wezard 2.4.0 → 2.5.0

package/dist/daemon/mirror-bridge.js

@@ -19,7 +19,7 @@ import { join, dirname } from "node:path";
 import { expandHome, sanitizeId } from "../shared/paths.js";
 import { augmentedPath } from "../shared/exec-path.js";
 import { sleep, truncateWithCount, mapLimit } from "../shared/std.js";
-import { activeBackends, backendForPath, primaryBackend, projectDirFor, projectDirsFor, } from "../shared/cli-backends.js";
+import { activeBackends, backendForPath, projectDirFor, projectDirsFor, } from "../shared/cli-backends.js";
 import { isModalPane, isAskqSubmitPage, parseModalOptions, pickModalAnswer, pickAutoAllowAnswer, parseConfirmContext, looksLikePicker } from "../shared/modal-pane.js";
 import { bindCardChannel, hasMirrorAskq, runMirrorAskqFlow, hasMirrorPlan, mootMirrorPlan, runMirrorPlanFlow, hasMirrorPicker, mootMirrorPicker, runMirrorPickerFlow } from "./approval.js";
 import { isAutoWindowActive } from "./session-cache.js";
@@ -34,10 +34,11 @@ import { wizardStore } from "./wizard.js";
 import { startSubagentWatch } from "./subagent-tail.js";
 import { knowsToolUse, recordTool, recordToolResult, recordMark, recordTurnStart, recordTurnQuery, recordTurnInject, recordTurnItem, recordTurnUsage, recordTurnClose, recordCloseOpenTurns, lastChannelOf, openTurnsOf, buildDetailUrl, buildChatUrl, roleUniq } from "./detail.js";
 import { errText } from "./last-response.js";
-import { labelFor, tagOfKey, baseOfKey, keyOf, stripSigil, displayName, withTagHeader, withLinkedTagHeader, linkedTagHead, linkTags, parseTagHeader, MAX_BODY_LINKS, isInternalKey } from "../shared/session-label.js";
+import { labelFor, tagOfKey, baseOfKey, keyOf, stripSigil, displayName, withTagHeader, withLinkedTagHeader, chatHead, headed, tagBadge, linkTags, parseTagHeader, MAX_BODY_LINKS, isInternalKey } from "../shared/session-label.js";
 import { splitMarkdown } from "../shared/md-chunk.js";
+import { isWxFrame } from "../shared/wx-text.js";
 import { randomTip } from "./tips.js";
-import { chatBaseOf, chatNameOf, listChatNames, parsePeerRef, peerAddress } from "./chat-name.js";
+import { chatBaseOf, chatNameOf, listChatNames, parsePeerRef, peerAddress, stewardTierOf } from "./chat-name.js";
 import { stripAnsi, paneIsBusy, summarizeTail, lastAssistantText, lastReply, replyToPeer as replyToPeerIn, unwrapPasted, lastContextTokens, lastModel, cacheTouchMs, cacheTtlSec, keepaliveStamps, openKeepalivePing, lastUserTurn, talkRounds, openToolUses, renderDialog } from "./peers.js";
 import { keepalivePingSigs, isKeepalivePingText } from "../shared/keepalive.js";
 // PATH augmentation: launchd / systemd start the daemon
@@ -1393,10 +1394,10 @@ const injectViaTmux = async (target, text, images, log, freshSpawn, backend, jso
     //     所以 codebuddy 走 C-v 必然失败 —— 改走 @<path> 文本提及，让 LLM
     //     调 Read 工具读图（Read 支持图片，见 codebuddy tools-reference.md）。
     //     v2.52.4 之后 @<path> 不自动转 image block，但 LLM 看到路径会 Read。
+    // dispatch 已对 codebuddy 把 refs 合进 text (images 传空), 这里的
+    // withImageRefs 是其它调用方直传 images 时的兜底 —— 幂等, 不会贴两份。
     if (backendName === "codebuddy") {
-        const refs = images.map((p) => `@${p}`);
-        const textWithRefs = refs.length ? (text ? `${refs.join("\n")}\n${text}` : refs.join("\n")) : text;
