wezard 1.3.18 → 1.3.19

package/CLAUDE.md

@@ -54,7 +54,7 @@ Three processes, all coordinating via 127.0.0.1:17890 and `~/.wezard/`:
 - `approval.ts` builds the button card, parks an HTTP request via `pending.ts`, and resolves it on `event.template_card_event`. Three click outcomes: `allow` / `allow_window` (auto-allow N min for that session) / `deny`. `session-cache.ts` short-circuits repeats.
 - `inbound.ts` routes `message.text` / `message.image` / `message.mixed` → either headless `cc-bridge.ts` or `mirror-bridge.ts`. Gates on `allowFrom`; one-shot exception is the bootstrap claim (see below).
 - `cc-bridge.ts` (mode=headless) spawns `claude -p --output-format stream-json --verbose`, parses `assistant`/`stream_event` lines, and pumps text via `client.replyStream`. Per-principal queue serializes turns; session id is captured from the `system/init` line and persisted to `~/.wezard/sessions.json` for `--resume`.
-- `mirror-bridge.ts` (mode=mirror) tails `<projectsDir>/<encoded(cwd)>/<sid>.jsonl` and pushes new `assistant` content to a chat. Inbound injection is either `claude --resume <sid> -p` (writes a turn into the same jsonl) or `tmux paste-buffer` against a live TTY (only path that makes WeCom messages visible in the user's CLI window). `recentInjects` ring buffer suppresses the WeCom→CLI echo when tailing.
+- `mirror-bridge.ts` (mode=mirror) tails `<projectsDir>/<encoded(cwd)>/<sid>.jsonl` and pushes new `assistant` content to a chat — but only for turns the chat itself started: a turn opened by a human's CLI keystroke is recorded to the turn store (chat-detail page) and kept out of the chat (`wrc.mirror.chatOriginOnly`, brief mode only); approval/askq cards and `[mirror]` notices bypass that gate. Inbound injection is either `claude --resume <sid> -p` (writes a turn into the same jsonl) or `tmux paste-buffer` against a live TTY (only path that makes WeCom messages visible in the user's CLI window). `recentInjects` ring buffer suppresses the WeCom→CLI echo when tailing.
 - `claim.ts` is a single-shot bootstrap: an exact-match magic phrase from any sender bypasses `allowFrom`, writes that principal as `defaultChat`, and appends to `allowFrom`. This is the **only** legitimate ingress for first-time setup.
 - `peers.ts` is the read side of multi-agent awareness: sibling sessions of one chat are exactly the target keys sharing a base principal (`baseOfKey`), i.e. the untagged default plus every `#tag`. Pure parsing only — bounded transcript-tail reads (`tailTurns` / `lastAssistantText` / `summarizeTail`, shared with `session-scan.ts`) and pane classification (`paneIsBusy`). **Busy detection matches the TUI spinner footer** (`… (12m 31s · …)` on current versions, `esc to interrupt` on older ones) in the last 8 non-blank rows — transcript mtime is useless here, it goes quiet during long tool calls.
 - `graph.ts` is the loop-graph runner: `nodes` = tagged sessions (own cli/model/cwd), `steps` = an ordered pipeline walked `rounds` times, each step injecting a prompt, waiting for that agent to go idle, then feeding its reply forward as `{{last}}` / `{{<tag>}}`. Stops early on the `until` sentinel. All effects go through injected deps backed by `mirror-bridge`, so a step is exactly what a human typing `#tag …` would do. Runs are **in-memory only** — a reload cancels them (panes survive).

package/config.example.jsonc

@@ -57,6 +57,9 @@
       // 中间 tool_use / tool_result / thinking 只写进详情页 (可折叠, Claude 网页风格)。
       // 授权卡照常发群。
       "brief": false,
+      // 出处门: CLI 手敲的那一轮只进终端 + chat 详情页, 不下发气泡。WeCom 发起的
+      // 轮次、审批卡、[mirror] 提示不受影响。只在 brief=true 下生效。
+      "chatOriginOnly": true,
       // 斜杠命令回执精简: /clear、/new 只回第一行 ack ("cleared"/"created"),
       // 不再附 📂 项目信息与 💡 tip。默认 false = 完整回执。
       "slashAckFirstLine": false

package/dist/daemon/mirror-bridge.js

@@ -28,7 +28,7 @@ import { labelFor, tagOfKey, baseOfKey, keyOf, withTagHeader, parseTagHeader } f
 import { splitMarkdown } from "../shared/md-chunk.js";
 import { randomTip } from "./tips.js";
 import { chatBaseOf, chatNameOf, listChatNames, normChatName, parsePeerRef, peerAddress } from "./chat-name.js";
-import { stripAnsi, compactPane, paneIsBusy, paneIsStalled, transcriptStalled, limitResetAt, summarizeTail, lastAssistantText, lastContextTokens, keepaliveStamps, tailTurns, renderDialog } from "./peers.js";
+import { stripAnsi, compactPane, paneIsBusy, paneIsStalled, transcriptStalled, summarizeTail, lastAssistantText, lastContextTokens, keepaliveStamps, tailTurns, renderDialog } from "./peers.js";
 // Same PATH augmentation logic as cc-bridge: launchd / systemd start the daemon
 // with a stripped PATH that often lacks nvm / homebrew, breaking spawn(claudeBin).
