功能层面我逐模块看实现细节，找的是真实的功能缺陷和缺口，不是罗列功能名。

The user asks about functionality aspects — likely wants an analysis of the features: what features it has, functional gaps/issues, functional completeness, maybe bugs. "功能方面呢" = "What about the functional aspects?" following up on compatibility/maintenance/security. So they want a functional review: feature inventory, functional defects/gaps, correctness issues.

Let me dig into the actual feature implementations to find functional issues rather than just listing features (already listed). I should identify:
- Feature completeness/bugs
- Correctness/logic issues
- UX limitations
- Missing capabilities

Let me examine key feature files: command.go, moderation.go, song.go, video.go, identity.go, report.go, messenger.go, route.go, gate.go, toolrun.go, tutorial. And look for functional bugs.

Let me read several relevant files in parallel. Focus on finding actual functional issues.

Let me read: internal/bot/command.go, internal/bot/song.go, internal/bot/video.go, internal/bot/moderation.go, internal/bot/identity.go, internal/bot/report.go, internal/bot/route.go, internal/bot/gate.go.

That's a lot. Let me prioritize. Read command.go, route.go, moderation.go first, then song.go, video.go, identity.go, report.go.

Let me batch reads.

Let me now look at the moderation (mute/kick parsing), song, video, identity, report, messenger (markdown card detection), gate. Let me read them.

Read moderation.go, song.go, video.go, report.go. Might be long. Let me read moderation.go and video.go and report.go and song.go.

Actually let me focus: read song.go (interaction flow), moderation.go (command parsing edge cases), video.go, report.go (card logic), messenger.go (truncation, media).

I'll batch a few.

Functional observations from moderation:
- `handleAutoModeration` returns false when not reached, so the message continues to LLM. Good.
- When muted, returns true. Good.
- `matchBannedWords` uses `strings.Contains` on lowercased text → **substring matching, no word boundaries**. "枪" would match inside unrelated words. Also for Chinese, "操" would match "操场". This is a real functional weakness: false positives. Also banned words split by punctuation.
- `parseMuteMinutes`: "禁言10abc" fine. But "禁言" + non-digit -> default. OK.
- `normalizeCommand` accepts "踢"/"解禁"/"禁言" as prefix anywhere if prefix is mention. But note: any group message starting with "踢" (e.g., "踢球去了") by a **moderator** would be treated as a kick command and if no mention → replies "请 @". For non-moderators, returns true? No: for non-moderators it replies "没有权限" and returns true — meaning a normal member saying "踢球" triggers a permission-denied reply. Actually wait — handleModeration returns true even for non-moderators (replies permission denied) and blocks the LLM. That's a functional false-trigger: anyone saying a sentence starting with "踢"/"禁言"/"解禁" in a group gets a permission-denied reply instead of a normal AI answer. That's a real UX bug/false positive. E.g. "踢球 好玩" from a member → bot says "你没有权限". Hmm wait, but the switch check comes first: if ModerationKick enabled (default true), then it proceeds; scope group; canModerate false → replies "没有权限使用该命令" returns true. Yes. So normal conversation starting with these chars is hijacked. Real functional issue.

Also "解禁" prefix matches "解禁" but "禁言" is checked before? switch order: HasPrefix "禁言", "解禁", "踢出", "踢", kick. "解禁xxx" → HasPrefix "禁言"? no. HasPrefix "解禁" yes → unmute. Fine. But a message "解释一下" starts with "解" not "解禁". OK.

But "禁言" only triggers if prefix. A normal sentence "禁言制度是怎样的" → HasPrefix 禁言 → mute command → non-moderator gets denied. False positive.

Also `handleModerationStep` calls handleModeration at priority 700, before command (600) and LLM (100). So it intercepts.

Let me confirm: for a normal user in group saying "踢足球" the bot replies permission denied. That's a functional issue worth noting.

Let me read song.go and video.go and report.go for more.

