diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2317a2d..94c751d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,32 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.9.8] - 2026-07-15
+
+### Added
+
+- 🏠 **Choose how `.cptr` is handled in projects.** Admins can now turn the automatic `.gitignore` entry on or off, so local chats, memory, caches, audio samples, and generated files stay private by default unless you choose otherwise.
+- 🤖 **A llama.cpp connection mode.** OpenAI-compatible connections can now be marked as llama.cpp, so Qwen and other local reasoning models can receive prior thinking in the format llama.cpp expects.
+- 🧩 **Skills can open their own companion files.** `view_skill` can now load files such as `references/guide.md` from the selected skill folder, without opening broader home-directory access.
+- 🌐 **Sized browser screenshots.** Browser tasks can request a screenshot at a specific size and Computer will restore the normal page view afterward.
+
+### Changed
+
+- ⚙️ **Admin switches behave like switches.** In Models, Agents, Connections, Messaging, and Tool Servers, click the row text to edit and click the switch only to turn something on or off.
+- 🔎 **Workspace search respects `.gitignore`.** File search and content search now skip ignored folders and files, keeping generated output, caches, and local Computer data out of normal results.
+- 🧠 **Reasoning replay is stricter by default.** Computer now keeps strict OpenAI-compatible connections clean, and only sends llama.cpp-style `reasoning_content` when the connection is set to llama.cpp.
+
+### Fixed
+
+- 🧠 **The chat size meter no longer freezes behind llama.cpp.** Computer now normalizes token usage from OpenAI-compatible streams and sends live context updates to the chat while a response is still running.
+- 💬 **Compacted chats survive regenerated answers better.** Summaries are saved on a stable kept user message when possible, so regenerating an assistant answer is less likely to lose the compacted history.
+- ✂️ **Manual compact keeps the latest turn available.** Compacting a chat now summarizes the older branch and leaves the newest message ready for the next response.
+- 🧠 **Prior reasoning is preserved more reliably.** Reasoning from ordinary replies, tool-call turns, Chat Completions streams, and Responses streams is less likely to be dropped before the next model request.
+- 🎨 **Token usage details match the chosen theme.** The usage popover now uses Computer's app surface colors instead of the wrong gray background.
+- ✨ **Messages render more faithfully.** Your own messages can show Markdown, and raw markup appears as text instead of disappearing.
+- 🗂️ **Commit suggestions are less brittle.** If the model returns a plain commit message instead of perfect structured output, Computer can still use the useful parts.
+- 📝 **File edits keep text encoding consistent.** Created and edited files are written as UTF-8, which avoids surprises with non-ASCII project files.
+
 ## [0.9.7] - 2026-07-12
 
 ### Fixed
diff --git a/cptr/env.py b/cptr/env.py
index 5d856a5..14ad9d4 100644
--- a/cptr/env.py
+++ b/cptr/env.py
@@ -66,6 +66,10 @@ def _env_int(name: str, default: int) -> int:
 # Claude SDK stdout JSON buffer; chat/tool output caps apply later after parsing.
 CLAUDE_CODE_MAX_BUFFER_SIZE = _env_int("CPTR_CLAUDE_CODE_MAX_BUFFER_SIZE", 128 * 1024 * 1024)
 
+# ── Workspace storage ───────────────────────────────────────
+WORKSPACE_AUTO_GITIGNORE_DOT_CPTR_ENV = os.environ.get("CPTR_AUTO_GITIGNORE_DOT_CPTR")
+WORKSPACE_AUTO_GITIGNORE_DOT_CPTR = _env_bool("CPTR_AUTO_GITIGNORE_DOT_CPTR", "true")
+
 # ── Execute timeout ─────────────────────────────────────────
 # Default wait (seconds) for run_command / check_task when the caller
 # doesn't pass an explicit wait value.  None = return immediately.
diff --git a/cptr/frontend/package-lock.json b/cptr/frontend/package-lock.json
index 1f05b5e..d258992 100644
--- a/cptr/frontend/package-lock.json
+++ b/cptr/frontend/package-lock.json
@@ -1,12 +1,12 @@
 {
 	"name": "frontend",
-	"version": "0.4.0",
+	"version": "0.4.1",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
 			"name": "frontend",
-			"version": "0.3.4",
+			"version": "0.4.1",
 			"dependencies": {
 				"@codemirror/lang-cpp": "^6.0.3",
 				"@codemirror/lang-css": "^6.3.1",
diff --git a/cptr/frontend/package.json b/cptr/frontend/package.json
index 8c07701..60d9fa0 100644
--- a/cptr/frontend/package.json
+++ b/cptr/frontend/package.json
@@ -93,4 +93,4 @@
 		"tippy.js": "^6.3.7",
 		"xlsx": "^0.18.5"
 	}
-}
\ No newline at end of file
+}
diff --git a/cptr/frontend/src/app.css b/cptr/frontend/src/app.css
index c29a90c..e03784f 100644
--- a/cptr/frontend/src/app.css
+++ b/cptr/frontend/src/app.css
@@ -29,8 +29,8 @@
 	--app-fg: #525252;
 	--app-fg-muted: color-mix(in oklab, var(--app-fg) 62%, var(--app-bg));
 	--app-fg-subtle: color-mix(in oklab, var(--app-fg) 48%, var(--app-bg));
-	--app-border: color-mix(in oklab, var(--app-fg) 7%, transparent);
-	--app-divider: color-mix(in oklab, var(--app-fg) 5%, transparent);
+	--app-border: color-mix(in oklab, var(--app-fg) 5%, transparent);
+	--app-divider: color-mix(in oklab, var(--app-fg) 3.5%, transparent);
 	--app-hover: color-mix(in oklab, var(--app-fg) 6%, transparent);
 	--app-active: color-mix(in oklab, var(--app-fg) 7%, transparent);
 	--app-ui-font: var(--font-sans);
diff --git a/cptr/frontend/src/lib/apis/admin.ts b/cptr/frontend/src/lib/apis/admin.ts
index a8140ff..b567453 100644
--- a/cptr/frontend/src/lib/apis/admin.ts
+++ b/cptr/frontend/src/lib/apis/admin.ts
@@ -130,6 +130,7 @@ export interface Connection {
 	name: string;
 	provider: string;
 	api_type: string;
+	provider_type?: 'default' | 'llama.cpp' | null;
 	prefix_id: string | null;
 	base_url: string | null;
 	api_key: string | null;
@@ -146,6 +147,7 @@ export const createConnection = (conn: {
 	name: string;
 	provider: string;
 	api_type?: string;
+	provider_type?: 'default' | 'llama.cpp' | null;
 	prefix_id?: string | null;
 	base_url?: string | null;
 	api_key?: string | null;
diff --git a/cptr/frontend/src/lib/apis/chat.ts b/cptr/frontend/src/lib/apis/chat.ts
index a5c8339..eb1e171 100644
--- a/cptr/frontend/src/lib/apis/chat.ts
+++ b/cptr/frontend/src/lib/apis/chat.ts
@@ -171,8 +171,11 @@ export const answerAskUser = (
 export const cancelTask = (chatId: string, messageId: string) =>
 	fetchJSON(`/api/chats/${chatId}/messages/${messageId}/cancel`, { method: 'POST' });
 
-export const compactChat = (chatId: string, modelId: string) =>
-	fetchJSON<CompactChatResult>(`/api/chats/${chatId}/compact`, jsonBody({ model_id: modelId }));
+export const compactChat = (chatId: string, modelId?: string | null) =>
+	fetchJSON<CompactChatResult>(
+		`/api/chats/${chatId}/compact`,
+		jsonBody({ model_id: modelId || null })
+	);
 
 export const updateCurrentMessage = (chatId: string, messageId: string) =>
 	fetchJSON<{ ok: boolean }>(`/api/chats/${chatId}/current`, jsonBody({ message_id: messageId }));
diff --git a/cptr/frontend/src/lib/components/Admin/Agents.svelte b/cptr/frontend/src/lib/components/Admin/Agents.svelte
index 59477f7..f081951 100644
--- a/cptr/frontend/src/lib/components/Admin/Agents.svelte
+++ b/cptr/frontend/src/lib/components/Admin/Agents.svelte
@@ -10,6 +10,7 @@
 	} from '$lib/apis/admin';
 	import Icon from '$lib/components/Icon.svelte';
 	import Spinner from '$lib/components/common/Spinner.svelte';
+	import ToggleSwitch from '$lib/components/common/ToggleSwitch.svelte';
 	import { t } from '$lib/i18n';
 	import { refreshChatState } from '$lib/stores/chat';
 	import AgentProfileModal from './AgentProfileModal.svelte';
@@ -179,8 +180,7 @@
 		modal = null;
 	}
 
-	function toggleProfile(e: Event, index: number) {
-		e.stopPropagation();
+	function toggleProfile(index: number) {
 		const profile = profiles[index];
 		if (!profile) return;
 		const mode = profile.mode === 'disabled' ? 'enabled' : 'disabled';
@@ -262,11 +262,12 @@
 	{:else}
 		<div class="divide-y divide-gray-100 dark:divide-white/5">
 			{#each profiles as profile, index (profile.id)}
-				<button
-					class="group flex items-center gap-2 w-full min-h-9 py-1.5 text-left"
-					onclick={() => (modal = { mode: 'edit', index, profile })}
-				>
-					<span class="flex-1 min-w-0">
+				<div class="group flex items-center gap-2 w-full min-h-9 py-1.5 text-left">
+					<button
+						type="button"
+						class="flex-1 min-w-0 text-left"
+						onclick={() => (modal = { mode: 'edit', index, profile })}
+					>
 						<span
 							class="block text-[0.8125rem] truncate {profile.mode === 'disabled'
 								? 'text-gray-400 dark:text-gray-600'
@@ -282,36 +283,15 @@
 								{detected[profile.id]?.message}
 							</span>
 						{/if}
-					</span>
+					</button>
 					<span
 						class="hidden sm:flex items-center gap-1.5 shrink-0 text-[0.6875rem] text-gray-400 dark:text-gray-600"
 					>
 						<span class="w-1.5 h-1.5 rounded-full {statusTone(profile)}"></span>
 						<span>{$t(statusKey(profile))}</span>
 					</span>
-					<!-- svelte-ignore a11y_no_static_element_interactions -->
-					<span
-						class="relative w-6 h-3.5 rounded-full shrink-0 cursor-pointer transition-colors duration-150
-							{profile.mode !== 'disabled' ? 'bg-gray-900 dark:bg-white' : 'bg-gray-200 dark:bg-gray-700'}"
-						role="switch"
-						tabindex="-1"
-						aria-checked={profile.mode !== 'disabled'}
-						onclick={(e) => toggleProfile(e, index)}
-						onkeydown={(e) => {
-							if (e.key === 'Enter' || e.key === ' ') {
-								e.preventDefault();
-								toggleProfile(e, index);
-							}
-						}}
-					>
-						<span
-							class="absolute top-0.5 w-2.5 h-2.5 rounded-full transition-all duration-150
-								{profile.mode !== 'disabled'
-								? 'left-3 bg-white dark:bg-gray-900'
-								: 'left-0.5 bg-white dark:bg-gray-500'}"
-						></span>
-					</span>
-				</button>
+					<ToggleSwitch value={profile.mode !== 'disabled'} onchange={() => toggleProfile(index)} />
+				</div>
 			{/each}
 		</div>
 
diff --git a/cptr/frontend/src/lib/components/Admin/Connections.svelte b/cptr/frontend/src/lib/components/Admin/Connections.svelte
index 94730a2..d1c261b 100644
--- a/cptr/frontend/src/lib/components/Admin/Connections.svelte
+++ b/cptr/frontend/src/lib/components/Admin/Connections.svelte
@@ -8,6 +8,7 @@
 	import { refreshChatState } from '$lib/stores/chat';
 	import { t } from '$lib/i18n';
 	import Spinner from '$lib/components/common/Spinner.svelte';
+	import ToggleSwitch from '$lib/components/common/ToggleSwitch.svelte';
 
 	let connections = $state<Connection[]>([]);
 	let loading = $state(true);
@@ -35,9 +36,7 @@
 		load();
 	}
 
