpacerPx, viewportEl]);
 
-  // Drop deferred spacer excess as soon as the user has scrolled far
-  // enough above the bottom that the shrink cannot clamp scrollTop.
-  // Keyed on viewportEl so the listener follows viewport remounts.
+  // Drop deferred spacer excess once the user has scrolled far enough above
+  // the bottom that the shrink cannot clamp scrollTop. Keyed on viewportEl
+  // so the listener follows viewport remounts.
   useEffect(() => {
     const el = viewportEl;
     if (!el) {
@@ -303,10 +299,10 @@ export const Thread: FC<{
   useAuiEvent("thread.initialize", releaseSpacerExcess);
   useAuiEvent("threadListItem.switchedTo", releaseSpacerExcess);
 
-  // Page-wide drag-and-drop: dropping a file anywhere on the chat page (not
-  // just on the composer) attaches it and shows the composer drop affordance.
-  // The composer's own dropzone still handles drops on the box itself; its
-  // handler calls preventDefault, so the page handler skips them (no double-add).
+  // Page-wide drag-and-drop: dropping a file anywhere on the chat page
+  // attaches it and shows the composer drop affordance. The composer's own
+  // dropzone handles drops on the box and calls preventDefault, so the page
+  // handler skips them (no double-add).
   const [pageDragging, setPageDragging] = useState(false);
   const dragDepth = useRef(0);
   const hasFiles = (e: ReactDragEvent) =>
@@ -332,8 +328,8 @@ export const Thread: FC<{
     // Compare panes hide this composer and use the shared composer's own
     // dropzone, so don't capture drops into a hidden composer here.
     if (hideComposer) return;
-    // Drops on the composer box are handled by its own dropzone, which calls
-    // preventDefault; skip those here so the file isn't added twice.
+    // Drops on the composer box are handled by its dropzone (preventDefault);
+    // skip those here so the file isn't added twice.
     if (e.defaultPrevented) return;
     const files = Array.from(e.dataTransfer.files);
     if (files.length === 0) return;
@@ -391,10 +387,9 @@ export const Thread: FC<{
               }}
             />
 
-            {/* Bottom slack so the last message has breathing room above the
-            sticky scroll-to-bottom button (and the floating composer in
-            single mode). Without this, content would butt against the
-            sticky footer and feel cramped. */}
+            {/* Bottom slack so the last message has room above the sticky
+            scroll-to-bottom button (and floating composer in single mode),
+            instead of butting against the footer. */}
             <AuiIf condition={({ thread }) => hideWelcome || !thread.isEmpty}>
               <div
                 ref={spacerRef}
@@ -562,8 +557,8 @@ const ThreadComposerDock: FC<{
 }> = ({ disabled, threadId, onHeightChange }) => {
   const { overlay } = useGeneratedImageOverlay();
 
-  // Report the dock's rendered height so the viewport can reserve matching
-  // scroll space when attachments or multiline input grow the composer.
+  // Report dock height so the viewport reserves matching scroll space when
+  // attachments or multiline input grow the composer.
   const dockRef = useRef<HTMLDivElement | null>(null);
   useEffect(() => {
     const el = dockRef.current;
@@ -608,12 +603,12 @@ const ThreadComposerDock: FC<{
 };
 
 const ThreadScrollToBottom: FC = () => {
-  // State and action both come from our IntentAwareScrollProvider (scoped
-  // per Thread, so compare panes are independent). We deliberately
-  // avoid `ThreadPrimitive.ScrollToBottom` + `useThreadViewport` to
-  // stay off assistant-ui's internal autoscroll path — see the hook
-  // for why. The button stays mounted and toggles via CSS; unmounting
-  // would trip the hook's MutationObserver as a content change.
+  // State and action both come from our IntentAwareScrollProvider (per-Thread
+  // scope, so compare panes are independent). We avoid
+  // `ThreadPrimitive.ScrollToBottom` + `useThreadViewport` to stay off
+  // assistant-ui's internal autoscroll path (see the hook). The button stays
+  // mounted and toggles via CSS; unmounting would trip the hook's
+  // MutationObserver as a content change.
   const isAtBottom = useIsThreadAtBottom();
   const scrollToBottom = useScrollThreadToBottom();
   return (
@@ -634,9 +629,9 @@ const ThreadScrollToBottom: FC = () => {
 const pickRandom = <T,>(arr: T[]): T =>
   arr[Math.floor(Math.random() * arr.length)];
 
-// Each greeting carries the sloth picture that best fits it, so a given line
-// always shows the same mascot. Greeting varies by local time; name-bearing
-// lines drop the name when none is set.
+// Each greeting carries its matching sloth picture so a line always shows the
+// same mascot. Greeting varies by local time; name-bearing lines drop the
+// name when none is set.
 type Welcome = { text: string; sloth: string };
 const DEFAULT_WELCOME: Welcome = {
   text: "What’s on your mind today?",
@@ -645,9 +640,8 @@ const DEFAULT_WELCOME: Welcome = {
 
 function buildWelcome(hour: number, name: string): Welcome {
   const g = (text: string, sloth: string): Welcome => ({ text, sloth });
-  // Use the name on roughly a third of the lines per time of day: only the
-  // direct salutations where it reads most naturally. Everything else stays
-  // name-free so the greeting doesn't feel repetitive.
+  // Use the name on ~a third of lines (only direct salutations where it reads
+  // naturally); the rest stay name-free so greetings don't feel repetitive.
   const base: Welcome[] = [
     g(name ? `Good to see you, ${name}.` : "Good to see you.", "large sloth wave.png"),
     g("Ready when you are.", "large sloth thumbs.png"),
@@ -696,7 +690,7 @@ const ThreadWelcome: FC<{
     <div className="aui-thread-welcome-root mx-auto my-auto flex w-full max-w-(--thread-max-width) grow flex-col">
       <div className="aui-thread-welcome-center flex w-full grow flex-col items-center justify-start pt-[28vh]">
         <div className="aui-thread-welcome-message flex w-full flex-col justify-center gap-9 px-4">
-          {/* Center the whole greeting (sloth + title) over the composer. */}
+          {/* Center the greeting (sloth + title) over the composer. */}
           <div className="flex flex-row items-center justify-center gap-[15px]">
             <img
               src={currentEmojiSrc}
@@ -784,8 +778,8 @@ const Composer: FC<{
   );
   const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
   const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
-  // With more than 4 pills showing, collapse them to icons only to cut clutter.
-  // Search and Code always show; Images, Canvas and MCP are conditional.
+  // More than 4 pills: collapse to icons only. Search and Code always show;
+  // Images, Canvas and MCP are conditional.
   const pillsCompact =
     2 +
       (supportsBuiltinImageGeneration ? 1 : 0) +
@@ -802,7 +796,7 @@ const Composer: FC<{
   // Expand only once the input wraps to a second line, not on first keystroke.
   // Latch until cleared so it can't flip-flop at the wrap boundary.
   const inputRef = useRef<HTMLTextAreaElement>(null);
-  // Cache line metrics so getComputedStyle runs once, not on every keystroke.
+  // Cache line metrics so getComputedStyle runs once, not per keystroke.
   const lineMetricsRef = useRef<{ lineHeight: number; padding: number } | null>(
     null,
   );
@@ -844,8 +838,8 @@ const Composer: FC<{
   const referenceThreadId = threadId ?? activeThreadId ?? null;
   const hasSendableContent =
     composerText.trim().length > 0 || hasAttachments || hasPendingAudio;
-  // Two-row layout shows once the input wraps or a tool is on. Tools pre-select
-  // before a model loads, so an active toggle expands the composer either way.
+  // Two-row layout shows once the input wraps or a tool is on. Tools can
+  // pre-select before a model loads, so an active toggle expands it either way.
   const composerExpanded =
     isMultiline ||
     hasAttachments ||
@@ -857,7 +851,7 @@ const Composer: FC<{
     mcpEnabledForChat;
   // react-textarea-autosize re-measures only on value change or window resize,
   // not on the width swap from expanding, so it keeps the taller height and
-  // leaves a stray blank row. Nudge a resize whenever the input width changes.
+  // leaves a stray blank row. Nudge a resize whenever input width changes.
   useEffect(() => {
     const el = inputRef.current;
     if (!el || typeof ResizeObserver === "undefined") {
@@ -872,8 +866,8 @@ const Composer: FC<{
         return;
       }
       lastWidth = width;
-      // Re-measure after layout settles. An immediate dispatch races autosize's
-      // own measurement (stale pre-expand width); 0ms + 64ms wins it, no flash.
+      // Re-measure after layout settles. An immediate dispatch races
+      // autosize's own measurement (stale pre-expand width); 0ms + 64ms wins.
       while (pending.length) {
         clearTimeout(pending.pop());
       }
@@ -893,8 +887,8 @@ const Composer: FC<{
       observer.disconnect();
     };
   }, []);
-  // Docked composer opens upward; the centered welcome composer opens downward
-  // by default and only flips up via collision detection when it would not fit.
+  // Docked composer opens upward; the welcome composer opens downward by
+  // default and only flips up via collision detection when it won't fit.
   const effectiveMenuSide = menuSide ?? "bottom";
   const shouldBlockSend = useCallback(
     () =>
@@ -1027,17 +1021,17 @@ const Composer: FC<{
       onSubmit={handleSubmit}
     >
       {isTauri ? (
-        // Phase 1 native model drops own Tauri local-path drops. Restore browser
-        // attachment drops in Tauri when Phase 1d adds attachment-token bridging.
+        // Phase 1 native model owns Tauri local-path drops. Restore browser
+        // attachment drops in Tauri once Phase 1d adds token bridging.
         <div className="aui-composer-attachment-dropzone unsloth-composer-surface">
           {composerContent}
         </div>
       ) : (
         <ComposerPrimitive.AttachmentDropzone className="group/dropzone aui-composer-attachment-dropzone unsloth-composer-surface relative">
           {composerContent}
-          {/* Gemini-style drop affordance: shown only while a file is dragged
-              over the composer. Absolutely positioned + pointer-events-none so
-              the dashed outline adds no layout shift and the drop still lands. */}
+          {/* Gemini-style drop affordance, shown while a file is dragged over
+              the composer. Absolute + pointer-events-none so the outline adds
+              no layout shift and the drop still lands. */}
           <div
             className={cn(
               "aui-composer-drop-overlay pointer-events-none absolute inset-0 z-20 flex flex-col items-center justify-center gap-1 overflow-hidden rounded-[32px] bg-background/90 opacity-0 backdrop-blur-sm transition-opacity duration-150 group-data-[dragging=true]/dropzone:opacity-100 dark:bg-card/90",
@@ -1063,13 +1057,12 @@ function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
 }
 
-// Fallback timeout for stuck IME composition. When Chrome on Windows talks
-// to a WSL-hosted Studio (issue #5546), `compositionend` never fires after
-// the candidate is committed, so `composingRef` stays true and Send stays
-// disabled. Every compositionupdate / non-composing input resets the timer;
-// only a true gap-after-commit lets it fire. 2500ms is well above a normal
-// candidate-window pause but short enough to recover before the user
-// notices the Send button is stuck.
+// Fallback timeout for stuck IME composition. With Chrome on Windows against
+// a WSL-hosted Studio (issue #5546), `compositionend` never fires after the
+// candidate commits, so `composingRef` stays true and Send stays disabled.
+// Every compositionupdate / non-composing input resets the timer; only a true
+// gap-after-commit lets it fire. 2500ms is above a normal candidate-window
+// pause but short enough to recover before the user notices Send is stuck.
 const IME_STUCK_TIMEOUT_MS = 2500;
 
 function useImeComposerInputHandlers() {
@@ -1153,13 +1146,12 @@ function useImeComposerInputHandlers() {
   );
 
   // If the watchdog cleared the composing flags during a long candidate-window
-  // pause, a subsequent IME keypress (browser-side isComposing=true / IME
-  // keyCode 229) would otherwise reach handleSubmit with composingRef=false
-  // and submit the preedit text. Re-arm composingRef synchronously from the
-  // native event so the form-submit gate keeps blocking until compositionend.
-  // Re-arm the watchdog at the same time — otherwise the WSL+Chrome path
-  // this PR targets (no compositionend, no follow-up input event) would
-  // leave composingRef pinned true indefinitely and Send blocked again.
+  // pause, a later IME keypress (isComposing=true / keyCode 229) would reach
+  // handleSubmit with composingRef=false and submit the preedit text. Re-arm
+  // composingRef synchronously from the native event so the submit gate keeps
+  // blocking until compositionend. Re-arm the watchdog too, or the WSL+Chrome
+  // path (no compositionend, no follow-up input) would pin composingRef true
+  // forever and block Send again.
   const onKeyDown = useCallback(
     (e: KeyboardEvent<HTMLTextAreaElement>) => {
       if (e.nativeEvent.isComposing || e.keyCode === 229) {
@@ -1207,8 +1199,8 @@ const ComposerAudioMenuItem: FC = () => {
   );
 
   // Build the input on document.body, not in the menu: selecting the item
-  // closes the dropdown, which would unmount a menu-rendered input before the
-  // OS picker returns and drop the file.
+  // closes the dropdown, unmounting a menu-rendered input before the OS picker
+  // returns and dropping the file.
   const pickAudio = useCallback(() => {
     const input = document.createElement("input");
     input.type = "file";
@@ -1573,9 +1565,8 @@ const WebSearchToggle: FC = () => {
   const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
   const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
   // External providers (OpenAI today) expose a server-side web_search tool
-  // even when the local tool runtime is unavailable — gate the Search pill
-  // on either source so it lights up on external models too. Mirror of
-  // shared-composer's searchDisabled.
+  // even without the local tool runtime; gate the pill on either source so it
+  // lights up on external models too. Mirror of shared-composer's searchDisabled.
   const supportsBuiltinWebSearch = useChatRuntimeStore(
     (s) => s.supportsBuiltinWebSearch,
   );
@@ -1594,7 +1585,7 @@ const WebSearchToggle: FC = () => {
       : undefined;
   const isKimiExternal = selectedExternalProvider?.providerType === "kimi";
   // Disable only when a loaded model lacks the capability; with no model the
-  // tool can still be pre-selected and reflected, matching the + menu.
+  // tool can still be pre-selected, matching the + menu.
   const disabled = modelLoaded && !(supportsTools || supportsBuiltinWebSearch);
 
   return (
@@ -1605,9 +1596,8 @@ const WebSearchToggle: FC = () => {
         const next = !toolsEnabled;
         setToolsEnabled(next);
         // Kimi's $web_search builtin requires thinking=disabled (see
-        // https://platform.kimi.ai/docs/guide/use-web-search). Keep
-        // the two pills mutually exclusive so the visible state always
-        // matches what the backend ends up sending.
+        // https://platform.kimi.ai/docs/guide/use-web-search). Keep the two
+        // pills mutually exclusive so visible state matches what's sent.
         if (isKimiExternal) {
           setReasoningEnabled(!next, { persist: false });
           applyQwenThinkingParams(!next);
@@ -1630,18 +1620,17 @@ const CodeToolsToggle: FC = () => {
     (s) => !!s.params.checkpoint && !s.modelLoading,
   );
   const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
-  // External providers have no local tool runtime, but Anthropic's
-  // Claude 4.x dispatches code_execution_20250825 server-side. The
-  // chat-page resolver stashes that capability in the runtime store
-  // (next to supportsBuiltinWebSearch). Mirror of shared-composer's
-  // codeDisabled so this pill lights up in active threads too.
+  // External providers have no local tool runtime, but Anthropic's Claude 4.x
+  // dispatches code_execution_20250825 server-side; the chat-page resolver
+  // stashes that capability in the runtime store (next to
+  // supportsBuiltinWebSearch). Mirror of shared-composer's codeDisabled.
   const supportsBuiltinCodeExecution = useChatRuntimeStore(
     (s) => s.supportsBuiltinCodeExecution,
   );
   const codeToolsEnabled = useChatRuntimeStore((s) => s.codeToolsEnabled);
   const setCodeToolsEnabled = useChatRuntimeStore((s) => s.setCodeToolsEnabled);
   // Disable only when a loaded model lacks the capability; with no model the
-  // tool can still be pre-selected and reflected, matching the + menu.
+  // tool can still be pre-selected, matching the + menu.
   const disabled = modelLoaded && !(supportsTools || supportsBuiltinCodeExecution);
 
   return (
@@ -1672,9 +1661,8 @@ const ImagesToggle: FC = () => {
     (s) => !!s.params.checkpoint && !s.modelLoading,
   );
   // OpenAI cloud Responses-API models advertise image_generation as a
-  // server-side tool; no local runtime fallback exists. Mirror of
-  // shared-composer's imageDisabled / showImagePill so the in-thread
-  // composer surfaces the same control as the empty-state composer.
+  // server-side tool; no local runtime fallback. Mirror of shared-composer's
+  // imageDisabled / showImagePill so this composer matches the empty state.
   const supportsBuiltinImageGeneration = useChatRuntimeStore(
     (s) => s.supportsBuiltinImageGeneration,
   );
@@ -1755,10 +1743,9 @@ const ToolStatusDisplay: FC = () => {
 
     setElapsed(0);
 
-    // Debounce badge visibility by 300ms when the badge is not
-    // already on screen. Once visible from a prior tool, consecutive
-    // tools show immediately so the badge does not flicker. Fast
-    // tool calls that all complete under 300ms never show the badge.
+    // Debounce visibility by 300ms when the badge isn't already on screen.
+    // Once visible from a prior tool, later tools show immediately so it
+    // doesn't flicker; tool calls under 300ms never show the badge.
     let showTimer: ReturnType<typeof setTimeout> | undefined;
     if (!visibleRef.current) {
       showTimer = setTimeout(() => setVisible(true), 300);
@@ -1790,8 +1777,8 @@ const ToolStatusDisplay: FC = () => {
     </div>
   );
 };
-// Plus menu: attachment and workflow actions. Opens downward in the centered
-// welcome composer; the docked composer passes side="top" to open upward.
+// Plus menu: attachment and workflow actions. Opens downward in the welcome
+// composer; the docked composer passes side="top" to open upward.
 const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   side = "bottom",
 }) => {
@@ -1806,7 +1793,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   const setMcpEnabledForChat = useChatRuntimeStore(
     (s) => s.setMcpEnabledForChat,
   );
-  // Capability gating, mirroring the visible pills so menu and pills agree on
+  // Capability gating mirrors the visible pills so menu and pills agree on
   // what a loaded model supports (a tool the backend drops must not look on).
   const modelLoaded = useChatRuntimeStore(
     (s) => !!s.params.checkpoint && !s.modelLoading,
@@ -1889,8 +1876,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
         sideOffset={2}
         avoidCollisions={true}
         className="unsloth-plus-menu w-[212px]"
-        // Don't refocus the + on close; the restored focus showed a stray
-        // focus-visible ring.
+        // Don't refocus the + on close; restored focus showed a stray ring.
         onCloseAutoFocus={(event) => event.preventDefault()}
       >
         <ComposerPrimitive.AddAttachment asChild={true}>
diff --git a/studio/frontend/src/components/assistant-ui/tool-call-spinner.tsx b/studio/frontend/src/components/assistant-ui/tool-call-spinner.tsx
index bc91408789..79789692cb 100644
--- a/studio/frontend/src/components/assistant-ui/tool-call-spinner.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-call-spinner.tsx
@@ -6,11 +6,7 @@
 import { Loader2Icon } from "lucide-react";
 import { cn } from "@/lib/utils";
 
-/**
- * Spinner shown while a tool call is running: a clean circular arc with a
- * rounded cap (lucide Loader2 / LoaderCircle), animated. Inherits the current
- * text color so it matches the surrounding font.
- */
+/** Spinner shown while a tool call runs. Inherits text color to match its surroundings. */
 export function ToolCallSpinner({ className }: { className?: string }) {
   return (
     <Loader2Icon
diff --git a/studio/frontend/src/components/assistant-ui/tool-group.tsx b/studio/frontend/src/components/assistant-ui/tool-group.tsx
index d30c0a8ee1..569e673da3 100644
--- a/studio/frontend/src/components/assistant-ui/tool-group.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-group.tsx
@@ -217,8 +217,8 @@ const ToolGroupImpl: FC<
       ),
   );
 
-  // Single tool calls and artifacts render directly so cards never hide inside
-  // a collapsed tool group.
+  // Render single tool calls and artifacts directly so cards never hide in a
+  // collapsed group.
   if (toolCount <= 1 || containsArtifactTool) {
     return <>{children}</>;
   }
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-code-execution.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-code-execution.tsx
index 2573509c34..e27441183e 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-code-execution.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-code-execution.tsx
@@ -23,12 +23,9 @@ import {
 } from "./tool-fallback";
 
 /**
- * Renders the synthetic `_toolEvent` chunks emitted by
- * `_stream_anthropic` when Anthropic's `code_execution_20250825` tool
- * fires. The backend collapses Anthropic's two sub-tools
- * (`bash_code_execution`, `text_editor_code_execution`) into a single
- * `tool_name: "code_execution"`, with `arguments.kind` ("bash" or
- * "text_editor") and a per-kind argument shape:
+ * Renders synthetic `_toolEvent` chunks from `_stream_anthropic` for the
+ * `code_execution_20250825` tool. The backend collapses Anthropic's two
+ * sub-tools into `tool_name: "code_execution"` with `arguments.kind`:
  *
  *   kind=bash:        { command: "<shell command>" }
  *   kind=text_editor: { command: "view"|"create"|"str_replace", path, ... }
@@ -141,9 +138,8 @@ const CodeExecutionToolUIImpl: ToolCallMessagePartComponent = ({
     completedLabel = commandLabel ? `Ran \`${commandLabel}\`` : "Ran command";
   }
 
-  // Collapse the card once the model has resumed streaming prose after
-  // the tool call. Mirrors WebSearchToolUI's behavior so the tool-card
-  // doesn't crowd the final answer once the run is done.
+  // Collapse the card once the model resumes streaming prose after the tool
+  // call (mirrors WebSearchToolUI) so it doesn't crowd the final answer.
   const hasText = useAuiState(({ message }) =>
     message.content.some(
       (p) =>
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
index c0ced3b844..edce6200d5 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
@@ -21,26 +21,14 @@ import {
 
 /**
  * Renders the synthetic `_toolEvent` chunks emitted by
- * `_stream_openai_responses` when OpenAI's Responses-API
- * `image_generation` tool fires. The backend stashes the base64
- * PNG/WebP/JPEG (the gpt-image backbone output) on an `image_b64`
- * field of the tool_end event so the JSON result stays small, and the
- * adapter repackages it into a structured `result` shape:
- *
- *   {
- *     image_b64: string,
- *     image_mime: string,        // e.g. "image/png"
- *     size?: string,             // "1024x1024" etc
- *     quality?: string,
- *     background?: string,
- *   }
- *
- * The corresponding `tool_start` carries the prompt as
- * `args.prompt` (after gpt-image's revision pass) plus `args.kind:
- * "image"`. Without this component the generic ToolFallback would
- * print the prompt as JSON args text with an empty Result block --
- * which is exactly the "no image" symptom users hit before this UI
- * landed.
+ * `_stream_openai_responses` when OpenAI's Responses-API `image_generation`
+ * tool fires. The backend stashes the base64 image on `image_b64` of the
+ * tool_end event (keeping the JSON small); the adapter repackages it into a
+ * structured `result` (image_b64, image_mime e.g. "image/png", size? e.g.
+ * "1024x1024", quality?, background?).
+ * The `tool_start` carries the revised prompt as `args.prompt` plus
+ * `args.kind: "image"`. Without this, ToolFallback would print the prompt as
+ * JSON with an empty Result block (the "no image" symptom).
  */
 interface ImageGenerationArgs {
   prompt?: string;
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
index 642340bb5f..1231ca4a7f 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
@@ -72,7 +72,7 @@ function CopyBtn({ text }: { text: string }) {
   );
 }
 
-/** Render code with syntax highlighting via Streamdown + shiki. No extra borders — inherits parent container. */
+/** Syntax-highlighted code via Streamdown + shiki; inherits parent container. */
 function HighlightedCode({ code: source, language }: { code: string; language: string }) {
   const markdown = useMemo(
     () => `\`\`\`${language}\n${truncate(source)}\n\`\`\``,
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
index d5b72e06ea..7b9d555946 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
@@ -17,8 +17,8 @@ import { BrowserIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { memo, useEffect } from "react";
 
-// Context7 assistant-ui docs: tool UIs can read streaming args via
-// useToolArgsStatus, so render_html does not need to wait for tool completion.
+// Per Context7 assistant-ui docs: tool UIs read streaming args via
+// useToolArgsStatus, so render_html need not wait for tool completion.
 type RenderHtmlArgs = Record<string, unknown> & {
   code?: string;
   title?: string;
@@ -41,9 +41,8 @@ const RenderHtmlToolUIImpl: ToolCallMessagePartComponent = ({
   const isRunning = status?.type === "running";
   const codeIsStreaming = propStatus.code === "streaming";
 
-  // Surface the backend error when the tool call completed with invalid
-  // args.  Backend success results start with "Rendered HTML artifact";
-  // error results start with "Error:".
+  // Surface the backend error when the tool call completed with invalid args.
+  // Success results start with "Rendered HTML artifact"; errors with "Error:".
   const errorText =
     status?.type === "complete" &&
     typeof result === "string" &&
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-web-search.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-web-search.tsx
index ef8cb4a89b..062d0b1370 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-web-search.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-web-search.tsx
@@ -25,9 +25,8 @@ const RE_URL = /URL:\s*(.+)/;
 const RE_SNIPPET = /Snippet:\s*(.+)/s;
 
 /**
- * Reject anything that is not a real http(s) URL. Web-search / web-fetch
- * output is provider-controlled, so hostile ``javascript:`` / ``data:``
- * lines must not reach the Source badge's <a href>.
+ * Reject non-http(s) URLs. Web-search/fetch output is provider-controlled,
+ * so hostile `javascript:` / `data:` lines must not reach a Source <a href>.
  */
 function isSafeHttpUrl(raw: string): boolean {
   const value = raw.trim();
diff --git a/studio/frontend/src/components/assistant-ui/use-intent-aware-autoscroll.tsx b/studio/frontend/src/components/assistant-ui/use-intent-aware-autoscroll.tsx
index e4c17a5021..9488ca1db9 100644
--- a/studio/frontend/src/components/assistant-ui/use-intent-aware-autoscroll.tsx
+++ b/studio/frontend/src/components/assistant-ui/use-intent-aware-autoscroll.tsx
@@ -18,57 +18,42 @@ import {
 /**
  * Intent-aware autoscroll for a Thread viewport.
  *
- * Why we don't reuse assistant-ui's built-in autoscroll:
- *   `useThreadViewportAutoScroll` runs unconditionally whenever
- *   `ThreadPrimitive.Viewport` is mounted. Even with every
- *   `scrollToBottomOn*` prop disabled, it still installs observers
- *   that write `isAtBottom` to the shared viewport store on every
- *   layout change. On sidebar toggles, browser resizes, or
- *   mobile↔desktop breakpoint crossings, that write races our scroll
- *   correction — whoever writes last wins, and the scroll-to-bottom
- *   button flickers or sticks depending on observer ordering.
+ * Why not assistant-ui's built-in: `useThreadViewportAutoScroll` always
+ * installs observers that write `isAtBottom` to the shared store on every
+ * layout change, even with `scrollToBottomOn*` disabled. On resizes /
+ * breakpoint crossings that write races our correction and the
+ * scroll-to-bottom button flickers depending on observer ordering.
  *
  * Strategy:
- *   - Own `isAtBottom` as local state (exposed via
- *     `useIsThreadAtBottom`). Upstream can still write to its own
- *     store; nobody reads from it.
- *   - Drive the viewport with a single rAF loop governed by a follow
- *     deadline (`followUntilRef`). Any signal that can invalidate
- *     bottom alignment (resize, mutation, AUI event, programmatic
- *     scroll) extends the deadline; the loop pins to the bottom and
- *     reports `isAtBottom=true` until it expires, then settles on
- *     pure DOM observation.
- *   - Detect user intent (wheel up, touch swipe up, scroll direction)
- *     to detach. While detached, resize/mutation signals don't extend
- *     the deadline, so the button appears and stays. Re-attach when
- *     the user scrolls *down* within 24px of the bottom.
+ *   - Own `isAtBottom` as local state (via `useIsThreadAtBottom`); nobody
+ *     reads upstream's store.
+ *   - Drive the viewport via a single rAF loop with a follow deadline
+ *     (`followUntilRef`). Any signal that can invalidate bottom alignment
+ *     extends the deadline; the loop pins and reports `isAtBottom=true`
+ *     until it expires, then settles on DOM observation.
+ *   - Detect user intent (wheel up, swipe up, scroll direction) to detach.
+ *     While detached, resize/mutation don't extend the deadline. Re-attach
+ *     when the user scrolls down within 24px of the bottom.
  */
 
-// 2px, not 1: subpixel rounding on HiDPI displays can leave a fractional
-// gap at the "true" bottom that a 1px threshold reports as not-at-bottom.
+// 2px, not 1: HiDPI subpixel rounding can leave a fractional gap that a
+// 1px threshold reports as not-at-bottom.
 const AT_BOTTOM_THRESHOLD_PX = 2;
 const RE_ATTACH_THRESHOLD_PX = 24;
 const TOUCH_MOVE_THRESHOLD_PX = 4;
-// Cumulative upward movement (summed across scroll events) that
-// counts as a deliberate detach. Summed rather than per-event so that
-// slow 1-px-per-event sources — middle-click autoscroll, scrollbar
-// drags, some trackpads — accumulate instead of slipping under a
-// per-event threshold forever.
+// Cumulative upward movement counting as a deliberate detach. Summed (not
+// per-event) so slow 1px-per-event sources (middle-click autoscroll,
+// scrollbar drags, some trackpads) accumulate instead of slipping under.
 const UPWARD_DETACH_THRESHOLD_PX = 2;
-// Window during which the viewport pins to the bottom through
-// layout/content races. Extends on every resize/mutation, so streaming
-// keeps the viewport pinned as long as content keeps arriving; settles
-// this long after the last change.
+// Window the viewport stays pinned through layout/content races. Extends
+// on every resize/mutation, so streaming keeps it pinned; settles this
+// long after the last change.
 const FOLLOW_SETTLE_MS = 600;
-// Maximum stabilizer compensation. The stabilizer is meant to absorb
-// sub-frame transients (~5-15px shiki re-renders, ~8px action-bar
-// reservation drift). Anything larger is almost certainly an intentional
-// content removal — message delete, regenerate's old-content clear,
-// reasoning-panel collapse — and should *not* be silently padded over,
-// which would leave persistent empty space below the last message.
-// Above this threshold we release the stabilizer immediately and let
-// the autoscroll re-pin to the new content height, which is the natural
-// behavior the user expects for those actions.
+// Max stabilizer compensation. Absorbs sub-frame transients (~5-15px shiki
+// re-renders, ~8px action-bar drift). Larger shrinks are intentional
+// content removals (delete, regenerate clear, reasoning collapse); padding
+// those over leaves empty space, so above this we release and let
+// autoscroll re-pin to the new height.
 const STABILIZER_MAX_PX = 64;
 
 export type ScrollToBottom = (behavior?: ScrollBehavior) => void;
@@ -78,11 +63,10 @@ type AutoScrollContextValue = {
   getIsAtBottom: () => boolean;
   subscribe: (listener: () => void) => () => void;
   /**
-   * Mark the user as detached from the bottom, as if they had scrolled
-   * up. Called when the composer grows and the bottom spacer grows with
-   * it: the chat is then above the new bottom, and observer-driven pins
-   * must not shove it up. Scrolling back to the bottom re-attaches;
-   * explicit pins (run start, scroll-to-bottom button) still work.
+   * Mark the user as detached, as if they scrolled up. Called when the
+   * composer (and bottom spacer) grow: the chat is then above the new
+   * bottom and observer-driven pins must not shove it up. Scrolling back
+   * re-attaches; explicit pins (run start, button) still work.
    */
   detachFromBottom: () => void;
 };
@@ -191,24 +175,18 @@ export function useIntentAwareAutoScroll(): {
       const atBottomStrict = (): boolean =>
         distanceFromBottom() <= AT_BOTTOM_THRESHOLD_PX;
 
-      // True only when there is room to scroll upward. Guards the
-      // wheel/touch detach paths: a wheel-up or swipe-down gesture on
-      // a viewport with nothing above (short thread, or already at
-      // the top) can't express intent to leave the bottom and must
-      // not flip userDetachedRef — otherwise later streaming updates
-      // skip extendFollow and auto-follow stays dead for the session.
+      // Room to scroll upward. Guards wheel/touch detach: a gesture on a
+      // viewport with nothing above can't express intent to leave the
+      // bottom and must not flip userDetachedRef, else later streaming
+      // skips extendFollow and auto-follow stays dead for the session.
       const canScrollUp = (): boolean => el.scrollTop > 0;
 
-      // True when a nested scrollable ancestor of the event target
-      // (e.g. the reasoning panel's own overflow-y-auto region, or
-      // any tool output with internal scroll) has room above and
-      // will therefore consume the upward delta before it reaches
-      // the viewport. Walking stops at the viewport element itself,
-      // so only intermediate inner scrollers count.
-      //
-      // Wheel and touchmove events bubble, so without this check a
-      // user reading back through a long reasoning pane mid-stream
-      // would falsely detach the outer viewport.
+      // True when a nested scrollable ancestor of the target (reasoning
+      // panel, tool output) has room above and will consume the upward
+      // delta before it reaches the viewport. Walk stops at the viewport,
+      // so only intermediate inner scrollers count. Without this, wheel/
+      // touchmove bubbling would falsely detach while reading a long
+      // reasoning pane mid-stream.
       const innerScrollWillConsumeUpward = (
         target: EventTarget | null,
       ): boolean => {
@@ -230,10 +208,8 @@ export function useIntentAwareAutoScroll(): {
         return false;
       };
 
-      // Stabilizer state — see `stabilize` below for the full
-      // explanation. Lives in this closure so it resets naturally
-      // whenever the viewport remounts (Compare-pane swap, thread
-      // switch with remount, etc.).
+      // Stabilizer state (see `stabilize`). In this closure so it resets
+      // when the viewport remounts (Compare-pane swap, thread switch).
       let stabilizerPx = 0;
       let maxContentHeight = 0;
 
@@ -255,11 +231,10 @@ export function useIntentAwareAutoScroll(): {
       const detach = (): void => {
         userDetachedRef.current = true;
         followUntilRef.current = 0;
-        // The stabilizer is only meaningful while we're actively
-        // pinning to the bottom. Once the user scrolls up, drop any
-        // residual padding so the bottom stays flush whenever they
-        // come back. Safe here because the user is mid-content —
-        // shrinking scrollHeight cannot cap their scrollTop.
+        // The stabilizer only matters while pinning. Once the user
+        // scrolls up, drop residual padding so the bottom stays flush on
+        // return. Safe mid-content: shrinking scrollHeight can't cap their
+        // scrollTop.
         releaseStabilizer();
         maxContentHeight = el.scrollHeight;
       };
@@ -270,12 +245,10 @@ export function useIntentAwareAutoScroll(): {
         }
       };
 
-      // Single rAF loop. While within the follow window and not
-      // detached, pin the viewport to the bottom and report
-      // isAtBottom=true every frame. Otherwise settle on whatever the
-      // DOM says. Scheduling is edge-triggered: scroll/resize/mutation
-      // events call requestTick(), and the loop self-perpetuates only
-      // as long as pinning is still active.
+      // Single rAF loop. Within the follow window and not detached, pin to
+      // bottom and report isAtBottom=true each frame; otherwise settle on
+      // the DOM. Edge-triggered: scroll/resize/mutation call requestTick(),
+      // and the loop self-perpetuates only while pinning is active.
       const tick = (): void => {
         rafId = null;
         const following =
@@ -304,8 +277,8 @@ export function useIntentAwareAutoScroll(): {
         requestTick();
       };
 
-      // Programmatic detach (see detachFromBottom). Same effect as the
-      // user scrolling up; the tick refresh updates isAtBottom.
+      // Programmatic detach (see detachFromBottom). Same as scrolling up;
+      // the tick refresh updates isAtBottom.
       detachImplRef.current = () => {
         detach();
         requestTick();
@@ -363,17 +336,12 @@ export function useIntentAwareAutoScroll(): {
             extendFollow();
           }
         } else if (delta < 0 && !userDetachedRef.current) {
-          // Upward: sum across events. Middle-click autoscroll and
-          // some trackpads deliver 1px-per-event scrolls that each
-          // slip under a per-event threshold; summing catches them.
-          //
-          // Count distance-from-bottom growth, not raw scrollTop
-          // delta. When content above collapses (reasoning panels
-          // auto-closing after streaming, tool outputs auto-hiding),
-          // browsers scroll-anchor to keep visible content stable:
-          // scrollTop decreases but scrollHeight decreases by the
-          // same amount, so distance is unchanged. Those layout-
-          // induced deltas must not flip user intent.
+          // Upward: sum across events to catch 1px-per-event sources
+          // (middle-click autoscroll, some trackpads) that slip under a
+          // per-event threshold. Count distance-from-bottom growth, not
+          // raw scrollTop delta: when content above collapses, browsers
+          // scroll-anchor so scrollTop and scrollHeight drop together and
+          // distance is unchanged. Those layout deltas must not flip intent.
           const distanceDelta = distanceNow - lastDistanceFromBottom;
           if (distanceDelta > 0) {
             upwardAccumulator += distanceDelta;
@@ -393,37 +361,25 @@ export function useIntentAwareAutoScroll(): {
 
       // Scroll stabilizer.
       //
-      // Problem: when a trailing code block finalizes at stream end
-      // (Streamdown flips `isAnimating` → false, shiki re-renders the
-      // <pre> with highlight spans), the block's rendered height
-      // briefly dips and then recovers a frame later. That dip shrinks
-      // `scrollHeight`, which the browser handles by *synchronously*
-      // capping `scrollTop` to the new (smaller) `scrollHeight −
-      // clientHeight`. The cap is visible as a one-frame upward jump;
-      // the recovery a frame or two later is the "snap back" the user
-      // perceives as a flicker. No amount of programmatic re-scrolling
-      // can prevent this — once `scrollHeight` drops, the cap has
-      // already happened and `scrollTop` cannot be pushed past the new
-      // max.
+      // Problem: when a trailing code block finalizes (shiki re-renders the
+      // <pre> with highlight spans), its height briefly dips then recovers.
+      // The dip shrinks `scrollHeight`, so the browser synchronously caps
+      // `scrollTop` to the new max — a one-frame upward jump, then a
+      // "snap back". Re-scrolling can't help: once scrollHeight drops the
+      // cap has happened and scrollTop can't exceed the new max.
       //
-      // Fix: keep `scrollHeight` monotonic across the follow window.
-      // We track the maximum *content* height (scrollHeight minus our
-      // own padding contribution) seen during follow, and compensate
-      // for any shortfall by writing the deficit into a CSS custom
-      // property `--aui-scroll-stabilizer`, which the viewport's
-      // `padding-bottom` reads. A 5px content shrink instantly grows
-      // the padding by 5px, so the browser sees no scrollHeight change
-      // and never caps scrollTop. As content naturally grows past its
-      // prior high-water mark (e.g. the next message streams in), the
-      // padding shrinks back toward zero.
+      // Fix: keep `scrollHeight` monotonic across the follow window. Track
+      // max content height (scrollHeight minus our padding) and write any
+      // shortfall into CSS var `--aui-scroll-stabilizer`, read by the
+      // viewport's padding-bottom. A 5px shrink grows padding 5px so the
+      // browser sees no scrollHeight change. Padding shrinks back to zero
+      // as content grows past its prior high-water mark.
       //
-      // Self-contained: lives entirely on the viewport element via a
-      // CSS variable. Doesn't touch the composer, the action bar, the
-      // message footer, the spacer, or any other UI.
+      // Self-contained: lives on the viewport element via a CSS variable,
+      // touching no other UI.
       //
-      // Returns the post-adjustment scrollHeight so a single layout
-      // read per observer callback can feed both stabilization and
-      // pinning, avoiding a redundant flush.
+      // Returns the post-adjustment scrollHeight so one layout read per
+      // observer callback feeds both stabilization and pinning.
       const stabilize = (): number => {
         const sh = el.scrollHeight;
         const currentContent = sh - stabilizerPx;
@@ -431,9 +387,8 @@ export function useIntentAwareAutoScroll(): {
           !userDetachedRef.current &&
           performance.now() < followUntilRef.current;
         if (!followActive) {
-          // Outside the follow window we stop adjusting, but we keep
-          // `maxContentHeight` aligned with reality so the next follow
-          // session starts from the current content size, not stale.
+          // Outside the follow window: stop adjusting but keep
+          // maxContentHeight current so the next session isn't stale.
           maxContentHeight = currentContent;
           return sh;
         }
@@ -441,14 +396,11 @@ export function useIntentAwareAutoScroll(): {
           maxContentHeight = currentContent;
         }
         const shrink = maxContentHeight - currentContent;
-        // Large shrinks (over STABILIZER_MAX_PX) are intentional content
-        // removals — message delete, regenerate clearing the old
-        // assistant turn, reasoning-panel collapse. Compensating for
-        // those would leave persistent empty space at the bottom of the
-        // viewport, which the user reads as "weird empty gap." Release
-        // the stabilizer instead and rebase the high-water mark; the
-        // pinIfFollowing call right after will smoothly re-anchor to
-        // the new (smaller) bottom.
+        // Large shrinks (over STABILIZER_MAX_PX) are intentional removals
+        // (delete, regenerate clear, reasoning collapse). Compensating
+        // would leave an empty gap, so release the stabilizer and rebase
+        // the high-water mark; the pinIfFollowing call below re-anchors to
+        // the new bottom.
         if (shrink > STABILIZER_MAX_PX) {
           maxContentHeight = currentContent;
           if (stabilizerPx !== 0) {
@@ -468,10 +420,9 @@ export function useIntentAwareAutoScroll(): {
         return currentContent + stabilizerPx;
       };
 
-      // Synchronous pin-to-bottom. Observer callbacks run in the event-
-      // loop's "update the rendering" step (after layout, before paint),
-      // so the scrollTo here is composited in the same frame as the
-      // mutation that triggered the observer.
+      // Synchronous pin-to-bottom. Observer callbacks run after layout,
+      // before paint, so this scrollTo composites in the same frame as the
+      // triggering mutation.
       const pinIfFollowing = (scrollHeight: number): void => {
         if (userDetachedRef.current) {
           return;
@@ -485,10 +436,8 @@ export function useIntentAwareAutoScroll(): {
         el.scrollTo({ top: scrollHeight, behavior: "instant" });
       };
 
-      // All three layout-change signals fan in here so there's a
-      // single place to understand "what runs when the viewport's
-      // content shape changes". Order matters: extend first so the
-      // stabilizer sees the follow window as active; stabilize before
+      // All three layout-change signals fan in here. Order matters: extend
+      // first so the stabilizer sees follow as active; stabilize before
       // pinning so we scroll to the post-adjustment scrollHeight.
       const onLayoutChange = (): void => {
         extendFollow();
@@ -501,18 +450,17 @@ export function useIntentAwareAutoScroll(): {
       const mutationObserver = new MutationObserver(onLayoutChange);
       const onViewportResize = onLayoutChange;
 
-      // Fresh attach (a new viewport element) always starts pinned.
-      // Rebinds to the SAME element must not pin or reset detach state:
-      // the Viewport composes refs with an identity that changes on
-      // re-render, so React re-runs the ref (null, then same element)
-      // on unrelated renders such as composer resizes. Pinning here
-      // would yank the chat to the bottom on every such render. The
-      // observers below are re-installed either way.
+      // Fresh attach always starts pinned. Rebinds to the SAME element must
+      // not pin or reset detach state: the Viewport's composed ref identity
+      // changes on re-render, so React re-runs the ref (null, then same
+      // element) on unrelated renders (composer resizes); pinning would
+      // yank the chat to bottom each time. Observers are re-installed either
+      // way.
       if (!isRebind) {
         userDetachedRef.current = false;
 
-        // Pin to bottom when the ref first attaches. Covers the case
-        // where `thread.initialize` fires before the ref is bound.
+        // Pin on first attach, covering thread.initialize firing before the
+        // ref is bound.
         extendFollow();
         if (el.scrollHeight > el.clientHeight) {
           el.scrollTo({ top: el.scrollHeight, behavior: "instant" });
@@ -521,23 +469,19 @@ export function useIntentAwareAutoScroll(): {
       }
       requestTick();
 
