opencode-visual-cache 1.2.9 → 1.2.10

package/README.md

@@ -134,6 +134,8 @@ npm install -g opencode-visual-cache@latest
 
 通过 `/cache-section` 切换后即时生效，无需重启。此外，该命令还可以开关面板的**外边框**——关闭后内容会顶格显示，释放额外空间。
 
+> **关于 Token 分布数值**：分布面板中"总计"为最后一次 API 调用的精确 token 数，"系统提示"/"用户"等分项为字符级 BPE 估算值。分项之和通常小于总计，差值主要来自 OpenCode 运行时注入的系统提示组成部分，包括环境信息、Skill 目录、工具 Schema 定义等（详见 [`system.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/system.ts)、[`tools.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/tools.ts)）。这些内容不在 agent 配置的 `prompt` 字段中，因此插件无法估算，属于预期行为。
+
 ---
 
 ## 5. 更新

package/dist/_version.d.ts

@@ -1 +1 @@
-export declare const PLUGIN_VERSION = "1.2.9";
+export declare const PLUGIN_VERSION = "1.2.10";

package/dist/_version.js

@@ -1,2 +1,2 @@
 // auto-generated
-export const PLUGIN_VERSION = "1.2.9";
+export const PLUGIN_VERSION = "1.2.10";

package/dist/index.js

@@ -216,7 +216,7 @@ function fmt(n) {
     if (n >= 1_000_000)
         return (n / 1_000_000).toFixed(1) + "M";
     if (n >= 10_000)
-        return Math.round(n / 1_000) + "K";
+        return (n / 1_000).toFixed(1) + "K";
     return n.toLocaleString("en-US");
 }
 function num(v) {
@@ -256,16 +256,19 @@ function estimateTokens(text) {
         else
             ascii++;
     }
-    // Tighten the ASCII ratio for structured content where punctuation is
-    // token-dense.  JSON key-value patterns and code keywords are strong
-    // signals that the default 4:1 ratio will materially under-estimate.
+    // Real BPE tokenizers (cl100k_base, o200k_base) average ~3.5-4.0
+    // ASCII chars/token for both JSON and source code — close to prose.
+    // The old 2.0 / 2.5 ratios matched minified-JS extremes, not typical
+    // payloads, and systematically over-estimated token counts.
     const trimmed = text.trimStart();
-    const jsonLike = (trimmed.startsWith("{") || trimmed.startsWith("["))
+    // Strip markdown code-fence prefix so that ```json … is detected as JSON
+    const strippedFence = trimmed.replace(/^\x60{3}\w*\s*\n?/, "");
+    const jsonLike = (strippedFence.startsWith("{") || strippedFence.startsWith("["))
         && /"[^"]+"\s*:/.test(text);
     const codeLike = !jsonLike
         && /```|^import |^export |^function |^const |^let |^var |^class |^interface |^type |^def |^fn |^pub |^use |^mod |^package /m.test(text);
-    const asciiPerToken = jsonLike ? 2 : codeLike ? 2.5 : 4;
-    return Math.max(1, Math.ceil(ascii / asciiPerToken + cjk / 1.5));
+    const asciiPerToken = jsonLike ? 3.5 : codeLike ? 3.5 : 4;
+    return Math.max(1, Math.ceil(ascii / asciiPerToken + cjk / 1.0));
 }
 const CURRENCIES = {
     USD: "$", CNY: "¥", EUR: "€", JPY: "JP¥", GBP: "£", KRW: "₩",
@@ -376,6 +379,7 @@ function TokenCachePanel(props) {
                 const sysPrompt = typeof agentCfg?.prompt === "string" ? agentCfg.prompt : "";
                 if (sysPrompt)
                     dist.system = estimateTokens(sysPrompt);
+                let lastAssMsg;
                 for (const msg of msgs) {
                     if (msg.role === "user") {
                         const um = msg;
@@ -409,14 +413,9 @@ function TokenCachePanel(props) {
                                 const tp = p;
                                 let rawInput = "";
                                 try {
-                                    rawInput = tp.state.raw ?? JSON.stringify(tp.state.input);
-                                }
-                                catch {
-                                    try {
-                                        rawInput = JSON.stringify(tp.state);
-                                    }
-                                    catch { }
+                                    rawInput = tp.state.raw ?? (tp.state.input != null ? JSON.stringify(tp.state.input) : "");
                                 }
+                                catch { }
                                 if (rawInput)
                                     dist.toolCall += estimateTokens(rawInput);
                                 if (tp.state.status === "completed") {
@@ -436,19 +435,23 @@ function TokenCachePanel(props) {
                                 const sub = p;
                                 dist.agent += estimateTokens(sub.prompt || sub.description || "");
                             }
-                            else if (p.type === "step-finish") {
-                                const sf = p;
-                                dist.apiInput += sf.tokens?.input ?? 0;
-                                dist.apiOutput += sf.tokens?.output ?? 0;
-                            }
                         }
                     }
                 }
-                const totalInput = dist.system + dist.user + dist.agent + dist.toolCall + dist.toolResult;
