cc-viewer 1.7.11 → 1.7.12

package/dist/index.html

@@ -21,7 +21,7 @@
     // 整体显示大小已弃用 CSS zoom：Electron 改用 webFrame.setZoomFactor（首屏抢占见
     // electron/tab-content-preload.js），纯浏览器交由用户用浏览器自带快捷键缩放，故此处不再设 zoom。
   </script>
-  <script type="module" crossorigin src="./assets/index-gH_W4sVT.js"></script>
+  <script type="module" crossorigin src="./assets/index-QCUTFwkw.js"></script>
   <link rel="modulepreload" crossorigin href="./assets/vendor-antd-DADYo_zg.js">
   <link rel="modulepreload" crossorigin href="./assets/vendor-codemirror-tF6HNoR6.js">
   <link rel="modulepreload" crossorigin href="./assets/vendor-mdxeditor-CFAmRN3Y.js">

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cc-viewer",
-  "version": "1.7.11",
+  "version": "1.7.12",
   "description": "Claude Code logging, visualization, and management toolkit — launch a web viewer alongside Claude Code with full request/response tracing, proxy, and mobile support",
   "license": "MIT",
   "main": "server.js",

package/server/lib/context-rules.js

@@ -27,6 +27,29 @@ export function parseContextSizeSuffix(modelName) {
   return m[2].toLowerCase() === 'm' ? num * 1000000 : num * 1000;
 }
 
+/**
+ * Resolve the model name to use for context-window classification (血条窗口判定专用).
+ *
+ * Precedence differs from getEffectiveModel (response-first) on one deliberate
+ * point: an EXPLICIT [Nk]/[Nm] suffix on the REQUEST model (`body.model`, which
+ * carries the user's hot-switch config / model selector intent) is authoritative
+ * and must NOT be overridden by the upstream response. Upstream APIs normalize
+ * the response `model` — e.g. hot-switching to `k3[1m]` makes Moonshot return
+ * `response.body.model: "k3"`, stripping the [1m] marker; a response-first read
+ * would then misclassify the window (bare k3 vs the configured 1M). So: request
+ * suffix wins; otherwise fall back to the response model, then the request name.
+ *
+ * @param {object|null|undefined} request log entry with body / response
+ * @returns {string|null}
+ */
+export function getCalibrationModel(request) {
+  const reqModel = request?.body?.model;
+  if (typeof reqModel === 'string' && parseContextSizeSuffix(reqModel) != null) return reqModel;
+  const respModel = request?.response?.body?.model;
+  if (typeof respModel === 'string' && respModel) return respModel;
+  return (typeof reqModel === 'string' && reqModel) ? reqModel : null;
+}
+
 // 模型家族 → 窗口档位表(有序,首条命中)。后缀解析在表外先行(见 getModelMaxTokens)。
 const MODEL_CONTEXT_SIZES = [
   // haiku 全系 200K,显式置于一切 1M 默认之前(claude-haiku-4-5 等)
@@ -48,6 +71,10 @@ const MODEL_CONTEXT_SIZES = [
   { match: /gpt-4o|o1|o3|o4/i, tokens: 128000 },
   { match: /gpt-4/i, tokens: 128000 },
   { match: /gpt-3/i, tokens: 16000 },
+  // Kimi 家族精确档:k2.x/k3 等带 kimi/moonshot 前缀的 → 256K;裸 'k3'(无前缀,
+  // 代理直连时的简写 model 名)→ 256K 精确档但 classifyContextWindow 不升 1M
+  // (见该函数的家族特判,裸 k3 归 200K 桶,超量由 adaptContextWindow 纠偏)。
+  { match: /kimi|moonshot|^k3$/i, tokens: 256000 },
   // deepseek-v4 defaults to 1M; placed before generic /deepseek/ so the
   // first-match-wins loop picks it up before falling through to 128K.
   { match: /deepseek-v4/i, tokens: 1000000 },
@@ -74,12 +101,19 @@ export function getModelMaxTokens(modelName) {
  * 不变量:只返回 1000000 或 200000(resolveCalibrationTokens 依赖此不变量)。
  * 裸 '1m' 子串(无方括号,如 deepseek-v3-1m)→ 1M 的宽松规则仅限本分类器,
  * 刻意不进 getModelMaxTokens(后者面向精确档位)。128K/16K 档归入 200K 桶。
+ * Kimi 家族特判:kimi/moonshot 前缀型号(k2.x/k3,真实窗口 256K)归 1M 桶 ——
+ * 避免会话中段从 200K 重标定到 256K/1M 的跳变;代价是相对真实 256K 上限
+ * 长期低估(约 4 倍刻度),可接受。裸 'k3' 同样归 1M:代理热切换到
+ * 'k3[1m]' 时上游会把响应 model 归一化成裸 'k3'(剥掉 [1m] 后缀),
+ * response-first 解析读到裸 'k3' 若归 200K 桶会与请求侧 1M 判定分裂,
+ * 血条分母错成 200K;且裸 'k3' 本就是 k3[1m] 的 1M 形态被剥后缀的产物。
  * @param {string} modelName
  * @returns {1000000|200000}
  */
 export function classifyContextWindow(modelName) {
   if (!modelName || typeof modelName !== 'string') return 200000;
   if (modelName.toLowerCase().includes('1m')) return 1000000;
+  if (/kimi|moonshot|^k3$/i.test(modelName)) return 1000000;
   return getModelMaxTokens(modelName) >= 1000000 ? 1000000 : 200000;
 }
 
