package main import ( "context" "fmt" "log/slog" "sync" "time" ) // roomMeta caches per-room classification we need to handle a message: member // counts (for the 1:1 test, F3), whether any member is outside ALLOWED_SERVERS, // and encryption state (F15). Lazily fetched from the CS-API on first need // (appservice transactions carry no room summary) and INVALIDATED whenever a // third party's membership changes, so a 1:1 that gains a member is reclassified // out of DM mode (no DM-mode third-party leak) and a newly added foreign member // is caught. All fields are guarded by Bot.mu — never read or written without it, // because the slow generation and the lazy CS-API probes run in concurrent per-room // goroutines. type roomMeta struct { joined, invited int countsKnown bool foreign bool // a joined/invited member is outside ALLOWED_SERVERS encrypted, encKnown bool } func (m *roomMeta) isDM() bool { return m.countsKnown && m.joined+m.invited == 2 } type Bot struct { cfg *Config log *slog.Logger mx *MatrixClient xai *XAIClient st *Store // mu guards the in-memory maps/sets below. Each transaction is acked to Synapse // immediately (appservice.go) and its events are processed in transaction order, // but the slow xAI generation runs in a per-room goroutine and the lazy probes run // off-lock, so several goroutines touch this shared state at once. mu is held only // for short map operations and is NEVER held across a network or xAI call — that // head-of-line hold was the root cause of the multi-minute silence. mu sync.Mutex seen *lruSet // event ids already handled (dedup within a session; self-locking) botSent *lruSet // event ids the bot itself sent (reply-parent detection; self-locking) meta map[string]*roomMeta buf map[string][]bufferedMsg inflight map[string]bool // roomID currently generating a reply (per-room single-flight) } func NewBot(ctx context.Context, cfg *Config, logger *slog.Logger) (*Bot, error) { mx := NewMatrixClient(cfg.HomeserverURL, cfg.ASToken, cfg.BotMXID) xai := NewXAIClient(cfg.XAIBaseURL, cfg.XAIAPIKey, logger) st, err := OpenStore(cfg.DatabaseURL) if err != nil { return nil, err } b := &Bot{ cfg: cfg, log: logger, mx: mx, xai: xai, st: st, seen: newLRUSet(5000), botSent: newLRUSet(5000), meta: make(map[string]*roomMeta), buf: make(map[string][]bufferedMsg), inflight: make(map[string]bool), } // Confirm the as_token + user_id resolves to BOT_MXID before serving. if err := b.verifyIdentity(ctx); err != nil { st.Close() return nil, err } // F23: ensure the profile has a display name (best-effort, idempotent). if err := mx.SetDisplayName(ctx, cfg.BotDisplayName); err != nil { logger.Warn("set display name failed (non-fatal)", "err", err) } return b, nil } func (b *Bot) Close() { if b.st != nil { _ = b.st.Close() } } func (b *Bot) verifyIdentity(ctx context.Context) error { who, err := b.mx.Whoami(ctx) if err != nil { return err } if who != b.cfg.BotMXID { return fmt.Errorf("as_token resolves to %q but BOT_MXID is %q", who, b.cfg.BotMXID) } b.log.Info("authenticated", "mxid", who) return nil } // Run starts the appservice transaction server and blocks until ctx is cancelled. func (b *Bot) Run(ctx context.Context) error { as := NewAppService(b.cfg, b.log, b.st, b.handleTransaction) return as.Serve(ctx) } // handleTransaction processes one already-acked transaction's events. It runs in a // background goroutine (the 200 has already been returned to Synapse). Events are // processed IN ORDER (dedup + classification are synchronous) so the per-room // single-flight claim is taken in arrival order; only the slow xAI generation is // spawned per room, so different rooms answer concurrently and a slow call never // blocks the next event or another room. Ordering within a room is therefore kept, // while the head-of-line freeze is gone. func (b *Bot) handleTransaction(ctx context.Context, events []Event) { for i := range events { b.handleEvent(ctx, &events[i]) } } // safego