dsh-connect-agnes-token-plan 0.10.2 → 0.11.0

package/CHANGELOG.md

@@ -2,6 +2,24 @@
 
 本文件只记**公开行为变化**（新增能力、破坏性改动、重要修复）。实现细节、重构与测试加固请直接看 `git log`。
 
+## [0.11.0] — 2026-10-16
+
+本版两条主线：新能力**对话框用量胶囊**（输入框右侧的第二条读数，按 provider 门控），与**面板操作体感**批次（ADR-014）落地。
+
+### 对话框用量胶囊：输入框右侧的第二条读数（两个上游、一个座席，provider 门控）
+
+- **新：注册 `conversation.input.right` composer 座席并挂载 `ComposerUsagePill`**——读数按当前会话模型的 provider 门控：`agnes-token-plan` 模型画控制台 5 小时 / 每周请求窗口的使用百分比，`agnescode` 模型画桌面端积分余额（画存量、不造百分比：网关只发积分不给上限，把余额换算成百分比需要发明一个它占谁的份额）；**其它 provider 的会话一律不挂载**（无读取、无定时器、无读数）——别的插件的会话绝不该带这些数字，`agnescode` 的额度宇宙也是桌面网关自己的，不能给它画控制台的窗口。
+- **门控坐在组件边界而非 hook 之后**：外层组件只决定挂哪个内层组件，轮询 hook 全在内层里——先调 hook 再早退会违反 Rules of Hooks，React 会把组件丢弃到会话结束（姊妹插件有专门钉住这条的回归 scenario）。
+- **读数与面板同源、不造第二把尺**：百分比走 `windowPercent`（平台自己的 `usagePct` 优先、`used/limit` 只作兜底、**从不钳制**——超额窗口报 >100 是事实不是渲染 bug），严重度点走 `usageTier`（90 / 70，与面板进度条同源）；AgnesCode 侧 `0` 是「耗尽」读数、字段缺席是「读不到」（破折号），不是零。
+- **软失败（§5 不变量②）**：读取失败保留最后读数并标记过期，不清空、不抛错；注册失败面板照常工作、该模块缺席。
+- 影响面：`src/client/{composer-usage-pill,const,apply,cards,agnescode-tab,i18n,styles,index}.ts` + `client.js`（随 src 同 commit 重建）；`test/composer-slot-gate.mjs`（新增：钉住座席 / 渲染器 / projection / 物化契约与两个 provider id 的双侧拼写）、`test/dom-render.test.mjs`（胶囊真 DOM scenario）、`test/render.test.mjs` / `test/contract.test.mjs`（§11c）；文档与 CI 门禁同步。
+
+### 面板操作体感：快照取数并行 + 保存回显不再拖全量快照（ADR-014）
+
+- **可感变化（账号失效 / 控制台降级场景）**：① 快照五源取数全部并行（原 `overview` 先行再批量；失效账号下两条 15 s 超时链叠加，单次快照最坏 30 s 起，现压到 ~15 s；死令牌锁号不重新引入——`getToken` 单飞 + `rejected` 集 + 逐请求 401 处置，测试 B 组按新形状重钉：每源至多一次出示 / 每 poll 一次登录）；② 接入 API tab 的 provider 开关 / 模型清单 / 出图·视频工具卡的「保存后全量快照 reload」改为把 POST 响应体就地并入快照 `llm` 块——关模型、切开关的反馈不再等下一轮慢轮询；③ AgnesCode 三动作（开关 / 检测本机登录态 / 保存模型清单）直接应用 POST 的完整响应，harvest 成功即时翻出已关联卡片（此前要再排一次最长 30 s 的 GET）；④ `/agnescode` 路由的 balance / catalog 两个 30 s 上游读改并行（最坏 60 s → 30 s）。工具卡「实际生效模型」徽标与额度数字仍随 30 s 轮询更新（Host 计算值，本地不造数）。
+- **归位说明**：此批次（①–④）在 0.10.2 发布（2026-10-10）之后的补记提交里被记在 `[0.10.2]` 节下，但已发布的 0.10.2 tarball 里**没有其中任何一条**（0.10.2 的 `overview` 仍先行单独取数、客户端半边尚无 `LlmEchoPatch`）；`[0.10.2]` 下的「面板操作体感」节移入本节订正。
+- 影响面：`src/host/{snapshot-aggregate,routes/agnescode}.ts`、`src/client/{panel-page,provider-controls,model-picker,wire,agnescode-tab}.ts`、`test/routes.test.mjs`、`docs/`（ADR-014、ARCHITECTURE §3、PROVIDER-HOT-RELOAD、AGNES-API、PITFALLS）、`lib/` + `client.js`（随 src 同 commit 重建）。
+
 ## [0.10.2] — 2026-10-10
 
 ### 修复：视频状态轮询断点续查、额度渲染与模型花名册的几处真实缺陷 + 凭据隔离加固

package/README.md

@@ -1,11 +1,11 @@
 # dsh-connect-agnes-token-plan
 
-把 Agnes 接进 DSH 的 **Plugins 页**插件卡。**DSH** 即 **DeepSeek Harness**；**Agnes**（爱思）是全模态大模型平台，本插件对接其**控制台 API** 与**推理 API**。
+把 Agnes 接进 DSH：**Plugins 页**的插件卡（三个 tab），外加**会话输入框右侧的一枚用量胶囊**。**DSH** 即 **DeepSeek Harness**；**Agnes**（爱思）是全模态大模型平台，本插件对接其**控制台 API** 与**推理 API**。
 
 
 > **第一次读这个仓库？先看这三句。**
 >
-> 1. **这是给 DeepSeek Harness 用的插件**，不是独立应用；所有入口都在 DSH 的 Plugins 页那张卡上（三个 tab，见下）。
+> 1. **这是给 DeepSeek Harness 用的插件**，不是独立应用；面板入口在 DSH 的 **Plugins 页**那张卡上（三个 tab，见下），另有一枚会话内的用量胶囊（见「对话框用量胶囊」）。
 > 2. **三种「红」是正常的，不是回归**：① `test/wiring.test.mjs` 等 peer 套件报 `cannot resolve the peer dependency` —— 那是 DSH runtime 自带的包不在本仓库，属环境问题，先查 `test/peer-roots.mjs`；② 没装 `dsh` CLI 时 e2e 会 SKIP，属预期；③ 若在 **23:00–00:00** 之间单独跑 `render.test.mjs` 报红，那是历史上的午夜窗口缺陷（已在 0.10.0 修复，见 [docs/PITFALLS.md](docs/PITFALLS.md) §48），当前安装版本已含，无需升级。
 > 3. **见到「奇怪」的写法先别定罪**——本仓多数刻意的取舍都在 [docs/PITFALLS.md](docs/PITFALLS.md) / [docs/ADR.md](docs/ADR.md) 有明文出处。**动手前先 grep 一遍 docs/**。
 >
@@ -58,6 +58,14 @@
 
