diff --git a/docs/zeus-prompt-cache-hud.md b/docs/zeus-prompt-cache-hud.md
new file mode 100644
index 00000000..e8966286
--- /dev/null
+++ b/docs/zeus-prompt-cache-hud.md
@@ -0,0 +1,52 @@
+# Prompt cache HUD
+
+A status bar item that shows AI cost and prompt-cache state in real time. The point is transparency: developers should be able to feel each AI call's cost so they can change behavior, not be surprised at the end of the month.
+
+This is positioned directly against Cursor's credit model, which most users describe as opaque. We show the raw numbers; users can hide it if they don't care.
+
+## Status bar layout
+
+```text
+⚡ 2 agents · 92% cache · $0.003 / $0.41 today
+```
+
+- **`⚡ N agents`**: number of currently-running subagents. Clickable → opens parallel-agents view
+- **`X% cache`**: rolling cache hit ratio over the last 100 requests. State (window + day totals) lives on the singleton **main-process** service `IAiCostService`, not directly in `IStorageService`, so multiple workbench windows opened against different workspaces stay consistent. The main process persists snapshots to `IStorageService.APPLICATION` (key `zeus.ai.cache.window`) on a 1s debounce so a hard kill doesn't lose more than ~1s of data. The debounce is durability-only — renderer windows stay in sync via IPC subscriptions, not by re-reading storage, so write frequency does not affect UI latency or cross-window consistency.
+- **`$X.XXX`**: cost of the most recent AI call
+- **`$X.XX today`**: cumulative cost for the local day (resets midnight)
+
+All visible strings (`agents`, `cache`, `today`) ship through `nls.localize` so translations land alongside other UI localisation. The currency symbol and decimal separator are formatted via `Intl.NumberFormat(locale, { style: 'currency', currency: 'USD' })`, with a setting `zeus.ai.hud.currency` to override the display currency for users whose Anthropic billing is in another currency.
+
+Hovering over each segment shows a tooltip with the breakdown (input tokens / output tokens / cached tokens / cost per million).
+
+## Source of data
+
+- Live agent count: `IAgentRuntime` event stream (`feat/agent-sdk`)
+- Per-call cost and cache state: Anthropic SDK `usage` field, mapped to current model pricing
+- Today's cumulative: owned by the `IAiCostService` singleton in the **main process** (ensures atomic updates across all open workbench windows — `IStorageService` writes from concurrent renderers would race and silently lose updates). The main process snapshots the state to `IStorageService.APPLICATION` (per-user, cross-workspace) under key `zeus.ai.cost` holding `{ date: "YYYY-MM-DD", total: number }`. Application scope (not workspace) because the user's per-day spend should not reset when switching between workspaces — the goal is to surface real cost, not per-project cost. When local midnight passes, the record is reset before the next write, so storage doesn't grow per day. A future `zeus.ai.hud.scope` setting can flip it to `WORKSPACE` if a user wants per-project tracking.
+
+## Configuration
+
+- `zeus.ai.hud.enabled` (default: `true`) — show the HUD at all
+- `zeus.ai.hud.detail` (`"compact" | "verbose"`) — controls the format
+- `zeus.ai.hud.todayLimit` (number | null) — soft cap in USD (matches the units shown in the status bar); turns the cost segment red when exceeded, no enforcement. `null` disables the colouring.
+- `zeus.ai.hud.stalePricingDays` (number | null, default `30`) — show the `⚠` stale-pricing glyph once the bundled pricing file is older than this many days. `null` disables the warning entirely for users in restricted environments (corporate-locked editor versions, offline installs) where update cadence is out of their control.
+
+The HUD is implemented as **multiple adjacent `StatusBarItem`s** (agents, cache, cost, today). VS Code's `StatusBarItem` API does not support per-segment coloring inside a single item, so the colored "over limit" treatment lives on its own item.
+
+## Why no enforcement
+
+Hard credit caps are what makes Cursor frustrating. Zeus shows the number; the user decides whether to stop. If you want enforcement, the local LLM path or a custom MCP proxy can give you that.
+
+## Acceptance criteria (real impl)
+
+- [ ] Status bar item appears when an Anthropic AI feature is configured
+- [ ] Real-time updates within ~200ms of each request completing
+- [ ] Hover tooltip shows token / cost breakdown
+- [ ] `today` value persists across editor restarts in the same local day
+- [ ] Setting `zeus.ai.hud.enabled = false` hides the item entirely
+- [ ] Pricing table lives in a bundled JSON file (`src/vs/workbench/contrib/aiHud/common/anthropicPricing.json`) shipped with the build. Updated by a dependabot-style PR when the upstream price page changes — see `script/refresh-pricing.mjs`. The HUD never makes a live network call for pricing on a hot path (latency + offline). If the file is older than `zeus.ai.hud.stalePricingDays` (default `30`) the HUD shows a small `⚠` glyph next to the cost segment and the tooltip says `"Pricing data from {date}; estimates may be stale — update Zeus"`. The user-visible numbers continue to use the bundled table; the warning is visible because the transparency goal of this feature is broken if users silently look at outdated estimates. Setting the threshold to `null` suppresses the glyph for users whose editor version cadence is out of their control.
+
+## Status
+
+Slot reserved at `src/vs/workbench/contrib/aiHud/`. Depends on `IAgentRuntime` (`feat/agent-sdk`).
diff --git a/src/vs/workbench/contrib/aiHud/README.md b/src/vs/workbench/contrib/aiHud/README.md
new file mode 100644
index 00000000..d41813bf
--- /dev/null
+++ b/src/vs/workbench/contrib/aiHud/README.md
@@ -0,0 +1,5 @@
+# `aiHud` contribution
+
+Slot for the status bar HUD that shows AI cost and prompt-cache state. Design at [`docs/zeus-prompt-cache-hud.md`](../../../../../docs/zeus-prompt-cache-hud.md).
+
+Reads from `IAgentRuntime` (`feat/agent-sdk`). No enforcement, just transparency.
diff --git a/src/vs/workbench/contrib/mcp/test/node/mcpStdioStateHandler.test.ts b/src/vs/workbench/contrib/mcp/test/node/mcpStdioStateHandler.test.ts
index 57283241..8bba4ab8 100644
--- a/src/vs/workbench/contrib/mcp/test/node/mcpStdioStateHandler.test.ts
+++ b/src/vs/workbench/contrib/mcp/test/node/mcpStdioStateHandler.test.ts
@@ -9,9 +9,18 @@ import * as assert from 'assert';
 import { McpStdioStateHandler } from '../../node/mcpStdioStateHandler.js';
 import { isWindows } from '../../../../../base/common/platform.js';
 
