wezard 1.2.9 → 1.2.11

package/dist/daemon/mirror-bridge.js

@@ -393,7 +393,7 @@ const renderToolInputCompact = (input, max) => {
 const renderToolUseGroupBody = (calls, deps) => {
     const renderOne = (c) => ({
         compact: safeForMarkdown(renderToolInputCompact(c.input, deps.toolUseInlineMaxChars)),
-        url: deps.detailUrlFor(c.toolUseId),
+        url: deps.detailUrlFor(c.toolUseId, deps.target),
     });
     if (calls.length === 1) {
         const c = calls[0];
@@ -489,7 +489,7 @@ const renderLine = (raw, deps) => {
                 if (toolUseId)
                     recordToolResult(toolUseId, full);
                 const compact = safeForMarkdown(oneLineSummary(full, 40));
-                const url = deps.detailUrlFor(toolUseId);
+                const url = deps.detailUrlFor(toolUseId, deps.target);
                 out.push({
                     kind: "tool_result",
                     toolUseId,
@@ -574,13 +574,12 @@ const renderLine = (raw, deps) => {
             const rawIn = u.input_tokens ?? 0;
             const cr = u.cache_read_input_tokens ?? 0;
             const cw = u.cache_creation_input_tokens ?? 0;
-            // CodeBuddy (claude-internal) 网关把 input_tokens 报成 cr+cw+fresh 的总和,
+            // 网关把 input_tokens 报成 cr+cw+fresh 的总和 (claude-4.7-opus, deepseek-v4-flash),
             // 而 Anthropic 官方 input_tokens 只含 fresh (与 cache_creation/cache_read disjoint)。
-            // 按模型命名风格判定源头: CodeBuddy 是 "claude-4.7-opus" (点号版本),
-            // Anthropic 官方是 "claude-opus-4-7" (纯连字符)。只对 CodeBuddy 风格反推,
-            // 普通 Anthropic-native 会话保持原样, 不影响其 usage 计算。
-            const isGatewayTotalized = typeof model === "string" && /\d\.\d/.test(model);
-            const input = isGatewayTotalized && rawIn >= cr + cw ? rawIn - cr - cw : rawIn;
+            // 判据用数据本身: input_tokens 是否已覆盖两个 cache 档 —— 覆盖即网关口径, 反推 fresh;
+            // 否则按原生口径 (input 即 fresh) 原样保留。不按模型名风格猜, deepseek 无点号版本。
+            const isTotalized = rawIn >= cr + cw;
+            const input = isTotalized ? rawIn - cr - cw : rawIn;
             out.push({
                 kind: "turn_usage",
                 model,
@@ -1289,6 +1288,8 @@ export const startMirror = (deps) => {
     // for >60s mid-turn and we don't want to drop the bubble while it's chewing.
     const FLUSH_MS = 250;
     const HARD_TIMEOUT_MS = 350_000;
+    /** 首条 CLI 产出到达前的早期超时: 3s 无响应则先收气泡为详情链接。 */
+    const EARLY_LINK_MS = 3_000;
     /** 软收口静默期。后端只能说"这条消息写完了"(codebuddy) 时, 等这么久没有新 item
      *  才认定一轮结束。取值只需盖住"叙述消息落盘 → 紧随其后的 function_call 落盘"
      *  这一段, 与模型思考/工具执行时长无关。 */
@@ -1647,7 +1648,8 @@ export const startMirror = (deps) => {
     const BRIEF_TURN_OPENERS = new Set(["text", "tool_use", "tool_result", "skill_output"]);
     // 链接落到 chat 视图: 默认选中本 turn 所属的 #tag, 贴底显示整条会话。turnId 依旧
     // 是凭据 (不可枚举), 只是页面从"一个 turn"扩成"这个 chat 的全部会话"。
-    const briefDetailLink = (turnId) => `✴️ [View chat details](${buildChatUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, turnId)})`;
+    // target 作为 ww_uniq 传下去, 让同 chat 的所有 turn 详情都复用一个 WeCom 窗口。
+    const briefDetailLink = (turnId, target) => `✴️ [View chat details](${buildChatUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, turnId, stripPrincipalPrefix(target))})`;
     // 收口一条 loading 气泡: finish=true 写入最终内容, 只生效一次。发送失败退回 standalone。
     // 排队中的 turn 也持有气泡, 所以气泡是显式入参而不是从 a 上取。
     const finishBubble = async (a, b, content) => {
@@ -1655,6 +1657,10 @@ export const startMirror = (deps) => {
             return;
         b.done = true;
         clearTimeout(b.hardTimer);
+        if (b.earlyTimer) {
+            clearTimeout(b.earlyTimer);
+            b.earlyTimer = undefined;
+        }
         if (a.briefBubble === b)
             a.briefBubble = undefined;
         try {
@@ -1700,8 +1706,16 @@ export const startMirror = (deps) => {
                 log.warn({ sessionId: a.sessionId, turnId }, "brief: queued turn timed out — force-closing the stuck one");
                 closeBriefTurn(a);
             }
