pi-llamacpp-infra 1.2.1 → 1.2.2

package/README.md

@@ -247,20 +247,20 @@ llama.cpp-family models are registered as reasoning models, exactly like a nativ
 
 ## Live Speed & Metrics (footer)
 
-When enabled, the speed reading appears in the footer's status line (no extra terminal row) whenever the active model is from llamacpp-infra, updating constantly while tokens flow:
+When enabled, the speed reading appears in the footer's status line (no extra terminal row) whenever the active model is from llamacpp-infra, updating constantly while tokens flow. Both entries are kept ultra-compact so they coexist with other extensions on pi's single status line (which truncates from the end):
 
 ```
-🦙 12 models · 3/3 ✓ 📊 · ⚡ prefill…                    (before the first token)
-🦙 12 models · 3/3 ✓ 📊 · 🔥 38.1 t/s · ⚡ 420 · 1.2k tok  (while streaming)
-🦙 12 models · 3/3 ✓ 📊 · 🔥 38.1 t/s · ⚡ 420 · 1.2k tok  (just after the answer ends)
-🦙 12 models · 3/3 ✓ 📊 · ⏸ idle                         (between turns)
-🦙 12 models · 3/3 ✓ 📊 · ▶ 2 · ⚡ 150 · 🔥 18.0 · server (pi idle, server busy for other clients)
+🦙(12) ⚡…            (before the first token)
+🦙(12) ⚡ 420 t/s 🔥 38.1 t/s  (while streaming)
+🦙(12) ⚡ 420 t/s 🔥 38.1 t/s  (just after the answer ends)
+🦙(12) ⏸             (between turns)
+🦙(12) ▶2 ⚡ 150 t/s 🔥 18.0 t/s  (pi idle, server busy for other clients)
 ```
 
-(The `🦙 …` prefix is the extension's model-count status; both live on the same footer line, so no extra row is consumed.)
+(`🦙(n)` is the extension's model-count status; both live on the same footer line, so no extra row is consumed.)
 
 - **Client measurement (always, no `--metrics` needed)** — prefill speed = `prompt tokens ÷ (request → first token)` (pi's `usage.input`, OpenAI-style `prompt_tokens` as fallback); generation speed = a moving 1.5 s window over per-token arrival samples. Updated ~every 100 ms while a stream is live (throttled, and unchanged text is skipped, so the footer never churns).
-- **Server supplement (only when pi is idle)** — the poller fetches the server's Prometheus `/metrics` endpoint (or JSON `/stats`) every `metricsPollMs` (default 5 s). If the server reports other clients processing, their ⚡/🔥 rates are shown; when the server is idle, the plain `⏸ idle` reading returns.
+- **Server supplement (only when pi is idle)** — the poller fetches the server's Prometheus `/metrics` endpoint (or JSON `/stats`) every `metricsPollMs` (default 5 s). If the server reports other clients processing, their ⚡/🔥 rates are shown (`▶n`); when the server is idle, the plain `⏸` reading returns.
 
 ## Architecture
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-llamacpp-infra",
-  "version": "1.2.1",
+  "version": "1.2.2",
   "description": "Discovery, metrics and control of llama.cpp-family servers for pi (llama.cpp, ZINC, DwarfStar/ds4, lucebox, LM Studio): scan machines (localhost, LAN, Tailscale), register models, live metrics, per-model thinking budgets, vision detection, native config UI.",
   "keywords": [
     "pi-package",

package/src/index.ts

@@ -225,7 +225,7 @@ export default function (pi: ExtensionAPI) {
 
 	async function rescan(ctx?: ExtensionContext) {
 		if (!extensionActive) return;
-		if (ctxHasUI(ctx)) ctx.ui.setStatus(STATUS_KEY, "🔎 scanning…");
+		if (ctxHasUI(ctx)) ctx.ui.setStatus(STATUS_KEY, "🔎");
 		stopPolling();
 		registerEmptyProvider();
 		const r = await discoverAndRegister();
@@ -237,11 +237,9 @@ export default function (pi: ExtensionAPI) {
 	function updateStatusFooter(ctx?: ExtensionContext) {
 		if (!ctxHasUI(ctx)) return;
 		if (shared.registeredCount > 0) {
-			const up = shared.lastScan?.serversUp ?? 0;
-			const total = shared.lastScan?.serversTotal ?? 0;
-			ctx.ui.setStatus(STATUS_KEY, `🦙 ${shared.registeredCount} models · ${up}/${total} ✓`);
+			ctx.ui.setStatus(STATUS_KEY, `🦙(${shared.registeredCount})`);
 		} else if (shared.lastScan?.endpoints.some((e) => e.loading)) {
-			ctx.ui.setStatus(STATUS_KEY, "⏳ loading…");
+			ctx.ui.setStatus(STATUS_KEY, "⏳");
 		} else if (shared.lastError) {
 			ctx.ui.setStatus(STATUS_KEY, "⚠️");
 		} else {