-        return injectViaTmuxText(target, textWithRefs, log, freshSpawn, witness);
+        return injectViaTmuxText(target, withImageRefs(text, images), log, freshSpawn, witness);
     }
     // Pump images first via clipboard+C-v so each one is attached as a separate
     // image content block. Each C-v needs a brief settle for Claude Code's TUI
@@ -1424,6 +1425,13 @@ const injectViaTmux = async (target, text, images, log, freshSpawn, backend, jso
     }
     return injectViaTmuxText(target, text, log, freshSpawn, witness);
 };
+// 图片以 @<path> 文本提及的形式进会话 (codebuddy 的 tmux 路径、一切 spawn 路径)。
+// 提升为共享 helper: dispatch 侧的印章 / 回显去重 / turn 记录必须拿到与 inject
+// 完全相同的字符串, 否则落盘行认不出是自己注入的。
+const withImageRefs = (text, images) => {
+    const refs = images.map((p) => `@${p}`).join("\n");
+    return refs ? (text ? `${refs}\n${text}` : refs) : text;
+};
 // 文本 paste + Enter 提交 + 自校验。从 injectViaTmux 抽出来，让 codebuddy
 // 后端的 @<path> 回退路径也能复用同样的 paste-verify 逻辑。
 //
@@ -1626,8 +1634,7 @@ const injectViaSpawn = (args) => {
     // Spawn-mode (no live TTY) can't do clipboard+C-v — fall back to `@<path>`,
     // which Claude parses at submit time and inlines as image content blocks
     // without a model-decided Read tool turn.
-    const refs = images.map((p) => `@${p}`).join("\n");
-    const finalText = refs ? (text ? `${refs}\n${text}` : refs) : text;
+    const finalText = withImageRefs(text, images);
     const cliArgs = [
         "-p",
         finalText,
@@ -1838,8 +1845,8 @@ export const startMirror = (deps) => {
     const detailCardFor = (s, target) => {
         if (s.tools.length === 0)
             return undefined;
-        const name = displayName(target);
-        const titlePrefix = name ? `${labelFor(name)} .${name} · ` : "";
+        const badge = tagBadge(target).trim();
+        const titlePrefix = badge ? `${badge} · ` : "";
         return {
             card_type: "button_interaction",
             main_title: { title: `${titlePrefix}本轮工具调用` },
@@ -1855,7 +1862,7 @@ export const startMirror = (deps) => {
         const content = s.acc;
         s.lastSent = content;
         try {
-            await client.replyStream(s.frame, s.streamId, withLinkedTagHeader(s.target, content || " ", chatTurnUrl(s.target, s.turnId)), false);
+            await client.replyStream(s.frame, s.streamId, withLinkedTagHeader(s.target, content || "…", chatTurnUrl(s.target, s.turnId), undefined, undefined, s.dest), false);
             log.debug({ turnId: s.turnId, len: content.length }, "stream flush ok");
         }
         catch (e) {
@@ -1929,11 +1936,14 @@ export const startMirror = (deps) => {
             turnId: newTurnId(),
             frame, streamId,
             target: a.target,
+            ...(a.channel ? { dest: a.channel } : {}),
             acc: "", lastSent: "",
             capped: false, closed: false, dead: false, cardSent: false,
             tools: [], sawTool: false,
         };
-        s.hardTimer = setTimeout(() => void finalizeStream(a, s, true), HARD_TIMEOUT_MS);
+        // 微信帧没有 stream 窗口: 到点收口只会多发一条「仍在处理中」, 白占它那几格出站额度。
+        if (!isWxFrame(frame))
+            s.hardTimer = setTimeout(() => void finalizeStream(a, s, true), HARD_TIMEOUT_MS);