-                const overhead = Math.max(0, dist.apiInput - totalInput);
-                if (overhead >= 50)
-                    dist.system += overhead;
-                hasDistData = totalInput > 0 || dist.apiOutput > 0 || dist.apiInput > 0;
+                // 从后往前找最后一条有 token 数据的 assistant 消息（避免取到 streaming 中未填充的消息）
+                for (let i = msgs.length - 1; i >= 0; i--) {
+                    if (msgs[i].role !== "assistant")
+                        continue;
+                    const t = msgs[i].tokens;
+                    if (t && (t.input > 0 || (t.cache?.read ?? 0) > 0)) {
+                        lastAssMsg = msgs[i];
+                        break;
+                    }
+                }
+                // 取最后一条有数据消息的总输入（含缓存读）作为当前 context 大小
+                dist.apiInput = num(lastAssMsg?.tokens?.input) + num(lastAssMsg?.tokens?.cache?.read);
+                dist.apiOutput = num(lastAssMsg?.tokens?.output);
+                hasDistData = dist.system + dist.user + dist.agent + dist.toolCall + dist.toolResult > 0 || dist.apiOutput > 0 || dist.apiInput > 0;
             }
             catch { }
             const finalDist = hasDistData ? dist : lastDist(), finalHasDist = hasDistData || lastHasDist();
@@ -626,7 +629,7 @@ function TokenCachePanel(props) {
             // boxEl.width may be undefined before the first measurement — guard with 0
             const w = boxEl ? Math.max(MIN_PANEL_WIDTH, boxEl.width ?? 0) : DEFAULT_PANEL_WIDTH;
             setPanelWidth((prev) => (prev === w ? prev : w));
-        }, children: [_jsxs("text", { onMouseUp: () => setOpen((o) => { const n = !o; persistFold("open", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: open() ? "\u25bc " : "\u25b6 " }), _jsxs("span", { style: { fg: pal().primary }, children: [_jsx("b", { children: T.title }), _jsx(Show, { when: open(), children: _jsxs("span", { style: { fg: pal().muted }, children: [" (v", PLUGIN_VERSION, ")"] }) })] }), _jsxs(Show, { when: !open() && data().hasData, children: [_jsxs(Show, { when: data().hasTrendData, children: [_jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - HEADER_PREFIX - visualWidth(T.title) - visualWidth(pct() + " " + T.hitFolded + " " + trendLabel(data().trend)))) }), _jsxs("span", { style: { fg: hitColor() }, children: [pct(), " ", T.hitFolded] }), _jsxs("span", { style: { fg: data().trend !== 0 ? (data().trend > 0 ? pal().success : pal().error) : pal().text }, children: [" ", trendLabel(data().trend)] })] }), _jsxs(Show, { when: !data().hasTrendData, children: [_jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - HEADER_PREFIX - visualWidth(T.title) - visualWidth(pct() + " " + T.hitFolded))) }), _jsxs("span", { style: { fg: hitColor() }, children: [pct(), " ", T.hitFolded] })] })] })] }), _jsx(Show, { when: open(), children: _jsxs(Show, { when: data().hasData, fallback: _jsxs(_Fragment, { children: [_jsx("text", { fg: pal().muted, children: sep() }), _jsxs("text", { children: [_jsx("span", { style: { fg: pal().muted }, children: "> " }), _jsx("span", { style: { fg: pal().muted }, children: T.noData })] })] }), children: [_jsx("text", { fg: pal().muted, children: sep() }), _jsxs("text", { children: [_jsxs("span", { style: { fg: pal().text }, children: [T.hit, " "] }), _jsxs("span", { style: { fg: hitColor() }, children: ["[", bar(), "] "] }), _jsx("span", { style: { fg: pal().text }, children: pct() }), _jsx(Show, { when: data().hasTrendData, children: _jsxs("span", { style: { fg: data().trend !== 0 ? (data().trend > 0 ? pal().success : pal().error) : pal().text }, children: [" ", trendLabel(data().trend)] }) })] }), _jsx("text", { fg: pal().muted, children: justify(T.totalHit, (Math.floor(data().sessionHitRate * 10) / 10).toFixed(1) + "%") }), _jsxs(Show, { when: sectionDetail(), children: [_jsxs("text", { onMouseUp: () => setDetailOpen((o) => { const n = !o; persistFold("detail", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: detailOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.secDetail }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((detailOpen() ? "\u25bc " : "\u25b6 ") + T.secDetail)) })] }), _jsxs(Show, { when: detailOpen(), children: [_jsx(Show, { when: data().read > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.read, fmt(data().read), T.tok) }) }), _jsx(Show, { when: data().write > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.write, fmt(data().write), T.tok) }) }), _jsx("text", { fg: pal().muted, children: justify(T.miss, fmt(data().freshInput), T.tok) }), _jsx("text", { fg: pal().muted, children: justify(T.out, fmt(data().output), T.tok) }), _jsx(Show, { when: data().saved > 0, children: _jsxs("text", { children: [_jsx("span", { style: { fg: pal().muted }, children: T.saved }), _jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - visualWidth(T.saved) - visualWidth("~" + fmtCost(data().saved, currencySymbol(), exchangeRate())))) }), _jsxs("span", { style: { fg: pal().success }, children: ["~", fmtCost(data().saved, currencySymbol(), exchangeRate())] })] }) })] })] }), _jsxs(Show, { when: sectionModel(), children: [_jsxs("text", { onMouseUp: () => setModelOpen((o) => { const n = !o; persistFold("model", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: modelOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.secModel }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((modelOpen() ? "\u25bc " : "\u25b6 ") + T.secModel)) })] }), _jsxs(Show, { when: modelOpen(), children: [_jsx("text", { fg: pal().text, children: justify(T.cost, fmtCost(data().cost, currencySymbol(), exchangeRate())) }), _jsx(Show, { when: data().providerName, children: _jsx("text", { fg: pal().muted, children: justify(T.provider, data().providerName) }) }), _jsx("text", { fg: pal().muted, children: justify(T.model, data().model) }), _jsxs(Show, { when: data().hasPricing, children: [_jsx("text", { fg: pal().muted, children: justify(T.rate, currencySymbol() + (data().inputRate * exchangeRate()).toFixed(2) + "/M " + T.inputRate) }), _jsx(Show, { when: data().cacheReadRate > 0, children: _jsx("text", { fg: pal().muted, children: justify("", currencySymbol() + (data().cacheReadRate * exchangeRate()).toFixed(2) + "/M " + T.cacheRate) }) }), _jsx(Show, { when: data().cacheWriteRate > 0, children: _jsx("text", { fg: pal().muted, children: justify("", currencySymbol() + (data().cacheWriteRate * exchangeRate()).toFixed(2) + "/M " + T.writeRate) }) })] })] })] }), _jsx(Show, { when: sectionDist(), children: _jsxs(Show, { when: data().hasDistData, children: [_jsxs("text", { onMouseUp: () => setDistOpen((o) => { const n = !o; persistFold("dist", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: distOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.distTitle }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((distOpen() ? "\u25bc " : "\u25b6 ") + T.distTitle)) })] }), _jsxs(Show, { when: distOpen(), children: [_jsx(Show, { when: data().dist.system > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distSys, fmt(data().dist.system), T.tok) }) }), _jsx(Show, { when: data().dist.user > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distUser, fmt(data().dist.user), T.tok) }) }), _jsx(Show, { when: data().dist.agent > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distAgent, fmt(data().dist.agent), T.tok) }) }), _jsx(Show, { when: data().dist.toolCall > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distTool, fmt(data().dist.toolCall), T.tok) }) }), _jsx(Show, { when: data().dist.toolResult > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distRes, fmt(data().dist.toolResult), T.tok) }) }), _jsx("text", { fg: pal().text, children: justify(T.distTotal, fmt(data().dist.system + data().dist.user + data().dist.agent + data().dist.toolCall + data().dist.toolResult), T.tok) })] })] }) })] }) })] }));
+        }, children: [_jsxs("text", { onMouseUp: () => setOpen((o) => { const n = !o; persistFold("open", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: open() ? "\u25bc " : "\u25b6 " }), _jsxs("span", { style: { fg: pal().primary }, children: [_jsx("b", { children: T.title }), _jsx(Show, { when: open(), children: _jsxs("span", { style: { fg: pal().muted }, children: [" (v", PLUGIN_VERSION, ")"] }) })] }), _jsxs(Show, { when: !open() && data().hasData, children: [_jsxs(Show, { when: data().hasTrendData, children: [_jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - HEADER_PREFIX - visualWidth(T.title) - visualWidth(pct() + " " + T.hitFolded + " " + trendLabel(data().trend)))) }), _jsxs("span", { style: { fg: hitColor() }, children: [pct(), " ", T.hitFolded] }), _jsxs("span", { style: { fg: data().trend !== 0 ? (data().trend > 0 ? pal().success : pal().error) : pal().text }, children: [" ", trendLabel(data().trend)] })] }), _jsxs(Show, { when: !data().hasTrendData, children: [_jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - HEADER_PREFIX - visualWidth(T.title) - visualWidth(pct() + " " + T.hitFolded))) }), _jsxs("span", { style: { fg: hitColor() }, children: [pct(), " ", T.hitFolded] })] })] })] }), _jsx(Show, { when: open(), children: _jsxs(Show, { when: data().hasData, fallback: _jsxs(_Fragment, { children: [_jsx("text", { fg: pal().muted, children: sep() }), _jsxs("text", { children: [_jsx("span", { style: { fg: pal().muted }, children: "> " }), _jsx("span", { style: { fg: pal().muted }, children: T.noData })] })] }), children: [_jsx("text", { fg: pal().muted, children: sep() }), _jsxs("text", { children: [_jsxs("span", { style: { fg: pal().text }, children: [T.hit, " "] }), _jsxs("span", { style: { fg: hitColor() }, children: ["[", bar(), "] "] }), _jsx("span", { style: { fg: pal().text }, children: pct() }), _jsx(Show, { when: data().hasTrendData, children: _jsxs("span", { style: { fg: data().trend !