-	async function toggleEnabled(e: Event, conn: Connection) {
-		e.stopPropagation();
-		const newVal = !conn.enabled;
+	async function toggleEnabled(conn: Connection, newVal: boolean) {
 		// Optimistic update
 		conn.enabled = newVal;
 		connections = [...connections];
@@ -72,37 +71,17 @@
 {:else}
 	<div>
 		{#each connections as conn}
-			<button
-				class="group flex items-center gap-2 w-full h-7 text-left"
-				onclick={() => (editConn = conn)}
-			>
-				<span
-					class="flex-1 text-[0.8125rem] truncate
+			<div class="group flex items-center gap-2 w-full h-7 text-left">
+				<button
+					type="button"
+					class="flex-1 min-w-0 text-left text-[0.8125rem] truncate
 					{conn.enabled ? 'text-gray-700 dark:text-gray-300' : 'text-gray-400 dark:text-gray-600'}"
+					onclick={() => (editConn = conn)}
 				>
 					{conn.name}
-				</span>
-				<!-- svelte-ignore a11y_no_static_element_interactions -->
-				<span
-					class="relative w-6 h-3.5 rounded-full shrink-0 cursor-pointer transition-colors duration-150
-						{conn.enabled ? 'bg-gray-900 dark:bg-white' : 'bg-gray-200 dark:bg-gray-700'}"
-					role="switch"
-					tabindex="-1"
-					aria-checked={conn.enabled}
-					onclick={(e) => toggleEnabled(e, conn)}
-					onkeydown={(e) => {
-						if (e.key === 'Enter' || e.key === ' ') {
-							e.preventDefault();
-							toggleEnabled(e, conn);
-						}
-					}}
-				>
-					<span
-						class="absolute top-0.5 w-2.5 h-2.5 rounded-full transition-all duration-150
-						{conn.enabled ? 'left-3 bg-white dark:bg-gray-900' : 'left-0.5 bg-white dark:bg-gray-500'}"
-					></span>
-				</span>
-			</button>
+				</button>
+				<ToggleSwitch value={conn.enabled} onchange={(value) => toggleEnabled(conn, value)} />
+			</div>
 		{/each}
 
 		{#if connections.length === 0}
diff --git a/cptr/frontend/src/lib/components/Admin/CreateConnectionModal.svelte b/cptr/frontend/src/lib/components/Admin/CreateConnectionModal.svelte
index 4c81224..a352ddd 100644
--- a/cptr/frontend/src/lib/components/Admin/CreateConnectionModal.svelte
+++ b/cptr/frontend/src/lib/components/Admin/CreateConnectionModal.svelte
@@ -17,6 +17,7 @@
 	let formName = $state('');
 	let formProvider = $state<'openai' | 'anthropic'>('openai');
 	let formApiType = $state<'chat_completions' | 'responses'>('chat_completions');
+	let formProviderType = $state<'default' | 'llama.cpp'>('default');
 	let formBaseUrl = $state('');
 	let formApiKey = $state('');
 	let formPrefixId = $state('');
@@ -49,6 +50,7 @@
 				name,
 				provider: formProvider,
 				api_type: formProvider === 'openai' ? formApiType : 'chat_completions',
+				provider_type: formProvider === 'openai' ? formProviderType : 'default',
 				prefix_id: formPrefixId.trim() || null,
 				base_url: formBaseUrl.trim(),
 				api_key: formApiKey.trim(),
@@ -115,6 +117,21 @@
 				<option value="chat_completions">{$t('connections.chatCompletions')}</option>
 				<option value="responses">{$t('connections.responses')}</option>
 			</select>
+
+			<label
+				for="create-provider-type"
+				class="text-[0.625rem] text-gray-400 dark:text-gray-600 mt-2"
+			>
+				Provider Type
+			</label>
+			<select
+				id="create-provider-type"
+				bind:value={formProviderType}
+				class="block w-full bg-transparent text-[0.8125rem] text-gray-700 dark:text-gray-300 outline-none py-0.5 cursor-pointer"
+			>
+				<option value="default">Default</option>
+				<option value="llama.cpp">llama.cpp</option>
+			</select>
 		{/if}
 
 		<label class="text-[0.625rem] text-gray-400 dark:text-gray-600 mt-2"
diff --git a/cptr/frontend/src/lib/components/Admin/EditConnectionModal.svelte b/cptr/frontend/src/lib/components/Admin/EditConnectionModal.svelte
index 424e061..e208414 100644
--- a/cptr/frontend/src/lib/components/Admin/EditConnectionModal.svelte
+++ b/cptr/frontend/src/lib/components/Admin/EditConnectionModal.svelte
@@ -27,6 +27,9 @@
 	let formApiType = $state<'chat_completions' | 'responses'>(
 		connection.api_type === 'responses' ? 'responses' : 'chat_completions'
 	);
+	let formProviderType = $state<'default' | 'llama.cpp'>(
+		connection.provider_type === 'llama.cpp' ? 'llama.cpp' : 'default'
+	);
 	let formBaseUrl = $state(connection.base_url || '');
 	let formApiKey = $state('');
 	let formPrefixId = $state(connection.prefix_id || '');
@@ -61,6 +64,7 @@
 				name,
 				provider: formProvider,
 				api_type: formProvider === 'openai' ? formApiType : 'chat_completions',
+				provider_type: formProvider === 'openai' ? formProviderType : 'default',
 				prefix_id: formPrefixId.trim() || null,
 				base_url: formBaseUrl.trim(),
 				models
@@ -151,6 +155,21 @@
 				<option value="chat_completions">{$t('connections.chatCompletions')}</option>
 				<option value="responses">{$t('connections.responses')}</option>
 			</select>
+
+			<label
+				for="edit-provider-type"
+				class="text-[0.625rem] text-gray-400 dark:text-gray-600 mt-2"
+			>
+				Provider Type
+			</label>
+			<select
+				id="edit-provider-type"
+				bind:value={formProviderType}
+				class="block w-full bg-transparent text-[0.8125rem] text-gray-700 dark:text-gray-300 outline-none py-0.5 cursor-pointer"
+			>
+				<option value="default">Default</option>
+				<option value="llama.cpp">llama.cpp</option>
+			</select>
 		{/if}
 
 		<label class="text-[0.625rem] text-gray-400 dark:text-gray-600 mt-2"
diff --git a/cptr/frontend/src/lib/components/Admin/Messaging.svelte b/cptr/frontend/src/lib/components/Admin/Messaging.svelte
index d55ca21..c18bf4e 100644
--- a/cptr/frontend/src/lib/components/Admin/Messaging.svelte
+++ b/cptr/frontend/src/lib/components/Admin/Messaging.svelte
@@ -6,6 +6,7 @@
 	import { listBots, deleteBot, startBot, stopBot, type BotData } from '$lib/apis/bots';
 	import { t } from '$lib/i18n';
 	import Spinner from '$lib/components/common/Spinner.svelte';
+	import ToggleSwitch from '$lib/components/common/ToggleSwitch.svelte';
 
 	let bots = $state<BotData[]>([]);
 	let loading = $state(true);
@@ -29,8 +30,7 @@
 		load();
 	}
 
-	async function toggleRunning(e: Event, bot: BotData) {
-		e.stopPropagation();
+	async function toggleRunning(bot: BotData) {
 		const wasRunning = bot.is_running;
 		// Optimistic
 		bot.is_running = !wasRunning;
@@ -108,27 +108,7 @@
 					<Icon name="trash" size={11} />
 				</button>
 
-				<!-- Toggle switch -->
-				<!-- svelte-ignore a11y_no_static_element_interactions -->
-				<span
-					class="relative w-6 h-3.5 rounded-full shrink-0 cursor-pointer transition-colors duration-150
-						{bot.is_running ? 'bg-gray-900 dark:bg-white' : 'bg-gray-200 dark:bg-gray-700'}"
-					role="switch"
-					tabindex="-1"
-					aria-checked={bot.is_running}
-					onclick={(e) => toggleRunning(e, bot)}
-					onkeydown={(e) => {
-						if (e.key === 'Enter' || e.key === ' ') {
-							e.preventDefault();
-							toggleRunning(e, bot);
-						}
-					}}
-				>
-					<span
-						class="absolute top-0.5 w-2.5 h-2.5 rounded-full transition-all duration-150
-						{bot.is_running ? 'left-3 bg-white dark:bg-gray-900' : 'left-0.5 bg-white dark:bg-gray-500'}"
-					></span>
-				</span>
+				<ToggleSwitch value={bot.is_running} onchange={() => toggleRunning(bot)} />
 			</div>
 		{/each}
 
diff --git a/cptr/frontend/src/lib/components/Admin/Models.svelte b/cptr/frontend/src/lib/components/Admin/Models.svelte
index 1b34ea7..45cba5b 100644
--- a/cptr/frontend/src/lib/components/Admin/Models.svelte
+++ b/cptr/frontend/src/lib/components/Admin/Models.svelte
@@ -278,9 +278,7 @@ Files:
 		}
 	}
 
-	async function toggleModel(e: Event, model: ModelEntry) {
-		e.stopPropagation();
-		const newVal = !model.is_active;
+	async function toggleModel(model: ModelEntry, newVal: boolean) {
 		model.is_active = newVal;
 		models = [...models];
 		try {
@@ -659,38 +657,18 @@ Files:
 
 			<!-- Per-model list -->
 			{#each models as model}
-				<button
-					class="group flex items-center gap-2 w-full h-7 text-left"
-					onclick={() => (selectedId = selectedId === model.id ? null : model.id)}
-				>
-					<span
-						class="flex-1 text-[0.8125rem] truncate {model.is_active
+				<div class="group flex items-center gap-2 w-full h-7 text-left">
+					<button
+						type="button"
+						class="flex-1 min-w-0 text-left text-[0.8125rem] truncate {model.is_active
 							? 'text-gray-700 dark:text-gray-300'
 							: 'text-gray-400 dark:text-gray-600'}"
+						onclick={() => (selectedId = selectedId === model.id ? null : model.id)}
 					>
 						{model.name}
-					</span>
-					<!-- svelte-ignore a11y_no_static_element_interactions -->
-					<span
-						class="relative w-6 h-3.5 rounded-full shrink-0 cursor-pointer transition-colors duration-150
-							{model.is_active ? 'bg-gray-900 dark:bg-white' : 'bg-gray-200 dark:bg-gray-700'}"
-						role="switch"
-						tabindex="-1"
-						aria-checked={model.is_active}
-						onclick={(e) => toggleModel(e, model)}
-						onkeydown={(e) => {
-							if (e.key === 'Enter' || e.key === ' ') {
-								e.preventDefault();
-								toggleModel(e, model);
-							}
-						}}
-					>
-						<span
-							class="absolute top-0.5 w-2.5 h-2.5 rounded-full transition-all duration-150
-							{model.is_active ? 'left-3 bg-white dark:bg-gray-900' : 'left-0.5 bg-white dark:bg-gray-500'}"
-						></span>
-					</span>
-				</button>
+					</button>
+					<ToggleSwitch value={model.is_active} onchange={(value) => toggleModel(model, value)} />
+				</div>
 
 				{#if selectedId === model.id}
 					{@render compactThresholdField(
diff --git a/cptr/frontend/src/lib/components/Admin/ToolServers.svelte b/cptr/frontend/src/lib/components/Admin/ToolServers.svelte
index 9a8afc3..1dac2f9 100644
--- a/cptr/frontend/src/lib/components/Admin/ToolServers.svelte
+++ b/cptr/frontend/src/lib/components/Admin/ToolServers.svelte
@@ -3,6 +3,7 @@
 	import Icon from '../Icon.svelte';
 	import Modal from '../Modal.svelte';
 	import Spinner from '$lib/components/common/Spinner.svelte';
+	import ToggleSwitch from '$lib/components/common/ToggleSwitch.svelte';
 	import { onMount } from 'svelte';
 	import {
 		listToolServers,
@@ -188,9 +189,7 @@
 		}
 	}
 
