Merge branch 'main' of https://github.com/andrewyng/openworker into feature/memory

surfaces/gui/src/App.tsx

@@ -31,10 +31,19 @@ import {
   type SurfaceVisibility,
   type WorkspaceCommandTrust,
 } from "./api";
-import type { ApprovalDecision, Attachment, Item, SessionInfo, TodoItem, WsEvent } from "./types";
+import type {
+  ApprovalDecision,
+  Attachment,
+  Item,
+  SessionInfo,
+  SessionUsage,
+  TodoItem,
+  WsEvent,
+} from "./types";
 import { isProjectScoped } from "./personaScope";
 import { baseName } from "./paths";
 import { itemsFromMessages } from "./itemsFromMessages";
+import { addTurnUsage, emptyUsage, usageFromMessages } from "./usage";
 import { streamMode } from "./streamGate";
 import { InboxItemCard } from "./components/InboxItemCard";
 import { isTauri, platformOS, startWindowDrag } from "./tauri";
@@ -156,6 +165,12 @@ export function App() {
   const [model, setModel] = useState("gpt-5.6-sol");
   const [models, setModels] = useState<string[]>([]);
   const [modelLabels, setModelLabels] = useState<Record<string, string>>({});
+  // {full model id → context window in tokens} from the curated matrix (verified only);
+  // drives the composer usage chip's context-fill meter.
+  const [modelContextWindows, setModelContextWindows] = useState<Record<string, number>>({});
+  // Per-session token usage (OPE-42): rebuilt from the transcript on session load,
+  // accumulated live from assistant_message events, reset with the transcript.
+  const [usage, setUsage] = useState<SessionUsage>(emptyUsage());
   const [surfaces, setSurfaces] = useState<SurfaceVisibility>({ cowork: true, chat: false, code: false });
   const [mode, setMode] = useState("interactive");
   const [connected, setConnected] = useState(false);
@@ -395,9 +410,12 @@ export function App() {
           setBranch(null);
         }
         try {
-          setItems(itemsFromMessages(await getSessionMessages(last.session_id)));
+          const messages = await getSessionMessages(last.session_id);
+          setItems(itemsFromMessages(messages));
+          setUsage(usageFromMessages(messages));
         } catch {
           setItems([]);
+          setUsage(emptyUsage());
         }
         setSessionId(last.session_id);
         setShowGate(false);
@@ -486,6 +504,7 @@ export function App() {
       .then((s) => {
         setModels(s.models || []);
         setModelLabels(s.model_labels || {});
+        setModelContextWindows(s.model_context_windows || {});
         setModelReady(s.model_ready);
         if (s.surfaces) setSurfaces(s.surfaces);
       })
@@ -601,6 +620,7 @@ export function App() {
           setReasoningStream(reasoningRef.current + (d.text || ""));
           break;
         case "assistant_message": {
+          if (d.usage) setUsage((u) => addTurnUsage(u, d.usage));
           // The event's reasoning is authoritative (covers background-delivered turns);
           // the local buffer is the fallback for older servers.
           const reasoning = d.reasoning || reasoningRef.current;
@@ -899,6 +919,7 @@ export function App() {
     const target = forAgent || agent;
     setSurface("session"); // return to the conversation view if we were on a sub-view
     setItems([]);
+    setUsage(emptyUsage());
     setStreaming("");
     setTodo([]);
     setRunning(false);
@@ -975,8 +996,10 @@ export function App() {
     try {
       const messages = await getSessionMessages(id);
       setItems(itemsFromMessages(messages));
+      setUsage(usageFromMessages(messages));
     } catch {
       setItems([]);
+      setUsage(emptyUsage());
     }
   };
   const switchAgent = async (name: string) => {
@@ -989,6 +1012,7 @@ export function App() {
 