 注意：检测不到登录态时，面板会逐条列出探测过哪些文件、各自为什么没成（「没装 App」「解不开密」「会话里没有令牌」是三种不同的处理）；
 
+## 对话框用量胶囊（自动出现）
+
+输入框右侧会出现一枚用量胶囊，**画哪条上游的额度由当前会话选中的模型决定**，前缀写明是谁的：选中 `agnes-token-plan` 的模型显示控制台的 **5 小时 / 本周**已用百分比（前缀 `Token Plan`）；选中 `agnescode`（桌面端）的模型显示那套**积分余额**（前缀 `AgnesCode`）。圆点分级（绿 / 黄 / 红 / 灰），余下的细节——重置时刻、账号昵称与时效 / 永久拆分、以及「为什么没有数字」——都在悬停提示里。
+
+- 读数与面板**同源**：控制台百分比优先用平台自己报的 `usage_pct`、缺失才回落 `已用 / 上限`，**不做钳制**（超额窗口就显示 >100%，那是事实）；桌面端画的是**存量**，所以**不造百分比**——网关只公布积分，没有「上限」可以当分母；
+- 两条上游**互不借用**：`agnescode` 的会话只读它自己的路由、绝不画控制台窗口，反之亦然；切到**别的 provider**（其它插件）的模型时胶囊直接消失，也不再发任何请求；
+- 读不到时**保留上一次数字并标「上次数据」**，不会清零或伪装成 0%；余额里的 `0` 是「烧完了」的读数（红色告警），`—` 才是「这次没读到」——这个区别是刻意的。
+
 ## 边界
 
 - 面板**不代你操作账务**——不改套餐、不代扣额度、不碰 Key 明文；数据来自 Agnes 控制台自己的 API，与网页控制台口径一致。
@@ -104,7 +112,7 @@ npm run doctor:json     # 机器读：JSON（可进你的巡检 / 工单脚本
 - [docs/API.md](docs/API.md) — 路由与控制台端点
 - [docs/AGNES-API.md](docs/AGNES-API.md) — Agnes 接口全集（控制台额度侧 + 推理侧）
 - [docs/TESTING.md](docs/TESTING.md) — 测试体系
-- [docs/PITFALLS.md](docs/PITFALLS.md) — 真实踩坑（55 条）
+- [docs/PITFALLS.md](docs/PITFALLS.md) — 真实踩坑（56 条）
 - [CHANGELOG.md](CHANGELOG.md) — 版本变化
 
 AI 协作会话请先读 [AGENTS.md](AGENTS.md)。人类贡献者先读 [SETUP.md](docs/SETUP.md) 与 [ARCHITECTURE.md](docs/ARCHITECTURE.md) §2 理解结构与改动代价，改代码前再读 [AGENTS.md](AGENTS.md) 的红线与 [TESTING.md](docs/TESTING.md) 的 `npm test` 门禁。

package/client.js

@@ -13,10 +13,12 @@ var dsh_connect_agnes_token_plan_client = (function() {
 
 //#endregion
 //#region src/client/const.ts
-	var PANEL_ID, NS, SNAPSHOT_PATH, ACCOUNT_PATH, API_KEY_PATH, PROVIDER_PATH, MODELS_PATH, DRAW_PATH, VIDEO_PATH, AGNESCODE_PATH, AGNES_SIGNUP_URL, AGNESCODE_SITE_URL;
+	var PANEL_ID, NS, PROVIDER_ID, AGNESCODE_PROVIDER_ID, SNAPSHOT_PATH, ACCOUNT_PATH, API_KEY_PATH, PROVIDER_PATH, MODELS_PATH, DRAW_PATH, VIDEO_PATH, AGNESCODE_PATH, AGNES_SIGNUP_URL, AGNESCODE_SITE_URL;
 	var init_const = __esmMin((() => {
 		PANEL_ID = "dsh-connect-agnes-token-plan";
 		NS = PANEL_ID;
+		PROVIDER_ID = "agnes-token-plan";
+		AGNESCODE_PROVIDER_ID = "agnescode";
 		SNAPSHOT_PATH = `/api/${PANEL_ID}/snapshot`;
 		ACCOUNT_PATH = `/api/${PANEL_ID}/account`;
 		API_KEY_PATH = `/api/${PANEL_ID}/api-key`;
@@ -108,6 +110,17 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			"quota.total.video": "视频",
 			"quota.total.activeDays": "活跃天数",
 			"quota.usageMissing": "{label}：暂未读到。",
+			"composer.planPrefix": "Token Plan",
+			"composer.window5h": "5 小时",
+			"composer.windowWeekly": "本周",
+			"composer.pillTitle": "Token Plan 额度",
+			"composer.notConnected": "控制台未连接",
+			"composer.noWindows": "控制台未返回额度窗口",
+			"composer.balancePrefix": "AgnesCode",
+			"composer.balanceTitle": "AgnesCode 积分余额",
+			"composer.balanceUnit": "积分",
+			"composer.balanceUnread": "本次没能读到余额",
+			"composer.stale": "上次数据",
 			"usage.none": "该区间内没有用量记录。",
 			"usage.period": "近 {days} 天",
 			"usage.periodUnknown": "用量区间",
@@ -330,6 +343,17 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			"quota.total.video": "Video",
 			"quota.total.activeDays": "Active days",
 			"quota.usageMissing": "{label}: not read yet.",
+			"composer.planPrefix": "Token Plan",
+			"composer.window5h": "5h",
+			"composer.windowWeekly": "this week",
+			"composer.pillTitle": "Token Plan quota",
+			"composer.notConnected": "console not connected",
+			"composer.noWindows": "the console returned no quota windows",
+			"composer.balancePrefix": "AgnesCode",
+			"composer.balanceTitle": "AgnesCode credit balance",
+			"composer.balanceUnit": "credits",
+			"composer.balanceUnread": "the balance could not be read this time",
+			"composer.stale": "last reading",
 			"usage.none": "No usage recorded in this range.",
 			"usage.period": "Last {days} days",
 			"usage.periodUnknown": "Usage range",
@@ -1344,7 +1368,42 @@ var dsh_connect_agnes_token_plan_client = (function() {
 				fontSize: 11,
 				lineHeight: "16px",
 				marginBottom: 2
-			}
+			},
+			pillTrigger: {
+				appearance: "none",
+				display: "inline-flex",
+				alignItems: "center",
+				gap: 5,
+				height: 28,
+				padding: "0 6px",
+				flexShrink: 0,
+				background: "none",
+				border: "none",
+				color: "inherit",
+				fontFamily: "inherit",
+				fontSize: 12,
+				lineHeight: "20px",
+				whiteSpace: "nowrap",
+				userSelect: "none",
+				cursor: "default"
+			},
+			pillDot: {
+				width: 7,
+				height: 7,
+				borderRadius: 999,
+				flexShrink: 0
+			},
+			pillPlan: {
+				color: "var(--dsw-alias-label-secondary)",
+				flexShrink: 0
+			},
+			pillReading: { fontVariantNumeric: "tabular-nums" },
+			pillStale: { color: "var(--dsw-alias-label-secondary)" },
+			pillErrorText: { color: "var(--dsw-alias-state-error-primary)" },
+			pillTierOk: { background: "var(--dsw-alias-state-success-primary, #22c55e)" },
+			pillTierWarn: { background: "var(--dsw-alias-state-warn-primary, #f59e0b)" },
+			pillTierError: { background: "var(--dsw-alias-state-error-primary, #ef4444)" },
+			pillTierUnknown: { background: "var(--dsw-alias-label-secondary)" }
 		};
 	}));
 