== 0 ? (data().trend > 0 ? pal().success : pal().error) : pal().text }, children: [" ", trendLabel(data().trend)] }) })] }), _jsx("text", { fg: pal().muted, children: justify(T.totalHit, (Math.floor(data().sessionHitRate * 10) / 10).toFixed(1) + "%") }), _jsxs(Show, { when: sectionDetail(), children: [_jsxs("text", { onMouseUp: () => setDetailOpen((o) => { const n = !o; persistFold("detail", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: detailOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.secDetail }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((detailOpen() ? "\u25bc " : "\u25b6 ") + T.secDetail)) })] }), _jsxs(Show, { when: detailOpen(), children: [_jsx(Show, { when: data().read > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.read, fmt(data().read), T.tok) }) }), _jsx(Show, { when: data().write > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.write, fmt(data().write), T.tok) }) }), _jsx("text", { fg: pal().muted, children: justify(T.miss, fmt(data().freshInput), T.tok) }), _jsx("text", { fg: pal().muted, children: justify(T.out, fmt(data().output), T.tok) }), _jsx(Show, { when: data().saved > 0, children: _jsxs("text", { children: [_jsx("span", { style: { fg: pal().muted }, children: T.saved }), _jsx("span", { children: " ".repeat(Math.max(1, panelWidth() - gutter() - visualWidth(T.saved) - visualWidth("~" + fmtCost(data().saved, currencySymbol(), exchangeRate())))) }), _jsxs("span", { style: { fg: pal().success }, children: ["~", fmtCost(data().saved, currencySymbol(), exchangeRate())] })] }) })] })] }), _jsxs(Show, { when: sectionModel(), children: [_jsxs("text", { onMouseUp: () => setModelOpen((o) => { const n = !o; persistFold("model", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: modelOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.secModel }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((modelOpen() ? "\u25bc " : "\u25b6 ") + T.secModel)) })] }), _jsxs(Show, { when: modelOpen(), children: [_jsx("text", { fg: pal().text, children: justify(T.cost, fmtCost(data().cost, currencySymbol(), exchangeRate())) }), _jsx(Show, { when: data().providerName, children: _jsx("text", { fg: pal().muted, children: justify(T.provider, data().providerName) }) }), _jsx("text", { fg: pal().muted, children: justify(T.model, data().model) }), _jsxs(Show, { when: data().hasPricing, children: [_jsx("text", { fg: pal().muted, children: justify(T.rate, currencySymbol() + (data().inputRate * exchangeRate()).toFixed(2) + "/M " + T.inputRate) }), _jsx(Show, { when: data().cacheReadRate > 0, children: _jsx("text", { fg: pal().muted, children: justify("", currencySymbol() + (data().cacheReadRate * exchangeRate()).toFixed(2) + "/M " + T.cacheRate) }) }), _jsx(Show, { when: data().cacheWriteRate > 0, children: _jsx("text", { fg: pal().muted, children: justify("", currencySymbol() + (data().cacheWriteRate * exchangeRate()).toFixed(2) + "/M " + T.writeRate) }) })] })] })] }), _jsx(Show, { when: sectionDist(), children: _jsxs(Show, { when: data().hasDistData, children: [_jsxs("text", { onMouseUp: () => setDistOpen((o) => { const n = !o; persistFold("dist", n); return n; }), children: [_jsx("span", { style: { fg: pal().muted }, children: distOpen() ? "\u25bc " : "\u25b6 " }), _jsx("span", { style: { fg: pal().primary }, children: _jsx("b", { children: T.distTitle }) }), _jsx("span", { style: { fg: pal().muted }, children: sep().slice(visualWidth((distOpen() ? "\u25bc " : "\u25b6 ") + T.distTitle)) })] }), _jsxs(Show, { when: distOpen(), children: [_jsx(Show, { when: data().dist.system > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distSys, fmt(data().dist.system), T.tok) }) }), _jsx(Show, { when: data().dist.user > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distUser, fmt(data().dist.user), T.tok) }) }), _jsx(Show, { when: data().dist.agent > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distAgent, fmt(data().dist.agent), T.tok) }) }), _jsx(Show, { when: data().dist.toolCall > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distTool, fmt(data().dist.toolCall), T.tok) }) }), _jsx(Show, { when: data().dist.toolResult > 0, children: _jsx("text", { fg: pal().muted, children: justify(T.distRes, fmt(data().dist.toolResult), T.tok) }) }), _jsx("text", { fg: pal().text, children: justify(T.distTotal, fmt(data().dist.apiInput), T.tok) })] })] }) })] }) })] }));
 }
 // ---------------------------------------------------------------------------
 // Plugin entry