     setAgent(name);
     setItems([]);
+    setUsage(emptyUsage());
     setStreaming("");
     setTodo([]);
     setRunning(false);
@@ -1017,9 +1041,12 @@ export function App() {
       else setShowGate(true);
       setSessionId(target.sessionId);
       try {
-        setItems(itemsFromMessages(await getSessionMessages(target.sessionId)));
+        const messages = await getSessionMessages(target.sessionId);
+        setItems(itemsFromMessages(messages));
+        setUsage(usageFromMessages(messages));
       } catch {
         setItems([]);
+        setUsage(emptyUsage());
       }
       return;
     }
@@ -1043,6 +1070,7 @@ export function App() {
     setShowGate(false);
     setGateCreate(false);
     setItems([]);
+    setUsage(emptyUsage());
     setStreaming("");
     setTodo([]);
     setSessionId(newId());
@@ -1055,6 +1083,7 @@ export function App() {
     const target = forAgent || agent;
     setSurface("session");
     setItems([]);
+    setUsage(emptyUsage());
     setStreaming("");
     setTodo([]);
     setRunning(false);
@@ -1079,6 +1108,7 @@ export function App() {
     // Archiving the open chat: leave it and start fresh (it moves to the Archived section).
     if (archived && id === sessionId) {
       setItems([]);
+      setUsage(emptyUsage());
       setStreaming("");
       setTodo([]);
       setRunning(false);
@@ -1091,6 +1121,7 @@ export function App() {
     refreshSessions();
     if (id === sessionId) {
       setItems([]);
+      setUsage(emptyUsage());
       setStreaming("");
       setTodo([]);
       setRunning(false);
@@ -1569,6 +1600,8 @@ export function App() {
               onUnattendedChange={agent !== "chat" ? toggleUnattended : undefined}
               prefill={composerPrefill}
               resetKey={sessionId}
+              usage={usage}
+              contextWindow={modelContextWindows[model]}
               placeholder={
                 agent === "code"
                   ? "Ask the coder to build, fix, or explain…  (drop or paste files)"

surfaces/gui/src/api.ts

@@ -141,6 +141,9 @@ export interface ConversationMessage {
   tool_calls?: any[];
   tool_call_id?: string;
   source?: MessageSource;
+  // Token counts for the round-trip that produced an assistant message
+  // ({model, input, output, cache_read, cache_write}); absent on older servers.
+  usage?: import("./types").TurnUsage;
   [key: string]: any;
 }
 
@@ -691,6 +694,9 @@ export interface ModelSettings {
   sessions_peek?: number;
   // Curated-matrix display names ({full id → "GLM-5.2 · via Together"}); custom models absent.
   model_labels?: Record<string, string>;
+  // {full id → context window in tokens}, verified matrix entries only — drives the
+  // composer's context-fill meter (absent id → the meter hides). Optional for older backends.
+  model_context_windows?: Record<string, number>;
   // Token savings (PDF attachments): fallback for models without native PDF support,
   // and attach-time thresholds. Optional so the GUI is robust to an older backend.
   pdf_fallback?: "text" | "images";

surfaces/gui/src/components/Composer.tsx

@@ -1,7 +1,8 @@
 import { useEffect, useLayoutEffect, useRef, useState, type ReactNode } from "react";
-import type { Attachment } from "../types";
+import type { Attachment, SessionUsage } from "../types";
 import { isPdfFile, readFile } from "../attach";
 import { getSettings, inspectPdf } from "../api";
+import { formatTokens, totalTokens } from "../usage";
 import { Dropdown, type Option } from "./Dropdown";
 import { Icon } from "./Icon";
 import { Toggle } from "./Toggle";
@@ -77,6 +78,12 @@ interface Props {
   resetKey?: string;
   // Surface-specific hint shown in the empty textarea.
   placeholder?: string;
+  // Per-session token usage (OPE-42) — absent/empty hides the usage chip entirely
+  // (older servers, backends that don't report usage, fresh sessions).
+  usage?: SessionUsage;
+  // Context-window size (tokens) of the ACTIVE model, from the curated matrix;
+  // undefined hides the fill meter (unverified/custom models) but keeps the counts.
+  contextWindow?: number;
 }
 
 export function Composer(props: Props) {
@@ -462,6 +469,18 @@ export function Composer(props: Props) {
 
           <span className="ml-auto" />
 
+          {/* token usage (OPE-42) — a quiet meter+count chip; hidden until the server
+              reports usage. Fill = context-window occupancy (bounded), count = session
+              consumption (unbounded, so never a fill). */}
+          {!dictation?.recording && props.usage && totalTokens(props.usage) > 0 && (
+            <UsageChip
+              usage={props.usage}
+              contextWindow={props.contextWindow}
+              model={props.model}
+              modelLabels={props.modelLabels}
+            />
+          )}
+
           {/* model — a quiet chip, now for the session's whole life (§17 rev 2026-07-22:
               mid-session switching shipped, so the picker stays actionable; the topbar
               subtitle still states the current model). */}
@@ -546,6 +565,135 @@ export function Composer(props: Props) {
   );
 }
 
