@bacnh85/pi-sub 0.1.33 → 0.1.36

package/CHANGELOG.md

@@ -1,3 +1,39 @@
+# Changelog
+
+## 0.1.36 (2026-09-07)
+
+### Fixed
+
+- **Removed the incorrect dead `tok-per-sec` module** (added in 0.1.35, never
+  wired in): its `withThinking` formula summed `reasoning` on top of `output`,
+  but Pi's `usage.output` already includes reasoning tokens (`reasoning` is a
+  subset), which would have inflated thinking-mode tok/s up to ~2×. The live
+  footer/footer tok/s math (`output / elapsed`) was and remains correct in both
+  thinking and normal mode.
+
+### Added
+
+- **`/sub` thinking/answer split** — when the model reasoned, the details line
+  now reads e.g. `Last response: 46 tok/s (36 think + 10 answer)` instead of a
+  bare total, using the correct subset math (`answer = output − reasoning`).
+  Footer keeps the single total number.
+
+## 0.1.35 (2026-09-05)
+
+### Added
+
+- **`zai-anthropic` provider usage tracking** — GLM through the Anthropic-compatible endpoint (`api.z.ai/api/anthropic`, registered by pi-model-tools) now shows the same quota footer as `zai`/`zai-coding-cn`: 5-hour/weekly windows, MCP monthly allowance, and model/tool breakdowns, keyed by the auth.json `zai-anthropic` credential, labeled `Z.ai (Anthropic)`.
+
+## 0.1.34 (2026-09-05)
+
+- Widen Pi SDK peer range to `>=0.80.8 <0.86.0` for Pi 0.85.0 compatibility (no breaking changes; peer cap widening only).
+
+## 0.1.33 (2026-08-29)
+
+### Added
+
+- `/sub` argument completion offers `refresh`.
+
 ## 0.1.32 (2026-08-22)
 
 - **DeepSeek via OmniRoute now shows the real USD balance** (e.g. `M:$18.25`)