Also messenger.go for markdown card / truncation.

Let me read song.go.

Song feature observations:
- `songKeyword` matches "点歌"/"点播" ANYWHERE in the text (idx anywhere, not just prefix). So a normal sentence containing "点歌" anywhere is hijacked → e.g., "我想点歌单里的第一首" or "你帮我推荐点歌" → treated as song search with garbage keyword. It returns everything after the keyword. Real false-positive risk. Actually more precisely: any message containing "点歌" anywhere triggers song flow. E.g. "这个点歌真的好用" → keyword "真的好用" → search fails → reply "没找到". That hijacks conversation. Functional issue.
- `songIndex`: bare number 1-99 selects pick if a pending list exists for the conversation. Note it's bound to conversation and requester (store binds sender). But `songPicks.has(convKey)` then `pick(convKey, sender, index)` checks sender. OK. But a bare "1" in a group with pending list from someone else → has(convKey) true, pick returns not-ok (different sender) → replies "没有可点播的歌曲编号..."? Wait handleSong: `if n, ok := songIndex(norm); ok && b.songPicks.has(convKey)` → then playPick. playPick: pick returns ok=false → replies "没有可点播的歌曲编号，请先发送「点歌 歌名」。" and returns true (handled). So a message "1" from another user gets a confusing reply. Also, bare numbers as normal chat (e.g., answering "1" to something) get intercepted. It claims the message (returns true) → LLM never sees it. Minor UX issue but notable: any standalone 1-99 while a list is pending is consumed.

Actually is that intended? The list has TTL 5 min. Within that window any bare number message in that conversation is captured. Could be annoying.

- Duration: QQ voice cap: MEMORY says 4+ min voice works, retries. Fine.
- Fallback: on total failure replies neutral. Good — doesn't leak direct URL (security-conscious).

Now video.go and report.go and messenger.go. Let me read video.go and report.go and messenger.go.

Let me batch.

Video feature observations:
- `ExtractLink` picks first http(s) link, detects platform (douyin/bilibili/kuaishou). If link isn't a supported platform, ExtractLink returns ok=false presumably. But note: it's gated by `mc.St.Features.VideoParse` and runs before @-gate → so in full-message mode it parses links in non-@ messages. If video_parse enabled and group is in full-message mode, bot will reply to every link posted. That could be spammy but it's intended.
- Concern: handleVideo runs before group @-gate (priority 900 vs gate 800). In mention-only group mode (GroupMentionOnly true), video handler still runs on un-mentioned messages because gate is AFTER video. Wait: handler order by priority desc: auto_moderation(1000), video_parse(900), group_gate(800), ... So video parse runs BEFORE the @ gate. That means even when "群聊仅 @ 回复" is ON, an un-mentioned group message containing a video link will be parsed and replied to. Is that intended? MEMORY says video runs before mention-only gate by design ("因此「接收全部消息」模式下未被 @ 的群链接也会被解析"). But that logic applies to full-message mode. In mention-only mode, QQ only pushes... hmm. Actually whether the bot receives un-mentioned group messages at all depends on QQ's "消息范围" setting (全量消息). If the bot only receives @ messages, then gate is moot. If the bot receives all messages (full mode) but user set "仅@回复", then video parse would still fire on non-@ links — contradicting the "仅 @" intent. That's a functional inconsistency: the mention-only switch doesn't suppress video parsing. Worth noting.

Similarly auto_moderation runs before gate (intended, to moderate all messages).

Let me verify: In MEMORY #161: "bot OnMessage 在 handleAutoModeration 之后、群聊 mention-only 门控之前调用 handleVideo(...)，因此「接收全部消息」模式下未被 @ 的群链接也会被解析。" So yes it's intentional that video parse ignores the mention-only gate. But it's arguably a functional surprise: turning on "只回复@" doesn't stop link parsing. I'll note as a design/UX consideration.

report.go: IsReportWith threshold logic. Fine. Note `hasReportMarker` uses `|` → any line with a pipe (e.g., "a|b") counts. Minor.