-            void finishBubble(a, bubble, briefDetailLink(turnId));
+            void finishBubble(a, bubble, briefDetailLink(turnId, a.target));
         }, HARD_TIMEOUT_MS);
+        // earlyTimer: 3s 无 CLI 产出 → 先把 loading 气泡收成详情链接, 用户可即时点入。
+        // 首条 item 到达时清除 (handleBriefItem 路径)。用显式 bubble ref 而非 a.briefBubble,
+        // 因为排队中的 turn 还没成为活跃 turn, a.briefBubble 指向的是前一个。
+        bubble.earlyTimer = setTimeout(() => {
+            if (bubble.done)
+                return;
+            void finishBubble(a, bubble, briefDetailLink(turnId, a.target));
+        }, EARLY_LINK_MS);
         if (a.briefTurnId) {
             (a.briefQueue ??= []).push(q);
             log.info({ sessionId: a.sessionId, turnId, queued: a.briefQueue.length }, "brief: turn queued (previous still running)");
@@ -1733,7 +1747,7 @@ export const startMirror = (deps) => {
         a.briefHadTool = false;
         a.briefConcluded = false;
         a.briefLastText = undefined;
-        enqueueStandalone(a, briefDetailLink(turnId));
+        enqueueStandalone(a, briefDetailLink(turnId, a.target));
         log.info({ sessionId: a.sessionId, turnId }, "brief: turn started (CLI-side, no bubble)");
     };
     // 本轮一旦出现工具调用 / tool_result / 非 final 文本, 就说明这不是"一句话答完"的
@@ -1744,7 +1758,7 @@ export const startMirror = (deps) => {
         if (!a.briefTurnId || !a.briefBubble || a.briefBubble.done)
             return;
         a.briefHadTool = true;
-        void finishBriefBubble(a, briefDetailLink(a.briefTurnId));
+        void finishBriefBubble(a, briefDetailLink(a.briefTurnId, a.target));
     };
     // 本轮结论落地, 只生效一次:
     //   • 无工具 → 结论 + 详情链接一并写进 loading 气泡 (仍是一条消息);
@@ -1758,7 +1772,7 @@ export const startMirror = (deps) => {
         if (a.briefHadTool || !a.briefBubble)
             sendStandalone(a, body);
         else
-            void finishBriefBubble(a, `${body}\n\n${briefDetailLink(turnId)}`);
+            void finishBriefBubble(a, `${body}\n\n${briefDetailLink(turnId, a.target)}`);
         // 排队中的 turn 已经建了记录但还没跑, 不能被这把扫帚扫成「已完成」。
         const keep = [turnId, ...(a.briefQueue ?? []).map((q) => q.turnId)];
         recordCloseOpenTurns({ target: a.target, sessionId: a.sessionId, exceptIds: keep });
@@ -1773,7 +1787,7 @@ export const startMirror = (deps) => {
         // 气泡还开着 (有工具的 turn: 结论只发了 standalone, 气泡留到这里收链接; 或空 turn
         // 没有结论) —— 收口: 统一写详情链接, 保证每个 turn 都有可点入口。
         if (a.briefBubble && !a.briefBubble.done) {
-            void finishBriefBubble(a, briefDetailLink(a.briefTurnId));
+            void finishBriefBubble(a, briefDetailLink(a.briefTurnId, a.target));
         }
         recordTurnClose(a.briefTurnId);
         log.info({ sessionId: a.sessionId, turnId: a.briefTurnId }, "brief: turn closed");
@@ -1793,6 +1807,11 @@ export const startMirror = (deps) => {
         const turnId = a.briefTurnId;
         if (!turnId)
             return;
+        // 首条 item 到达 → 清除 earlyTimer (CLI 已有响应, 不需要超时收链接了)。
+        if (a.briefBubble?.earlyTimer) {
+            clearTimeout(a.briefBubble.earlyTimer);
+            a.briefBubble.earlyTimer = undefined;
+        }
         const now = Date.now();
         if (item.kind === "tool_use") {
             a.briefHadTool = true; // 有工具 → 气泡收口写详情链接, 结论另发 standalone
@@ -1922,6 +1941,9 @@ export const startMirror = (deps) => {
         a.keepaliveQuiet = undefined;
     };
     const onItem = (a, item) => {
+        // 新 spawn 的 pane 在首次 inject 落地前吞掉所有初始输出 (greeting/system)。
+        if (a.muteUntilInject)
+            return;
         // Keepalive ping turns are cache-warmers: swallow every item from the WeCom
         // paths so the ping/pong never reaches chat. But record the REAL exchange
         // into its chat-detail turn — the actual assistant reply (expected: just
@@ -2216,8 +2238,10 @@ export const startMirror = (deps) => {
     // Click-to-detail URL for a tool_use id. Cached on the bridge so both
     // attach() and migrateAttachment() share the same closure. Returns ""
     // when disabled in config — renderLine then drops the markdown wrapping.