+// Token-usage chip + popover (OPE-42). Trigger: a tiny context-fill meter (only when the
+// active model's window is known) + the session's total token count. Click → per-model
+// breakdown. Tokens only, never dollars (true cost is unknowable client-side — discounted
+// pricing, per-provider cache billing).
+function UsageChip({
+  usage,
+  contextWindow,
+  model,
+  modelLabels,
+}: {
+  usage: SessionUsage;
+  contextWindow?: number;
+  model: string;
+  modelLabels?: Record<string, string>;
+}) {
+  const [open, setOpen] = useState(false);
+  const total = totalTokens(usage);
+  const pct = contextWindow
+    ? Math.min(100, Math.round((usage.context / contextWindow) * 100))
+    : null;
+  const labelFor = (id: string) =>
+    id === "unknown" ? "Unknown model" : modelLabels?.[id] || shortModel(id);
+  // One field per line, session-summed (owner ask 2026-07-28). Values are cumulative
+  // across the whole session, never just the last turn; "Input" is the fresh
+  // (uncached) share — the cached share sits in the cache rows at its own price.
+  const stat = (label: string, value: number) => (
+    <div className="flex items-baseline justify-between text-[11.5px] leading-snug">
+      <span className="text-faint">{label}</span>
+      <span className="text-ink tabular-nums">{formatTokens(value)}</span>
+    </div>
+  );
+  return (
+    <div className="relative">
+      <button
+        className="inline-flex items-center gap-1.5 px-2 py-1 rounded-lg text-[11.5px] text-muted hover:text-ink hover:bg-paper shrink-0"
+        onClick={() => setOpen((v) => !v)}
+        aria-haspopup="menu"
+        aria-expanded={open}
+        aria-label="Token usage"
+        title={
+          pct !== null
+            ? `Token usage — ${pct}% of the context window used`
+            : "Token usage this session"
+        }
+        data-testid="usage-chip"
+      >
+        {pct !== null && (
+          <span className="w-7 h-1 rounded-full bg-line overflow-hidden" aria-hidden="true">
+            <span
+              className="block h-full bg-accent transition-all"
+              style={{ width: `${Math.max(pct, 4)}%` }}
+            />
+          </span>
+        )}
+        <span className="tabular-nums">{formatTokens(total)}</span>
+      </button>
+      {open && (
+        <>
+          <div className="fixed inset-0 z-30" onClick={() => setOpen(false)} />
+          <div
+            className="absolute z-40 bottom-full mb-1 right-0 w-[280px] rounded-xl border border-line bg-panel shadow-2xl p-3"
+            role="menu"
+            data-testid="usage-popover"
+          >
+            {contextWindow ? (
+              <div className="mb-2.5">
+                <div className="text-[10.5px] uppercase tracking-[0.06em] text-faint font-semibold mb-1">
+                  Context window
+                </div>
+                <div className="h-1.5 rounded-full bg-line overflow-hidden">
+                  <div
+                    className="h-full bg-accent transition-all"
+                    style={{ width: `${pct}%` }}
+                  />
+                </div>
+                <div className="mt-1 text-[11.5px] text-muted tabular-nums">
+                  {formatTokens(usage.context)} of {formatTokens(contextWindow)} · {pct}%
+                </div>
+              </div>
+            ) : usage.context > 0 ? (
+              <div className="mb-2.5 text-[11.5px] text-muted tabular-nums">
+                In context now: {formatTokens(usage.context)} tokens
+              </div>
+            ) : null}
+            <div className="text-[10.5px] uppercase tracking-[0.06em] text-faint font-semibold mb-1">
+              Session totals
+            </div>
+            <div className="flex flex-col gap-1.5">
+              {Object.entries(usage.byModel).map(([id, t]) => (
+                <div key={id}>
+                  <div className="text-[12px] text-ink font-medium truncate" title={id}>
+                    {labelFor(id)}
+                  </div>
+                  {/* Every row is a session sum. With a cache split, the input rows are
+                      the three BILLING CLASSES of input (each priced differently) and
+                      read as components: uncached + cache reads + cache writes = total.
+                      Without one (Ollama, compat vendors), plain "Input" says it all. */}
+                  <div className="mt-0.5 flex flex-col gap-0.5">
+                    {t.cache_read + t.cache_write > 0 ? (
+                      <>
+                        {stat("Uncached input", t.input)}
+                        {stat("Cache reads", t.cache_read)}
+                        {stat("Cache writes", t.cache_write)}
+                        {stat("Total input", t.input + t.cache_read + t.cache_write)}
+                      </>
+                    ) : (
+                      stat("Input", t.input)
+                    )}
+                    {stat("Output", t.output)}
+                  </div>
+                </div>
+              ))}
+            </div>
+            <div className="mt-2 pt-2 border-t border-line flex items-baseline justify-between text-[11.5px]">
+              <span className="text-faint">Total</span>
+              <span className="text-ink tabular-nums">{formatTokens(total)} tokens</span>
+            </div>
+            {model && !modelLabels?.[model] && contextWindow === undefined && (
+              <div className="mt-1 text-[10.5px] text-faint leading-snug">
+                Context meter unavailable for custom models.
+              </div>
+            )}
+          </div>
+        </>
+      )}
+    </div>
+  );
+}
+
 // The composer's Mode menu (§22): a quiet "Mode ⌄" chip opening the five permission options with
 // the current one marked, plus — when the session supports it — the "Send approvals to Inbox"
 // toggle at the bottom (the old standalone InboxControl, folded in).