 const NODE_BIN_DIR = dirname(process.execPath);
@@ -2027,6 +2027,21 @@ export const startMirror = (deps) => {
         const tag = tagOfKey(target);
         return tag ? `[${labelFor(tag)} #${tag}](${url})` : `[🧙](${url})`;
     };
+    // ── 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
+    // 在看。群里保留的是需要在群里发生的事: WeCom 发起的轮次、审批/提问卡、
+    // `[mirror]` 系统提示 —— 它们都不走这道门。
+    // 只在 brief 下生效: 抑制的前提是详情页兜得住内容, 而 turn store 只有 brief 在写。
+    const turnSilent = (a) => cfg.wrc.mirror.brief && cfg.wrc.mirror.chatOriginOnly && a.turnFromChat === false;
+    // 详情页入口一次性告知。CLI 轮全程静默后, 一个从未收到过气泡的会话在群里将没有
+    // 任何链接可点 —— 每个 attachment 发一条 (且仅一条) 带链接的提示解决这个死角。
+    const noticeCliSilent = (a, turnId) => {
+        if (a.cliSilentNoticed)
+            return;
+        a.cliSilentNoticed = true;
+        sendRaw(a, `${briefDetailLink(turnId, a.target)} CLI 侧对话不再下发到群里，点标签看实时详情。`);
+    };
     // 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。
     // raw=true skips withSessionTag (used when content already contains the linked tag header).
     // WeCom 客户端收到 finish=true 后仍有打字机动画要播放, 如果紧接着就下发
@@ -2089,6 +2104,8 @@ export const startMirror = (deps) => {
     const startBriefTurn = async (a, frame, streamId, isSlash = false, userQuery = "") => {
         const turnId = newTurnId();
         a.queryEpoch = (a.queryEpoch ?? 0) + 1; // WeCom 侧的新一轮同样是 query 边界
+        a.turnFromChat = true; // 出处确凿: 这一轮有 frame, 群里就是它的主场
+        a.pendingFromCli = false; // 人改从聊天里说话了, 之前那条 CLI 输入不再是出处
         recordTurnStart({ id: turnId, target: a.target, sessionId: a.sessionId, cwd: a.runningCwd || undefined, userQuery: userQuery.trim() || undefined, cut: consumeCut(a), origin: consumeOrigin(a) });
         // 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 };
@@ -2130,7 +2147,12 @@ export const startMirror = (deps) => {
         const turnId = newTurnId();
         const query = a.pendingBriefQuery?.replace(/^> ?/gm, "").trim();
         a.pendingBriefQuery = undefined;
-        recordTurnStart({ id: turnId, target: a.target, sessionId: a.sessionId, userQuery: query || undefined, cut: consumeCut(a), origin: consumeOrigin(a) });
+        // 出处继承: 有真人 CLI 输入行开头 → CLI 轮; 否则 (收口后的补写、peer/graph 注入
+        // 开出的轮) 沿用上一轮 —— 那些续写属于同一场对话, 不该改变可见性。
+        const fromCli = a.pendingFromCli === true;
+        a.pendingFromCli = false;
+        a.turnFromChat = fromCli ? false : a.turnFromChat ?? true;
+        recordTurnStart({ id: turnId, target: a.target, sessionId: a.sessionId, cwd: a.runningCwd || undefined, userQuery: query || undefined, cut: consumeCut(a), origin: consumeOrigin(a) });
         a.briefTurnId = turnId;
         a.briefBubble = undefined;
         a.briefIsSlash = false;
@@ -2140,8 +2162,12 @@ export const startMirror = (deps) => {
         clearCot(a);
         // 链接暂存, 不再单独发 — 让首条 standalone body 取走拼到前缀, 一条消息解决。
         // turn 始终会通过 concludeBriefTurn / closeBriefTurn 收口, header 不会泄漏。
-        a.pendingBriefHeader = briefDetailLink(turnId, a.target);
-        log.info({ sessionId: a.sessionId, turnId }, "brief: turn started (CLI-side, no bubble)");
+        // 静默轮不留 header: 它没有任何 standalone 可以搭车, 留着只是个悬空的链接。
+        if (turnSilent(a))
+            noticeCliSilent(a, turnId);
+        else
+            a.pendingBriefHeader = briefDetailLink(turnId, a.target);
+        log.info({ sessionId: a.sessionId, turnId, fromCli, silent: turnSilent(a) }, "brief: turn started (CLI-side, no bubble)");