-const GRACE_TIME = 100;
+// 1000ms gives the child enough time on slow CI runners to handle
+// SIGTERM and flush stdout before the parent escalates to SIGKILL. 100ms
+// was racy on Linux containers; 250ms still tripped under load. The
+// test's `delay >= GRACE_TIME` assertion still scales correctly.
+const GRACE_TIME = 1000;
 
-suite('McpStdioStateHandler', () => {
+// `sigkill after grace` waits at least GRACE_TIME * 2 = 2000ms, which
+// collides with mocha's default 2000ms test timeout. Raise the suite
+// timeout so the SIGTERM-then-SIGKILL path has room without becoming
+// flaky again.
+suite('McpStdioStateHandler', function () {
+	this.timeout(10_000);
 	const store = ensureNoDisposablesAreLeakedInTestSuite();
 
 	function run(code: string) {
@@ -53,11 +62,7 @@ suite('McpStdioStateHandler', () => {
 		assert.strictEqual(result.trim(), 'Data received: Hello MCP!');
 	});
 
-	// Flaky on shared CI: the child process can exit before its
-	// post-SIGTERM stdout flush lands, so the test sees 'stdin ended'
-	// only and not 'stdin ended\nSIGTERM received'. Skip on CI until
-	// the upstream subprocess flush race is properly fixed.
-	if (!isWindows && !process.env.CI && !process.env.GITHUB_ACTIONS) {
+	if (!isWindows) {
 		test('sigterm after grace', async () => {
 			const { handler, output } = run(`
 			setInterval(() => {}, 1000);