surfaces/gui/src/types.ts

@@ -39,6 +39,25 @@ export interface TodoItem {
   status: "pending" | "in_progress" | "done";
 }
 
+// Per-round-trip token counts, as attached by the server to assistant messages and
+// the assistant_message event (`{model, input, output, cache_read, cache_write}`).
+// Absent on older servers and on backends that don't report usage.
+export interface TurnUsage {
+  model?: string | null;
+  input: number;
+  output: number;
+  cache_read: number;
+  cache_write: number;
+}
+
+// Per-session accumulation, keyed by model id (multiple models when the user
+// switched mid-session). `context` = the latest round-trip's prompt-side total —
+// what currently occupies the active model's context window.
+export interface SessionUsage {
+  byModel: Record<string, TurnUsage>;
+  context: number;
+}
+
 export interface SessionInfo {
   session_id: string;
   title?: string;

surfaces/gui/src/usage.test.ts

@@ -0,0 +1,78 @@
+import { describe, expect, it } from "vitest";
+import { addTurnUsage, emptyUsage, formatTokens, totalTokens, usageFromMessages } from "./usage";
+
+const turn = (over: Record<string, unknown> = {}) => ({
+  model: "anthropic:claude-fable-5",
+  input: 100,
+  output: 20,
+  cache_read: 50,
+  cache_write: 10,
+  ...over,
+});
+
+describe("addTurnUsage", () => {
+  it("accumulates per model and tracks the latest prompt-side total as context", () => {
+    let u = addTurnUsage(emptyUsage(), turn());
+    u = addTurnUsage(u, turn({ input: 40, output: 5, cache_read: 200, cache_write: 0 }));
+    const m = u.byModel["anthropic:claude-fable-5"];
+    expect(m).toEqual({
+      model: "anthropic:claude-fable-5",
+      input: 140,
+      output: 25,
+      cache_read: 250,
+      cache_write: 10,
+    });
+    // context = LAST turn's input + cache_read + cache_write, not a sum.
+    expect(u.context).toBe(240);
+  });
+
+  it("keys separate models separately (mid-session switch)", () => {
+    let u = addTurnUsage(emptyUsage(), turn());
+    u = addTurnUsage(u, turn({ model: "gpt-5.5", input: 7 }));
+    expect(Object.keys(u.byModel).sort()).toEqual(["anthropic:claude-fable-5", "gpt-5.5"]);
+  });
+
+  it("ignores malformed payloads and clamps negatives", () => {
+    expect(addTurnUsage(emptyUsage(), null)).toEqual(emptyUsage());
+    expect(addTurnUsage(emptyUsage(), "x")).toEqual(emptyUsage());
+    const u = addTurnUsage(emptyUsage(), turn({ input: -5, output: "9" }));
+    expect(u.byModel["anthropic:claude-fable-5"].input).toBe(0);
+    expect(u.byModel["anthropic:claude-fable-5"].output).toBe(9);
+  });
+
+  it("buckets a missing model id under 'unknown'", () => {
+    const u = addTurnUsage(emptyUsage(), turn({ model: undefined }));
+    expect(u.byModel["unknown"].input).toBe(100);
+  });
+});
+
+describe("usageFromMessages", () => {
+  it("folds assistant usage sidecars and ignores everything else", () => {
+    const u = usageFromMessages([
+      { role: "user", content: "hi" },
+      { role: "assistant", content: "a", usage: turn() as any },
+      { role: "tool", content: "r", usage: turn() as any }, // wrong role — ignored
+      { role: "assistant", content: "b" }, // no sidecar (older server) — ignored
+      { role: "assistant", content: "c", usage: turn({ input: 300 }) as any },
+    ]);
+    expect(u.byModel["anthropic:claude-fable-5"].input).toBe(400);
+    expect(u.context).toBe(360);
+  });
+});
+
+describe("totalTokens / formatTokens", () => {
+  it("totals across models and directions", () => {
+    let u = addTurnUsage(emptyUsage(), turn());
+    u = addTurnUsage(u, turn({ model: "gpt-5.5" }));
+    expect(totalTokens(u)).toBe(360);
+  });
+
+  it("formats humane numbers", () => {
+    expect(formatTokens(0)).toBe("0");
+    expect(formatTokens(980)).toBe("980");
+    expect(formatTokens(12_400)).toBe("12.4k");
+    expect(formatTokens(982_000)).toBe("982k");
+    expect(formatTokens(1_240_000)).toBe("1.24M");
+    expect(formatTokens(NaN)).toBe("0");
+  });
+});