-      // Observe the border box, not the content box. The stabilizer
-      // writes `padding-bottom`, which shrinks the content box; if we
-      // observed that, every stabilizer adjustment would echo back as
-      // a resize and re-enter onLayoutChange. Border-box stays put
-      // through padding changes but still tracks parent-driven
-      // resizes (window, sidebar toggle) — which is all we need.
+      // Observe the border box, not content box. The stabilizer writes
+      // padding-bottom (shrinks the content box); observing that would echo
+      // every adjustment back as a resize into onLayoutChange. Border-box is
+      // stable through padding changes but still tracks parent resizes.
       resizeObserver.observe(el, { box: "border-box" });
       mutationObserver.observe(el, {
         childList: true,
         subtree: true,
         characterData: true,
-        // Layout-affecting attributes only. Excludes `style`, which
-        // elements may mutate in response to viewport state (feedback
-        // loop). `class` catches Tailwind show/hide; `hidden` /
-        // `aria-hidden` / `aria-expanded` / `data-state` catch Radix
-        // and native collapsible toggles that change scrollHeight
-        // without triggering the viewport ResizeObserver.
+        // Layout-affecting attributes only. Excludes `style` (elements may
+        // mutate it in response to viewport state → feedback loop). `class`
+        // catches Tailwind show/hide; the rest catch Radix/native
+        // collapsibles that change scrollHeight without a viewport resize.
         attributes: true,
         attributeFilter: [
           "class",
@@ -551,10 +495,9 @@ export function useIntentAwareAutoScroll(): {
       el.addEventListener("touchstart", onTouchStart, { passive: true });
       el.addEventListener("touchmove", onTouchMove, { passive: true });
       el.addEventListener("scroll", onScroll, { passive: true });
-      // ResizeObserver above covers browser-window resizes (they resize
-      // the viewport element). visualViewport.resize is the only signal
-      // for iOS software-keyboard changes, where the visual viewport
-      // shrinks without the viewport element's clientHeight changing.
+      // ResizeObserver covers window resizes. visualViewport.resize is the
+      // only signal for iOS software-keyboard changes, where the visual
+      // viewport shrinks without the element's clientHeight changing.
       window.visualViewport?.addEventListener("resize", onViewportResize);
 
       return () => {
@@ -580,10 +523,9 @@ export function useIntentAwareAutoScroll(): {
     [setIsAtBottom],
   );
 
-  // Thread lifecycle moments that always pin to the bottom, regardless
-  // of prior detach state. "auto" respects CSS smooth scrolling for
-  // runStart so new turns glide in; "instant" snaps for load/switch
-  // where any animation would just be wasted motion.
+  // Thread lifecycle moments that always pin, regardless of detach state.
+  // "auto" respects CSS smooth scroll for runStart (new turns glide in);
+  // "instant" snaps for load/switch where animation is wasted.
   const pinToBottom = useCallback((behavior: ScrollBehavior) => {
     userDetachedRef.current = false;
     scrollImplRef.current(behavior);
diff --git a/studio/frontend/src/components/shutdown-dialog.tsx b/studio/frontend/src/components/shutdown-dialog.tsx
index dfeafb33eb..e3e6ca9920 100644
--- a/studio/frontend/src/components/shutdown-dialog.tsx
+++ b/studio/frontend/src/components/shutdown-dialog.tsx
@@ -18,10 +18,9 @@ import {
 interface ShutdownDialogProps {
   open: boolean;
   onOpenChange: (open: boolean) => void;
-  /** Called after the shutdown API returns success, right before we replace
-   *  document.body with the "Server stopped" page. Callers use this to remove
-   *  their beforeunload listener — otherwise the browser would prompt
-   *  "Leave site?" when the user tries to close the final tab. */
+  /** Called after shutdown succeeds, before we replace document.body. Lets
+   *  callers remove their beforeunload listener so the browser doesn't prompt
+   *  "Leave site?" when closing the final tab. */
   onAfterShutdown?: () => void;
 }
 
@@ -44,7 +43,7 @@ export function ShutdownDialog({
         return;
       }
     } catch {
-      // Network error — shutdown request never reached the server
+      // Network error: request never reached the server
       toastError("Could not reach server");
       setStopping(false);
       return;
diff --git a/studio/frontend/src/components/tauri/window-titlebar.tsx b/studio/frontend/src/components/tauri/window-titlebar.tsx
index b3180e6663..2ef3258902 100644
--- a/studio/frontend/src/components/tauri/window-titlebar.tsx
+++ b/studio/frontend/src/components/tauri/window-titlebar.tsx
@@ -137,7 +137,7 @@ export function WindowTitlebar({
       const appWindow = await getAppWindow();
       setMaximized(await appWindow.isMaximized());
     } catch {
-      // If a window permission is not ready yet, keep the previous visual state.
+      // Window permission not ready yet: keep previous visual state.
     }
   }, [enabled]);
 
diff --git a/studio/frontend/src/config/env.ts b/studio/frontend/src/config/env.ts
index 3839706d25..92dad2ae27 100644
--- a/studio/frontend/src/config/env.ts
+++ b/studio/frontend/src/config/env.ts
@@ -11,7 +11,7 @@ export const env = {
   BASE_URL: import.meta.env.BASE_URL,
 } as const;
 
-// ── Platform / device type ──────────────────────────────────
+// Platform / device type
 
 export type DeviceType = "mac" | "windows" | "linux" | string;
 
@@ -22,7 +22,7 @@ interface PlatformState {
   isChatOnly: () => boolean;
 }
 
-// Client-side platform detection as fallback when backend isn't ready yet.
+// Client-side fallback when backend isn't ready yet.
 function detectLocalPlatform(): DeviceType {
   if (typeof navigator === "undefined") return "linux";
   const platform = navigator.platform.toLowerCase();
@@ -55,9 +55,9 @@ export async function fetchDeviceType(): Promise<DeviceType> {
       return deviceType;
     }
   } catch {
-    // Backend not ready — use client-side detection so chat-only guard
-    // still works on initial load (important for macOS). Keep fetched=false
-    // so a later call retries against the backend.
+    // Backend not ready: use client-side detection so chat-only guard works
+    // on initial load (important for macOS). Keep fetched=false so a later
+    // call retries against the backend.
     const deviceType = detectLocalPlatform();
     const chatOnly = deviceType === "mac";
     usePlatformStore.setState({ deviceType, chatOnly, fetched: false });
diff --git a/studio/frontend/src/features/auth/api.ts b/studio/frontend/src/features/auth/api.ts
index 95296d3ab9..6356f89378 100644
--- a/studio/frontend/src/features/auth/api.ts
+++ b/studio/frontend/src/features/auth/api.ts
@@ -168,9 +168,8 @@ export async function authFetch(
     });
   } catch (err) {
     if (err instanceof TypeError) {
-      // fetch TypeError = offline | backend down | CORS/DNS. In Tauri
-      // it's always backend-down; in the web build distinguish offline
-      // so the user gets the right recovery path.
+      // fetch TypeError = offline | backend down | CORS/DNS. Tauri is always
+      // backend-down; the web build distinguishes offline for the right message.
       if (!isTauri && typeof navigator !== "undefined" && navigator.onLine === false) {
         throw new Error(
           "You appear to be offline. Check your network connection and try again.",
@@ -227,10 +226,9 @@ async function postLogout(accessToken: string | null): Promise<Response | null>
 }
 
 export async function logout(): Promise<void> {
-  // Server-side revoke. If the access token is expired the 401 fires
-  // BEFORE revoke runs; rotate via the refresh token and retry so the
-  // refresh family is actually revoked. Generation bump in finally
-  // invalidates any in-flight refresh from before this call.
+  // Server-side revoke. If the access token is expired, the 401 fires before
+  // revoke runs; rotate via the refresh token and retry so the refresh family
+  // is revoked. The finally generation bump invalidates in-flight refreshes.
   try {
     let response = await postLogout(getAuthToken());
     if (response && response.status === 401 && getRefreshToken()) {
diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx
index c1606c7020..f2017ff0b6 100644
--- a/studio/frontend/src/features/auth/components/auth-form.tsx
+++ b/studio/frontend/src/features/auth/components/auth-form.tsx
@@ -12,8 +12,8 @@ import type { ReactElement } from "react";
 import type { SyntheticEvent } from "react";
 import { refreshSession } from "../api";
 
-// Bootstrap credentials injected into index.html by the backend
-// (only present while default admin must_change_password is true)
+// Bootstrap credentials injected into index.html by the backend (only present
+// while default admin must_change_password is true)
 declare global {
   interface Window {
     __UNSLOTH_BOOTSTRAP__?: { username: string; password: string };
@@ -93,9 +93,9 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
     let canceled = false;
 
     async function initializeAuthForm(): Promise<void> {
-      // Always check the server first — localStorage flags can be stale
-      // (e.g. tokens from a previous install attempt).  The server's
-      // /api/auth/status is the source of truth for requires_password_change.
+      // Always check the server first; localStorage flags can be stale (e.g.
+      // tokens from a previous install). /api/auth/status is the source of
+      // truth for requires_password_change.
       try {
         const response = await fetch(apiUrl("/api/auth/status"));
         if (!response.ok) throw new Error("Failed to load auth status.");
@@ -109,7 +109,7 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
             setMustChangePassword(result.requires_password_change);
           }
 
-          // Redirect between login ↔ change-password based on server state
+          // Redirect between login / change-password per server state
           if (mode === "login" && result.requires_password_change) {
             navigate({ to: "/change-password" });
             return;
@@ -119,8 +119,8 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
             return;
           }
 
-          // On login page, if user already has a valid session and no
-          // password change is required, skip straight to the app.
+          // On login, skip to the app if a valid session exists and no
+          // password change is required.
           if (isLoginMode && !result.requires_password_change) {
             if (hasRefreshToken()) {
               const refreshed = await refreshSession();
@@ -293,12 +293,12 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
       storeAuthTokens(token.access_token, token.refresh_token);
       navigate({ to: getPostAuthRoute() });
     } catch (err: unknown) {
-      // The backend already returns the correct, PATH-based command
-      // ("unsloth studio reset-password"), which the installer puts on PATH on
-      // every platform. Do NOT rewrite it to a relative Windows path like
-      // ".\unsloth_studio\Scripts\unsloth.exe ..." -- that only resolves when the
-      // terminal happens to be inside the Studio home dir, so it fails with
-      // CommandNotFoundException everywhere else. Show the backend message as-is.
+      // The backend returns the correct PATH-based command ("unsloth studio
+      // reset-password"), which the installer puts on PATH on every platform.
+      // Do NOT rewrite it to a relative Windows path like
+      // ".\unsloth_studio\Scripts\unsloth.exe ..." -- that only resolves inside
+      // the Studio home dir and fails with CommandNotFoundException elsewhere.
+      // Show the backend message as-is.
       const msg = err instanceof Error ? err.message : "Auth failed.";
       setError(msg);
     } finally {
diff --git a/studio/frontend/src/features/auth/session.ts b/studio/frontend/src/features/auth/session.ts
index 1e3234590a..e398ee0608 100644
--- a/studio/frontend/src/features/auth/session.ts
+++ b/studio/frontend/src/features/auth/session.ts
@@ -39,9 +39,9 @@ export function storeAuthTokens(
   accessToken: string,
   refreshToken: string,
 ): void {
-  // Callers set must_change_password via setMustChangePassword(). Routing it
-  // through here would let CodeQL trace the boolean to localStorage and flag
-  // the (deliberate) JWT writes as sensitive-info storage.
+  // must_change_password is set via setMustChangePassword(), not here: routing
+  // it through would let CodeQL trace the boolean into localStorage and flag the
+  // deliberate JWT writes as sensitive-info storage.
   if (!canUseStorage()) return;
   localStorage.setItem(AUTH_TOKEN_KEY, accessToken);
   localStorage.setItem(AUTH_REFRESH_TOKEN_KEY, refreshToken);
@@ -54,10 +54,9 @@ export function clearAuthTokens(): void {
   localStorage.removeItem(AUTH_MUST_CHANGE_PASSWORD_KEY);
 }
 
-// Encode the flag as key presence (literal "1" or absence) so localStorage
-// receives a constant, not a derived boolean. Breaks the CodeQL data flow
-// from TokenResponse.must_change_password into localStorage.setItem; the
-// stored value is a route hint (/change-password vs /chat), not a secret.
+// Flag stored as key presence (constant "1" or absence), not a derived boolean,
+// so CodeQL doesn't flow must_change_password into localStorage.setItem. The
+// value is a route hint (/change-password vs /chat), not a secret.
 export function mustChangePassword(): boolean {
   if (!canUseStorage()) return false;
   return localStorage.getItem(AUTH_MUST_CHANGE_PASSWORD_KEY) !== null;
diff --git a/studio/frontend/src/features/auth/tauri-auto-auth.ts b/studio/frontend/src/features/auth/tauri-auto-auth.ts
index 44884796d3..760af26395 100644
--- a/studio/frontend/src/features/auth/tauri-auto-auth.ts
+++ b/studio/frontend/src/features/auth/tauri-auto-auth.ts
@@ -20,8 +20,8 @@ type TauriAutoAuthOptions = {
   force?: boolean;
 };
 
-// Concurrency guard: multiple route guards can call tauriAutoAuth simultaneously.
-// Without this, the first-launch password-change could race with itself.
+// Concurrency guard: multiple route guards can call tauriAutoAuth at once;
+// without this the first-launch password-change could race with itself.
 let pending: { promise: Promise<boolean>; force: boolean } | null = null;
 let lastTauriAuthFailure: string | null = null;
 
@@ -61,7 +61,7 @@ async function doTauriAutoAuth(options: TauriAutoAuthOptions): Promise<boolean>
     return true;
   }
 
-  // Try refreshing existing session
+  // Try refreshing an existing session.
   if (!options.force && hasRefreshToken()) {
     const refreshed = await refreshSession();
     if (refreshed && hasAuthToken() && !mustChangePassword()) {
diff --git a/studio/frontend/src/features/chat/api-provider-logo.tsx b/studio/frontend/src/features/chat/api-provider-logo.tsx
index 09794e3acf..f9d574b8fa 100644
--- a/studio/frontend/src/features/chat/api-provider-logo.tsx
+++ b/studio/frontend/src/features/chat/api-provider-logo.tsx
@@ -7,8 +7,8 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import { isCustomProviderType } from "./external-providers";
 
 /**
- * Registry logos live at `public/provider-logos/{provider_type}.{ext}` where `provider_type`
- * matches `PROVIDER_REGISTRY` keys exactly (lowercase). Extension varies by asset (svg preferred).
+ * Registry logos at `public/provider-logos/{provider_type}.{ext}`; key matches
+ * `PROVIDER_REGISTRY` (lowercase). Extension varies per asset (svg preferred).
  */
 const PROVIDER_LOGO_EXT: Record<string, "svg" | "png" | "jpg"> = {
   openai: "svg",
@@ -41,9 +41,8 @@ interface ApiProviderLogoProps {
 }
 
 /**
- * Renders the logo for a registry provider type when `provider_type.{ext}` exists under
- * `public/provider-logos/`.
- * OpenAI's asset is black-on-transparent; it is inverted in dark mode for contrast.
+ * Renders a registry provider's logo when its asset exists under
+ * `public/provider-logos/`. OpenAI's is inverted in dark mode for contrast.
  */
 export function ApiProviderLogo({ providerType, className, title }: ApiProviderLogoProps) {
   const src = apiProviderLogoSrc(providerType);
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 44b8507906..31f73c37bd 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -80,7 +80,7 @@ interface ServerUsage {
   completion_tokens: number;
   total_tokens: number;
   // External prompt-cache fields (see _build_usage_chunk in
-  // external_provider.py). cache_creation is Anthropic-only.
+  // external_provider.py); cache_creation is Anthropic-only.
   prompt_tokens_details?: {
     cached_tokens?: number;
   };
@@ -107,9 +107,9 @@ type RunMessage = RunMessages[number];
 /** Tracks which user messages were sent with an audio file (messageId → filename). */
 export const sentAudioNames = new Map<string, string>();
 
-// Synthetic provider-side tool names; backend stamps args._server_tool
-// so user functions with the same name aren't dropped. Mirror of
-// _SERVER_SIDE_BUILTIN_TOOL_NAMES on the backend.
+// Synthetic provider-side tool names; backend stamps args._server_tool so
+// user functions with the same name aren't dropped. Mirror of backend
+// _SERVER_SIDE_BUILTIN_TOOL_NAMES.
 const SERVER_SIDE_BUILTIN_TOOL_NAMES = new Set<string>([
   "web_search",
   "web_fetch",
@@ -118,10 +118,10 @@ const SERVER_SIDE_BUILTIN_TOOL_NAMES = new Set<string>([
 ]);
 
 /**
- * Whether a persisted tool-call part is provider-side synthetic and
- * should be stripped from outbound history. Match on the
- * args._server_tool marker or a Gemini native_part payload — no shape
- * heuristic, because user functions can legitimately share a name.
+ * Whether a persisted tool-call part is provider-side synthetic and should
+ * be stripped from outbound history. Matches on the args._server_tool marker
+ * or a Gemini native_part payload (no shape heuristic, since user functions
+ * can legitimately share a name).
  */
 function isServerSideBuiltinToolPart(
   toolNameLower: string,
@@ -176,23 +176,19 @@ export function useThreadAutosaveHandle(): ThreadAutosaveHandle {
 }
 
 /**
- * Match error messages that indicate the request filled or would fill
- * the KV cache, so the UI can show a dedicated toast pointing at the
- * ``Context Length`` setting.
+ * Match error messages indicating the request filled (or would fill) the KV
+ * cache, so the UI can toast a pointer at the ``Context Length`` setting.
  *
  * Two wordings reach the client and both must hit:
- *
- *   1. The raw llama-server text when ``--no-context-shift`` trips --
+ *   1. Raw llama-server text when ``--no-context-shift`` trips:
  *      "the request exceeds the available context size (N tokens)".
- *   2. The rewritten friendly text emitted by
- *      ``backend/routes/inference.py::_friendly_error`` -- "Message too
- *      long: X tokens exceeds the Y-token context window. Try
- *      increasing the Context Length ..." This is the one most users
- *      see on the streaming GGUF path.
+ *   2. The friendly rewrite from
+ *      ``backend/routes/inference.py::_friendly_error``: "Message too long:
+ *      ... context window. Try increasing the Context Length ..." (the one
+ *      most users see on the streaming GGUF path).
  *
- * We match on substrings rather than full regexes because both layers
- * have drifted across versions (llama.cpp master has tweaked the
- * phrasing; ``_friendly_error`` has gone through several copy edits).
+ * Match substrings, not full regexes: both layers have drifted across
+ * versions (llama.cpp phrasing and _friendly_error copy edits).
  */
 export function isContextLimitError(message: string): boolean {
   if (!message) return false;
@@ -224,10 +220,10 @@ async function updateStoredChatThreadEventually(
 }
 
 /**
- * Return ``raw`` when it is a safe-to-navigate http(s) URL, or "" otherwise.
+ * Return ``raw`` when it is a safe-to-navigate http(s) URL, else "".
  * Rejects non-string input, CR/LF (header injection), and non-http(s)
- * schemes (``javascript:`` / ``data:`` / ``vbscript:``) so provider /
- * tool-controlled strings cannot land in an <a href>.
+ * schemes (``javascript:`` / ``data:`` / ``vbscript:``) so provider/tool-
+ * controlled strings cannot land in an <a href>.
  */
 function isSafeNavigableSourceUrl(raw: unknown): string {
   if (typeof raw !== "string") return "";
@@ -264,10 +260,9 @@ function documentCitationToSource(
     "";
   const docIndex =
     typeof cit.document_index === "number" ? cit.document_index : undefined;
-  // Only treat ``source`` as a navigable URL when it is real http(s);
+  // Only treat ``source`` as navigable when it is real http(s);
   // search_result_location can carry a free-form id (e.g. ``kb-doc-42``)
-  // or a hostile ``javascript:`` / ``data:`` / ``vbscript:`` string.
-  // Fall back to a stable doc anchor otherwise.
+  // or a hostile scheme. Fall back to a stable doc anchor otherwise.
   const url =
     isSafeNavigableSourceUrl(source) || `#anthropic-doc-${docIndex ?? fallbackIdx}`;
   const title = docTitle || source || `Document ${fallbackIdx + 1}`;
@@ -277,9 +272,8 @@ function documentCitationToSource(
   const description =
     cited.length > 240 ? `${cited.slice(0, 240)}...` : cited;
   // Anthropic numbers inline [N] per citation, not per source URL.
-  // Fold citation type + position-bearing fields into the id so two
-  // distinct citations on the same source (or two search_result_locations
-  // with different search_result_index) keep separate Sources entries.
+  // Fold citation type + position-bearing fields into the id so distinct
+  // citations on the same source keep separate Sources entries.
   const citationType =
     typeof cit.type === "string" ? String(cit.type) : "";
   const positionParts = [
@@ -333,8 +327,8 @@ function parseSourcesFromResult(raw: string): {
     const snippetMatch = block.match(/Snippet:\s*(.+)/);
     if (titleMatch && urlMatch) {
       // Drop blocks whose ``URL:`` is not safe http(s); provider/tool
-      // output is attacker-controllable so a hostile ``javascript:`` /
-      // ``data:`` line must not reach the Sources panel <a href>.
+      // output is attacker-controllable, so a hostile scheme must not
+      // reach the Sources panel <a href>.
       const url = isSafeNavigableSourceUrl(urlMatch[1]);
       if (!url) continue;
       const snippet = snippetMatch?.[1]?.trim();
@@ -363,21 +357,18 @@ function estimateTokenCount(text: string): number | undefined {
  * Normalize a streamed `delta.content` to a plain text string.
  *
  * OpenAI Chat Completions originally typed `delta.content` as a string, but
- * a number of providers now emit it as an array of structured content parts.
- * Concatenating that with `cumulativeText += delta` would stringify each
- * part as `[object Object]` — this function is the guard against that.
+ * some providers now emit an array of structured content parts; concatenating
+ * those directly would stringify each as `[object Object]`. This guards that.
  *
  * Handled part shapes:
  *   { type: "text" | "output_text", text | content: "..." }   → text body
  *   { type: "thinking" | "reasoning", thinking | text: "..." } → wrapped as
- *       inline `<think>...</think>` so the downstream parser
- *       (`parseAssistantContent`) lifts it into a reasoning part the same way
- *       it does for providers that emit thinking inline. Without this wrap,
- *       Mistral magistral and similar reasoning-part providers would lose
- *       their thinking panel.
+ *       inline `<think>...</think>` so `parseAssistantContent` lifts it into
+ *       a reasoning part (else Mistral magistral and similar reasoning-part
+ *       providers lose their thinking panel).
  *
- * Unknown part types are skipped — better to drop a stray field than to
- * stringify an object and pollute the rendered chat with `[object Object]`.
+ * Unknown part types are skipped — better to drop a stray field than
+ * stringify an object into the rendered chat.
  */
 function extractDeltaText(delta: unknown): string {
   const extractReasoningText = (payload: unknown): string => {
@@ -564,8 +555,8 @@ function toOpenAIImageEditReferenceMessage(
 
 // Refusal flag stamped on assistant metadata when the backend emits the
 // `anthropic_refusal` _toolEvent. We drop the refused pair from the next
-// request body (Anthropic guidance: leaving refusals in context keeps
-// refusing). Metadata (not text) prevents content from spoofing a reset.
+// request body (Anthropic: leaving refusals in context keeps refusing).
+// Using metadata, not text, prevents content from spoofing a reset.
 function isAnthropicRefusalMessage(message: RunMessage): boolean {
   if (message.role !== "assistant") return false;
   const metadata = (message as { metadata?: unknown }).metadata as
@@ -618,8 +609,8 @@ function collectAssistantToolCalls(
       hasNativePart,
     );
     if (isServerSideBuiltin) {
-      // Gemini code_execution / image_generation still need to round-
-      // trip the native_part payload for native replay; drop the rest.
+      // Gemini code_execution / image_generation must round-trip the
+      // native_part payload for native replay; drop the rest.
       if (!hasNativePart) continue;
     }
     const argumentsStr =
@@ -640,8 +631,8 @@ function collectAssistantToolCalls(
       },
     };
     // Promote args.google to extra_content.google so the backend
-    // native_part replay branch can find it. The backend only inspects
-    // extra_content, not function.arguments.
+    // native_part replay branch finds it (it only inspects
+    // extra_content, not function.arguments).
     if (tc.extra_content !== undefined) {
       entry.extra_content = tc.extra_content;
     } else if (argsGoogle) {
@@ -670,7 +661,7 @@ function collectToolResultMessages(
     if (part.type !== "tool-call") continue;
     const tc = part as ToolCallMessagePart;
     const result = (tc as { result?: unknown }).result;
-    // Skip provider-side builtins; see isServerSideBuiltinToolPart.
+    // Skip provider-side builtins; see isServerSideBuiltinToolPart().
     const argsObj =
       tc.args && typeof tc.args === "object"
         ? (tc.args as Record<string, unknown>)
@@ -702,8 +693,8 @@ function collectToolResultMessages(
     let content: string;
     if (typeof result === "string") {
       // Backend ChatMessage validator rejects role="tool" with empty
-      // content; serialise a sentinel JSON so legitimately empty tool
-      // outputs still round-trip the follow-up turn to the provider.
+      // content; serialise sentinel JSON so legitimately empty tool
+      // outputs still round-trip to the provider.
       content = result.length > 0 ? result : JSON.stringify({ result: "" });
     } else {
       try {
@@ -774,8 +765,8 @@ function toOpenAIMessages(message: RunMessage): SerializedMessage[] {
       "[audio]",
     );
     if (isAnthropicRefusalMessage(message)) {
-      // Prune refused assistant turn from outbound history; the
-      // rendered transcript still shows the user-visible notice.
+      // Prune refused assistant turn from outbound history; the rendered
+      // transcript still shows the user-visible notice.
       return [];
     }
   }
@@ -795,9 +786,9 @@ function toOpenAIMessages(message: RunMessage): SerializedMessage[] {
   };
   if (toolCalls.length > 0) {
     base.tool_calls = toolCalls;
-    // OpenAI requires content === null on assistant turns whose
-    // payload is entirely tool_calls (matches the wire shape Gemini
-    // expects for the next functionCall replay).
+    // OpenAI requires content === null on assistant turns that are
+    // entirely tool_calls (matches the wire shape Gemini expects for
+    // the next functionCall replay).
     if (!textContent && imageParts.length === 0) {
       base.content = null;
     }
@@ -830,7 +821,7 @@ function findLatestUserImageBase64(messages: RunMessages): string | undefined {
       continue;
     }
 
-    // Image in message.content (e.g. compare view appends content with image parts)
+    // Image in message.content (e.g. compare view).
     for (const part of message.content ?? []) {
       if (part.type === "image" && "image" in part) {
         const encoded = extractImageBase64(part.image);
@@ -838,7 +829,7 @@ function findLatestUserImageBase64(messages: RunMessages): string | undefined {
       }
     }
 
-    // Image in message.attachments (e.g. chat composer)
+    // Image in message.attachments (e.g. chat composer).
     if ("attachments" in message && (message.attachments?.length ?? 0) > 0) {
       for (const attachment of message.attachments ?? []) {
         for (const part of attachment.content ?? []) {
@@ -858,7 +849,7 @@ function findLatestUserImageBase64(messages: RunMessages): string | undefined {
 }
 
 function findLatestUserAudioBase64(messages: RunMessages): string | undefined {
-  // Check message content parts (from compare view's CompareMessagePart with type: "audio")
+  // Message content parts (compare view CompareMessagePart type: "audio").
   for (let i = messages.length - 1; i >= 0; i -= 1) {
     const message = messages[i];
     if (!message || message.role !== "user") continue;
@@ -877,7 +868,7 @@ function findLatestUserAudioBase64(messages: RunMessages): string | undefined {
     }
   }
 
-  // Check the runtime store (from main composer's audio upload)
+  // Runtime store (main composer's audio upload).
   const pendingAudio = useChatRuntimeStore.getState().pendingAudioBase64;
   return pendingAudio ?? undefined;
 }
@@ -893,8 +884,8 @@ async function resolveUseAdapter(
     if (!thread?.pairId) {
       return undefined;
     }
-    // model1/model2 threads don't use the adapter toggle — each side
-    // loads its own model via /api/inference/load before generation.
+    // model1/model2 threads skip the adapter toggle — each side loads
+    // its own model via /api/inference/load before generation.
     if (thread.modelType === "model1" || thread.modelType === "model2") {
       return undefined;
     }
@@ -962,9 +953,9 @@ function waitForModelReady(abortSignal?: AbortSignal): Promise<void> {
 }
 
 /**
- * Auto-load the smallest downloaded model when the user tries to chat
- * without selecting one. Prefers GGUF (picks smallest cached variant),
- * falls back to smallest cached safetensors model.
+ * Auto-load the smallest downloaded model when the user chats without
+ * selecting one. Prefers GGUF (smallest cached variant), then smallest
+ * cached safetensors model.
  */
 // Cap cascade so broken cached repos can't spam /api/inference/load.
 const MAX_AUTO_LOAD_ATTEMPTS = 3;
@@ -1009,8 +1000,7 @@ async function autoLoadSmallestModel(): Promise<{
       listCachedModels().catch(() => []),
     ]);
 
-    // Try GGUF first: pick the repo with the smallest total size,
-    // then pick its smallest downloaded variant.
+    // GGUF first: smallest-total-size repo, then its smallest variant.
     if (ggufRepos.length > 0) {
       const sorted = [...ggufRepos].sort((a, b) => a.size_bytes - b.size_bytes);
       for (const repo of sorted) {
@@ -1053,7 +1043,7 @@ async function autoLoadSmallestModel(): Promise<{
               ...store.params,
               maxTokens: loadResp.context_length ?? 131072,
             });
-            // Add model to store so the selector shows the name
+            // Add to store so the selector shows the name.
             const autoModel: ChatModelSummary = {
               id: repo.repo_id,
               name: loadResp.display_name ?? repo.repo_id,
@@ -1101,7 +1091,7 @@ async function autoLoadSmallestModel(): Promise<{
       }
     }
 
-    // Fall back to safetensors models
+    // Fall back to safetensors models.
     if (modelRepos.length > 0) {
       const sorted = [...modelRepos].sort(
         (a, b) => a.size_bytes - b.size_bytes,
@@ -1171,7 +1161,7 @@ async function autoLoadSmallestModel(): Promise<{
       }
     }
 
-    // Cap also gates the default download so the total /api/inference/load
+    // Cap also gates the default download, so total /api/inference/load
     // budget across cached + fallback is MAX_AUTO_LOAD_ATTEMPTS, not +1.
     if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) {
       toast.dismiss(toastId);
@@ -1182,7 +1172,7 @@ async function autoLoadSmallestModel(): Promise<{
       };
     }
 
-    // No cached models found — try downloading a small default GGUF
+    // No cached models — try downloading a small default GGUF.
     toast("Downloading a small model…", {
       id: toastId,
       description:
@@ -1275,8 +1265,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
     async *run({ messages, abortSignal, unstable_threadId }) {
       await useChatRuntimeStore.getState().hydratePersistedSettings();
       let runtime = useChatRuntimeStore.getState();
-      // Capture the thread ID once at the start so it stays stable even if
-      // the user switches chats while waiting for model load / auto-load.
+      // Capture the thread ID once so it stays stable even if the user
+      // switches chats while waiting for model load / auto-load.
       const resolvedThreadId =
         (unstable_threadId ?? runtime.activeThreadId) || undefined;
       const sandboxSessionId = await resolveSandboxSessionId(resolvedThreadId);
@@ -1305,7 +1295,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         }
       };
 
-      // Wait for in-progress model load to finish before inferring
+      // Wait for in-progress model load before inferring.
       if (runtime.modelLoading) {
         toast.info("Waiting for model to finish loading…");
         try {
@@ -1317,7 +1307,6 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       }
 
       if (!useChatRuntimeStore.getState().params.checkpoint) {
-        // Auto-load the smallest downloaded model
         let loaded: boolean;
         let blockedByTrustRemoteCode: boolean;
         try {
@@ -1342,7 +1331,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         }
       }
 
-      // Re-read store after potential auto-load / model ready wait
+      // Re-read store after auto-load / model-ready wait.
       runtime = useChatRuntimeStore.getState();
       const { params } = runtime;
       const {
@@ -1405,8 +1394,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         throw new Error("Missing connection API key.");
       }
 
-      // Image-generation flag (OpenAI cloud + Responses-capable model).
-      // Computed first so Gemini image mode can suppress Search/Code.
+      // Image-generation flag (OpenAI cloud + Responses-capable model);
+      // computed first so Gemini image mode can suppress Search/Code.
       const imageGenerationEnabledForThisTurn = Boolean(
         externalProvider &&
           externalSelection &&
@@ -1444,9 +1433,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           ),
       );
       // Fetch pill is independent of Search (Anthropic bills web_fetch
-      // separately from web_search). Sourced from `webFetchToolsEnabled`;
-      // on providers without web_fetch the toggle is forced off in
-      // chat-page's runtime setState.
+      // separately). Sourced from `webFetchToolsEnabled`; on providers
+      // without web_fetch the toggle is forced off in chat-page setState.
       const webFetchEnabledForThisTurn = Boolean(
         externalProvider &&
           webFetchToolsEnabled &&
@@ -1482,7 +1470,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
 
       // toOpenAIMessages emits assistant tool_calls + role="tool"
       // follow-ups; the backend Gemini translator rebuilds the
-      // functionCall/functionResponse parts (with thoughtSignature).
+      // functionCall / functionResponse parts (with thoughtSignature).
       const outboundMessages = survivingMessages
         .flatMap(toOpenAIMessages)
         .filter((message): message is NonNullable<typeof message> =>
@@ -1508,9 +1496,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           }
         }
         // OpenAIChatMessage is a structural superset of SerializedMessage
-        // for the role/content axis the outbound pipeline consumes; cast
-        // through unknown since referenceMessage carries no tool_calls
-        // (the image_edit reference is a plain assistant turn).
+        // on the role/content axis; cast through unknown since
+        // referenceMessage carries no tool_calls (plain assistant turn).
         outboundMessages.splice(
           insertAt,
           0,
@@ -1545,10 +1532,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         const webLabel = providerShipsWebFetch
           ? "web search or web fetch"
           : "web search";
-        // Treat search and fetch as a single "any web tool" axis so
-        // the guard only warns when neither pill is on; checking
-        // webSearchEnabledForThisTurn alone mis-fired when only Fetch
-        // was on and suppressed live web_fetch calls.
+        // Treat search and fetch as one "any web tool" axis so the guard
+        // only warns when neither pill is on; checking webSearch alone
+        // mis-fired when only Fetch was on and suppressed web_fetch.
         const anyWebEnabledForThisTurn =
           webSearchEnabledForThisTurn || webFetchEnabledForThisTurn;
         if (
@@ -1625,15 +1611,14 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       }
 
       // Scan post-prune history so a refused user turn's image/audio
-      // doesn't gate or mis-attribute the next non-refused turn.
+      // doesn't gate or mis-attribute the next turn.
       const imageBase64 = findLatestUserImageBase64(survivingMessages);
       const audioBase64 = findLatestUserAudioBase64(survivingMessages);
       const hasOutboundImage = Boolean(imageBase64);
 
       // Keep render_html local-only and mirror the backend image-turn gate.
-      // Artifacts are independent of Search/Code: if a local tool-capable
-      // model has Artifacts enabled, expose render_html even when no other
-      // tool pills are active.
+      // Artifacts are independent of Search/Code: a local tool-capable model
+      // with Artifacts on exposes render_html even with no other pills active.
       const renderHtmlToolEnabledForThisTurn = Boolean(
         !isExternalRequest &&
           supportsTools &&
@@ -1652,10 +1637,10 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       addSystemInstruction(outboundMessages, effectiveDisabledToolGuard);
       addSystemInstruction(outboundMessages, artifactInstruction);
 
-      // Block when ANY image is in the outbound payload (current or
-      // prior turns) and the loaded model can't process images. Keeps
-      // the gate simple: once a chat contains an image, a non-vision
-      // model can't respond — user starts a new chat to switch models.
+      // Block when ANY image is in the outbound payload (current or prior
+      // turns) and the loaded model can't process images. Once a chat
+      // contains an image, a non-vision model can't respond — the user
+      // starts a new chat to switch models.
       if (imageBase64) {
         const activeModel = runtime.models.find(
           (m) => m.id === params.checkpoint,
@@ -1672,10 +1657,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         });
         if (imageGateReason) {
           toast.error(imageGateReason);
-          // Flip the per-thread running flag on→off so the compare-mode
-          // waitForRunEnd resolves instead of hanging. This gate fires
-          // before the streaming path's setThreadRunning(true), so the
-          // wait promise would otherwise never settle.
+          // Flip the per-thread running flag on→off so compare-mode
+          // waitForRunEnd resolves instead of hanging: this gate fires
+          // before the streaming path's setThreadRunning(true).
           const gatedThreadKey = resolvedThreadId || "__default";
           runtime.setThreadRunning(gatedThreadKey, true);
           runtime.setThreadRunning(gatedThreadKey, false);
@@ -1683,7 +1667,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           throw new Error(imageGateReason);
         }
       }
-      // Clear pending audio from store after extracting (consumed on send)
+      // Clear pending audio from store after extracting (consumed on send).
       if (audioBase64) {
         const audioName = runtime.pendingAudioName;
         if (audioName) {
@@ -1785,8 +1769,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       let reasoningStartAt: number | null = null;
       let reasoningDuration = 0;
       // True while wrapping a `delta.reasoning_content` stream in
-      // <think>...</think> for parseAssistantContent. Lives outside
-      // the SSE loop because the close tag fires when content arrives.
+      // <think>...</think> for parseAssistantContent. Lives outside the
+      // SSE loop because the close tag fires when content arrives.
       let reasoningContentOpen = false;
       // Tool call parts, cumulative; result lands on tool_end.
       const toolCallParts: ToolCallMessagePart[] = [];
@@ -1819,9 +1803,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         );
         return [...otherToolParts, ...textParts, ...imageToolParts];
       };
-      // Anthropic document_citations tool_event payload, converted to
-      // Sources-panel source parts at end-of-stream so the inline [N]
-      // markers have matching entries.
+      // Anthropic document_citations payload, converted to Sources-panel
+      // parts at end-of-stream so inline [N] markers have matching entries.
       const documentCitationParts: Array<{
         type: "source";
         sourceType: "url";
@@ -1830,16 +1813,16 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         title: string;
         metadata?: { description: string };
       }> = [];
-      // Latched on the `anthropic_refusal` tool event; stamped onto the
-      // final assistant metadata as `custom.anthropicRefusal` to drive
-      // the history-prune above.
+      // Latched on the `anthropic_refusal` tool event; stamped onto final
+      // assistant metadata as `custom.anthropicRefusal` to drive the
+      // history-prune above.
       let anthropicRefusalSeen = false;
       let serverMetadata: {
         usage?: ServerUsage;
         timings?: ServerTimings;
       } | null = null;
 
-      // Per-run cancellation token so a delayed stop POST cannot match
+      // Per-run cancellation token so a delayed stop POST can't match
       // the next run on the same thread.
       const cancelId =
         typeof crypto !== "undefined" && "randomUUID" in crypto
@@ -1854,10 +1837,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         // Plain fetch, not authFetch: authFetch redirects to login on
         // 401, which would kick the user out mid-stop.
         const token = getAuthToken();
-        // Use apiUrl so the cancel POST reaches the right origin in
-        // Tauri production builds (where the webview origin is not the
-        // backend at 127.0.0.1:<port>). Browser/dev builds get the empty
-        // base, so the path is unchanged there.
+        // Use apiUrl so the cancel POST reaches the right origin in Tauri
+        // production builds (webview origin != backend at 127.0.0.1:<port>).
+        // Browser/dev builds get the empty base, so the path is unchanged.
         void fetch(apiUrl("/api/inference/cancel"), {
           method: "POST",
           headers: {
@@ -1948,8 +1930,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 anthropicCodeExecContainerId = null;
               }
               // Pre-send container validation (OpenAI). Stale ids drop
-              // silently and fall through to lazy-create. On list-call
-              // failure, skip and rely on the backend's retry path.
+              // silently and fall through to lazy-create; on list-call
+              // failure, rely on the backend's retry path.
               let activeContainerIds: Set<string> | null = null;
               if (externalProvider.providerType === "openai") {
                 try {
@@ -1972,8 +1954,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   openaiCodeExecContainerId = null;
                 }
               }
-              // Cross-thread inheritance: reuse the most recently used
-              // container from any other thread; opt-out via the picker.
+              // Cross-thread inheritance: reuse the most recent container
+              // from any other thread; opt-out via the picker.
               if (
                 !openaiCodeExecContainerId &&
                 externalProvider.providerType === "openai"
@@ -1985,7 +1967,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   for (const t of others) {
                     if (t.id === resolvedThreadId) continue;
                     if (!t.openaiCodeExecContainerId) continue;
-                    // Skip ids not in active set; null on source thread so
+                    // Skip ids not in active set; null the source thread so
                     // the next pass doesn't re-pick a dead id.
                     if (
                       activeContainerIds &&
@@ -2006,9 +1988,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   /* fall through to lazy-create below */
                 }
               }
-              // Pre-create our own container (vs container_auto) so it
-              // shows up in the picker with a friendly name and the
-              // configured TTL. Falls back to container_auto on failure.
+              // Pre-create our own container (vs container_auto) so it shows
+              // in the picker with a friendly name and the configured TTL.
+              // Falls back to container_auto on failure.
               if (
                 !openaiCodeExecContainerId &&
                 externalProvider.providerType === "openai"
@@ -2022,10 +2004,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                       baseUrl: externalProvider.baseUrl || null,
                     },
                     {
-                      // Friendly English-word name so the container
-                      // is human-readable in the picker list (e.g.
-                      // "kestrel-3f9c") instead of a thread-id slug
-                      // or OpenAI's default blank name.
+                      // Friendly English-word name so the container is
+                      // human-readable in the picker (e.g. "kestrel-3f9c")
+                      // instead of a thread-id slug or blank default.
                       name: pickFriendlyContainerName(),
                       ttlMinutes: ttlToUse,
                     },
@@ -2035,10 +2016,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                     openaiCodeExecContainerId: created.id,
                   }).catch(() => {});
                 } catch {
-                  // Fall back to backend's container_auto path on
-                  // failure — keeps the chat moving; the next turn
-                  // can retry. The auto-created container will be
-                  // unnamed, but the chat doesn't break.
+                  // Fall back to the backend's container_auto path on
+                  // failure — keeps the chat moving (the auto-created
+                  // container is unnamed); the next turn can retry.
                   openaiCodeExecContainerId = null;
                 }
               }
@@ -2047,8 +2027,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               model: externalSelection.modelId,
               messages: outboundMessages,
               stream: true,
-              // Reasoning-class models (OpenAI gpt-5.x / o3) reject temperature
-              // and top_p; only forward when the active provider supports them.
+              // Reasoning-class models (OpenAI gpt-5.x / o3) reject
+              // temperature and top_p; forward only when supported.
               ...(externalCapabilities?.temperature !== false
                 ? { temperature: params.temperature }
                 : {}),
@@ -2072,8 +2052,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               ...(externalCapabilities?.presencePenalty
                 ? { presence_penalty: params.presencePenalty }
                 : {}),
-              // Compose the enabled_tools list from the active pills;
-              // backend maps each name to the provider's tool schema.
+              // enabled_tools from active pills; backend maps each name
+              // to the provider's tool schema.
               ...(webSearchEnabledForThisTurn ||
               webFetchEnabledForThisTurn ||
               codeExecEnabledForThisTurn ||
@@ -2128,8 +2108,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 ? { prompt_cache_ttl: externalProvider.promptCacheTtl }
                 : {}),
               // Anthropic fast mode (Opus 4.6 / 4.7 only); backend
-              // silently drops on unsupported models as a second
-              // line of defence.
+              // silently drops on unsupported models as a backstop.
               ...(params.fastMode &&
               providerSupportsFastMode(
                 externalProvider.providerType,
@@ -2230,8 +2209,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               }
 