         log.info({ sessionId: a.sessionId, turnId: s.turnId, streamId }, "stream open");
         return s;
     };
@@ -1945,8 +1955,8 @@ export const startMirror = (deps) => {
         ? buildChatUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, turnId, roleUniq(target))
         : undefined;
     // `[mirror]` / `[wezard]` 系统提示不是答给谁的话, 不标方向。
-    const withLinkedTag = (a, content, seq, turnId) => withLinkedTagHeader(a.target, content, chatTurnUrl(a.target, turnId ?? a.briefTurnId), seq, /^\[(mirror|wezard)\]/.test(content.trim()) ? undefined : a.replyTo);
-    // 空正文一票否决 (近源拦截): 剥掉可能存在的路由头 (`🦊 #tag` / `[🧙](url)`)
+    const withLinkedTag = (a, content, seq, turnId) => withLinkedTagHeader(a.target, content, chatTurnUrl(a.target, turnId ?? a.briefTurnId), seq, /^\[(mirror|wezard)\]/.test(content.trim()) ? undefined : a.replyTo, a.channel || undefined);
+    // 空正文一票否决 (近源拦截): 剥掉可能存在的路由头 (`.name:` / `[.name](url):`)
     // 后没有可见内容就不发。中央 chat-gate (last-response 的 SDK 包装) 是最后一道
     // 防线; 这里拦在源头 —— 空内容不进 standalonePending FIFO、不占防抖 buf、
     // 不重置计时器, 免得"空 part 入队 → flush 时 join 出空串"的死角。
@@ -2282,7 +2292,7 @@ export const startMirror = (deps) => {
             return;
         a.cotLastSent = text;
         try {
-            await client.replyStream(b.frame, b.streamId, `${briefDetailLink(turnId, a.target)} \`${text}\``, false);
+            await client.replyStream(b.frame, b.streamId, briefLine(a, turnId, `\`${text}\``), false);
         }
         catch (e) {
             log.debug({ sessionId: a.sessionId, turnId, err: e.message }, "brief: cot refresh failed");
@@ -2304,7 +2314,8 @@ export const startMirror = (deps) => {
     // 链接落到 chat 视图: 默认选中本 turn 所属的 #tag, 贴底显示整条会话。turnId 依旧
     // 是凭据 (不可枚举), 只是页面从"一个 turn"扩成"这个 chat 的全部会话"。
     // wizard 名字作为 ww_uniq 传下去, 让同一个 wizard 的所有 turn 详情都复用一个 WeCom 窗口。
-    const briefDetailLink = (turnId, target) => linkedTagHead(target, buildChatUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, turnId, roleUniq(target)));
+    // 头怎么写归 chatHead (默认 wizard 在 home 群: 企微只留名字链接, 微信不写头); 头与正文都空时留 `…`, 不发空气泡。
+    const briefLine = (a, turnId, rest, dest = a.channel) => headed(chatHead(a.target, buildChatUrl(cfg.daemon.detailPublicBase, cfg.daemon.host, cfg.daemon.port, turnId, roleUniq(a.target)), undefined, dest || undefined), rest) || "…";
     // ── 出处门 (chatOriginOnly) ───────────────────────────────────────────
     // 人在 CLI 里手敲的一轮, 镜像只发生在两处: 他眼前的终端, 和 chat 详情页 (turn
     // store 全量记录 + SSE 实时刷新)。聊天里不再下发它的气泡 —— 那份内容他本来就
@@ -2492,7 +2503,7 @@ export const startMirror = (deps) => {
     };
     /** 人那句被并进了一轮私聊: 它预建的气泡与轮次收口成一个链接 —— 答复不进群, 也不让气泡挂到超时。 */
     const mergedAway = (a, q) => {
-        void finishBubble(a, q.bubble, `${briefDetailLink(q.turnId, a.target)} （这句和一段私聊并成了一轮, 答复只在 rolepage）`, true);
+        void finishBubble(a, q.bubble, briefLine(a, q.turnId, "（这句和一段私聊并成了一轮, 答复只在 rolepage）"), true);
         recordTurnClose(q.turnId);
     };
     /** 人在聊天里说的那一句接成活跃 turn: 收掉在跑的那一轮, 频道与发话人取它自己的印章。
@@ -2556,15 +2567,17 @@ export const startMirror = (deps) => {
     const startBriefTurn = async (a, frame, streamId, isSlash = false, userQuery = "", channel, speaker) => {
         const turnId = newTurnId();
         // hardTimer 兜底: turn 若无终句 / turn_end 收口 (卡死/漏收), 到点仍收气泡。