surfaces/gui/src/usage.ts

@@ -0,0 +1,75 @@
+// Per-session token-usage accumulation (OPE-42). Pure functions so the reducer is
+// unit-testable without the app. The server attaches a `usage` sidecar
+// ({model, input, output, cache_read, cache_write}) to assistant messages and to the
+// assistant_message event; older servers and non-reporting backends send none, and
+// everything here no-ops gracefully in that case.
+
+import type { ConversationMessage } from "./api";
+import type { SessionUsage, TurnUsage } from "./types";
+
+export function emptyUsage(): SessionUsage {
+  return { byModel: {}, context: 0 };
+}
+
+const num = (v: any): number => {
+  const n = Number(v);
+  return Number.isFinite(n) && n > 0 ? n : 0;
+};
+
+/** Fold one turn's usage sidecar into the session accumulation. */
+export function addTurnUsage(prev: SessionUsage, raw: any): SessionUsage {
+  if (!raw || typeof raw !== "object") return prev;
+  const turn: TurnUsage = {
+    model: typeof raw.model === "string" && raw.model ? raw.model : null,
+    input: num(raw.input),
+    output: num(raw.output),
+    cache_read: num(raw.cache_read),
+    cache_write: num(raw.cache_write),
+  };
+  const key = turn.model || "unknown";
+  const cur = prev.byModel[key];
+  return {
+    byModel: {
+      ...prev.byModel,
+      [key]: {
+        model: turn.model,
+        input: (cur?.input || 0) + turn.input,
+        output: (cur?.output || 0) + turn.output,
+        cache_read: (cur?.cache_read || 0) + turn.cache_read,
+        cache_write: (cur?.cache_write || 0) + turn.cache_write,
+      },
+    },
+    // Prompt-side total of the LATEST round-trip = what currently sits in the
+    // context window (not a sum — each request resends the whole history).
+    context: turn.input + turn.cache_read + turn.cache_write,
+  };
+}
+
+/** Rebuild the accumulation from a replayed transcript (session load/switch). */
+export function usageFromMessages(messages: ConversationMessage[]): SessionUsage {
+  let acc = emptyUsage();
+  for (const m of messages || []) {
+    if (m.role === "assistant" && m.usage) acc = addTurnUsage(acc, m.usage);
+  }
+  return acc;
+}
+
+/** All tokens consumed this session, across models and directions (chip headline). */
+export function totalTokens(u: SessionUsage): number {
+  return Object.values(u.byModel).reduce(
+    (sum, t) => sum + t.input + t.output + t.cache_read + t.cache_write,
+    0,
+  );
+}
+
+/** 980 → "980", 12_400 → "12.4k", 982_000 → "982k", 1_240_000 → "1.24M". */
+export function formatTokens(n: number): string {
+  if (!Number.isFinite(n) || n <= 0) return "0";
+  if (n < 1000) return String(Math.round(n));
+  if (n < 1_000_000) {
+    const k = n / 1000;
+    return (k < 100 ? k.toFixed(1).replace(/\.0$/, "") : String(Math.round(k))) + "k";
+  }
+  const m = n / 1_000_000;
+  return (m < 100 ? m.toFixed(2).replace(/\.?0+$/, "") : String(Math.round(m))) + "M";
+}