-    const detailUrlFor = (id) => cfg.daemon.detailLinksInMirror && id
-        ? buildDetailUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, id)
+    // `principal` flows through to ww_uniq so all detail links from the same
+    // chat reuse one WeCom inner-browser window.
+    const detailUrlFor = (id, principal) => cfg.daemon.detailLinksInMirror && id
+        ? buildDetailUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, id, principal ? stripPrincipalPrefix(principal) : undefined)
         : "";
     const attach = ({ sessionId, jsonlPath, target: targetOverride, tmuxPane, tmuxSession, cwd, pendingCwd }) => {
         const target = resolveTarget(targetOverride);
@@ -2601,8 +2625,16 @@ export const startMirror = (deps) => {
         // 迁移是所有会话轮换的唯一漏斗 (注入的 /clear、TUI 里手打的 /clear、resume fork、
         // worktree drift)。只有目标 jsonl 首条用户行是 /clear 的那种才是真清空 —— fork /
         // drift 上下文是延续的, 让 store 去推它那条中性的 "switch"。
-        if (jsonlIsPostClearChild(newJsonlPath))
+        if (jsonlIsPostClearChild(newJsonlPath)) {
             a.ctxCut = "clear";
+            // 真清空 = 缓存里已无任何值得保温的内容。像 /stop 一样暂停保活 —— 与
+            // dispatch 里 WeCom 注入 /clear 的暂停对齐;TUI 手打 /clear 只走这个漏斗,
+            // 不在这里停下的话旧时钟会继续 ping 一个空上下文。busy-resume(过 grace)
+            // 或真实 inbound 解除。下方 store.set 一并落盘。
+            a.keepaliveOff = true;
+            a.keepaliveOffAt = Date.now();
+            a.keepalive = undefined;
+        }
         a.tail.stop();
         bySessionId.delete(oldSessionId);
         a.sessionId = newSessionId;
@@ -2634,6 +2666,10 @@ export const startMirror = (deps) => {
             tmuxPane: a.tmuxPane || undefined,
             cwd: a.runningCwd || undefined,
             pendingCwd: a.pendingCwd || undefined,
+            // store.set 是整记录替换 —— 必须带上暂停态,否则迁移会把 dispatch 刚落盘
+            // 的 /clear 暂停(或之前的 /stop 暂停)从盘上抹掉,reload 后保活复活。
+            keepaliveOff: a.keepaliveOff || undefined,
+            keepaliveOffAt: a.keepaliveOffAt || undefined,
         });
         log.info({ target: a.target, oldSessionId, newSessionId, oldJsonlPath, newJsonlPath, startOffset }, "mirror migrated session");
     };
@@ -2994,6 +3030,7 @@ export const startMirror = (deps) => {
         const spawned = byTarget.get(target);
         if (spawned) {
             spawned.justSpawned = true;
+            spawned.muteUntilInject = true;
             // 只有换过 pane 才是断点 —— 首次 /wrc 建会话时 prev 为空, 那不是"不连续", 是开局。
             if (prev)
                 spawned.ctxCut = "new";
@@ -3201,11 +3238,17 @@ export const startMirror = (deps) => {
     };
     const fireKeepalive = async (a) => {
         const kc = cfg.wrc.mirror.keepalive;
+        // Only the streak's FIRST ping carries the full instruction; consecutive
+        // pings shrink to a bare "ping" — the instruction is already in context,
+        // and the smaller delta means a smaller cache-write and less pane clutter.
+        // keepaliveStamps matches both forms, so the bare one never reads as real
+        // activity. (round is incremented below, after the inject lands.)
+        const text = (a.keepalive?.round ?? 0) > 0 ? "ping" : kc.ping;
         // Suppress the pane→WeCom echo of the ping user line, then swallow the whole
         // ping turn (reply included). The 60s fail-safe clears the quiet window if
         // the turn somehow never emits turn_end, so a later real turn is never muted;
         // it also closes the detail turn so it can't hang open in the chat timeline.