     };
     // 本轮出现过工具调用 / tool_result / 非 final 文本 —— 只记状态: 详情链接从 ack 起
     // 就挂在气泡里, 无需再推一次。气泡仍然不关, 等正文到来后以 `链接 正文` 覆盖收口。
@@ -2190,6 +2216,11 @@ export const startMirror = (deps) => {
             void finishBriefBubble(a, `${briefDetailLink(turnId, a.target)} ${body}`, true);
             a.pendingBriefHeader = undefined;
         }
+        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。
+            a.pendingBriefHeader = undefined;
+            log.info({ sessionId: a.sessionId, turnId }, "brief: CLI-origin conclusion kept out of chat");
+        }
         else {
             const header = a.pendingBriefHeader;
             a.pendingBriefHeader = undefined;
@@ -2476,6 +2507,10 @@ export const startMirror = (deps) => {
                 void finishBriefBubble(a, item.body);
                 a.pendingBriefHeader = undefined;
             }
+            else if (turnSilent(a)) {
+                // CLI 里手敲的 /clear、/model 之类, 回执给敲的人看就够了 —— 只进详情页。
+                a.pendingBriefHeader = undefined;
+            }
             else {
                 // CLI-driven turn: 把 ensureBriefTurn 暂存的详情链接拼到 body 前缀, 避免
                 // "只发了一条 link 渲染成空文本" 的前置消息。
@@ -2670,6 +2705,11 @@ export const startMirror = (deps) => {
             // 清掉防残留阻塞新 turn 的软收口 (guard 的 defer 判断依据)。
             voidTurnAgents(a);
             a.queryEpoch = (a.queryEpoch ?? 0) + 1;
+            // 这一行是人在 CLI 里敲的 —— 它既是下一轮的 query, 也是下一轮的出处。
+            // quiet (includeUser=false, 默认) 同样记: includeUser 只决定气泡渲不渲染,
+            // 不决定详情页记不记 —— 聊天不再下发 CLI 轮之后, 详情页是唯一的落点。
+            a.pendingBriefQuery = item.body;
+            a.pendingFromCli = true;
         }
         // 后台派发的结束信号 —— 它不是 function_call_result (那只是 spawn 句柄)。放在
         // keepalive 门之前: 吞没窗口里到达也必须销账, 否则账本永久悬空。
@@ -2877,13 +2917,9 @@ export const startMirror = (deps) => {
         }
         // Brief 模式下没有活跃 turn 时, 任何 assistant 侧产出都补开一个无气泡 turn ——
         // 覆盖 CLI 侧直接开的新一轮 (WeCom 没参与, 拿不到 frame) 与收口后的零星补写。
-        // user_text 只记下来当下一轮的 query, 不开 turn (CLI 敲字 ≠ 一定有回复)。
-        if (cfg.wrc.mirror.brief && !a.briefTurnId) {
-            if (item.kind === "user_text")
-                a.pendingBriefQuery = item.body;
-            else if (BRIEF_TURN_OPENERS.has(item.kind))
-                ensureBriefTurn(a);
-        }
+        // user 行不开 turn (CLI 敲字 ≠ 一定有回复), 它的 query/出处已在上面记下。
+        if (cfg.wrc.mirror.brief && !a.briefTurnId && BRIEF_TURN_OPENERS.has(item.kind))
+            ensureBriefTurn(a);
         // Usage snapshots never flow into WeCom bubbles — brief 分支下 handleBriefItem
         // 会把它写进 turn store, 非 brief 下直接吞掉, 保持 onItem 主流程只处理有渲染
         // 输出的 item 类型。
@@ -4352,10 +4388,8 @@ export const startMirror = (deps) => {
                 // Stall recovery, decided by RULE only (no model self-judgment): the last
                 // transcript turn is a synthetic API-error/limit line, or the idle pane
                 // still shows an error banner ⇒ a turn died mid-work. Send the resume
-                // instruction instead of the plain warmer. A limit-parked episode is
-                // excluded: retrying before resetsAt only buys another 429 line —
-                // limitResumeTick owns that recovery, anchored on the reset moment.
-                const stalled = kc.resumeOnStall && !a.limitResume && (transcriptStalled(a.jsonlPath) || paneIsStalled(paneTail));
+                // instruction instead of the plain warmer.