-	async function toggleEnabled(e: Event, s: ToolServer) {
-		e.stopPropagation();
-		const newVal = !s.enabled;
+	async function toggleEnabled(s: ToolServer, newVal: boolean) {
 		s.enabled = newVal;
 		servers = [...servers];
 		try {
@@ -222,10 +221,7 @@
 {:else}
 	<div>
 		{#each servers as s}
-			<button
-				class="group flex items-center gap-2 w-full h-7 text-left"
-				onclick={() => openEdit(s)}
-			>
+			<div class="group flex items-center gap-2 w-full h-7 text-left">
 				<span
 					class="text-[0.625rem] font-mono shrink-0
 					{s.type === 'mcp' || s.type === 'mcp_stdio'
@@ -234,33 +230,16 @@
 				>
 					{s.type === 'mcp' ? 'MCP' : s.type === 'mcp_stdio' ? 'STDIO' : 'API'}
 				</span>
-				<span
-					class="flex-1 text-[0.8125rem] truncate
+				<button
+					type="button"
+					class="flex-1 min-w-0 text-left text-[0.8125rem] truncate
 					{s.enabled ? 'text-gray-700 dark:text-gray-300' : 'text-gray-400 dark:text-gray-600'}"
+					onclick={() => openEdit(s)}
 				>
 					{s.name || s.command || s.url}
-				</span>
-				<!-- svelte-ignore a11y_no_static_element_interactions -->
-				<span
-					class="relative w-6 h-3.5 rounded-full shrink-0 cursor-pointer transition-colors duration-150
-						{s.enabled ? 'bg-gray-900 dark:bg-white' : 'bg-gray-200 dark:bg-gray-700'}"
-					role="switch"
-					tabindex="-1"
-					aria-checked={s.enabled}
-					onclick={(e) => toggleEnabled(e, s)}
-					onkeydown={(e) => {
-						if (e.key === 'Enter' || e.key === ' ') {
-							e.preventDefault();
-							toggleEnabled(e, s);
-						}
-					}}
-				>
-					<span
-						class="absolute top-0.5 w-2.5 h-2.5 rounded-full transition-all duration-150
-						{s.enabled ? 'left-3 bg-white dark:bg-gray-900' : 'left-0.5 bg-white dark:bg-gray-500'}"
-					></span>
-				</span>
-			</button>
+				</button>
+				<ToggleSwitch value={s.enabled} onchange={(value) => toggleEnabled(s, value)} />
+			</div>
 		{/each}
 
 		{#if servers.length === 0}
diff --git a/cptr/frontend/src/lib/components/Admin/Workspace.svelte b/cptr/frontend/src/lib/components/Admin/Workspace.svelte
new file mode 100644
index 0000000..a2859ee
--- /dev/null
+++ b/cptr/frontend/src/lib/components/Admin/Workspace.svelte
@@ -0,0 +1,69 @@
+<script lang="ts">
+	import { onMount } from 'svelte';
+	import { toast } from 'svelte-sonner';
+	import { getAdminConfig, updateConfig } from '$lib/apis/admin';
+	import { t } from '$lib/i18n';
+	import Spinner from '$lib/components/common/Spinner.svelte';
+	import ToggleSwitch from '$lib/components/common/ToggleSwitch.svelte';
+
+	let loading = $state(true);
+	let saving = $state(false);
+	let autoGitignoreDotCptr = $state(true);
+
+	onMount(async () => {
+		try {
+			const config = await getAdminConfig();
+			autoGitignoreDotCptr = config['workspace.auto_gitignore_dot_cptr'] !== false;
+		} catch {
+			toast.error($t('admin.failedToLoadConfig'));
+		} finally {
+			loading = false;
+		}
+	});
+
+	async function save() {
+		saving = true;
+		try {
+			await updateConfig({ 'workspace.auto_gitignore_dot_cptr': autoGitignoreDotCptr });
+			toast.success($t('settings.saved'));
+		} catch {
+			toast.error($t('admin.failedToSave'));
+		} finally {
+			saving = false;
+		}
+	}
+</script>
+
+<div class="flex flex-col min-h-full">
+	<h2 class="text-sm font-medium text-gray-900 dark:text-white mb-4">{$t('admin.workspace')}</h2>
+
+	{#if loading}
+		<div class="flex justify-center py-8"><Spinner size={16} /></div>
+	{:else}
+		<div class="flex flex-col gap-2.5">
+			<label class="flex items-center justify-between cursor-pointer">
+				<span class="text-xs text-gray-600 dark:text-gray-400"
+					>{$t('admin.workspaceAutoGitignoreDotCptr')}</span
+				>
+				<ToggleSwitch
+					value={autoGitignoreDotCptr}
+					onchange={(value) => {
+						autoGitignoreDotCptr = value;
+					}}
+					disabled={saving}
+				/>
+			</label>
+			<p class="text-[0.6875rem] text-gray-400 dark:text-gray-600 -mt-1">
+				{$t('admin.workspaceAutoGitignoreDotCptrHint')}
+			</p>
+		</div>
+
+		<div class="mt-auto pt-6 flex justify-end">
+			<button
+				class="text-[0.8125rem] text-gray-600 dark:text-gray-400 hover:text-gray-900 dark:hover:text-white transition-colors duration-100 disabled:opacity-50"
+				onclick={save}
+				disabled={saving}>{$t('settings.save')}</button
+			>
+		</div>
+	{/if}
+</div>
diff --git a/cptr/frontend/src/lib/components/SettingsModal.svelte b/cptr/frontend/src/lib/components/SettingsModal.svelte
index 7d3c7d0..44154bb 100644
--- a/cptr/frontend/src/lib/components/SettingsModal.svelte
+++ b/cptr/frontend/src/lib/components/SettingsModal.svelte
@@ -21,6 +21,7 @@
 	import AdminWeb from './Admin/Web.svelte';
 	import ToolServers from './Admin/ToolServers.svelte';
 	import Subagents from './Admin/Subagents.svelte';
+	import Workspace from './Admin/Workspace.svelte';
 	import { session } from '$lib/session';
 	import { t } from '$lib/i18n';
 
@@ -43,7 +44,8 @@
 		| 'images'
 		| 'web'
 		| 'toolservers'
-		| 'subagents';
+		| 'subagents'
+		| 'workspace';
 
 	interface Props {
 		onclose: () => void;
@@ -71,6 +73,7 @@
 		'web',
 		'toolservers',
 		'subagents',
+		'workspace',
 		'memory',
 		'skills'
 	];
@@ -99,6 +102,7 @@
 		{ id: 'web', label: $t('admin.web'), icon: 'globe' },
 		{ id: 'toolservers', label: $t('admin.toolServers'), icon: 'plug' },
 		{ id: 'subagents', label: $t('admin.subagents'), icon: 'user' },
+		{ id: 'workspace', label: $t('admin.workspace'), icon: 'folder' },
 		{ id: 'memory', label: $t('settings.memory'), icon: 'brain' },
 		{ id: 'skills', label: 'Skills', icon: 'spark' }
 	]);
@@ -218,6 +222,8 @@
 			<ToolServers />
 		{:else if activeTab === 'subagents'}
 			<Subagents />
+		{:else if activeTab === 'workspace'}
+			<Workspace />
 		{/if}
 	</div>
 </Modal>
diff --git a/cptr/frontend/src/lib/components/chat/AssistantMessage.svelte b/cptr/frontend/src/lib/components/chat/AssistantMessage.svelte
index f4e8f86..991be66 100644
--- a/cptr/frontend/src/lib/components/chat/AssistantMessage.svelte
+++ b/cptr/frontend/src/lib/components/chat/AssistantMessage.svelte
@@ -754,7 +754,7 @@
 							onmouseenter={() => (showUsageTooltip = true)}
 							onmouseleave={() => (showUsageTooltip = false)}
 							aria-label={$t('chat.usageInfo')}
-							use:tooltip={$t('chat.usageInfo')}
+							use:tooltip={showUsageTooltip ? null : $t('chat.usageInfo')}
 						>
 							<svg
 								xmlns="http://www.w3.org/2000/svg"
@@ -774,18 +774,15 @@
 						{#if showUsageTooltip}
 							<div
 								class="absolute bottom-full left-1/2 -translate-x-1/2 mb-1.5 z-50
-									bg-gray-900 dark:bg-gray-800 text-gray-100 dark:text-gray-100
-									rounded-lg shadow-lg px-2.5 py-1.5 text-[0.625rem] font-mono
+									app-surface rounded-lg shadow-lg px-2.5 py-1.5 text-[0.625rem] font-mono
 									whitespace-nowrap pointer-events-none
-									min-w-[10rem] border border-transparent dark:border-gray-700"
+									min-w-[10rem] border"
 							>
 								<div class="space-y-0.5">
 									{#each Object.entries(usage) as [key, value]}
 										<div class="flex justify-between gap-4">
-											<span class="text-gray-400 dark:text-gray-400">{formatUsageLabel(key)}</span>
-											<span class="tabular-nums text-white dark:text-gray-200"
-												>{formatUsageValue(key, value)}</span
-											>
+											<span class="app-muted">{formatUsageLabel(key)}</span>
+											<span class="tabular-nums">{formatUsageValue(key, value)}</span>
 										</div>
 									{/each}
 								</div>
@@ -794,7 +791,8 @@
 									class="absolute top-full left-1/2 -translate-x-1/2 w-0 h-0
 									border-l-[0.3125rem] border-l-transparent
 									border-r-[0.3125rem] border-r-transparent
-									border-t-[0.3125rem] border-t-gray-900 dark:border-t-gray-800"
+									border-t-[0.3125rem]"
+									style="border-top-color: var(--app-bg);"
 								></div>
 							</div>
 						{/if}
diff --git a/cptr/frontend/src/lib/components/chat/ChatPanel.svelte b/cptr/frontend/src/lib/components/chat/ChatPanel.svelte
index 0b16f7e..1f64586 100644
--- a/cptr/frontend/src/lib/components/chat/ChatPanel.svelte
+++ b/cptr/frontend/src/lib/components/chat/ChatPanel.svelte
@@ -532,6 +532,7 @@
 		delta?: string;
 		output?: any;
 		tasks?: ChatTask[];
+		context_usage?: ContextUsage | null;
 		done?: boolean;
 		error?: string;
 		pending_inputs_processed?: boolean;
@@ -574,6 +575,9 @@
 			chatTitle = data.title;
 			updateTab(tabId, data.chat_id, data.title);
 		}
+		if (data.context_usage) {
+			contextUsage = data.context_usage;
+		}
 