@@ -1861,14 +1920,16 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* The curated model allow-list: which of this key's models get pushed to
 	* DSH's model list.
 	*
-	* Hook-based like `ApiKeyForm`, so the render suite exercises the secret-
+	* Hook-based like `ApiKeyForm`, so the Node render suite exercises the secret-
 	* free half it draws - `ModelRoster` and the counts - instead of this
-	* state machine. The draft/dirty/saved machinery now lives in
-	* `roster-draft.ts`, shared with `AgnescodeModelPicker`; what stays here is
-	* only what this roster does differently — it POSTs to `MODELS_PATH` itself
-	* and words its errors with the `llm.roster*` keys.
+	* state machine. The hook path itself (search filter, tick→dirty footer, save
+	* POST) is mounted for real by `test/dom-render.test.mjs` (jsdom + real
+	* react-dom over the shipped artifact). The draft/dirty/saved machinery now
+	* lives in `roster-draft.ts`, shared with `AgnescodeModelPicker`; what stays
+	* here is only what this roster does differently — it POSTs to `MODELS_PATH`
+	* itself and words its errors with the `llm.roster*` keys.
 	*/
-	function ModelPicker({ llm, onDone, tt }) {
+	function ModelPicker({ llm, onDone, onEcho, tt }) {
 		const models = Array.isArray(llm?.models) ? llm.models : [];
 		const hostIds = Array.isArray(llm?.enabledModelIds) ? llm.enabledModelIds : [];
 		const { ids, busy, setBusy, query, setQuery, setSavedKey, notice, setNotice, dirty, justSaved, visible, tickedCount, bulk, toggle, discard } = useRosterDraft(models, hostIds);
@@ -1878,9 +1939,15 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			setNotice(null);
 			const posted = JSON.stringify(ids);
 			try {
-				await postJsonOrThrow(MODELS_PATH, { enabledModelIds: ids });
+				const body = await postJsonOrThrow(MODELS_PATH, { enabledModelIds: ids });
 				setSavedKey(posted);
-				onDone?.();
+				if (onEcho) onEcho({
+					enabledModelIds: Array.isArray(body.enabledModelIds) ? body.enabledModelIds : ids,
+					registerProvider: body.registerProvider === true,
+					providerRegistered: body.providerRegistered === true,
+					providerError: typeof body.providerError === "string" ? body.providerError : ""
+				});
+				else onDone?.();
 			} catch (error) {
 				setNotice(format(tt("llm.rosterError"), { error: error instanceof Error ? error.message : String(error) }));
 			} finally {
@@ -1890,6 +1957,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			busy,
 			ids,
 			onDone,
+			onEcho,
 			tt,
 			setBusy,
 			setNotice,
@@ -2213,10 +2281,12 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* restart. Hook-based like `ApiKeyForm`, so the render suite (which
 	* cannot mount hooks) exercises the secret-free status lines instead;
 	* the route itself is covered by `routes.test.mjs`. The state shown is
-	* the SNAPSHOT's effective value, never local optimism — the poll after
-	* `onDone` repaints whatever the Host actually reports.
+	* the SNAPSHOT's effective value, never local optimism: the flip's OWN
+	* POST answer (the Host's post-write state) is merged into that snapshot
+	* by `onEcho`, and the plain `onDone` full reload is only the fallback
+	* for mounts without an echo hook.
 	*/
-	function ProviderSwitch({ llm, onDone, tt }) {
+	function ProviderSwitch({ llm, onDone, onEcho, tt }) {
 		const [busy, setBusy] = useState(false);
 		const [switchError, setSwitchError] = useState(null);
 		const enabled = llm?.registerProvider === true;
@@ -2224,8 +2294,14 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			setBusy(true);
 			setSwitchError(null);
 			try {
-				await postJsonOrThrow(PROVIDER_PATH, { enabled: !enabled });
-				onDone?.();
+				const body = await postJsonOrThrow(PROVIDER_PATH, { enabled: !enabled });
+				if (onEcho) onEcho({
+					registerProvider: body.registerProvider === true,
+					registerSource: typeof body.registerSource === "string" ? body.registerSource : "",
+					providerRegistered: body.providerRegistered === true,
+					providerError: typeof body.providerError === "string" ? body.providerError : ""
+				});
+				else onDone?.();
 			} catch (error) {
 				setSwitchError(format(tt("llm.switchError"), { error: error instanceof Error ? error.message : String(error) }));
 			} finally {
@@ -2234,6 +2310,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		}, [
 			enabled,
 			onDone,
+			onEcho,
 			tt
 		]);
 		return h("div", null, h(Switch, {
@@ -2267,7 +2344,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* `useCallback` through, so `render.test.mjs` pins the picker's row markup —
 	* the `modelRowHead` wrap below is that contract.
 	*/
-	function ToolSwitch({ path, prefix, modelKey, enabled, hasKey, candidates, preferred, effective, extraNote, onDone, tt }) {
+	function ToolSwitch({ path, prefix, modelKey, enabledKey, sourceKey, preferredKey, enabled, hasKey, candidates, preferred, effective, extraNote, onDone, onEcho, tt }) {
 		const [busy, setBusy] = useState(false);
 		const [switchError, setSwitchError] = useState(null);
 		const k = (suffix) => tt(`${prefix}.${suffix}`);
@@ -2275,8 +2352,15 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			setBusy(true);
 			setSwitchError(null);
 			try {
-				await postJsonOrThrow(path, { enabled: !enabled });
-				onDone?.();
+				const body = await postJsonOrThrow(path, { enabled: !enabled });
+				if (onEcho) {
+					const patch = {};
+					if (enabledKey === "drawEnabled") patch.drawEnabled = body[enabledKey] === true;
+					else patch.videoEnabled = body[enabledKey] === true;
+					if (sourceKey === "drawSource") patch.drawSource = typeof body[sourceKey] === "string" ? body[sourceKey] : "";
+					else patch.videoSource = typeof body[sourceKey] === "string" ? body[sourceKey] : "";
+					onEcho(patch);
+				} else onDone?.();
 			} catch (error) {
 				setSwitchError(format(k("switchError"), { error: error instanceof Error ? error.message : String(error) }));
 			} finally {
@@ -2285,16 +2369,25 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		}, [
 			path,
 			prefix,
+			enabledKey,
+			sourceKey,
 			enabled,
 			onDone,
+			onEcho,
 			tt
 		]);
 		const saveModel = useCallback(async (id) => {
 			setBusy(true);
 			setSwitchError(null);
 			try {
-				await postJsonOrThrow(path, { [modelKey]: id });
-				onDone?.();
+				const body = await postJsonOrThrow(path, { [modelKey]: id });
+				if (onEcho) {
+					const patch = {};
+					const picked = typeof body[modelKey] === "string" ? body[modelKey] : "";
+					if (preferredKey === "drawPreferredModel") patch.drawPreferredModel = picked === "" ? void 0 : picked;
+					else patch.videoPreferredModel = picked === "" ? void 0 : picked;
+					onEcho(patch);
+				} else onDone?.();
 			} catch (error) {
 				setSwitchError(format(k("switchError"), { error: error instanceof Error ? error.message : String(error) }));
 			} finally {
@@ -2304,7 +2397,9 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			path,
 			modelKey,
 			prefix,
+			preferredKey,
 			onDone,
+			onEcho,
 			tt
 		]);
 		const statusText = enabled ? hasKey ? k("onList") : k("needsKey") : k("off");
@@ -2384,17 +2479,21 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* shows the effective value immediately. The route is covered by
 	* `routes.test.mjs`.
 	*/
-	function DrawSwitch({ llm, onDone, tt }) {
+	function DrawSwitch({ llm, onDone, onEcho, tt }) {
 		return h(ToolSwitch, {
 			path: DRAW_PATH,
 			prefix: "draw",
 			modelKey: "drawModelId",
+			enabledKey: "drawEnabled",
+			sourceKey: "drawSource",
+			preferredKey: "drawPreferredModel",
 			enabled: llm?.drawEnabled === true,
 			hasKey: llm?.hasApiKey === true,
 			candidates: rosterIds(llm?.drawCandidateIds),
 			preferred: llm?.drawPreferredModel != null ? String(llm.drawPreferredModel) : null,
 			effective: String(llm?.drawModel ?? ""),
 			onDone,
+			onEcho,
 			tt
 		});
 	}
@@ -2411,7 +2510,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* nearest whole second. The card's note line says so, rather than leaving a
 	* reader wondering why a seconds-based model sits next to a frame-based one.
 	*/
-	function VideoSwitch({ llm, onDone, tt }) {
+	function VideoSwitch({ llm, onDone, onEcho, tt }) {
 		const excluded = rosterIds(llm?.video25ModelIds);
 		const extraNote = excluded.length > 0 ? h("div", { style: {
 			...S.muted,
@@ -2425,6 +2524,9 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			path: VIDEO_PATH,
 			prefix: "video",
 			modelKey: "videoModelId",
+			enabledKey: "videoEnabled",
+			sourceKey: "videoSource",
+			preferredKey: "videoPreferredModel",
 			enabled: llm?.videoEnabled === true,
 			hasKey: llm?.hasApiKey === true,
 			candidates: rosterIds(llm?.videoCandidateIds),
@@ -2432,6 +2534,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			effective: String(llm?.videoModel ?? ""),
 			extraNote,
 			onDone,
+			onEcho,
 			tt
 		});
 	}
@@ -2443,14 +2546,16 @@ var dsh_connect_agnes_token_plan_client = (function() {
 	* constituents (`ProviderSwitch`, `ProviderRegStatus`, and the `ModelPicker`
 	* it composes) rather than in the key editor's file.
 	*/
-	function ProviderForm({ llm, onDone, tt }) {
+	function ProviderForm({ llm, onDone, onEcho, tt }) {
 		return h("div", null, h(ProviderSwitch, {
 			llm,
 			onDone,
+			onEcho,
 			tt
 		}), h(ModelPicker, {
 			llm,
 			onDone,
+			onEcho,
 			tt
 		}), h(ProviderRegStatus, {
 			llm,
@@ -2607,20 +2712,63 @@ var dsh_connect_agnes_token_plan_client = (function() {
 