runs fn in a goroutine with panic recovery. The slow per-room work is // detached from the HTTP handler, so an unrecovered panic there would crash the // whole process and silence the bot for EVERY room — recover + log instead, so one // malformed event can never take the bot down. func (b *Bot) safego(what string, fn func()) { go func() { defer func() { if r := recover(); r != nil { b.log.Error("recovered panic in handler goroutine", "what", what, "panic", r) } }() fn() }() } func (b *Bot) handleEvent(ctx context.Context, ev *Event) { if ev.EventID == "" || ev.RoomID == "" { return } if !b.markSeen(ev.EventID) { return // already handled (in-memory or durable dedup) } b.log.Debug("event", "type", ev.Type, "room", ev.RoomID, "sender", ev.Sender, "id", ev.EventID) switch ev.Type { case "m.room.member": if ev.StateKey != nil && *ev.StateKey == b.cfg.BotMXID { b.safego("self-membership", func() { b.handleSelfMembership(ctx, ev) }) } else { // A third party's membership changed: counts + foreign flag are now stale. // Re-probe on the next message so a 1:1 that gains a member drops out of DM // mode (no third-party leak) and a new foreign member is caught. b.invalidateCounts(ev.RoomID) } case "m.room.encryption": b.setEncrypted(ev.RoomID) case "m.room.message": // Synchronous (in transaction order) up to the single-flight claim; only the // slow generation inside handleMessage is spawned as a goroutine. This keeps // per-room event order — the earlier message wins the claim — while different // rooms still run concurrently. b.handleMessage(ctx, ev) } } // markSeen records an event id in both the in-memory set and the durable store and // reports whether it is NEW (first time). The in-memory Add is atomic, and SeenEvent // is an atomic INSERT … ON CONFLICT DO NOTHING, so two racing goroutines for the same // event can never both proceed. On a durable-store error we fall through (the in-memory set still // guards this session). func (b *Bot) markSeen(eventID string) bool { if !b.seen.Add(eventID) { return false } isNew, err := b.st.SeenEvent(eventID) if err != nil { b.log.Error("durable dedup check failed", "id", eventID, "err", err) return true } return isNew } // handleSelfMembership reacts to membership changes for the bot user: auto-join // invites from allowed servers (F11), reject others, forget rooms we leave. Runs in // its own goroutine because JoinRoom/LeaveRoom are network calls. func (b *Bot) handleSelfMembership(ctx context.Context, ev *Event) { switch ev.membershipOf() { case "invite": if !b.cfg.AllowedServers[serverOf(ev.Sender)] { b.log.Warn("rejecting invite (server not allowed)", "room", ev.RoomID, "sender", ev.Sender) if err := b.mx.LeaveRoom(ctx, ev.RoomID); err != nil { b.log.Error("leave (reject) failed", "room", ev.RoomID, "err", err) } return } b.log.Info("accepting invite", "room", ev.RoomID, "sender", ev.Sender) if err := b.mx.JoinRoom(ctx, ev.RoomID); err != nil { b.log.Error("join failed", "room", ev.RoomID, "err", err) return } // Fully-on-allowed-servers gate: a vojo.chat inviter can still pull the bot // into a room that already holds federated third parties — leave at once. if _, _, foreign := b.ensureCounts(ctx, ev.RoomID); foreign { b.leaveForeign(ctx, ev.RoomID) } case "leave", "ban": b.forgetRoom(ev.RoomID) } } // leaveForeign leaves a room that contains a member outside ALLOWED_SERVERS, so // the bot only ever operates in rooms hosted entirely on allowed homeservers. func (b *Bot) leaveForeign(ctx context.Context, roomID string) { b.log.Warn("leaving room — a member is outside ALLOWED_SERVERS", "room", roomID) if err := b.mx.LeaveRoom(ctx, roomID); err != nil { b.log.Error("leave (foreign) failed", "room", roomID, "err", err) } } func (b *Bot) handleMessage(ctx context.Context, ev *Event) { roomID := ev.RoomID // A9/F15: re-check encryption; if (or once) encrypted, react once and skip — the // bot can't read it. The probe runs