               // Emit tool-call content parts for assistant-ui.
-              // On tool_start: add a new tool-call part (renders in "running" state).
-              // On tool_end: set result on the existing part (transitions to "complete").
+              // tool_start: add a part (renders "running").
+              // tool_end: set result on the part (transitions to "complete").
               const toolEvent = (
                 chunk as unknown as { _toolEvent?: Record<string, unknown> }
               )._toolEvent;
@@ -2253,8 +2232,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   continue;
                 }
                 if (toolEvent.type === "document_citations") {
-                  // Convert Anthropic citations_delta footnotes into
-                  // Sources-panel entries matching the inline [N] markers.
+                  // Convert citations_delta footnotes into Sources-panel
+                  // entries matching the inline [N] markers.
                   const cits = toolEvent.citations;
                   if (Array.isArray(cits)) {
                     cits.forEach((entry, idx) => {
@@ -2286,8 +2265,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   continue;
                 }
                 if (toolEvent.type === "anthropic_refusal") {
-                  // Latch the backend refusal signal so the final
-                  // message metadata can drive the prune.
+                  // Latch the backend refusal signal so final message
+                  // metadata can drive the prune.
                   anthropicRefusalSeen = true;
                   continue;
                 }
@@ -2346,7 +2325,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                       imageB64
                     ) {
                       // Backend keeps base64 on separate image_b64 /
-                      // image_mime fields so logs stay small; repackage.
+                      // image_mime fields so logs stay small; repackage here.
                       parsedResult = {
                         image_b64: imageB64,
                         image_mime:
@@ -2359,8 +2338,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                       };
                     } else if (imgIdx !== -1) {
                       const text = rawResult.slice(0, imgIdx);
-                      // Fall back to "_default" to match the backend sandbox directory
-                      // used when no session_id is provided (see tools.py _get_workdir).
+                      // Fall back to "_default" to match the backend sandbox
+                      // dir used when no session_id (see tools.py _get_workdir).
                       const sessionId = sandboxSessionId || "_default";
                       try {
                         const images = JSON.parse(
@@ -2386,7 +2365,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                     // Merge tool_end native_part into args.google so the
                     // outbound translator replays both start (executableCode)
                     // and end (result / inlineData) on the same turn.
-                    // Concatenate parts so each keeps its own thoughtSignature.
+                    // Concatenate so each part keeps its own thoughtSignature.
                     const endGoogle = (
                       toolEvent as { google?: { native_part?: unknown } }
                     ).google;
@@ -2410,9 +2389,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                         string,
                         unknown
                       >;
-                      // Extract part entries from either parts:[...] or
-                      // legacy single-object native_part. Legacy
-                      // thoughtSignature always belongs on executableCode.
+                      // Extract part entries from parts:[...] or legacy
+                      // single-object native_part. Legacy thoughtSignature
+                      // always belongs on executableCode.
                       const collectParts = (
                         native: Record<string, unknown>,
                       ): Record<string, unknown>[] => {
@@ -2466,8 +2445,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                     };
                   }
                 }
-                // Cumulative yield. orderAssistantContent puts search/
-                // code before text and generated images after.
+                // Cumulative yield; orderAssistantContent puts search/code
+                // before text and generated images after.
                 const textParts = pinTextThoughtSignature(
                   parseAssistantContent(cumulativeText),
                 );
@@ -2485,7 +2464,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 continue;
               }
 
-              // OpenAI-standard usage chunk: choices=[], usage populated
+              // OpenAI-standard usage chunk: choices=[], usage populated.
               if (chunk.choices?.length === 0 && chunk.usage) {
                 serverMetadata = {
                   usage: chunk.usage,
@@ -2497,8 +2476,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               }
 
               totalChunks += 1;
-              // Latch the chunk's `model` field so the openrouter/free
-              // chip can show the chosen underlying model.
+              // Latch the chunk's `model` field so the openrouter/free chip
+              // shows the chosen underlying model.
               if (
                 isExternalRequest &&
                 externalProvider?.providerType === "openrouter" &&
@@ -2517,7 +2496,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 }
               }
               const rawDelta = chunk.choices?.[0]?.delta?.content;
-              // Normalize structured delta.content (mistral magistral) to text.
+              // Normalize structured delta.content (mistral magistral).
               const delta = extractDeltaText(rawDelta);
               // Latest Gemini text-part thoughtSignature for next-turn replay.
               const deltaExtraContent = (
@@ -2539,15 +2518,15 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   }
                 }
               }
-              // Kimi / DeepSeek stream thinking via delta.reasoning_content.
-              // Wrap inline as <think>...</think> for parseAssistantContent.
+              // Kimi / DeepSeek stream thinking via delta.reasoning_content;
+              // wrap inline as <think>...</think> for parseAssistantContent.
               const rawReasoning = (
                 chunk.choices?.[0]?.delta as
                   | { reasoning_content?: unknown }
                   | undefined
               )?.reasoning_content;
               // OpenRouter ships reasoning as delta.reasoning_details[]
-              // regardless of underlying provider; merge into the same wrap path.
+              // regardless of provider; merge into the same wrap path.
               const rawReasoningDetails = (
                 chunk.choices?.[0]?.delta as
                   | { reasoning_details?: unknown }
@@ -2567,7 +2546,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 reasoningFromDetails;
               // OpenAI delta.tool_calls: streams fragments by index;
               // accumulate into one part. extra_content carries Gemini 3
-              // thoughtSignature for next-turn replay.
+              // thoughtSignature for replay.
               const rawDeltaToolCalls = (
                 chunk.choices?.[0]?.delta as
                   | { tool_calls?: unknown }
@@ -2588,9 +2567,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                   const idx =
                     typeof call.index === "number" ? call.index : undefined;
                   const stableId = call.id;
-                  // Match an existing fragment by id first (canonical),
-                  // then by index slot. Fall back to a freshly-minted
-                  // tool_call_<n> id for streams that send neither.
+                  // Match an existing fragment by id first (canonical), then
+                  // by index slot; fall back to a minted tool_call_<n> id
+                  // for streams that send neither.
                   let existing = stableId
                     ? toolCallParts.find((p) => p.toolCallId === stableId)
                     : undefined;
@@ -2777,9 +2756,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             throw streamError;
           }
         }
-        // If the stream ended while we were still inside a
-        // delta.reasoning_content block (Kimi / DeepSeek path), close
-        // the open <think> tag so the reasoning panel parses cleanly.
+        // If the stream ended inside a delta.reasoning_content block
+        // (Kimi / DeepSeek), close the open <think> tag so the reasoning
+        // panel parses cleanly.
         if (reasoningContentOpen) {
           cumulativeText += "</think>";
           reasoningContentOpen = false;
@@ -2787,9 +2766,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         settleFirstTokenOk();
 
         // Extract source parts from completed web_search and web_fetch
-        // tool calls. Both emit the same `Title:` / `URL:` / `Snippet:`
-        // block shape from the Anthropic backend, so the parser does
-        // not need to branch on tool name.
+        // calls. Both emit the same `Title:` / `URL:` / `Snippet:` block
+        // shape, so the parser need not branch on tool name.
         const sourceParts = toolCallParts.flatMap((tc) => {
           if (
             (tc.toolName !== "web_search" && tc.toolName !== "web_fetch") ||
@@ -2818,8 +2796,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         const cacheWriteTokens = meta?.usage?.cache_creation_input_tokens ?? 0;
 
         // Gate on the captured checkpoint still being active so a late
-        // completion from provider A doesn't populate the bar after the
-        // user switched to provider B mid-stream.
+        // completion from provider A doesn't populate the bar after a
+        // mid-stream switch to provider B.
         if (
           meta?.usage &&
           typeof meta.usage.prompt_tokens === "number" &&
@@ -2882,10 +2860,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         if (!abortSignal.aborted) {
           const msg = err instanceof Error ? err.message : String(err);
           if (isContextLimitError(msg)) {
-            // llama-server was launched with --no-context-shift, so it
-            // returns a hard error instead of silently dropping old
-            // turns from the KV cache. Point the user at the exact
-            // control that raises the ceiling.
+            // llama-server runs with --no-context-shift, returning a hard
+            // error instead of silently dropping old KV-cache turns. Point
+            // the user at the control that raises the ceiling.
             toast.error("Context limit reached", {
               description:
                 "The conversation has filled the model's context window. " +
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 7a71850d66..9d0939c37d 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -141,9 +141,8 @@ export interface DownloadProgressResponse {
   expected_bytes: number;
   progress: number;
   /**
-   * Resolved on-disk path of the snapshot dir (or cache repo root if no
-   * snapshot exists yet). Null when nothing has been written to the
-   * cache for this repo.
+   * On-disk path of the snapshot dir (or cache repo root if no snapshot yet).
+   * Null when nothing has been written to the cache for this repo.
    */
   cache_path: string | null;
 }
@@ -168,9 +167,8 @@ export type ModelLoadPhase = "mmap" | "ready" | null;
 
 export interface LoadProgressResponse {
   /**
-   * Load phase: ``"mmap"`` while the llama-server subprocess is paging
-   * weight shards into RAM, ``"ready"`` once it has reported healthy,
-   * or ``null`` when no load is in flight.
+   * Load phase: "mmap" while llama-server pages weight shards into RAM,
+   * "ready" once healthy, or null when no load is in flight.
    */
   phase: ModelLoadPhase;
   bytes_loaded: number;
@@ -179,10 +177,9 @@ export interface LoadProgressResponse {
 }
 
 /**
- * Fetch the active GGUF load's mmap/upload progress. Complements
- * ``getDownloadProgress`` / ``getGgufDownloadProgress`` for the window
- * between "download complete" and "chat ready", which for large MoE
- * models can be several minutes of otherwise-opaque spinning.
+ * Fetch the active GGUF load's mmap/upload progress. Complements the download
+ * progress endpoints for the "download complete" -> "chat ready" window, which
+ * for large MoE models can be several minutes of otherwise-opaque spinning.
  */
 export async function getLoadProgress(): Promise<LoadProgressResponse> {
   const response = await authFetch(`/api/inference/load-progress`);
@@ -551,16 +548,14 @@ export async function buildBackendChatExport(): Promise<{
   return parseJsonOrThrow(response);
 }
 
-// Legacy-Dexie import ledger. The server-side source of truth that
-// replaces the boolean localStorage sentinel
-// (`unsloth_chat_legacy_imported_to_studio_db`) so a studio.db wipe
-// makes the import recoverable.
+// Legacy-Dexie import ledger: server-side source of truth replacing the
+// boolean localStorage sentinel, so a studio.db wipe keeps the import
+// recoverable.
 export async function listChatImportLedger(): Promise<Set<string>> {
   const response = await authFetch("/api/chat/import-ledger");
-  // Backend deployments that don't have this endpoint yet behave the
-  // same as an empty ledger -- caller treats every legacy thread as
-  // un-imported and tries to import. The UPSERT semantics in
-  // syncChatMessages prevent duplicates, so this fallback is safe.
+  // Backends without this endpoint behave like an empty ledger -- caller
+  // re-imports every legacy thread. syncChatMessages UPSERTs prevent
+  // duplicates, so this fallback is safe.
   if (response.status === 404 || response.status === 405) return new Set();
   const data = await parseJsonOrThrow<{ threadIds: string[] }>(response);
   return new Set(data.threadIds);
@@ -569,9 +564,9 @@ export async function listChatImportLedger(): Promise<Set<string>> {
 export interface RecordChatImportLedgerResult {
   accepted: number;
   inserted: number;
-  // false when the backend predates /api/chat/import-ledger (404/405/501)
-  // so the caller can avoid poisoning the localStorage perf hint -- the
-  // next launch will retry the (idempotent) import.
+  // false when the backend predates /api/chat/import-ledger (404/405/501) so
+  // the caller avoids poisoning the localStorage perf hint; next launch
+  // retries the (idempotent) import.
   supported: boolean;
 }
 
@@ -633,11 +628,10 @@ export async function browseFolders(
   if (path !== undefined && path !== null) params.set("path", path);
   if (showHidden) params.set("show_hidden", "true");
   const qs = params.toString();
-  // Forward the AbortSignal through authFetch -> fetch so that a
-  // navigation cancelled in the FolderBrowser (rapid breadcrumb / row /
-  // hidden-toggle clicks) actually cancels the in-flight HTTP request
-  // server-side, instead of merely dropping the response client-side
-  // while the backend keeps walking large directory trees.
+  // Forward the AbortSignal through authFetch -> fetch so a cancelled
+  // FolderBrowser navigation actually cancels the in-flight request
+  // server-side, instead of just dropping the response while the backend
+  // keeps walking large directory trees.
   const response = await authFetch(
     `/api/models/browse-folders${qs ? `?${qs}` : ""}`,
     signal ? { signal } : undefined,
diff --git a/studio/frontend/src/features/chat/api/openai-containers.ts b/studio/frontend/src/features/chat/api/openai-containers.ts
index 6463b5dc7b..f95b5a6349 100644
--- a/studio/frontend/src/features/chat/api/openai-containers.ts
+++ b/studio/frontend/src/features/chat/api/openai-containers.ts
@@ -2,11 +2,10 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Wrappers for the three OpenAI shell-tool container management
- * endpoints exposed by the backend (studio/backend/routes/inference.py).
- * Each one proxies to OpenAI's /v1/containers REST surface using the
- * user's encrypted API key. Backend rejects any base URL that isn't
- * api.openai.com — the shell tool only exists on the managed cloud.
+ * Wrappers for the backend's three OpenAI shell-tool container endpoints
+ * (studio/backend/routes/inference.py). Each proxies to OpenAI's
+ * /v1/containers using the user's encrypted API key. Backend rejects any
+ * base URL but api.openai.com (the shell tool is managed-cloud only).
  */
 
 import { authFetch } from "@/features/auth";
@@ -110,9 +109,8 @@ export async function deleteOpenAIContainer(
       }),
     },
   );
-  // 404 = container already gone (deleted elsewhere, or expired-then-purged).
-  // Treat as idempotent success so a stale list entry doesn't surface as a
-  // confusing error — the caller will refresh and the entry will disappear.
+  // 404 = container already gone; treat as idempotent success so a stale
+  // list entry doesn't surface as a confusing error.
   if (!response.ok && response.status !== 204 && response.status !== 404) {
     throw new Error(await parseError(response));
   }
diff --git a/studio/frontend/src/features/chat/api/providers-api.ts b/studio/frontend/src/features/chat/api/providers-api.ts
index 9625c4c217..d52e5e855e 100644
--- a/studio/frontend/src/features/chat/api/providers-api.ts
+++ b/studio/frontend/src/features/chat/api/providers-api.ts
@@ -140,10 +140,9 @@ export async function deleteProviderConfig(providerId: string): Promise<void> {
   const response = await authFetch(`/api/providers/${providerId}`, {
     method: "DELETE",
   });
-  // Treat 404 as success: another browser (or tab) already deleted this
-  // provider on the backend, so locally pruning the stale cache is the
-  // correct follow-up. Without this, the caller would throw and the user
-  // would be stuck with an entry they cannot remove from the UI.
+  // Treat 404 as success: another tab already deleted this provider, so pruning
+  // the stale cache is correct. Otherwise the caller throws and the user is stuck
+  // with an entry they cannot remove from the UI.
   if (response.status === 404) {
     return;
   }
diff --git a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
index b4ddf38aaa..a685515253 100644
--- a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
+++ b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
@@ -46,10 +46,9 @@ function buildHtmlFence(source: string): string {
   const fence = "`".repeat(longestBacktickRun + 1);
   return `${fence}html\n${source}\n${fence}`;
 }
-// Sandboxed artifact iframes are intentionally excluded from the overlay focus
-// trap. Granting same-origin sandbox privileges would weaken isolation, so
-// keyboard users can reach Studio controls here while fully interactive artifact
-// content remains a known sandbox limitation.
+// Sandboxed artifact iframes are deliberately outside the overlay focus trap:
+// granting same-origin sandbox privileges would weaken isolation, so reaching
+// interactive artifact content via keyboard is a known sandbox limitation.
 const FOCUSABLE_SELECTOR =
   'a[href], button:not([disabled]), textarea:not([disabled]), input:not([disabled]), select:not([disabled]), [tabindex]:not([tabindex="-1"])';
 
diff --git a/studio/frontend/src/features/chat/artifacts/html-frame.tsx b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
index dc8c14c0ed..041dc4dc56 100644
--- a/studio/frontend/src/features/chat/artifacts/html-frame.tsx
+++ b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
@@ -60,9 +60,8 @@ export function ArtifactHtmlFrame({
     return apiUrl(`/api/inference/artifact-preview-frame?${query.toString()}`);
   }, [allowNetworkAccess, code]);
   const postArtifactHtml = useCallback(() => {
-    // The sandboxed frame intentionally has an opaque origin ("null").
-    // A wildcard target is required here;
-    // the payload is sent only to this iframe's contentWindow.
+    // Sandboxed frame has an opaque origin ("null"), so a wildcard target is
+    // required; the payload only reaches this iframe's contentWindow.
     iframeRef.current?.contentWindow?.postMessage(
       { type: "unsloth:artifact-html", html: artifactHtml },
       "*",
diff --git a/studio/frontend/src/features/chat/artifacts/types.ts b/studio/frontend/src/features/chat/artifacts/types.ts
index 86d42fd733..5ce22cc68a 100644
--- a/studio/frontend/src/features/chat/artifacts/types.ts
+++ b/studio/frontend/src/features/chat/artifacts/types.ts
@@ -44,8 +44,8 @@ export function hashArtifactCode(code: string): string {
 export function createArtifactId(input: ChatArtifactInput): string {
   const threadSegment = input.threadId || "no-thread";
   const messageSegment = input.sourceMessageId || "transient";
-  // Backend tool call IDs (call_0, call_1, …) reset per request, so
-  // the message ID is needed to scope them to a specific turn.
+  // Backend tool call IDs (call_0, call_1, …) reset per request, so the
+  // message ID is needed to scope them to a specific turn.
   const parts = [input.source, threadSegment, messageSegment];
 
   if (input.source === "tool" && input.sourceToolCallId) {
diff --git a/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx b/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
index c706d971ac..f78784b436 100644
--- a/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
@@ -84,10 +84,9 @@ function isValidAddress(value: string): boolean {
       return false;
     }
   }
-  // Anything else is treated as a local command (stdio); the backend gates
-  // whether stdio servers are allowed on this host. Reject other URL schemes
-  // only when the command itself is a URL; "://" is fine inside an argument
-  // (e.g. a database connection string passed to the server).
+  // Otherwise it's a local command (stdio); the backend gates whether those
+  // are allowed. Reject only when the command itself is a URL; "://" is fine
+  // inside an argument (e.g. a DB connection string passed to the server).
   return !trimmed.split(/\s+/)[0].includes("://");
 }
 
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 834d0baf34..50650416a4 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -367,9 +367,8 @@ function modelMatchesDeleted(
 }
 
 /**
- * Detect if this is a LoRA base-vs-fine-tuned compare.
- * Returns true when the loaded checkpoint is a LoRA — in that case
- * we use the fast simultaneous base/lora adapter-toggle path.
+ * True when the loaded checkpoint is a LoRA, meaning a base-vs-fine-tuned
+ * compare that uses the fast simultaneous adapter-toggle path.
  */
 function useIsLoraCompare(): boolean {
   return useChatRuntimeStore((s) => {
@@ -424,15 +423,12 @@ const CompareContent = memo(function CompareContent({
 });
 
 /**
- * A single column in the compare layout. Hosts one ChatRuntimeProvider
- * and one Thread rendered with hideComposer — the composer is shared
- * across both panes and rendered outside the pane flex.
+ * A single column in the compare layout: one ChatRuntimeProvider and one
+ * Thread with hideComposer (the composer is shared across panes).
  *
- * Each pane is a flex item with `flex-1 basis-0 min-h-0 min-w-0` so on
- * mobile (flex-col) they share height equally, and on desktop (flex-row)
- * they share width equally. The `min-*` constraints are required for
- * the inner viewport to scroll internally instead of spilling into the
- * page.
+ * Each pane is `flex-1 basis-0 min-h-0 min-w-0` so panes share height
+ * (mobile flex-col) or width (desktop flex-row) equally. The `min-*`
+ * constraints let the inner viewport scroll instead of spilling.
  */
 function ComparePane({
   modelType,
@@ -476,16 +472,13 @@ function ComparePane({
 }
 
 /**
- * Shared shell for both compare variants. A vertical flex column with
- * the two panes as siblings and the shared composer docked at the
- * bottom. On mobile the panes stack (flex-col); on desktop they sit
- * side by side (md:flex-row).
+ * Shared shell for both compare variants: a flex column with the two panes
+ * as siblings and the shared composer docked at the bottom. Panes stack on
+ * mobile (flex-col), sit side by side on desktop (md:flex-row).
  *
- * Flex is used rather than CSS grid for the pane container so that
- * viewport sizing stays stable across viewport-size transitions. Grid
- * rows with 1fr were triggering resize thrash in assistant-ui's
- * autoscroll hook on breakpoint crossings, leaving it stuck in a
- * scroll-to-bottom loop.
+ * Flex, not grid, for the pane container: grid rows with 1fr triggered
+ * resize thrash in assistant-ui's autoscroll on breakpoint crossings,
+ * leaving it stuck in a scroll-to-bottom loop.
  */
 function CompareShell({
   handlesRef,
@@ -594,10 +587,9 @@ const LoraCompareContent = memo(function LoraCompareContent({
 });
 
 /**
- * Per-pane header rendered inside GeneralCompareContent. Contains the
- * model selector aligned with the global topbar height. The left pane
- * reserves room for the mobile sidebar trigger; the right pane reserves
- * room for the global settings button.
+ * Per-pane header (inside GeneralCompareContent) with the model selector,
+ * aligned to the global topbar height. Left pane reserves room for the
+ * mobile sidebar trigger; right pane for the global settings button.
  */
 function GeneralCompareHeader({
   models,
@@ -1087,9 +1079,8 @@ export function ChatPage(): ReactElement {
   const [modelSelectorOpen, setModelSelectorOpen] = useState(false);
   const [modelSelectorLocked, setModelSelectorLocked] = useState(false);
   const viewBeforeCompareRef = useRef<ChatSearch | null>(null);
-  // Tracks the latest non-compare view so exiting compare can restore it even
-  // when compare was opened from a path that does not set viewBeforeCompareRef
-  // (e.g. the composer + menu).
+  // Latest non-compare view, so exiting compare can restore it even when
+  // compare was opened from a path that doesn't set viewBeforeCompareRef.
   const lastNonCompareViewRef = useRef<ChatSearch | null>(null);
   useEffect(() => {
     if (!search.compare) {
@@ -1249,15 +1240,11 @@ export function ChatPage(): ReactElement {
       preferredEffort,
       effortLevels,
     );
-    // Per-provider default effort. Anthropic gets the highest available
-    // level (xhigh on 4.6/4.7, high on 4.5) since Claude's adaptive
-    // thinking adjusts cost per turn — sitting at the top of the dial
-    // gives users the strongest answers and the model can still skip
-    // thinking when the turn is trivial. OpenAI gets "high" by default
-    // — the gpt-5.x reasoning models accept high across the board and
-    // it's the right cost/quality sweet spot for Responses-API tools
-    // (web search included). Everyone else gets "medium" as a balanced
-    // default. Users can pick another level via the Think dropdown.
+    // Per-provider default effort. Anthropic gets the highest level since
+    // Claude's adaptive thinking adjusts cost per turn (top of dial =
+    // strongest answers, still skips thinking when trivial). OpenAI gets
+    // "high" (gpt-5.x accept it across the board; good cost/quality for
+    // Responses-API tools). Everyone else "medium". Overridable via Think.
     const isAnthropic = provider?.providerType === "anthropic";
     const isOpenAI = provider?.providerType === "openai";
     const anthropicTopEffort = effortLevels.includes("xhigh")
@@ -1298,19 +1285,15 @@ export function ChatPage(): ReactElement {
     const supportsBuiltinWebFetch = providerSupportsBuiltinWebFetch(
       provider?.providerType,
     );
-    // Kimi's k2.6/k2.5 default to thinking enabled on the server side
-    // (per https://platform.kimi.ai/docs/models). Mirror that default
-    // in the UI so the Think pill comes up clicked when the user picks
-    // a Kimi model. The Search pill stays off by default; the mutual-
-    // exclusion handlers in the composer flip the two when needed.
+    // Kimi's k2.6/k2.5 default to thinking enabled server-side (per
+    // https://platform.kimi.ai/docs/models). Mirror that so the Think pill
+    // comes up clicked for Kimi models. Search stays off; the composer's
+    // mutual-exclusion handlers flip the two when needed.
     const isKimi = provider?.providerType === "kimi";
-    // Web search is on by default for the two providers we trust most
-    // for it: Anthropic (web_search_20250305 server tool, structured
-    // citations) and OpenAI (/v1/responses web_search, structured
-    // citations). Other providers stay off-by-default — OpenRouter's
-    // plugins shape and Kimi's $web_search builtin still work when the
-    // user opts in via the pill, but they're a notch less reliable so
-    // we don't pre-enable them.
+    // Web search on by default only for the two providers we trust most:
+    // Anthropic and OpenAI (both with structured citations). Others stay
+    // off-by-default; OpenRouter and Kimi work on opt-in but are less
+    // reliable, so we don't pre-enable them.
     const searchOnByDefault =
       supportsBuiltinWebSearch &&
       (provider?.providerType === "anthropic" ||
@@ -1345,15 +1328,11 @@ export function ChatPage(): ReactElement {
           : true
         : state.reasoningEnabled,
       supportsPreserveThinking: false,
-      // External models never give us a local tool runtime (no
-      // python sandbox), so `supportsTools` must be false. The three
-      // `supportsBuiltin*` flags pick up the slack for providers that
-      // run the tool server-side: `supportsBuiltinWebSearch` lights
-      // up the Search pill (OpenAI / Anthropic / OpenRouter / Kimi),
-      // `supportsBuiltinCodeExecution` lights up the Code pill
-      // (Anthropic Claude 4.x and OpenAI gpt-5.5), and
-      // `supportsBuiltinImageGeneration` lights up the Images pill
-      // (OpenAI cloud Responses-API models only).
+      // External models have no local tool runtime, so `supportsTools` is
+      // false. The `supportsBuiltin*` flags cover providers that run tools
+      // server-side: WebSearch lights the Search pill (OpenAI/Anthropic/
+      // OpenRouter/Kimi), CodeExecution the Code pill (Claude 4.x, gpt-5.5),
+      // ImageGeneration the Images pill (OpenAI cloud Responses-API only).
       supportsTools: false,
       supportsBuiltinWebSearch,
       supportsBuiltinCodeExecution,
@@ -1485,9 +1464,9 @@ export function ChatPage(): ReactElement {
   useEffect(() => {
     if (view.mode !== "single") return;
     if (view.threadId || view.newThreadNonce || !selectedArtifact) return;
-    // view intentionally excludes __LOCALID_ threads (they fall through to
-    // { mode: "single" } with no threadId/nonce).  Don't close an artifact
-    // whose thread is the currently active local thread.
+    // view excludes __LOCALID_ threads (they fall through to mode:"single"
+    // with no threadId/nonce). Don't close an artifact whose thread is the
+    // active local thread.
     if (
       selectedArtifact.threadId &&
       selectedArtifact.threadId === activeThreadId
@@ -1593,9 +1572,8 @@ export function ChatPage(): ReactElement {
           preferredEffort,
           effortLevels,
         );
-        // Same per-provider default policy as the useEffect path above:
-        // Anthropic picks the highest available level, OpenAI picks
-        // "high", everyone else picks "medium".
+        // Same per-provider default policy as the useEffect above:
+        // Anthropic highest level, OpenAI "high", everyone else "medium".
         const isAnthropic = selectedProvider?.providerType === "anthropic";
         const isOpenAI = selectedProvider?.providerType === "openai";
         const anthropicTopEffort = effortLevels.includes("xhigh")
@@ -1617,9 +1595,8 @@ export function ChatPage(): ReactElement {
                 ? "medium"
                 : clampedEffort
           : store.reasoningEffort;
-        // Clear any cached router-picked openrouter/free model unless the
-        // user is staying on openrouter/free — otherwise the chip would
-        // keep showing a stale ":<chosen>" suffix from a previous model.
+        // Clear any cached router-picked openrouter/free model unless staying
+        // on openrouter/free, else the chip keeps a stale ":<chosen>" suffix.
         const stillOnOpenRouterFree =
           selectedProvider?.providerType === "openrouter" &&
           selectedExternal?.modelId === "openrouter/free";
@@ -1644,13 +1621,12 @@ export function ChatPage(): ReactElement {
         const supportsBuiltinWebFetch = providerSupportsBuiltinWebFetch(
           selectedProvider?.providerType,
         );
-        // See sibling useEffect above: Kimi's k2.x default to thinking
-        // enabled, so the Think pill comes up clicked. Search pill stays
-        // off by default; mutual exclusion flips them via the composer.
+        // See sibling useEffect: Kimi's k2.x default to thinking enabled
+        // (Think pill clicked). Search stays off; the composer's mutual
+        // exclusion flips them.
         const isKimi = selectedProvider?.providerType === "kimi";
-        // Mirror of sibling useEffect: Anthropic and OpenAI get Search
-        // on-by-default since their server tools emit structured
-        // citations end-to-end. OpenRouter and Kimi stay off-by-default.
+        // Mirror of sibling useEffect: Anthropic/OpenAI get Search on by
+        // default (structured citations end-to-end); others stay off.
         const searchOnByDefault =
           supportsBuiltinWebSearch &&
           (selectedProvider?.providerType === "anthropic" ||
@@ -1676,9 +1652,8 @@ export function ChatPage(): ReactElement {
           ggufMaxContextLength: null,
           ggufNativeContextLength: null,
           activeNativePathToken: null,
-          // Clear previous-model counters; the relaxed external-provider
-          // render gate would otherwise show stale stats until the next
-          // completion overwrites them.
+          // Clear previous-model counters, else the relaxed external-provider
+          // render gate shows stale stats until the next completion.
           contextUsage: null,
           supportsReasoning: reasoningCaps.supportsReasoning,
           reasoningAlwaysOn: reasoningCaps.reasoningAlwaysOn,
@@ -1694,14 +1669,10 @@ export function ChatPage(): ReactElement {
               : true
             : store.reasoningEnabled,
           supportsPreserveThinking: false,
-          // External models have no local tool runtime → supportsTools
-          // stays false. The three supportsBuiltin* flags carry the
-          // server-side capability info for each pill:
-          //   - Search → providerSupportsBuiltinWebSearch
-          //   - Code   → providerSupportsBuiltinCodeExecution
-          //              (Anthropic Claude 4.x + OpenAI gpt-5.5)
-          //   - Images → providerSupportsBuiltinImageGeneration
-          //              (OpenAI cloud Responses-API models)
+          // External models have no local tool runtime → supportsTools false.
+          // The supportsBuiltin* flags carry server-side capability per pill:
+          // Search, Code (Claude 4.x + gpt-5.5), Images (OpenAI cloud
+          // Responses-API).
           supportsTools: false,
           supportsBuiltinWebSearch,
           supportsBuiltinCodeExecution,
@@ -1813,8 +1784,8 @@ export function ChatPage(): ReactElement {
   }, [currentProjectId, navigate, search]);
 
   const exitCompare = useCallback(() => {
-    // Prefer the explicit save; fall back to the last non-compare view so the
-    // composer + menu path also returns to where the user started.
+    // Prefer the explicit save; fall back to the last non-compare view so
+    // the composer + menu path also returns where the user started.
     const saved = viewBeforeCompareRef.current ?? lastNonCompareViewRef.current;
     // No saved view (compare opened by direct URL); fall back to a fresh chat.
     if (!saved) {
@@ -1823,9 +1794,8 @@ export function ChatPage(): ReactElement {
     }
     viewBeforeCompareRef.current = null;
     navigate({ to: "/chat", search: saved });
-    // Restore usage from the last assistant message, but only if it
-    // matches the currently active checkpoint. Without this guard the
-    // relaxed render gate would show stale stats from another model.
+    // Restore usage from the last assistant message, only if it matches the
+    // active checkpoint, else the relaxed render gate shows stale stats.
     const threadId =
       saved.thread ?? useChatRuntimeStore.getState().activeThreadId;
     if (threadId) {
@@ -1845,7 +1815,7 @@ export function ChatPage(): ReactElement {
           const usageModelId =
             (usage as { modelId?: unknown }).modelId;
           // Scope by modelId when present; reject if no active checkpoint
-          // (model-scoped usage cannot be attributed to "nothing").
+          // (model-scoped usage can't be attributed to "nothing").
           if (typeof usageModelId === "string" && usageModelId) {
             if (!activeCheckpoint || usageModelId !== activeCheckpoint) {
               return;
@@ -1895,15 +1865,13 @@ export function ChatPage(): ReactElement {
         .flatMap((provider) =>
           provider.models.map((model) => {
             // For OpenRouter's free router we know which underlying free
-            // model the gateway actually picked once a stream completes
-            // (chat-adapter latches `chunk.model` into the runtime store).
-            // Render the chip as `openrouter:<short-chosen>` — drop the
-            // redundant `/free` from the router id and the org prefix
-            // from the chosen id (e.g.
-            //   openrouter/free + inclusionai/ring-2.6-1t-20260508:free
-            //     -> openrouter:ring-2.6-1t-20260508:free
-            // ). The `:free` suffix on the chosen id already conveys
-            // 'free model', so the leading `/free` is noise.
+            // model the gateway picked once a stream completes (chat-adapter
+            // latches `chunk.model`). Render the chip as
+            // `openrouter:<short-chosen>`, dropping the redundant `/free`
+            // and the chosen id's org prefix (e.g. openrouter/free +
+            // inclusionai/ring-2.6-1t-20260508:free ->
+            // openrouter:ring-2.6-1t-20260508:free). The `:free` suffix
+            // already conveys "free model".
             let displayName = model;
             if (
               provider.providerType === "openrouter" &&
@@ -2117,8 +2085,8 @@ export function ChatPage(): ReactElement {
       <GuidedTour {...tour.tourProps} />
       <div className="relative flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-hidden">
         <NativeModelDropOverlay state={nativeModelDropState} />
-        {/* Bottom fade under the top bar so messages dissolve as they scroll
-            beneath it (Gemini / unsloth-sidebar style), instead of a hard cut. */}
+        {/* Fade under the top bar so messages dissolve as they scroll
+            beneath it, instead of a hard cut. */}
         {view.mode !== "compare" && (
           <div
             aria-hidden
diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
index c2e4eaab79..42f8e56ac5 100644
--- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
@@ -326,9 +326,8 @@ export function ChatProvidersSettings({
       return;
     }
     // Seed default_models only for curated providers (catalog too large to
-    // enumerate — defaults are the suggestion shortlist). Remote-mode cloud
-    // providers and local OpenAI-compat presets stay empty until the user
-    // clicks "Load available models".
+    // enumerate). Remote cloud providers and local OpenAI-compat presets stay
+    // empty until the user clicks "Load available models".
     const seedDefaults = entry.model_list_mode === "curated";
     setAvailableModels(seedDefaults ? [...entry.default_models] : []);
     setSelectedModelIds([]);
@@ -425,10 +424,9 @@ export function ChatProvidersSettings({
               updatedAt,
             };
           });
-        // Trust the backend response when it succeeds. An empty array means
-        // every connection was removed (often from another browser/tab) and
-        // the local cache should mirror that, otherwise the stale entries
-        // become un-removable in this browser until localStorage is cleared.
+        // Trust the backend response. An empty array means every connection was
+        // removed (often from another tab); mirror that locally, else stale
+        // entries become un-removable here until localStorage is cleared.
         onProvidersChange(syncedProviders);
       } catch (error) {
         // Only surface a toast for real failures, not for the silent
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index bddcb02c69..3b192927d6 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -127,19 +127,11 @@ export function InfoHint({ children }: { children: ReactNode }) {
 }
 
 /**
- * Editable numeric value display.
- *
- * Renders as a single <input> that *looks* like text by default —
- * transparent background, no border, no ring — and only shows a faint
- * surface tint on hover/focus to signal editability. When unfocused,
- * the input shows the formatted display string (`displayValue ?? value`,
- * so labels like "Off" / "Max" still render); on focus, it switches to
- * the raw numeric value, selects it, and accepts free text input.
- * Commit happens on blur or Enter; Escape reverts. The clamp-to-range
- * happens on commit so users can type intermediate values without the
- * input fighting them mid-keystroke. Single component shared by every
- * slider value and the Context Length input so the click-to-edit
- * affordance is consistent across the panel.
+ * Editable numeric value display, shared by every slider value and the Context
+ * Length input. An <input> that looks like text (shows `displayValue ?? value`,
+ * so "Off"/"Max" labels render) until focus, when it swaps to the raw number,
+ * selects it, and accepts free text. Commits on blur/Enter, reverts on Escape.
+ * Clamping happens on commit so typing intermediate values isn't fought.
  */
 function snapToStep(
   value: number,
@@ -205,7 +197,7 @@ function NumericValueInput({
         cancelBlurCommitRef.current = false;
         setDraft(String(value));
         setFocused(true);
-        // Defer the select() so it runs after the value swap above.
+        // Defer select() so it runs after the value swap above.
         const target = e.currentTarget;
         requestAnimationFrame(() => target.select());
       }}
@@ -331,10 +323,9 @@ function CollapsibleSection({
 }: {
   label: string;
   /**
-   * When set, the label text becomes an external link (e.g. to the feature's
-   * GitHub PR) instead of part of the collapse toggle. The chevron still
-   * toggles open/close, so we render the two as siblings rather than nesting
-   * an <a> inside the <button> (invalid HTML).
+   * When set, the label becomes an external link (e.g. the feature's GitHub PR)
+   * instead of part of the toggle. The chevron still toggles, so link and button
+   * are siblings rather than an <a> nested in a <button> (invalid HTML).
    */
   labelHref?: string;
   children?: ReactNode;
@@ -412,17 +403,16 @@ interface ChatSettingsPanelProps {
   onParamsChange: (params: InferenceParams) => void;
   isExternalModel?: boolean;
   /**
-   * Sampling-param capability set for the active external provider, or `null`
-   * for local models (in which case every knob is rendered). Drives the
-   * per-param visibility in the sampling section.
+   * Sampling-param capabilities for the active external provider, or `null` for
+   * local models (every knob rendered). Drives per-param sampling visibility.
    */
   providerCapabilities?: ProviderCapabilities | null;
   activeExternalProvider?: ExternalProviderConfig | null;
   onExternalProviderChange?: (provider: ExternalProviderConfig) => void;
   /**
    * Backend provider type for the active external model (e.g. "kimi",
-   * "anthropic", "openai"), or `null` for local models. Drives the
-   * per-provider Max Tokens floor in the slider.
+   * "anthropic", "openai"), or `null` for local models. Drives the per-provider
+   * Max Tokens floor in the slider.
    */
   externalProviderType?: string | null;
   onReloadModel?: () => void;
@@ -440,9 +430,8 @@ export function ChatSettingsPanel({
   externalProviderType = null,
   onReloadModel,
 }: ChatSettingsPanelProps) {
-  // For non-external (local) models we show every knob — providerCapabilities
-  // is only consulted when `isExternalModel` is true. An external model with an
-  // unknown provider falls back to the OpenAI-compat shape via
+  // Local models show every knob; providerCapabilities is only consulted when
+  // isExternalModel. Unknown providers fall back to the OpenAI-compat shape via
   // getProviderCapabilities, so these flags never undercount support.
   const showTemperature =
     !isExternalModel || Boolean(providerCapabilities?.temperature);
@@ -724,8 +713,7 @@ export function ChatSettingsPanel({
   const settingsContent = (
     <>
       <div className="flex h-full min-h-0 flex-col">
-      {/* Header sits outside the scroll area so the scrollbar never shifts the
-          close button. */}
+      {/* Header is outside the scroll area so the scrollbar never shifts the close button. */}
       <div className="flex h-[48px] shrink-0 items-start gap-2 bg-panel-surface pl-[18px] pr-[16px] pt-[11px]">
         {isMobile ? (
           <span className="flex h-[34px] flex-1 items-center text-[16px] font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
diff --git a/studio/frontend/src/features/chat/components/context-usage-bar.tsx b/studio/frontend/src/features/chat/components/context-usage-bar.tsx
index e15a3e0b4f..16dee2a624 100644
--- a/studio/frontend/src/features/chat/components/context-usage-bar.tsx
+++ b/studio/frontend/src/features/chat/components/context-usage-bar.tsx
@@ -28,7 +28,7 @@ function getSeverityColor(percent: number): {
 
 export const ContextUsageBar: FC<{
   used: number;
-  // null on external providers (no known window); bar then hides the ratio.
+  // null on external providers (unknown window); bar hides the ratio.
   total?: number | null;
   cached?: number;
   // Anthropic-only (billed at the write premium).
diff --git a/studio/frontend/src/features/chat/components/model-load-status.tsx b/studio/frontend/src/features/chat/components/model-load-status.tsx
index 75d92f4581..fd78866ece 100644
--- a/studio/frontend/src/features/chat/components/model-load-status.tsx
+++ b/studio/frontend/src/features/chat/components/model-load-status.tsx
@@ -17,15 +17,10 @@ function clampProgress(value: number): number {
 }
 
 /**
- * Split a composed progress label like
- *   "22.8 of 122.3 GB • 330.1 MB/s • 5m 9s left"
- * into a primary chunk ("22.8 of 122.3 GB") that can sit next to the
- * percent, and a secondary chunk ("330.1 MB/s • 5m 9s left") that can
- * live on its own row. This keeps either line from overflowing into a
- * ragged wrap when the rate/ETA part shows up mid-download.
- *
- * Labels without " • " (e.g. "22.8 GB downloaded") are returned
- * primary-only, so the secondary row simply doesn't render.
+ * Split a composed progress label like "22.8 of 122.3 GB • 330.1 MB/s • 5m 9s
+ * left" into a primary chunk (next to the percent) and a secondary chunk (its
+ * own row), so neither line wraps raggedly once rate/ETA appears mid-download.
+ * Labels without " • " return primary-only, so the secondary row doesn't render.
  */
 function splitProgressLabel(
   label: string | null | undefined,
@@ -46,8 +41,7 @@ export function ModelLoadDescription({
   progressLabel,
 }: ModelLoadDescriptionProps) {
   const hasProgress = typeof progressPercent === "number";
-  // Split once at the top of the render so the JSX below stays flat --
-  // no IIFE required. splitProgressLabel is a trivial string op.
+  // Split once at the top so the JSX below stays flat (no IIFE).
   const { primary: labelPrimary, secondary: labelSecondary } =
     splitProgressLabel(progressLabel);
 
@@ -116,9 +110,8 @@ export function ModelLoadInlineStatus({
             className="flex shrink-0 items-center gap-1 text-[10px] font-medium tracking-[0.08em] text-muted-foreground/80"
             title={progressLabel ?? undefined}
           >
-            {/* Inline layout is horizontal and tight -- show only the
-                primary (bytes) chunk; the full label (with rate/ETA)
-                stays available via the tooltip. */}
+            {/* Tight inline layout: show only the primary (bytes) chunk;
+                full label (rate/ETA) stays in the tooltip. */}
             <span>{splitProgressLabel(progressLabel).primary}</span>
             <span className="tabular-nums">
               {Math.round(clampProgress(progressPercent))}%
diff --git a/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx b/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
index 29e5260de2..7dfa03672d 100644
--- a/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
+++ b/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
@@ -2,29 +2,21 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Settings-sheet section for OpenAI shell-tool container management.
- * Renders only when:
- *   - active provider is OpenAI cloud (api.openai.com base URL), AND
- *   - the active model is gpt-5.5 or gpt-5.5-pro (the only families
- *     where the shell tool is wired through today).
+ * Settings-sheet section for OpenAI shell-tool container management. Renders
+ * only when the active provider is OpenAI cloud (api.openai.com) and the model
+ * is gpt-5.5 / gpt-5.5-pro (the only families wired to the shell tool today).
  *
- * Surfaces three controls:
- *   1. Default container idle-timeout (minutes). Persists on the
- *      provider record; pre-fills the create dialog and is used by
- *      the chat-adapter's lazy-create path on the first turn of a
- *      thread.
- *   2. Container picker for the *active thread* — pick any of the
- *      user's existing OpenAI containers, or "Auto-create per thread"
- *      (default; lets the auto-create path manage it).
- *   3. Create-new-container inline form. Refresh + delete actions
- *      per row.
+ * Controls:
+ *   1. Default container idle-timeout (minutes). Persisted on the provider
+ *      record; pre-fills the create dialog and feeds the chat-adapter's
+ *      lazy-create path on a thread's first turn.
+ *   2. Container picker for the active thread (an existing OpenAI container or
+ *      auto-create per thread).
+ *   3. Create-new-container inline form, with per-row refresh + delete.
  *
- * State persistence:
- *   - TTL → ExternalProviderConfig.openaiContainerTtlMinutes
- *   - Active container for this thread → ThreadRecord.openaiCodeExecContainerId
- *
- * No new global stores — list is fetched on open / refresh and held
- * in component state.
+ * Persistence: TTL -> ExternalProviderConfig.openaiContainerTtlMinutes;
+ * active container -> ThreadRecord.openaiCodeExecContainerId. No global stores;
+ * the list is fetched on open / refresh and held in component state.
  */
 