 //#endregion
 //#region src/client/cards.ts
-/** The bar fill and figure tone for a usage percentage: 70 warn / 90 danger. */
+/**
+	* One percentage's severity: 90 danger / 70 warn, else ok.
+	*
+	* Its own function because a SECOND surface reads the same thresholds — the
+	* composer usage pill paints a status dot with them (`composer-usage-pill.ts`).
+	* Two copies of "90 / 70" would be two answers to "when is a window in
+	* trouble", and the panel's tonal vocabulary is not something to re-invent per
+	* surface.
+	*/
+	function usageTier(pct) {
+		if (pct >= 90) return "error";
+		if (pct >= 70) return "warn";
+		return "ok";
+	}
+	/**
+	* The bar fill for a usage percentage, by {@link usageTier}.
+	*
+	* The fill is the whole answer: the bar is the only surface that colours by
+	* tier, and the figure keeps the panel's one text colour. (This used to also
+	* return a `color` "for the figure", which no caller ever read — a second
+	* answer to a question nobody asked.)
+	*/
 	function usageTone(pct) {
-		if (pct >= 90) return {
-			fill: S.barFillError,
-			color: "var(--dsw-alias-state-error-primary)"
-		};
-		if (pct >= 70) return {
-			fill: S.barFillWarn,
-			color: "var(--dsw-alias-state-warn-primary)"
-		};
-		return {
-			fill: S.barFill,
-			color: "var(--dsw-alias-label-secondary)"
-		};
+		const tier = usageTier(pct);
+		if (tier === "error") return S.barFillError;
+		if (tier === "warn") return S.barFillWarn;
+		return S.barFill;
+	}
+	/**
+	* One window's consumption percentage — the panel's single reading of it.
+	*
+	* The percentage is the platform's own `usagePct`, quoted verbatim — never a
+	* re-derivation. The Host parses it from the subscription response and ships it
+	* down for exactly this purpose (docs/API.md §快照, AGNES-API.md: the bar is a
+	* transcription, not a computation). Two reasons it is NOT `used/limit` first:
+	*
+	*  - the platform's denominator need not be `limit` (it may be a separate
+	*    quota ceiling for the same window), so the arithmetic silently shows a
+	*    different number than the one the console shows;
+	*  - clamping to 100 would rewrite the platform's own value. A window in
+	*    overage reports >100, and that is a fact, not a rendering bug.
+	*
+	* `used / limit` remains only as the fallback for a Host that predates the
+	* field or omits it for this window — a shape mismatch must still render.
+	* `null` means "no reading", which is a different claim from `0`.
+	*
+	* Exported because the composer pill quotes the same two windows and must not
+	* grow a second opinion about what a window's percentage is.
+	*/
+	function windowPercent(window) {
+		if (window === null || typeof window !== "object") return null;
+		const source = window;
+		const platformPct = typeof source.usagePct === "number" && Number.isFinite(source.usagePct) ? source.usagePct : null;
+		if (platformPct !== null) return platformPct;
+		const limit = Number(source.limit) || 0;
+		const used = typeof source.used === "number" && Number.isFinite(source.used) ? source.used : null;
+		return used !== null && limit > 0 ? used / limit * 100 : null;
 	}
 	/**
 	* The window's own name, from the key the Host emits.
@@ -2643,18 +2791,26 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		return format(tt("quota.perHours"), { hours: value });
 	}
 	/**
-	* The unit a dimension is counted in — `""` for video.
+	* The unit a dimension is counted in — `""` only where naming one would be an
+	* invention.
 	*
 	* Video is deliberately unitless. The platform's limit field is
 	* `video_daily_limit` while its usage field is `total_video_seconds`, and it
 	* never says which one the limit is expressed in, so printing "500 秒" would
 	* assert something nobody verified.
+	*
+	* An UNKNOWN unit falls back to the raw name, the same rule `windowLabel`
+	* follows for an unknown key: a Host that grew a dimension this bundle has
+	* never heard of must show up as itself. Returning `""` there would drop the
+	* noun and print a naked "1500" — the exact "reads as available capacity"
+	* shape this card exists to avoid, and silent where an unknown KEY is loud.
 	*/
 	function unitOf(unit, tt) {
 		const name = String(unit ?? "");
 		if (name === "requests") return tt("quota.unit.requests");
 		if (name === "images") return tt("quota.unit.images");
-		return "";
+		if (name === "video") return "";
+		return name;
 	}
 	/** `monthly` / `yearly` as the panel phrases them. */
 	function cycleLabel(cycle, tt) {
@@ -2687,9 +2843,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		const limit = Number(source.limit) || 0;
 		const unit = unitOf(source.unit, tt);
 		const used = typeof source.used === "number" && Number.isFinite(source.used) ? source.used : null;
-		const platformPct = typeof source.usagePct === "number" && Number.isFinite(source.usagePct) ? source.usagePct : null;
-		const pct = platformPct !== null ? platformPct : used !== null && limit > 0 ? used / limit * 100 : null;
-		const tone = usageTone(pct ?? 0);
+		const pct = windowPercent(source);
 		const resetAt = typeof source.resetAt === "number" && Number.isFinite(source.resetAt) ? source.resetAt : null;
 		const resetInSeconds = typeof source.resetInSeconds === "number" && Number.isFinite(source.resetInSeconds) ? source.resetInSeconds : null;
 		const resetLine = resetAt !== null ? format(tt("quota.resetAt"), { time: clockSameDay(resetAt) }) : resetInSeconds !== null ? format(tt("quota.resetCountdown"), { minutes: Math.max(1, Math.round(resetInSeconds / 60)) }) : null;
@@ -2699,9 +2853,10 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			"aria-label": `${label} ${tt("quota.used")} ${pct.toFixed(1)}%`,
 			"aria-valuenow": pct.toFixed(1),
 			"aria-valuemin": 0,
-			"aria-valuemax": 100
+			"aria-valuemax": Math.max(100, Math.ceil(pct)),
+			"aria-valuetext": `${pct.toFixed(1)}%`
 		}, h("div", { style: {
-			...tone.fill,
+			...usageTone(pct),
 			width: `${pct}%`
 		} })), (pct === null || used === null) && resetLine === null ? null : h("div", { style: S.quotaFoot }, pct === null || used === null ? null : h("span", { style: S.quotaUsed }, `${tt("quota.used")} ${count(used)} / ${count(limit)}`), resetLine === null ? null : h("span", { style: {
 			...S.muted,
@@ -3084,11 +3239,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 					action: "switch",
 					enabled
 				});
-				if (alive.current) setState((current) => current ? {
-					...current,
-					enabled: body.enabled === true,
-					providerRegistered: body.providerRegistered === true
-				} : current);
+				if (alive.current) setState(body);
 			} catch (why) {
 				if (alive.current) setNote(format(tt("agnescode.switchError"), { error: why instanceof Error ? why.message : String(why) }));
 			}
@@ -3099,32 +3250,23 @@ var dsh_connect_agnes_token_plan_client = (function() {
 			try {
 				const body = await postJson(AGNESCODE_PATH, { action: "harvest" });
 				if (alive.current) {
-					const harvestBlock = body?.harvest;
-					const attempts = Array.isArray(harvestBlock?.attempts) ? harvestBlock.attempts : [];
-					setState((current) => current ? {
-						...current,
-						harvest: {
-							ok: body?.ok === true,
-							attempts
-						}
-					} : current);
+					if (body !== null) setState(body);
 					if (body?.ok === true) setNote(tt("agnescode.harvestOk"));
 					else setNote(tt("agnescode.harvestFail"));
-					load();
 				}
 			} catch (why) {
 				if (alive.current) setNote(format(tt("agnescode.error"), { error: why instanceof Error ? why.message : String(why) }));
 			} finally {
 				if (alive.current) setHarvestBusy(false);
 			}
-		}, [load, tt]);
+		}, [tt]);
 		const saveModels = useCallback(async (ids) => {
-			await postJsonOrThrow(AGNESCODE_PATH, {
+			const body = await postJsonOrThrow(AGNESCODE_PATH, {
 				action: "saveModels",
 				enabledModelIds: ids
 			});
-			if (alive.current) load();
-		}, [load]);
+			if (alive.current) setState(body);
+		}, []);
 		const enabled = state?.enabled === true;
 		const view = agnescodeView(state, error);
 		const loggedIn = view.linked;
@@ -3560,6 +3702,21 @@ var dsh_connect_agnes_token_plan_client = (function() {
 				if (inFlight.current === controller) inFlight.current = null;
 			}
 		}, []);
+		const dataRef = useRef(null);
+		dataRef.current = data;
+		const echoLlm = useCallback((patch) => {
+			if (dataRef.current === null) {
+				load();
+				return;
+			}
+			setData((current) => current === null ? current : {
+				...current,
+				llm: {
+					...current.llm ?? {},
+					...patch
+				}
+			});
+		}, [load]);
 		const toggleSection = useCallback((key) => {
 			setOpenSections((current) => ({
 				...current,
@@ -3749,6 +3906,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		}, h(ProviderForm, {
 			llm: data?.llm ?? null,
 			onDone: () => void load(),
+			onEcho: echoLlm,
 			tt
 		})), h(SectionCard, {
 			title: tt("draw.title"),
@@ -3758,6 +3916,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		}, h(DrawSwitch, {
 			llm: data?.llm ?? null,
 			onDone: () => void load(),
+			onEcho: echoLlm,
 			tt
 		})), h(SectionCard, {
 			title: tt("video.title"),
@@ -3767,6 +3926,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		}, h(VideoSwitch, {
 			llm: data?.llm ?? null,
 			onDone: () => void load(),
+			onEcho: echoLlm,
 			tt
 		})), h(SectionCard, {
 			title: tt("llm.title"),
@@ -3816,6 +3976,404 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		init_agnescode_tab();
 	}));
 