 		// Follow-up state changed server-side: reload to see new transcript/generation state.
 		if (data.pending_inputs_processed || data.async_subagent_pending) {
@@ -1024,12 +1028,12 @@
 			toast.message($t('chat.compactNoChat'));
 			return;
 		}
-		if (!selectedModel || sending || streaming) return;
+		if (sending || streaming) return;
 		sending = true;
 		inputText = '';
 		const toastId = toast.loading($t('chat.compacting'));
 		try {
-			const result = await apiCompactChat(chatId, selectedModel);
+			const result = await apiCompactChat(chatId, selectedModel || null);
 			contextUsage = result.context_usage ?? contextUsage;
 			if (result.compacted) {
 				toast.success($t('chat.compactDone'), { id: toastId });
diff --git a/cptr/frontend/src/lib/components/chat/UserMessage.svelte b/cptr/frontend/src/lib/components/chat/UserMessage.svelte
index ca01034..ce58ac7 100644
--- a/cptr/frontend/src/lib/components/chat/UserMessage.svelte
+++ b/cptr/frontend/src/lib/components/chat/UserMessage.svelte
@@ -1,6 +1,7 @@
 <script lang="ts">
 	import { tick } from 'svelte';
 	import Icon from '../Icon.svelte';
+	import MarkdownRenderer from '$lib/components/markdown/MarkdownRenderer.svelte';
 	import MessageTimestamp from './MessageTimestamp.svelte';
 	import { fileIconName } from '$lib/utils/fileIcon';
 	import { openFileTab, setFileBrowserCwd, setActiveTab } from '$lib/stores';
@@ -245,11 +246,10 @@
 		{/if}
 		<div class="flex justify-end">
 			<div class="max-w-[90%] px-4 py-2 rounded-3xl app-interactive-active">
-				<div
-					class="text-[0.8125rem] leading-relaxed text-gray-900 dark:text-gray-200 whitespace-pre-wrap break-words"
-				>
-					{#each parsedContent as segment}{#if segment.type === 'text'}{segment.value}{:else}{@const isDir =
-								segment.path.endsWith('/')}{@const cleanPath = isDir
+				<div class="text-[0.8125rem] leading-relaxed text-gray-900 dark:text-gray-200 break-words">
+					{#each parsedContent as segment}{#if segment.type === 'text'}<MarkdownRenderer
+								content={segment.value}
+							/>{:else}{@const isDir = segment.path.endsWith('/')}{@const cleanPath = isDir
 								? segment.path.slice(0, -1)
 								: segment.path}<button
 								class="inline-flex items-center gap-0.5 bg-blue-500/10 text-blue-400 rounded px-1.5 py-px mx-0.5 text-xs font-mono align-baseline border-none cursor-pointer hover:bg-blue-500/20 transition-colors"
diff --git a/cptr/frontend/src/lib/components/common/ToggleSwitch.svelte b/cptr/frontend/src/lib/components/common/ToggleSwitch.svelte
index fe4cb35..ed1793f 100644
--- a/cptr/frontend/src/lib/components/common/ToggleSwitch.svelte
+++ b/cptr/frontend/src/lib/components/common/ToggleSwitch.svelte
@@ -3,18 +3,25 @@
 		value: boolean;
 		onchange?: (value: boolean) => void;
 		disabled?: boolean;
+		ariaLabel?: string;
+	}
+	let { value, onchange, disabled = false, ariaLabel = 'Toggle' }: Props = $props();
+
+	function toggle(event: MouseEvent) {
+		event.stopPropagation();
+		onchange?.(!value);
 	}
-	let { value, onchange, disabled = false }: Props = $props();
 </script>
 
 <button
 	type="button"
 	class="relative w-7 h-4 rounded-full transition-colors duration-150 shrink-0
 		{value ? 'bg-gray-900 dark:bg-white' : 'bg-gray-300 dark:bg-gray-700'}"
-	onclick={() => onchange?.(!value)}
+	onclick={toggle}
 	{disabled}
 	role="switch"
 	aria-checked={value}
+	aria-label={ariaLabel}
 >
 	<span
 		class="absolute top-[0.125rem] w-3 h-3 rounded-full transition-all duration-150
diff --git a/cptr/frontend/src/lib/components/markdown/BlockRenderer.svelte b/cptr/frontend/src/lib/components/markdown/BlockRenderer.svelte
index 1b34bf9..ff27e50 100644
--- a/cptr/frontend/src/lib/components/markdown/BlockRenderer.svelte
+++ b/cptr/frontend/src/lib/components/markdown/BlockRenderer.svelte
@@ -153,7 +153,7 @@
 			{token.raw}
 		{/if}
 	{:else if token.type === 'html'}
-		<!-- Skip raw HTML for safety -->
+		{(token as Tokens.HTML).text || token.raw}
 	{:else if token.type === 'def'}
 		<!-- Link reference definition, no visual output -->
 	{:else if token.type === 'space'}
diff --git a/cptr/frontend/src/lib/components/markdown/InlineRenderer.svelte b/cptr/frontend/src/lib/components/markdown/InlineRenderer.svelte
index 1cec7e5..3fb004a 100644
--- a/cptr/frontend/src/lib/components/markdown/InlineRenderer.svelte
+++ b/cptr/frontend/src/lib/components/markdown/InlineRenderer.svelte
@@ -141,6 +141,8 @@
 				class="text-blue-500 dark:text-blue-400 bg-blue-500/8 dark:bg-blue-400/10 rounded px-1 cursor-pointer hover:underline transition-colors"
 				title={$t('markdown.linkTo', { target: wl.target })}>{wl.label}</span
 			>
+		{:else}
+			{item.raw}
 		{/if}
 	{/if}
 {/each}
diff --git a/cptr/frontend/src/lib/i18n/locales/en.json b/cptr/frontend/src/lib/i18n/locales/en.json
index ce65fc1..ad952ae 100644
--- a/cptr/frontend/src/lib/i18n/locales/en.json
+++ b/cptr/frontend/src/lib/i18n/locales/en.json
@@ -274,6 +274,9 @@
 	"admin.minCharsPassword": "Min 6 characters",
 	"admin.connections": "Connections",
 	"admin.agents": "Agents",
+	"admin.workspace": "Workspace",
+	"admin.workspaceAutoGitignoreDotCptr": "Add .cptr to .gitignore",
+	"admin.workspaceAutoGitignoreDotCptrHint": "Keeps local chats, memory, caches, and generated assets out of git by default. Turn this off if you want to track .cptr and manage exclusions inside .cptr/.gitignore.",
 	"admin.agentsRefresh": "Refresh",
 	"admin.agentsAddProfile": "Add profile",
 	"admin.agentsDeleteProfile": "Delete",
diff --git a/cptr/routers/admin.py b/cptr/routers/admin.py
index c84c55c..7921e70 100644
--- a/cptr/routers/admin.py
+++ b/cptr/routers/admin.py
@@ -235,6 +235,9 @@ async def create_connection(body: CreateConnectionRequest, request: Request):
         "name": body.name,
         "provider": body.provider,
         "api_type": body.api_type,  # "chat_completions" | "responses" (openai only)
+        "provider_type": "llama.cpp"
+        if body.provider == "openai" and body.provider_type == "llama.cpp"
+        else "default",
         "prefix_id": body.prefix_id,
         "base_url": body.base_url,
         "api_key": encrypt_key(body.api_key, _get_jwt_secret()) if body.api_key else None,
@@ -260,8 +263,16 @@ async def update_connection(conn_id: str, body: UpdateConnectionRequest, request
         conn["name"] = body.name
     if body.provider is not None:
         conn["provider"] = body.provider
+        if body.provider != "openai":
+            conn["provider_type"] = "default"
     if body.api_type is not None:
         conn["api_type"] = body.api_type
+    if "provider_type" in body.model_fields_set:
+        conn["provider_type"] = (
+            "llama.cpp"
+            if conn.get("provider") == "openai" and body.provider_type == "llama.cpp"
+            else "default"
+        )
     if body.prefix_id is not None:
         conn["prefix_id"] = body.prefix_id or None
     if body.base_url is not None:
@@ -389,6 +400,7 @@ class CreateConnectionRequest(BaseModel):
     name: str
     provider: str  # "anthropic" | "openai"
     api_type: str = "chat_completions"  # "chat_completions" | "responses" (openai only)
+    provider_type: Optional[str] = None  # None | "default" | "llama.cpp" (openai only)
     prefix_id: Optional[str] = None  # e.g. "openrouter" → "openrouter/model-id"
     base_url: Optional[str] = None
     api_key: Optional[str] = None
@@ -400,6 +412,7 @@ class UpdateConnectionRequest(BaseModel):
     name: Optional[str] = None
     provider: Optional[str] = None
     api_type: Optional[str] = None
+    provider_type: Optional[str] = None
     prefix_id: Optional[str] = None
     base_url: Optional[str] = None
     api_key: Optional[str] = None
diff --git a/cptr/routers/audio.py b/cptr/routers/audio.py
index e0ba52a..d63dc66 100644
--- a/cptr/routers/audio.py
+++ b/cptr/routers/audio.py
@@ -72,8 +72,8 @@ def _get_user(request: Request) -> str:
 def _workspace_audio_cache_dir(workspace: str | None, kind: str) -> Path | None:
     # Voice samples are project context, so they live under the workspace .cptr
     # folder and move with the project. That keeps STT and TTS artifacts available
-    # for reuse, debugging, and the local data flywheel while ensure_cptr_gitignored
-    # keeps them out of git.
+    # for reuse, debugging, and the local data flywheel. By default,
+    # ensure_cptr_gitignored keeps them out of git.
     if not workspace:
         return None
     ws = Path(workspace).expanduser().resolve()
diff --git a/cptr/routers/chat.py b/cptr/routers/chat.py
index aec71e8..326b851 100644
--- a/cptr/routers/chat.py
+++ b/cptr/routers/chat.py
@@ -457,6 +457,7 @@ async def _get_chat_context_usage(chat, model_id: str | None = None) -> dict | N
         estimate_context_usage,
         estimate_messages_tokens,
         load_compact_token_threshold,
+        usage_context_tokens,
     )
 
     messages, existing_summary = await _load_message_history(chat.id, message_id)
@@ -470,9 +471,8 @@ async def _get_chat_context_usage(chat, model_id: str | None = None) -> dict | N
     usage_checkpoint = await _get_latest_usage_checkpoint(chat.id, message_id)
     if usage_checkpoint:
         trailing_messages, usage = usage_checkpoint
-        tokens = usage.get("input_tokens")
-        if isinstance(tokens, int) and tokens > 0:
-            tokens += usage.get("output_tokens", 0)
+        tokens = usage_context_tokens(usage)
+        if tokens > 0:
             if trailing_messages:
                 tokens += estimate_messages_tokens(
                     [{"role": m.role, "content": m.content or ""} for m in trailing_messages]
@@ -495,6 +495,8 @@ async def _infer_chat_model(chat_id: str) -> str:
 async def _get_latest_usage_checkpoint(
     chat_id: str, message_id: str
 ) -> tuple[list[ChatMessage], dict] | None:
+    from cptr.utils.context import normalize_usage, usage_context_tokens
+
     all_msgs = await ChatMessage.get_all_by_chat(chat_id)
     msg_map = {m.id: m for m in all_msgs}
     chain = []
@@ -508,8 +510,8 @@ async def _get_latest_usage_checkpoint(
         message = chain[index]
         if message.chat_summary:
             return None
-        usage = message.usage or {}
-        if message.role == "assistant" and usage.get("input_tokens"):
+        usage = normalize_usage(message.usage)
+        if message.role == "assistant" and usage_context_tokens(usage):
             return chain[index + 1 :], usage
     return None
 
@@ -700,7 +702,7 @@ class SendMessageRequest(BaseModel):
 
 
 class CompactRequest(BaseModel):
-    model_id: str
+    model_id: Optional[str] = None
 
 
 @router.post("")
@@ -864,13 +866,16 @@ async def compact_chat(chat_id: str, body: CompactRequest, request: Request):
         raise HTTPException(404, "chat not found")
     if await _chat_has_active_generation(chat_id):
         raise HTTPException(409, "wait for the current response to finish before compacting")
+    model_id = body.model_id or await _infer_chat_model(chat_id)
+    if not model_id:
+        raise HTTPException(400, "choose a model before compacting")
 
     message_id = await _get_context_leaf_message_id(chat)
     if not message_id:
         return {"ok": True, "compacted": False, "reason": "empty", "context_usage": None}
     current_msg = await ChatMessage.get_by_id(message_id)
     if not current_msg or not current_msg.parent_id:
-        usage = await _get_chat_context_usage(chat, body.model_id)
+        usage = await _get_chat_context_usage(chat, model_id)
         return {"ok": True, "compacted": False, "reason": "too_short", "context_usage": usage}
 
     from cptr.utils.model_targets import (
@@ -879,16 +884,18 @@ async def compact_chat(chat_id: str, body: CompactRequest, request: Request):
         resolve_model_target,
     )
 
-    target = await resolve_model_target(body.model_id, request.app.state)
+    target = await resolve_model_target(model_id, request.app.state)
     if not isinstance(target, ApiModelTarget):
         target = await first_api_model_target(request.app.state)
 
-    from cptr.utils.chat_task import _load_message_history
+    from cptr.utils.chat_task import _load_message_history, _summary_checkpoint_message_id
     from cptr.utils.summarize import summarize_messages
 
     messages, existing_summary = await _load_message_history(chat_id, current_msg.parent_id)
-    if not messages:
-        usage = await _get_chat_context_usage(chat, body.model_id)
+    compacted_messages = messages[:-1]
+    keep_zone = messages[-1:]
+    if not compacted_messages or not keep_zone:
+        usage = await _get_chat_context_usage(chat, model_id)
         return {"ok": True, "compacted": False, "reason": "too_short", "context_usage": usage}
 
     connection = target.connection
@@ -901,7 +908,7 @@ async def compact_chat(chat_id: str, body: CompactRequest, request: Request):
     api_type = connection.get("api_type", "chat_completions")
 
     summary = await summarize_messages(
-        messages,
+        compacted_messages,
         existing_summary,
         provider,
         base_url,
@@ -909,12 +916,13 @@ async def compact_chat(chat_id: str, body: CompactRequest, request: Request):
         runtime_model,
         api_type=api_type,
     )
-    await ChatMessage.update(message_id, chat_summary=summary)
+    checkpoint_message_id = _summary_checkpoint_message_id(keep_zone, message_id)
+    await ChatMessage.update(checkpoint_message_id, chat_summary=summary)
 
     from cptr.utils.chat_export import export_chat_to_file
 
     await export_chat_to_file(chat_id)
-    usage = await _get_chat_context_usage(chat, body.model_id)
+    usage = await _get_chat_context_usage(chat, model_id)
     return {
         "ok": True,
         "compacted": True,
diff --git a/cptr/routers/git.py b/cptr/routers/git.py
index 90319f0..23d2b71 100644
--- a/cptr/routers/git.py
+++ b/cptr/routers/git.py
@@ -11,6 +11,7 @@
 from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel
 
+from cptr.utils.json_parser import extract_json
 from cptr.utils.git import (
     GitError,
     branches,
@@ -153,6 +154,23 @@ class CommitRequest(BaseModel):
 The description should be one short paragraph explaining the meaningful change. Do not invent details."""
 