@@ -88,7 +122,9 @@ export function classifyContextWindow(modelName) {
  * 一个真正的 200K 模型,其输入上下文(input + cache_creation + cache_read)物理上不可能
  * 超过 200K —— 超了 API 直接拒收。所以一旦真实输入用量越过 200K 还被判成 200K,必然是
  * model 名识别错了(误判),此时自动升到 1M,免得血条卡死在 100%、百分比与真实进度脱节。
- * 仅做 200K→1M 这一个方向的纠偏;其余判定(1M、各家 200K 真值等)一律原样返回。
+ * One-way upgrades only: 200K→1M and 256K→1M (the kimi exact tier used by the
+ * server-side SSE path); every other classification (1M, 128K/16K tiers, true
+ * 200K values) is returned unchanged — 128K is deliberately never promoted.
  * 注意:usedContextTokens 必须是"输入侧"用量(sumUsageInputTokens,不含 output_tokens),
  * 否则大输出会误触发。
  * @param {number} classifiedTokens classifyContextWindow / getModelMaxTokens 的结果
@@ -97,6 +133,7 @@ export function classifyContextWindow(modelName) {
  */
 export function adaptContextWindow(classifiedTokens, usedContextTokens) {
   if (classifiedTokens === 200000 && usedContextTokens > 200000) return 1000000;
+  if (classifiedTokens === 256000 && usedContextTokens > 256000) return 1000000;
   return classifiedTokens;
 }
 

package/server/lib/context-watcher.js

@@ -2,7 +2,7 @@ import { readFileSync, existsSync, realpathSync } from 'node:fs';
 import { join } from 'node:path';
 import { homedir } from 'node:os';
 import { getClaudeConfigDir } from '../../findcc.js';
-import { getModelMaxTokens, adaptContextWindow, sumUsageInputTokens, sumUsageContextTokens } from './context-rules.js';
+import { getModelMaxTokens, adaptContextWindow, sumUsageInputTokens, sumUsageContextTokens, getCalibrationModel } from './context-rules.js';
 