Now messenger.go (how replies are sent: markdown card, truncation, media) and gate.go (the @ detection). Let me read messenger.go and gate.go.

Also identity.go (name extraction) and maybe tutorial. Let me read messenger.go and gate.go and identity.go.

Let me batch.

gate.go is actually the speaker label file (misnamed?). The @-gate is in pipeline.go handleGroupGateStep. Let me find deliver function — it's in messenger.go? No, deliver is at end of gate.go after line 40? The Read stopped at 40 because file has more. Let me read gate.go fully. Also read reply.go (deliver, reply, markdown logic) and event.go (BotMentioned, MentionedMembers).

Let me read gate.go rest and reply.go.

Actually "deliver" comment starts at line 38-40 of gate.go but function body continues. Let me read gate.go from 40.

Let me read reply.go and the rest of gate.go.

gate.go seems truncated — the file only has 40 lines and the deliver comment is a dangling comment? That means deliver function is declared elsewhere, or the file got cut. Wait, gate.go content: speakerLabel, userLabel, then a comment block ending at line 40 "it falls back to plain text so the user still gets a reply." with no function following — that's odd. Perhaps the file is exactly that and the comment is orphaned (dead comment). Actually a dangling comment at EOF is legal in Go. So the `deliver` function is... where? deliver is defined in reply.go (lines 9-23). So gate.go's trailing comment is an orphaned leftover. Minor code cleanliness. Actually the comment says "deliver sends the assistant text..." but deliver lives in reply.go — so the comment got separated from its function. Not a functional bug but code smell.

Hmm, but wait: multiple files declare `deliver`? No, only reply.go. So gate.go has an orphan comment. Fine.

Now note: `deliver` calls `b.reply(... truncateRunes(content, maxReplyRunes))`. maxReplyRunes=1500. But replies are truncated; markdown 4000. So long LLM answers get truncated without notice "…". Minor.

Also identity shows openid to the user (in group, their own openid) — privacy-ish but it's their own.

Now the tools: toolrun.go and tools.go for functional. And identity.go name learning regex.

Let me read identity.go and toolrun.go quickly. And api.go SendText for markdown/keyboard and content length limits. Actually enough.

Let me read identity.go (name extraction) and tools.go (tool definitions).

Let me batch read identity.go and toolrun.go.

Identity name extraction: regex based. Known functional risk: false positives. E.g., "我是说真的" → matches `我是` pattern → name "说真的"? Actually pattern `(?:我是|我是叫)\s*(...)` — "我是说真的" → name = "说真的". That's a false name. Or "我是你的粉丝" → name "你的粉丝"? It excludes... no, `你的粉丝` would be captured (not in stopwords). So the bot would record a wrong name. Also "我叫你" → name "你". Also "我是群主" → name "群主"? Not much. This name-learning is heuristic and will produce wrong profile names. Real functional weakness but low impact (only affects a display label).

Also: profiles store name; `/whoami` shows it.

Now tools definition — basic tools: system_status, uptime, disk_usage (all users). Advanced: processes/ports/services (admin). Fine.

Let me also check the event parsing (BotMentioned / MentionedMembers) and the group mention gate in pipeline.go (to confirm the video-before-gate claim and mention detection). I already read pipeline.go earlier but content was cleared. Let me re-read pipeline.go to confirm priorities and the gate.

Actually earlier I read pipeline.go and it had triggerRegistry global, priorities, handleGroupGateStep. Let me re-read to get exact details for the functional analysis (priority order, gate behavior, mention detection).

Let me read pipeline.go again.

Confirmed pipeline priorities. Key functional observations:

1. **`handleModerationStep` logs "inbound message" for EVERY message** (line 141-143), and handleModeration returns false for non-moderation. So it logs inbound info. Fine but logs at Info per message.