 "use client";
@@ -63,10 +55,9 @@ import {
 const DEFAULT_TTL_MINUTES = 20;
 const TTL_MIN = 1;
 const TTL_MAX = 20; // OpenAI hard cap on expires_after.minutes
-// Cadence for re-fetching the container list while the section is
-// mounted. OpenAI's container TTL flips at minute granularity, so 30s
-// is fast enough that an expired container loses its ACTIVE pill within
-// half a minute without hammering /v1/containers.
+// Re-fetch cadence while the section is mounted. OpenAI's TTL flips at minute
+// granularity, so 30s drops an expired container's ACTIVE pill within half a
+// minute without hammering /v1/containers.
 const REFRESH_POLL_MS = 30_000;
 
 function ageLabel(epochSeconds: number | null | undefined): string {
@@ -89,10 +80,9 @@ function shortContainerId(id: string): string {
 }
 
 function isContainerRunning(c: OpenAIContainerSummary): boolean {
-  // OpenAI's containers API reports `status: "running"` while idle TTL is
-  // valid and `status: "expired"` once the idle window has passed. Treat
-  // a missing status as running so we don't false-positive on any older
-  // payloads that didn't include the field.
+  // OpenAI reports `status: "running"` while idle TTL is valid and "expired"
+  // afterward. Treat a missing status as running so older payloads without the
+  // field don't false-positive.
   return c.status == null || c.status === "running";
 }
 
@@ -114,32 +104,27 @@ export function OpenAICodeExecSection({
   const [creating, setCreating] = useState(false);
   const [createOpen, setCreateOpen] = useState(false);
   const [createName, setCreateName] = useState("");
-  // Ids that have been deleted in this session. Once tombstoned, an id
-  // stays hidden from the picker for the lifetime of the page — OpenAI's
-  // /containers list can keep returning a freshly-deleted id for an
-  // undocumented and variable amount of time, and an automatic re-show
-  // creates more confusion than it solves. Refreshing the page resets
-  // the tombstone naturally.
+  // Ids deleted this session. A tombstoned id stays hidden for the page's
+  // lifetime: OpenAI's /containers list can keep returning a freshly-deleted
+  // id for an undocumented, variable time, and auto-reshowing it confuses more
+  // than it helps. A page refresh resets the tombstone.
   const [tombstones, setTombstones] = useState<Set<string>>(() => new Set());
-  // Ids optimistically inserted after a successful create but not yet
-  // confirmed by a /v1/containers list response. OpenAI's list endpoint
-  // is eventually consistent — a freshly-created container can be absent
-  // for several seconds. We render the row immediately with a "Creating"
-  // pill, then drop it from this set once a refresh sees the id.
+  // Ids optimistically inserted after a create but not yet confirmed by a
+  // /v1/containers list. That endpoint is eventually consistent (a new
+  // container can be absent for several seconds), so we render the row at once
+  // with a "Creating" pill and drop it once a refresh sees the id.
   const [pendingIds, setPendingIds] = useState<Set<string>>(() => new Set());
-  // Ref mirror so `refresh()` can read the current pending set without
-  // re-binding when it changes (the callback is in a useEffect dep).
+  // Ref mirror so `refresh()` reads the current pending set without re-binding
+  // when it changes (the callback is a useEffect dep).
   const pendingIdsRef = useRef<Set<string>>(pendingIds);
   useEffect(() => {
     pendingIdsRef.current = pendingIds;
   }, [pendingIds]);
-  // One-shot follow-up refresh scheduled after a create, to catch the
-  // common case where the server list lags the create response by a few
-  // seconds. Tracked so we can clear it on unmount.
+  // One-shot follow-up refresh after a create, for when the server list lags
+  // the create response by a few seconds. Tracked so we clear it on unmount.
   const pendingRetryRef = useRef<number | null>(null);
-  // Target row for the destructive confirmation dialog. Held in state
-  // (rather than blocking with window.confirm) so the dialog sits inside
-  // the settings sheet instead of a native browser alert.
+  // Target row for the delete-confirmation dialog. Held in state (not
+  // window.confirm) so the dialog sits inside the settings sheet.
   const [pendingDelete, setPendingDelete] =
     useState<OpenAIContainerSummary | null>(null);
   const [deleting, setDeleting] = useState(false);
@@ -174,15 +159,15 @@ export function OpenAICodeExecSection({
   }, [activeThreadId]);
 
   // Hide just-deleted containers even if OpenAI's list still returns them.
-  // This is the single chokepoint — every downstream view (sorted picker,
-  // auto-bind candidate, all-containers list) derives from visibleContainers.
+  // Single chokepoint: every downstream view (sorted picker, auto-bind
+  // candidate, all-containers list) derives from visibleContainers.
   const visibleContainers = useMemo(() => {
     if (tombstones.size === 0) return containers;
     return containers.filter((c) => !tombstones.has(c.id));
   }, [containers, tombstones]);
 
-  // Containers sorted newest-first by lastActiveAt so the dropdown's
-  // default (auto-bind target) shows up first.
+  // Newest-first by lastActiveAt so the dropdown default (auto-bind target)
+  // shows up first.
   const sortedContainers = useMemo(
     () =>
       [...visibleContainers].sort(
@@ -191,19 +176,18 @@ export function OpenAICodeExecSection({
     [visibleContainers],
   );
 
-  // First running container by lastActiveAt — the auto-bind target and
-  // also what we surface visually before Dexie catches up.
+  // First running container by lastActiveAt: the auto-bind target and what we
+  // surface visually before Dexie catches up.
   const firstRunningContainer = useMemo(
     () => sortedContainers.find(isContainerRunning) ?? null,
     [sortedContainers],
   );
 
-  // What the picker should treat as "active" right now. We decouple
-  // this from `activeContainerId` (Dexie state) so the user immediately
-  // sees the most-recent running container while the auto-bind effect's
-  // async write propagates. If the Dexie-bound container has since
-  // expired, fall back to the first running candidate — the stale-bind
-  // sweeper below will clear Dexie shortly after.
+  // What the picker treats as "active" now. Decoupled from `activeContainerId`
+  // (Dexie state) so the user sees the most-recent running container while the
+  // auto-bind effect's async write propagates. If the Dexie-bound container
+  // expired, fall back to the first running candidate; the stale-bind sweeper
+  // clears Dexie shortly after.
   const boundContainer = useMemo(
     () => sortedContainers.find((c) => c.id === activeContainerId) ?? null,
     [sortedContainers, activeContainerId],
@@ -223,8 +207,8 @@ export function OpenAICodeExecSection({
       });
       const serverIds = new Set(list.map((c) => c.id));
       setContainers((prev) => {
-        // Preserve optimistic inserts the server hasn't acknowledged
-        // yet so they don't disappear on the reconciling refresh.
+        // Preserve optimistic inserts the server hasn't acknowledged yet so
+        // they don't disappear on the reconciling refresh.
         const orphans = prev.filter(
           (c) => !serverIds.has(c.id) && pendingIdsRef.current.has(c.id),
         );
@@ -248,11 +232,9 @@ export function OpenAICodeExecSection({
     }
   }, [apiKey, provider.baseUrl]);
 
-  // Fetch once when the section mounts (or provider changes), then
-  // poll on a low cadence so an expired container's ACTIVE pill clears
-  // without the user clicking the refresh button. Also re-fetch when
-  // the tab regains visibility — covers the common case of leaving the
-  // sheet open across a long idle period.
+  // Fetch on mount (or provider change), then poll on a low cadence so an
+  // expired container's ACTIVE pill clears without a manual refresh. Also
+  // re-fetch when the tab regains visibility (sheet left open while idle).
   useEffect(() => {
     void refresh();
     const interval = window.setInterval(() => {
@@ -276,21 +258,17 @@ export function OpenAICodeExecSection({
     };
   }, [refresh]);
 
-  // Auto-bind the active thread to the most-recently-active container
-  // whenever the thread has none set and at least one container exists
-  // on the user's OpenAI account. Sorting by `lastActiveAt` matches
-  // what feels "most recent" from the user's perspective.
+  // Auto-bind the active thread to the most-recently-active container when the
+  // thread has none and at least one container exists on the account. Sorting
+  // by `lastActiveAt` matches what feels "most recent" to the user.
   //
-  // We eagerly materialize the thread row via `ensureThreadRecord` so
-  // the bind lands before the user has sent a first
-  // message. This does NOT create anything at OpenAI — only a local
-  // ThreadRecord — so it does not bypass the user's expectation that
-  // a fresh OpenAI container is not created until first send.
+  // `ensureThreadRecord` eagerly materializes the thread row so the bind lands
+  // before the first message. This creates nothing at OpenAI (only a local
+  // ThreadRecord), so a fresh OpenAI container is still not created until first
+  // send.
   //
-  // If `containers` is empty (no OpenAI containers exist yet), this
-  // effect short-circuits: the picker renders an empty-state hint and
-  // the chat-adapter's lazy-create path will mint the first container
-  // on first send.
+  // If no containers exist yet, this short-circuits: the picker shows an
+  // empty-state hint and the chat-adapter mints the first container on send.
   useEffect(() => {
     if (
       !activeThreadId ||
@@ -330,14 +308,12 @@ export function OpenAICodeExecSection({
 
   const onPick = async (value: string) => {
     if (!activeThreadId || !value) return;
-    // value is always a container id now — the "Auto-create per thread"
-    // option has been removed in favour of always defaulting to the
-    // most-recently-active container. The chat-adapter still handles
-    // the no-containers-exist case (lazy-create on first send).
+    // value is always a container id now; "Auto-create per thread" was removed
+    // in favour of defaulting to the most-recently-active container. The
+    // chat-adapter still handles the no-containers case (lazy-create on send).
     //
-    // ensureThreadRecord materializes the thread row eagerly (modelType
-    // "base" — settings sheet is single-thread-mode only) so the update
-    // actually lands when the user hasn't sent a message yet.
+    // ensureThreadRecord eagerly materializes the thread row (modelType "base":
+    // settings sheet is single-thread only) so the update lands before send.
     try {
       await ensureThreadRecord({ threadId: activeThreadId, modelType: "base" });
       const updated = await updateStoredChatThread(activeThreadId, {
@@ -360,9 +336,9 @@ export function OpenAICodeExecSection({
       toast.error("Container name is required");
       return;
     }
-    // TTL inherits from the section-level "Idle timeout" control —
-    // there is no per-container override on the form. Read it at
-    // submit time so a last-second change to the TTL row applies.
+    // TTL inherits from the section-level "Idle timeout" control (no
+    // per-container override). Read at submit time so a last-second TTL change
+    // applies.
     const ttlMinutes =
       provider.openaiContainerTtlMinutes ?? DEFAULT_TTL_MINUTES;
     setCreating(true);
@@ -374,10 +350,9 @@ export function OpenAICodeExecSection({
       toast.success(`Created container ${name}`);
       setCreateName("");
       setCreateOpen(false);
-      // Optimistic insert + "Creating" pill. OpenAI's /v1/containers
-      // list endpoint is eventually consistent and can omit the new
-      // container for several seconds — without this, the row only
-      // shows up on the next 30s poll or a manual refresh.
+      // Optimistic insert + "Creating" pill. The /v1/containers list is
+      // eventually consistent and can omit the new container for seconds;
+      // without this the row only appears on the next poll or manual refresh.
       setContainers((prev) =>
         prev.some((c) => c.id === created.id) ? prev : [created, ...prev],
       );
@@ -387,9 +362,8 @@ export function OpenAICodeExecSection({
         next.add(created.id);
         return next;
       });
-      // Follow-up refresh ~5s later to reconcile the optimistic row
-      // with the server's view once /v1/containers catches up. One
-      // shot; the regular poll covers any longer tail.
+      // Follow-up refresh ~5s later to reconcile the optimistic row once
+      // /v1/containers catches up. One shot; the poll covers a longer tail.
       if (pendingRetryRef.current != null) {
         window.clearTimeout(pendingRetryRef.current);
       }
@@ -397,9 +371,8 @@ export function OpenAICodeExecSection({
         pendingRetryRef.current = null;
         void refresh();
       }, 5000);
-      // Auto-bind the just-created container to the active thread.
-      // ensureThreadRecord first so the bind lands even when the user
-      // creates a container before sending the first message.
+      // Auto-bind the new container to the active thread. ensureThreadRecord
+      // first so the bind lands even if no message has been sent yet.
       if (activeThreadId) {
         try {
           await ensureThreadRecord({
@@ -420,8 +393,7 @@ export function OpenAICodeExecSection({
     } finally {
       setCreating(false);
       // Refresh even on failure: the request may have partially succeeded
-      // server-side (created container, lost response), and a re-fetch
-      // keeps the picker in sync with OpenAI's actual state.
+      // (container created, response lost); a re-fetch keeps the picker in sync.
       await refresh();
     }
   };
@@ -435,8 +407,8 @@ export function OpenAICodeExecSection({
         { apiKey, baseUrl: provider.baseUrl || null },
         id,
       );
-      // Tombstone the id so the picker hides it immediately even if
-      // OpenAI's list keeps returning it for a while.
+      // Tombstone the id so the picker hides it at once even if OpenAI's list
+      // keeps returning it for a while.
       setTombstones((prev) => {
         if (prev.has(id)) return prev;
         const next = new Set(prev);
@@ -460,9 +432,8 @@ export function OpenAICodeExecSection({
     } finally {
       setDeleting(false);
       setPendingDelete(null);
-      // Always refresh so a stale list entry (e.g. container deleted
-      // elsewhere, or already expired) is purged from the UI even when
-      // the delete call itself errored.
+      // Always refresh so a stale list entry (deleted elsewhere or expired) is
+      // purged even when the delete call errored.
       await refresh();
     }
   };
@@ -496,9 +467,8 @@ export function OpenAICodeExecSection({
         />
       </div>
 
-      {/* Single container list. Clicking a row binds it to the active
-          thread and the ACTIVE pill marks which one — no separate
-          picker needed. */}
+      {/* Container list. Clicking a row binds it to the active thread; the
+          ACTIVE pill marks which one (no separate picker). */}
       <div className="flex flex-col gap-1.5">
         <div className="flex items-center justify-between gap-2">
           <span className="text-[11px] uppercase tracking-wider text-muted-foreground">
@@ -518,10 +488,9 @@ export function OpenAICodeExecSection({
           </Button>
         </div>
         {sortedContainers.length === 0 ? (
-          // Quiet placeholder with the same muted border as row cards
-          // so an empty section doesn't masquerade as an active control.
-          // The first container is minted by the chat-adapter on first
-          // send (lazy-create) and appears here after the next refresh.
+          // Quiet placeholder with the same muted border as row cards so an
+          // empty section doesn't look like an active control. The first
+          // container is lazy-created on first send and appears after refresh.
           <div className="flex h-9 w-full items-center rounded-md border border-dashed border-border/60 bg-muted/20 px-2 text-xs text-muted-foreground">
             None yet - one will be created on first send.
           </div>
@@ -566,9 +535,8 @@ export function OpenAICodeExecSection({
                         : undefined
                   }
                 >
-                  {/* min-w-0 + truncate keeps long OpenAI container ids
-                      from spilling under the trash button on narrow
-                      settings sheets. */}
+                  {/* min-w-0 + truncate keeps long container ids from
+                      spilling under the trash button on narrow sheets. */}
                   <div className="flex min-w-0 flex-1 flex-col gap-0.5">
                     <div className="flex min-w-0 items-center gap-1.5">
                       <span className="min-w-0 truncate font-medium">
@@ -619,10 +587,9 @@ export function OpenAICodeExecSection({
         )}
       </div>
 
-      {/* Create new — inline single-row edit that visually echoes a
-          container card. TTL is inherited from the section's top
-          "Idle timeout" control (no per-container override), which
-          keeps the form light and avoids a duplicated input. */}
+      {/* Create new: inline single-row edit echoing a container card. TTL is
+          inherited from the top "Idle timeout" control (no per-container
+          override), keeping the form light. */}
       {createOpen ? (
         <div className="flex items-center gap-1 rounded-md border border-border/60 bg-muted/20 px-1.5 py-1">
           <Input
diff --git a/studio/frontend/src/features/chat/db.ts b/studio/frontend/src/features/chat/db.ts
index 2aa19ecd9f..928d1722e0 100644
--- a/studio/frontend/src/features/chat/db.ts
+++ b/studio/frontend/src/features/chat/db.ts
@@ -45,9 +45,8 @@ export { db };
 /**
  * Wraps Dexie liveQuery for React state updates.
  *
- * Important: include every semantic query input in `deps` (filters, sort keys,
- * IDs, etc). `querier` identity is intentionally ignored to avoid re-subscribing
- * on every render when callers pass inline functions.
+ * Include every semantic query input in `deps` (filters, sort keys, IDs). `querier`
+ * identity is ignored to avoid re-subscribing every render on inline functions.
  */
 export function useLiveQuery<T>(
   querier: () => Promise<T>,
diff --git a/studio/frontend/src/features/chat/external-providers.ts b/studio/frontend/src/features/chat/external-providers.ts
index b10253c280..300d384cdf 100644
--- a/studio/frontend/src/features/chat/external-providers.ts
+++ b/studio/frontend/src/features/chat/external-providers.ts
@@ -17,33 +17,29 @@ export interface ExternalProviderConfig {
   /** Whether to ask supported hosted providers to use prompt caching. */
   enablePromptCaching?: boolean;
   /**
-   * Anthropic prompt-cache TTL bucket. Only meaningful when
-   * `enablePromptCaching` is true and the provider supports the choice
-   * (Anthropic today). Maps to `prompt_cache_ttl` on the backend, which
-   * attaches `cache_control.ttl` to the cache marker. Omitted = inherit
-   * Anthropic's default 5-minute pool, same as before this knob existed.
+   * Anthropic prompt-cache TTL bucket. Only meaningful when `enablePromptCaching`
+   * is true and the provider supports the choice (Anthropic today). Maps to
+   * backend `prompt_cache_ttl`, which sets `cache_control.ttl` on the cache
+   * marker. Omitted = inherit Anthropic's default 5-minute pool.
    */
   promptCacheTtl?: "5m" | "1h";
   /** User-pinned: the loaded vLLM model supports `enable_thinking`. */
   isReasoningModel?: boolean;
   /**
-   * Default idle-timeout (in minutes) for newly created OpenAI shell
-   * containers. Pre-fills the "Create container" dialog and is the
-   * TTL the auto-create-per-thread path POSTs to /v1/containers with.
-   * OpenAI's hard default is 20. Only meaningful for OpenAI cloud.
+   * Default idle-timeout (minutes) for new OpenAI shell containers. Pre-fills the
+   * "Create container" dialog and is the TTL the auto-create-per-thread path POSTs
+   * to /v1/containers. OpenAI's hard default is 20. Only for OpenAI cloud.
    */
   openaiContainerTtlMinutes?: number;
   createdAt: number;
   updatedAt: number;
 }
 
-// Gemini supports prompt caching, but the wire flow requires a
-// separate POST to /v1beta/cachedContents to create the cache before
-// the generateContent call can reference it; the boolean Studio
-// currently emits on enable_prompt_caching is not enough on its own.
-// Until that two-step orchestration ships we keep the picker off so
-// the toggle does not silently no-op for Gemini users. See
-// https://ai.google.dev/gemini-api/docs/caching.
+// Gemini supports prompt caching, but the wire flow needs a separate POST to
+// /v1beta/cachedContents before generateContent can reference the cache; the
+// enable_prompt_caching boolean alone isn't enough. Until that two-step flow
+// ships, keep the picker off so the toggle doesn't silently no-op for Gemini.
+// See https://ai.google.dev/gemini-api/docs/caching.
 const PROMPT_CACHING_PROVIDER_TYPES = new Set(["openai", "anthropic"]);
 
 export function supportsProviderPromptCaching(
@@ -53,10 +49,9 @@ export function supportsProviderPromptCaching(
 }
 
 /**
- * Whether the provider lets the user choose between a short and a long
- * prompt-cache pool. Anthropic exposes both a 5m and a 1h ephemeral
- * pool via `cache_control.ttl`; OpenAI's automatic prompt cache has no
- * equivalent user-selectable knob, so it stays off the picker.
+ * Whether the provider lets the user choose between a short and long prompt-cache
+ * pool. Anthropic exposes 5m and 1h ephemeral pools via `cache_control.ttl`;
+ * OpenAI's automatic cache has no equivalent knob, so it stays off the picker.
  */
 const PROMPT_CACHE_TTL_PROVIDER_TYPES = new Set(["anthropic"]);
 
@@ -74,8 +69,8 @@ export function isPromptCacheTtl(value: unknown): value is "5m" | "1h" {
   return typeof value === "string" && PROMPT_CACHE_TTL_VALUES.has(value as "5m" | "1h");
 }
 
-// Provider types that expose the connection-level "reasoning model"
-// toggle. vLLM's OpenAI-compat endpoint doesn't advertise this per model.
+// Provider types exposing the connection-level "reasoning model" toggle.
+// vLLM's OpenAI-compat endpoint doesn't advertise this per model.
 const REASONING_TOGGLE_PROVIDER_TYPES = new Set(["vllm"]);
 
 export function supportsProviderReasoningToggle(
@@ -245,10 +240,9 @@ export function toExternalBackendProviderType(
   providerType: string | null | undefined,
 ): string | undefined {
   if (!providerType) return undefined;
-  // vLLM's /v1/responses applies the loaded model's chat template, which
-  // 400s on strict-alternation templates (e.g. Gemma 3). Pass the actual
-  // type through so the backend routes vLLM to /v1/chat/completions instead
-  // of the OpenAI Responses path used for gpt-5.x.
+  // vLLM's /v1/responses applies the loaded model's chat template, which 400s on
+  // strict-alternation templates (e.g. Gemma 3). Pass the type through so the
+  // backend routes vLLM to /v1/chat/completions instead of the Responses path.
   if (providerType === "vllm") return "vllm";
   if (providerType === "ollama") return "ollama";
   if (providerType === "llama_cpp") return "llama_cpp";
@@ -419,10 +413,8 @@ export function loadExternalProviders(): ExternalProviderConfig[] {
   }
 }
 
-/**
- * Load the raw (encrypted or legacy plaintext) key map from localStorage.
- * Values are opaque strings — either AES-GCM ciphertext or legacy plaintext.
- */
+/** Load the raw key map from localStorage. Values are opaque strings: either
+ * AES-GCM ciphertext or legacy plaintext. */
 function loadRawKeyMap(): Record<string, string> {
   if (!canUseStorage()) return {};
   try {
@@ -457,7 +449,7 @@ export function saveExternalProviders(
   if (!canUseStorage()) return;
   try {
     localStorage.setItem(EXTERNAL_PROVIDERS_KEY, JSON.stringify(providers));
-    // Prune keys for removed providers — works on raw ciphertext, no decryption needed
+    // Prune keys for removed providers (works on raw ciphertext, no decryption)
     const allowedIds = new Set(providers.map((provider) => provider.id));
     const keys = loadRawKeyMap();
     const pruned: Record<string, string> = {};
@@ -472,10 +464,7 @@ export function saveExternalProviders(
   }
 }
 
-/**
- * Retrieve a provider API key from localStorage.
- * Returns "" if no key is stored.
- */
+/** Retrieve a provider API key from localStorage; "" if none stored. */
 export function getExternalProviderApiKey(
   providerId: string,
 ): string {
@@ -483,9 +472,7 @@ export function getExternalProviderApiKey(
   return keys[providerId] ?? "";
 }
 
-/**
- * Store a provider API key in localStorage.
- */
+/** Store a provider API key in localStorage. */
 export function setExternalProviderApiKey(
   providerId: string,
   apiKey: string,
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index d1dab924cb..a152cbb358 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -167,11 +167,9 @@ function getTrustRemoteCodeRequiredMessage(modelName: string): string {
   return `${modelName} needs custom code enabled to load. Turn on "Enable custom code" in Chat Settings, then try again.`;
 }
 
-// Canonicalises any value the backend reports (or persisted state holds)
-// onto the five UI-facing modes the Speculative Decoding dropdown
-// understands: "auto" / "mtp" / "ngram" / "mtp+ngram" / "off" / null.
-// Mirrors backend _canonicalize_spec_mode so old persisted "default" /
-// "draft-mtp" / "ngram-mod" / chain values round-trip cleanly.
+// Canonicalises any backend/persisted value onto the Speculative Decoding
+// dropdown's modes ("auto"/"mtp"/"ngram"/"mtp+ngram"/"off"/null). Mirrors
+// backend _canonicalize_spec_mode so legacy persisted values round-trip.
 function normalizeSpeculativeType(v: string | null | undefined): string | null {
   if (v == null) return null;
   const s = String(v).trim().toLowerCase();
@@ -276,9 +274,8 @@ export function useChatModelRuntime() {
         listLoras(),
       ]);
 
-      // Cancellation can land while the requests above are in flight (e.g. the
-      // user cancels a load during this refresh). Bail before writing any
-      // backend state back into the store -- cancelLoading already cleared it.
+      // Cancellation can land while the requests above are in flight. Bail
+      // before writing backend state back -- cancelLoading already cleared it.
       if (signal?.aborted) return;
 
       setModels(listRes.models.map(toChatModelSummary));
@@ -331,12 +328,10 @@ export function useChatModelRuntime() {
         const currentSpecType = normalizeSpeculativeType(
           statusRes.speculative_type,
         );
-        // Refresh runs both on F5 (fresh store needs hydration) AND right
-        // after a fresh load (store was already set by the load path). For
-        // the user-configurable model params we only hydrate when the shadow
-        // `loaded*` field is still null -- that signals "not yet hydrated".
-        // Otherwise we'd clobber the values the load path just applied and
-        // the UI would appear to revert the user's changes.
+        // Refresh runs on F5 (needs hydration) and right after a load (store
+        // already set). For user-configurable params, only hydrate when the
+        // shadow `loaded*` field is null ("not yet hydrated"); otherwise we'd
+        // clobber what the load path just applied and revert the user.
         const prevState = useChatRuntimeStore.getState();
         const clampedReasoningEffort = clampLocalReasoningEffort(
           prevState.reasoningEffort,
@@ -355,15 +350,12 @@ export function useChatModelRuntime() {
           supportsPreserveThinking,
           supportsTools,
           // Reset per-turn reasoning flag so:
-          //   1. models that do not support reasoning do not inherit a stale
-          //      off state from a prior model, and
-          //   2. local reasoning-effort models (where the composer hides
-          //      the Off option via supportsReasoningOff=false) cannot end
-          //      up with reasoningEnabled=false carried over from an
-          //      external model where Off was selected — the composer would
-          //      keep showing "Think: <level>" via effectiveReasoningEnabled,
-          //      but the chat-adapter would omit the kwarg and the Harmony
-          //      template would fall back to its own default effort.
+          //   1. non-reasoning models don't inherit a stale off state, and
+          //   2. local reasoning-effort models (Off hidden via
+          //      supportsReasoningOff=false) don't carry reasoningEnabled=false
+          //      from an external model where Off was selected -- the composer
+          //      would still show "Think: <level>" but the adapter would omit
+          //      the kwarg, so Harmony falls back to its default effort.
           reasoningEnabled: supportsReasoning
             ? reasoningStyle === "reasoning_effort"
               ? true
@@ -600,15 +592,12 @@ export function useChatModelRuntime() {
             }
             if (abortCtrl.signal.aborted) throw new Error("Cancelled");
 
-            // Reset Speculative Decoding to Auto whenever the user
-            // switches to a different model. Spec strategy is a
-            // per-model decision: a sub-3B non-MTP GGUF that ran with
-            // "Off" should not carry that choice into a 27B MTP GGUF
-            // where Auto would auto-promote to draft-mtp. The user can
-            // still pick a forced mode on the new model; this just
-            // clears the stale prior-model choice so the backend's
-            // platform-aware path runs by default. Same applies to
-            // spec_draft_n_max which is MTP-only.
+            // Reset Speculative Decoding to Auto on model switch: spec
+            // strategy is per-model, so a sub-3B non-MTP GGUF's "Off" must
+            // not carry into a 27B MTP GGUF where Auto auto-promotes to
+            // draft-mtp. Clears the stale prior choice so the backend's
+            // platform-aware path runs by default; same for spec_draft_n_max
+            // (MTP-only). The user can still force a mode on the new model.
             if (currentCheckpoint && currentCheckpoint !== modelId) {
               useChatRuntimeStore.setState({
                 speculativeType: null,
@@ -852,17 +841,15 @@ export function useChatModelRuntime() {
         );
         loadToastIdRef.current = toastId;
 
-        // Poll download progress for non-cached models (GGUF and non-GGUF).
-        // Then, once the download wraps (or for already-cached models),
-        // poll the llama-server mmap phase so "Starting model..." no
-        // longer looks frozen for several minutes on large MoE models.
+        // Poll download progress for non-cached models, then (after download
+        // or for cached models) poll the llama-server mmap phase so "Starting
+        // model..." doesn't look frozen for minutes on large MoE models.
         let progressInterval: ReturnType<typeof setInterval> | null = null;
         const expectedBytes =
           typeof selection !== "string" ? selection.expectedBytes ?? 0 : 0;
 
-        // Rolling window of byte samples for rate / ETA estimation.
-        // Shared across download + mmap phases so the estimator doesn't
-        // reset when the phase flips.
+        // Rolling window of byte samples for rate/ETA estimation, shared
+        // across download + mmap phases so it survives phase flips.
         type Sample = { t: number; b: number };
         const MIN_SAMPLES = 3;
         const MIN_WINDOW = 3_000; // ms
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts b/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
index f8e1535e04..dc9226fb8e 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
@@ -95,8 +95,8 @@ async function buildIndex(): Promise<ChatSearchItem[]> {
   );
 
   // Legacy-only chats can exist before server-side history import finishes.
-  // Fill just the missing ids from the legacy-aware path instead of issuing
-  // one request per thread up front.
+  // Fill only the missing ids via the legacy path instead of one request per
+  // thread up front.
   const missingThreadIds = allThreadIds.filter(
     (threadId) => !messagesByThread.has(threadId),
   );
@@ -163,8 +163,7 @@ export function useChatSearchIndex(enabled: boolean): {
       setLoading(true);
       buildIndex()
         .then((result) => {
-          // Drop out-of-order responses so a slower rebuild can't clobber
-          // a fresher one.
+          // Drop out-of-order responses so a slower rebuild can't clobber a fresher one.
           if (cancelled || seq !== requestSeqRef.current) return;
           setItems(result);
         })
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
index 54aa5d82dc..c2607d6b7b 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
@@ -61,9 +61,8 @@ export function groupThreads(threads: ThreadRecord[]): SidebarItem[] {
   return items.sort((a, b) => b.createdAt - a.createdAt);
 }
 
-// Streaming fires CHAT_HISTORY_UPDATED_EVENT per chunk. Debounce so
-// each quiet window produces at most one O(N) fetch; requestSeq
-// discards stale responses.
+// Streaming fires CHAT_HISTORY_UPDATED_EVENT per chunk. Debounce so each quiet
+// window produces at most one O(N) fetch; requestSeq discards stale responses.
 const SIDEBAR_REFRESH_DEBOUNCE_MS = 300;
 
 export function useChatSidebarItems(options?: {
diff --git a/studio/frontend/src/features/chat/hooks/use-pill-activation-order.ts b/studio/frontend/src/features/chat/hooks/use-pill-activation-order.ts
index e159101a7f..76987a86d3 100644
--- a/studio/frontend/src/features/chat/hooks/use-pill-activation-order.ts
+++ b/studio/frontend/src/features/chat/hooks/use-pill-activation-order.ts
@@ -3,9 +3,8 @@
 
 import { useEffect, useState } from "react";
 
-// Tracks which of the given keys are active and the order in which each became
-// active, so opt-in composer pills (Canvas, MCP) render in the order they were
-// toggled on rather than a fixed order.
+// Tracks which keys are active and their activation order, so opt-in
+// composer pills (Canvas, MCP) render in toggle-on order, not a fixed one.
 export function usePillActivationOrder(states: Record<string, boolean>): string[] {
   const [order, setOrder] = useState<string[]>(() =>
     Object.keys(states).filter((key) => states[key]),
diff --git a/studio/frontend/src/features/chat/hooks/use-transfer-stats.ts b/studio/frontend/src/features/chat/hooks/use-transfer-stats.ts
index 5681680842..b1f03354e3 100644
--- a/studio/frontend/src/features/chat/hooks/use-transfer-stats.ts
+++ b/studio/frontend/src/features/chat/hooks/use-transfer-stats.ts
@@ -2,19 +2,17 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Compute rate (bytes/sec) and ETA (seconds) from a time-series of
- * cumulative ``bytes`` values, using a rolling window of recent samples.
+ * Compute rate (bytes/sec) and ETA (seconds) from a time-series of cumulative
+ * ``bytes`` values, using a rolling window of recent samples.
  *
- * Shared between the chat-flow download toast, the training-start
- * overlay, and the model-load phase UI. All three have the same shape:
- * a counter that rises monotonically from 0 toward ``totalBytes``, polled
- * on an interval. The derived stats are identical regardless of whether
- * the bytes came from an HTTP download or an mmap page-in.
+ * Shared by the chat-flow download toast, training-start overlay, and
+ * model-load phase UI: all three are a counter rising monotonically from 0
+ * toward ``totalBytes``, polled on an interval, regardless of HTTP download or
+ * mmap page-in.
  *
- * Stability rule: ``stable`` stays ``false`` until we've observed at
- * least 3 samples spanning ≥3 seconds. That keeps the UI from flashing
- * wildly varying rates during the first tick or two when the denominator
- * is effectively zero.
+ * Stability: ``stable`` stays ``false`` until at least 3 samples spanning >=3s,
+ * so the UI doesn't flash wild rates during the first tick or two when the
+ * denominator is ~0.
  */
 
 import { useEffect, useRef, useState } from "react";
diff --git a/studio/frontend/src/features/chat/lib/friendly-names.ts b/studio/frontend/src/features/chat/lib/friendly-names.ts
index 503008f761..79744b3181 100644
--- a/studio/frontend/src/features/chat/lib/friendly-names.ts
+++ b/studio/frontend/src/features/chat/lib/friendly-names.ts
@@ -2,25 +2,17 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Friendly default names for auto-created OpenAI shell containers.
- * Used by the chat-adapter when the lazy-create path fires (Code pill
- * on, no thread container yet, user has set a non-default TTL). The
- * goal is a human-memorable label like "otter" or "harbor" instead of
- * "chat-abc12345" — the user can still rename via the Studio-side
- * alias map.
+ * Friendly default names for auto-created OpenAI shell containers, used by the
+ * chat-adapter's lazy-create path (Code pill on, no thread container, non-default
+ * TTL). Goal: a memorable label like "otter" instead of "chat-abc12345"; users
+ * can still rename via the Studio alias map.
  *
- * The list is curated to:
- *   - Be unambiguous, non-offensive nouns from natural categories
- *     (animals, plants, geography, materials, weather).
- *   - Avoid technical / political / brand words that might read as
- *     odd in a chat UI.
- *   - Stay reasonably small so the bundle cost is negligible (~200
- *     entries × ~7 bytes ≈ 1.5 KB).
+ * The list is curated to be unambiguous, non-offensive nouns from natural
+ * categories (animals, plants, geography, materials, weather), avoid
+ * technical/political/brand words, and stay small (~1.5 KB).
  *
- * Collisions are tolerated — the container's real unique key is its
- * ``cntr_*`` id, not its name. A short random hex suffix is appended
- * to make accidental same-name collisions visually distinct in the
- * picker list.
+ * Collisions are tolerated: the real unique key is the ``cntr_*`` id, not the
+ * name. A short random hex suffix keeps same-name picks visually distinct.
  */
 
 const WORDS = [
@@ -223,20 +215,15 @@ function randomHexSuffix(): string {
   ) {
     return crypto.randomUUID().replace(/-/g, "").slice(0, 4);
   }
-  // Older browser fallback. Math.random is fine here — this is a
-  // display suffix, not a security token.
+  // Older-browser fallback. Math.random is fine: display suffix, not a token.
   return Math.floor(Math.random() * 0xffff)
     .toString(16)
     .padStart(4, "0");
 }
 
 /**
- * Returns a single English-word name with a short random hex suffix.
- *
- * Example output: "kestrel-3f9c", "harbor-a012".
- *
- * The suffix keeps containers visually distinguishable in the picker
- * when the same word recurs across creations.
+ * Returns an English-word name with a short random hex suffix (e.g.
+ * "kestrel-3f9c"), so repeated words stay distinguishable in the picker.
  */
 export function pickFriendlyContainerName(): string {
   const word = WORDS[Math.floor(Math.random() * WORDS.length)] ?? "container";
diff --git a/studio/frontend/src/features/chat/mcp-composer-button.tsx b/studio/frontend/src/features/chat/mcp-composer-button.tsx
index 1e59960d07..8ff9385484 100644
--- a/studio/frontend/src/features/chat/mcp-composer-button.tsx
+++ b/studio/frontend/src/features/chat/mcp-composer-button.tsx
@@ -80,8 +80,8 @@ const MCP_PRESETS: readonly McpPreset[] = [
   },
 ] as const;
 
-// mcp_servers has no UNIQUE(url); dedupe by normalized URL so a preset
-// toggle reuses its row instead of creating duplicates.
+// mcp_servers has no UNIQUE(url); dedupe by normalized URL so a preset toggle
+// reuses its row instead of duplicating.
 function normalizeMcpUrl(url: string): string {
   return (url || "").trim().toLowerCase().replace(/\/+$/, "");
 }
@@ -110,8 +110,8 @@ export function McpComposerButton({
   const [pendingUrl, setPendingUrl] = useState<string | null>(null);
   const [hintKey, setHintKey] = useState<string | null>(null);
 
-  // Grey out only when a loaded model lacks tool support; with no model yet MCP
-  // can still be pre-selected, matching the other composer tools.
+  // Grey out only when a loaded model lacks tool support; with no model yet,
+  // MCP can still be pre-selected, like the other composer tools.
   const usable = !modelLoaded || supportsTools;
 
   const refresh = useCallback(async () => {
@@ -178,10 +178,9 @@ export function McpComposerButton({
     }
   }
 
-  // One dropdown row. Enabled rows get a green underlay and a tick that
-  // becomes an X on hover so a click removes them. A hint shows as a tooltip
-  // driven by row hover; the tooltip anchor is pointer-events-none so the whole
-  // row stays clickable (a Radix TooltipTrigger would swallow the select).
+  // One dropdown row. Enabled rows get a green underlay and a tick that becomes
+  // an X on hover (click removes). The hint tooltip anchor is pointer-events-none
+  // so the row stays clickable (a Radix TooltipTrigger would swallow the select).
   const renderRow = (opts: {
     key: string;
     label: string;
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index f550279826..07c953fecc 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -4,13 +4,10 @@
 /**
  * Per-provider sampling parameter capability matrix.
  *
- * Values are derived from each provider's published chat-completion docs as of
- * 2026-05. They describe which of our UI knobs map cleanly onto the provider's
- * request body; the panel hides params a provider does not accept so users
- * cannot dial a value that gets silently dropped or rejected.
- *
- * "Local" models (anything that is not an external provider) are represented by
- * a null capability — every knob renders for them.
+ * Derived from each provider's published chat-completion docs (2026-05). The
+ * panel hides params a provider does not accept so users cannot dial a value
+ * that gets silently dropped or rejected. "Local" models (non-external) use a
+ * null capability — every knob renders for them.
  */
 
 export interface ProviderCapabilities {
@@ -48,9 +45,8 @@ export type ExternalReasoningCapabilities = {
 };
 
 /**
- * Prefer a stored reasoning effort level that exists in ``effortLevels``,
- * mapping legacy "xhigh" to "max" when the model only exposes the latter
- * (Claude 4.6 adaptive thinking).
+ * Pick a stored reasoning effort level that exists in `effortLevels`, mapping
+ * legacy "xhigh" to "max" when only the latter is exposed (Claude 4.6).
  */
 export function clampReasoningEffortToLevels(
   preferred: ExternalReasoningCapabilities["reasoningEffortLevels"][number],
@@ -122,8 +118,8 @@ const EXTERNAL_MAX_OUTPUT_TOKENS_BY_MODEL: Array<{
 
 /**
  * Documented per-model output cap; unknown ids fall back to
- * `EXTERNAL_MAX_OUTPUT_TOKENS` (32k). OpenRouter ids are
- * `provider/model`; the prefix is stripped before matching.
+ * `EXTERNAL_MAX_OUTPUT_TOKENS` (32k). OpenRouter `provider/model` ids have the
+ * prefix stripped before matching.
  */
 export function getExternalMaxOutputTokens(
   providerType: string | null | undefined,
@@ -192,16 +188,14 @@ export function providerSupportsBuiltinWebSearch(
   modelId?: string | null | undefined,
   baseUrl?: string | null | undefined,
 ): boolean {
-  // Gemini ships grounded search via `tools: [{googleSearch: {}}]` on
-  // every chat-capable model. Most image-tier ids (`-image`,
-  // `nano-banana`) reject text-tool wiring because the
-  // responseModalities path is mutually exclusive with text tools, but
-  // Google explicitly documents Search grounding on the Gemini 3 image
-  // family (gemini-3-pro-image-preview, gemini-3.1-flash-image-preview,
-  // nano-banana-pro). Allow Search on those; hide on older image ids.
-  // Custom Gemini OpenAI-compat proxies (non-Google bases) skip the
-  // native translator on the backend, so native tool envelopes never
-  // reach them -- hide the pill there.
+  // Gemini ships grounded search via `tools: [{googleSearch: {}}]` on every
+  // chat-capable model. Most image-tier ids reject text-tool wiring (the
+  // responseModalities path excludes text tools), but Google documents Search
+  // grounding on the Gemini 3 image family (gemini-3-pro-image-preview,
+  // gemini-3.1-flash-image-preview, nano-banana-pro), so allow it there and
+  // hide on older
+  // image ids. Custom Gemini OpenAI-compat proxies skip the backend's native
+  // translator, so native tool envelopes never reach them -- hide the pill.
   if (providerType === "gemini") {
     if (isGeminiCustomOpenAICompatBase(baseUrl)) return false;
     const normalized = modelId?.trim().toLowerCase() ?? "";
@@ -247,8 +241,8 @@ export function providerSupportsFastMode(
 ): boolean {
   if (providerType !== "anthropic") return false;
   if (!modelId) return false;
-  // Family boundary ("" or "-") required so IDs like "claude-opus-4-70"
-  // / "claude-opus-4-7b" do not match.
+  // Family boundary ("" or "-") required so IDs like "claude-opus-4-70" or
+  // "claude-opus-4-7b" do not match.
   return ANTHROPIC_FAST_MODE_MODEL_PREFIXES.some(
     (prefix) => modelId === prefix || modelId.startsWith(`${prefix}-`),
   );
@@ -269,15 +263,13 @@ export function providerSupportsFastMode(
  *     thread's stored `openaiCodeExecContainerId`. Documented at
  *       https://developers.openai.com/api/docs/guides/tools-shell
  *
- * Returns false for every other provider. The backend additionally
- * gates the OpenAI shell tool on `is_openai_cloud` so custom
- * OpenAI-compat servers (ollama / llama.cpp / vLLM) that also report
- * `provider_type="openai"` never receive the tool — but in practice
- * none of those catalogs surface the `gpt-5.5` ids anyway, so the
- * frontend prefix match is enough.
+ * Returns false for every other provider. The backend also gates the OpenAI
+ * shell tool on `is_openai_cloud` so custom OpenAI-compat servers reporting
+ * `provider_type="openai"` never receive it; in practice those catalogs don't
+ * surface the `gpt-5.5` ids, so the frontend prefix match suffices.
  *
- * v1 wires the tools themselves; file uploads (Anthropic
- * `container_upload` / OpenAI `input_file`) are a deliberate follow-up.
+ * v1 wires the tools only; file uploads (Anthropic `container_upload` / OpenAI
+ * `input_file`) are a deliberate follow-up.
  */
 const ANTHROPIC_CODE_EXECUTION_MODEL_PREFIXES = [
   "claude-opus-4-7",
@@ -286,29 +278,26 @@ const ANTHROPIC_CODE_EXECUTION_MODEL_PREFIXES = [
   "claude-opus-4-5",
   "claude-sonnet-4-5",
   "claude-haiku-4-5",
-  // Deprecated upstream but the registry still exposes the ids, so the
-  // pill should remain functional for users on those snapshots.
+  // Deprecated upstream but the registry still exposes these ids; keep the
+  // pill working for users on those snapshots.
   "claude-opus-4-1",
   "claude-opus-4",
   "claude-sonnet-4",
 ] as const;
 
-// OpenAI cloud shell-tool gating. Docs only explicitly demonstrate
-// gpt-5.5; gpt-5.5-pro is included because the family share the same
-// /v1/responses contract. `gpt-5.5-pro` is checked first so the prefix
-// match doesn't collide with a hypothetical `gpt-5.5-turbo` etc.
+// OpenAI cloud shell-tool gating. Docs only show gpt-5.5; gpt-5.5-pro shares
+// the same /v1/responses contract. `gpt-5.5-pro` is checked first so the prefix
+// match doesn't collide with e.g. a hypothetical `gpt-5.5-turbo`.
 const OPENAI_CODE_EXECUTION_MODEL_PREFIXES = [
   "gpt-5.5-pro",
   "gpt-5.5",
 ] as const;
 