-        const bubble = { frame, streamId, hardTimer: undefined, done: false };
+        const bubble = { frame, streamId, done: false };
         const q = { turnId, bubble, isSlash };
-        bubble.hardTimer = setTimeout(() => {
-            // WeCom ~6min stream 窗口快到, 必须 finish=true 收口。finish 是整条替换 ——
-            // 定格在一行过期的 CoT 进度上看着像卡死, 光链接又像说完了; 气泡仍是本轮活跃
-            // 气泡 (turn 还在跑) 就明说「仍在处理中」, 已被顶替的旧轮只留链接。
-            const live = a.briefBubble === bubble && !a.briefConcluded;
-            void finishBubble(a, bubble, `${briefDetailLink(turnId, a.target)}${live ? ` ${STILL_WORKING}` : ""}`, true);
-        }, HARD_TIMEOUT_MS);
+        // 微信帧不挂: 没有 stream 窗口要赶, 气泡等终句来了再一次发出 (见 openStream 同一条)。
+        if (!isWxFrame(frame))
+            bubble.hardTimer = setTimeout(() => {
+                // WeCom ~6min stream 窗口快到, 必须 finish=true 收口。finish 是整条替换 ——
+                // 定格在一行过期的 CoT 进度上看着像卡死, 光链接又像说完了; 气泡仍是本轮活跃
+                // 气泡 (turn 还在跑) 就明说「仍在处理中」, 已被顶替的旧轮只留链接。
+                const live = a.briefBubble === bubble && !a.briefConcluded;
+                void finishBubble(a, bubble, briefLine(a, turnId, live ? STILL_WORKING : "", channel), true);
+            }, HARD_TIMEOUT_MS);
         const seal = { keys: sealKeys(userQuery), at: Date.now(), query: userQuery, channel, speaker, fresh: true, turn: q };
         // 对方闲着 → 这一句就是下一轮, 当场接上。正忙 (在跑别人的回执 / 派活 / 上一句) 就先排着:
         // 那一轮的回复不能流进这一句的气泡, 等它在 transcript 里露面再接 (bindLine)。turn 记录
@@ -2587,7 +2600,7 @@ export const startMirror = (deps) => {
         // 也不必猜后端: codebuddy 那种几十秒后才落盘的轮, 入口从第一秒就在。
         // 尾随 `…` 表示"正文还没到", 正文到位时整条内容被 `链接 正文` 覆盖。
         try {
-            await client.replyStream(frame, streamId, `${briefDetailLink(turnId, a.target)} …`, false);
+            await client.replyStream(frame, streamId, briefLine(a, turnId, "…", channel), false);
         }
         catch (e) {
             log.warn({ sessionId: a.sessionId, turnId, err: e.message }, "brief: turn ack failed");
@@ -2698,7 +2711,7 @@ export const startMirror = (deps) => {
         a.lastConcludedEpoch = a.queryEpoch ?? 0;
         a.lastConcludedAt = Date.now();
         if (a.briefBubble && !a.briefBubble.done) {
-            void finishBriefBubble(a, `${briefDetailLink(turnId, a.target)} ${linkTags(a.target, body)}`, true);
+            void finishBriefBubble(a, briefLine(a, turnId, linkTags(a.target, body)), true);
         }
         else if (turnSilent(a)) {
             // CLI 手敲的一轮: 终稿只留在 turn store, 详情页照常实时可见。
@@ -3364,7 +3377,7 @@ export const startMirror = (deps) => {
                 // codebuddy 对 AskUserQuestion 不在提问时触发 PreToolUse hook — 先弹本地
                 // 面板, hook 只在面板被提交后才到达 (实测可延迟数小时)。这里从 jsonl 提前
                 // 看到 function_call, 直接驱动 vote 卡让远端可答; 点选后注入一段触发文本
-                // 提交本地面板, hook 到达时以 deny+reason 覆盖答案, 与 claude 路径同产物。
+                // 提交本地面板, hook 到达时以 deny+reason 覆盖答案 (claude 路径改走 allow+updatedInput)。
                 // claude 系 hook 即时触发, 不走此路 (否则双重发卡)。
                 if (c.name === "AskUserQuestion"
                     && backendForPath(a.jsonlPath).name === "codebuddy"
@@ -4455,32 +4468,14 @@ export const startMirror = (deps) => {
             return undefined;
         }
     };