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-visual-cache",
-  "version": "1.2.9",
+  "version": "1.2.10",
   "description": "OpenCode TUI plugin displaying real-time token cache hit rate in the sidebar",
   "type": "module",
   "main": "dist/index.js",

package/src/_version.ts

@@ -1,2 +1,2 @@
 // auto-generated
-export const PLUGIN_VERSION="1.2.9";
+export const PLUGIN_VERSION="1.2.10";

package/src/index.tsx

@@ -241,7 +241,7 @@ function progressBar(percent: number, width: number): string {
 
 function fmt(n: number): string {
   if (n >= 1_000_000) return (n / 1_000_000).toFixed(1) + "M"
-  if (n >= 10_000) return Math.round(n / 1_000) + "K"
+  if (n >= 10_000) return (n / 1_000).toFixed(1) + "K"
   return n.toLocaleString("en-US")
 }
 
@@ -277,17 +277,20 @@ function estimateTokens(text: string): number {
     else ascii++
   }
 
-  // Tighten the ASCII ratio for structured content where punctuation is
-  // token-dense.  JSON key-value patterns and code keywords are strong
-  // signals that the default 4:1 ratio will materially under-estimate.
+  // Real BPE tokenizers (cl100k_base, o200k_base) average ~3.5-4.0
+  // ASCII chars/token for both JSON and source code — close to prose.
+  // The old 2.0 / 2.5 ratios matched minified-JS extremes, not typical
+  // payloads, and systematically over-estimated token counts.
   const trimmed = text.trimStart()
-  const jsonLike = (trimmed.startsWith("{") || trimmed.startsWith("["))
+  // Strip markdown code-fence prefix so that ```json … is detected as JSON
+  const strippedFence = trimmed.replace(/^\x60{3}\w*\s*\n?/, "")
+  const jsonLike = (strippedFence.startsWith("{") || strippedFence.startsWith("["))
     && /"[^"]+"\s*:/.test(text)
   const codeLike = !jsonLike
     && /```|^import |^export |^function |^const |^let |^var |^class |^interface |^type |^def |^fn |^pub |^use |^mod |^package /m.test(text)
 