 
+def _parse_commit_message_response(text: str) -> tuple[str, str]:
+    message = extract_json(text)
+    summary = str(message.get("summary", "")).strip() if isinstance(message, dict) else ""
+    description = str(message.get("description", "")).strip() if isinstance(message, dict) else ""
+    if summary:
+        return summary.splitlines()[0][:72], description
+
+    lines = [
+        line.strip()
+        for line in text.strip().splitlines()
+        if line.strip() and not line.strip().startswith("```")
+    ]
+    if not lines:
+        return "", ""
+    return lines[0][:72], "\n".join(lines[1:]).strip()
+
+
 class CheckoutRequest(BaseModel):
     root: str
     branch: str
@@ -260,7 +278,6 @@ async def generate_commit_message(body: CommitMessageRequest, request: Request):
     from cptr.utils.chat_task import _default_base_url
     from cptr.utils.config import _get_jwt_secret
     from cptr.utils.crypto import decrypt_key
-    from cptr.utils.json_parser import extract_json
     from cptr.utils.model_targets import ApiModelTarget, resolve_model_target
     from cptr.models import Config
 
@@ -285,12 +302,10 @@ async def generate_commit_message(body: CommitMessageRequest, request: Request):
         )
     except Exception as e:
         raise HTTPException(status_code=502, detail="Could not generate a commit message") from e
-    message = extract_json(text)
-    summary = str(message.get("summary", "")).strip().splitlines()[0] if isinstance(message, dict) else ""
-    description = str(message.get("description", "")).strip() if isinstance(message, dict) else ""
+    summary, description = _parse_commit_message_response(text)
     if not summary:
         raise HTTPException(status_code=502, detail="Could not generate a commit message")
-    return {"summary": summary[:72], "description": description}
+    return {"summary": summary, "description": description}
 
 
 @router.post("/checkout")
diff --git a/cptr/routers/workspace.py b/cptr/routers/workspace.py
index 0438de4..37da275 100644
--- a/cptr/routers/workspace.py
+++ b/cptr/routers/workspace.py
@@ -18,6 +18,8 @@
 from fastapi import APIRouter, HTTPException, Query
 from pydantic import BaseModel
 
+from cptr.utils.gitignore import is_gitignored, load_gitignore
+
 router = APIRouter(prefix="/api/workspace/files", tags=["workspace-files"])
 
 
@@ -355,22 +357,35 @@ def _is_search_ignored(name: str) -> bool:
     return name in SEARCH_IGNORE_DIRS or name.endswith(".egg-info")
 
 
-def _walk_match_entries(root: Path, show_hidden: bool):
+def _walk_match_entries(
+    root: Path,
+    show_hidden: bool,
+    gitignore: tuple[Path, tuple] | None = None,
+):
     """Yield visible files and directories without following symlinks."""
+    if gitignore is None:
+        gitignore = load_gitignore(root)
+    ignore_base, ignore_patterns = gitignore
+
     try:
         entries = sorted(root.iterdir(), key=lambda item: item.name.lower())
     except (OSError, PermissionError):
         return
 
     for item in entries:
-        if _is_search_ignored(item.name) or (not show_hidden and item.name.startswith(".")):
-            continue
         try:
+            item_is_dir = item.is_dir()
+            if (
+                _is_search_ignored(item.name)
+                or (not show_hidden and item.name.startswith("."))
+                or is_gitignored(item, ignore_base, ignore_patterns, is_dir=item_is_dir)
+            ):
+                continue
             if item.is_symlink():
                 yield item, "file"
-            elif item.is_dir():
+            elif item_is_dir:
                 yield item, "directory"
-                yield from _walk_match_entries(item, show_hidden)
+                yield from _walk_match_entries(item, show_hidden, gitignore)
             elif item.is_file():
                 yield item, "file"
         except (OSError, PermissionError):
@@ -408,7 +423,6 @@ def _content_matches_with_rg(
         "--column",
         "--max-count",
         str(MAX_CONTENT_MATCHES_PER_FILE + 1),
-        "--no-ignore",
     ]
     if show_hidden:
         args.append("--hidden")
@@ -489,7 +503,9 @@ def find_file_matches(
     files = {path.resolve() for path, kind in entries if kind == "file" and not path.is_symlink()}
     content_result = _content_matches_with_rg(root, query, query_lower, show_hidden, files)
     content_matches, _ = (
-        content_result if content_result is not None else _content_matches_with_python(files, query_lower)
+        content_result
+        if content_result is not None
+        else _content_matches_with_python(files, query_lower)
     )
 
     matches: list[tuple[int, int, FileMatch]] = []
@@ -593,13 +609,19 @@ def walk_and_rank_files(root: Path, query: str, limit: int = 20) -> list[SearchR
     query_lower = query.strip().lower().replace("\\", "/")
     matches: list[tuple[int, int, SearchResult]] = []
     max_collect = limit * 10
+    ignore_base, ignore_patterns = load_gitignore(root)
 
     def walk(directory: Path, depth: int = 0):
         if depth > 8 or len(matches) >= max_collect:
             return
         try:
             for item in sorted(directory.iterdir(), key=lambda p: p.name.lower()):
-                if item.name in SEARCH_IGNORE_DIRS or item.name.startswith("."):
+                item_is_dir = item.is_dir()
+                if (
+                    item.name in SEARCH_IGNORE_DIRS
+                    or item.name.startswith(".")
+                    or is_gitignored(item, ignore_base, ignore_patterns, is_dir=item_is_dir)
+                ):
                     continue
                 if len(matches) >= max_collect:
                     return
@@ -619,7 +641,7 @@ def walk(directory: Path, depth: int = 0):
                             SearchResult(
                                 path=str(item),
                                 name=item.name,
-                                type="directory" if item.is_dir() else "file",
+                                type="directory" if item_is_dir else "file",
                             ),
                         )
                     )
@@ -631,12 +653,12 @@ def walk(directory: Path, depth: int = 0):
                             SearchResult(
                                 path=str(item),
                                 name=item.name,
-                                type="directory" if item.is_dir() else "file",
+                                type="directory" if item_is_dir else "file",
                             ),
                         )
                     )
 
-                if item.is_dir():
+                if item_is_dir:
                     walk(item, depth + 1)
         except (PermissionError, OSError):
             pass
@@ -646,7 +668,6 @@ def walk(directory: Path, depth: int = 0):
     return [m[2] for m in matches[:limit]]
 
 
-
 # ── File management ──────────────────────────────────────────────
 
 
diff --git a/cptr/utils/ai.py b/cptr/utils/ai.py
index c858624..f922bb5 100644
--- a/cptr/utils/ai.py
+++ b/cptr/utils/ai.py
@@ -390,26 +390,40 @@ async def stream_anthropic(
 
 
 def _reasoning_items_to_content(items: list[dict]) -> str:
-    """Convert Responses API reasoning items to a reasoning_content string.
+    """Convert replayable reasoning items to a reasoning_content string.
 
     Reasoning items may expose text in content or summary blocks. We extract
     those text blocks and join them for Chat Completions-compatible providers.
     """
     texts: list[str] = []
     for item in items:
-        blocks = item.get("content") or item.get("summary") or []
-        for block in blocks:
-            if isinstance(block, dict) and block.get("type") == "text" and block.get("text"):
-                texts.append(block["text"])
+        for blocks in (item.get("content"), item.get("summary")):
+            if isinstance(blocks, str):
+                if blocks:
+                    texts.append(blocks)
+            elif isinstance(blocks, list):
+                for block in blocks:
+                    if isinstance(block, str):
+                        if block:
+                            texts.append(block)
+                    elif (
+                        isinstance(block, dict)
+                        and block.get("type") in ("text", "output_text", "summary_text")
+                        and block.get("text")
+                    ):
+                        texts.append(block["text"])
     return "\n".join(texts)
 
 
-def _to_openai_messages(messages: list[dict], instructions: str) -> list[dict]:
+def _to_openai_messages(
+    messages: list[dict], instructions: str, *, provider_type: str = "default"
+) -> list[dict]:
     """Canonical messages → OpenAI Chat Completions format.
 
-    Converts reasoning_items (Responses API format) to reasoning_content
-    (Chat Completions format) for providers like Kimi that require it.
-    Strips other non-standard fields (fc_id in tool_calls).
+    Strict default OpenAI-compatible requests do not receive provider-specific
+    reasoning fields.  llama.cpp compatibility replays text reasoning as
+    reasoning_content.  Other non-standard fields (fc_id in tool_calls) are
+    stripped.
     """
     result = []
     if instructions:
@@ -431,33 +445,42 @@ def _to_openai_messages(messages: list[dict], instructions: str) -> list[dict]:
                     data_uri = f"data:{block.get('media_type', 'image/jpeg')};base64,{block.get('base64', '')}"
                     formatted_content.append({"type": "image_url", "image_url": {"url": data_uri}})
             new_m = dict(m)
+            new_m.pop("reasoning_items", None)
+            if new_m.get("tool_calls"):
+                new_m["tool_calls"] = [
+                    {k: v for k, v in tc.items() if k != "fc_id"}
+                    for tc in new_m["tool_calls"]
+                ]
             new_m["content"] = formatted_content
-            # Convert reasoning_items → reasoning_content for round-tripping
-            ri = new_m.pop("reasoning_items", None)
-            if ri and new_m.get("role") == "assistant":
+            ri = m.get("reasoning_items")
+            if provider_type == "llama.cpp" and ri and new_m.get("role") == "assistant":
                 rc = _reasoning_items_to_content(ri)
                 if rc:
                     new_m["reasoning_content"] = rc
             result.append(new_m)
         else:
             out = dict(m)
-            # Convert reasoning_items → reasoning_content for round-tripping
-            ri = out.pop("reasoning_items", None)
-            if ri and out.get("role") == "assistant":
-                rc = _reasoning_items_to_content(ri)
-                if rc:
-                    out["reasoning_content"] = rc
-            # Clean tool_calls: remove fc_id which is Responses-API-only
+            out.pop("reasoning_items", None)
             if out.get("tool_calls"):
                 out["tool_calls"] = [
                     {k: v for k, v in tc.items() if k != "fc_id"} for tc in out["tool_calls"]
                 ]
+            ri = m.get("reasoning_items")
+            if provider_type == "llama.cpp" and ri and out.get("role") == "assistant":
+                rc = _reasoning_items_to_content(ri)
+                if rc:
+                    out["reasoning_content"] = rc
             result.append(out)
     return result
 