 /**
- * Strict check that a provider configuration points at OpenAI's
- * managed cloud (api.openai.com) or Azure OpenAI Foundry
- * (*.openai.azure.com), as opposed to a custom OpenAI-compat backend
- * (ollama / llama.cpp / vLLM / generic "custom" preset). The shell and
- * image-generation tools only exist on cloud backends; sending them to
- * anything else 400s the request. Mirror of the backend's
+ * Strict check that a provider config points at OpenAI managed cloud
+ * (api.openai.com) or Azure OpenAI Foundry (*.openai.azure.com), not a custom
+ * OpenAI-compat backend. The shell and image-generation tools exist only on
+ * cloud backends; sending them elsewhere 400s. Mirrors the backend's
  * `_is_openai_family_cloud` host check.
  */
 function isOpenAICloudBaseUrl(baseUrl: string | null | undefined): boolean {
@@ -357,16 +346,13 @@ export function providerSupportsBuiltinCodeExecution(
 }
 
 /**
- * Whether the selected external provider/model exposes OpenAI's
- * Responses-API server-side image_generation tool. Lit on for OpenAI
- * cloud (`api.openai.com`) when the picked model is a Responses-API
- * family id (gpt-5.x today). The backend additionally gates on
- * `is_openai_cloud`; mirror that here so the pill is hidden on custom
- * OpenAI-compat backends (ollama / llama.cpp / vLLM) that report
- * `provider_type="openai"` but would 400 on a `{type:"image_generation"}`
- * tool. See backend/core/inference/external_provider.py near line 2770
- * for the dispatch and backend/tests/test_openai_image_generation.py
- * for the round-trip coverage.
+ * Whether the selected provider/model exposes OpenAI's Responses-API
+ * image_generation tool. On for OpenAI cloud (`api.openai.com`) with a
+ * Responses-API family id (gpt-5.x). Mirrors the backend's `is_openai_cloud`
+ * gate so the pill hides on custom OpenAI-compat backends reporting
+ * `provider_type="openai"` that would 400 on a `{type:"image_generation"}` tool.
+ * See backend/core/inference/external_provider.py (~line 2770) for dispatch and
+ * backend/tests/test_openai_image_generation.py for round-trip coverage.
  */
 const OPENAI_IMAGE_GENERATION_MODEL_PREFIXES = [
   "gpt-5.5-pro",
@@ -394,14 +380,11 @@ export function providerSupportsBuiltinImageGeneration(
     );
   }
   if (providerType === "gemini") {
-    // Gemini's Nano Banana image-output ids carry either `-image` (e.g.
-    // `gemini-2.5-flash-image`, `gemini-3.1-flash-image-preview`) or the
-    // `nano-banana` alias (`nano-banana-pro-preview`). The backend flips
-    // generationConfig.responseModalities to ["TEXT", "IMAGE"] when one
-    // is picked, and translates inlineData parts into the same image_b64
-    // tool_end envelope the OpenAI path emits so the chat UI renders the
-    // picture inline. Custom Gemini OpenAI-compat proxies skip the
-    // native translator on the backend, so hide the image pill there.
+    // Gemini Nano Banana image-output ids carry `-image` or the `nano-banana`
+    // alias. The backend flips responseModalities to ["TEXT", "IMAGE"] and maps
+    // inlineData parts into the same image_b64 tool_end envelope as the OpenAI
+    // path so the UI renders inline. Custom Gemini OpenAI-compat proxies skip
+    // the native translator, so hide the image pill there.
     // See https://ai.google.dev/gemini-api/docs/image-generation.
     if (isGeminiCustomOpenAICompatBase(baseUrl)) return false;
     return normalized.includes("-image") || normalized.includes("nano-banana");
@@ -420,13 +403,11 @@ function isGeminiImageModel(modelId: string): boolean {
 }
 
 /**
- * Whether the saved Gemini connection points at a custom
- * OpenAI-compatible gateway (any non-Google host). The backend
- * `_is_openai_compatible` mirrors this to route those connections
- * through `/chat/completions` instead of the native translator, so
- * native Gemini tool envelopes (googleSearch, codeExecution,
- * responseModalities) never reach them. Hide the corresponding
- * Studio pills here so the request, builder, and UI agree.
+ * Whether the saved Gemini connection points at a custom OpenAI-compat gateway
+ * (any non-Google host). The backend `_is_openai_compatible` routes these
+ * through `/chat/completions` instead of the native translator, so native Gemini
+ * tool envelopes never reach them. Hide the matching Studio pills here so the
+ * request, builder, and UI agree.
  */
 export function isGeminiCustomOpenAICompatBase(
   baseUrl: string | null | undefined,
@@ -458,17 +439,13 @@ function geminiImageModelAllowsGoogleSearch(modelId: string): boolean {
 }
 
 /**
- * Per-provider minimum on the outbound max_tokens. Kimi's docs require
- * `max_tokens >= 16000` whenever a thinking model is in use so the
- * reasoning_content and final answer both fit in the budget — anything
- * lower truncates the response mid-stream. Other providers don't have a
- * documented floor, so they fall through to the generic min of 64 in
- * the slider.
+ * Per-provider minimum on the outbound max_tokens. Kimi requires
+ * `max_tokens >= 16000` for thinking models so reasoning_content and the answer
+ * both fit; lower truncates mid-stream. Others fall through to the generic 64.
  *
- * The chat-adapter resolves the effective floor on send and bumps the
- * outbound max_tokens up to this value if the user's stored maxTokens
- * sits below it. The settings panel reflects the same floor as the
- * slider min so the displayed value never drifts from what's sent.
+ * The chat-adapter resolves the floor on send and bumps maxTokens up to it if
+ * the stored value is below; the settings slider min reflects the same floor so
+ * the displayed value never drifts from what's sent.
  */
 const EXTERNAL_MIN_OUTPUT_TOKENS_BY_PROVIDER: Record<string, number> = {
   kimi: 16000,
@@ -528,11 +505,11 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     presencePenalty: false,
   },
   mistral: OPENAI_COMPAT_BASE,
-  // Gemini's native generationConfig accepts temperature, topP, topK and
-  // presencePenalty (plus a separate frequencyPenalty we do not surface
-  // today). minP and repetitionPenalty are not part of the contract --
-  // see https://ai.google.dev/api/rest/v1beta/GenerationConfig. Backend
-  // request shaping lives in _stream_gemini in
+  // Gemini's generationConfig accepts temperature, topP, topK and
+  // presencePenalty (plus a frequencyPenalty we don't surface). minP and
+  // repetitionPenalty aren't part of the contract --
+  // see https://ai.google.dev/api/rest/v1beta/GenerationConfig. Backend shaping
+  // lives in _stream_gemini in
   // studio/backend/core/inference/external_provider.py.
   gemini: {
     temperature: true,
@@ -542,11 +519,9 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     repetitionPenalty: false,
     presencePenalty: true,
   },
-  // Kimi k2.5/k2.6 are reasoning-class — the API locks temperature and
-  // top_p to fixed defaults and 400s on any other value:
-  //   "invalid temperature: only 1 is allowed for this model".
-  // Hide both sliders so the user is not offered knobs the model
-  // silently overrides. Backend additionally strips these fields via
+  // Kimi k2.5/k2.6 are reasoning-class — the API locks temperature and top_p to
+  // fixed defaults and 400s on any other value ("invalid temperature: only 1 is
+  // allowed for this model"). Hide both sliders. Backend also strips these via
   // PROVIDER_REGISTRY['kimi']['body_omit'].
   kimi: {
     temperature: false,
@@ -570,9 +545,9 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
   // OpenRouter silently drops params the target model does not support, so we
   // surface every knob and let the gateway handle the per-model fan-out.
   openrouter: ALL_SUPPORTED,
-  // Local OpenAI-compatible connections are proxied through the OpenAI backend
-  // path, but vLLM/Ollama/llama.cpp users often want top_k / min_p /
-  // repetition controls, so be permissive.
+  // Local OpenAI-compat connections go through the OpenAI backend path, but
+  // vLLM/Ollama/llama.cpp users often want top_k/min_p/repetition controls, so
+  // be permissive.
   custom: ALL_SUPPORTED,
   vllm: ALL_SUPPORTED,
   ollama: ALL_SUPPORTED,
@@ -733,11 +708,11 @@ function withReasoningEffortStyle(caps: ReasoningCaps): ExternalReasoningCapabil
 }
 
 function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCapabilities {
-  // Kimi exposes a boolean thinking toggle rather than an effort scale.
-  //   - kimi-k2.6:        thinking enabled by default, toggleable
-  //                       via extra_body: {thinking: {type: enabled|disabled}}
-  //   - kimi-k2-thinking: thinking always on, no off switch
-  //   - kimi-k2.5 (and anything else): no thinking
+  // Kimi exposes a boolean thinking toggle, not an effort scale.
+  //   - kimi-k2.6:        on by default, toggleable via
+  //                       extra_body: {thinking: {type: enabled|disabled}}
+  //   - kimi-k2-thinking: always on, no off switch
+  //   - kimi-k2.5 (and others): no thinking
   if (modelId === "kimi-k2-thinking") {
     return withEnableThinkingStyle({
       supportsReasoning: true,
@@ -754,16 +729,13 @@ function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCap
 }
 
 // Gemini's thinking ladder.
-//   - Gemini 3.x (3 / 3.1 / 3.5, Pro + Flash + Flash-Lite) and the
-//     gemini-pro-latest / gemini-flash-latest aliases use the new
-//     `thinkingConfig.thinkingLevel` string field (LOW/MEDIUM/HIGH/
-//     MINIMAL). Pro tier rejects MINIMAL.
-//   - Gemini 2.5 Flash + 2.5 Pro stay on the integer
-//     `thinkingConfig.thinkingBudget` (0=off on Flash, -1=dynamic,
-//     N>0=cap; Pro rejects 0).
-//   - 2.5 Flash-Lite: no native thinking surfaced; leave it off.
-//   - Image-tier ids (`*-image*`, `nano-banana-pro-preview`): image
-//     generation path -- no reasoning controls.
+//   - Gemini 3.x (Pro + Flash + Flash-Lite) and the gemini-pro-latest /
+//     gemini-flash-latest aliases use the string `thinkingConfig.thinkingLevel`
+//     (LOW/MEDIUM/HIGH/MINIMAL); Pro tier rejects MINIMAL.
+//   - Gemini 2.5 Flash + 2.5 Pro use the integer `thinkingConfig.thinkingBudget`
+//     (0=off on Flash, -1=dynamic, N>0=cap; Pro rejects 0).
+//   - 2.5 Flash-Lite: no native thinking surfaced; leave off.
+//   - Image-tier ids: image generation path -- no reasoning controls.
 const GEMINI3_PRO_PREFIXES = [
   "gemini-3.5-pro",
   "gemini-3.1-pro",
@@ -795,9 +767,8 @@ function resolveGeminiReasoningCapabilities(
     // Image generation; no thinking knob.
     return withEnableThinkingStyle();
   }
-  // Gemini 2.5 Flash-Lite supports `thinkingBudget` with `0` = off and
-  // a positive range starting at 512 (the backend maps "minimal" to
-  // that floor at external_provider._stream_gemini). Check this branch
+  // Gemini 2.5 Flash-Lite: `thinkingBudget` 0 = off, positive range from 512
+  // (backend maps "minimal" to that floor in _stream_gemini). Check this branch
   // BEFORE the broader `gemini-2.5-flash` prefix.
   // https://ai.google.dev/gemini-api/docs/thinking
   if (m.startsWith("gemini-2.5-flash-lite")) {
@@ -815,10 +786,10 @@ function resolveGeminiReasoningCapabilities(
     });
   }
   if (GEMINI3_PRO_PREFIXES.some((p) => m.startsWith(p))) {
-    // Gemini 3.x Pro: thinkingLevel supports low/medium/high per
+    // Gemini 3.x Pro: thinkingLevel low/medium/high; cannot fully disable, and
+    // "minimal" is rejected on Pro. Refs:
     // https://ai.google.dev/gemini-api/docs/thinking and
     // https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-1-pro.
-    // Cannot fully disable thinking; "minimal" is rejected on Pro.
     return withReasoningEffortStyle({
       supportsReasoning: true,
       supportsReasoningOff: false,
@@ -826,8 +797,8 @@ function resolveGeminiReasoningCapabilities(
     });
   }
   if (GEMINI3_FLASH_PREFIXES.some((p) => m.startsWith(p))) {
-    // Gemini 3 Flash: thinkingLevel minimal/low/medium/high. Minimal
-    // is the closest to "off" Google offers on Gemini 3.
+    // Gemini 3 Flash: thinkingLevel minimal/low/medium/high. Minimal is the
+    // closest to "off" Google offers on Gemini 3.
     return withReasoningEffortStyle({
       supportsReasoning: true,
       supportsReasoningOff: false,
@@ -840,9 +811,9 @@ function resolveGeminiReasoningCapabilities(
     });
   }
   if (GEMINI25_PRO_PREFIXES.some((p) => m.startsWith(p))) {
-    // Gemini 2.5 Pro: thinkingBudget cannot be 0 (API rejects with
-    // "only works in thinking mode"); backend coerces to a small
-    // positive budget. The picker still hides the off switch.
+    // Gemini 2.5 Pro: thinkingBudget cannot be 0 (API rejects "only works in
+    // thinking mode"); backend coerces to a small positive budget. Hide the off
+    // switch in the picker.
     return withReasoningEffortStyle({
       supportsReasoning: true,
       supportsReasoningOff: false,
@@ -907,9 +878,8 @@ function resolveConnectionLevelReasoning(
 }
 
 /**
- * resolve external-model thinking capabilities.
- * provider-specific matching lives in the OpenAI/Anthropic resolvers.
- * other providers default to no reasoning controls.
+ * Resolve external-model thinking capabilities. Provider-specific matching lives
+ * in the per-provider resolvers; others default to no reasoning controls.
  */
 export function getExternalReasoningCapabilities(
   providerType: string | null | undefined,
@@ -951,10 +921,9 @@ export function getExternalReasoningCapabilities(
   const isMistralProvider = normalizedProvider === "mistral";
   const isOpenRouterProvider = normalizedProvider === "openrouter";
   if (isOpenRouterProvider) {
-    // OpenRouter's unified `reasoning` parameter is accepted on every
-    // chat-completion request; the gateway silently no-ops for models
-    // that don't reason. Mandatory-reasoning ids are handled by the
-    // early guard above; everything else exposes a toggleable control.
+    // OpenRouter's unified `reasoning` param is accepted on every request; the
+    // gateway no-ops for non-reasoning models. Mandatory-reasoning ids are
+    // handled by the early guard; everything else gets a toggleable control.
     return {
       supportsReasoning: true,
       reasoningStyle: "enable_thinking",
@@ -966,10 +935,9 @@ export function getExternalReasoningCapabilities(
   if (isKimiProvider) return resolveKimiReasoningCapabilities(modelForMatching);
   if (isMistralProvider) return resolveMistralReasoningCapabilities(modelForMatching);
   if (normalizedProvider === "gemini") {
-    // Custom Gemini OAI-compat gateways (LiteLLM, proxies) route
-    // through /chat/completions which drops the Gemini-native
-    // thinkingConfig payload. Hide the native thinking ladder so the
-    // UI does not advertise a control the backend cannot honor.
+    // Custom Gemini OAI-compat gateways route through /chat/completions, which
+    // drops the native thinkingConfig payload. Hide the native thinking ladder
+    // so the UI doesn't advertise a control the backend can't honor.
     if (isGeminiCustomOpenAICompatBase(options?.baseUrl)) {
       return withEnableThinkingStyle();
     }
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 5e5780ec2a..8d9e5a10e3 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -194,9 +194,8 @@ class PDFAttachmentAdapter implements AttachmentAdapter {
 
 class TextAttachmentAdapter implements AttachmentAdapter {
   // MIME is unreliable for source files, so also match by extension
-  // (assistant-ui's fileMatchesAccept supports ".ext" entries). Covers
-  // svg, code, config and other plain-text formats; html keeps its own
-  // adapter below.
+  // (assistant-ui's fileMatchesAccept supports ".ext" entries). Covers svg, code,
+  // config and other plain-text formats; html keeps its own adapter below.
   accept = [
     "text/plain,text/markdown,text/csv,text/xml,text/json,text/css",
     "application/json,application/xml,image/svg+xml",
@@ -258,9 +257,7 @@ class HtmlAttachmentAdapter implements AttachmentAdapter {
 
   async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
     const html = await attachment.file.text();
-    // Strip HTML tags to extract readable text
     const doc = new DOMParser().parseFromString(html, "text/html");
-    // Remove script and style elements
     for (const el of doc.querySelectorAll("script, style")) el.remove();
     const text = (doc.body.textContent ?? "").replace(/\s+/g, " ").trim();
     return {
@@ -584,9 +581,9 @@ export async function ensureThreadRecord({
   try {
     await saveStoredChatThread(record);
   } catch (error) {
-    // assistant-ui can issue overlapping first-message persistence calls.
-    // If another call created the same thread while this one was waiting,
-    // treat initialization as successful and let the message write continue.
+    // assistant-ui can issue overlapping first-message persistence calls. If
+    // another call created the same thread while this one waited, treat init as
+    // successful and let the message write continue.
     const existingAfterRace = await listStoredChatThreads({
       includeArchived: true,
     }).catch(() => []);
@@ -866,7 +863,7 @@ function useStudioRuntimeAdapters(
           return a.id < b.id ? -1 : a.id > b.id ? 1 : 0;
         });
 
-        // Restore context usage from last assistant message if model matches
+        // Restore context usage from last assistant message if model matches.
         const lastAssistant = [...msgs]
           .reverse()
           .find((m) => m.role === "assistant");
@@ -882,14 +879,14 @@ function useStudioRuntimeAdapters(
             }
           | undefined;
         const store = useChatRuntimeStore.getState();
-        // Window check applies only when a local GGUF window is known;
-        // external providers have ggufContextLength === null.
+        // Window check applies only when a local GGUF window is known; external
+        // providers have ggufContextLength === null.
         const withinLocalLimit =
           !store.ggufContextLength ||
           (savedUsage?.totalTokens ?? 0) <= store.ggufContextLength;
-        // Legacy unscoped usage (no modelId) is only trusted when a
-        // known local window bounds the totals, so we can't misattribute
-        // an old local turn to a newly-selected external provider.
+        // Legacy unscoped usage (no modelId) is trusted only when a known local
+        // window bounds the totals, so an old local turn can't be misattributed
+        // to a newly-selected external provider.
         const modelMatches = savedUsage?.modelId
           ? savedUsage.modelId === store.params.checkpoint
           : typeof store.ggufContextLength === "number" &&
@@ -898,12 +895,10 @@ function useStudioRuntimeAdapters(
           store.setContextUsage(savedUsage);
         }
 
-        // If any message has a stored parentId, reconstruct the tree
-        // so retries/regenerations load as branches instead of being
-        // unrolled into a flat list.  For mixed legacy/new threads
-        // (old messages without parentId + new messages with), infer
-        // sequential parents for old messages to preserve the chain.
-        // Fall back to fromArray for fully legacy threads.
+        // If any message has a stored parentId, reconstruct the tree so
+        // retries/regenerations load as branches rather than a flat list. For
+        // mixed legacy/new threads, infer sequential parents for old messages to
+        // preserve the chain. Fall back to fromArray for fully legacy threads.
         const hasParentIds = msgs.some((m) => m.parentId != null);
         if (hasParentIds) {
           let previousId: string | null = null;
@@ -931,7 +926,7 @@ function useStudioRuntimeAdapters(
             return;
           }
           // Keep single-chat runtime state in sync once a new chat is first
-          // persisted. Compare panes intentionally do not write global activeThreadId.
+          // persisted. Compare panes intentionally don't write global activeThreadId.
           if (modelType === "base" && !pairId) {
             const store = useChatRuntimeStore.getState();
             if (store.activeThreadId !== remoteId) {
@@ -1071,8 +1066,8 @@ function ThreadNewChatSwitch({
     if (isLoading) {
       return;
     }
-    // Switch to a fresh local thread without persisting it yet.
-    // Persistence still happens on first message append.
+    // Switch to a fresh local thread without persisting it yet; persistence
+    // still happens on first message append.
     void aui.threads().switchToNewThread();
     useChatRuntimeStore.getState().setActiveThreadId(null);
   }, [aui, isLoading, nonce]);
@@ -1099,8 +1094,8 @@ function ActiveThreadSync({
 }
 
 // Exposes the current thread's cancelRun() via the shared store so external
-// surfaces (e.g. the sidebar trash button) can stop an in-flight stream
-// before deleting the thread — mirroring the Stop → Trash sequence.
+// surfaces (e.g. the sidebar trash button) can stop an in-flight stream before
+// deleting the thread, mirroring the Stop -> Trash sequence.
 function CancelRegistrar(): ReactElement | null {
   const aui = useAui();
   const mainThreadId = useAuiState(({ threads }) => threads.mainThreadId);
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index 1717152ba0..e54111862c 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -102,7 +102,7 @@ export interface CompareHandle {
 const IMAGE_ACCEPT = "image/jpeg,image/png,image/webp,image/gif";
 const MAX_IMAGE_SIZE = 20 * 1024 * 1024;
 
-// Inlined to avoid a new icon dependency. Kept in sync with the main composer.
+// Inlined to avoid a new icon dep. Kept in sync with the main composer.
 const ArrowDownStandardIcon: FC<{ className?: string }> = ({ className }) => (
   <svg
     className={className}
@@ -149,9 +149,9 @@ function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
 }
 
-// Mirrors the threshold in thread.tsx — see the comment there. Chrome on
-// Windows-over-WSL (issue #5546) never fires `compositionend` after the
-// IME commit, so the compose flag would otherwise stay true forever.
+// Mirrors the threshold in thread.tsx. Chrome on Windows-over-WSL (#5546)
+// never fires `compositionend` after IME commit, so the compose flag would
+// otherwise stay true forever.
 const IME_STUCK_TIMEOUT_MS = 2500;
 
 function fileToBase64DataURL(file: File): Promise<string> {
@@ -187,8 +187,8 @@ function formatReasoningDisabledLabel(
   modelId?: string,
 ): string {
   const normalized = modelId?.trim().toLowerCase() ?? "";
-  // Magistral keeps the "none" wire value, but UX should present this floor
-  // as "Medium" rather than a disabled state label.
+  // Magistral keeps the "none" wire value, but UX presents this floor as
+  // "Medium" rather than a disabled-state label.
   if (normalized.includes("magistral-medium-latest")) return "Medium";
   return supportsReasoningOff && isExternalOpenAIReasoning ? "None" : "Off";
 }
@@ -374,7 +374,7 @@ type CompareModelSelection = {
   ggufVariant?: string;
 };
 
-// Tool icon plus an X overlay the CSS reveals on hover when the pill is active.
+// Tool icon plus an X overlay CSS reveals on hover when the pill is active.
 function PillGlyph({ children }: { children: ReactNode }) {
   return (
     <span className="composer-pill-glyph">
@@ -396,8 +396,7 @@ export function SharedComposer({
   onExitCompare?: () => void;
 }): ReactElement {
   const navigate = useNavigate();
-  // Exit compare. Uses the parent's restore handler, or a fresh chat when
-  // compare was opened by direct URL.
+  // Exit compare: parent's restore handler, or fresh chat if opened by URL.
   const handleExitCompare = useCallback(() => {
     if (onExitCompare) {
       onExitCompare();
@@ -472,7 +471,7 @@ export function SharedComposer({
   const setMcpEnabledForChat = useChatRuntimeStore(
     (s) => s.setMcpEnabledForChat,
   );
-  // Three most recently updated projects for the quick-access submenu.
+  // Three most recently updated projects for the quick-access submenu
   const { projects } = useChatProjects();
   const recentProjects = [...projects]
     .sort((a, b) => b.updatedAt - a.updatedAt)
@@ -507,11 +506,10 @@ export function SharedComposer({
     modelLoaded,
   });
   const isCompareMode = Boolean(model1?.id || model2?.id);
-  // Attach-time gate. Compare mode defers to send: the catalog can lag
-  // behind a model's real capabilities (e.g., a GGUF whose mmproj
-  // arrives after the catalog snapshot), and we only sync the models[]
-  // entry after ensureModelLoaded runs at send time. Single mode uses
-  // the loaded model's runtime capability.
+  // Attach-time gate. Compare mode defers to send: the catalog can lag a
+  // model's real capabilities (e.g. a GGUF whose mmproj arrives after the
+  // snapshot), and models[] only syncs after ensureModelLoaded at send time.
+  // Single mode uses the loaded model's runtime capability.
   const attachUnavailableReason = isCompareMode ? null : imageUnavailableReason;
   const effectiveExternalModelId =
     selectedExternalProvider?.providerType === "openrouter" &&
@@ -547,11 +545,10 @@ export function SharedComposer({
   const reasoningLockedOn =
     effectiveSupportsReasoning &&
     (effectiveReasoningAlwaysOn || !effectiveSupportsReasoningOff);
-  // Kimi's $web_search builtin mandates thinking=disabled per the docs at
-  // https://platform.kimi.ai/docs/guide/use-web-search. Both pills stay
-  // clickable for Kimi, but turning one on flips the other off — the
-  // click handlers below enforce this mutual exclusion so the visible
-  // state always matches what the backend actually sends.
+  // Kimi's $web_search builtin mandates thinking=disabled
+  // (https://platform.kimi.ai/docs/guide/use-web-search). Both pills stay
+  // clickable, but turning one on flips the other off; the click handlers
+  // below enforce this so the visible state matches what the backend sends.
   const isKimiExternal = selectedExternalProvider?.providerType === "kimi";
   const effectiveReasoningEnabled = reasoningLockedOn ? true : reasoningEnabled;
   const effectiveReasoningVisualEnabled =
@@ -563,15 +560,13 @@ export function SharedComposer({
   const thinkingActiveLook = isEffort
     ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !reasoningDisabled)
     : reasoningLockedOn || (effectiveReasoningEnabled && !reasoningDisabled);
-  // Two-pill gating: Search pill lights up when the runtime has either
-  // a local tool runtime (supportsTools, gives us our Code/python + local
-  // web_search) OR a server-side web_search the provider runs for us
-  // (supportsBuiltinWebSearch, currently OpenAI / Anthropic / OpenRouter
-  // / Kimi). Code pill lights up on the local runtime OR when Anthropic
-  // is selected with a model that accepts the server-side
-  // code_execution_20250825 tool — see
-  // providerSupportsBuiltinCodeExecution. Anthropic is the only external
-  // provider that ships a code-execution tool today.
+  // Two-pill gating: Search lights up on a local tool runtime (supportsTools:
+  // Code/python + local web_search) OR a provider-run server-side web_search
+  // (supportsBuiltinWebSearch: OpenAI/Anthropic/OpenRouter/Kimi). Code lights
+  // up on the local runtime OR Anthropic with a model accepting the
+  // server-side code_execution_20250825 tool (see
+  // providerSupportsBuiltinCodeExecution). Anthropic is the only external
+  // provider shipping a code-execution tool today.
   const supportsBuiltinCodeExecution = providerSupportsBuiltinCodeExecution(
     selectedExternalProvider?.providerType,
     effectiveExternalModelId,
@@ -586,24 +581,22 @@ export function SharedComposer({
     selectedExternalProvider?.providerType,
   );
   // Gemini rejects codeExecution alongside image modalities. Search is
-  // blocked on older Gemini image ids but allowed on Gemini 3 image
-  // models -- supportsBuiltinWebSearch already encodes the per-model
-  // allowance, so we only disable Code unconditionally in Gemini
-  // image mode.
+  // blocked on older Gemini image ids but allowed on Gemini 3 image models
+  // (supportsBuiltinWebSearch encodes the per-model allowance), so we only
+  // disable Code unconditionally in Gemini image mode.
   const isExternalGemini = selectedExternalProvider?.providerType === "gemini";
   const imageDisabled = !modelLoaded || !supportsBuiltinImageGeneration;
   const imageModeDisablesCode =
     isExternalGemini && imageToolsEnabled && !imageDisabled;
   // Image-tier Gemini models always reject codeExecution and reject
-  // web_search on older ids (Gemini 3.x Pro/Flash allow it -- encoded
-  // in supportsBuiltinWebSearch). Don't let the local `supportsTools`
-  // runtime flag re-enable a pill the Gemini backend will silently
-  // drop. Detect "external provider is Gemini AND model is image-tier"
-  // and gate strictly on the provider builtin support.
+  // web_search on older ids (Gemini 3.x Pro/Flash allow it, encoded in
+  // supportsBuiltinWebSearch). Don't let local `supportsTools` re-enable a
+  // pill the Gemini backend silently drops: detect image-tier Gemini and
+  // gate strictly on provider builtin support.
   const isGeminiImageTier =
     isExternalGemini && supportsBuiltinImageGeneration;
   // Disable only when a loaded model lacks the capability; with no model the
-  // tool can still be pre-selected and reflected, matching the + menu.
+  // tool can still be pre-selected, matching the + menu.
   const searchDisabled =
     modelLoaded &&
     (isGeminiImageTier
@@ -615,14 +608,14 @@ export function SharedComposer({
         ? true
         : !(supportsTools || supportsBuiltinCodeExecution))) ||
     imageModeDisablesCode;
-  // Images pill is only ever lit on OpenAI cloud's Responses-API models
-  // and Gemini Nano Banana family. No local tool runtime fallback.
+  // Images pill lights only on OpenAI cloud Responses-API models and the
+  // Gemini Nano Banana family. No local tool runtime fallback.
   const showImagePill = supportsBuiltinImageGeneration;
   // Fetch pill: Anthropic-only (web_fetch_20250910 / web_fetch_20260209).
   const webFetchDisabled = !modelLoaded || !supportsBuiltinWebFetch;
   const showWebFetchPill = supportsBuiltinWebFetch;
-  // With more than 4 pills showing, collapse them to icons only to cut clutter.
-  // Compare, Search and Code always show; the rest are conditional.
+  // Above 4 pills, collapse to icons only to cut clutter. Compare, Search and
+  // Code always show; the rest are conditional.
   const pillsCompact =
     3 +
       (showImagePill ? 1 : 0) +
@@ -630,8 +623,8 @@ export function SharedComposer({
       (artifactsEnabled ? 1 : 0) +
       (mcpEnabledForChat ? 1 : 0) >
     4;
-  // Backwards-compatible alias for any other call site that may still
-  // reference `toolsDisabled` (rare; both pills used it before).
+  // Backwards-compatible alias for call sites still referencing
+  // `toolsDisabled` (rare; both pills used it before).
   const toolsDisabled = codeDisabled;
   const setPendingAudioStore = useChatRuntimeStore((s) => s.setPendingAudio);
   const clearPendingAudioStore = useChatRuntimeStore(
@@ -753,11 +746,11 @@ export function SharedComposer({
     const isGeneralizedCompare =
       hasCompareHandles && Boolean(model1?.id && model2?.id);
 
-    // Generalized compare requires both panes to have a model. A
-    // half-selected send either races to an empty bubble with bogus
-    // tok/s (#5569) or leaves the empty pane with a dangling prompt.
-    // hasCompareHandles is true only in GeneralCompareContent, so
-    // LoraCompare and single-pane chats are unaffected.
+    // Generalized compare requires both panes to have a model. A half-
+    // selected send either races to an empty bubble with bogus tok/s (#5569)
+    // or leaves the empty pane with a dangling prompt. hasCompareHandles is
+    // true only in GeneralCompareContent, so LoraCompare and single-pane
+    // chats are unaffected.
     if (hasCompareHandles && !isGeneralizedCompare) {
       toast.error("Pick a model in each pane to compare", {
         description:
@@ -771,10 +764,10 @@ export function SharedComposer({
       !isGeneralizedCompare &&
       imageUnavailableReason
     ) {
-      // Single mode: the loaded model's runtime capability is known
-      // here. Compare mode defers — each ensureModelLoaded below sets
-      // loadedIsMultimodal for its side, and the chat-adapter's
-      // pre-stream gate runs per-side against that fresh state.
+      // Single mode: the loaded model's runtime capability is known here.
+      // Compare mode defers: each ensureModelLoaded sets loadedIsMultimodal
+      // for its side, and the chat-adapter's pre-stream gate runs per-side
+      // against that fresh state.
       toast.error(imageUnavailableReason);
       return;
     }
@@ -869,10 +862,9 @@ export function SharedComposer({
           supportsTools: resp.supports_tools ?? false,
           loadedIsMultimodal: isMultimodalResponse(resp),
         });
-        // Sync the models[] entry with the load response so the
-        // attach/send gates read fresh capabilities. /api/models/list
-        // can lag behind a model's actual state (e.g., a GGUF whose
-        // mmproj was downloaded after the catalog snapshot).
+        // Sync the models[] entry with the load response so attach/send gates
+        // read fresh capabilities. /api/models/list can lag a model's actual
+        // state (e.g. a GGUF whose mmproj arrived after the snapshot).
         const currentModels = useChatRuntimeStore.getState().models;
         const idx = currentModels.findIndex((m) => m.id === sel.id);
         const synced = {
@@ -982,13 +974,12 @@ export function SharedComposer({
   const busy = running || comparing;
 
   function onKeyDown(e: KeyboardEvent) {
-    // IME composition (Japanese/Chinese/Korean): Enter commits the candidate.
-    // Don't hijack it. See issue #5318. Re-pin composingRef in case the stuck
-    // watchdog (#5546) cleared it during a long candidate-window pause; this
-    // keeps a follow-up click-Send from submitting preedit text. Re-arm the
-    // watchdog on the same path — without it the WSL+Chrome no-compositionend
-    // case would leave composingRef pinned forever after an IME keypress and
-    // re-lock Send.
+    // IME composition (JP/CN/KR): Enter commits the candidate, don't hijack it
+    // (#5318). Re-pin composingRef in case the stuck watchdog (#5546) cleared
+    // it during a long candidate-window pause, so a follow-up click-Send won't
+    // submit preedit text. Re-arm the watchdog on the same path; without it the
+    // WSL+Chrome no-compositionend case pins composingRef forever after an IME
+    // keypress and re-locks Send.
     if (e.nativeEvent.isComposing || e.keyCode === 229) {
       composingRef.current = true;
       refreshStuckImeTimer();
@@ -1019,8 +1010,8 @@ export function SharedComposer({
       }}
       onDragLeave={() => setDragging(false)}
       onDrop={(e) => {
-        // Phase 1 native model drops own Tauri local-path drops. Restore browser
-        // attachment drops in Tauri when Phase 1d adds attachment-token bridging.
+        // Phase 1 native model drops own Tauri local-path drops. Restore
+        // browser attachment drops in Tauri once Phase 1d adds token bridging.
         if (isTauri) return;
         e.preventDefault();
         setDragging(false);
@@ -1071,11 +1062,11 @@ export function SharedComposer({
         value={text}
         onChange={(e) => {
           // ALWAYS mirror the DOM value into React state, even during IME
-          // composition. The controlled `value` prop must match the DOM at
-          // all times, otherwise any unrelated parent re-render reconciles
-          // the textarea back to the stored value mid-composition — wiping
-          // the IME preedit AND prior committed text (e.g. Tab cycling
-          // candidates erases earlier words). Issue #5318.
+          // composition: the controlled `value` must match the DOM at all
+          // times, else an unrelated parent re-render reconciles the textarea
+          // back to the stored value mid-composition, wiping the IME preedit
+          // AND prior committed text (e.g. Tab-cycling candidates erases
+          // earlier words). #5318.
           setCompositionState(isNativeComposing(e.nativeEvent));
           setText(e.target.value);
         }}
@@ -1093,8 +1084,8 @@ export function SharedComposer({
         placeholder="Send to both models..."
         className="composer-input"
         rows={1}
-        // dir="auto" auto-detects RTL (Arabic / Hebrew / Persian / Urdu)
-        // from the first strong character; no effect on LTR scripts.
+        // dir="auto" detects RTL (Arabic/Hebrew/Persian/Urdu) from the first
+        // strong character; no effect on LTR scripts.
         dir="auto"
       />
       <div className="composer-action-wrapper">
@@ -1127,8 +1118,8 @@ export function SharedComposer({
             open={newProjectOpen}
             onOpenChange={setNewProjectOpen}
           />
-          {/* Same + side menu as the single-chat composer (ComposerToolsMenu),
-              wired to the compare composer's own file/audio inputs and tools. */}
+          {/* Same + menu as single-chat (ComposerToolsMenu), wired to the
+              compare composer's own file/audio inputs and tools. */}
           <DropdownMenu>
             <DropdownMenuTrigger asChild={true}>
               <button
@@ -1241,8 +1232,8 @@ export function SharedComposer({
                 {mcpEnabledForChat ? <CheckIcon className="ml-auto" /> : null}
               </DropdownMenuItem>
               {/* RAG hidden temporarily */}
-              {/* Always active: this menu only renders in compare mode.
-                  Ticked like Web search/Code; click toggles it off. */}
+              {/* Always active: this menu only renders in compare mode. Ticked
+                  like Web search/Code; click toggles it off. */}
               <DropdownMenuItem
                 className="text-primary font-medium"
                 onSelect={handleExitCompare}
@@ -1303,7 +1294,7 @@ export function SharedComposer({
               const next = !toolsEnabled;
               setToolsEnabled(next);
               // Kimi's $web_search builtin requires thinking=disabled
-              // (https://platform.kimi.ai/docs/guide/use-web-search). Toggle
+              // (https://platform.kimi.ai/docs/guide/use-web-search): toggle
               // the Think pill off when Search is on, mirroring the backend.
               if (isKimiExternal) {
                 setReasoningEnabled(!next, { persist: false });
@@ -1564,9 +1555,8 @@ export function SharedComposer({
                   const next = !reasoningEnabled;
                   setReasoningEnabled(next);
                   applyQwenThinkingParams(next);
-                  // Mutual exclusion: Kimi's $web_search builtin
-                  // requires thinking off, so turning thinking on flips
-                  // the Search pill off (and vice versa).
+                  // Mutual exclusion: Kimi's $web_search builtin requires
+                  // thinking off, so turning thinking on flips Search off.
                   if (isKimiExternal && next && toolsEnabled) {
                     setToolsEnabled(false, { persist: false });
                   }
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index ebcd1892d1..320cf4c962 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -37,12 +37,11 @@ export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
 
 // External provider selection is encoded into `params.checkpoint` as
-// `external::<providerId>::<modelId>`. PersistedChatSettings deliberately
-// Omits `checkpoint` because the local-model side is mirrored by the
-// backend's `/api/inference/status.active_model` response. External
-// selections have no such backend mirror, so without explicit
-// localStorage persistence here the user's external pick is silently
-// reset to the default on every page refresh.
+// `external::<providerId>::<modelId>`. PersistedChatSettings omits `checkpoint`
+// because the local-model side is mirrored by the backend's
+// /api/inference/status.active_model. External selections have no such mirror,
+// so without explicit localStorage persistence here the user's external pick
+// is reset to the default on every refresh.
 const LAST_EXTERNAL_CHECKPOINT_KEY = "unsloth_chat_last_external_checkpoint";
 
 function loadLastExternalCheckpoint(): string | null {
@@ -61,13 +60,13 @@ function saveLastExternalCheckpoint(value: string | null): void {
     if (value && isExternalModelId(value)) {
       window.localStorage.setItem(LAST_EXTERNAL_CHECKPOINT_KEY, value);
     } else {
-      // Clearing on a switch to a local / empty checkpoint means the
-      // next refresh won't override the now-active local selection.
+      // Clear on switch to a local/empty checkpoint so the next refresh
+      // won't override the now-active local selection.
       window.localStorage.removeItem(LAST_EXTERNAL_CHECKPOINT_KEY);
     }
   } catch {
-    // Storage quota / private-mode failures are non-fatal -- the
-    // selection just won't survive the refresh.
+    // Storage quota / private-mode failures are non-fatal; selection just
+    // won't survive the refresh.
   }
 }
 
@@ -103,10 +102,9 @@ function warnSettingsPersistenceFailure(): void {
   });
 }
 
-// Coalesce setting writes into one pendingPatch (deep merge for nested
-// keys), flush on a trailing-edge debounce, flush on beforeunload so a
-// pending patch survives tab close. Slider drag ticks now produce one
-// HTTP write per quiet window instead of one per tick.
+// Coalesce setting writes into one pendingPatch (deep merge for nested keys),
+// flush on a trailing-edge debounce and on beforeunload so a pending patch
+// survives tab close. Slider drags produce one HTTP write per quiet window.
 type SettingsPatch = Parameters<typeof savePersistedChatSettingsPatch>[0];
 
 const SETTINGS_DEBOUNCE_MS = 400;
@@ -156,9 +154,9 @@ function saveSettingsPatch(patch: SettingsPatch): void {
   }, SETTINGS_DEBOUNCE_MS);
 }
 
-// Best-effort flush of any pending patch when the tab closes. keepalive
-// lets the PUT outlive the unload; without it the browser cancels the
-// fetch and the user's last slider drag is dropped.
+// Best-effort flush of any pending patch on tab close. keepalive lets the PUT
+// outlive the unload; without it the browser cancels the fetch and the user's
+// last slider drag is dropped.
 if (typeof window !== "undefined") {
   window.addEventListener("beforeunload", () => {
     if (pendingTimer !== null) clearTimeout(pendingTimer);
@@ -191,9 +189,9 @@ export function loadOptionalBool(key: string): boolean | null {
 