package/README.md

@@ -51,6 +51,14 @@ The built-in `zai-coding-cn` provider targets the domestic BigModel endpoint (`o
 (Z.ai (CN) key#1a2b3c4d) R:55%/2H W:80%/3D 42 tok/s
 ```
 
+### Z.ai via Anthropic endpoint
+
+The `zai-anthropic` provider (registered by pi-model-tools, GLM through `api.z.ai/api/anthropic`) is tracked the same way — same api.z.ai quota monitor as the international `zai` provider, keyed by the auth.json `zai-anthropic` credential, labeled `Z.ai (Anthropic)`:
+
+```text
+(Z.ai (Anthropic) key#1a2b3c4d) R:55%/2H W:80%/3D 42 tok/s
+```
+
 ### Router (pi-router — formerly 9router)
 
 For **OmniRoute** instances the footer shows real usage: `GET <origin>/api/usage/om-usage`

package/extensions/index.ts

@@ -38,6 +38,13 @@ const ZAI_PROVIDER = "zai";
 const ZAI_USAGE_URL = "https://api.z.ai/api/monitor/usage/quota/limit";
 const ZAI_CODING_CN_PROVIDER = "zai-coding-cn";
 const ZAI_CODING_CN_USAGE_URL = "https://open.bigmodel.cn/api/monitor/usage/quota/limit";
+// GLM via the Anthropic-compatible endpoint (pi-model-tools `zai-anthropic`
+// provider). Same api.z.ai host and quota monitor as the `zai` provider.
+// ponytail: usage URL is fixed to api.z.ai — if ZAI_ANTHROPIC_BASE_URL is
+// overridden to BigModel/zcode-plan, quota still reads from api.z.ai (correct
+// for the z.ai coding-plan key; BigModel-plan keys should use zai-coding-cn).
+const ZAI_ANTHROPIC_PROVIDER = "zai-anthropic";
+const ZAI_ANTHROPIC_USAGE_URL = ZAI_USAGE_URL;
 const ROUTER_PROVIDER = "router";
 const LEGACY_9ROUTER_PROVIDER = "9router";
 // pi-router (formerly pi-9router): URL lives in settings.json `router.baseUrl`
@@ -120,6 +127,7 @@ interface State {
   debounceTimer?: NodeJS.Timeout;
   responseStartTime?: number;
   lastTokPerSec?: number;
+  lastTokPerSecLabel?: string;
   cumulativeOutput: number;
   cumulativeDurationMs: number;
   cumulativeCost: number;
@@ -142,6 +150,10 @@ function isZaiCodingCnModel(model: ModelLike): boolean {
   return (model?.provider?.toLowerCase() ?? "") === ZAI_CODING_CN_PROVIDER;
 }
 
+function isZaiAnthropicModel(model: ModelLike): boolean {
+  return (model?.provider?.toLowerCase() ?? "") === ZAI_ANTHROPIC_PROVIDER;
+}
+
 function isRouterModel(model: ModelLike, provider: string = ROUTER_PROVIDER): boolean {
   return (model?.provider?.toLowerCase() ?? "") === provider;
 }
@@ -451,6 +463,9 @@ if (process.env.PI_SUB_SELF_CHECK === "1") {
   assert(p.personalWeekly?.remaining === 90, "personal weekly 90");
   assert(p.session?.remaining === 47, "session 47");
   assert(p.providerWeekly?.remaining === 28, "provider weekly 28");
+  // tok/s split label: usage.reasoning ⊂ usage.output, never summed.
+  assert(tokPerSecLabel(3200, 2500, 70_000) === "46 tok/s (36 think + 10 answer)", "tok/s split label");
+  assert(tokPerSecLabel(200, 0, 10_000) === "20 tok/s", "tok/s plain label");
   assert(p.personalDaily?.resetLabel?.includes("15h") === true, "daily reset label");
   const disabled = parseOmniUsageText("Usage command is disabled for this API key.");
   assert(Object.keys(disabled).length === 0, "disabled text parses empty");
@@ -991,11 +1006,14 @@ function zaiUsageAdapter(providerId: string, usageUrl: string, displayName: stri
   return { fetchUsage };
 }
 
-function supportedAdapter(model: ModelLike): SubscriptionProviderAdapter | undefined {
+// Exported for the adapter-wiring regression test (provider-id string ↔
+// adapter id ↔ usage URL are exactly what a typo silently breaks).
+export function supportedAdapter(model: ModelLike): SubscriptionProviderAdapter | undefined {
   if (isCodexModel(model)) return { id: CODEX_PROVIDER, displayName: "Codex", fetchUsage: fetchCodexUsage };
   if (isOpenCodeGoModel(model)) return { id: OPC_PROVIDER, displayName: "OpenCode Go", fetchUsage: fetchOpenCodeGoUsage };
   if (isZaiModel(model)) return { id: ZAI_PROVIDER, displayName: "Z.ai", ...zaiUsageAdapter(ZAI_PROVIDER, ZAI_USAGE_URL, "Z.ai") };
   if (isZaiCodingCnModel(model)) return { id: ZAI_CODING_CN_PROVIDER, displayName: "Z.ai (CN)", ...zaiUsageAdapter(ZAI_CODING_CN_PROVIDER, ZAI_CODING_CN_USAGE_URL, "Z.ai (CN)") };
+  if (isZaiAnthropicModel(model)) return { id: ZAI_ANTHROPIC_PROVIDER, displayName: "Z.ai (Anthropic)", ...zaiUsageAdapter(ZAI_ANTHROPIC_PROVIDER, ZAI_ANTHROPIC_USAGE_URL, "Z.ai (Anthropic)") };
   if (isRouterModel(model)) {
     const prefix = routerUpstreamPrefix(model);
     return {
@@ -1179,11 +1197,21 @@ function pad(value: string, width: number): string {
   return value.length >= width ? value : value + " ".repeat(width - value.length);
 }
 
+/** "46 tok/s (36 think + 10 answer)" — split shown only when the model
+ *  reasoned. usage.reasoning is a subset of usage.output (Pi SDK contract),
+ *  so answer speed = (output − reasoning)/s, never output + reasoning. */
+function tokPerSecLabel(output: number, thinking: number, elapsedMs: number): string {
+  const total = Math.round(output / (elapsedMs / 1000));
+  if (thinking <= 0) return `${total} tok/s`;
+  const secs = elapsedMs / 1000;
+  return `${total} tok/s (${Math.round(thinking / secs)} think + ${Math.round((output - thinking) / secs)} answer)`;
+}
+
 function buildDetails(snapshot: SubscriptionUsageSnapshot | undefined, state: State): string {
   if (!state.adapter) {
     const header = `Provider: ${state.model?.provider ?? "unknown"}${state.model?.id ? ` · Model: ${state.model.id}` : ""}`;
     if (state.lastTokPerSec === undefined) return `${header}\nSubscription tracking inactive for this provider.`;
-    const tokPerSecLine = `Last response: ${state.lastTokPerSec} tok/s` +
+    const tokPerSecLine = `Last response: ${state.lastTokPerSecLabel}` +
       (state.cumulativeDurationMs > 0
         ? ` · Session avg: ${Math.round(state.cumulativeOutput / (state.cumulativeDurationMs / 1000))} tok/s`
         : "");
@@ -1226,7 +1254,7 @@ function buildDetails(snapshot: SubscriptionUsageSnapshot | undefined, state: St
 
   const costLine = state.cumulativeCost > 0 ? `\nSession cost: $${state.cumulativeCost.toFixed(2)}` : "";
   const tokPerSecLine = state.lastTokPerSec !== undefined
-    ? `\nLast response: ${state.lastTokPerSec} tok/s` +
+    ? `\nLast response: ${state.lastTokPerSecLabel}` +
       (state.cumulativeDurationMs > 0
         ? ` · Session avg: ${Math.round(state.cumulativeOutput / (state.cumulativeDurationMs / 1000))} tok/s`
         : "")
@@ -1263,11 +1291,15 @@ export default function (pi: ExtensionAPI) {
     if (event.message.role === "assistant") {
       state.cumulativeCost += (event.message.usage as any)?.cost?.total ?? 0;
       if (state.responseStartTime) {
+        // usage.output already includes reasoning tokens (Pi SDK contract) —
+        // this is total tok/s in both thinking and normal mode.
         const output = (event.message.usage as any)?.output ?? 0;
+        const reasoning = (event.message.usage as any)?.reasoning ?? 0;
         const elapsed = Date.now() - state.responseStartTime;
         state.responseStartTime = undefined;
         if (elapsed > 0 && output > 0) {
           state.lastTokPerSec = Math.round(output / (elapsed / 1000));
+          state.lastTokPerSecLabel = tokPerSecLabel(output, reasoning, elapsed);
           state.cumulativeOutput += output;
           state.cumulativeDurationMs += elapsed;
         }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bacnh85/pi-sub",
-  "version": "0.1.33",
+  "version": "0.1.36",
   "description": "Pi extension showing subscription usage for supported model providers.",
   "type": "module",
   "license": "MIT",
@@ -39,6 +39,8 @@
     "node": ">=20.3.0"
   },
   "peerDependencies": {
-    "@earendil-works/pi-coding-agent": ">=0.80.8 <0.85.0"
+    "@earendil-works/pi-coding-agent": ">=0.80.8 <0.86.0"
+  },
+  "devDependencies": {
   }
 }