+//#endregion
+//#region src/client/composer-usage-pill.ts
+/**
+	* The provider the session is currently pointed at, `next` winning over
+	* `lastUsed` — with one repair: an EMPTY `next` (the transient frame right after
+	* a picker switch, before it carries the new selection) must not shadow a
+	* still-usable `lastUsed`. A bare `next ?? lastUsed` reads a provider off `{}`,
+	* finds none, and hides the pill for OWNED models on exactly the frames a switch
+	* produces.
+	*/
+	function selectedProviderOf(projection) {
+		const usable = (candidate) => candidate !== null && candidate !== void 0 && (typeof candidate.provider === "string" || typeof candidate.model === "string");
+		const p = projection ?? {};
+		const provider = (usable(p.next) ? p.next : usable(p.lastUsed) ? p.lastUsed : void 0)?.provider;
+		return typeof provider === "string" && provider !== "" ? provider : void 0;
+	}
+	/**
+	* One window by key, skipping rows that are not objects at all.
+	*
+	* A shape-drifted row must not throw inside the composer: the panel already
+	* reports shape warnings, and a pill that crashes would take the toolbar down
+	* with it.
+	*/
+	function windowByKey(windows, key) {
+		if (!Array.isArray(windows)) return void 0;
+		for (const row of windows) {
+			if (row === null || typeof row !== "object") continue;
+			if (row.key === key) return row;
+		}
+	}
+	/**
+	* The worse of several readings, with "unknown" only when NONE is readable.
+	*
+	* `null` (no reading) is skipped rather than treated as zero: a window the Host
+	* could not measure must not drag the dot down to a healthy-looking grade.
+	*/
+	function worstTier(percentages) {
+		const rank = {
+			error: 3,
+			warn: 2,
+			ok: 1,
+			unknown: 0
+		};
+		let worst = "unknown";
+		for (const pct of percentages) {
+			if (pct === null) continue;
+			const tier = usageTier(pct);
+			if (rank[tier] > rank[worst]) worst = tier;
+		}
+		return worst;
+	}
+	/**
+	* A percentage as the pill writes it — the platform's own figure at one decimal,
+	* the same precision the window card leads with. `—` is reserved for "no reading",
+	* which is a different claim from `0.0%`.
+	*/
+	function percentText(pct) {
+		return pct === null ? "—" : `${pct.toFixed(1)}%`;
+	}
+	/**
+	* The pill's own route reader.
+	*
+	* It exists because the pill and the panel are mounted independently: the card
+	* is a Plugins-page surface that may never be opened in a given session, while
+	* the pill sits in the composer of every conversation whose model is one of this
+	* plugin's. Sharing one poller would mean either polling for a card nobody
+	* opened or rendering a card from a read the pill happened to pay for.
+	*
+	* ONE core, TWO routes. The generation guard, the abort, the stated cadence and
+	* the hidden-tab pause are identical for both upstreams; only the route, the
+	* body reading and the default rate differ. A second copy of this body is
+	* exactly what `dup-gate` exists to refuse, and two hand-kept copies would drift
+	* the first time one of them is repaired.
+	*
+	* What it does NOT carry is the panel's policy: the failure taxonomy (guidance
+	* lines, the sign-in form gate), the refresh spinner, and the `llm` echo merge
+	* all belong to a screen with forms on it. Here a failure is one bit — "this
+	* reading is stale" — and the numbers stay.
+	*
+	* @param route - module-scope stable: a fresh literal per render would rebuild
+	*   `load` and restart the poll interval with it.
+	* @param initialCadenceMs - the rate to follow before any answer arrives;
+	*   `null` means "no timer until something states one".
+	*/
+	function useRemoteRead(route, initialCadenceMs) {
+		const [reading, setReading] = useState({
+			data: null,
+			updatedAt: 0,
+			cadenceMs: initialCadenceMs,
+			loadedOnce: false,
+			failed: false
+		});
+		const cadenceRef = useRef(initialCadenceMs);
+		cadenceRef.current = reading.cadenceMs;
+		const generation = useRef(0);
+		const inFlight = useRef(null);
+		const load = useCallback(async () => {
+			generation.current += 1;
+			const mine = generation.current;
+			const isCurrent = () => generation.current === mine;
+			inFlight.current?.abort?.();
+			const controller = typeof AbortController === "function" ? new AbortController() : null;
+			inFlight.current = controller;
+			const fail = () => {
+				if (isCurrent()) setReading((current) => ({
+					...current,
+					failed: true,
+					loadedOnce: true
+				}));
+			};
+			try {
+				const response = await fetch(route.path, {
+					headers: { accept: "application/json" },
+					cache: "no-store",
+					...controller ? { signal: controller.signal } : {}
+				});
+				if (!isCurrent()) return;
+				if (!response.ok) {
+					fail();
+					return;
+				}
+				const body = await response.json().catch(() => null);
+				if (!isCurrent()) return;
+				const read = route.parse(body, response.status);
+				if (read.data === null) {
+					fail();
+					return;
+				}
+				setReading({
+					data: read.data,
+					updatedAt: Date.now(),
+					cadenceMs: read.cadenceMs !== void 0 ? read.cadenceMs : cadenceRef.current,
+					loadedOnce: true,
+					failed: false
+				});
+			} catch {
+				fail();
+			} finally {
+				if (inFlight.current === controller) inFlight.current = null;
+			}
+		}, [route]);
+		useEffect(() => {
+			load();
+		}, [load]);
+		useEffect(() => {
+			const cadenceMs = reading.cadenceMs;
+			if (cadenceMs === null) return void 0;
+			let timer = null;