2. **Group gate only applies to `GROUP_MESSAGE_CREATE`** (full-message mode). In mention-only mode, QQ sends `GROUP_AT_MESSAGE_CREATE`. So gate only relevant in full mode. OK, consistent.

3. **Video before gate**: In full-message mode with `GroupMentionOnly=true`, an unmentioned message containing a video link is STILL parsed (video runs at 900 before gate 800). This contradicts "仅@回复". Intentional per comment, but a functional surprise. Similarly auto-moderation runs on all messages (intended).

4. **C2C**: gate doesn't apply. In C2C, moderation command "禁言" → scope != group → replies "该命令仅支持群聊" and returns true → blocks LLM! So a user in private chat saying "禁言" or "踢" gets "该命令仅支持群聊" and their AI conversation is hijacked. Minor false positive. Actually more: any private message starting with "踢"/"禁言"/"解禁" → hits handleModeration (priority 700 > command 600 > LLM 100), cmd set, feature enabled, scope!=group → reply "该命令仅支持群聊" return true. So private chat false positive. Note this.

Wait, does command handler run before moderation? No: moderation 700 > command 600. So "/help" is fine (doesn't start with 踢). But "禁言" in private. Edge.

5. **False-positive command capture in groups for non-moderators**: as noted, non-moderator saying a sentence starting with "踢"/"禁言"/"解禁" gets a "你没有权限" reply and loses the LLM answer. This is a real UX functional issue.

6. **Song false positive**: any message containing "点歌"/"点播" anywhere hijacks. And bare 1-99 while a pick list pending.

7. **`handleCommandStep`** uses `b.set()` inside handleCommand for IdentityQuery. Fine.

8. **`/tools`** lists tools; fine.

9. Name learning false positives.

10. **Trigger registry race** (already noted).

Now, let me also check the LLM answerWithLLM for functional issues: context window, memory trimming, group speaker labels, system prompt. And whether group conversations share context (convKey = group openid → all members share one context? Actually group conv key is group openid, so multiple users share memory — the bot maintains multi-speaker context with speaker labels). Let me read llm answer file (internal/bot/llm.go).

Let me read internal/bot/llm.go.

Let me find the file with answerWithLLM.

LLM turn observations:
- Group shared memory: convKey = group openid → all group members share one conversation memory. Speaker labels A/B/C. The bot replies to everyone, sharing context. This is intended (multi-participant). But a functional limitation: in group, the LLM sees everyone's messages, so one user's conversation bleeds into another's. Could be a privacy/UX issue but by design.
- `RecordLLM(...,0,0,0)` records a call but tokens recorded via usage scope (llm.WithUsageScope). OK.
- No streaming; no retry.
- `b.store.Reset(convKey)` when memory off.
- Compression: threshold/keep. If Summarize fails, turns re-appended (could loop but ok).

Now I have plenty of functional detail. Let me also quickly check the api send for markdown keyboard, and the C2C/group message content length limits; and check `SendMarkdown` usage. Also the tutorial custom text. Enough.

Let me also check the "无关消息/图片消息" handling: QQ pushes media-only messages. Does bot handle image messages (contents with <image>)? Let me check event.go for content sanitization and whether images/voice are stripped. And whether it handles the case where user sends only a picture (no text) → text may be "<image ...>" placeholder → LLM sees weird text. Let me check event.go ExtractText/content.

Let me read event.go.

`BotMentioned()` returns true if ANY mention has Bot=true. But mentions may include OTHER bots (e.g., another bot mentioned). Since QQ's author is the sender, and mentions include the current bot when @'d. But if a user @s a different bot, this returns true incorrectly. Edge case—QQ only includes the current bot if mentioned? Actually mentions list includes all mentioned users, including other bots. So `BotMentioned` could be true when a different bot is @'d. Minor.

Also `SenderOpenID` uses UserOpenID for group? In group messages, author.user_openid is present. In C2C, user_openid. Fine.