 export const CONTEXT_WINDOW_FILE = join(getClaudeConfigDir(), 'context-window.json');
 export const CLAUDE_SETTINGS_FILE = join(getClaudeConfigDir(), 'settings.json');
@@ -43,10 +43,25 @@ export function readModelContextSize() {
 /**
  * Get context size for a given API model name (e.g. 'claude-opus-4-6-20250514').
  * Uses startup cache to avoid re-reading the file.
- * @param {string} apiModelName - model name from req.body.model
+ * Accepts either a bare model-name string (legacy path) or a full log entry.
+ * Entry input resolves the model via getCalibrationModel (context-rules.js):
+ * an explicit [Nk]/[Nm] suffix on the REQUEST model wins (the user's hot-switch
+ * config intent, e.g. k3[1m]); otherwise the upstream response.body.model is
+ * authoritative. The startup cache is request-side static info, stale after a
+ * hot-switch, so entry resolution skips it and goes straight to the family
+ * rules table. String input keeps legacy cache-first behavior unchanged.
+ * @param {string|object} modelOrEntry - model name, or log entry with body/response
  * @returns {number} context window size in tokens
  */
-export function getContextSizeForModel(apiModelName) {
+export function getContextSizeForModel(modelOrEntry) {
+  const isEntry = modelOrEntry !== null && typeof modelOrEntry === 'object';
+  // Entry input: calibration-aware resolution (request [Nk]/[Nm] suffix wins,
+  // else response model). Authoritative over the stale startup cache.
+  if (isEntry) {
+    const model = getCalibrationModel(modelOrEntry);
+    return model ? getModelMaxTokens(model) : (_startupContextSize || 200000);
+  }
+  const apiModelName = modelOrEntry;
   if (!apiModelName) return _startupContextSize || 200000;
   const lower = apiModelName.toLowerCase();
   // Extract base: 'claude-opus-4-6-20250514' → 'opus-4-6'
@@ -56,7 +71,7 @@ export function getContextSizeForModel(apiModelName) {
     return _startupContextSize;
   }
   // 完整档位表见 server/lib/context-rules.js(与前端同源;含 haiku/旧 opus/3-opus 200K、
-  // deepseek-v4 1M、gpt/deepseek 等三方档位,默认 200K)
+  // deepseek-v4 1M、kimi/moonshot 256K、gpt/deepseek 等三方档位,默认 200K)
   return getModelMaxTokens(apiModelName);
 }
 

package/server/lib/log-watcher.js

@@ -156,7 +156,7 @@ export function processWatchedEntry(parsed, ctx) {
     if (cached) sendEventToClients(clients, 'kv_cache_content', cached);
     const usage = parsed.response?.body?.usage;
     if (usage) {
-      const contextSize = getContextSizeForModel(parsed.body?.model);
+      const contextSize = getContextSizeForModel(parsed);
       const cwData = buildContextWindowEvent(usage, contextSize);
       if (cwData) sendEventToClients(clients, 'context_window', cwData);
     }

package/server/routes/events.js

@@ -320,7 +320,7 @@ async function events(req, res, parsedUrl, isLocal, deps) {
       if (!latestContextWindow) {
         const usage = entry.response?.body?.usage;
         if (usage) {
-          const contextSize = getContextSizeForModel(entry.body?.model);
+          const contextSize = getContextSizeForModel(entry);
           const cw = buildContextWindowEvent(usage, contextSize);
           if (cw) latestContextWindow = cw;
         }

package/src/utils/effectiveModel.js

@@ -7,6 +7,10 @@
  * server-reported model in `response.body.model` (authoritative under proxy
  * hot-switch) over the client-supplied `body.model`. Returns null when both
  * are missing — callers should fall back to a sensible default.
+ *
+ * KEEP IN SYNC: server/lib/context-watcher.js getContextSizeForModel reuses
+ * this precedence for its entry path — changing the priority here must be
+ * mirrored there (and vice versa).
  */
 export function getEffectiveModel(request) {
   return request?.response?.body?.model || request?.body?.model || null;

package/src/utils/helpers.js

@@ -20,8 +20,9 @@ export {
   sumCacheCreationTokens,
   sumUsageInputTokens,
   sumUsageContextTokens,
+  getCalibrationModel,
 } from '../../server/lib/context-rules.js';
-import { classifyContextWindow, adaptContextWindow } from '../../server/lib/context-rules.js';
+import { classifyContextWindow, adaptContextWindow, getCalibrationModel } from '../../server/lib/context-rules.js';
 
