pi-subagents-lite 1.4.5 → 1.4.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-subagents-lite",
-  "version": "1.4.5",
+  "version": "1.4.6",
   "description": "Lightweight sub-agents for pi — spawn specialized agents with isolated sessions, tools, and models.",
   "keywords": [
     "pi-package",

package/src/agents/agent-manager.ts

@@ -8,6 +8,7 @@ import { randomUUID } from "node:crypto";
 import type { AgentSession, ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { runAgent } from "./agent-runner.js";
 import { AgentOutputLog } from "./output-file.js";
+import { getStore } from "../shell.js";
 import {
   type AgentRecord,
   type AgentStatus,
@@ -18,7 +19,7 @@ import {
   type ToolActivity,
 } from "../types.js";
 import type { SubagentType } from "./types.js";
-import { addUsage, getLifetimeTotal, getSessionContextPercent, type LifetimeUsage } from "./usage.js";
+import { addUsage, getLifetimeTotal, getSessionContextPercent, type AgentUsage } from "./usage.js";
 import { errorMessage } from "../utils.js";
 
 /** How often to check for expired agent records (milliseconds). */
@@ -373,7 +374,7 @@ export class AgentManager {
     options?: Pick<SpawnOptions, "onToolActivity" | "onAssistantUsage" | "onCompaction">,
   ): {
     onToolActivity: (activity: ToolActivity) => void;
-    onAssistantUsage: (usage: LifetimeUsage) => void;
+    onAssistantUsage: (usage: AgentUsage) => void;
     onCompaction: (info: CompactionInfo) => void;
   } {
     return {
@@ -382,7 +383,17 @@ export class AgentManager {
         options?.onToolActivity?.(activity);
       },
       onAssistantUsage: (usage) => {
-        addUsage(record.stats.lifetimeUsage, usage);
+        // vLLM doesn't report cache hits, so usage.input is full prompt_tokens.
+        // Estimate new tokens as delta from previous message's input.
+        const deltaEnabled = getStore().agent.deltaInputTokens;
+        const cacheRead = usage.cacheRead;
+        let inputDelta = usage.input;
+        if (deltaEnabled && cacheRead === 0 && record.stats.prevInputTokens != null && usage.input > record.stats.prevInputTokens) {
+          inputDelta = usage.input - record.stats.prevInputTokens;
+        }
+        record.stats.prevInputTokens = usage.input;
+
+        addUsage(record.stats.lifetimeUsage, { ...usage, input: inputDelta });
         options?.onAssistantUsage?.(usage);
       },
       onCompaction: (info) => {

package/src/agents/agent-runner.ts

@@ -20,7 +20,7 @@ import {
 } from "@earendil-works/pi-coding-agent";
 import { getAgentConfig, getConfig, getToolNamesForType, resolveVisibleTools } from "./agent-types.js";
 import { extractText } from "../prompt/context.js";
-import type { LifetimeUsage } from "./usage.js";
+import type { AgentUsage } from "./usage.js";
 import { findModelInRegistry, GIT_EXEC_TIMEOUT_MS } from "../utils.js";
 import { DEFAULT_AGENTS } from "./default-agents.js";
 import { buildAgentPrompt, type PromptExtras } from "../prompt/prompts.js";
@@ -106,13 +106,14 @@ function forwardAbortSignal(session: AgentSession, signal?: AbortSignal): () =>
  * assistant messages at runtime, but this shape isn't reflected in the
  * AgentSessionEvent public types.
  */
-function usageFromAssistantMessage(msg: Record<string, unknown>): LifetimeUsage | undefined {
+function usageFromAssistantMessage(msg: Record<string, unknown>): AgentUsage | undefined {
   const usage = msg.usage as Record<string, unknown> | undefined;
   if (!usage) return undefined;
   return {
     input: (usage.input as number) ?? 0,
     output: (usage.output as number) ?? 0,
     cacheWrite: (usage.cacheWrite as number) ?? 0,
+    cacheRead: (usage.cacheRead as number) ?? 0,
     cost: ((usage.cost as Record<string, unknown>)?.total as number) ?? 0,
   };
 }

package/src/agents/usage.ts

@@ -9,6 +9,13 @@
  */
 export type LifetimeUsage = { input: number; output: number; cacheWrite: number; cost: number };
 
+/**
+ * A single per-turn usage event as emitted upstream. Adds `cacheRead`, which
+ * LifetimeUsage omits from totals (see issue #38). Used to estimate input
+ * deltas for providers like vLLM that don't report cache hits.
+ */
+export type AgentUsage = LifetimeUsage & { cacheRead: number };
+
 /** Sum of lifetime usage components (including cost), or 0 if undefined. */
 export function getLifetimeTotal(u?: LifetimeUsage): number {
   return u ? u.input + u.output + u.cacheWrite + u.cost : 0;

package/src/config/config-io.ts

@@ -42,6 +42,7 @@ const DEFAULT_AGENT: SubagentsConfig["agent"] = {
   showContext: true,
   showCost: false,
   showTime: true,
+  deltaInputTokens: false,
 };
 