Also text sanitization: does the bot strip `<image src=...>` / voice content placeholders? Message.Content may include "<@!...>" mentions and possibly media placeholders. handleCommand strips leading `<@...>`. But for LLM, mc.Text is passed raw including `<@!openid>` mention tokens. Actually buildMessages uses mc.Text which still contains the mention token like "<@!1234>" if the bot mention is inline. Hmm — stripBotMention is only applied in command/song/moderation. For LLM, raw text with `<@!...>` is sent. That means the model sees the mention token. Minor cosmetic.

Let me check where mc.Text is set — in bot.go OnMessage. Earlier read cleared. Let me check: Text likely = normalize/trimmed content. Let me grep for "Text:" assignment.

Let me check bot.go around building MessageCtx and whether it strips media placeholder.

Let me grep.

So text = trimmed raw content, no stripping of `<@!...>` mention tokens or media placeholders like `<image ...>`/`<voice ...>`. So:
- If a group message is "@bot 你好", content = "<@!botid> 你好" (QQ leaves the structured mention). The LLM sees "<@!botid> 你好". Cosmetic but could confuse the model or leak openid into messages. Actually the bot's own openid is in the mention token → sent to LLM. Minor.
- Image/voice-only messages: QQ content for a pure image message may be empty or contain a placeholder. If empty, text="" → maybe skipped? Let's check OnMessage handling of empty text. Let me read bot.go 170-230.

Let me read that region.

Confirmed. text empty → dropped (so image-only messages with no text are silently ignored). Content still contains mention tokens for LLM.

Now I have a comprehensive functional picture. Let me also quickly check the settings.feature for defaults and panel feature switches to see if there are functional gaps/duplicated logic. And check the tutorial/command duplication (handleCommand's Help text hardcoded is duplicated from replyTutorial → maintenance).

I think I have enough. Let me write the functional analysis.

Let me structure the answer as:

## 功能实现质量总评 (brief)
Then sections:
A. 功能覆盖与成熟度（哪些扎实、哪些半成品）
B. 具体的功能缺陷 / 边界问题（带证据）
C. 功能缺口（没有的能力）
D. 建议

Let me be concrete and cite files.

Key functional defects to report:

1. 命令触发误伤（false positive）——群/私聊里任何以「踢」「禁言」「解禁」开头的普通句子会被当管理命令，非管理员/私聊场景直接回「无权限/仅群聊」并**吞掉**这条消息（不再走 LLM）。因为 moderation(700) 早于 llm(100) 且返回 true。证据：[pipeline.go](...#L87-L96)、[moderation.go](...#L141-L185)、[command.go]...
   即便非管理员：handleModeration 在 canModerate 失败时仍 return true。→ "踢球/禁言制度/解释..." 都会被劫持。

2. 点歌误触发——`songKeyword` 用 `strings.Index` 在**整条消息任意位置**找「点歌/点播」，任意消息只要含这两个词就进入点歌流程并回复"没找到…"，吞掉消息（prioSong 500 > LLM 100）。例：「你推荐的这首歌点歌单里有吗」。另外 `songIndex`：列表 pending 期间，对话里任意 "1"~"99" 的独立数字都会被当作选歌并 return true（即使选不到也回"请先点歌"）。

3. 昵称/身份学习的误识别——正则 `我是…`/`我叫…` 会把「我是说真的」「我是你的粉丝」等当成名字写进档案，且会显示在 /whoami 与提示词里。低危但会污染档案。

4. 群内上下文全群共享——convKey=群 openid，所有群成员的对话进同一份记忆并共享摘要；A 的私聊式对话会进入 B 的上下文。是设计取舍，但会造成串话/隐私感问题，且没有 scope 隔离开关。

5. 视频解析绕过「仅 @ 回复」——video_parse(900) 在 group_gate(800) 之前执行，[pipeline.go](...#L117-L124) 注释也承认「full-message 模式下未 @ 的群链接也会被解析」。所以用户打开「群聊仅 @ 回复」后，群里任何没 @ 的链接仍会被解析回复，与开关语义不一致。