 
 async def stream_openai_completions(
-    form_data: ChatCompletionForm, url: str, key: str, *, request_params: dict | None = None
+    form_data: ChatCompletionForm,
+    url: str,
+    key: str,
+    *,
+    request_params: dict | None = None,
+    provider_type: str = "default",
 ) -> AsyncIterator[dict]:
     tools = [
         {
@@ -473,7 +496,9 @@ async def stream_openai_completions(
 
     body: dict = {
         "model": form_data.model,
-        "messages": _to_openai_messages(form_data.messages, form_data.instructions),
+        "messages": _to_openai_messages(
+            form_data.messages, form_data.instructions, provider_type=provider_type
+        ),
         "stream": True,
         "stream_options": {"include_usage": True},
     }
@@ -511,17 +536,21 @@ async def stream_openai_completions(
                         )
                     resp.raise_for_status()
                     tool_calls: dict[int, dict] = {}
-                    reasoning_buffer = ""  # Accumulate reasoning_content deltas (Kimi, DeepSeek)
+                    reasoning_buffer = ""
                     reasoning_item: dict | None = None
+                    reasoning_details: list = []
 
                     def complete_reasoning_item() -> dict | None:
-                        nonlocal reasoning_buffer, reasoning_item
+                        nonlocal reasoning_buffer, reasoning_item, reasoning_details
                         if reasoning_item is None:
                             return None
                         reasoning_item["status"] = "completed"
+                        if reasoning_details:
+                            reasoning_item["reasoning_details"] = copy.deepcopy(reasoning_details)
                         item = copy.deepcopy(reasoning_item)
                         reasoning_item = None
                         reasoning_buffer = ""
+                        reasoning_details = []
                         return item
 
                     async for line in resp.aiter_lines():
@@ -531,9 +560,25 @@ def complete_reasoning_item() -> dict | None:
                         choices = chunk.get("choices", [])
                         delta = choices[0].get("delta", {}) if choices else {}
 
-                        # Capture reasoning_content from providers like Kimi / DeepSeek
-                        if delta.get("reasoning_content"):
-                            reasoning_buffer += delta["reasoning_content"]
+                        reasoning_delta = (
+                            delta.get("reasoning_content")
+                            or delta.get("reasoning")
+                            or delta.get("thinking")
+                        )
+                        if delta.get("reasoning_details"):
+                            details = delta["reasoning_details"]
+                            if isinstance(details, list):
+                                reasoning_details.extend(copy.deepcopy(details))
+                            else:
+                                reasoning_details.append(copy.deepcopy(details))
+                            if reasoning_item is None:
+                                reasoning_item = _reasoning_output_item()
+                            reasoning_item["reasoning_details"] = copy.deepcopy(reasoning_details)
+                            emitted = True
+                            yield {"type": "output", "item": copy.deepcopy(reasoning_item)}
+
+                        if reasoning_delta:
+                            reasoning_buffer += reasoning_delta
                             if reasoning_item is None:
                                 reasoning_item = _reasoning_output_item()
                             reasoning_item["content"][0]["text"] = reasoning_buffer
@@ -617,7 +662,29 @@ def complete_reasoning_item() -> dict | None:
 # ── OpenAI Responses API ─────────────────────────────────────
 
 
-def _to_responses_input(messages: list[dict], instructions: str) -> list[dict]:
+def _replayable_reasoning_items(items: list[dict] | None, *, provider_type: str) -> list[dict]:
+    replayable: list[dict] = []
+    for item in items or []:
+        if (
+            not isinstance(item, dict)
+            or item.get("type") != "reasoning"
+            or item.get("status") not in (None, "completed")
+            or str(item.get("id", "")).startswith("reasoning-")
+        ):
+            continue
+        text = _reasoning_items_to_content([item])
+        if not text and not item.get("encrypted_content") and not item.get("reasoning_details"):
+            continue
+        out = copy.deepcopy(item)
+        if provider_type == "llama.cpp" and text:
+            out["content"] = [{"type": "text", "text": text}]
+        replayable.append(out)
+    return replayable
+
+
+def _to_responses_input(
+    messages: list[dict], instructions: str, *, provider_type: str = "default"
+) -> list[dict]:
     """Canonical messages → Responses API input items."""
     # Pre-collect all tool result call_ids so we can validate function_calls
     # have matching outputs.  This prevents orphaned function_calls (from
@@ -649,7 +716,9 @@ def _to_responses_input(messages: list[dict], instructions: str) -> list[dict]:
             )
         elif role == "assistant" and m.get("tool_calls"):
             # Emit reasoning items before function calls (required by reasoning models)
-            for ri in m.get("reasoning_items", []):
+            for ri in _replayable_reasoning_items(
+                m.get("reasoning_items"), provider_type=provider_type
+            ):
                 items.append(ri)
             for tc in m["tool_calls"]:
                 call_id = tc.get("id", "")
@@ -677,6 +746,11 @@ def _to_responses_input(messages: list[dict], instructions: str) -> list[dict]:
                     }
                 )
         else:
+            if role == "assistant":
+                for ri in _replayable_reasoning_items(
+                    m.get("reasoning_items"), provider_type=provider_type
+                ):
+                    items.append(ri)
             content = m.get("content", "")
             if isinstance(content, list):
                 formatted_content = []
@@ -695,7 +769,12 @@ def _to_responses_input(messages: list[dict], instructions: str) -> list[dict]:
 
 
 async def stream_openai_responses(
-    form_data: ChatCompletionForm, url: str, key: str, *, request_params: dict | None = None
+    form_data: ChatCompletionForm,
+    url: str,
+    key: str,
+    *,
+    request_params: dict | None = None,
+    provider_type: str = "default",
 ) -> AsyncIterator[dict]:
     tools = [
         {
@@ -709,7 +788,9 @@ async def stream_openai_responses(
 
     body: dict = {
         "model": form_data.model,
-        "input": _to_responses_input(form_data.messages, form_data.instructions),
+        "input": _to_responses_input(
+            form_data.messages, form_data.instructions, provider_type=provider_type
+        ),
         "stream": True,
     }
     if form_data.instructions:
@@ -749,6 +830,20 @@ async def stream_openai_responses(
                             error_body.decode(errors="replace"),
                         )
                     resp.raise_for_status()
+                    output_items_by_id: dict[str, dict] = {}
+                    active_reasoning_item: dict | None = None
+
+                    def get_reasoning_item(event: dict) -> dict:
+                        nonlocal active_reasoning_item
+                        item_id = event.get("item_id") or event.get("output_item_id") or "rs_stream"
+                        if item_id in output_items_by_id:
+                            active_reasoning_item = output_items_by_id[item_id]
+                        if active_reasoning_item is None:
+                            active_reasoning_item = _reasoning_output_item()
+                            active_reasoning_item["id"] = item_id
+                            output_items_by_id[item_id] = active_reasoning_item
+                        return active_reasoning_item
+
                     async for line in resp.aiter_lines():
                         if not line.startswith("data: "):
                             continue
@@ -762,8 +857,53 @@ async def stream_openai_responses(
                             emitted = True
                             yield {"type": "text_delta", "content": event["delta"]}
 
+                        elif etype == "response.output_item.added":
+                            item = event.get("item") or {}
+                            item_id = item.get("id")
+                            if item_id:
+                                output_items_by_id[item_id] = copy.deepcopy(item)
+                            if item.get("type") == "reasoning":
+                                active_reasoning_item = output_items_by_id.get(item_id) or item
+
+                        elif etype in (
+                            "response.reasoning.delta",
+                            "response.reasoning_text.delta",
+                        ):
+                            item = get_reasoning_item(event)
+                            delta = event.get("delta", "")
+                            content = item.setdefault("content", [{"type": "text", "text": ""}])
+                            if not content:
+                                content.append({"type": "text", "text": ""})
+                            if isinstance(content[0], dict):
+                                content[0]["type"] = content[0].get("type") or "text"
+                                content[0]["text"] = f"{content[0].get('text', '')}{delta}"
+                            emitted = True
+                            yield {"type": "output", "item": copy.deepcopy(item)}
+
+                        elif etype == "response.reasoning_summary_text.delta":
+                            item = get_reasoning_item(event)
+                            delta = event.get("delta", "")
+                            summary = item.setdefault("summary", [{"type": "summary_text", "text": ""}])
+                            if not summary:
+                                summary.append({"type": "summary_text", "text": ""})
+                            if isinstance(summary[0], dict):
+                                summary[0]["type"] = summary[0].get("type") or "summary_text"
+                                summary[0]["text"] = f"{summary[0].get('text', '')}{delta}"
+                            emitted = True
+                            yield {"type": "output", "item": copy.deepcopy(item)}
+
+                        elif etype == "response.reasoning_summary_part.added":
+                            item = get_reasoning_item(event)
+                            part = event.get("part")
+                            if part:
+                                item.setdefault("summary", []).append(copy.deepcopy(part))
+                            emitted = True
+                            yield {"type": "output", "item": copy.deepcopy(item)}
+
                         elif etype == "response.output_item.done":
                             item = event["item"]
+                            if item.get("id"):
+                                output_items_by_id[item["id"]] = copy.deepcopy(item)
                             if item["type"] == "function_call":
                                 emitted = True
                                 yield {
@@ -775,6 +915,11 @@ async def stream_openai_responses(
                                 }
                             elif item["type"] == "reasoning":
                                 # Reasoning items must be round-tripped for reasoning models
+                                if (
+                                    active_reasoning_item is not None
+                                    and active_reasoning_item.get("id") == item.get("id")
+                                ):
+                                    active_reasoning_item = None
                                 emitted = True
                                 yield {
                                     "type": "output",
@@ -787,6 +932,13 @@ async def stream_openai_responses(
                             raise RuntimeError(f"Responses API error: {msg}")
 
                         elif etype == "response.completed":
+                            if (
+                                active_reasoning_item is not None
+                                and _reasoning_items_to_content([active_reasoning_item])
+                            ):
+                                active_reasoning_item["status"] = "completed"
+                                emitted = True
+                                yield {"type": "output", "item": copy.deepcopy(active_reasoning_item)}
                             usage = event.get("response", {}).get("usage", {})
                             if usage:
                                 emitted = True
diff --git a/cptr/utils/browser/cdp.py b/cptr/utils/browser/cdp.py
index ce2f16e..aed7c26 100644
--- a/cptr/utils/browser/cdp.py
+++ b/cptr/utils/browser/cdp.py
@@ -325,12 +325,26 @@ async def scroll(self, direction: str = "down", amount: int = 3) -> None:
 
     # ── Observation ────────────────────────────────────────
 
-    async def screenshot(self) -> bytes:
+    async def screenshot(self, width: int | None = None, height: int | None = None) -> bytes:
         """Capture a screenshot of the current viewport. Returns PNG bytes."""
-        result = await self._send(
-            "Page.captureScreenshot", {"format": "png", "quality": 80}
-        )
-        return base64.b64decode(result["data"])
+        if width is not None and height is not None:
+            await self._send(
+                "Emulation.setDeviceMetricsOverride",
+                {
+                    "width": width,
+                    "height": height,
+                    "deviceScaleFactor": 1,
+                    "mobile": False,
+                },
+            )
+        try:
+            result = await self._send(
+                "Page.captureScreenshot", {"format": "png", "quality": 80}
+            )
+            return base64.b64decode(result["data"])
+        finally:
+            if width is not None and height is not None:
+                await self._send("Emulation.clearDeviceMetricsOverride")
 
     async def get_text(self) -> str:
         """Extract visible text content from the page."""
diff --git a/cptr/utils/chat_task.py b/cptr/utils/chat_task.py
index ece9122..598d8c0 100644
--- a/cptr/utils/chat_task.py
+++ b/cptr/utils/chat_task.py
@@ -14,7 +14,13 @@
 
 from cptr.events import EVENTS, publish_event
 from cptr.env import CHAT_MAX_ITERATIONS, CHAT_TOOL_COMMAND_MAX_CHARS, CHAT_TOOL_MAX_CHARS
-from cptr.utils.context import resolve_compact_token_threshold, should_compact
+from cptr.utils.context import (
+    build_context_usage,
+    normalize_usage,
+    resolve_compact_token_threshold,
+    should_compact,
+    usage_context_tokens,
+)
 from cptr.utils.skills import (
     bump_skill_view,
     discover_skills,
@@ -754,7 +760,7 @@ async def generate_chat_title(
 