-        rememberInject(kc.ping);
+        rememberInject(text);
         a.keepaliveQuiet = setTimeout(() => {
             a.keepaliveQuiet = undefined;
             if (a.keepaliveTurnId) {
@@ -3214,7 +3257,7 @@ export const startMirror = (deps) => {
             }
         }, 60_000);
         const r = await inject({
-            text: kc.ping, images: [], cfg,
+            text, images: [], cfg,
             log: log.child({ target: a.target, sessionId: a.sessionId, sub: "keepalive" }),
             sessionId: a.sessionId, jsonlPath: a.jsonlPath, tmuxTarget: a.tmuxPane,
         });
@@ -3227,11 +3270,7 @@ export const startMirror = (deps) => {
         }
         const tokens = lastContextTokens(a.jsonlPath);
         const size = tokens > 0 ? `~${fmtTokens(tokens)} tokens` : "未知";
-        // Denominator tracks the ACTUAL ping cadence (idleTriggerMs = ttlSec - marginSec),
-        // not the nominal ttlSec — pings fire marginSec early, so ttlSec would undercount
-        // and let `n` exceed `N` (a 9/8). floor keeps the last shown round the last one that fires.
-        const cadenceSec = Math.max(30, kc.ttlSec - kc.marginSec);
-        const totalRounds = Math.max(1, Math.floor(kc.maxIdleSec / cadenceSec));
+        const totalRounds = Math.max(1, kc.rounds);
         const round = a.keepalive ? (a.keepalive.round += 1) : 1;
         log.info({ target: a.target, tokens, round, totalRounds }, "keepalive: ping injected");
         // Open a chat-detail turn for the real heartbeat exchange: userQuery is the
@@ -3240,7 +3279,7 @@ export const startMirror = (deps) => {
         // closed on the ping's turn_end (or the fail-safe). Kept out of chat.
         const turnId = newTurnId();
         a.keepaliveTurnId = turnId;
-        recordTurnStart({ id: turnId, target: a.target, sessionId: a.sessionId, userQuery: kc.ping });
+        recordTurnStart({ id: turnId, target: a.target, sessionId: a.sessionId, userQuery: text });
         if (kc.notify)
             sendStandalone(a, `KeepAlive ${round}/${totalRounds} · ${size}`);
     };
@@ -3253,7 +3292,6 @@ export const startMirror = (deps) => {
         try {
             const idleTriggerMs = Math.max(30, kc.ttlSec - kc.marginSec) * 1000;
             const ttlMs = kc.ttlSec * 1000;
-            const maxIdleMs = Math.max(kc.ttlSec, kc.maxIdleSec) * 1000;
             const pingSig = normAssistant(kc.ping).slice(0, 40);
             const now = Date.now();
             for (const a of byTarget.values()) {
@@ -3322,13 +3360,12 @@ export const startMirror = (deps) => {
                 if (a.keepaliveOff)
                     continue; // paused by /stop until real activity returns
                 const idleSinceTouch = k.lastMs ? now - k.lastMs : now - mtime; // last model turn = cache touch
-                const realIdle = k.lastRealMs ? now - k.lastRealMs : Infinity; // no real turn in window ⇒ dead
                 if (idleSinceTouch < idleTriggerMs)
                     continue; // cache still comfortably warm
                 if (idleSinceTouch >= ttlMs)
                     continue; // cache already cold — a ping would cold-rewrite for nothing
-                if (realIdle >= maxIdleMs)
-                    continue; // real work too old — stop, let it die (你的「超5分钟不保活」)
+                if (k.round >= kc.rounds)
+                    continue; // budget spent — let it go cold
                 k.pinging = true;
                 k.pingMtime = k.lastMs; // settles when a newer turn (the ping's own) appears
                 await fireKeepalive(a);
@@ -3823,6 +3860,7 @@ export const startMirror = (deps) => {
                     sessionId: sid, jsonlPath: a.jsonlPath, tmuxTarget: a.tmuxPane, freshSpawn,
                 });
                 if (!r.ok) {
+                    a.muteUntilInject = false;
                     if (s) {
                         s.acc = s.acc ? `${s.acc}\n\n[mirror] ✗ ${r.reason ?? "failed"}` : `[mirror] ✗ ${r.reason ?? "failed"}`;
                         await finalizeStream(a, s);
@@ -3846,6 +3884,8 @@ export const startMirror = (deps) => {
                     }
                     return;
                 }
+                // inject 成功 → 解除初始输出静默, 后续 onItem 正常流转。
+                a.muteUntilInject = false;
                 // Inject "succeeded" but the input box never cleared — the target
                 // session may be busy / not consuming input (long task, full context).
                 // Hint the user once so a silently-dropped message isn't mistaken for a