6. 消息文本未净化——`text = TrimSpace(msg.Content)` 直接把 QQ 结构化 mention 令牌 `<@!openid>` 交给 LLM 与日志（[bot.go](...#L175-L215)）。模型会看到 `<@!xxx> 你好`；同时 bot 自己的 openid 被带进提示词。另外纯图片/语音消息（content 为空）被直接丢弃（[bot.go](...#L176-L178)），无法响应「发图问我」这类用法。

7. 回复被静默截断——普通文本 1500 字、Markdown 4000 字上限（reply.go truncateRunes / maxReplyRunes），超长回答直接截断且**无「已截断」提示**。

8. LLM 调用无重试/无超时控制细化——单次失败即回「稍后再试」（llm_turn.go），无 429/超时重试；也没有流式，长回答要等完整生成。

9. 群管理强依赖平台权限——踢人需 11253 白名单，mute 需机器人是群管理员；代码已把错误码映射成友好提示（做得不错），但功能在多数未申请权限的 bot 上实际不可用，属平台限制而非代码问题。

10. BotMentioned 逻辑：只要 mentions 里**任何**一个 u.Bot==true 就判定被 @，若群里同时 @ 了别的机器人会误判（[event.go](...#L67-L75)）。

C. 功能缺口：
- 无群成员/列表、无欢迎新人、无定时/主动推送（群发）、无多轮图片理解（多模态）、无管理员 CLI、无群级别差异化配置（全局一套 settings）、无模板/消息自定义（除 tutorial 外回复文案硬编码）、无黑名单用户、无静默模式（撤回自己消息）、无幂等/重试队列（DROP 而非排队）。
- 会话记忆只按 convKey 单一维度，无按用户隔离选项。
- 无 i18n（全中文硬编码）。
- 插件/脚本引擎只是预留接口（RegisterTrigger），未实现。

Also note dedup: markSeen cap 10000; if overflow maybe eviction. Fine.

Let me verify `songPicks` TTL to be accurate. Not critical.

Also handleCommand help text duplicates tutorial → two sources of truth.

I'll present concisely. Use tables. Provide file links with line numbers where I'm confident. I have line numbers from reads:
- pipeline.go priorities L87-96; video L117-124; gate L129-136.
- moderation.go handleModeration L141-185; normalizeCommand L304-321.
- song.go songKeyword L487-500; songIndex L505-523.
- bot.go L175-215.
- reply.go L9-23 (deliver), truncation.
- llm_turn.go L39-44 (no retry).
- event.go L67-75.

Good. Let me write final answer. Keep it tight but complete. Avoid emoji.

I should probably offer to fix. End with offer.

Let me write.

功能层面我把核心流程都读了一遍。下面是**功能实现质量 + 真实缺陷 + 能力缺口**三部分，都带代码证据。

## 一、总评

功能完成度高、主链路健壮：点歌的三级降级（语音→转MP3→文件→纯文本）、短视频/图文解析、违禁词阶梯提醒+自动禁言、@ 门控、多说话人匿名标签、记忆压缩、工具分级、Markdown 卡片自适应——这些都是**完整实现且有降级兜底**的，不是空壳。问题主要出在**关键词触发过宽**和**若干边界语义不一致**上。