         # Do not overwrite a title manually set while this request was running.
         chat = await Chat.get_by_id(chat_id)
-        fallback = user_content[:50].strip() or "New Chat"
+        fallback = user_message[:50].strip() or "New Chat"
         if not chat or chat.title != fallback:
             return
 
@@ -812,7 +818,7 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
         # (truly empty placeholder on first run)
         if m.id == message_id and not m.done and not m.content and not m.output:
             continue
-        entry: dict = {"role": m.role, "content": m.content or ""}
+        entry: dict = {"id": m.id, "role": m.role, "content": m.content or ""}
 
         # Transform uploaded images into base64 multimodal blocks; inline text files
         if m.role == "user":
@@ -903,6 +909,16 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
             for item in m.output:
                 itype = item.get("type")
                 if itype == "reasoning":
+                    if (
+                        item.get("status") not in (None, "completed")
+                        or str(item.get("id", "")).startswith("reasoning-")
+                        or (
+                            not item.get("encrypted_content")
+                            and not item.get("reasoning_details")
+                            and _reasoning_text_len(item) <= 0
+                        )
+                    ):
+                        continue
                     # A reasoning item after we already have outputs means
                     # a new API iteration started — flush the current turn.
                     if current_turn["outputs"]:
@@ -945,7 +961,7 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
                 # Skip other types (message, artifact, pending calls, etc.)
 
             # Don't forget the last turn
-            if current_turn["calls"] or current_turn["outputs"]:
+            if current_turn["reasoning"] or current_turn["calls"] or current_turn["outputs"]:
                 turns.append(current_turn)
 
             if not turns:
@@ -959,6 +975,18 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
                     call_names = {
                         tc["id"]: tc.get("function", {}).get("name", "") for tc in turn["calls"]
                     }
+                    if turn["reasoning"] and not matched_calls and not turn["outputs"]:
+                        if ti == 0:
+                            entry["reasoning_items"] = turn["reasoning"]
+                        else:
+                            result.append(entry)
+                            entry = {
+                                "id": m.id,
+                                "role": "assistant",
+                                "content": "",
+                                "reasoning_items": turn["reasoning"],
+                            }
+                        continue
                     if matched_calls:
                         if ti == 0:
                             # First turn: attach to the existing entry
@@ -970,6 +998,7 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
                             # result from previous turn) before creating a new one.
                             result.append(entry)
                             entry = {
+                                "id": m.id,
                                 "role": "assistant",
                                 "content": "",
                                 "tool_calls": matched_calls,
@@ -979,6 +1008,7 @@ async def _load_message_history(chat_id: str, message_id: str) -> tuple[list[dic
                     for out in turn["outputs"]:
                         result.append(entry)
                         entry = {
+                            "id": m.id,
                             "role": "tool",
                             "tool_call_id": out["call_id"],
                             "content": _tool_result_for_model(
@@ -1116,7 +1146,6 @@ def _tool_result_for_model(tool_name: str, result: str) -> str:
         }
     )
 
-
 def _append_tool_to_messages(
     messages: list[dict],
     event: dict,
@@ -1281,10 +1310,18 @@ def _scrub_incomplete_items(output_items: list[dict]) -> None:
 
 def _reasoning_text_len(item: dict) -> int:
     """Return displayable reasoning text length for diagnostics."""
-    blocks = item.get("summary") or item.get("content") or []
-    if not isinstance(blocks, list):
-        return 0
-    return sum(len(block.get("text") or "") for block in blocks if isinstance(block, dict))
+    total = 0
+    for blocks in (item.get("summary"), item.get("content")):
+        if isinstance(blocks, str):
+            total += len(blocks)
+        elif isinstance(blocks, list):
+            total += sum(
+                len(block.get("text") or "")
+                for block in blocks
+                if isinstance(block, dict)
+                and block.get("type") in ("text", "output_text", "summary_text")
+            )
+    return total
 
 
 def _output_debug_stats(output_items: list[dict] | None) -> tuple[dict[str, int], int, int]:
@@ -1390,6 +1427,16 @@ def _find_safe_split(messages: list[dict], target_keep: int) -> int:
     return min(split, n - 2)  # always keep at least 2
 
 
+def _summary_checkpoint_message_id(keep_zone: list[dict], fallback: str) -> str:
+    for message in keep_zone:
+        if message.get("role") == "user" and message.get("id"):
+            return message["id"]
+    for message in keep_zone:
+        if message.get("id"):
+            return message["id"]
+    return fallback
+
+
 # ── Connection resolution ───────────────────────────────────
 
 
@@ -2046,7 +2093,10 @@ async def _finish_reasoning_item():
                     api_type=api_type,
                 )
 
-                await ChatMessage.update(summary_message_id, chat_summary=summary)
+                checkpoint_message_id = _summary_checkpoint_message_id(
+                    keep_zone, summary_message_id
+                )
+                await ChatMessage.update(checkpoint_message_id, chat_summary=summary)
                 loaded_summary = summary
 
                 # Append summary to system prompt (works for all providers)
@@ -2077,7 +2127,7 @@ async def _finish_reasoning_item():
                 logger.info(
                     "[task %s] compacted: checkpoint=%s dropped %d msgs, kept %d, summary=%d chars",
                     message_id[:8],
-                    summary_message_id[:8],
+                    checkpoint_message_id[:8],
                     len(drop_zone),
                     len(keep_zone),
                     len(summary),
@@ -2114,6 +2164,7 @@ async def _finish_reasoning_item():
                             ],
                         }
                     )
+            api_messages = [{k: v for k, v in m.items() if k != "id"} for m in api_messages]
 
             form_data = ChatCompletionForm(
                 model=model,
@@ -2121,6 +2172,11 @@ async def _finish_reasoning_item():
                 instructions=system,
                 tools=tools,
             )
+            provider_type = (
+                "llama.cpp"
+                if provider == "openai" and connection.get("provider_type") == "llama.cpp"
+                else "default"
+            )
 
             if provider == "anthropic":
                 stream = stream_anthropic(
@@ -2128,11 +2184,19 @@ async def _finish_reasoning_item():
                 )
             elif connection.get("api_type") == "responses":
                 stream = stream_openai_responses(
-                    form_data, base_url, api_key, request_params=request_params
+                    form_data,
+                    base_url,
+                    api_key,
+                    request_params=request_params,
+                    provider_type=provider_type,
                 )
             else:
                 stream = stream_openai_completions(
-                    form_data, base_url, api_key, request_params=request_params
+                    form_data,
+                    base_url,
+                    api_key,
+                    request_params=request_params,
+                    provider_type=provider_type,
                 )
 
             restart = False
@@ -2188,13 +2252,16 @@ async def _finish_reasoning_item():
                     await emit(output=item)
 
                 elif event["type"] == "usage":
-                    usage = {k: v for k, v in event.items() if k != "type"}
-                    if "total_tokens" not in usage:
-                        usage["total_tokens"] = usage.get("input_tokens", 0) + usage.get(
-                            "output_tokens", 0
-                        )
+                    usage = normalize_usage({k: v for k, v in event.items() if k != "type"})
                     last_usage = usage
                     new_messages_since = 0
+                    tokens = usage_context_tokens(usage)
+                    if tokens > 0:
+                        await emit(
+                            context_usage=build_context_usage(
+                                tokens, threshold=compact_token_threshold, source="estimated"
+                            )
+                        )
 
                 elif event["type"] == "done":
                     # Stream ended. Usage may have arrived earlier, multiple times, or never.
diff --git a/cptr/utils/context.py b/cptr/utils/context.py
index 1a4e12c..ad04667 100644
--- a/cptr/utils/context.py
+++ b/cptr/utils/context.py
@@ -46,6 +46,7 @@ def should_compact(
     """
     resolved_threshold = threshold or _get_threshold()
 
+    last_usage = normalize_usage(last_usage)
     if last_usage and last_usage.get("input_tokens"):
         # Real base from last API call + estimate only new additions
         base = last_usage["input_tokens"] + last_usage.get("output_tokens", 0)
@@ -72,6 +73,36 @@ def build_context_usage(tokens: int, *, threshold: int | None = None, source: st
     }
 
 
+def normalize_usage(usage: dict | None) -> dict | None:
+    """Return provider usage with cptr's canonical token field names."""
+    if not usage:
+        return None
+    normalized = dict(usage)
+    input_tokens = _parse_nonnegative_int(
+        normalized.get("input_tokens", normalized.get("prompt_tokens"))
+    )
+    output_tokens = _parse_nonnegative_int(
+        normalized.get("output_tokens", normalized.get("completion_tokens"))
+    )
+    total_tokens = _parse_nonnegative_int(normalized.get("total_tokens"))
+    if total_tokens == 0:
+        total_tokens = input_tokens + output_tokens
+    normalized["input_tokens"] = input_tokens
+    normalized["output_tokens"] = output_tokens
+    normalized["total_tokens"] = total_tokens
+    return normalized
+
+
+def usage_context_tokens(usage: dict | None) -> int:
+    """Best available context token count from normalized or provider usage."""
+    usage = normalize_usage(usage)
+    if not usage:
+        return 0
+    if usage.get("input_tokens"):
+        return usage["input_tokens"] + usage.get("output_tokens", 0)
+    return usage.get("total_tokens", 0)
+
+
 def estimate_context_usage(
     messages: list[dict], system_prompt: str, *, threshold: int | None = None
 ) -> dict:
@@ -89,6 +120,14 @@ def _parse_positive_int(value: object) -> int | None:
     return parsed if parsed > 0 else None
 
 
+def _parse_nonnegative_int(value: object) -> int:
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return 0
+    return max(0, parsed)
+
+
 def resolve_compact_token_threshold(
     model: str | None = None,
     *,
diff --git a/cptr/utils/gitignore.py b/cptr/utils/gitignore.py
new file mode 100644
index 0000000..033366d
--- /dev/null
+++ b/cptr/utils/gitignore.py
@@ -0,0 +1,83 @@
+"""Small .gitignore matcher for fallback file walkers."""
+
+from __future__ import annotations
+
+from fnmatch import fnmatch
+from pathlib import Path
+
+GitignorePattern = tuple[str, bool, bool, bool]
+
+
+def load_gitignore(root: Path) -> tuple[Path, tuple[GitignorePattern, ...]]:
+    """Load the nearest repo/root .gitignore patterns for a search root."""
+    base = _ignore_base(root.resolve())
+    gitignore = base / ".gitignore"
+    if not gitignore.is_file():
+        return base, ()
+
+    patterns: list[GitignorePattern] = []
+    for raw in gitignore.read_text(encoding="utf-8", errors="replace").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#"):
+            continue
+        if line.startswith("\\#") or line.startswith("\\!"):
+            line = line[1:]
+        negated = line.startswith("!")
+        if negated:
+            line = line[1:].strip()
+        directory_only = line.endswith("/")
+        pattern = line.strip("/")
+        if pattern:
+            patterns.append((pattern, negated, directory_only, "/" in pattern))
+    return base, tuple(patterns)
+
+
+def is_gitignored(
+    path: Path,
+    base: Path,
+    patterns: tuple[GitignorePattern, ...],
+    *,
+    is_dir: bool,
+) -> bool:
+    if not patterns:
+        return False
+    try:
+        rel = path.resolve().relative_to(base).as_posix()
+    except ValueError:
+        return False
+
+    ignored = False
+    name = path.name
+    for pattern, negated, directory_only, anchored in patterns:
+        if directory_only and not is_dir:
+            continue
+        if _matches_gitignore_pattern(pattern, rel, name, directory_only, anchored):
+            ignored = not negated
+    return ignored
+
+
+def _ignore_base(root: Path) -> Path:
+    current = root if root.is_dir() else root.parent
+    fallback = current
+    for parent in (current, *current.parents):
+        if (parent / ".git").exists():
+            return parent
+        if (parent / ".gitignore").is_file():
+            fallback = parent
+    return fallback
+
+
+def _matches_gitignore_pattern(
+    pattern: str,
+    rel: str,
+    name: str,
+    directory_only: bool,
+    anchored: bool,
+) -> bool:
+    if anchored:
+        return (
+            rel == pattern
+            or (directory_only and rel.startswith(f"{pattern}/"))
+            or fnmatch(rel, pattern)
+        )
+    return any(fnmatch(part, pattern) for part in rel.split("/")) or fnmatch(name, pattern)
diff --git a/cptr/utils/skills.py b/cptr/utils/skills.py
index 53d4b7f..7b6aa85 100644
--- a/cptr/utils/skills.py
+++ b/cptr/utils/skills.py
@@ -685,6 +685,7 @@ def format_skill_content(skill: SkillContent) -> str:
         "",
         f"Skill directory: {skill_dir}",
         "Relative paths in this skill are relative to the skill directory.",
+        "Load listed resources with view_skill using this skill name and file_path.",
     ]
 