+			const start = () => {
+				if (timer === null) timer = setInterval(() => {
+					load();
+				}, cadenceMs);
+			};
+			const stop = () => {
+				if (timer !== null) {
+					clearInterval(timer);
+					timer = null;
+				}
+			};
+			start();
+			const onVisibility = () => {
+				if (typeof document === "undefined") return;
+				if (document.visibilityState === "hidden") stop();
+				else {
+					load();
+					start();
+				}
+			};
+			if (typeof document !== "undefined" && typeof document.addEventListener === "function") document.addEventListener("visibilitychange", onVisibility);
+			return () => {
+				stop();
+				if (typeof document !== "undefined" && typeof document.removeEventListener === "function") document.removeEventListener("visibilitychange", onVisibility);
+			};
+		}, [load, reading.cadenceMs]);
+		useEffect(() => () => {
+			generation.current += 1;
+			inFlight.current?.abort?.();
+		}, []);
+		return reading;
+	}
+	/**
+	* "These are the last numbers we managed to get."
+	*
+	* Two causes: the last read failed, or the reading has aged past two cadences —
+	* expressed as a multiple of the rate being followed, so this file holds no
+	* second spelling of any poll interval. Without a cadence there is no expected
+	* freshness to judge against, so only a failure can mark it.
+	*/
+	function isStale(reading, now) {
+		const aged = reading.cadenceMs !== null && reading.updatedAt > 0 && now - reading.updatedAt > reading.cadenceMs * 2;
+		return reading.loadedOnce && (reading.failed || aged);
+	}
+	/** The console reader (the pill's `/snapshot` loop). */
+	function useConsoleQuota() {
+		return useRemoteRead(SNAPSHOT_ROUTE, null);
+	}
+	/** The desktop-app reader (the pill's `/agnescode` loop). */
+	function useAgnescodeState() {
+		return useRemoteRead(AGNESCODE_ROUTE, AGNESCODE_POLL_MS);
+	}
+	/** The shared shell: the tier dot, the upstream prefix, the reading. */
+	function PillBody(props) {
+		const { tt, planLabel, tier, stale, title, children } = props;
+		return h("button", {
+			type: "button",
+			className: PILL_CLASS,
+			title,
+			"aria-label": title,
+			style: S.pillTrigger
+		}, h("span", {
+			className: `${PILL_CLASS}-dot`,
+			"aria-hidden": "true",
+			style: {
+				...S.pillDot,
+				...TIER_INK[tier]
+			}
+		}), h("span", {
+			className: `${PILL_CLASS}-plan`,
+			style: S.pillPlan
+		}, `${planLabel} · `), h("span", {
+			className: `${PILL_CLASS}-text`,
+			style: tier === "error" ? S.pillErrorText : void 0
+		}, children, stale ? h("span", { style: S.pillStale }, ` · ${tt("composer.stale")}`) : null));
+	}
+	/** The first of several epoch-second instants that is a usable number, else 0. */
+	function firstInstant(...values) {
+		for (const value of values) if (typeof value === "number" && Number.isFinite(value) && value > 0) return value;
+		return 0;
+	}
+	/**
+	* The console reading: the two request windows and the tooltip that carries what
+	* the body has no room for (which plan, which quota, when the window resets, and
+	* whether this reading is the last known one).
+	*/
+	function ConsoleQuotaPill({ tt }) {
+		const reading = useConsoleQuota();
+		const quota = reading.data?.quota ?? null;
+		const windows = Array.isArray(quota?.windows) ? quota.windows : [];
+		const fiveHour = windowByKey(windows, WINDOW_5H);
+		const weekly = windowByKey(windows, WINDOW_WEEKLY);
+		const p5 = windowPercent(fiveHour);
+		const pWeekly = windowPercent(weekly);
+		const tier = worstTier([p5, pWeekly]);
+		const stale = isStale(reading, Date.now());
+		const disconnected = quota !== null && quota.consoleConnected === false;
+		const planName = String(quota?.plan?.displayName || quota?.plan?.name || "");
+		const resetAt = firstInstant(fiveHour?.resetAt, weekly?.resetAt);
+		const title = [
+			tt("composer.pillTitle"),
+			planName === "" ? null : planName,
+			!reading.loadedOnce ? tt("panel.loading") : disconnected ? tt("composer.notConnected") : null,
+			reading.loadedOnce && !disconnected && windows.length === 0 ? tt("composer.noWindows") : null,
+			resetAt > 0 ? format(tt("quota.resetAt"), { time: clockSameDay(resetAt) }) : null,
+			stale ? tt("composer.stale") : null
+		].filter(Boolean).join(" · ");
+		const windowText = (label, pct) => `${label} ${percentText(pct)}`;
+		return h(PillBody, {
+			tt,
+			planLabel: tt("composer.planPrefix"),
+			tier,
+			stale,
+			title,
+			children: [
+				h("span", {
+					key: WINDOW_5H,
+					style: S.pillReading
+				}, windowText(tt("composer.window5h"), p5)),
+				" · ",
+				h("span", {
+					key: WINDOW_WEEKLY,
+					style: S.pillReading
+				}, windowText(tt("composer.windowWeekly"), pWeekly))
+			]
+		});
+	}
+	/**
+	* One credit balance as a number, or `null` when the state declared none.
+	*
+	* The AgnesCode gateway publishes STOCKS: `totalBalance` plus its time-limited
+	* and permanent parts. There is no cap in that universe, so the pill draws the
+	* balance — never a percentage, which would require inventing the limit it is a
+	* share of.
+	*
+	* `0` is a READING (a drained balance is the alarm, and it must read `0` in the
+	* error tier), while a missing or non-finite field is ABSENCE — the grey dash.
+	* The panel's own rule for this tab ("a rendered zero would present an unread
+	* figure as a measurement") is the same one, applied to a field-level gap the
+	* panel never has to face because it renders the block only when it is there.
+	*/
+	function balanceOf(state) {