## 二、具体的功能缺陷

| # | 问题 | 证据 | 影响 |
|---|---|---|---|
| 1 | **管理命令误劫持**：任何以「踢 / 禁言 / 解禁」开头的句子都进入管理流程，且非管理员/私聊时会**回一条拒绝并吞掉消息**（不再走 LLM） | [moderation.go](file:///workspace/analysis/extracted/internal/bot/moderation.go#L141-L185)、[pipeline.go](file:///workspace/analysis/extracted/internal/bot/pipeline.go#L87-L96) | 群员说"踢球好玩""禁言制度"→ 机器人回"你没有权限"；私聊说"禁言"→ 回"仅支持群聊"。普通对话被打断 |
| 2 | **点歌误劫持**：`songKeyword` 在整条消息**任意位置**找「点歌/点播」，含词即点数 | [song.go](file:///workspace/analysis/extracted/internal/bot/song.go#L487-L500) | "这首歌点歌单里有吗"→ 回"没找到…"，消息被吞 |
| 3 | **数字选歌误吞**：只要有 pending 歌单，对话里任何独立的 1–99 数字都被当选歌，选不到也回"请先点歌" | [song.go](file:///workspace/analysis/extracted/internal/bot/song.go#L505-L523) | 群聊里回答"1"会被截走 |
| 4 | **「仅 @ 回复」不生效于视频解析**：video_parse(900) 在群门控(800)之前 | [pipeline.go](file:///workspace/analysis/extracted/internal/bot/pipeline.go#L117-L124)、[L129-L136](file:///workspace/analysis/extracted/internal/bot/pipeline.go#L129-L136) | 开了"只回复@"后，群里没 @ 的链接照样被解析回复 |
| 5 | **文本未净化**：直接把 QQ 结构化 mention 令牌 `<@!openid>` 送进 LLM 与日志；纯图片/语音消息（content 空）被直接丢弃 | [bot.go](file:///workspace/analysis/extracted/internal/bot/bot.go#L175-L215) | 模型看到 `<@!xxx> 你好`、bot openid 进提示词；"发张图问我"无从响应 |
| 6 | **回复静默截断**：普通文本 1500 字 / Markdown 4000 字上限，无"已截断"提示 | [reply.go](file:///workspace/analysis/extracted/internal/bot/reply.go#L9-L23) | 长答案被砍尾，用户以为模型答完 |
| 7 | **姓名学习误识别**：`我是说真的`/`我是你的粉丝` 会被当名字写入档案 | [identity.go](file:///workspace/analysis/extracted/internal/bot/identity.go#L10-L15) | 档案与 /whoami、提示词被污染（低危） |
| 8 | **群记忆全群共享**：convKey=群 openid，所有成员进同一份记忆与摘要，无按用户隔离选项 | [llm_turn.go](file:///workspace/analysis/extracted/internal/bot/llm_turn.go#L24-L28) | A 的对话会进入 B 的上下文，存在串话/隐私观感 |
| 9 | **LLM 无重试/无流式**：单次失败即回"稍后再试"，无 429/超时重试 | [llm_turn.go](file:///workspace/analysis/extracted/internal/bot/llm_turn.go#L39-L44) | 偶发抖动直接丢一次回答；长回答需整段等待 |
| 10 | **@ 判定过宽**：mentions 里任意 `bot:true` 都算被 @ | [event.go](file:///workspace/analysis/extracted/internal/event/event.go#L67-L75) | 群里同时 @ 了别的机器人会误判 |
| 11 | **过载即丢消息**：并发上限 16，满了直接丢弃不排队 | [bot.go](file:///workspace/analysis/extracted/internal/bot/bot.go#L220-L225) | 高峰时用户消息静默丢失 |

## 三、功能缺口（目前没有的能力）

- **主动能力**：无群欢迎新人、无定时群发/提醒、无公告推送（只有被动应答）。
- **配置粒度**：settings 是**全局一套**，没有"按群/按用户差异化"配置；除教程文案外，回复文案全硬编码，不可自定义。
- **名单管理**：无用户黑名单/白名单（只有管理员名单），无"我的信息"之外的自助查询。
- **会话维度**：记忆只按 convKey 单维度，无法切"每人独立记忆"或"临时会话/静默模式（撤回自己上一句）"。

- **i18n**：全中文硬编码，无多语言。
- **运维**：无健康检查端点、无指标导出（Prometheus），只有面板日志。

## 四、给的结论