 // getEffectiveModel moved to ./effectiveModel.js (pure, node-testable — sessionMerge/sessionManager
 // import it without helpers' Vite-only svg imports); re-exported here to keep import paths stable.
@@ -57,7 +58,9 @@ const CALIBRATION_TOKEN_MAP = {
 export function resolveCalibrationTokens(calibrationModel, lastMainAgent, projectModelHint = null) {
   const direct = CALIBRATION_TOKEN_MAP[calibrationModel];
   if (direct) return direct;
-  const lastModel = lastMainAgent ? getEffectiveModel(lastMainAgent) : null;
+  // 校准用 getCalibrationModel:请求名带显式 [Nk]/[Nm] 后缀时优先(用户热切换配置的
+  // 1M 意图),不被上游响应归一化(如 k3[1m]→裸 k3)覆盖;其余回退 response-first。
+  const lastModel = lastMainAgent ? getCalibrationModel(lastMainAgent) : null;
   // 优先用真实 mainAgent 信号；haiku 一律视为 init ping 噪声，跳过
   if (typeof lastModel === 'string' && lastModel && !/haiku/i.test(lastModel)) {
     return classifyContextWindow(lastModel);
@@ -334,7 +337,7 @@ const MODEL_PROVIDERS = [
     match: /kimi|moonshot|^k3$/i,
     name: 'Kimi',
     color: 'var(--bg-model-avatar)',
-    svg: '<svg t="1771495664798" class="icon" viewBox="0 0 1024 1024" version="1.1" xmlns="http://www.w3.org/2000/svg" p-id="6649" width="200" height="200"><path d="M731.062857 590.262857v318.317714h-148.589714V443.977143a146.285714 146.285714 0 0 1-146.285714 146.285714l-214.491429-0.036571v318.354285H73.142857V165.814857h148.553143v275.858286h180.882286l119.771428-275.858286h167.753143l-67.84 156.269714a255.524571 255.524571 0 0 1-106.678857 119.588572h66.925714a148.553143 148.553143 0 0 1 148.516572 148.589714z m120.758857-473.965714a99.035429 99.035429 0 0 1 0 198.070857h-99.035428V215.332571a99.035429 99.035429 0 0 1 99.035428-99.035428z" fill="currentColor" p-id="6650"></path></svg>',
+    svg: '<svg class="icon" viewBox="0 0 1024 1024" version="1.1" xmlns="http://www.w3.org/2000/svg" width="200" height="200"><path d="M932.096 0a82.048 82.048 0 1 1 0 164.096h-72.352a9.568 9.568 0 0 1-9.664-9.632V82.016A82.048 82.048 0 0 1 932.096 0z" fill="#1783FF"></path><path d="M472.064 477.856l309.76-307.2c5.888-5.792 2.56-17.472-4.96-17.472h-166.72a7.008 7.008 0 0 0-5.056 2.112L271.456 486.24c-5.12 5.12-12.8 0.576-12.8-7.68V162.976c0-5.376-3.584-9.792-7.936-9.792H135.936c-4.352 0-7.936 4.416-7.936 9.792V843.52c0 5.44 3.584 9.824 7.936 9.824h114.784c4.352 0 7.936-4.288 7.936-9.824v-138.656c0-2.912 1.024-5.728 2.88-7.616l103.424-102.656a6.72 6.72 0 0 1 8.8-0.928l276.64 203.616a328.64 328.64 0 0 0 147.296 54.688c4.608 0.512 8.544-4 8.544-9.824V711.68c0-4.96-2.912-9.056-7.008-9.632a214.528 214.528 0 0 1-86.432-34.496l-239.456-173.472c-5.024-3.328-5.632-11.936-1.184-16.224h-0.096z" fill="currentColor"></path></svg>',
   },
   {
     match: /glm|chatglm/i,