-    // 确切模型名的证据 (见 exactModelId): transcript 里 API 真答过的 model id, 只认同一个
-    // CLI 的。先问最可能的那几个会话 (自己、被克隆者), 再按新近扫其余 —— 惰性的,
-    // 第一个对上就停, 绝大多数时候只读一两个 transcript 的尾巴。
-    function* modelSightings(cli, likely) {
-        const name = cli ?? primaryBackend().name;
-        const pathOf = (t) => byTarget.get(t)?.jsonlPath ?? expandHome(deps.store.get(t)?.jsonlPath ?? "");
-        const head = likely.map(pathOf).filter(Boolean);
-        // 生成器体到第一次取值才跑: 前几个就对上时, 其余几百个 stat 一个也不做。
-        const paths = function* () {
-            yield* head;
-            yield* Object.values(deps.store.all())
-                .map((r) => expandHome(r.jsonlPath))
-                .filter((p) => !head.includes(p))
-                .map((p) => ({ p, m: mtimeOf(p) }))
-                .filter((x) => x.m > 0)
-                .sort((x, y) => y.m - x.m)
-                .map((x) => x.p);
-        };
-        for (const path of paths()) {
-            if (!existsSync(path) || backendForPath(path).name !== name)
-                continue;
-            const id = lastModel(path);
-            if (id)
-                yield id;
-        }
-    }
+    /** 管家该跑的那一档 (模型 + effort); 不是管家 = undefined。 */
+    const stewardSpec = (target) => {
+        const tier = stewardTierOf(cfg, target);
+        return tier ? { model: cfg.models.tiers[tier].model || undefined, effort: cfg.models.tiers[tier].effort } : undefined;
+    };
+    /** 死 pane 重生 (`--resume` 同一个会话) 的模型: 绑定里记着的优先 —— set_model 换上的也记在那;
+     *  没记 (老绑定 / CLI 默认) 的管家补上管家档。 */
+    const respawnSpec = (a) => a.model ? { model: a.model, effort: parseEffort(a.effort) } : stewardSpec(a.target) ?? { effort: parseEffort(a.effort) };
     const newSession = async (target, windowName, cli, opts) => {
         const prev = byTarget.get(target);
         // Resolution precedence (all chat-scoped except the running-cwd fallback):
@@ -4512,11 +4507,12 @@ export const startMirror = (deps) => {
         const baseBound = base === target ? undefined : byTarget.get(base)?.jsonlPath ?? deps.store.get(base)?.jsonlPath;
         const boundPath = prev?.jsonlPath ?? rec?.jsonlPath ?? baseBound;
         const effCli = cli ?? (boundPath ? backendForPath(expandHome(boundPath)).name : undefined);
-        // 群的默认会话 (管家) 没点名模型 / 档位时用 models.router。只管这里新起的 —— 死 pane
-        // 重生走绑定里记下的模型, 不经过这里; 已在跑的管家也不动 (换模型会让缓存整份重读)。
-        const router = !tagOfKey(target) && !isInternalKey(target) ? cfg.models.router : undefined;
-        const effModel = opts?.model?.trim() || router?.model || undefined;
-        const effEffort = opts?.effort ?? parseEffort(prev?.effort ?? deps.store.get(target)?.effort) ?? router?.effort;
+        // 管家 (群的默认会话) 新起的会话按管家档起 (models.steward / chatPolicy), 档压过绑定里记下的
+        // effort —— `/new`、换目录、交接重开都是让改过的档位生效的那一下。点名了模型就整档让开。
+        // 已在跑的管家不动 (换模型会让缓存整份重读); 死 pane 重生见 respawnSpec。
+        const steward = opts?.model?.trim() ? undefined : stewardSpec(target);
+        const effModel = opts?.model?.trim() || steward?.model || undefined;
+        const effEffort = opts?.effort ?? steward?.effort ?? parseEffort(prev?.effort ?? deps.store.get(target)?.effort);
         if (prev?.tmuxPane) {
             // Best-effort kill; ignore errors (pane may already be dead).