 /**

package/src/config/config-store.ts

@@ -74,6 +74,8 @@ export interface ResolvedAgentSettings {
   readonly showContext: boolean;
   /** Whether to show elapsed time in widget stats line. */
   readonly showTime: boolean;
+  /** Whether to estimate input token delta for vLLM (no cache reporting). */
+  readonly deltaInputTokens: boolean;
   /** Buffer size for streaming thinking blocks to output file. 0 = disabled. */
   readonly outputThinkingBufferSize: number;
 }
@@ -133,6 +135,7 @@ export class ConfigStore {
       showOutput: a.showOutput !== false,
       showContext: a.showContext !== false,
       showTime: a.showTime !== false,
+      deltaInputTokens: a.deltaInputTokens !== false,
       outputThinkingBufferSize: a.outputThinkingBufferSize ?? 0,
     };
   }
@@ -265,6 +268,10 @@ export class ConfigStore {
       setShowOutput: (enabled: boolean) => this.setAgentVisibility("showOutput", enabled),
       setShowContext: (enabled: boolean) => this.setAgentVisibility("showContext", enabled),
       setShowTime: (enabled: boolean) => this.setAgentVisibility("showTime", enabled),
+      setDeltaInputTokens: (enabled: boolean): void => {
+        this.config.agent.deltaInputTokens = enabled;
+        this.persist();
+      },
       setOutputThinkingBufferSize: (size: number): void => {
         this.config.agent.outputThinkingBufferSize = size;
         this.persist();

package/src/config/types.ts

@@ -10,6 +10,7 @@ export const CONFIG_AGENT_NON_MODEL_KEYS = [
   "showOutput",
   "showContext",
   "showTime",
+  "deltaInputTokens",
   "widgetMaxLines",
   "widgetMaxLinesCompact",
   "widgetDescLengthFull",

package/src/types.ts

@@ -5,7 +5,7 @@
 import type { Model } from "@earendil-works/pi-ai";
 import type { AgentSession } from "@earendil-works/pi-coding-agent";
 import type { AgentOutputLog } from "./agents/output-file.js";
-import type { LifetimeUsage } from "./agents/usage.js";
+import type { LifetimeUsage, AgentUsage } from "./agents/usage.js";
 import type { SubagentType, AgentInvocation } from "./agents/types.js";
 
 /** Thinking level for agent models. */
@@ -59,7 +59,7 @@ export interface RunCallbacks {
   onTextDelta?: (delta: string, fullText: string) => void;
   onSessionCreated?: (session: AgentSession) => void;
   onTurnEnd?: (turnCount: number) => void;
-  onAssistantUsage?: (usage: LifetimeUsage) => void;
+  onAssistantUsage?: (usage: AgentUsage) => void;
   onCompaction?: (info: CompactionInfo) => void;
 }
 
@@ -156,6 +156,8 @@ export interface AgentAccumulatedStats {
   maxTurns?: number;
   /** Number of times this agent's session has compacted. Initialized to 0 at spawn. */
   compactionCount: number;
+  /** Previous input token count for delta estimation (vLLM doesn't report cache hits). */
+  prevInputTokens?: number;
   /** Last-known context usage percentage (0–100), captured at completion. */
   contextPercent?: number | null;
 }

package/src/ui/menu/menu-widget-settings.ts

@@ -26,6 +26,7 @@ function buildStatConfig(store: ReturnType<typeof getStore>) {
     ["showTools", { label: "Show tools", get: () => store.agent.showTools, set: (v) => store.mutate.agent.setShowTools(v) }],
     ["showTurns", { label: "Show turns", get: () => store.agent.showTurns, set: (v) => store.mutate.agent.setShowTurns(v) }],
     ["showInput", { label: "Show input tokens", get: () => store.agent.showInput, set: (v) => store.mutate.agent.setShowInput(v) }],
+    ["deltaInputTokens", { label: "Delta input tokens", get: () => store.agent.deltaInputTokens, set: (v) => store.mutate.agent.setDeltaInputTokens(v) }],
     ["showOutput", { label: "Show output tokens", get: () => store.agent.showOutput, set: (v) => store.mutate.agent.setShowOutput(v) }],
     ["showContext", { label: "Show context %", get: () => store.agent.showContext, set: (v) => store.mutate.agent.setShowContext(v) }],
     ["showCost", { label: "Show cost", get: () => store.agent.showCost, set: (v) => store.mutate.agent.setShowCost(v) }],
@@ -66,6 +67,7 @@ export async function showWidgetSettingsMenu(ctx: ExtensionCommandContext): Prom
       showTools: "Show tool count (🛠 ) in the widget.",
       showTurns: "Show turn count (⟳ ) in the widget.",
       showInput: "Show input tokens (↑) in the widget.",
+      deltaInputTokens: "Estimate input token delta for vLLM (no cache reporting).",
       showOutput: "Show output tokens (↓) in the widget.",
       showContext: "Show context-fill percent (%) in the widget.",
       showCost: "Show dollar cost ($) in the widget.",