-  const asciiPerToken = jsonLike ? 2 : codeLike ? 2.5 : 4
-  return Math.max(1, Math.ceil(ascii / asciiPerToken + cjk / 1.5))
+  const asciiPerToken = jsonLike ? 3.5 : codeLike ? 3.5 : 4
+  return Math.max(1, Math.ceil(ascii / asciiPerToken + cjk / 1.0))
 }
 
 interface TokenDist {
@@ -436,6 +439,7 @@ function TokenCachePanel(props: {
         const agentCfg = agents?.[agentName] as Record<string, unknown> | undefined
         const sysPrompt = typeof agentCfg?.prompt === "string" ? agentCfg.prompt : ""
         if (sysPrompt) dist.system = estimateTokens(sysPrompt)
+        let lastAssMsg: AssistantMessage | undefined
         for (const msg of msgs) {
           if (msg.role === "user") {
             const um = msg as UserMessage; if (um.system) dist.system += estimateTokens(um.system)
@@ -445,24 +449,31 @@ function TokenCachePanel(props: {
               else if (p.type === "file") { const fp = p as any; if (fp.source?.text?.value) dist.user += estimateTokens(fp.source.text.value) }
             }
           } else if (msg.role === "assistant") {
-            const am = msg as AssistantMessage; dist.output += num(am.tokens?.output)
+            const am = msg as AssistantMessage
+            dist.output += num(am.tokens?.output)
             let parts: readonly Part[] = []; try { parts = props.api.state.part(msg.id) } catch {}
             for (const p of parts) {
               if (p.type === "tool") {
                 const tp = p as any; let rawInput = ""
-                try { rawInput = tp.state.raw ?? JSON.stringify(tp.state.input) } catch { try { rawInput = JSON.stringify(tp.state) } catch {} }
+                try { rawInput = tp.state.raw ?? (tp.state.input != null ? JSON.stringify(tp.state.input) : "") } catch {}
                 if (rawInput) dist.toolCall += estimateTokens(rawInput)
                 if (tp.state.status === "completed") { const c = tp.state; if (c.output) dist.toolResult += estimateTokens(c.output) }
                 else if (tp.state.status === "error") { const e = tp.state; if (e.error) dist.toolResult += estimateTokens(e.error) }
               } else if (p.type === "reasoning") dist.agent += estimateTokens((p as any).text)
               else if (p.type === "subtask") { const sub = p as any; dist.agent += estimateTokens(sub.prompt || sub.description || "") }
-              else if (p.type === "step-finish") { const sf = p as any; dist.apiInput += sf.tokens?.input ?? 0; dist.apiOutput += sf.tokens?.output ?? 0 }
             }
           }
         }
-        const totalInput = dist.system + dist.user + dist.agent + dist.toolCall + dist.toolResult
-        const overhead = Math.max(0, dist.apiInput - totalInput); if (overhead >= 50) dist.system += overhead
-        hasDistData = totalInput > 0 || dist.apiOutput > 0 || dist.apiInput > 0
+        // 从后往前找最后一条有 token 数据的 assistant 消息（避免取到 streaming 中未填充的消息）
+        for (let i = msgs.length - 1; i >= 0; i--) {
+          if (msgs[i].role !== "assistant") continue
+          const t = (msgs[i] as AssistantMessage).tokens
+          if (t && (t.input > 0 || (t.cache?.read ?? 0) > 0)) { lastAssMsg = msgs[i] as AssistantMessage; break }
+        }
+        // 取最后一条有数据消息的总输入（含缓存读）作为当前 context 大小
+        dist.apiInput = num(lastAssMsg?.tokens?.input) + num(lastAssMsg?.tokens?.cache?.read)
+        dist.apiOutput = num(lastAssMsg?.tokens?.output)
+        hasDistData = dist.system + dist.user + dist.agent + dist.toolCall + dist.toolResult > 0 || dist.apiOutput > 0 || dist.apiInput > 0
       } catch {}
       const finalDist = hasDistData ? dist : lastDist(), finalHasDist = hasDistData || lastHasDist()
       return { finalDist, finalHasDist }
@@ -479,7 +490,7 @@ function TokenCachePanel(props: {
 
   const data = createMemo(() => {
     return dataSignal()
-  })
+  })
 
   // Persist the last valid distribution so that data() can fall back
   // to it while api.state.part() is re-hydrating after a view switch.
@@ -819,7 +830,7 @@ function TokenCachePanel(props: {
               </text>
             </Show>
             <text fg={pal().text}>
-              {justify(T.distTotal, fmt(data().dist.system + data().dist.user + data().dist.agent + data().dist.toolCall + data().dist.toolResult), T.tok)}
+              {justify(T.distTotal, fmt(data().dist.apiInput), T.tok)}
             </text>
             </Show>
           </Show>