             void runTmux(["kill-pane", "-t", prev.tmuxPane]);
@@ -4534,7 +4530,6 @@ export const startMirror = (deps) => {
             cwdOverride: eff,
             cli: effCli,
             model: effModel,
-            knownModels: modelSightings(effCli, [target, base]),
             effort: effEffort,
             systemPrompt: opts?.systemPrompt ?? charterFor(target, { cwd: eff }),
         }).catch((e) => { inFlightSids.delete(freshSid); throw e; });
@@ -4687,7 +4682,6 @@ export const startMirror = (deps) => {
             cwdOverride: parent.runningCwd,
             cli: parentCli,
             model,
-            knownModels: modelSightings(parentCli, [args.parent]),
             effort,
             systemPrompt: args.systemPrompt,
         });
@@ -4877,6 +4871,37 @@ export const startMirror = (deps) => {
         return eff === expandedDefaultCwd;
     };
     // ── Peer graph (sibling sessions of one chat) ────────────────────────
+    // 关一个会话的 pane。`drop` = /kill: 连绑定一起删, 下一条消息另起白板; 否则只是睡着 (同 pane cap 的 reap), 绑定留着、`--resume` 带全上下文复活。
+    const downPane = async (target, drop) => {
+        // paneOf (not byTarget.get) so a binding surviving only in the persisted
+        // store — e.g. after a reload that couldn't re-attach — is still killable.
+        const a = byTarget.get(target);
+        const pane = a?.tmuxPane || paneOf(target);
+        if (!a && !pane)
+            return { ok: false, reason: "no session bound to target" };
+        if (pane && (await tmuxPaneAlive(pane))) {
+            // Esc first, like /stop: a mid-generation CLI gets a beat to unwind and
+            // flush its transcript before the TTY is yanked out from under it.
+            await runTmux(["send-keys", "-t", pane, "Escape"]);
+            await sleep(250);
+            const r = await runTmux(["kill-pane", "-t", pane]);
+            if (r.code !== 0)
+                return { ok: false, reason: `kill-pane failed: ${r.stdout.slice(-200) || r.code}` };
+        }
+        if (a)
+            detach(a, drop ? "/kill" : "sleep");
+        if (!drop) {
+            log.info({ target, sessionId: a?.sessionId, pane }, "mirror sleep — pane killed, binding kept");
+            return { ok: true };
+        }
+        // Drop the persisted record too — keeping it would let the next inbound
+        // resurrect this very session via the dead-pane `--resume` self-heal,
+        // which is the opposite of what /kill means. The chat then auto-spawns a
+        // fresh session on its next message.
+        deps.store.drop(target);
+        log.info({ target, sessionId: a?.sessionId, pane }, "mirror /kill — pane killed, binding dropped");
+        return { ok: true };
+    };
     // A chat's sessions are exactly the keys sharing its base principal: the
     // untagged default plus every `#tag`. Live attachments are the truth; the
     // persisted store fills in cold bindings so a peer nobody has talked to since
@@ -5492,6 +5517,7 @@ export const startMirror = (deps) => {
         isBusy,
         idleNow,
         parkedNow,
+        idleVerdict,
         untilIdle,
         setModel: async (target, wanted, scope) => {
             const a = byTarget.get(target);
@@ -5735,13 +5761,17 @@ export const startMirror = (deps) => {
                 // Same resume-fork hazard as dispatch: snapshot before spawn, re-bind
                 // onto the forked jsonl once it appears (EOF offset — fork is seeded).
                 const resumeBaseline = listJsonls(dirname(a.jsonlPath));
-                const r = await spawnTmuxClaude({ cfg, log: log.child({ sub: "respawn-init", sessionId: sid }), resumeSessionId: sid, windowName: displayName(target) || target, cwdOverride: a.runningCwd, cli: backendForPath(a.jsonlPath).name, model: a.model || undefined, knownModels: modelSightings(backendForPath(a.jsonlPath).name, [target]), effort: parseEffort(a.effort), systemPrompt: charterFor(target, { cwd: a.runningCwd }) });
+                const r = await spawnTmuxClaude({ cfg, log: log.child({ sub: "respawn-init", sessionId: sid }), resumeSessionId: sid, windowName: displayName(target) || target, cwdOverride: a.runningCwd, cli: backendForPath(a.jsonlPath).name, ...respawnSpec(a), systemPrompt: charterFor(target, { cwd: a.runningCwd }) });
                 if (!r.ok || !r.tmuxPane)
                     return { ok: false, reason: `respawn failed: ${r.reason ?? "unknown"}` };
                 a.tmuxPane = r.tmuxPane;
                 a.tmuxSession = r.tmuxSession ?? a.tmuxSession;
                 if (r.cwd)
                     a.runningCwd = r.cwd;
+                // What the pane really landed on: a recorded string the CLI refused was
+                // relaunched past, and must not cost the same extra launch next time.