+		const total = state?.balance?.totalBalance;
+		return typeof total === "number" && Number.isFinite(total) ? total : null;
+	}
+	/** One part of the balance (time-limited / permanent), or `null` when unreadable. */
+	function partOf(state, key) {
+		const value = state?.balance?.[key];
+		return typeof value === "number" && Number.isFinite(value) ? value : null;
+	}
+	/**
+	* The desktop-app reading: the gateway's own credit balance.
+	*
+	* It is a STOCK, not a meter — see {@link balanceOf} — so the body carries the
+	* figure and the tooltip carries the three facts that have no room there: which
+	* account (`nickname`), how the total splits (time-limited / permanent), and —
+	* when there is no figure at all — WHICH fact is missing. "The desktop App is
+	* not signed in", "the provider switch is off" and "the gateway did not answer
+	* this time" are three different problems, and a bare dash beside them left the
+	* reader guessing (the same repair the sibling's capsule made for its gateway).
+	*/
+	function AgnescodeBalancePill({ tt }) {
+		const reading = useAgnescodeState();
+		const state = reading.data;
+		const balance = balanceOf(state);
+		const tier = balance === null ? "unknown" : balance <= 0 ? "error" : "ok";
+		const stale = isStale(reading, Date.now());
+		const pending = !reading.loadedOnce && balance === null;
+		const why = !reading.loadedOnce || balance !== null ? null : state === null ? tt("composer.balanceUnread") : state.enabled !== true ? tt("agnescode.unregistered") : state.loggedIn !== true ? tt("agnescode.notLogged") : tt("composer.balanceUnread");
+		const timeLimited = partOf(state, "timeSensitiveBalance");
+		const permanent = partOf(state, "permanentBalance");
+		const split = timeLimited === null || permanent === null ? null : `${tt("agnescode.balanceTimeLabel")} ${count(timeLimited)} · ${tt("agnescode.balancePermanentLabel")} ${count(permanent)}`;
+		const title = [
+			tt("composer.balanceTitle"),
+			state?.loggedIn === true && String(state.nickname ?? "") !== "" ? format(tt("agnescode.loggedIn"), { nick: String(state.nickname) }) : null,
+			split,
+			why,
+			stale ? tt("composer.stale") : null
+		].filter(Boolean).join(" · ");
+		return h(PillBody, {
+			tt,
+			planLabel: tt("composer.balancePrefix"),
+			tier,
+			stale,
+			title,
+			children: h("span", { style: S.pillReading }, pending ? "…" : balance === null ? "—" : `${count(balance)} ${tt("composer.balanceUnit")}`)
+		});
+	}
+	/**
+	* The composer pill.
+	*
+	* Renders nothing unless the session's model belongs to this plugin — so no other
+	* provider's conversation carries a foreign readout, and no fetch loop runs while
+	* it is absent (the inner reader is not mounted). WHICH upstream it belongs to is
+	* the same decision as whether one mounts: the console provider reads the
+	* snapshot, the desktop-app provider reads its own route, and neither is mounted
+	* for the other's models.
+	*/
+	function ComposerUsagePill(props) {
+		const { tt, useProjection } = props;
+		const provider = selectedProviderOf((useProjection ?? (() => void 0))("modelSelection"));
+		if (provider === "agnes-token-plan") return h(ConsoleQuotaPill, { tt });
+		if (provider === "agnescode") return h(AgnescodeBalancePill, { tt });
+		return null;
+	}
+	var PILL_CLASS, WINDOW_5H, WINDOW_WEEKLY, TIER_INK, SNAPSHOT_ROUTE, AGNESCODE_ROUTE;
+	var init_composer_usage_pill = __esmMin((() => {
+		init_runtime();
+		init_styles();
+		init_format();
+		init_const();
+		init_snapshot();
+		init_cards();
+		init_agnescode_tab();
+		PILL_CLASS = "dsh-agnes-composer-pill";
+		WINDOW_5H = "requests5h";
+		WINDOW_WEEKLY = "requestsWeekly";
+		TIER_INK = {
+			error: S.pillTierError,
+			warn: S.pillTierWarn,
+			ok: S.pillTierOk,
+			unknown: S.pillTierUnknown
+		};
+		SNAPSHOT_ROUTE = {
+			path: SNAPSHOT_PATH,
+			parse: (body, status) => {
+				if (status < 200 || status >= 300) return { data: null };
+				const read = interpretSnapshot(body);
+				if (read.data === null) return { data: null };
+				const stated = read.data.pollSeconds;
+				return {
+					data: read.data,
+					cadenceMs: typeof stated === "number" && Number.isFinite(stated) && stated > 0 ? Math.floor(stated) * 1e3 : null
+				};
+			}
+		};
+		AGNESCODE_ROUTE = {
+			path: AGNESCODE_PATH,
+			parse: (body, status) => {
+				if (status < 200 || status >= 300) return { data: null };
+				const state = body;
+				if (state === null || typeof state !== "object" || state.ok !== true) return { data: null };
+				return {
+					data: state,
+					cadenceMs: AGNESCODE_POLL_MS
+				};
+			}
+		};
+	}));
+
 //#endregion
 //#region src/client/apply.ts
 /**
@@ -3861,6 +4419,17 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		} catch (error) {
 			console.warn(`[${NS}] config card registration failed:`, error);
 		}
+		try {
+			disposers.push(ctx.slots.inject("conversation.input.right", () => ctx.slots.register({
+				name: "conversation.input.right",
+				id: `${NS}-composer-pill`,
+				locale: NS,
+				order: 4,
+				inject: () => ({ tt })
+			}, ComposerUsagePill)));
+		} catch (error) {
+			console.warn(`[${NS}] composer pill registration failed:`, error);
+		}
 		ctx.effect(() => () => {
 			for (const dispose of disposers.splice(0)) try {
 				dispose();
@@ -3872,6 +4441,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		init_const();
 		init_i18n();
 		init_panel_page();
+		init_composer_usage_pill();
 		inject = ["slots", "locale"];
 	}));
 