 /**
  * Resolve the web-search / code-execution pill state to apply when a model
- * loads. Honors the user's persisted preference so loading a tool-capable
- * model never silently re-enables a pill the user turned off; falls back to
- * the model's capability only when no preference has been expressed.
+ * loads. Honors the user's persisted preference so a tool-capable model never
+ * re-enables a pill the user turned off; falls back to the model's capability
+ * only when no preference has been expressed.
  */
 export function resolveToolsEnabledOnLoad(supportsTools: boolean): {
   toolsEnabled: boolean;
@@ -255,12 +253,10 @@ type ChatRuntimeStore = {
   reasoningAlwaysOn: boolean;
   reasoningEnabled: boolean;
   /**
-   * The model id the OpenRouter router actually picked for the most recent
-   * stream when the active checkpoint is the openrouter/free meta-model.
-   * Updated each time a chunk arrives carrying a non-empty `model` field
-   * that differs from the requested id. Cleared when a non-OpenRouter
-   * model is selected. Used purely for UI display — appended after
-   * `openrouter/free:` in the active model chip.
+   * The model id the OpenRouter router picked for the most recent stream when
+   * the active checkpoint is the openrouter/free meta-model. Updated when a
+   * chunk's `model` field differs from the requested id; cleared on a
+   * non-OpenRouter model. UI display only (appended after `openrouter/free:`).
    */
   lastOpenRouterChosenModel: string | null;
   reasoningStyle: ReasoningStyle;
@@ -271,35 +267,29 @@ type ChatRuntimeStore = {
   preserveThinking: boolean;
   supportsTools: boolean;
   /**
-   * Whether the active external provider exposes a server-side
-   * web_search tool (OpenAI's /v1/responses today). Distinct from
-   * `supportsTools` — that flag governs the local tool runtime (Code,
-   * python sandbox, our DuckDuckGo web_search). This one only enables
-   * the chat composer's Search pill for external models. Local models
-   * keep `supportsTools` only.
+   * Whether the active external provider exposes a server-side web_search tool
+   * (OpenAI's /v1/responses today). Distinct from `supportsTools` (the local
+   * tool runtime): this only enables the composer's Search pill for external
+   * models. Local models keep `supportsTools` only.
    */
   supportsBuiltinWebSearch: boolean;
   /**
-   * Whether the active external provider exposes a server-side
-   * code-execution tool (Anthropic's `code_execution_20250825` on the
-   * Claude 4.x family). Distinct from `supportsTools` for the same
-   * reason as `supportsBuiltinWebSearch`: external providers don't
-   * give us a local tool runtime, but Anthropic dispatches code
-   * execution server-side. Read by both composers' Code pill gate.
+   * Whether the active external provider exposes a server-side code-execution
+   * tool (Anthropic's `code_execution_20250825` on Claude 4.x). Distinct from
+   * `supportsTools` like supportsBuiltinWebSearch: Anthropic dispatches it
+   * server-side. Read by both composers' Code pill gate.
    */
   supportsBuiltinCodeExecution: boolean;
   /**
-   * Whether the active external provider exposes a server-side
-   * image-generation tool (OpenAI's Responses-API `image_generation`
-   * today). Gates the chat composer's Images pill. Local models never
-   * receive the tool because their runtime cannot dispatch it.
+   * Whether the active external provider exposes a server-side image-generation
+   * tool (OpenAI's Responses-API `image_generation`). Gates the composer's
+   * Images pill. Local models never receive it (their runtime can't dispatch it).
    */
   supportsBuiltinImageGeneration: boolean;
   /**
-   * Whether the active external provider exposes a server-side
-   * web_fetch tool (Anthropic's `web_fetch_20250910` /
-   * `web_fetch_20260209`). Gates the composer's Fetch pill,
-   * independent of Search.
+   * Whether the active external provider exposes a server-side web_fetch tool
+   * (Anthropic's `web_fetch_20250910` / `web_fetch_20260209`). Gates the
+   * composer's Fetch pill, independent of Search.
    */
   supportsBuiltinWebFetch: boolean;
   toolsEnabled: boolean;
@@ -596,10 +586,9 @@ function setScalarSettingVersion<K extends ScalarSettingKey>(
 
 export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   settingsHydrated: false,
-  // Hydrate the last external checkpoint into params.checkpoint so the
-  // external picker selection survives a page refresh. Local model
-  // checkpoints are re-derived from the backend in useChatModelRuntime
-  // and intentionally NOT persisted here.
+  // Hydrate the last external checkpoint so the external picker survives a
+  // refresh. Local checkpoints are re-derived from the backend in
+  // useChatModelRuntime and intentionally NOT persisted here.
   params: (() => {
     const persistedExternal = loadLastExternalCheckpoint();
     return persistedExternal
@@ -699,9 +688,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
           return nextState;
         });
       } catch {
-        // Hydrate failed: treat as hydrated-with-defaults so future
-        // setParams calls reach saveSettingsPatch (which surfaces its
-        // own toast on real network failure).
+        // Hydrate failed: treat as hydrated-with-defaults so future setParams
+        // calls reach saveSettingsPatch (which toasts on real network failure).
         warnSettingsPersistenceFailure();
         set({ settingsHydrated: true });
       } finally {
@@ -715,16 +703,15 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     set({ modelRequiresTrustRemoteCode }),
   setParams: (params) =>
     set((state) => {
-      // Bump version unconditionally so a late hydration response
-      // won't clobber a pre-hydrate user edit; only the HTTP write
-      // is gated on settingsHydrated.
+      // Bump version unconditionally so a late hydration response won't clobber
+      // a pre-hydrate user edit; only the HTTP write is gated on settingsHydrated.
       const changedParams = getChangedInferenceParams(params, state.params);
       if (state.settingsHydrated && hasKeys(changedParams)) {
         saveSettingsPatch({ inferenceParams: changedParams });
       }
-      // Mirror setCheckpoint: the local model load path can mutate
-      // params.checkpoint via setParams() before setCheckpoint runs,
-      // leaving stale per-turn counters under the new checkpoint.
+      // Mirror setCheckpoint: the local load path can mutate params.checkpoint
+      // via setParams() before setCheckpoint runs, leaving stale per-turn
+      // counters under the new checkpoint.
       const checkpointChanged = state.params.checkpoint !== params.checkpoint;
       return {
         params,
@@ -787,19 +774,17 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   setModelsError: (modelsError) => set({ modelsError }),
   setCheckpoint: (modelId, ggufVariant) =>
     set((state) => {
-      // Persist external selections so they survive a page refresh.
-      // Local model ids are NOT persisted here -- they get re-derived
-      // from the backend's `/api/inference/status.active_model` on
-      // mount, and a stale persisted local id would race against the
-      // freshly-loaded model. See LAST_EXTERNAL_CHECKPOINT_KEY notes.
+      // Persist external selections so they survive a refresh. Local ids are
+      // NOT persisted -- they're re-derived from the backend on mount, and a
+      // stale persisted local id would race the freshly-loaded model. See
+      // LAST_EXTERNAL_CHECKPOINT_KEY notes.
       saveLastExternalCheckpoint(isExternalModelId(modelId) ? modelId : null);
-      // Clear stale per-turn usage when the model changes; the relaxed
-      // external-provider render gate would otherwise show old counters
-      // until the next completion overwrites them.
+      // Clear stale per-turn usage on model change; the relaxed external-provider
+      // render gate would otherwise show old counters until the next completion.
       const checkpointChanged = state.params.checkpoint !== modelId;
-      // Clamp maxTokens to the new model's cap on switch into an
-      // external model so a value carried over from a prior local
-      // session does not render above the slider's max.
+      // Clamp maxTokens to the new model's cap when switching into an external
+      // model so a value carried over from a local session doesn't exceed the
+      // slider's max.
       let nextMaxTokens = state.params.maxTokens;
       if (checkpointChanged && isExternalModelId(modelId)) {
         const parsed = parseExternalModelId(modelId);
@@ -831,10 +816,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   setActiveProjectId: (activeProjectId) => set({ activeProjectId }),
   setSettingsPanelOpen: (settingsPanelOpen) => set({ settingsPanelOpen }),
   clearCheckpoint: () => {
-    // Mirror setCheckpoint's persistence behavior: dropping the
-    // checkpoint must also clear any stored external selection so
-    // the next refresh doesn't snap back to a model the user
-    // intentionally cleared.
+    // Mirror setCheckpoint's persistence: dropping the checkpoint must also
+    // clear any stored external selection so the next refresh doesn't snap
+    // back to a model the user intentionally cleared.
     saveLastExternalCheckpoint(null);
     return set((state) => ({
       params: {
diff --git a/studio/frontend/src/features/chat/tour/steps.tsx b/studio/frontend/src/features/chat/tour/steps.tsx
index e4f37685cc..de27e93c88 100644
--- a/studio/frontend/src/features/chat/tour/steps.tsx
+++ b/studio/frontend/src/features/chat/tour/steps.tsx
@@ -62,8 +62,8 @@ export function buildChatTourSteps({
   ];
 
   if (canCompare) {
-    // Compare now lives in the + menu, so there is no sidebar button to anchor
-    // to; the view step enters compare on its own and explains it.
+    // Compare lives in the + menu (no sidebar button to anchor to); this step
+    // enters compare on its own and explains it.
     steps.push({
       id: "compare-view",
       target: "chat-compare-view",
diff --git a/studio/frontend/src/features/chat/types.ts b/studio/frontend/src/features/chat/types.ts
index a2bd395065..ef4dd7dbcb 100644
--- a/studio/frontend/src/features/chat/types.ts
+++ b/studio/frontend/src/features/chat/types.ts
@@ -37,29 +37,24 @@ export interface ThreadRecord {
   archived: boolean;
   createdAt: number;
   /**
-   * OpenAI shell tool container id captured from a prior response on
-   * this thread. When set, the next turn reuses it via
-   * `environment.type="container_reference"` so the model can read
-   * files it wrote earlier in the conversation. When null/undefined,
-   * the next turn auto-creates a fresh container.
+   * OpenAI shell tool container id from a prior response. When set, the
+   * next turn reuses it via `environment.type="container_reference"` so
+   * the model can read files it wrote earlier; else auto-creates one.
    *
-   * OpenAI containers expire after ~20 min of inactivity by default;
-   * if a stale id is sent, the backend surfaces an
-   * `_toolEvent.type="container_invalidated"` and the chat-adapter
-   * clears this field so the following turn falls back to auto-create.
+   * Containers expire after ~20 min idle; on a stale id the backend
+   * emits `_toolEvent.type="container_invalidated"` and the chat-adapter
+   * clears this field so the next turn falls back to auto-create.
    */
   openaiCodeExecContainerId?: string | null;
   /**
-   * Anthropic code_execution container id captured from a prior
-   * response on this thread. When set, the next turn sends a
-   * top-level `container` field on /v1/messages so filesystem state
-   * (files, packages, variables) persists across turns. When
-   * null/undefined, Anthropic auto-creates a fresh container.
+   * Anthropic code_execution container id from a prior response. When
+   * set, the next turn sends a top-level `container` on /v1/messages so
+   * filesystem state (files, packages, variables) persists; else auto-
+   * creates one.
    *
-   * Anthropic containers expire after ~1 hour by default; on a stale
-   * id the backend surfaces `_toolEvent.type="container_invalidated"`
-   * and the chat-adapter clears this field so the following turn
-   * falls back to auto-create.
+   * Containers expire after ~1 hour; on a stale id the backend emits
+   * `_toolEvent.type="container_invalidated"` and the chat-adapter
+   * clears this field so the next turn falls back to auto-create.
    */
   anthropicCodeExecContainerId?: string | null;
 }
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 92ea1500bd..48efdbba2e 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -44,18 +44,16 @@ export interface LoadModelRequest {
   chat_template_override?: string | null;
   cache_type_kv?: string | null;
   /**
-   * Speculative decoding mode for GGUF models. Canonical values:
-   * "auto" (platform-aware: MTP on MTP GGUFs, ngram-mod fallback for
-   * sub-3B), "mtp" (force draft-mtp only on both GPU and CPU), "ngram"
-   * (force ngram-mod only), "mtp+ngram" (force ngram-mod + draft-mtp
-   * chain on both platforms), or "off". Legacy values "default" /
-   * "draft-mtp" / "ngram-mod" / "ngram-simple" are still accepted by
-   * the backend.
+   * Speculative decoding mode for GGUF models. Canonical values: "auto"
+   * (platform-aware: MTP on MTP GGUFs, ngram-mod fallback for sub-3B), "mtp"
+   * (force draft-mtp), "ngram" (force ngram-mod), "mtp+ngram" (ngram-mod +
+   * draft-mtp chain), "off". Legacy "default"/"draft-mtp"/"ngram-mod"/
+   * "ngram-simple" are still accepted by the backend.
    */
   speculative_type?: string | null;
   /**
-   * Override --spec-draft-n-max for MTP speculative decoding. Only
-   * applied when speculative_type resolves to "mtp" or "mtp+ngram".
+   * Override --spec-draft-n-max for MTP speculative decoding. Applied only
+   * when speculative_type resolves to "mtp" or "mtp+ngram".
    */
   spec_draft_n_max?: number | null;
 }
@@ -221,13 +219,11 @@ export type OpenAIMessageContentPart =
 export type OpenAIMessageContent = string | OpenAIMessageContentPart[];
 
 /**
- * OpenAI Chat Completions tool_call shape. Assistant turns echo back
- * function/tool calls as `tool_calls`; the matching tool result rides
- * on a separate `role="tool"` message keyed by `tool_call_id`.
- * `extra_content.google.thought_signature` is the Gemini-specific
- * round-trip field the backend translator both emits (on `delta.
- * tool_calls`) and consumes (when rebuilding the native functionCall
- * part on the next turn).
+ * OpenAI Chat Completions tool_call shape. Assistant turns echo function calls
+ * as `tool_calls`; the matching result rides on a separate `role="tool"`
+ * message keyed by `tool_call_id`. `extra_content.google.thought_signature` is
+ * the Gemini round-trip field the backend translator emits (on `delta.
+ * tool_calls`) and consumes (when rebuilding the functionCall part next turn).
  */
 export interface OpenAIToolCallPart {
   id?: string;
@@ -289,36 +285,28 @@ export interface OpenAIChatCompletionsRequest {
   encrypted_api_key?: string;
   provider_base_url?: string | null;
   /**
-   * Boolean toggle for OpenAI/Anthropic ephemeral cache_control. For
-   * Gemini the backend also accepts the cached-content resource name
-   * (`cachedContents/...`) as a string, which is forwarded as
-   * `generationConfig.cachedContent` on the native streamGenerateContent
-   * request.
+   * Boolean toggle for OpenAI/Anthropic ephemeral cache_control. For Gemini the
+   * backend also accepts a cached-content resource name (`cachedContents/...`)
+   * string, forwarded as `generationConfig.cachedContent`.
    */
   enable_prompt_caching?: boolean | string | null;
   /**
-   * OpenAI shell-tool container id captured from the prior response in
-   * this chat thread. When set and the Code pill is on, the backend
-   * routes the next /v1/responses call with
-   * `environment.type="container_reference"` so filesystem state
-   * persists across turns. Unset → backend uses
-   * `environment.type="container_auto"` and OpenAI creates a fresh
-   * container. Only meaningful for OpenAI cloud + gpt-5.5 family.
+   * OpenAI shell-tool container id from the prior response in this thread. When
+   * set and the Code pill is on, the backend routes the next /v1/responses with
+   * `environment.type="container_reference"` so filesystem state persists; unset
+   * → `container_auto` (fresh container). OpenAI cloud + gpt-5.5 family only.
    */
   openai_code_exec_container_id?: string | null;
   /**
-   * Anthropic code_execution container id captured from the prior
-   * response in this chat thread. When set and the Code pill is on,
-   * the backend forwards a top-level `container` field on
-   * /v1/messages so filesystem state persists across turns. Unset →
-   * Anthropic auto-creates a fresh container. Only meaningful for
-   * the Anthropic provider with `code_execution` in `enabled_tools`.
+   * Anthropic code_execution container id from the prior response in this
+   * thread. When set and the Code pill is on, the backend forwards a top-level
+   * `container` on /v1/messages so filesystem state persists; unset →
+   * auto-created. Anthropic provider with `code_execution` in `enabled_tools`.
    */
   anthropic_code_exec_container_id?: string | null;
   /**
-   * Anthropic fast-mode toggle. Opus 4.6 / 4.7 only; backend drops
-   * silently on every other model + provider. See
-   * https://platform.claude.com/docs/en/build-with-claude/fast-mode
+   * Anthropic fast-mode toggle. Opus 4.6 / 4.7 only; dropped silently elsewhere.
+   * See https://platform.claude.com/docs/en/build-with-claude/fast-mode
    */
   fast_mode?: boolean | null;
 }
@@ -327,16 +315,15 @@ export interface OpenAIChatDelta {
   role?: string;
   content?: string | null;
   /**
-   * Streamed assistant tool calls. The Gemini and OpenAI Responses
-   * translators emit incremental `tool_calls` deltas (function name +
-   * arguments fragments) so the chat-adapter can render tool cards as
-   * they arrive.
+   * Streamed assistant tool calls. The Gemini and OpenAI Responses translators
+   * emit incremental deltas (function name + arguments fragments) so the
+   * chat-adapter can render tool cards as they arrive.
    */
   tool_calls?: OpenAIToolCallPart[];
   /**
-   * Provider-specific passthrough. Gemini ships `thoughtSignature`,
-   * citations, `native_part`, etc., here so the round-trip can replay
-   * them on follow-up turns without bleeding into other providers.
+   * Provider-specific passthrough. Gemini ships `thoughtSignature`, citations,
+   * `native_part`, etc., here so the round-trip can replay them on follow-up
+   * turns without bleeding into other providers.
    */
   extra_content?: Record<string, unknown>;
 }
diff --git a/studio/frontend/src/features/chat/types/runtime.ts b/studio/frontend/src/features/chat/types/runtime.ts
index 24286c94fb..f521fd192f 100644
--- a/studio/frontend/src/features/chat/types/runtime.ts
+++ b/studio/frontend/src/features/chat/types/runtime.ts
@@ -15,8 +15,8 @@ export interface InferenceParams {
   /** Allow loading models with custom code (e.g. NVIDIA Nemotron). Only enable for repos you trust. */
   trustRemoteCode?: boolean;
   /**
-   * Anthropic fast-mode toggle. Opus 4.6 / 4.7 only; higher OTPS at
-   * 6x standard Opus pricing. Default false.
+   * Anthropic fast-mode toggle. Opus 4.6 / 4.7 only; higher OTPS at 6x Opus
+   * pricing. Default false.
    * https://platform.claude.com/docs/en/build-with-claude/fast-mode
    */
   fastMode?: boolean;
diff --git a/studio/frontend/src/features/chat/utils/chat-history-storage.ts b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
index 6dc53cf48e..0e9b30a5cc 100644
--- a/studio/frontend/src/features/chat/utils/chat-history-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
@@ -43,11 +43,10 @@ type ThreadListArgs = {
 };
 
 // localStorage perf-hint that the Dexie -> studio.db import already
-// finished in a previous session. NOT consulted by the import gate
-// itself -- the server-side ledger (chat_legacy_imports) is the source
-// of truth so a studio.db wipe stays recoverable. The hint only short-
-// circuits the listing paths' "should I also surface Dexie threads?"
-// branches once the ledger has covered everything.
+// finished. NOT the import gate -- the server-side ledger
+// (chat_legacy_imports) is the source of truth so a studio.db wipe stays
+// recoverable. The hint only short-circuits the listing paths' "also
+// surface Dexie threads?" branches once the ledger has covered everything.
 const LEGACY_CHAT_IMPORT_KEY = "unsloth_chat_legacy_imported_to_studio_db";
 
 let legacyChatImportPromise: Promise<void> | null = null;
@@ -265,9 +264,9 @@ async function backfillLegacyThreadFields(
   }
 }
 
-// Fast-path: ask IndexedDB whether the "unsloth-chat" database exists
-// without opening it. Modern Chromium / Firefox / Safari support this;
-// older browsers return undefined and we fall through to the next probe.
+// Fast-path: check whether the "unsloth-chat" DB exists without opening
+// it. Supported on modern Chromium/Firefox/Safari; older browsers return
+// undefined and we fall through to the next probe.
 async function dexieDbAbsent(): Promise<boolean> {
   if (typeof indexedDB === "undefined") return true;
   const dbs = (indexedDB as IDBFactory).databases;
@@ -291,24 +290,22 @@ async function dexieIsEmpty(): Promise<boolean> {
     ]);
     return threadCount === 0 && messageCount === 0;
   } catch {
-    // Dexie threw (corrupted DB / version mismatch / quota). Returning
-    // false forces the slow path, which uses the same Dexie under the
-    // hood; that path will throw too and the import promise gets reset
-    // so the next caller can retry rather than silently doing nothing.
+    // Dexie threw (corrupt DB / version mismatch / quota). Returning
+    // false forces the slow path (same Dexie underneath); it'll throw
+    // too and reset the import promise so the next caller can retry.
     return false;
   }
 }
 
 async function importLegacyChatsIfNeeded(): Promise<void> {
-  // Session-level cache: same tab, repeated sidebar mounts share one
-  // import. localStorage is NOT consulted here -- the server-side ledger
-  // is the source of truth so a studio.db wipe still re-triggers the
-  // import even if the browser kept its old hint.
+  // Session-level cache: repeated sidebar mounts in the same tab share
+  // one import. localStorage is NOT consulted -- the server-side ledger
+  // is the source of truth, so a studio.db wipe re-triggers the import
+  // even if the browser kept its old hint.
   if (legacyChatImportPromise) return legacyChatImportPromise;
 
   legacyChatImportPromise = (async () => {
-    // Fast-path: no Dexie database at all. New user, never had the
-    // browser-only Studio. ~0.1 ms, zero network.
+    // Fast-path: no Dexie DB -- new user, never had browser-only Studio.
     if (await dexieDbAbsent()) {
       markLegacyChatImportDone();
       return;
@@ -336,10 +333,9 @@ async function importLegacyChatsIfNeeded(): Promise<void> {
     const unimportedIds: string[] = [];
     const unimportedThreads: ThreadRecord[] = [];
 
-    // "Unimported" = missing from the ledger. We also include threads
-    // already present in the backend (without a ledger row) so the ledger
-    // gets backfilled for old-FE-then-new-FE users -- otherwise the next
-    // launch would redo the diff for the same threads forever.
+    // "Unimported" = missing from the ledger. Include threads already in
+    // the backend (without a ledger row) so the ledger gets backfilled
+    // for old-FE-then-new-FE users; else the next launch re-diffs forever.
     for (const thread of legacyThreads) {
       if (isChatThreadDeleted(thread.id)) continue;
       if (importedThreadIds.has(thread.id)) continue;
@@ -406,14 +402,12 @@ async function importLegacyChatsIfNeeded(): Promise<void> {
       result = await recordChatImportLedger(newlyImportedIds);
     } catch {
       // Network error: leave the perf hint alone so the next launch
-      // retries. The import itself is idempotent via UPSERT, no
-      // duplicates.
+      // retries. Import is idempotent via UPSERT, so no duplicates.
       return;
     }
-    // Only flip the localStorage hint when the backend actually has the
-    // ledger. On older deployments (404/405/501) the hint would lie:
-    // "import done" while the ledger stays empty, defeating recovery
-    // when studio.db gets wiped later.
+    // Only flip the hint when the backend actually has the ledger. On
+    // older deployments (404/405/501) it would lie ("import done" with an
+    // empty ledger), defeating recovery after a studio.db wipe.
     if (result.supported) {
       markLegacyChatImportDone();
     }
@@ -576,8 +570,8 @@ export async function listStoredChatThreadsWithMessages(
 ): Promise<ThreadRecord[]> {
   const threads = await listStoredChatThreads(args);
   if (threads.length === 0) return [];
-  // One batched HTTP call instead of N. Per-thread legacy Dexie
-  // fallback only fires when the batch result is empty.
+  // One batched HTTP call instead of N. Per-thread legacy Dexie fallback
+  // only fires when the batch result is empty.
   const threadIds = threads.map((t) => t.id);
   let backendByThread: Map<string, MessageRecord[]>;
   try {
@@ -728,8 +722,8 @@ export interface ClearStoredChatsResult {
 }
 
 export async function clearStoredChats(): Promise<ClearStoredChatsResult> {
-  // Clear both sides independently and report each outcome so the
-  // toast can distinguish full vs partial success.
+  // Clear both sides independently and report each outcome so the toast
+  // can distinguish full vs partial success.
   const [backendThreadsResult, legacyThreads] = await Promise.all([
     listChatThreads()
       .then((threads) => ({ ok: true as const, threads }))
@@ -752,8 +746,8 @@ export async function clearStoredChats(): Promise<ClearStoredChatsResult> {
     failedThreadIds: [],
   };
   try {
-    // Defer the history refresh until Dexie clear and tombstone state are
-    // finalized, so listeners never observe the composite clear mid-flight.
+    // Defer the history refresh until Dexie clear and tombstones finalize,
+    // so listeners never observe the composite clear mid-flight.
     await clearBackendChats({ notify: false });
     result.backend = "cleared";
   } catch (error) {
diff --git a/studio/frontend/src/features/chat/utils/chat-thread-tombstones.ts b/studio/frontend/src/features/chat/utils/chat-thread-tombstones.ts
index d9a6f2d501..4d2a0da653 100644
--- a/studio/frontend/src/features/chat/utils/chat-thread-tombstones.ts
+++ b/studio/frontend/src/features/chat/utils/chat-thread-tombstones.ts
@@ -2,10 +2,9 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Tombstones mask deleted threads in the Dexie read fallback. Each
- * carries a `deletedAt` timestamp so old entries can be GC'd, keeping
- * localStorage bounded. Reads accept both the legacy plain-string
- * format and the new {id, deletedAt} tuple form.
+ * Tombstones mask deleted threads in the Dexie read fallback. Each carries a
+ * `deletedAt` so old entries can be GC'd, keeping localStorage bounded. Reads
+ * accept both the legacy plain-string format and the {id, deletedAt} tuple.
  */
 
 interface Tombstone {
@@ -62,8 +61,7 @@ function gc(): void {
   for (const [id, t] of deletedThreads) {
     if (t.deletedAt < cutoff) deletedThreads.delete(id);
   }
-  // Cap absolute size: drop oldest if we somehow exceed the limit
-  // (e.g. a script clearing thousands of threads at once).
+  // Cap size: drop oldest if we exceed the limit (e.g. a bulk thread clear).
   if (deletedThreads.size > TOMBSTONE_MAX_COUNT) {
     const sorted = Array.from(deletedThreads.entries()).sort(
       (a, b) => a[1].deletedAt - b[1].deletedAt,
diff --git a/studio/frontend/src/features/chat/utils/delete-thread-message.ts b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
index 37d4732d93..3b439eb779 100644
--- a/studio/frontend/src/features/chat/utils/delete-thread-message.ts
+++ b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
@@ -2,16 +2,15 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * assistant-ui does not expose a public `deleteMessage` on `ThreadRuntime` / `MessageRuntime`
- * in our version, but it already implements branch-safe deletion inside `MessageRepository`.
- * We import that helper from `@assistant-ui/core/internal`, the package's exported internal
- * surface. Avoid importing the deeper `runtime/utils/message-repository` path directly: newer
- * `@assistant-ui/core` releases no longer export arbitrary deep paths.
+ * assistant-ui exposes no public `deleteMessage` in our version, but
+ * `MessageRepository` already does branch-safe deletion. We import it from
+ * `@assistant-ui/core/internal` (the exported internal surface); avoid the
+ * deeper `runtime/utils/message-repository` path since newer releases no
+ * longer export arbitrary deep paths.
  *
- * **Maintainability:** treat this file as the only place that imports `MessageRepository` from
- * `@assistant-ui/core`. When bumping `@assistant-ui/react` / `@assistant-ui/core`, re-run chat
- * delete + reload smoke tests; the path or API may change without a semver signal on “public”
- * surface area.
+ * Keep this file the only importer of `MessageRepository`. When bumping
+ * `@assistant-ui/react` / `core`, re-run chat delete + reload smoke tests;
+ * the path or API may change without a semver signal.
  */
 import { MessageRepository } from "@assistant-ui/core/internal";
 import type {
diff --git a/studio/frontend/src/features/chat/utils/format-transfer.ts b/studio/frontend/src/features/chat/utils/format-transfer.ts
index 737f26c8cb..b6e23519c7 100644
--- a/studio/frontend/src/features/chat/utils/format-transfer.ts
+++ b/studio/frontend/src/features/chat/utils/format-transfer.ts
@@ -2,14 +2,9 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Format a byte-per-second rate as a human-readable string.
- *
- *   512            → "512 B/s"
- *   1_234_567      → "1.2 MB/s"
- *   1_234_567_890  → "1.15 GB/s"
- *
- * Returns `"--"` for non-finite or non-positive inputs so callers can
- * render the label safely before the first stable sample arrives.
+ * Format a byte-per-second rate (e.g. 1_234_567 → "1.2 MB/s"). Returns `"--"`
+ * for non-finite or non-positive inputs so the label renders safely before the
+ * first stable sample.
  */
 export function formatRate(bytesPerSecond: number): string {
   if (!Number.isFinite(bytesPerSecond) || bytesPerSecond <= 0) return "--";
@@ -21,13 +16,8 @@ export function formatRate(bytesPerSecond: number): string {
 }
 
 /**
- * Format an ETA (in seconds) as a short human-readable string.
- *
- *   47       → "47s"
- *   125      → "2m 5s"
- *   3725     → "1h 2m"
- *
- * Returns `"--"` for non-finite or non-positive inputs.
+ * Format an ETA in seconds as a short string (e.g. 125 → "2m 5s", 3725 →
+ * "1h 2m"). Returns `"--"` for non-finite or non-positive inputs.
  */
 export function formatEta(seconds: number): string {
   if (!Number.isFinite(seconds) || seconds <= 0) return "--";
diff --git a/studio/frontend/src/features/chat/utils/image-input-support.ts b/studio/frontend/src/features/chat/utils/image-input-support.ts
index fb39303e29..d1ea31fc76 100644
--- a/studio/frontend/src/features/chat/utils/image-input-support.ts
+++ b/studio/frontend/src/features/chat/utils/image-input-support.ts
@@ -14,8 +14,8 @@ export function getImageInputUnavailableReason({
   activeModel?: ChatModelSummary;
   isExternalModel: boolean;
   // true/false = caller knows; null/undefined = unknown (default-allow).
-  // External selections aren't in runtime.models[], so callers should
-  // resolve provider-type capability and pass it here.
+  // External selections aren't in runtime.models[], so callers resolve
+  // provider-type capability and pass it here.
   externalSupportsVision?: boolean | null;
   // Fallback toast label when activeModel is missing.
   externalModelLabel?: string | null;
@@ -40,11 +40,10 @@ export function getImageInputUnavailableReason({
     return null;
   }
   if (!modelLoaded) return "Load a model before adding images.";
-  // loadedIsMultimodal is true for vision OR audio. Can't tell them apart
-  // from that one flag, so only block when activeModel confirms
-  // audio-only: audio capability set AND isVision === false. Otherwise
-  // trust the load response. The models-list entry might be stale, or
-  // not even there yet (gets auto-injected after load).
+  // loadedIsMultimodal is true for vision OR audio; that one flag can't tell
+  // them apart, so only block when activeModel confirms audio-only (audio
+  // capability set AND isVision === false). Otherwise trust the load
+  // response: the models-list entry may be stale or not yet injected.
   if (loadedIsMultimodal) {
     const isAudioOnly =
       Boolean(activeModel?.isAudio || activeModel?.hasAudioInput) &&
diff --git a/studio/frontend/src/features/chat/utils/parse-assistant-content.ts b/studio/frontend/src/features/chat/utils/parse-assistant-content.ts
index 36655ca81e..515fb0e1dd 100644
--- a/studio/frontend/src/features/chat/utils/parse-assistant-content.ts
+++ b/studio/frontend/src/features/chat/utils/parse-assistant-content.ts
@@ -8,11 +8,9 @@ type ContentPart = NonNullable<ChatModelRunResult["content"]>[number];
 const THINK_OPEN_TAG = "<think>";
 const THINK_CLOSE_TAG = "</think>";
 
-// ContentPart from @assistant-ui/react has readonly fields, so we cannot
-// do `last.text += text` to coalesce adjacent same-type parts — tsc fails
-// with TS2540 "Cannot assign to 'text' because it is a read-only property".
-// Instead, replace the last element with a fresh merged object: same
-// allocation cost as the mutation path but type-safe.
+// ContentPart from @assistant-ui/react has readonly fields, so coalescing via
+// `last.text += text` fails (TS2540). Instead replace the last element with a
+// fresh merged object: same allocation cost as mutation but type-safe.
 
 function appendTextPart(parts: ContentPart[], text: string): void {
   if (!text) return;
diff --git a/studio/frontend/src/features/chat/utils/transfer-stats.ts b/studio/frontend/src/features/chat/utils/transfer-stats.ts
index 7aecf29eb0..ca6c082a1a 100644
--- a/studio/frontend/src/features/chat/utils/transfer-stats.ts
+++ b/studio/frontend/src/features/chat/utils/transfer-stats.ts
@@ -4,13 +4,9 @@
 /**
  * Pure, framework-free math behind {@link useTransferStats}.
  *
- * Split out so it can be unit-tested without a DOM/React renderer, and
- * so the training-start overlay, the chat download toast, and the
- * model-load phase UI all share the exact same rate/ETA semantics.
- *
- * No React, no `useRef`/`useState`, no timers -- the caller owns the
- * sample buffer and clock. This is the "what is the rate right now
- * given these samples" question, nothing more.
+ * Split out for unit-testing without React, and so the training-start overlay,
+ * chat download toast, and model-load UI share identical rate/ETA semantics.
+ * No React/timers -- the caller owns the sample buffer and clock.
  */
 
 export type TransferSample = { t: number; b: number };
@@ -19,10 +15,9 @@ export type TransferStats = {
   rateBytesPerSecond: number;
   etaSeconds: number;
   /**
-   * False until the window has at least {@link MIN_SAMPLES} samples
-   * spanning ≥ {@link MIN_WINDOW_SECONDS} with strictly forward progress.
-   * Consumers should hide rate/ETA while this is false so the UI doesn't
-   * flicker "123 GB/s" during the first tick.
+   * False until the window has ≥ {@link MIN_SAMPLES} samples spanning ≥
+   * {@link MIN_WINDOW_SECONDS} with forward progress. Hide rate/ETA while false
+   * so the UI doesn't flicker "123 GB/s" during the first tick.
    */
   stable: boolean;
 };
@@ -32,11 +27,9 @@ export const MIN_WINDOW_SECONDS = 3;
 export const MAX_WINDOW_SECONDS = 15;
 
 /**
- * Mutate ``samples`` in place: append the new sample, drop any that
- * fell out of the rolling window, and clear the buffer if the counter
- * went backwards (user cancelled + restarted a download, etc.).
- *
- * Returns the same array for chain-ability.
+ * Mutate ``samples`` in place: append the sample, drop any out of the rolling
+ * window, and clear the buffer if the counter went backwards (cancel + restart).
+ * Returns the same array for chaining.
  */
 export function appendSample(
   samples: TransferSample[],
@@ -56,13 +49,11 @@ export function appendSample(
 }
 
 /**
- * Derive {@link TransferStats} from a window of cumulative-byte samples
- * plus the known total.
- *
+ * Derive {@link TransferStats} from a window of cumulative-byte samples plus the
+ * known total.
  *   * Needs ≥ {@link MIN_SAMPLES} samples spanning ≥ {@link MIN_WINDOW_SECONDS}
- *     seconds before it will report ``stable: true``.
- *   * ETA is clamped to 0 when: no progress, no total known, or the
- *     counter already hit the total.
+ *     seconds before reporting ``stable: true``.
+ *   * ETA clamps to 0 when there's no progress, no total, or total is hit.
  */
 export function computeTransferStats(
   samples: readonly TransferSample[],
diff --git a/studio/frontend/src/features/export/api/export-api.ts b/studio/frontend/src/features/export/api/export-api.ts
index 56f8c8129b..58df511366 100644
--- a/studio/frontend/src/features/export/api/export-api.ts
+++ b/studio/frontend/src/features/export/api/export-api.ts
@@ -37,10 +37,8 @@ export interface ExportOperationResponse {
   success: boolean;
   message: string;
   /**
-   * Optional extras returned by the backend. The export endpoints set
-   * `details.output_path` to the resolved on-disk directory of the
-   * saved model when a local save was requested. Hub-only pushes leave
-   * `details` undefined.
+   * Optional backend extras. Local saves set `details.output_path` to the
+   * saved model's on-disk directory; hub-only pushes leave it undefined.
    */
   details?: { output_path?: string | null } & Record<string, unknown>;
 }
diff --git a/studio/frontend/src/features/export/components/export-dialog.tsx b/studio/frontend/src/features/export/components/export-dialog.tsx
index 176b10a52f..c529cadab6 100644
--- a/studio/frontend/src/features/export/components/export-dialog.tsx
+++ b/studio/frontend/src/features/export/components/export-dialog.tsx
@@ -26,9 +26,8 @@ import { streamExportLogs, type ExportLogEntry } from "../api/export-api";
 import { collapseAnim } from "../anim";
 import { EXPORT_METHODS, type ExportMethod } from "../constants";
 
-// Max log lines kept in the dialog's local state. Matches the backend
-// ring buffer's maxlen so the UI shows the full scrollback captured
-// server side.
+// Max log lines kept in local state. Matches the backend ring buffer's maxlen
+// so the UI shows the full server-side scrollback.
 const MAX_LOG_LINES = 4000;
 
 interface UseExportLogsResult {
@@ -38,14 +37,10 @@ interface UseExportLogsResult {
 }
 
 /**
- * Subscribe to the live export log SSE stream while `exporting` is
- * true, and accumulate lines in local state. Lines from a previous
- * action are cleared:
- *
- *   - when a new export starts (`exporting` flips to true), and
- *   - when the user switches export method, dialog opens fresh, or
- *     the dialog closes — so re-opening into a different action's
- *     screen doesn't show the prior screen's saved output.
+ * Subscribe to the live export log SSE stream while `exporting` is true and
+ * accumulate lines. Prior-action lines are cleared when a new export starts and
+ * when the user switches method / reopens / closes the dialog, so reopening
+ * into a different action doesn't show stale output.
  */
 function useExportLogs(
   exporting: boolean,
@@ -56,10 +51,9 @@ function useExportLogs(
   const [connected, setConnected] = useState(false);
   const [error, setError] = useState<string | null>(null);
 
-  // Reset log state whenever the user moves to a different screen --
-  // either by switching export method or by reopening the dialog -- so
-  // each (open × method) tuple shows only its own run history. The
-  // streaming effect below additionally clears on new export start.
+  // Reset log state on screen change (method switch or dialog reopen) so each
+  // (open × method) shows only its own run history. The streaming effect below
+  // additionally clears on new export start.
   useEffect(() => {
     setLines([]);
     setError(null);
@@ -74,17 +68,15 @@ function useExportLogs(
 
     const abortCtrl = new AbortController();
     let cancelled = false;
-    // Track the highest seq we've observed on a `log` event so we can
-    // resume the stream via `since=` / `Last-Event-ID` after a drop.
-    // The backend's SSE `id:` field carries this as ExportLogEvent.id.
+    // Highest seq seen on a `log` event, to resume via `since=` / Last-Event-ID
+    // after a drop. The backend's SSE `id:` field carries this as ExportLogEvent.id.
     let lastSeq: number | null = null;
-    // Exponential backoff with jitter, capped. Reset on every
-    // successful connection so flaky networks don't accumulate delay.
+    // Capped exponential backoff with jitter; reset on each successful connect
+    // so flaky networks don't accumulate delay.
     let backoffMs = 500;
     const MAX_BACKOFF_MS = 5000;
-    // Flipped by a terminal event (explicit `complete` from the
-    // backend or a non-transient error we choose not to retry). Stops
-    // the outer reconnect loop even if `exporting` is still true.
+    // Set by a terminal event (backend `complete` or a non-retryable error) to
+    // stop the reconnect loop even while `exporting` is still true.
     let terminated = false;
 
     const run = async () => {
@@ -115,9 +107,8 @@ function useExportLogs(
                   return next;
                 });
               } else if (event.event === "complete") {
-                // Backend signalled the run is fully drained -- stop
-                // trying to reconnect even though `exporting` may not
-                // have flipped false yet on this tick.
+                // Run fully drained -- stop reconnecting even if `exporting`
+                // hasn't flipped false yet on this tick.
                 terminated = true;
               } else if (event.event === "error" && event.error) {
                 setError(event.error);
@@ -128,17 +119,15 @@ function useExportLogs(
           if (cancelled) return;
           if (err instanceof DOMException && err.name === "AbortError") return;
           setError(err instanceof Error ? err.message : String(err));
-          // Fall through to the backoff path below; a fetch-level
-          // failure is retryable the same way a clean EOF is.
+          // Fall through to backoff; a fetch-level failure retries like a clean EOF.
         }
 
         setConnected(false);
         if (cancelled || terminated) return;
 
-        // Exponential backoff with jitter before reconnecting. The
-        // backend's ring buffer plus Last-Event-ID resume means we
-        // don't lose lines across the retry as long as the reconnect
-        // happens within the buffer's lifetime (~4000 lines).
+        // Exponential backoff with jitter before reconnecting. Ring buffer +
+        // Last-Event-ID resume means no lines lost across the retry as long as
+        // the reconnect lands within the buffer's lifetime (~4000 lines).
         const delay = backoffMs + Math.floor(Math.random() * 250);
         backoffMs = Math.min(backoffMs * 2, MAX_BACKOFF_MS);
         try {
@@ -163,9 +152,8 @@ function useExportLogs(
       }
     };
 
-    // run()'s own try/catch handles every failure path we care about;
-    // swallow anything that somehow escapes so React's dev overlay
-    // doesn't flag an unhandled rejection on dialog close.
+    // run() handles every failure path; swallow stragglers so React's dev
+    // overlay doesn't flag an unhandled rejection on dialog close.
     void run().catch(() => {});
 
     return () => {
@@ -179,10 +167,9 @@ function useExportLogs(
 }
 
 /**
- * Tick every second while `exporting` is true and report elapsed
- * seconds. Powers the "Working… 27s" badge in the log header so the
- * panel doesn't look frozen during long single-step phases (cache
- * file copy, GGUF conversion) when no new lines are arriving.
+ * Tick every second while `exporting` is true and report elapsed seconds.
+ * Powers the "Working… 27s" badge so the panel doesn't look frozen during long
+ * single-step phases (cache copy, GGUF conversion) with no new lines.
  */
 function useElapsedSeconds(exporting: boolean): number {
   const [elapsed, setElapsed] = useState(0);
@@ -209,8 +196,7 @@ function formatElapsed(seconds: number): string {
 }
 
 function formatLogLine(entry: ExportLogEntry): string {
-  // Strip trailing carriage returns that tqdm-style progress leaves
-  // in the stream so the scrollback doesn't render funky boxes.
+  // Strip trailing CRs from tqdm-style progress so scrollback doesn't render boxes.
   return entry.line.replace(/\r+$/g, "");
 }
 
@@ -240,9 +226,8 @@ interface ExportDialogProps {
   exportError: string | null;
   exportSuccess: boolean;
   /**
-   * Resolved on-disk realpath of the most recent successful export.
-   * Surfaced on the Export Complete screen so users can find their
-   * model. Null when the export only pushed to the Hub.
+   * Resolved on-disk realpath of the latest successful export, shown on the
+   * Export Complete screen. Null when the export only pushed to the Hub.
    */
   exportOutputPath: string | null;
 }
@@ -272,8 +257,7 @@ export function ExportDialog({
   exportSuccess,
   exportOutputPath,
 }: ExportDialogProps) {
-  // Live log capture is useful for any export path executed by the
-  // backend worker, including LoRA adapter-only export.
+  // Live log capture applies to any backend-worker export path, incl. LoRA-only.
   const showLogPanel =
     exportMethod === "merged" ||
     exportMethod === "gguf" ||
@@ -285,8 +269,7 @@ export function ExportDialog({
   const elapsedSeconds = useElapsedSeconds(exporting && showLogPanel);
 
   const logScrollRef = useRef<HTMLDivElement | null>(null);
-  // Auto-scroll to bottom whenever a new line arrives, unless the
-  // user has scrolled up to read earlier output.
+  // Auto-scroll to bottom on each new line, unless the user scrolled up.
   const [followTail, setFollowTail] = useState(true);
 
   useEffect(() => {
@@ -461,9 +444,8 @@ export function ExportDialog({
               )}
             </AnimatePresence>
 
-            {/* Success banner for log-driven exports.
-                Keep users on the log screen after completion so they can
-                inspect conversion output before closing. */}
+            {/* Success banner: keep users on the log screen after completion so
+                they can inspect conversion output before closing. */}
             {exportSuccess && showLogPanel && (
               <div className="flex items-start gap-2 rounded-lg bg-emerald-500/10 p-3 text-sm text-emerald-700 dark:text-emerald-300">
                 <HugeiconsIcon icon={CheckmarkCircle02Icon} className="mt-0.5 size-4 shrink-0" />
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 233164b59a..b1a379b582 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -121,9 +121,8 @@ export function ExportPage() {
   const [exporting, setExporting] = useState(false);
   const [exportError, setExportError] = useState<string | null>(null);
   const [exportSuccess, setExportSuccess] = useState(false);
-  // Resolved on-disk path of the most recent successful export, surfaced
-  // on the Export Complete screen so the user can find their model
-  // without digging through the server log. Null for Hub-only pushes.
+  // On-disk path of the last successful export, shown on the Export Complete
+  // screen so the user can find their model. Null for Hub-only pushes.
   const [exportOutputPath, setExportOutputPath] = useState<string | null>(null);
 
   const hfComboboxAnchorRef = useRef<HTMLDivElement>(null);
@@ -438,11 +437,9 @@ export function ExportPage() {
         });
       }
 
-      // 2. Run export based on method. Capture the resolved output_path
-      // (when the backend wrote a local copy) so the success screen can
-      // show the user the realpath of their saved model. For multi-quant
-      // GGUF runs, the directory is the same for every quant so we just
-      // keep the last response.
+      // 2. Run export. Capture the resolved output_path (when the backend
+      // wrote a local copy) for the success screen. Multi-quant GGUF runs
+      // share one directory, so we just keep the last response.
       let lastOutputPath: string | null = null;
       if (exportMethod === "merged") {
         if (isAdapter) {
diff --git a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
index 337023ad33..a1174186af 100644
--- a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
@@ -39,10 +39,7 @@ function formatLR(value: number): string {
   return `${rounded}e${exp}`;
 }
 
-/**
- * Step learning rate up in a scientific-notation-friendly sequence:
- * 1e-4 -> 2e-4 -> 3e-4 -> ... -> 9e-4 -> 1e-3 -> 2e-3 -> ...
- */
+/** Step the LR in a scientific sequence: 1e-4 -> 2e-4 -> ... -> 9e-4 -> 1e-3 -> ... */
 function stepLR(value: number, direction: 1 | -1): number {
   if (value <= 0) return 1e-5;
   const exp = Math.floor(Math.log10(value) + 1e-9);
@@ -116,8 +113,7 @@ export function HyperparametersStep() {
   const maxStepsSliderMax = Math.max(500, maxSteps, 30);
   const epochsSliderMax = Math.max(10, epochs, 1);
 
-  // Use model's max_position_embeddings to cap context length options.
-  // Fall back to 65536 (64K) if not available.
+  // Cap context length by model's max_position_embeddings; fall back to 64K.
   const maxCtx = maxPositionEmbeddings ?? 65536;
   const contextLengthOptions = useMemo(
     () => CONTEXT_LENGTHS.filter((len) => len <= maxCtx),
diff --git a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
index 52f61700e2..8e15fc6195 100644
--- a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
@@ -115,7 +115,7 @@ export function ModelSelectionStep() {
     return applyPriorityOrdering(ids);
   }, [hfResults]);
 