+                const stalled = kc.resumeOnStall && (transcriptStalled(a.jsonlPath) || paneIsStalled(paneTail));
                 k.pinging = true;
                 k.pingMtime = k.lastMs; // settles when a newer turn (the ping's own) appears
                 await fireKeepalive(a, stalled);
@@ -4369,83 +4403,6 @@ export const startMirror = (deps) => {
         }
     };
     const keepaliveTimer = setInterval(() => void keepaliveTick(), 15_000);
-    // ── Rate-limit auto-resume ────────────────────────────────────────────
-    // Detection is poll-derived from the transcript tail every tick (limitResetAt:
-    // "last message turn IS a 429 line"), so there is nothing to persist — a
-    // daemon reload, a human retry in the TTY, or our own inject all converge
-    // naturally: any newer turn makes detection stop firing and drops the state.
-    const fmtClock = (ms) => {
-        const d = new Date(ms);
-        return `${String(d.getHours()).padStart(2, "0")}:${String(d.getMinutes()).padStart(2, "0")}`;
-    };
-    let limitResumeTicking = false;
-    const limitResumeTick = async () => {
-        const lr = cfg.wrc.mirror.limitResume;
-        if (!lr.enabled || limitResumeTicking)
-            return;
-        limitResumeTicking = true;
-        try {
-            const now = Date.now();
-            for (const a of byTarget.values()) {
-                const resetsAt = limitResetAt(a.jsonlPath, now);
-                if (!resetsAt) {
-                    a.limitResume = undefined;
-                    continue;
-                }
-                if (a.migrationWatcher)
-                    continue; // session rotating — judge after it settles
-                if (a.keepaliveOff)
-                    continue; // /stop = deliberately quieted; honor it
-                if (a.limitResume?.resetsAt !== resetsAt)
-                    a.limitResume = { resetsAt, notified: false, retried: false };
-                const st = a.limitResume;
-                const dueAt = resetsAt + lr.delaySec * 1000;
-                if (!st.notified) {
-                    st.notified = true;
-                    log.info({ target: a.target, resetsAt, dueAt }, "limit-resume: parked, retry scheduled");
-                    sendStandalone(a, `[mirror] ⏳ 已触限额 (reset ${fmtClock(resetsAt)}) — ${fmtClock(dueAt)} 自动注入 \`${lr.text}\` 续跑`);
-                }
-                if (st.retried || now < dueAt)
-                    continue;
-                if (a.liveStream && !a.liveStream.closed)
-                    continue; // mid typewriter — don't inject
-                if (a.tmuxPane) {
-                    // Pane mode: a dead pane means the human closed shop — auto-respawning
-                    // a TUI at 3am would be a surprise, not a service. Busy means someone
-                    // (human, or another turn) is already driving; detection clears itself
-                    // once their turn lands. Spawn-mode has no pane and no such hazards —
-                    // inject's `claude --resume -p` fallback IS its normal operation.
-                    if (!(await tmuxPaneAlive(a.tmuxPane)))
-                        continue;
-                    if (paneIsBusy(await capturePaneTail(a.tmuxPane, 16)))
-                        continue;
-                }
-                st.retried = true;
-                // Unlike keepalive there is no swallow: the resumed work is real work —
-                // the tail mirrors its output to chat normally. rememberInject only
-                // suppresses the echo of the injected user line itself.
-                rememberInject(lr.text);
-                const r = await inject({
-                    text: lr.text, images: [], cfg,
-                    log: log.child({ target: a.target, sessionId: a.sessionId, sub: "limit-resume" }),
-                    sessionId: a.sessionId, jsonlPath: a.jsonlPath, tmuxTarget: a.tmuxPane,
-                });
-                if (!r.ok) {
-                    st.retried = false; // roll back so the next tick retries the inject
-                    log.warn({ target: a.target, reason: r.reason }, "limit-resume: inject failed");
-                    continue;
-                }
-                log.info({ target: a.target, resetsAt }, "limit-resume: resume injected");
-            }
-        }
-        catch (e) {
-            log.warn({ err: e.message }, "limit-resume tick failed");
-        }
-        finally {
-            limitResumeTicking = false;
-        }
-    };
-    const limitResumeTimer = setInterval(() => void limitResumeTick(), 30_000);
     // ── 原生权限 picker 哨兵 ──────────────────────────────────────────────
     // CodeBuddy 对一部分调用先弹自家权限 picker、PreToolUse hook 在人点掉之后才
     // 触发 / 根本不触发 (与 AskUserQuestion 的「本地面板先行」同源; 实测命中: 子代理
@@ -4961,7 +4918,6 @@ export const startMirror = (deps) => {
         shutdown: () => {
             clearInterval(paneDriftTimer);
             clearInterval(keepaliveTimer);
-            clearInterval(limitResumeTimer);
             clearInterval(pickerTimer);
             for (const a of bySessionId.values()) {
                 if (a.outbound?.kind === "deferred")