+                if (r.model !== undefined)
+                    a.model = r.model;
                 deps.store.set(target, { sessionId: sid, jsonlPath: a.jsonlPath, tmuxSession: a.tmuxSession, tmuxPane: a.tmuxPane, cwd: a.runningCwd || undefined, model: a.model || undefined, effort: a.effort || undefined, pendingCwd: a.pendingCwd || undefined });
                 startMigrationWatcher(a, resumeBaseline, forkOf(a.jsonlPath));
                 fresh = true;
@@ -5822,32 +5852,8 @@ export const startMirror = (deps) => {
                 return { ok: false, reason: escReason };
             return { ok: true, torndown, escOk, escReason: escReason || undefined };
         },
-        killPane: async (target) => {
-            // paneOf (not byTarget.get) so a binding surviving only in the persisted
-            // store — e.g. after a reload that couldn't re-attach — is still killable.
-            const a = byTarget.get(target);
-            const pane = a?.tmuxPane || paneOf(target);
-            if (!a && !pane)
-                return { ok: false, reason: "no session bound to target" };
-            if (pane && (await tmuxPaneAlive(pane))) {
-                // Esc first, like /stop: a mid-generation CLI gets a beat to unwind and
-                // flush its transcript before the TTY is yanked out from under it.
-                await runTmux(["send-keys", "-t", pane, "Escape"]);
-                await sleep(250);
-                const r = await runTmux(["kill-pane", "-t", pane]);
-                if (r.code !== 0)
-                    return { ok: false, reason: `kill-pane failed: ${r.stdout.slice(-200) || r.code}` };
-            }
-            if (a)
-                detach(a, "/kill");
-            // Drop the persisted record too — keeping it would let the next inbound
-            // resurrect this very session via the dead-pane `--resume` self-heal,
-            // which is the opposite of what /kill means. The chat then auto-spawns a
-            // fresh session on its next message.
-            deps.store.drop(target);
-            log.info({ target, sessionId: a?.sessionId, pane }, "mirror /kill — pane killed, binding dropped");
-            return { ok: true };
-        },
+        killPane: (target) => downPane(target, true),
+        sleepPane: (target) => downPane(target, false),
         lastActivity: (target) => {
             const jsonl = jsonlOf(target);
             try {
@@ -6135,6 +6141,16 @@ export const startMirror = (deps) => {
             // 任意 slash 命令 (/context, /cost, /status…): 产物走 skill_output standalone,
             // 详情链接是纯噪声 —— brief 起 turn 时不推链接。
             const isSlash = /^\/[a-z]/i.test(text.trim());
+            // codebuddy 没有可用的剪贴板图片通道 (tmux 里 C-v 会被它自己的
+            // syncTmuxToSystemClipboard 冲掉, 见 injectViaTmux 分流注释) —— 图片只能以
+            // @<path> 文本进会话。既然如此, 问话 / 印章 / 回显去重 / 静默分叉跟随都按这个
+            // 最终落盘形态记 (与 inject 拿到的同一个字符串); 否则落盘的 `@<path>` 行认不出
+            // 是自己注入的: 被当成 CLI 手敲另开幽灵轮、触发 chatOriginOnly 吞回复, 而 IM 侧
+            // 预建的 turn 永远等不到问话改写, rolepage 的入消息丢了图 (rolepage 一律按 CLI
+            // 最终展示的那条消息渲染)。slash / /clear 判定仍用原始 text —— 命令不带图。
+            const refsFirst = backendForPath(a.jsonlPath).name === "codebuddy" && (images?.length ?? 0) > 0;
+            const effText = refsFirst ? withImageRefs(text, images ?? []) : text;
+            const effImages = refsFirst ? [] : images;
             // Auto-upgrade /clear → /new when the user has queued a project switch:
             // a plain /clear would only rotate sessionId in the same pane, which sits
             // in the OLD cwd. Killing+respawning is the only way to honor the switch.