-  // Match Studio behavior: only show exception signals (OOM/TIGHT) in training flows.
+  // Match Studio: only show exception signals (OOM/TIGHT) in training flows.
   const vramMap = useMemo(() => {
     const fitMap = buildModelVramMap(
       hfResults,
diff --git a/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
index 4ab9fedf10..2a286a8c80 100644
--- a/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
@@ -192,7 +192,7 @@ export function ModelTypeStep(): ReactElement {
             )}
           >
             <CardContent className="flex items-center gap-4 py-4">
-              {/* Invisible spacer matching RadioGroupItem (size-4 flex) in other cards */}
+              {/* Invisible spacer matching RadioGroupItem (size-4) in other cards */}
               <div className="size-4 shrink-0" aria-hidden="true" />
               <div
                 className={cn(
diff --git a/studio/frontend/src/features/profile/utils/jwt-subject.ts b/studio/frontend/src/features/profile/utils/jwt-subject.ts
index 9c7596966a..a6b71811f6 100644
--- a/studio/frontend/src/features/profile/utils/jwt-subject.ts
+++ b/studio/frontend/src/features/profile/utils/jwt-subject.ts
@@ -1,9 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-/**
- * Read the JWT `sub` claim for display purposes only (not verified).
- */
+/** Read the JWT `sub` claim for display only (not verified). */
 export function decodeJwtSubject(token: string | null): string | null {
   if (!token) return null;
   try {
diff --git a/studio/frontend/src/features/profile/utils/resize-image-file.ts b/studio/frontend/src/features/profile/utils/resize-image-file.ts
index c420497395..7df5f45972 100644
--- a/studio/frontend/src/features/profile/utils/resize-image-file.ts
+++ b/studio/frontend/src/features/profile/utils/resize-image-file.ts
@@ -50,9 +50,7 @@ function encodeCanvasWithinLimit(
   return dataUrl.length <= MAX_DATA_URL_LENGTH ? dataUrl : null;
 }
 
-/**
- * Downscale the image and keep transparency when present while staying localStorage-friendly.
- */
+/** Downscale the image, preserving transparency, to stay localStorage-friendly. */
 export async function resizeImageFileToDataUrl(file: File): Promise<string> {
   const img = await loadImage(file);
   const w = img.naturalWidth;
diff --git a/studio/frontend/src/features/recipe-studio/components/executions/executions-view.tsx b/studio/frontend/src/features/recipe-studio/components/executions/executions-view.tsx
index cfc0e61338..3488ac0595 100644
--- a/studio/frontend/src/features/recipe-studio/components/executions/executions-view.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/executions/executions-view.tsx
@@ -108,11 +108,9 @@ export function ExecutionsView({
     [datasetColumnNames, hiddenDatasetColumns],
   );
 
-  // Columns where at least one row has text long enough that it would wrap at
-  // the default narrow width. We give those columns a wider min-width so the
-  // text is readable without clicking anything. The table's wrapper already
-  // scrolls horizontally, so a few wide columns just add a horizontal
-  // scrollbar instead of squeezing everything into the viewport.
+  // Columns with at least one long-text row get a wider min-width so the text
+  // is readable without clicking. The wrapper scrolls horizontally, so wide
+  // columns just add a scrollbar instead of squeezing the viewport.
   const wideColumns = useMemo(() => {
     const result = new Set<string>();
     if (!selectedExecution) {
diff --git a/studio/frontend/src/features/recipe-studio/components/inline/inline-model.tsx b/studio/frontend/src/features/recipe-studio/components/inline/inline-model.tsx
index 1d98d18b56..b7354c1805 100644
--- a/studio/frontend/src/features/recipe-studio/components/inline/inline-model.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/inline/inline-model.tsx
@@ -55,9 +55,9 @@ export function InlineModel(props: InlineModelProps): ReactElement {
     );
   }
 
-  // model_config branch - mirror the local-aware provider sync from the
-  // dialog path so inline edits clear stale local-only metadata without
-  // synthesizing the legacy "local" placeholder.
+  // model_config branch: mirror the dialog path's local-aware provider sync so
+  // inline edits clear stale local-only metadata without synthesizing the
+  // legacy "local" placeholder.
   const localNames = props.localProviderNames ?? new Set<string>();
   const modelConfig = props.config;
   const isLinkedToLocal = localNames.has(modelConfig.provider);
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
index 00b9fa1b46..9b91bfe334 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
@@ -556,8 +556,8 @@ export function LocalRecipeModelSelector({
             return;
           }
         } catch {
-          // Non-GGUF local models commonly have no variant endpoint. Fall
-          // through to regular selection so users can still choose them.
+          // Non-GGUF local models commonly have no variant endpoint;
+          // fall through to regular selection so they stay choosable.
         } finally {
           setProbingVariantModelId(null);
         }
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/models/model-config-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/models/model-config-dialog.tsx
index 68f912bc57..64f203fe4f 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/models/model-config-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/models/model-config-dialog.tsx
@@ -49,10 +49,8 @@ export function ModelConfigDialog({
   const skipHealthCheckId = `${config.id}-skip-health-check`;
   const providerAnchorRef = useRef<HTMLDivElement>(null);
   const providerInputRef = useRef(config.provider);
-  // Sync providerInputRef with the current provider value. Updating a ref in
-  // an effect (vs reading/writing it during render) satisfies the
-  // react-hooks/refs rule and keeps the combobox blur path stable across
-  // re-renders.
+  // Sync providerInputRef in an effect (not during render) to satisfy the
+  // react-hooks/refs rule and keep the combobox blur path stable.
   useEffect(() => {
     providerInputRef.current = config.provider;
   }, [config.provider]);
diff --git a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-executions.ts b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-executions.ts
index 19a6a3004d..f11fed6e13 100644
--- a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-executions.ts
+++ b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-executions.ts
@@ -563,10 +563,9 @@ export function useRecipeExecutions({
 
     resetForRecipe();
 
-    // Seed previewRows from the recipe's original run.rows (read from the
-    // loaded JSON, not the rebuilt payload (the builder hardcodes 5).
-    // Templates ship their own suggested preview size (e.g. GitHub Support
-    // Bot: 10); we honor it so users don't see a surprise 5.
+    // Seed previewRows from the recipe's original run.rows (the loaded JSON, not
+    // the rebuilt payload, which hardcodes 5). Templates ship a suggested preview
+    // size (e.g. GitHub Support Bot: 10); honor it so users don't see a surprise 5.
     if (
       typeof initialRunRows === "number" &&
       Number.isFinite(initialRunRows) &&
@@ -794,8 +793,8 @@ export function useRecipeExecutions({
       }
 
       // Flip to the Runs pane before validation starts. Validation can re-crawl
-      // the seed (multiple seconds for the github_repo reader), and runExecution()
-      // later no-ops this callback if the view has already been flipped.
+      // the seed (seconds for the github_repo reader); runExecution() later no-ops
+      // this callback if the view has already been flipped.
       onExecutionStart?.();
 
       const normalizedRows = sanitizeExecutionRows(rows, kind);
@@ -811,11 +810,11 @@ export function useRecipeExecutions({
         return false;
       }
 
-      // Recipe and Chat share one singleton local inference backend. This
-      // direct load is a point-in-time handoff to job creation, not a lease:
-      // if Chat swaps models after this succeeds, the backend will reject or
-      // run against the active backend state. A future generation token should
-      // be validated across this load and the `/jobs` loaded-model gate.
+      // Recipe and Chat share one singleton local inference backend. This direct
+      // load is a point-in-time handoff to job creation, not a lease: if Chat
+      // swaps models after this succeeds, the backend runs against current state.
+      // A future generation token should be validated across this load and the
+      // `/jobs` loaded-model gate.
       const restorePrevious = await prepareLocalModelForExecution(payload);
       if (restorePrevious === false) {
         return false;
diff --git a/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx b/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
index c1001c0322..ec9358ee0f 100644
--- a/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
+++ b/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
@@ -78,16 +78,15 @@ const EDGE_TYPES: EdgeTypes = {
 const COMPLETE_ISLAND_VISIBLE_MS = 7_000;
 const TAB_SWITCH_FIT_DELAY_MS = 110;
 /**
- * Maximum RAF iterations to wait for React Flow's ResizeObserver to populate
- * `node.measured` dimensions before calling fitView. ~20 frames ≈ 333 ms at
- * 60 fps — more than enough for the render → layout → ResizeObserver cycle.
+ * Max RAF iterations to wait for React Flow's ResizeObserver to populate
+ * `node.measured` before calling fitView. ~20 frames ≈ 333 ms at 60 fps,
+ * ample for the render → layout → ResizeObserver cycle.
  */
 const MAX_FIT_VIEW_RETRIES = 20;
 /**
- * After all target nodes appear measured, wait this many extra stable frames
- * before firing fitView. This absorbs `updateNodeInternals` calls from
- * InternalsSync and individual node mount effects that can transiently reset
- * measurements.
+ * Extra stable frames to wait after target nodes appear measured before
+ * firing fitView, absorbing `updateNodeInternals` calls from InternalsSync
+ * and node mount effects that can transiently reset measurements.
  */
 const FIT_VIEW_STABLE_FRAMES = 3;
 
@@ -227,10 +226,10 @@ export function RecipeStudioPage({
     },
     [viewModeStorageKey],
   );
-  // Easy mode has no canvas overlay/progress island, so once a run starts the
-  // user sees the Run button stuck on "Running..." with nothing else changing.
-  // Flip to the Runs pane so they land where progress is actually rendered.
-  // Advanced (editor) keeps its island and stays put.
+  // Easy mode has no canvas overlay/progress island, so a started run would
+  // leave the Run button stuck on "Running..." with nothing else changing.
+  // Flip to the Runs pane where progress is rendered. Advanced (editor) keeps
+  // its island and stays put.
   const handleExecutionStart = useCallback(() => {
     setActiveView((currentView) =>
       currentView === "easy" ? "executions" : currentView,
@@ -396,10 +395,9 @@ export function RecipeStudioPage({
   const runBusy = previewLoading || fullLoading || executionLocked;
   const islandExecution = activeExecution ?? recentCompletedExecution;
 
-  // Easy mode runs a full run (artifact persisted, tracked in Runs pane)
-  // using runFull. runFull requires a non-empty fullRunName but the Easy form
-  // has no run-name input, so seed a default here as soon as Easy is active.
-  // User can still rename it from Advanced/Runs dialogs before clicking Run.
+  // Easy mode uses runFull (artifact persisted, tracked in Runs pane), which
+  // requires a non-empty fullRunName. The Easy form has no run-name input, so
+  // seed a default once Easy is active; user can rename from Advanced/Runs.
   useEffect(() => {
     if (!supportsEasyMode) return;
     if (activeView !== "easy") return;
@@ -527,22 +525,21 @@ export function RecipeStudioPage({
           return;
         }
         if (retries >= MAX_FIT_VIEW_RETRIES) {
-          // Timed out waiting — fit with whatever we have (graceful fallback).
+          // Timed out: fit with whatever we have (graceful fallback).
           doFit();
           return;
         }
         const targets = getFitViewTargetNodes(reactFlowInstance.getNodes());
         if (allTargetsMeasured(targets)) {
           stableCount++;
-          // Wait a few extra frames after measurements appear to let
-          // updateNodeInternals (InternalsSync, node mount effects) settle.
+          // Extra frames after measurements appear let updateNodeInternals
+          // (InternalsSync, node mount effects) settle.
           if (stableCount >= FIT_VIEW_STABLE_FRAMES) {
             doFit();
             return;
           }
         } else {
-          // Measurements were reset (e.g. by updateNodeInternals) — restart
-          // the stability counter.
+          // Measurements reset (e.g. by updateNodeInternals): restart counter.
           stableCount = 0;
         }
         retries++;
@@ -808,10 +805,9 @@ export function RecipeStudioPage({
                 updateConfig={updateConfig}
                 onRun={() => {
                   // Easy mode is a full run (artifact persisted, tracked in
-                  // the Runs pane) capped at the user's row count. runFull
-                  // requires a non-empty fullRunName; the effect below
-                  // populates one on mount so the closure in runFull is
-                  // already up to date by the time the user clicks Run.
+                  // Runs) capped at the user's row count. runFull requires a
+                  // non-empty fullRunName; the effect above populates one on
+                  // mount so runFull's closure is current by click time.
                   void runFull();
                 }}
                 runLoading={fullLoading || executionLocked}
diff --git a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
index 52ce1329bf..8659cad4fb 100644
--- a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
+++ b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
@@ -756,10 +756,9 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         configs = applyRenameToConfigs(configs, oldName, newName);
       }
 
-      // When a provider toggles between local and external, keep already
-      // linked model_config nodes in sync. applyRenameToConfigs above has
-      // already propagated any name change, so providerName here is the
-      // post-rename value.
+      // When a provider toggles local/external, keep linked model_config nodes
+      // in sync. applyRenameToConfigs above already propagated any name change,
+      // so providerName here is the post-rename value.
       if (current.kind === "model_provider" && next.kind === "model_provider") {
         const prevIsLocal = current.is_local === true;
         const nextIsLocal = next.is_local === true;
diff --git a/studio/frontend/src/features/recipe-studio/utils/graph/fit-view.ts b/studio/frontend/src/features/recipe-studio/utils/graph/fit-view.ts
index a9da180db4..0ee541e02d 100644
--- a/studio/frontend/src/features/recipe-studio/utils/graph/fit-view.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/graph/fit-view.ts
@@ -27,12 +27,9 @@ function isAuxNode(node: Node): boolean {
 /**
  * Returns the primary workflow nodes that fitView should target.
  *
- * Excludes markdown notes and aux (LLM input overlay) nodes so the viewport
- * is framed around the primary workflow blocks. Falls back to all nodes if
- * filtering would leave an empty set.
- *
- * The returned array contains full {@link Node} objects so callers can inspect
- * `node.measured` without a second lookup pass.
+ * Excludes markdown notes and aux (LLM input overlay) nodes; falls back to
+ * all nodes if filtering would leave an empty set. Returns full {@link Node}
+ * objects so callers can inspect `node.measured` without a second lookup.
  */
 export function getFitViewTargetNodes(nodes: Node[]): Node[] {
   const primary = nodes.filter(
@@ -42,9 +39,9 @@ export function getFitViewTargetNodes(nodes: Node[]): Node[] {
 }
 
 /**
- * Builds a standard {@link FitViewOptions} object targeting the primary
- * workflow nodes. Every call site that invokes `fitView` should go through
- * this helper so zoom, padding, and node filtering stay consistent.
+ * Builds a standard {@link FitViewOptions} targeting the primary workflow
+ * nodes. All `fitView` call sites go through this so zoom, padding, and
+ * node filtering stay consistent.
  */
 export function buildFitViewOptions(
   nodes: Node[],
diff --git a/studio/frontend/src/features/recipe-studio/utils/graph/recipe-graph-connection.ts b/studio/frontend/src/features/recipe-studio/utils/graph/recipe-graph-connection.ts
index 059ddf186f..25e4d848c2 100644
--- a/studio/frontend/src/features/recipe-studio/utils/graph/recipe-graph-connection.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/graph/recipe-graph-connection.ts
@@ -387,10 +387,10 @@ export function applyRecipeConnection(
     nextBaseEdges,
   );
   if (source.kind === "model_provider" && target.kind === "model_config") {
-    // Keep model_config.provider in sync when a graph drag changes the link.
-    // Local providers now require an explicit selected load id; do not synthesize
-    // the legacy "local" placeholder. External relinks clear local-only GGUF
-    // metadata, while legacy placeholders are normalized back to empty.
+    // Keep model_config.provider in sync when a drag changes the link.
+    // Local providers need an explicit load id; don't synthesize the legacy
+    // "local" placeholder. External relinks clear local-only GGUF metadata;
+    // legacy placeholders normalize back to empty.
     const isSourceLocal = source.is_local === true;
     const isLegacyLocalPlaceholder =
       target.model.trim().toLowerCase() === "local";
diff --git a/studio/frontend/src/features/recipe-studio/utils/layout.ts b/studio/frontend/src/features/recipe-studio/utils/layout.ts
index 6b357c9028..3b76900e73 100644
--- a/studio/frontend/src/features/recipe-studio/utils/layout.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/layout.ts
@@ -17,10 +17,7 @@ type LayoutOptions = {
   configs?: Record<string, NodeConfig>;
 };
 
-/**
- * Pipeline rank order used to enforce a logical flow even for disconnected nodes.
- * Lower rank = earlier in the pipeline.
- */
+/** Pipeline rank for logical flow; lower = earlier. */
 function getPipelineRank(config: NodeConfig | undefined): number {
   if (!config) {
     return 2;
@@ -65,17 +62,15 @@ function getEdgeWeight(edgeType: string | undefined): number {
 
 /**
  * Build phantom edges between disconnected data-pipeline nodes so dagre
- * respects the pipeline rank order even when blocks aren't wired together.
- *
- * Groups nodes by rank, then inserts invisible edges from the last node of
- * rank N to the first node of rank N+1 when no real edge already connects them.
+ * respects pipeline rank order even when blocks aren't wired together.
+ * Groups by rank, then inserts invisible edges from the last node of rank N
+ * to the first of rank N+1 when no real edge already connects them.
  */
 function buildPhantomEdges(
   nodes: Node[],
   edges: Edge[],
   configs: Record<string, NodeConfig>,
 ): Edge[] {
-  // Group nodes by rank
   const byRank = new Map<number, string[]>();
   for (const node of nodes) {
     const rank = getPipelineRank(configs[node.id]);
@@ -94,7 +89,6 @@ function buildPhantomEdges(
       continue;
     }
 
-    // Check if any real edge already connects these rank groups
     const hasRealEdge = edges.some(
       (e) => currentIds.includes(e.source) && nextIds.includes(e.target),
     );
@@ -102,7 +96,7 @@ function buildPhantomEdges(
       continue;
     }
 
-    // Insert one phantom edge from last node in current rank to first in next
+    // One phantom edge: last node in current rank to first in next.
     phantoms.push({
       id: `phantom-${ranks[i]}-${ranks[i + 1]}`,
       source: currentIds[currentIds.length - 1],
@@ -129,7 +123,7 @@ export function getLayoutedElements<TNode extends Node>(
     configs,
   } = options;
 
-  // When configs are provided, filter out infra and aux nodes from dagre
+  // With configs, filter infra and aux nodes out of dagre.
   const hasConfigs = configs && Object.keys(configs).length > 0;
   const dataNodes = hasConfigs
     ? nodes.filter((n) => !(isInfraNode(n.id, configs) || isAuxNode(n.id)))
@@ -146,7 +140,7 @@ export function getLayoutedElements<TNode extends Node>(
       )
     : edges;
 
-  // Build phantom edges to enforce pipeline rank ordering for disconnected nodes
+  // Phantom edges enforce rank ordering for disconnected nodes.
   const phantomEdges = hasConfigs
     ? buildPhantomEdges(dataNodes, dataEdges, configs)
     : [];
@@ -175,7 +169,7 @@ export function getLayoutedElements<TNode extends Node>(
 
   dagre.layout(graph);
 
-  // Build position map from dagre results (data nodes only)
+  // Position map from dagre results (data nodes only).
   const layoutedPositions = new Map<string, { x: number; y: number }>();
   for (const node of dataNodes) {
     const pos = graph.node(node.id);
@@ -187,7 +181,7 @@ export function getLayoutedElements<TNode extends Node>(
     });
   }
 
-  // Apply positions: data nodes get dagre positions, infra/aux keep original
+  // Data nodes get dagre positions; infra/aux keep original.
   const layoutedNodes = nodes.map((node) => {
     const position = layoutedPositions.get(node.id);
     if (!position) {
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/builders-model.ts b/studio/frontend/src/features/recipe-studio/utils/payload/builders-model.ts
index 1575919705..0a85ef618f 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/builders-model.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/builders-model.ts
@@ -8,11 +8,10 @@ export function buildModelProvider(
   config: ModelProviderConfig,
   errors: string[],
 ): Record<string, unknown> {
-  // Local providers do not use any of the advanced request overrides -
-  // the backend overrides endpoint/api_key/provider_type and strips the
-  // extra fields in _inject_local_providers. Skip parsing the hidden
-  // JSON inputs here so imported or hydrated recipes with stale extra
-  // headers/body cannot block the client-side validation step.
+  // Local providers ignore advanced request overrides: the backend overrides
+  // endpoint/api_key/provider_type and strips extras in _inject_local_providers.
+  // Skip parsing hidden JSON inputs so stale headers/body in imported recipes
+  // can't block client-side validation.
   if (config.is_local === true) {
     return {
       name: config.name,
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/builders-sampler.ts b/studio/frontend/src/features/recipe-studio/utils/payload/builders-sampler.ts
index 596450d1e1..70e345f3b3 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/builders-sampler.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/builders-sampler.ts
@@ -134,7 +134,7 @@ function buildSamplerParams(
     if (!raw) {
       return {};
     }
-    // UI historically used "uuid4" as a "format". data_designer uuid sampler is always uuid4.
+    // UI historically used "uuid4" as a format; data_designer uuid is always uuid4.
     if (raw.toLowerCase() === "uuid4") {
       return {};
     }
diff --git a/studio/frontend/src/features/settings/components/create-key-form.tsx b/studio/frontend/src/features/settings/components/create-key-form.tsx
index 93802bfad8..c375bc280c 100644
--- a/studio/frontend/src/features/settings/components/create-key-form.tsx
+++ b/studio/frontend/src/features/settings/components/create-key-form.tsx
@@ -36,9 +36,8 @@ export function CreateKeyForm({
       onCreated(result.key);
       setName("");
     } catch {
-      // API helpers in ../api/api-keys.ts throw generic English Error
-      // messages; always use the translated message so zh-CN users do not
-      // see English text bleed through from internal exceptions.
+      // api-keys.ts throws English Error messages; use the translated one so
+      // zh-CN users don't see English bleed through from internal exceptions.
       onError(t("settings.apiKeys.createError"));
     } finally {
       setLoading(false);
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index a1048b202e..26efcde7ea 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -116,21 +116,20 @@ export function SettingsDialog() {
         showCloseButton={false}
         overlayClassName="bg-background/40"
         onCloseAutoFocus={(e) => {
-          // Restore focus to the element that triggered openDialog().
-          // Radix's FocusScope races our rAF-scheduled tab-button focus
-          // and loses the previous-focus reference, so we restore by hand.
+          // Restore focus to the element that triggered openDialog(). Radix's
+          // FocusScope races our rAF-scheduled tab focus and loses the
+          // previous-focus reference, so restore it by hand.
           if (opener && opener.isConnected) {
             e.preventDefault();
             opener.focus({ preventScroll: true });
           }
         }}
         className={cn(
-          // Cap at 820px but shrink to the viewport so we don't clip
-          // on iPad-portrait widths (640-820px) where the fixed
-          // `w-[820px]` overflows by 26px on each side.
+          // Cap at 820px but shrink to the viewport so it doesn't clip on
+          // iPad-portrait widths (640-820px) where fixed `w-[820px]` overflows.
           "settings-surface !max-w-[min(820px,calc(100vw-2rem))] h-[560px] w-[min(820px,calc(100vw-2rem))] p-0 overflow-hidden",
-          // Soft shadow only, no outline ring. Pin --radius to the light value
-          // so the corner rounding is the same in dark mode.
+          // Soft shadow, no outline ring. Pin --radius to the light value so
+          // corner rounding matches in dark mode.
           "shadow-border rounded-xl ring-0 [--radius:1.1rem]",
           "max-sm:h-dvh max-sm:w-dvw max-sm:!max-w-none max-sm:rounded-none",
         )}
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 7ac5422ff1..1f4ac9c372 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -15,10 +15,10 @@ export type SettingsTab =
 interface SettingsDialogState {
   open: boolean;
   activeTab: SettingsTab;
-  // Element focused at the moment openDialog() ran. Radix's FocusScope
-  // would normally track this, but the rAF-scheduled focus() in
-  // settings-dialog.tsx races its previous-focus capture, leaving focus
-  // on <body> after close. We restore explicitly via onCloseAutoFocus.
+  // Element focused when openDialog() ran. Radix's FocusScope normally tracks
+  // this, but the rAF-scheduled focus() in settings-dialog.tsx races its
+  // previous-focus capture, leaving focus on <body> after close. We restore
+  // explicitly via onCloseAutoFocus.
   opener: HTMLElement | null;
   openDialog: (tab?: SettingsTab) => void;
   closeDialog: () => void;
diff --git a/studio/frontend/src/features/settings/stores/theme-store.ts b/studio/frontend/src/features/settings/stores/theme-store.ts
index f73fc51e1c..9d05582c77 100644
--- a/studio/frontend/src/features/settings/stores/theme-store.ts
+++ b/studio/frontend/src/features/settings/stores/theme-store.ts
@@ -32,9 +32,8 @@ function resolveTheme(theme: Theme): ResolvedTheme {
 
 function applyToDocument(resolved: ResolvedTheme) {
   if (typeof document === "undefined") return;
-  // Keep "dark"/"light" mutually exclusive. next-themes (via Sonner)
-  // adds "light" on first mount; without the explicit toggle we'd end
-  // up with `class="light dark"` after a switch.
+  // Keep "dark"/"light" mutually exclusive: next-themes (via Sonner) adds
+  // "light" on first mount, so without this toggle we'd get "light dark".
   const cl = document.documentElement.classList;
   cl.toggle("dark", resolved === "dark");
   cl.toggle("light", resolved === "light");
@@ -72,15 +71,14 @@ function getServerSnapshot(): Theme {
 }
 
 /**
- * Single source of truth for setting the theme. All writers (the Settings
- * dialog's segmented control AND the sidebar dropdown's animated toggler)
- * must route through this so the DOM class, localStorage, and React
- * subscribers stay in sync.
+ * Single source of truth for setting the theme. All writers (Settings dialog
+ * control, sidebar dropdown toggler) route through this so the DOM class,
+ * localStorage, and React subscribers stay in sync.
  */
 export function setTheme(next: Theme): void {
   if (typeof window === "undefined") return;
-  // Persist "system" explicitly so next-themes (mounted with
-  // defaultTheme="light") doesn't clobber the choice on reload.
+  // Persist "system" explicitly so next-themes (defaultTheme="light")
+  // doesn't clobber the choice on reload.
   try {
     window.localStorage.setItem(STORAGE_KEY, next);
   } catch {
diff --git a/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx b/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
index d562b585b1..601a81b799 100644
--- a/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
@@ -32,9 +32,8 @@ export function ApiKeysTab() {
     ? { duration: 0 }
     : { duration: 0.18, ease: [0.165, 0.84, 0.44, 1] as const };
 
-  // API helpers in ../api/api-keys.ts throw generic English Error messages
-  // ("Failed to load API access", etc.). Always use the translated message
-  // so zh-CN users do not see those English strings bleed through.
+  // ../api/api-keys.ts throws generic English errors; use the translated
+  // message so zh-CN users don't see English strings bleed through.
   const load = useCallback(async () => {
     setLoading(true);
     setError(null);
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index aaaf0a7115..57a52b83d1 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -30,15 +30,10 @@ import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 
 // Keys cleared by "Reset all local preferences".
-//
-// NEVER include auth / session keys here — resetting them would log the user
-// out, which is not what users expect from a "reset preferences" button.
-//
-// Explicitly EXCLUDED:
-//   - "unsloth_auth_token"                 (auth: access token)
-//   - "unsloth_auth_refresh_token"         (auth: refresh token)
-//   - "unsloth_auth_must_change_password"  (auth: forced password change flag)
-//   - "unsloth_onboarding_done"            (session: would force re-onboarding)
+// NEVER include auth/session keys here — clearing them would log the user out
+// or force re-onboarding. Explicitly excluded: unsloth_auth_token,
+// unsloth_auth_refresh_token, unsloth_auth_must_change_password,
+// unsloth_onboarding_done.
 const PREFS_KEYS: string[] = [
   // Appearance
   "theme",
@@ -70,10 +65,8 @@ const PREFS_KEYS: string[] = [
   "tour:studio:v1",
 ];
 
-// Set to true from resetAllPrefs so the unmount-commit effect skips writing
-// back the in-memory draft — otherwise the cleanup would re-persist the old
-// HF token into localStorage after it was just cleared, and the subsequent
-// reload would read the re-written value.
+// Set by resetAllPrefs so the unmount-commit effect skips writing back the
+// in-memory draft, else cleanup would re-persist the just-cleared HF token.
 let resetInProgress = false;
 
 function resetAllPrefs() {
diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx
index 504e251bc5..e958f2ba0d 100644
--- a/studio/frontend/src/features/studio/historical-training-view.tsx
+++ b/studio/frontend/src/features/studio/historical-training-view.tsx
@@ -95,7 +95,7 @@ export function HistoricalTrainingView({
   const [detail, setDetail] = useState<TrainingRunDetailResponse | null>(null);
   const [error, setError] = useState<string | null>(null);
 
-  // Derive loading from detail/error -- no separate state needed
+  // Derive loading from detail/error; no separate state.
   const loading = detail === null && error === null;
 
   useEffect(() => {
@@ -114,7 +114,7 @@ export function HistoricalTrainingView({
       });
     return () => {
       controller.abort();
-      // Reset on runId change so loading derives correctly for the next fetch
+      // Reset on runId change so loading derives correctly for the next fetch.
       setDetail(null);
       setError(null);
     };
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index 75ec43f20d..dbe7fff01b 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -125,7 +125,7 @@ function Sparkline({
   const w = 120;
   const gradientId = `sparkFill-${id}`;
 
-  // Build points with vertical padding so the stroke isn't clipped
+  // Vertical padding so the stroke isn't clipped.
   const pts = values.map((v, i) => ({
     x: (i / (values.length - 1)) * w,
     y: pad + (1 - (v - min) / range) * (h - pad * 2),
@@ -206,7 +206,7 @@ export function HistoryCardGrid({
   }, [runs.length]);
 
   const fetchRuns = useCallback(async (offset = 0, append = false, limit = PAGE_SIZE) => {
-    // Cancel any in-flight poll so its stale response can't clobber this fresher fetch
+    // Cancel any in-flight poll so its stale response can't clobber this fetch.
     pollControllerRef.current?.abort();
     userControllerRef.current?.abort();
     const controller = new AbortController();
@@ -261,7 +261,7 @@ export function HistoryCardGrid({
     };
   }, [fetchRuns]);
 
-  // Poll while any run is still "running" so the card shows live progress
+  // Poll while any run is "running" so cards show live progress.
   const hasRunningRun = runs.some((r) => r.status === "running");
   const visibleCount = runs.length;
   useEffect(() => {
@@ -275,11 +275,11 @@ export function HistoryCardGrid({
       try {
         const limit = Math.max(PAGE_SIZE, visibleCount);
         const result = await listTrainingRuns(limit, 0, controller.signal);
-        if (pollIdRef.current !== pid) return; // stale poll — discard
+        if (pollIdRef.current !== pid) return; // stale poll
         setRuns(result.runs);
         setTotal(result.total);
       } catch {
-        // silently handle — poll will retry
+        // ignore; poll will retry
       }
     }, RUNNING_POLL_INTERVAL_MS);
     return () => {
@@ -294,11 +294,11 @@ export function HistoryCardGrid({
     try {
       await deleteTrainingRun(deleteTarget);
       emitTrainingRunDeleted(deleteTarget);
-      // Re-fetch preserving visible count so offsets stay consistent for "Load more"
+      // Re-fetch preserving visible count so "Load more" offsets stay consistent.
       const currentCount = runs.length - 1;
       const limit = Math.max(PAGE_SIZE, currentCount);
       fetchRuns(0, false, limit).catch(() => {
-        // Refresh failed — card is already removed, no stale display
+        // Refresh failed; card is already removed, no stale display.
       });
     } catch {
       setDeleteError(translate("studio.history.deleteError"));
diff --git a/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx b/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
index 4a99c4f702..4f9ea2c98f 100644
--- a/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
@@ -303,8 +303,8 @@ const FROM_CANONICAL: Record<string, Record<string, string>> = {
 };
 
 /**
- * Remap a column→role mapping between formats.
- * Normalises every role to canonical chatml first, then maps to the target format.
+ * Remap a column→role mapping between formats: normalise to canonical chatml,
+ * then map to the target format.
  */
 export function remapRolesForFormat(
   mapping: Record<string, string>,
diff --git a/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx b/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
index d05a8cf242..6710099400 100644
--- a/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
@@ -86,8 +86,7 @@ export function DatasetPreviewDialog({
   );
   const { isStarting, startError, startTrainingRun } = useTrainingActions();
 
-  // If the backend reports image data, treat as VLM even if the prop
-  // hasn't caught up yet (isDatasetImage may still be null in the store).
+  // Treat backend-reported image data as VLM even if the prop hasn't caught up.
   const effectiveIsAudio = !!data?.is_audio;
   const effectiveIsVlm = isVlm || !!data?.is_image;
 
@@ -119,7 +118,7 @@ export function DatasetPreviewDialog({
       });
 
       if (result.success && result.suggested_mapping) {
-        // Remap from chatml roles (user/assistant/system) to format-specific roles
+        // Remap chatml roles to format-specific roles
         const table = ROLE_REMAP[datasetFormat];
         const mapped: Record<string, string> = {};
         for (const [col, role] of Object.entries(result.suggested_mapping)) {
@@ -155,14 +154,12 @@ export function DatasetPreviewDialog({
     setManualMapping(remapRolesForFormat(manualMapping, datasetFormat));
   }, [datasetFormat]); // eslint-disable-line react-hooks/exhaustive-deps
 
-  // Handle role change for a column
   const handleRoleChange = useCallback(
     (colName: string, role: string | undefined) => {
       const next = { ...manualMapping };
-      // Remove this column's previous role
       delete next[colName];
       if (role) {
-        // Remove any other column that had this role (each role can only be assigned once)
+        // Each role maps to one column, so drop any other column holding it
         for (const [col, r] of Object.entries(next)) {
           if (r === role) delete next[col];
         }
@@ -230,7 +227,6 @@ export function DatasetPreviewDialog({
   const rows = data?.preview_samples ?? [];
   const columns = data?.columns ?? [];
 
-  // Determine source label
   const sourceLabel = useMemo(() => {
     if (!datasetName) return "";
     if (datasetSource === "huggingface") {
@@ -243,7 +239,6 @@ export function DatasetPreviewDialog({
     return `Local Files (${datasetName})`;
   }, [datasetName, datasetSource, datasetSubset, datasetSplit]);
 
-  // Build TanStack Table columns from the column names
   const tableColumns = useMemo<ColumnDef<Record<string, unknown>>[]>(() => {
     if (!columns.length) return [];
 
@@ -507,10 +502,7 @@ export function DatasetPreviewDialog({
   );
 }
 
-// ---------------------------------------------------------------------------
 // Metadata row
-// ---------------------------------------------------------------------------
-
 function MetaRow({
   label,
   value,
diff --git a/studio/frontend/src/features/studio/sections/model-section.tsx b/studio/frontend/src/features/studio/sections/model-section.tsx
index 055437a1f5..cc5717692a 100644
--- a/studio/frontend/src/features/studio/sections/model-section.tsx
+++ b/studio/frontend/src/features/studio/sections/model-section.tsx
@@ -233,12 +233,8 @@ export function ModelSection() {
     });
   }, [localMetaById, localModelInput, localResultIds]);
 
-  // Pre-compute VRAM fit status for every model in the current result set.
-  // Keyed by model id so the render callback is a simple O(1) lookup.
-  //
-  // Pre-compute VRAM fit status for every model in the current result set.
-  // Keyed by model id so the render callback is a simple O(1) lookup.
-  // Re-computes when the training method changes (QLoRA=4-bit vs LoRA/Full=fp16).
+  // VRAM fit status per model, keyed by id for O(1) render lookups.
+  // Recomputes on training-method change (QLoRA=4-bit vs LoRA/Full=fp16).
   const vramMap = useMemo(() => {
     const fitMap = buildModelVramMap(
       hfResults,
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index f52f3bcb99..67800df27b 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -122,7 +122,7 @@ export function ProgressSection({
   const [stopDialogOpen, setStopDialogOpen] = useState(false);
   const [stopRequestedLocal, setStopRequestedLocal] = useState(false);
 
-  // Auto-reset when training stops -- no useEffect needed
+  // Auto-resets when training stops; no useEffect needed
   const stopRequested = data.isTrainingRunning && stopRequestedLocal;
 
   const pct =
diff --git a/studio/frontend/src/features/studio/sections/training-section.tsx b/studio/frontend/src/features/studio/sections/training-section.tsx
index adeb1a7b14..d1ef132efb 100644
--- a/studio/frontend/src/features/studio/sections/training-section.tsx
+++ b/studio/frontend/src/features/studio/sections/training-section.tsx
@@ -81,15 +81,14 @@ export function TrainingSection() {
   };
 
   const handleSaveConfig = () => {
-    // isDatasetImage is null in three windows: before a dataset check
-    // completes, after dataset edits, and on import. Treat all three as
-    // "save it" so the user's choice is never silently dropped while we
-    // wait to confirm the dataset type. Only a confirmed text-only dataset
+    // isDatasetImage is null before a dataset check completes, after edits, and
+    // on import. Treat all three as "save it" so the user's choice isn't dropped
+    // while the type is unconfirmed. Only a confirmed text-only dataset
     // (=== false) suppresses the vision fields.
     const includeVisionFields =
       store.isVisionModel && store.isDatasetImage !== false;
-    // DeepSeek OCR ignores vision_image_size; don't emit it to YAML either,
-    // or a later import on a non-DeepSeek model would activate the stale value.
+    // DeepSeek OCR ignores vision_image_size; don't emit it, or a later import
+    // on a non-DeepSeek model would activate the stale value.
     const selectedModelLower = (store.selectedModel ?? "").toLowerCase();
     const isDeepseekOcr =
       selectedModelLower.includes("deepseek") &&
diff --git a/studio/frontend/src/features/studio/studio-page.tsx b/studio/frontend/src/features/studio/studio-page.tsx
index d5368fba58..4635a6ce0e 100644
--- a/studio/frontend/src/features/studio/studio-page.tsx
+++ b/studio/frontend/src/features/studio/studio-page.tsx
@@ -61,9 +61,9 @@ export function StudioPage(): ReactElement {
     return () => setSelectedHistoryRunId(null);
   }, [setSelectedHistoryRunId]);
 
-  // Derive activeTab: auto-switch to "current-run" only while training is
-  // genuinely running.  Once training ends, honour whatever tab the user clicks.
-  // If requestedTab is "current-run" but there's nothing to show, fall back to "configure".
+  // Auto-switch to "current-run" only while training runs; afterward honour
+  // the user's clicked tab. If "current-run" has nothing to show, use
+  // "configure".
   const activeTab =
     isTrainingRunning && requestedTab !== "history"
       ? "current-run"
diff --git a/studio/frontend/src/features/studio/training-start-overlay.tsx b/studio/frontend/src/features/studio/training-start-overlay.tsx
index 0205394a53..ba80432896 100644
--- a/studio/frontend/src/features/studio/training-start-overlay.tsx
+++ b/studio/frontend/src/features/studio/training-start-overlay.tsx
@@ -38,9 +38,8 @@ import { useT } from "@/i18n";
 const HF_REPO_REGEX = /^[A-Za-z0-9._-]+\/[A-Za-z0-9._-]+$/;
 
 // Tracks which jobs have already played the terminal intro animation. The
-// overlay unmounts when you navigate away from the training page, so without
-// this its typing/fade-in would replay on every return even though the run
-// itself is still going. Module-level so it survives remounts.
+// overlay unmounts on navigation away, so without this its typing/fade-in would
+// replay on every return mid-run. Module-level so it survives remounts.
 const animatedJobs = new Set<string>();
 
 function formatBytes(n: number): string {
@@ -90,12 +89,9 @@ function coerceCachedStateReady(state: DownloadState): DownloadState {
 type Fetcher = (repoId: string) => Promise<DownloadProgressResponse>;
 
 /**
- * Polls a HF repo's download progress on a 1.5s tick. Used for both
- * model weights (`/api/models/download-progress`) and dataset blobs
- * (`/api/datasets/download-progress`) by swapping the fetcher.
- *
- * Stops polling once `progress >= 1.0` -- the bar freezes at the final
- * value rather than disappearing, mirroring the existing chat flow.
+ * Polls a HF repo's download progress on a 1.5s tick. Serves both model weights
+ * and dataset blobs by swapping the fetcher. Stops once `progress >= 1.0`; the
+ * bar freezes at the final value rather than disappearing, matching chat flow.
  */
 function useHfDownloadProgress(
   repoId: string | null,
@@ -179,9 +175,8 @@ type DownloadRowProps = {
 
 function DownloadRow({ label, state }: DownloadRowProps): ReactElement | null {
   const t = useT();
-  // Compute a rolling-window rate + ETA from the same cumulative-byte
-  // series the poll hook already produces, so we can show
-  // "5.2 / 20.7 GB • 85.3 MB/s • 3m 12s left" instead of just the pair.
+  // Rolling-window rate + ETA from the cumulative-byte series the poll hook
+  // produces, so we show "5.2 / 20.7 GB • 85.3 MB/s • 3m 12s left", not just the pair.
   const stats = useTransferStats(state.downloadedBytes, state.totalBytes);
 
   if (state.downloadedBytes <= 0 && !state.cachePath) return null;
@@ -267,8 +262,8 @@ export function TrainingStartOverlay({
   const configuredModel = useTrainingConfigStore((s) => s.selectedModel);
   const datasetSource = useTrainingConfigStore((s) => s.datasetSource);
   const dataset = useTrainingConfigStore((s) => s.dataset);
-  // Only HF datasets have a download phase to track. Uploaded files are
-  // already on disk by the time the overlay shows up.
+  // Only HF datasets have a download phase to track; uploaded files are already
+  // on disk by the time the overlay shows up.
   const hfDatasetName = datasetSource === "huggingface" ? dataset : null;
   const hasStartResources = startModelName !== null;
   const useConfiguredResources = !isStarting && !hasStartResources;
@@ -305,9 +300,8 @@ export function TrainingStartOverlay({
     }
   }, [isStarting]);
 
-  // Play the intro animation only the first time we mount for a given job.
-  // On later remounts (e.g. leaving the training page and coming back) the
-  // terminal renders its final state instantly so the logs don't restart.
+  // Play the intro animation only on the first mount per job. On later remounts
+  // the terminal renders its final state instantly so the logs don't restart.
   const alreadyAnimated = jobId != null && animatedJobs.has(jobId);
   useEffect(() => {
     if (jobId != null) {
diff --git a/studio/frontend/src/features/tour/lib/confetti-fireworks.ts b/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
index 1904436c69..6eaf8e031e 100644
--- a/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
+++ b/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
@@ -11,10 +11,9 @@ type FireworksOpts = {
   zIndex?: number;
 };
 
-// CSP blocks canvas-confetti's default blob: worker, so reuse a single
-// overlay canvas via `confetti.create(..., { useWorker: false })`.
-// Caller-provided canvases ignore the per-fire `zIndex`; stacking is
-// driven by `_sharedCanvas.style.zIndex` instead (set in fireConfettiFireworks).
+// CSP blocks canvas-confetti's default blob: worker, so reuse a single overlay
+// canvas via `confetti.create(..., { useWorker: false })`. Caller canvases
+// ignore per-fire `zIndex`; stacking is driven by `_sharedCanvas.style.zIndex`.
 const DEFAULT_FIREWORKS_Z_INDEX = 99999;
 let _sharedCanvas: HTMLCanvasElement | null = null;
 let _sharedFire: ConfettiInstance | null = null;
diff --git a/studio/frontend/src/features/training/api/models-api.ts b/studio/frontend/src/features/training/api/models-api.ts
index 8fecfaf7b8..1e9a135fd3 100644
--- a/studio/frontend/src/features/training/api/models-api.ts
+++ b/studio/frontend/src/features/training/api/models-api.ts
@@ -92,10 +92,7 @@ interface LocalModelListResponse {
   models: LocalModelInfo[];
 }
 
-/**
- * Check whether a model is a vision model by asking the backend.
- * Calls GET /api/models/check-vision/{model_name}.
- */
+/** Ask the backend whether a model is a vision model (GET /api/models/check-vision/{model_name}). */
 export async function checkVisionModel(modelName: string): Promise<boolean> {
   const encoded = encodeURIComponent(modelName);
   const response = await authFetch(`/api/models/check-vision/${encoded}`);
@@ -107,10 +104,7 @@ export async function checkVisionModel(modelName: string): Promise<boolean> {
   return data.is_vision;
 }
 
-/**
- * Check whether a model is an embedding model by asking the backend.
- * Calls GET /api/models/check-embedding/{model_name}.
- */
+/** Ask the backend whether a model is an embedding model (GET /api/models/check-embedding/{model_name}). */
 export async function checkEmbeddingModel(
   modelName: string,
 ): Promise<boolean> {
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 3d2cb64022..8069ff9bbf 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -75,8 +75,8 @@ export function useTrainingActions() {
           isVlm,
         });
 