@@ -3892,6 +4462,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 		init_api_key_form();
 		init_model_picker();
 		init_panel_page();
+		init_composer_usage_pill();
 		init_agnescode_tab();
 		function clientFactory(loaderRequire) {
 			provideClientReact(loaderRequire("react"));
@@ -3935,7 +4506,14 @@ var dsh_connect_agnes_token_plan_client = (function() {
 					allowListFor,
 					toggleModelIn,
 					setAllModelsIn,
-					bulkModelsIn
+					bulkModelsIn,
+					selectedProviderOf,
+					windowByKey,
+					worstTier,
+					percentText,
+					windowPercent,
+					usageTier,
+					balanceOf
 				}),
 				components: Object.freeze({
 					PlanCard,
@@ -3956,6 +4534,7 @@ var dsh_connect_agnes_token_plan_client = (function() {
 					ModelRoster,
 					ModelPicker,
 					PanelPage,
+					ComposerUsagePill,
 					AgnescodeTab,
 					AgnescodeRoster,
 					AgnescodeModelPicker

package/lib/index.js

@@ -4638,8 +4638,8 @@ async function buildSnapshotBody({ settings, cache, inflight, tokenStore, apiKey
 	const providerState = publisher.state;
 	const resolveApiKey = async () => (await apiKeyStore.resolve()).value;
 	const { startDate, endDate } = usageWindow(settings.usageDays);
-	const overview = await soft(() => fetchConsole(settings, USAGE_OVERVIEW_PATH, void 0, settings.cacheSeconds * 1e3, cache, inflight, tokenStore));
-	const [series, subscription, plans, catalog] = await Promise.all([
+	const [overview, series, subscription, plans, catalog] = await Promise.all([
+		soft(() => fetchConsole(settings, USAGE_OVERVIEW_PATH, void 0, settings.cacheSeconds * 1e3, cache, inflight, tokenStore)),
 		soft(() => fetchConsole(settings, USAGE_SERIES_PATH, {
 			range: "custom",
 			start_date: startDate,
@@ -5801,16 +5801,15 @@ function registerAgnescodeRoute(ctx, wiring) {
 						bffBase = state?.bffBase ?? "";
 						expiresAtMs = state?.expiresAtMs ?? null;
 						credential = (await agnescodeStore.resolve().catch(() => ({ credential: null }))).credential ?? null;
-						if (loggedIn && credential?.accessToken) balance = await fetchAgnescodeBalance(credential).catch(() => null);
 					}
 				} catch (why) {
 					error = redactSecrets(why instanceof Error ? why.message : String(why));
 				}
 				let models = null;
-				try {
-					if (credential?.accessToken) models = await fetchAgnescodeCatalog(credential).catch(() => null);
-				} catch {
-					models = null;
+				if (credential?.accessToken) {
+					const [b, m] = await Promise.all([loggedIn ? fetchAgnescodeBalance(credential).catch(() => null) : Promise.resolve(null), fetchAgnescodeCatalog(credential).catch(() => null)]);
+					balance = b;
+					models = m;
 				}
 				const roster = models !== null && Array.isArray(models) && models.length > 0 ? models : AGNESCODE_FALLBACK_MODELS;
 				const publisherState = agnescodePublisher?.state ?? null;

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-connect-agnes-token-plan",
   "displayName": "DSH Connect Agnes Token Plan",
   "icon": "./icon.svg",
-  "version": "0.10.2",
+  "version": "0.11.0",
   "description": "Agnes connect bundle for the Harness Web UI: Token Plan quota panel (account-wide quota windows, cumulative usage, console sign-in) plus optional provider registration that brings Agnes models into DSH conversation, with host-side rate-limit self-healing.",
   "keywords": [
     "dsh-plugin",
@@ -95,6 +95,9 @@
     "@commitlint/config-conventional": "^21.2.3",
     "@types/node": "^26.6.3",
     "jscpd": "^5.4.0",
+    "jsdom": "^30.1.3",
+    "react": "^18.3.1",
+    "react-dom": "^18.3.1",
     "tsdown": "^0.23.0",
     "typescript": "^5.9.3"
   }