@@ -6217,7 +6233,7 @@ export const startMirror = (deps) => {
                 // 挂 loading 气泡并起新 turn —— 上一 turn 若还在跑, startBriefTurn 内部
                 // 会立刻把它收掉 (对话边界策略), 本轮直接成为活跃 turn。
                 // 后续 onItem 走 handleBriefItem, 不再走 stream / defer 路径。
-                await startBriefTurn(a, frame, streamId, isSlash, text, channel ?? baseOfKey(principal), speaker);
+                await startBriefTurn(a, frame, streamId, isSlash, effText, channel ?? baseOfKey(principal), speaker);
             }
             else if (!armMigration && !eagerOpen) {
                 enterDeferred(a, frame, streamId);
@@ -6252,13 +6268,15 @@ export const startMirror = (deps) => {
                     const resumeBaseline = !armMigration ? listJsonls(dirname(a.jsonlPath)) : undefined;
                     // Respawn in the binding's runningCwd (pendingCwd doesn't apply to a
                     // mid-turn reincarnation — only /new and /clear-with-pending swap cwd).
-                    const r = await spawnTmuxClaude({ cfg, log: log.child({ sub: "respawn", sessionId: sid }), resumeSessionId: sid, windowName: displayName(a.target) || a.target, cwdOverride: a.runningCwd, cli: backendForPath(a.jsonlPath).name, model: a.model || undefined, knownModels: modelSightings(backendForPath(a.jsonlPath).name, [a.target]), effort: parseEffort(a.effort), systemPrompt: charterFor(a.target, { cwd: a.runningCwd }) });
+                    const r = await spawnTmuxClaude({ cfg, log: log.child({ sub: "respawn", sessionId: sid }), resumeSessionId: sid, windowName: displayName(a.target) || a.target, cwdOverride: a.runningCwd, cli: backendForPath(a.jsonlPath).name, ...respawnSpec(a), systemPrompt: charterFor(a.target, { cwd: a.runningCwd }) });
                     if (r.ok && r.tmuxPane && r.tmuxSession) {
                         a.tmuxPane = r.tmuxPane;
                         a.tmuxSession = r.tmuxSession;
                         if (r.cwd)
                             a.runningCwd = r.cwd;
                         freshSpawn = true;
+                        if (r.model !== undefined)
+                            a.model = r.model; // 同 respawn-init: 记真落地的模型
                         deps.store.set(a.target, {
                             sessionId: sid,
                             jsonlPath: a.jsonlPath,
@@ -6302,7 +6320,7 @@ export const startMirror = (deps) => {
                     log.warn({ target: a.target, sessionId: sid, myGen, gen: a.injectGen }, "inject aborted — superseded by /stop or watchdog");
                     return;
                 }
-                rememberInject(text);
+                rememberInject(effText);
                 // 解除静默必须在 inject 之前: inject 内部的 freshSpawn 验证循环可达 10+s
                 // (paste-verify + settle + waitForCleared + retry), 而 LLM 的回复可能在 Enter
                 // 发出后 2-3s 就落盘 —— tail 在 mute 期间读到的 item 会被 onItem 永久丢弃,
@@ -6311,7 +6329,7 @@ export const startMirror = (deps) => {
                 // 用户行有 isOwnInject 过滤 —— 提前解除不会泄漏脏数据。
                 a.muteUntilInject = false;
                 const r = await inject({
-                    text, images, cfg, log: log.child({ principal, sessionId: sid }),
+                    text: effText, images: effImages, cfg, log: log.child({ principal, sessionId: sid }),
                     sessionId: sid, jsonlPath: a.jsonlPath, tmuxTarget: a.tmuxPane, freshSpawn,
                 });
                 if (!r.ok) {
@@ -6359,7 +6377,7 @@ export const startMirror = (deps) => {
                 // stuck on the dead old jsonl. /clear via WeCom (armMigration) has its
                 // own watcher; skip spawn-mode (no shared pane to fork under us).
                 if (!armMigration && a.tmuxPane)
-                    armSilentForkRebind(a, text);
+                    armSilentForkRebind(a, effText);
                 // Confirm the /model picker. Early Enter (before the picker renders)
                 // would leave it unconfirmed, so settle first; a late Enter is a
                 // harmless no-op in the empty input box. On confirm, claude writes