-        // Backend auto-detects image/audio from dataset content.
-        // Sync these flags into the store so buildTrainingStartPayload picks them up.
+        // Backend auto-detects image/audio from dataset content; sync the flags
+        // into the store so buildTrainingStartPayload picks them up.
         const isAudio = !!check.is_audio;
         const isImage = !!check.is_image;
 
diff --git a/studio/frontend/src/features/training/hooks/use-training-history-sidebar.ts b/studio/frontend/src/features/training/hooks/use-training-history-sidebar.ts
index 92ffd68466..6569214c40 100644
--- a/studio/frontend/src/features/training/hooks/use-training-history-sidebar.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-history-sidebar.ts
@@ -60,9 +60,8 @@ export function useTrainingHistorySidebarItems(enabled: boolean) {
     }
   }, [fetchRuns]);
 
-  // Initial load: bounded retry-with-backoff, then surface a toast on
-  // final failure with a Retry action so the user isn't stuck staring
-  // at an empty sidebar after F5 if the backend was slow to come up.
+  // Initial load: bounded retry-with-backoff, then a toast with Retry on final
+  // failure so the user isn't stuck on an empty sidebar if the backend is slow.
   useEffect(() => {
     if (!enabled) {
       return;
@@ -124,10 +123,9 @@ export function useTrainingHistorySidebarItems(enabled: boolean) {
     };
   }, [enabled, fetchRuns]);
 
-  // Poll while there's a running run, but only when the tab is visible.
-  // Browsers throttle background timers but don't pause them — gating on
-  // visibility avoids hammering the API for tabs left open in the
-  // background, which is common during long training runs.
+  // Poll while a run is active, but only when the tab is visible. Browsers
+  // throttle background timers but don't pause them, so gating on visibility
+  // avoids hammering the API for tabs left open during long training runs.
   const hasRunning = items.some((r) => r.status === "running");
   useEffect(() => {
     if (!enabled || !hasRunning) {
diff --git a/studio/frontend/src/features/training/hooks/use-training-unload-guard.ts b/studio/frontend/src/features/training/hooks/use-training-unload-guard.ts
index 1808cf5888..879200ade3 100644
--- a/studio/frontend/src/features/training/hooks/use-training-unload-guard.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-unload-guard.ts
@@ -31,10 +31,9 @@ export function useTrainingUnloadGuard() {
 }
 
 /**
- * Removes the active beforeunload guard (if any).
- * Call this before intentionally ending the session (e.g. shutting down
- * the Studio server) so the "Server stopped" page can render without
- * the browser prompting the user to confirm leaving.
+ * Removes the active beforeunload guard (if any). Call before intentionally
+ * ending the session (e.g. shutting down the server) so the "Server stopped"
+ * page renders without the browser prompting to confirm leaving.
  */
 export function removeTrainingUnloadGuard() {
   if (currentHandler) {
diff --git a/studio/frontend/src/features/training/lib/yaml-config.ts b/studio/frontend/src/features/training/lib/yaml-config.ts
index 1deabf3547..d25f5a56f8 100644
--- a/studio/frontend/src/features/training/lib/yaml-config.ts
+++ b/studio/frontend/src/features/training/lib/yaml-config.ts
@@ -27,10 +27,10 @@ export function parseYamlConfig(text: string): BackendModelConfig {
     console.warn("Ignored unknown YAML keys:", unknownKeys.join(", "));
   }
 
-  // File import is authoritative: forge vision_image_size = null when the
-  // training section is missing, malformed, or missing the key, so a stale
-  // store value cannot survive an import. (Same-model defaults reloads
-  // preserve user choice via Object.hasOwn in model-defaults.ts.)
+  // File import is authoritative: force vision_image_size = null when the
+  // training section is missing/malformed/lacks the key, so a stale store
+  // value can't survive an import. (Same-model defaults reloads preserve
+  // user choice via Object.hasOwn in model-defaults.ts.)
   const rawTraining = raw.training;
   const isPlainTrainingObject =
     rawTraining != null &&
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 137669f8b8..7de78096f9 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -19,11 +19,8 @@ const MIN_STEP: StepNumber = 1;
 const MAX_STEP: StepNumber = STEPS.length as StepNumber;
 
 /**
- * Auto-select LoRA (16-bit) vs QLoRA (4-bit) based on model size and GPU memory.
- *
- * Rule: if model_size_gb * 1.5 * context_scale fits in free VRAM, use "lora" (16-bit).
- * Otherwise use "qlora" (4-bit).
- *
+ * Auto-select LoRA (16-bit) vs QLoRA (4-bit) by model size and GPU memory.
+ * Use "lora" if model_size_gb * 1.5 * context_scale fits in free VRAM, else "qlora".
  * Context scale: <=8192 = 1.0, >8192 = 1.7, >=16384 = 2.0, >=32768 = 4.0
  */
 async function autoSelectTrainingMethod(
@@ -97,17 +94,16 @@ let _datasetCheckController: AbortController | null = null;
 // AbortController for in-flight model default loads.
 let _modelConfigController: AbortController | null = null;
 
-// Track whether the user has manually toggled trainOnCompletions
-// since the last auto-set (model load or dataset change).
+// Has the user manually toggled trainOnCompletions since the last auto-set
+// (model load or dataset change)?
 let _trainOnCompletionsManuallySet = false;
 
-// Track whether the user has manually edited the learning rate
-// since the last model load. When false, switching training method
-// auto-sets LR to 2e-4 (LoRA/QLoRA) or 2e-5 (full fine-tune).
+// Has the user manually edited the LR since the last model load? When false,
+// switching method auto-sets LR to 2e-4 (LoRA/QLoRA) or 2e-5 (full fine-tune).
 let _learningRateManuallySet = false;
 
-// Stash the model-config-provided (YAML) learning rate so that
-// setTrainingMethod can restore it when switching back from full to adapter.
+// Stash the YAML learning rate so setTrainingMethod can restore it when
+// switching back from full to adapter.
 let _yamlLearningRate: number | undefined = undefined;
 
 // Track whether entering CPT auto-forced datasetFormat="raw" so that
@@ -303,20 +299,18 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             _yamlLearningRate = undefined;
             const patch = mapBackendModelConfigToTrainingPatch(modelDetails.config);
 
-            // If the model config provides a specific learning rate, treat
-            // it as authoritative so the async auto-select does not overwrite it.
+            // Treat a model-config LR as authoritative so async auto-select
+            // won't overwrite it.
             const modelConfigHasLR = patch.learningRate !== undefined;
             _yamlLearningRate = patch.learningRate;
 
-            // YAML learning rates are tuned for adapter methods (LoRA/QLoRA).
-            // If the user is currently on full fine-tune, override with the
-            // full-finetune default instead of applying the YAML adapter LR.
+            // YAML LRs are tuned for adapters (LoRA/QLoRA); on full fine-tune,
+            // use the full-finetune default instead of the YAML adapter LR.
             if (modelConfigHasLR && !isAdapterMethod(get().trainingMethod)) {
               patch.learningRate = LR_DEFAULT_FULL;
             }
 
-            // If vision model + image dataset already known, override
-            // trainOnCompletions to false regardless of backend default.
+            // Vision model + known image dataset: force trainOnCompletions off.
             if (modelDetails.is_vision && get().isDatasetImage === true) {
               patch.trainOnCompletions = false;
             }
@@ -331,17 +325,13 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               patch.trainOnCompletions = false;
             }
 
-            // Use backend-provided model_type when available, otherwise
-            // infer from capability flags.
+            // Use backend model_type when available, else infer from flags.
             const isEmbedding = !!modelDetails.is_embedding;
             const inferredModelType: ModelType = modelDetails.model_type
               ?? (isEmbedding ? "embeddings" : modelDetails.is_vision ? "vision" : modelDetails.is_audio ? "audio" : "text");
 
-            // Auto-select training method based on model size vs GPU memory.
-            // If model_size * 1.5 * context_scale fits in free VRAM, use LoRA 16-bit.
-            // Otherwise use QLoRA 4-bit.
-            // Auto-select LoRA vs QLoRA based on GPU memory.
-            // Skip if user has manually chosen CPT -- don't override it.
+            // Auto-select LoRA vs QLoRA by model size vs GPU memory (see
+            // autoSelectTrainingMethod). Skip if the user chose CPT.
             const modelSizeBytes = modelDetails.model_size_bytes;
             if (modelSizeBytes && modelSizeBytes > 0 && get().trainingMethod !== "cpt") {
               void autoSelectTrainingMethod(modelSizeBytes, patch.contextLength ?? get().contextLength)
@@ -358,7 +348,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             }
 
             // Preserve CPT hyperparams: YAML adapter defaults (r/alpha/targets/LR)
-            // are tuned for standard LoRA and would otherwise clobber CPT settings.
+            // are tuned for standard LoRA and would clobber CPT settings.
             const cptOverrides =
               get().trainingMethod === "cpt"
                 ? getCptModelDefaultsPatch()
diff --git a/studio/frontend/src/hooks/use-collapse-scroll-lock.ts b/studio/frontend/src/hooks/use-collapse-scroll-lock.ts
index fb7baa062c..5ad8853839 100644
--- a/studio/frontend/src/hooks/use-collapse-scroll-lock.ts
+++ b/studio/frontend/src/hooks/use-collapse-scroll-lock.ts
@@ -4,12 +4,12 @@
 import { type RefObject, useCallback, useEffect, useRef } from "react";
 
 /**
- * Locks the nearest scrollable ancestor's scrollTop for the duration of a
- * collapsible animation so the page doesn't jump when content height changes.
+ * Locks the nearest scrollable ancestor's scrollTop during a collapsible
+ * animation so the page doesn't jump when content height changes.
  *
- * Unlike @assistant-ui/react's `useScrollLock`, this hook does NOT toggle
- * `scrollbar-width: none` on the container. Hiding the scrollbar mid-animation
- * caused a visible disappear/reappear flicker on tool-call collapsibles.
+ * Unlike @assistant-ui/react's `useScrollLock`, this does NOT toggle
+ * `scrollbar-width: none` on the container; hiding the scrollbar mid-animation
+ * caused a visible flicker on tool-call collapsibles.
  */
 export function useCollapseScrollLock(
   animatedElementRef: RefObject<HTMLElement | null>,
diff --git a/studio/frontend/src/hooks/use-gpu-info.ts b/studio/frontend/src/hooks/use-gpu-info.ts
index a7ee416112..c7a2852067 100644
--- a/studio/frontend/src/hooks/use-gpu-info.ts
+++ b/studio/frontend/src/hooks/use-gpu-info.ts
@@ -54,10 +54,8 @@ async function fetchGpuOnce(): Promise<GpuInfo> {
 }
 
 /**
- * Fetch GPU info from the backend /api/system endpoint.
- *
- * The result is cached at module level -- only one network request is made
- * regardless of how many components call this hook.
+ * Fetch GPU info from /api/system. Cached at module level, so only one request
+ * is made no matter how many components call this hook.
  */
 export function useGpuInfo(): GpuInfo {
   const [gpu, setGpu] = useState<GpuInfo>(cachedGpu ?? DEFAULT_GPU);
diff --git a/studio/frontend/src/hooks/use-gpu-utilization.ts b/studio/frontend/src/hooks/use-gpu-utilization.ts
index 1262dcbc2f..1a9f6102dd 100644
--- a/studio/frontend/src/hooks/use-gpu-utilization.ts
+++ b/studio/frontend/src/hooks/use-gpu-utilization.ts
@@ -31,10 +31,8 @@ const DEFAULT: GpuUtilization = {
 };
 
 /**
- * Poll `GET /api/train/hardware` for live GPU utilization stats.
- *
- * Only polls while `enabled` is true (i.e. training is running).
- * Polling interval defaults to 10 000 ms.
+ * Poll `GET /api/train/hardware` for live GPU utilization stats while
+ * `enabled` (training running). Interval defaults to 10 000 ms.
  */
 export function useGpuUtilization(
     enabled: boolean,
diff --git a/studio/frontend/src/hooks/use-hardware-info.ts b/studio/frontend/src/hooks/use-hardware-info.ts
index 1e9210be5a..e988830d70 100644
--- a/studio/frontend/src/hooks/use-hardware-info.ts
+++ b/studio/frontend/src/hooks/use-hardware-info.ts
@@ -49,7 +49,7 @@ async function fetchOnce(): Promise<HardwareInfo> {
             cached = info;
             return info;
         } catch {
-            // Reset promise so subsequent calls retry (e.g. backend wasn't ready)
+            // Reset so subsequent calls retry (e.g. backend wasn't ready).
             fetchPromise = null;
             return DEFAULT;
         }
@@ -59,10 +59,8 @@ async function fetchOnce(): Promise<HardwareInfo> {
 }
 
 /**
- * Fetch hardware info from `GET /api/system/hardware`.
- *
- * The result is cached at module level — only one network request is made
- * regardless of how many components call this hook.
+ * Fetch hardware info from `GET /api/system/hardware`. Cached at module level,
+ * so only one request is made regardless of how many components call this hook.
  */
 export function useHardwareInfo(): HardwareInfo {
     const [info, setInfo] = useState<HardwareInfo>(cached ?? DEFAULT);
diff --git a/studio/frontend/src/hooks/use-hf-dataset-search.ts b/studio/frontend/src/hooks/use-hf-dataset-search.ts
index 78fe2d68d1..f8863f328f 100644
--- a/studio/frontend/src/hooks/use-hf-dataset-search.ts
+++ b/studio/frontend/src/hooks/use-hf-dataset-search.ts
@@ -332,8 +332,8 @@ function isOcrOrVisionTextDataset(dataset: HfDatasetResult): boolean {
 }
 
 function toCuratedDatasetResult(id: string): HfDatasetResult {
-  // Curated defaults are id-only. This adapter satisfies the shared result shape
-  // used by downstream combobox/ranking code without making extra HF requests.
+  // Curated defaults are id-only. Adapts them to the shared result shape used
+  // by downstream combobox/ranking code without extra HF requests.
   return {
     id,
     downloads: 0,
diff --git a/studio/frontend/src/hooks/use-hf-dataset-splits.ts b/studio/frontend/src/hooks/use-hf-dataset-splits.ts
index 2be5cad309..44c76779e4 100644
--- a/studio/frontend/src/hooks/use-hf-dataset-splits.ts
+++ b/studio/frontend/src/hooks/use-hf-dataset-splits.ts
@@ -169,9 +169,8 @@ export function useHfDatasetSplits(
   // Derive unique subsets
   const subsets = Array.from(new Set(entries.map((e) => e.config)));
 
-  // Derive splits for the active subset.
-  // If dataset has >1 subset and none is selected yet, return no splits so UI
-  // doesn't auto-pick/show a split before subset is chosen.
+  // Splits for the active subset. With >1 subset and none selected, return no
+  // splits so the UI doesn't auto-pick before a subset is chosen.
   const activeSubset =
     selectedSubset ?? (subsets.length === 1 ? subsets[0] : null);
   const filteredEntries = activeSubset
diff --git a/studio/frontend/src/hooks/use-hf-model-search.ts b/studio/frontend/src/hooks/use-hf-model-search.ts
index efe4d726be..ba52b654b8 100644
--- a/studio/frontend/src/hooks/use-hf-model-search.ts
+++ b/studio/frontend/src/hooks/use-hf-model-search.ts
@@ -42,8 +42,8 @@ const EXCLUDED_TAGS_MLX = new Set([
   "ctranslate2",
 ]);
 
-// Embedding / sentence-transformer models ship with onnx/openvino as additional
-// export formats — they should not be excluded by the tag check above.
+// Embedding / sentence-transformer models ship onnx/openvino as extra
+// export formats, so the tag check above must not exclude them.
 const EMBEDDING_TAGS = new Set([
   "sentence-transformers",
   "feature-extraction",
@@ -123,17 +123,16 @@ function makeMapModel(excludeGguf: boolean, excludedTags: Set<string>) {
 
 /** Number of unsloth results to pull up-front before yielding general results. */
 const UNSLOTH_PREFETCH = 20;
-/** When the user searched for a specific publisher, show fewer unsloth results
- *  before the pinned (original publisher) model. */
+/** Fewer unsloth results before the pinned model when a publisher is searched. */
 const UNSLOTH_PINNED_PREFETCH = 4;
 /** Matches a valid "owner/repo" identifier (exactly two non-empty segments). */
 const PUBLISHER_RE = /^([^/\s]+)\/([^/\s]+)$/;
 
 /**
- * Prime the hf-cache from a listModels result. For public (non-gated,
- * non-private) models, also prime the anonymous slot so the VRAM hook
- * gets cache hits without re-fetching. Gated/private models are only
- * cached under the caller's token to avoid auth leakage.
+ * Prime the hf-cache from a listModels result. For public models, also
+ * prime the anonymous slot so the VRAM hook gets cache hits without
+ * re-fetching. Gated/private models are cached only under the caller's
+ * token to avoid auth leakage.
  */
 function primeFromListing(
   name: string,
@@ -173,8 +172,8 @@ async function* mergedModelIterator(
     ...common,
   });
 
-  // Start pinned model lookup immediately so it can run in parallel with
-  // the Phase 1 unsloth iteration instead of blocking Phase 2.
+  // Start the pinned lookup now so it runs in parallel with Phase 1
+  // instead of blocking Phase 2.
   const pinnedPromise = pinnedId
     ? cachedModelInfo({
         name: pinnedId,
@@ -203,9 +202,8 @@ async function* mergedModelIterator(
   if (pinnedId && !seen.has(pinnedId) && pinnedPromise) {
     const pinned = await pinnedPromise;
     if (pinned) {
-      // Record both the raw input and the canonical name returned by HF
-      // so phase 2 deduplication works even when casing differs
-      // (e.g. user typed "OpenAI/gpt-oss-20b", HF returns "openai/gpt-oss-20b").
+      // Record both the raw input and HF's canonical name so phase 2 dedupe
+      // works across casing (e.g. "OpenAI/..." vs HF's "openai/...").
       seen.add(pinnedId);
       const canonicalName = (pinned as { name?: string }).name;
       if (canonicalName && canonicalName !== pinnedId) {
@@ -255,7 +253,7 @@ async function* priorityThenListingIterator(
   for (const result of settled) {
     if (result.status === "fulfilled") {
       const m = result.value as { name?: string; pipeline_tag?: string };
-      // Skip models that don't match the selected task filter
+      // Skip models that don't match the task filter.
       if (task && m.pipeline_tag && m.pipeline_tag !== task) continue;
       if (m.name) seen.add(m.name);
       yield result.value;
@@ -288,8 +286,8 @@ export function useHfModelSearch(
 ) {
   const { task, accessToken, excludeGguf = false, priorityIds } = options ?? {};
 
-  // Parse publisher detection once and share between the iterator factory
-  // and the secondary sort gate (avoids duplicating the regex + logic).
+  // Detect the publisher once, shared by the iterator factory and the
+  // secondary sort gate (avoids duplicating the regex + logic).
   const { isPublisherQuery, searchQuery, pinnedId, trimmed } = useMemo(() => {
     const t = query.trim();
     const m = PUBLISHER_RE.exec(t);
@@ -305,7 +303,7 @@ export function useHfModelSearch(
   const createIter = useCallback(
     () => {
       if (!trimmed) {
-        // No query: show priority models first (with full metadata), then general unsloth listing
+        // No query: priority models first (full metadata), then general unsloth.
         if (priorityIds && priorityIds.length > 0) {
           return priorityThenListingIterator(priorityIds, task, accessToken) as AsyncGenerator<unknown>;
         }
@@ -316,13 +314,11 @@ export function useHfModelSearch(
           ...(accessToken ? { credentials: { accessToken } } : {}),
         }) as AsyncGenerator<unknown>;
       }
-      // Typed query: disable task filter so explicitly searched models still
-      // appear even if HF task metadata is wrong/missing.
-      // If the query is a valid "owner/repo" identifier (exactly two non-empty,
-      // slash-free, space-free segments), strip the org prefix so unsloth
-      // variants surface, then pin the original publisher model after a small
-      // batch of unsloth results.  Queries for unsloth-owned models are left
-      // as-is so they get the full 20-result prefetch and secondary sort.
+      // Typed query: disable the task filter so explicitly searched models
+      // appear even with wrong/missing HF task metadata. For a valid
+      // "owner/repo" query, strip the org prefix so unsloth variants
+      // surface, then pin the original publisher model after a small batch.
+      // unsloth-owned queries are left as-is for the full prefetch + sort.
       return mergedModelIterator(searchQuery, undefined, accessToken, pinnedId) as AsyncGenerator<unknown>;
     },
     [trimmed, searchQuery, pinnedId, task, accessToken, priorityIds],
@@ -333,10 +329,9 @@ export function useHfModelSearch(
   const mapModel = useMemo(() => makeMapModel(excludeGguf, excludedTags), [excludeGguf, excludedTags]);
   const search = useHfPaginatedSearch(createIter, mapModel);
 
-  // Secondary sort guarantee: unsloth models always float to the top.
-  // Skip when the user searched for a specific non-unsloth publisher
-  // (e.g. "openai/gpt-oss-20b") -- the iterator already handles the
-  // pinned ordering in that case.
+  // Secondary sort: unsloth models always float to the top. Skip for a
+  // specific non-unsloth publisher query (e.g. "openai/gpt-oss-20b") --
+  // the iterator already handles pinned ordering there.
   const results = useMemo(
     () =>
       isPublisherQuery
diff --git a/studio/frontend/src/hooks/use-hf-token-validation.ts b/studio/frontend/src/hooks/use-hf-token-validation.ts
index f13be6b10a..11f56a2151 100644
--- a/studio/frontend/src/hooks/use-hf-token-validation.ts
+++ b/studio/frontend/src/hooks/use-hf-token-validation.ts
@@ -18,9 +18,8 @@ const INITIAL: HfTokenValidationState = {
 };
 
 /**
- * Validates the Hugging Face token by calling the whoami-v2 API.
- * Debounces the token to avoid excessive requests while typing.
- * Returns validation state: isValid (null = not checked), error message, and isChecking.
+ * Validates the HF token via the whoami-v2 API, debounced to avoid excessive
+ * requests while typing. isValid is null until checked.
  */
 export function useHfTokenValidation(token: string): HfTokenValidationState {
   const debouncedToken = useDebouncedValue(
diff --git a/studio/frontend/src/hooks/use-recommended-model-vram.ts b/studio/frontend/src/hooks/use-recommended-model-vram.ts
index 8826b38da9..39d58fe600 100644
--- a/studio/frontend/src/hooks/use-recommended-model-vram.ts
+++ b/studio/frontend/src/hooks/use-recommended-model-vram.ts
@@ -48,8 +48,8 @@ export function useRecommendedModelVram(ids: string[]) {
         }),
       );
       if (!canceled) {
-        // Merge with previous state so that VRAM badges for already-visible
-        // models are preserved while newly-visible models are still loading.
+        // Merge with previous state so VRAM badges for already-visible models
+        // persist while newly-visible models are still loading.
         setParamCountById((prev) => new Map([...prev, ...next]));
         setIsLoading(false);
       }
diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts
index c7628c0fd0..b376cb3a37 100644
--- a/studio/frontend/src/hooks/use-tauri-backend.ts
+++ b/studio/frontend/src/hooks/use-tauri-backend.ts
@@ -277,12 +277,11 @@ export function useTauriBackend() {
 
     try {
       const { invoke } = await import("@tauri-apps/api/core");
-      // backend/run.py keeps the existing 8888-8908 fallback via
-      // server-port/TAURI_PORT.
+      // backend/run.py keeps the 8888-8908 fallback via server-port/TAURI_PORT.
       await invoke("start_managed_server", { port: 8888 });
 
-      // Wait for the owned backend's server-port event. Do not attach to an
-      // external backend if the managed start does not report a port.
+      // Wait for the owned backend's server-port event. Don't attach to an
+      // external backend if the managed start doesn't report a port.
       const startupResult = await waitForManagedServerReady(
         invoke,
         () => portRef.current,
@@ -348,8 +347,7 @@ export function useTauriBackend() {
 
   async function stopServer() {
     if (isExternalServer) {
-      // We attached to a server we didn't spawn — can't kill it,
-      // just disconnect the UI.
+      // We attached to a server we didn't spawn: can't kill it, just disconnect the UI.
       startingRef.current = false;
       setIsExternalServer(false);
       stopExternalServerPoll();
@@ -373,19 +371,18 @@ export function useTauriBackend() {
     const { invoke } = await import("@tauri-apps/api/core");
     try {
       await invoke("start_install");
-      // Install completed — start the managed backend we just installed.
-      // Do not run the general preflight here: it can attach to an unrelated
-      // already-running CLI/backend server before launching our managed one.
-      // The install-complete event listener does NOT call startServer() to
-      // avoid a double-start race condition.
+      // Install done: start the managed backend we just installed. Don't run the
+      // general preflight here, it can attach to an unrelated running CLI/backend
+      // before launching ours. The install-complete listener does NOT call
+      // startServer() to avoid a double-start race.
       setBackendStatus("starting");
       elevationResumeRef.current = null;
       await startServer();
     } catch (e) {
       const msg = String(e);
-      // NEEDS_ELEVATION is not a real error — the Rust side also emits
-      // install-needs-elevation which sets needs-elevation status.
-      // Don't race with it by setting install-error here.
+      // NEEDS_ELEVATION is not a real error: the Rust side also emits
+      // install-needs-elevation (sets needs-elevation status). Don't race with it
+      // by setting install-error here.
       if (msg.includes("NEEDS_ELEVATION")) return;
       setBackendError(msg, "install-error");
     }
@@ -511,8 +508,8 @@ export function useTauriBackend() {
         setLogs((prev) => [...prev.slice(-499), e.payload]);
       });
 
-      // install-complete is informational only — does NOT trigger startServer.
-      // The invoke("start_install") success path handles that to avoid races.
+      // install-complete is informational only; does NOT trigger startServer. The
+      // invoke("start_install") success path handles that to avoid races.
       register<void>("install-complete", () => {
         setCurrentStepIndex(999); // all steps done
       });
diff --git a/studio/frontend/src/lib/copy-to-clipboard.ts b/studio/frontend/src/lib/copy-to-clipboard.ts
index cce9ea7783..2cbecbbdab 100644
--- a/studio/frontend/src/lib/copy-to-clipboard.ts
+++ b/studio/frontend/src/lib/copy-to-clipboard.ts
@@ -2,9 +2,8 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 /**
- * Copy text to clipboard in a way that works on Mac/Safari.
- * Uses a synchronous textarea + execCommand fallback so the copy runs in the
- * same user gesture as the click (required by Safari's clipboard security).
+ * Synchronous textarea + execCommand copy so it runs in the same user gesture
+ * as the click (required by Safari's clipboard security).
  */
 function copyWithExecCommand(text: string): boolean {
   if (typeof document === "undefined" || !document.body) return false;
@@ -44,8 +43,7 @@ export async function copyToClipboard(text: string): Promise<boolean> {
       return true;
     } catch (error) {
       console.warn("Async clipboard API failed, falling back to execCommand", error);
-      // Clipboard API rejected (NotAllowedError, insecure context, etc.)
-      // Fall through to execCommand fallback.
+      // Rejected (NotAllowedError, insecure context, etc.); fall through.
     }
   }
 
diff --git a/studio/frontend/src/lib/format-fastapi-error.ts b/studio/frontend/src/lib/format-fastapi-error.ts
index ed4a5ece7d..79d53f469d 100644
--- a/studio/frontend/src/lib/format-fastapi-error.ts
+++ b/studio/frontend/src/lib/format-fastapi-error.ts
@@ -5,11 +5,9 @@
  * Render a FastAPI error response body into a human-readable string.
  *
  * FastAPI emits 422s as `{detail: Array<{loc, msg, type, input?}>}`. The
- * naive `body.detail || body.message` pattern truthy-coerces the array
- * and stringifies it as `[object Object]`, which is unhelpful and was
- * exactly the regression #5409 fixed inside `train-api.ts`. Lift the
- * helper here so chat, export, history, datasets, and recipe-studio
- * can all share it.
+ * naive `body.detail || body.message` stringifies the array as
+ * `[object Object]` (the #5409 regression). Shared here across chat,
+ * export, history, datasets, and recipe-studio.
  *
  * Falls back through: array detail -> string detail -> message -> null.
  */
@@ -38,10 +36,9 @@ export function formatFastApiDetail(detail: unknown): string | null {
 }
 
 /**
- * Convert a Response (likely a non-ok response from a fetch) into the
- * best human-readable error message available. Used by *-api.ts wrappers
- * so toast text reads as `field: msg` instead of `[object Object]` or
- * `Request failed (422)`.
+ * Convert a (likely non-ok) Response into the best human-readable error
+ * message. Used by *-api.ts wrappers so toasts read `field: msg` instead
+ * of `[object Object]` or `Request failed (422)`.
  */
 export async function readFastApiError(
   response: Response,
diff --git a/studio/frontend/src/lib/hf-cache.ts b/studio/frontend/src/lib/hf-cache.ts
index d87345eb61..eebb3ce4b2 100644
--- a/studio/frontend/src/lib/hf-cache.ts
+++ b/studio/frontend/src/lib/hf-cache.ts
@@ -19,8 +19,8 @@ const MAX_CONCURRENT = 3;
 
 // ── Cache & in-flight maps ──────────────────────────────────────
 
-// Extend ModelEntry with the additional fields we always request so callers
-// do not need unsafe casts to access safetensors/tags.
+// Extend ModelEntry with the fields we always request, so callers avoid
+// unsafe casts to access safetensors/tags.
 export type CachedResult = ModelEntry & {
   safetensors?: { total?: number; parameters?: Record<string, number> };
   tags?: string[];
@@ -140,10 +140,9 @@ export async function cachedModelInfo(
       });
       const entry = { data: result as CachedResult, ts: Date.now() };
       cache.set(key, entry);
-      // For public (non-gated, non-private) models, also prime the anonymous
-      // cache slot so the VRAM hook (which reads without credentials) gets a
-      // cache hit. We skip gated/private models to avoid leaking auth-scoped
-      // metadata into the anonymous slot.
+      // For public models, also prime the anonymous cache slot so the VRAM
+      // hook (which reads without credentials) hits cache. Skip gated/private
+      // models to avoid leaking auth-scoped metadata into the anonymous slot.
       const r = result as CachedResult & { gated?: false | "auto" | "manual"; private?: boolean };
       if (token && !r.private && !r.gated) {
         const anonKey = cacheKey(params.name, undefined);
diff --git a/studio/frontend/src/lib/latex.ts b/studio/frontend/src/lib/latex.ts
index 0092d51be0..1a4ebdace8 100644
--- a/studio/frontend/src/lib/latex.ts
+++ b/studio/frontend/src/lib/latex.ts
@@ -15,8 +15,8 @@ const CURRENCY_REGEX =
   /(?<![\\$])\$(?!\$)(?=\d+(?:,\d{3})*(?:\.\d+)?[KMBkmb]?(?:\s|$|[^a-zA-Z\d]))/g;
 
 /**
- * Find regions inside code blocks (``` ... ``` and ` ... `) so we can skip them.
- * Returns sorted array of [start, end] index pairs.
+ * Find code-block regions (``` ... ``` and ` ... `) to skip.
+ * Returns a sorted array of [start, end] index pairs.
  */
 function findCodeBlockRegions(content: string): Array<[number, number]> {
   const regions: Array<[number, number]> = [];
@@ -28,12 +28,11 @@ function findCodeBlockRegions(content: string): Array<[number, number]> {
     regions.push([match.index, match.index + match[0].length]);
   }
 
-  // Inline code: `...` (but not inside fenced blocks -- we filter below)
+  // Inline code: `...` (skip spans inside fenced blocks, filtered below)
   const inlineRe = /`[^`\n]+`/g;
   while ((match = inlineRe.exec(content)) !== null) {
     const start = match.index;
     const end = start + match[0].length;
-    // Skip if this backtick span falls inside a fenced block
     let inside = false;
     for (const [rs, re] of regions) {
       if (start >= rs && end <= re) {
@@ -46,7 +45,7 @@ function findCodeBlockRegions(content: string): Array<[number, number]> {
     }
   }
 
-  // Sort by start position for binary search
+  // Sort by start position for binary search.
   regions.sort((a, b) => a[0] - b[0]);
   return regions;
 }
@@ -74,51 +73,42 @@ function isInCodeBlock(
   return false;
 }
 
-/**
- * A token (no whitespace) that looks purely like currency, e.g. `5`,
- * `1,000`, `5.99`, `100K`, `3.5M`.
- */
+/** A whitespace-free token that looks purely like currency, e.g. `5`, `1,000`, `5.99`, `100K`, `3.5M`. */
 const CURRENCY_BODY_RE = /^\d+(?:,\d{3})*(?:\.\d+)?[KMBkmb]?$/;
 
 /** Body characters that almost always indicate real LaTeX. */
 const LATEX_CHAR_RE = /[\\^_{}]/;
 
 /**
- * Operators that strongly suggest a math expression. We deliberately
- * leave out `^` and `_` because `LATEX_CHAR_RE` already short-circuits
- * on those before we ever consult this regex.
+ * Operators that strongly suggest math. Omits `^` and `_` since
+ * `LATEX_CHAR_RE` short-circuits on those before this regex runs.
  */
 const MATH_OP_RE = /[=+\-<>/*]/;
 
 /**
- * Trailing chars stripped before the currency check. Includes prose
- * punctuation plus the connectors `-` and `/` that appear in compact
- * currency ranges like `$5-$10` or `$5/$10`; without them the body
- * `5-` or `5/` would slip through the single-token math shortcut.
+ * Trailing chars stripped before the currency check: prose punctuation
+ * plus `-` and `/` from compact ranges like `$5-$10`; without them the
+ * body `5-` or `5/` would slip through the single-token math shortcut.
  */
 const TRAIL_PUNCT_RE = /[.,;:!?\-/]+$/;
 
 /**
- * A standalone single letter (variable name) inside the body. We require
- * that the letter is not part of a longer word so that prose like
- * "5 to attend" doesn't get misread as a math expression with the
- * variable `t`.
+ * A standalone single-letter variable, not part of a longer word, so
+ * prose like "5 to attend" isn't misread as math with variable `t`.
  */
 const LONE_LETTER_RE = /(?<![a-zA-Z])[a-zA-Z](?![a-zA-Z])/;
 
 /**
- * Numeric or single-letter operands joined by math operators, with
- * optional whitespace. Matches `2 + 2`, `100 < 200`, `1,000 - 500`,
- * `2.5 * 3`, `x + y`. Lets us recognise numeric-only expressions like
- * `$2 + 2$` as math without forcing a lone variable token.
+ * Numeric or single-letter operands joined by math operators (optional
+ * whitespace): `2 + 2`, `100 < 200`, `1,000 - 500`, `x + y`. Recognises
+ * numeric-only expressions like `$2 + 2$` without a lone variable token.
  */
 const SIMPLE_MATH_RE =
   /^(?:\d+(?:,\d{3})*(?:\.\d+)?|[a-zA-Z])(?:\s*[=+\-<>/*]\s*(?:\d+(?:,\d{3})*(?:\.\d+)?|[a-zA-Z]))+$/;
 
 /**
- * Return true if the substring between two `$` delimiters looks like a
- * LaTeX expression rather than a span of prose between two currency
- * tokens.
+ * True if the substring between two `$` delimiters looks like LaTeX
+ * rather than prose between two currency tokens.
  *
  * Rule of thumb:
  *   - `$30^\circ$`  -> math (LaTeX chars)
@@ -142,14 +132,12 @@ function looksLikeMathBody(body: string): boolean {
 }
 
 /**
- * Return true if the `$` at `offset` opens a balanced inline math span
- * (`$...$`) on the same line. The closer must be unescaped, must not be
- * part of `$$`, and must lie inside a 200-character window. The body
- * must look like LaTeX so we don't pair two currency tokens that share
- * a line (e.g. "$5 to $10"). Bold-wrapped spans (`**$X$**` and the
- * underscore equivalent `__$X$__`) are always treated as math because
- * LLMs reach for that pattern when they want "bold math" and the
- * heuristic would otherwise reject prose-shaped bodies like "90 - x".
+ * True if the `$` at `offset` opens a balanced inline math span (`$...$`)
+ * on the same line. The closer must be unescaped, not part of `$$`, and
+ * within 200 chars. The body must look like LaTeX so we don't pair two
+ * currency tokens on a line (e.g. "$5 to $10"). Bold-wrapped spans
+ * (`**$X$**`, `__$X$__`) are always math: LLMs use that for "bold math"
+ * and the heuristic would otherwise reject prose-shaped bodies like "90 - x".
  */
 function hasInlineMathCloser(content: string, offset: number): boolean {
   const MAX_SPAN = 200;
@@ -163,10 +151,9 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
       i++;
       continue;
     }
-    // A `$` immediately followed by a digit is far more likely the start
-    // of another currency token than the closer for the current span. Keep
-    // scanning so prose like `Starts at $5 + a $10 add-on` doesn't pair
-    // the two currency markers as a math span.
+    // A `$` followed by a digit is more likely another currency token than
+    // the closer. Keep scanning so prose like `$5 + a $10 add-on` doesn't
+    // pair the two currency markers as a math span.
     if (/\d/.test(content[i + 1] ?? "")) {
       continue;
     }
diff --git a/studio/frontend/src/lib/open-link.ts b/studio/frontend/src/lib/open-link.ts
index b7e6a0fa6e..46c9a8dd59 100644
--- a/studio/frontend/src/lib/open-link.ts
+++ b/studio/frontend/src/lib/open-link.ts
@@ -1,21 +1,20 @@
 import { isTauri } from "@/lib/api-base";
 
 /**
- * Open a URL in the system browser (Tauri) or new tab (web).
- * Handles anchor links and mailto: natively without the opener plugin.
- * Returns true when the caller should preventDefault; false when the
- * browser's native navigation should proceed (relative URLs, empty, etc.).
+ * Open a URL in the system browser (Tauri) or new tab (web). Handles anchor
+ * and mailto: links natively. Returns true when the caller should
+ * preventDefault; false to let native navigation proceed (relative, empty).
  */
 export function openLink(url: string): boolean {
   if (!url) return false;
 
-  // Anchor links — scroll within the page, don't open externally
+  // Anchor links scroll within the page, don't open externally
   if (url.startsWith("#")) {
     window.location.hash = url;
     return true;
   }
 
-  // Relative URLs — let the browser / router handle them natively
+  // Relative URLs: let the browser / router handle them natively
   if (!url.includes("://") && !url.startsWith("mailto:")) {
     return false;
   }
diff --git a/studio/frontend/src/lib/vram.ts b/studio/frontend/src/lib/vram.ts
index fad2c438bd..1c443134d9 100644
--- a/studio/frontend/src/lib/vram.ts
+++ b/studio/frontend/src/lib/vram.ts
@@ -4,56 +4,42 @@
 /**
  * VRAM estimation for model loading (4-bit quantization via bitsandbytes).
  *
- * Estimates the total driver-level VRAM (what nvidia-smi reports) needed to
- * load a model in 4-bit with Unsloth / bitsandbytes.  This determines
- * whether a model will fit on the user's GPU before any training begins.
+ * Estimates the total driver-level VRAM (what nvidia-smi reports) to load a
+ * model in 4-bit with Unsloth / bitsandbytes, to check it fits the GPU before
+ * training.
  *
- * Formula:  totalParams * 0.90  +  1.4 GB
+ * Formula: totalParams * 0.90 + 1.4 GB
  *
- * Calibrated against isolated Unsloth model loads on RTX 5070 Ti (2026.2):
+ * Calibrated against isolated Unsloth loads on RTX 5070 Ti (2026.2):
  *   Qwen2.5-0.5B  (0.49B) : est 1.8 vs actual 1.86 GB  (-3%)
  *   Llama-3.2-1B  (1.24B) : est 2.5 vs actual 2.54 GB   (-1%)
  *   Llama-3.2-3B  (3.21B) : est 4.3 vs actual 4.40 GB   (-2%)
  *   Llama-3.1-8B  (8.03B) : est 8.6 vs actual 8.14 GB   (+6%)
  *
- * Accuracy: within 3% for 0.5B-3B models, within 6% for 8B.
+ * Accuracy: within 3% for 0.5B-3B, within 6% for 8B.
  */
 
-// ---------------------------------------------------------------------------
 // Constants (exported for testing)
-// ---------------------------------------------------------------------------
 
 /**
- * Effective bytes per parameter for 4-bit model weights at driver level.
- *
- * Raw bnb 4-bit is ~0.5 bytes/param, but embedding and lm_head layers remain
- * in fp16 and bnb adds per-block quantization metadata, bringing the
- * effective rate to ~0.84-0.93 across tested architectures.  0.9 is the
- * calibrated middle ground.
+ * Effective bytes per parameter for 4-bit weights at driver level. Raw bnb
+ * 4-bit is ~0.5, but embedding/lm_head stay fp16 and bnb adds per-block
+ * metadata, giving ~0.84-0.93 across architectures; 0.9 is the calibrated mid.
  */
 export const BNB_4BIT_LOADING_BYTES = 0.9;
 
 /**
- * Fixed overhead (GB) for the CUDA driver context and PyTorch runtime.
- *
- * This is independent of model size -- it is the baseline GPU memory consumed
- * before any model weights are loaded.  Measured at 1.34-1.46 GB across
- * tested models; we use 1.4 as the default.
+ * Fixed overhead (GB) for the CUDA driver context and PyTorch runtime,
+ * independent of model size. Measured at 1.34-1.46 GB; we use 1.4.
  */
 export const LOADING_OVERHEAD_GB = 1.4;
 
-// ---------------------------------------------------------------------------
-// Estimation
-// ---------------------------------------------------------------------------
-
 export type VramFitStatus = "fits" | "tight" | "exceeds";
 
 /**
- * Bytes per parameter when loading a model at fp16/bf16 (LoRA, full FT).
- *
- * This is the theoretical value (2 bytes = 16 bits).  Not yet calibrated
- * against actual measurements -- the real driver-level usage may be slightly
- * higher due to buffers and metadata, similar to how 4-bit is 0.9 vs 0.5.
+ * Bytes per parameter at fp16/bf16 (LoRA, full FT). Theoretical (2 bytes = 16
+ * bits); not yet calibrated, so real usage may run slightly higher (as 4-bit
+ * is 0.9 vs 0.5).
  */
 export const FP16_LOADING_BYTES = 2.0;
 
@@ -65,15 +51,9 @@ function usesQuantizedLoading(method: TrainingMethod, modelId?: string): boolean
 }
 
 /**
- * Estimate VRAM (GB) needed to load a model with Unsloth.
- *
- * The bytes-per-param rate depends on the training method:
- *   - QLoRA  : 4-bit quantized via bnb  -> 0.90 bytes/param (calibrated)
- *   - LoRA   : fp16                      -> 2.0  bytes/param (theoretical)
- *   - Full   : fp16                      -> 2.0  bytes/param (theoretical)
- *   - CPT    : fp16 LoRA (16-bit base)   -> 2.0  bytes/param (theoretical)
- *
- * Formula:  totalParams * bytesPerParam + 1.4 GB overhead
+ * Estimate VRAM (GB) to load a model with Unsloth. Bytes/param by method:
+ *   QLoRA: 4-bit bnb -> 0.90 (calibrated); LoRA/Full/CPT: fp16 -> 2.0
+ *   (theoretical). Formula: totalParams * bytesPerParam + 1.4 GB overhead.
  */
 export function estimateLoadingVram(
   totalParams: number,
@@ -88,11 +68,8 @@ export function estimateLoadingVram(
 }
 
 /**
- * Check whether a model fits in the available GPU VRAM.
- *
- *   fits     -  uses <= 75% of available
- *   tight    -  uses 75-100% of available
- *   exceeds  -  uses > 100% of available
+ * Check whether a model fits in available GPU VRAM.
+ *   fits: <= 75%; tight: 75-100%; exceeds: > 100%.
  */
 export function checkVramFit(
   requiredGb: number,
diff --git a/studio/frontend/src/main.tsx b/studio/frontend/src/main.tsx
index a924c1652e..0922e764bb 100644
--- a/studio/frontend/src/main.tsx
+++ b/studio/frontend/src/main.tsx
@@ -12,8 +12,8 @@ import { initializeLocale } from "./i18n";
 const globalCrypto = globalThis.crypto as Crypto | undefined;
 
 if (globalCrypto && typeof globalCrypto.randomUUID !== "function") {
-  // Some envs ship `crypto` but no `randomUUID()` (or a non-function stub).
-  // Provide a best-effort v4 UUID using `getRandomValues` when available.
+  // Some envs ship `crypto` without `randomUUID()`. Provide a best-effort v4
+  // UUID using `getRandomValues` when available.
   const cryptoRef = globalCrypto;
 
   function getRandomByte(): number {
diff --git a/studio/frontend/src/utils/index.ts b/studio/frontend/src/utils/index.ts
index fef4e946d9..e8eccd2e29 100644
--- a/studio/frontend/src/utils/index.ts
+++ b/studio/frontend/src/utils/index.ts
@@ -1,5 +1,4 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Utility functions
 export { normalizeNonEmptyName } from "./strings";