     if skill.resources:
diff --git a/cptr/utils/tools.py b/cptr/utils/tools.py
index 0e8ec0a..066a248 100644
--- a/cptr/utils/tools.py
+++ b/cptr/utils/tools.py
@@ -18,9 +18,11 @@
 import os
 import time
 import uuid
-from pathlib import Path
+from pathlib import Path, PureWindowsPath
 from typing import Any, Literal, Optional, get_args, get_origin, get_type_hints
+
 from cptr.env import CHAT_TOOL_COMMAND_MAX_CHARS, CHAT_TOOL_MAX_CHARS, EXECUTE_TIMEOUT
+from cptr.utils.gitignore import is_gitignored, load_gitignore
 
 try:
     import fcntl
@@ -716,13 +718,22 @@ def _read_text_for_search(fpath: Path) -> str | None:
     def _walk_and_search():
         results = []
         ignore = {".git", "node_modules", "__pycache__", ".venv", "venv"}
+        ignore_base, ignore_patterns = load_gitignore(full)
         q = query.lower() if case_insensitive else query
 
         for root, dirs, files in os.walk(full):
-            dirs[:] = [d for d in dirs if d not in ignore]
+            root_path = Path(root)
+            dirs[:] = [
+                d
+                for d in dirs
+                if d not in ignore
+                and not is_gitignored(root_path / d, ignore_base, ignore_patterns, is_dir=True)
+            ]
             for fname in files:
                 fpath = Path(root) / fname
-                if _is_dotenv(fpath):
+                if _is_dotenv(fpath) or is_gitignored(
+                    fpath, ignore_base, ignore_patterns, is_dir=False
+                ):
                     continue
                 text = _read_text_for_search(fpath)
                 if text is None:
@@ -767,7 +778,7 @@ async def create_file(
         artifact_path = artifact_dir / f"{ts}_{artifact_type}.md"
 
         def _write_artifact():
-            artifact_path.write_text(content)
+            artifact_path.write_text(content, encoding="utf-8")
 
         await asyncio.to_thread(_write_artifact)
 
@@ -793,7 +804,7 @@ def _write_artifact():
 
     def _write():
         full.parent.mkdir(parents=True, exist_ok=True)
-        full.write_text(content)
+        full.write_text(content, encoding="utf-8")
 
     await asyncio.to_thread(_write)
     return f"Created {path} ({len(content)} bytes, {len(content.splitlines())} lines)"
@@ -820,7 +831,7 @@ async def create_artifact(
     artifact_path = artifact_dir / f"{ts}_{artifact_type}.md"
 
     def _write():
-        artifact_path.write_text(content)
+        artifact_path.write_text(content, encoding="utf-8")
 
     await asyncio.to_thread(_write)
 
@@ -972,7 +983,7 @@ async def write_file(path: str, content: str, *, workspace: str) -> str:
 
     def _write():
         full.parent.mkdir(parents=True, exist_ok=True)
-        full.write_text(content)
+        full.write_text(content, encoding="utf-8")
 
     await asyncio.to_thread(_write)
     return f"Wrote {len(content)} bytes to {path}"
@@ -1043,7 +1054,7 @@ async def edit_file(
         return f"Error: file not found: {path}"
 
     def _edit():
-        content = full.read_text(errors="replace")
+        content = full.read_text(encoding="utf-8", errors="replace")
 
         if start_line > 0 or end_line > 0:
             lines = content.splitlines(keepends=True)
@@ -1075,7 +1086,7 @@ def _edit():
                 )
             new_content = content.replace(target, replacement, 1)
 
-        full.write_text(new_content)
+        full.write_text(new_content, encoding="utf-8")
         return None  # success sentinel
 
     result = await asyncio.to_thread(_edit)
@@ -1115,7 +1126,7 @@ async def multi_edit_file(
         return "Error: edits must be a non-empty JSON array"
 
     def _apply():
-        content = full.read_text(errors="replace")
+        content = full.read_text(encoding="utf-8", errors="replace")
         applied = 0
 
         for i, edit in enumerate(edit_list):
@@ -1138,7 +1149,7 @@ def _apply():
             content = content.replace(target, replacement, 1)
             applied += 1
 
-        full.write_text(content)
+        full.write_text(content, encoding="utf-8")
         return f"Applied {applied} edits to {path}"
 
     return await asyncio.to_thread(_apply)
@@ -1605,11 +1616,13 @@ async def delete_automation(
 
 async def view_skill(
     skill_name: str,
+    file_path: str = "",
     *,
     workspace: str,
 ) -> str:
     """Load the full instructions and resource listing for an available skill.
     :param skill_name: The name of the skill to load (from the <available_skills> catalog).
+    :param file_path: Optional relative path to a bundled skill file (for example references/api.md).
     """
     from cptr.models import Config
     from cptr.utils.skills import bump_skill_view, load_skill, format_skill_content
@@ -1617,14 +1630,51 @@ async def view_skill(
     if (await Config.get("skills.enabled")) in (False, "false", "0"):
         return "Error: skills are disabled by the administrator."
 
-    # Deduplication: if already activated, return short notice
-    if skill_name in _activated_skills:
-        return f"Skill '{skill_name}' is already loaded in this session. Refer to the existing <skill_content> above."
-
     skill = load_skill(workspace, skill_name)
     if not skill:
         return f"Error: skill '{skill_name}' not found. Check <available_skills> for valid names."
 
+    if file_path:
+        skill_dir = Path(skill.location).parent
+        p = Path(file_path)
+        windows_path = PureWindowsPath(file_path)
+        if p.is_absolute() or windows_path.is_absolute() or windows_path.drive:
+            return "Error: skill file path must be relative to the skill directory."
+        if ".." in p.parts or ".." in windows_path.parts:
+            return "Error: skill file path cannot contain '..' traversal components."
+
+        target = (skill_dir / file_path).resolve()
+        try:
+            target.relative_to(skill_dir.resolve())
+        except (ValueError, OSError):
+            return "Error: skill file path escapes the skill directory."
+        if _is_dotenv(target):
+            return _DOTENV_ERROR
+        if not target.is_file():
+            return f"Error: skill file not found: {file_path}"
+        if target.suffix.lower() in IMAGE_EXTENSIONS:
+            return await asyncio.to_thread(_read_image_file, target, file_path)
+
+        def _read_skill_file():
+            size = target.stat().st_size
+            if size > 500_000:
+                return f"Error: skill file too large ({size} bytes, max 500KB)"
+            try:
+                content = target.read_text(errors="strict")
+            except (UnicodeDecodeError, ValueError):
+                return f"Error: binary skill file ({target.suffix}), cannot read as text"
+            return (
+                f'<skill_file name="{skill.name}" path="{file_path}">\n'
+                f"{content}\n"
+                "</skill_file>"
+            )
+
+        return await asyncio.to_thread(_read_skill_file)
+
+    # Deduplication: if already activated, return short notice
+    if skill_name in _activated_skills:
+        return f"Skill '{skill_name}' is already loaded in this session. Refer to the existing <skill_content> above."
+
     _activated_skills.add(skill_name)
     bump_skill_view(workspace, skill_name, skill.source)
     return format_skill_content(skill)
@@ -1805,15 +1855,24 @@ async def browser_type(ref: str, text: str, *, __context__: dict) -> str:
     return await client.snapshot()
 
 
-async def browser_screenshot(*, __context__: dict) -> str:
-    """Take a screenshot of the current browser page. Saves the image to the workspace."""
+async def browser_screenshot(
+    width: Optional[int] = None, height: Optional[int] = None, *, __context__: dict
+) -> str:
+    """Take a screenshot of the current browser page. Saves the image to the workspace.
+    :param width: Optional screenshot viewport width in CSS pixels.
+    :param height: Optional screenshot viewport height in CSS pixels.
+    """
     cfg = await _get_browser_config()
     if cfg.get("provider", "local") != "local":
         return "Error: browser_screenshot requires Local CDP provider."
+    if (width is None) != (height is None):
+        return "Error: browser_screenshot width and height must be provided together."
+    if width is not None and height is not None and (width <= 0 or height <= 0):
+        return "Error: browser_screenshot width and height must be positive integers."
 
     chat_id = __context__.get("chat_id", "default")
     client = await _get_cdp_session(chat_id)
-    png_bytes = await client.screenshot()
+    png_bytes = await client.screenshot(width=width, height=height)
 
     # Save to workspace
     workspace = __context__.get("workspace", ".")
diff --git a/cptr/utils/workspace.py b/cptr/utils/workspace.py
index c71ae34..1d22624 100644
--- a/cptr/utils/workspace.py
+++ b/cptr/utils/workspace.py
@@ -4,9 +4,30 @@
 
 from pathlib import Path
 
+from cptr import env
+from cptr.utils.config import load_config
+
+
+def auto_gitignore_cptr_enabled() -> bool:
+    if env.WORKSPACE_AUTO_GITIGNORE_DOT_CPTR_ENV is not None:
+        return env.WORKSPACE_AUTO_GITIGNORE_DOT_CPTR
+
+    config = load_config()
+    value = None
+    workspace = config.get("workspace", {})
+    if isinstance(workspace, dict):
+        value = workspace.get("auto_gitignore_dot_cptr")
+    app_config = config.get("app_config", {})
+    if isinstance(app_config, dict):
+        value = app_config.get("workspace.auto_gitignore_dot_cptr", value)
+    return _bool_config(value, default=True)
+
 
 def ensure_cptr_gitignored(workspace: str | Path) -> None:
     """If workspace is a git repo, ensure .cptr is listed in .gitignore."""
+    if not auto_gitignore_cptr_enabled():
+        return
+
     ws = Path(workspace)
     if not (ws / ".git").exists():
         return
@@ -26,3 +47,15 @@ def ensure_cptr_gitignored(workspace: str | Path) -> None:
         gitignore.write_text(content, encoding="utf-8")
     else:
         gitignore.write_text(f"{entry}\n", encoding="utf-8")
+
+
+def _bool_config(value: object, default: bool) -> bool:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in {"true", "1", "yes", "on"}:
+            return True
+        if normalized in {"false", "0", "no", "off"}:
+            return False
+    return default
diff --git a/pyproject.toml b/pyproject.toml
index dbbb454..4926e96 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "cptr"
-version = "0.9.7"
+version = "0.9.8"
 description = "Your computer, from anywhere. Code, manage, and control your machine from the web."
 license = {file = "LICENSE"}
 readme = "README.md"
diff --git a/uv.lock b/uv.lock
index 66b52a9..18d5688 100644
--- a/uv.lock
+++ b/uv.lock
@@ -284,7 +284,7 @@ wheels = [
 
 [[package]]
 name = "cptr"
-version = "0.9.7"
+version = "0.9.8"
 source = { editable = "." }
 dependencies = [
     { name = "aiosqlite" },