package/src/speed.ts

@@ -6,10 +6,15 @@
 //   prefill t/s  = prompt tokens / (before_provider_request → first token)
 //   gen t/s      = moving window over per-token arrival samples
 //
+// Footer format (kept ultra-compact so it coexists with other extensions on
+// pi's single status line, which is truncated at the end):
+//
+//   ⚡ {prefill} t/s 🔥 {gen} t/s
+//
 // The server-side /metrics polling (metrics.ts) only supplements this:
 // when the client is idle but the server reports other clients processing,
-// their rate is shown. The client measurement works even when the server has
-// no --metrics endpoint.
+// their rate is shown as `▶{n} ⚡ … t/s 🔥 … t/s`. The client measurement
+// works even when the server has no --metrics endpoint.
 
 import { debugLog, METRICS_STATUS_KEY } from "./core.ts";
 import type { AssistantMessageEvent, ExtensionContext, ServerMetricsState, ThemeFg } from "./types.ts";
@@ -126,58 +131,58 @@ export function createSpeedTracker(deps: SpeedDeps) {
 		return tokens / (span / 1000);
 	}
 
-	function buildLine(ctx: ExtensionContext | undefined, now: number): string[] {
+	function buildLine(ctx: ExtensionContext | undefined, now: number): string {
 		const f = fg(ctx);
-		const parts: string[] = [f("accent", "📊")];
+		const parts: string[] = [];
 
 		switch (state) {
 			case "prefill":
-				parts.push(f("warning", "⚡ prefill…"));
+				parts.push(f("warning", "⚡…"));
 				break;
 
 			case "streaming": {
+				// Prefill rate of the previous call (only known once it ended).
+				if (lastPrefillTps !== undefined) {
+					parts.push(f("muted", `⚡ ${Math.round(lastPrefillTps)} t/s`));
+				}
 				const gen = genRateAt(now);
 				if (gen !== undefined) {
 					lastGenTps = gen;
 					parts.push(f(rateColor(gen, "gen"), `🔥 ${formatRate(gen)} t/s`));
 				}
-				if (lastPrefillTps !== undefined) {
-					parts.push(f("muted", `⚡ ${Math.round(lastPrefillTps)}`));
-				}
-				parts.push(f("muted", `${tokenCount} tok`));
 				break;
 			}
 
 			case "done": {
-				if (lastGenTps !== undefined) {
-					parts.push(f("muted", `🔥 ${formatRate(lastGenTps)} t/s`));
-				}
 				if (lastPrefillTps !== undefined) {
-					parts.push(f("muted", `⚡ ${Math.round(lastPrefillTps)}`));
+					parts.push(f(rateColor(lastPrefillTps, "prefill"), `⚡ ${Math.round(lastPrefillTps)} t/s`));
+				}
+				if (lastGenTps !== undefined) {
+					parts.push(f(rateColor(lastGenTps, "gen"), `🔥 ${formatRate(lastGenTps)} t/s`));
 				}
-				parts.push(f("muted", `${tokenCount} tok`));
 				break;
 			}
 
 			case "idle": {
 				// Server supplement: this endpoint is busy for *other* clients.
 				if (server && server.processing > 0) {
-					parts.push(f("success", `▶ ${server.processing}`));
+					parts.push(f("success", `▶${server.processing}`));
 					if (server.promptTps !== undefined && server.promptTps > 0) {
-						parts.push(f(rateColor(server.promptTps, "prefill"), `⚡ ${Math.round(server.promptTps)}`));
+						parts.push(f(rateColor(server.promptTps, "prefill"), `⚡ ${Math.round(server.promptTps)} t/s`));
 					}
 					if (server.genTps !== undefined && server.genTps > 0) {
-						parts.push(f(rateColor(server.genTps, "gen"), `🔥 ${formatRate(server.genTps)}`));
+						parts.push(f(rateColor(server.genTps, "gen"), `🔥 ${formatRate(server.genTps)} t/s`));
 					}
-					parts.push(f("muted", "server"));
 				} else {
-					parts.push(f("muted", "⏸ idle"));
+					parts.push(f("muted", "⏸"));
 				}
 				break;
 			}
 		}
 
-		return [parts.join(" · ")];
+		// A stream is live but no rate is computable yet (first ~300 ms of a call).
+		if (parts.length === 0) parts.push(f("muted", "🔥…"));
+		return parts.join(" ");
 	}
 
 	function render(ctx: ExtensionContext | undefined, now: number, force = false): void {
@@ -186,7 +191,7 @@ export function createSpeedTracker(deps: SpeedDeps) {
 			return;
 		}
 		if (!deps.hasUI(ctx) || !deps.isOurs(ctx)) return;
-		const line = buildLine(ctx, now).join(" · ");
+		const line = buildLine(ctx, now);
 		if (!force && now - lastRenderAt < RENDER_THROTTLE_MS) return;
 		lastRenderAt = now;
 		// No visual change → skip (avoids a full UI re-render on the footer).