without the lock. if b.ensureEncryption(ctx, roomID) { b.log.Debug("skip: encrypted room", "room", roomID) b.reactEncryptedOnce(ctx, roomID, ev.EventID) return } mc, ok := ev.DecodeMessage() if !ok { // A message addressed to the bot that we can't decode shouldn't vanish without // a trace (no silent drops): log at WARN so it's visible at the default level. b.log.Warn("skip: undecodable message content", "room", roomID, "sender", ev.Sender, "id", ev.EventID) return } // Edits re-carry m.mentions; never re-trigger or replay them (F16). if mc.IsReplace() { return } if ev.Sender == b.cfg.BotMXID { return // our own message (the reply is buffered when we send it, not on echo-back) } if mc.MsgType == "m.notice" { return // anti-loop: ignore notices (ours and other bots') } // Media / non-text is handled only once we know the message is addressed (below), // so a stray image in a group the bot isn't mentioned in stays silent (correct), // while media in a 1:1 or an @-mention gets a clear "text only" reaction. isMedia := mc.MsgType != "m.text" && mc.MsgType != "m.emote" countsKnown, isDM, foreign := b.ensureCounts(ctx, roomID) // Stay only in rooms hosted entirely on allowed servers — never operate in (or // "leak" the bot into) a federated room with non-consenting third parties. if foreign { b.leaveForeign(ctx, roomID) return } replyParentIsBot := mc.RelatesTo != nil && mc.RelatesTo.InReplyTo != nil && b.botSent.Has(mc.RelatesTo.InReplyTo.EventID) mentioned := mentionsBot(mc, b.cfg.BotMXID, replyParentIsBot) if !(isDM || mentioned) { if !countsKnown { // We couldn't classify the room (member probe failed) and the message isn't // an explicit mention, so we can't tell a 1:1 (answer everything) from a // group (answer only on mention). Log at WARN — not a silent Debug drop — // so it's visible; we don't react because reacting in a group the bot isn't // addressed in would be wrong. Re-probed on the next message. b.log.Warn("skip: room unclassified (member probe failed), message not an explicit mention", "room", roomID, "sender", ev.Sender) } else { b.log.Debug("skip: not addressed (group without mention)", "room", roomID, "sender", ev.Sender, "dm", isDM, "mentioned", mentioned) } return } // Addressed but not text: react "text only" (no silent drop). if isMedia { b.log.Debug("skip: non-text msgtype (reacted)", "room", roomID, "sender", ev.Sender, "msgtype", mc.MsgType) b.react(ctx, roomID, ev.EventID, reactMedia) return } // Per-room single-flight: while a generation is in flight for this room, drop // further messages (like a chat — no queue, no new session, no token burn). No // reaction here: the typing indicator is already showing for this room, which is // the language-free "I'm busy" signal. The claim is taken here, synchronously and // in transaction order, so the FIRST message for a room wins and later ones are // dropped until release — never the reverse. if !b.tryClaim(roomID) { b.log.Debug("drop: room busy generating", "room", roomID, "sender", ev.Sender) return } // Snapshot the room history (excludes this trigger) under the claim, then run the // slow generation in its own goroutine so this transaction's remaining events and // other rooms are not blocked by the xAI call. respond appends the trigger+answer // to the buffer itself, only on success (see sendReply), and releases the claim. history := b.snapshotBuf(roomID) b.safego("respond", func() { defer b.release(roomID) b.respond(ctx, roomID, isDM, ev, mc, history) }) } // unlimitedCap is the effective per-user cap for UNLIMITED_USERS — high enough to // never trip the per-user gate, while the global DAILY_USD_CEILING still applies. const unlimitedCap = 1 << 30 func (b *Bot) respond(ctx context.Context, roomID string, isDM bool, ev *Event, mc *MessageContent, history []bufferedMsg) { perUserCap := b.cfg.PerUserDailyCap if b.cfg.UnlimitedUsers[ev.Sender] { perUserCap = unlimitedCap } switch res, err := b.st.Reserve(ev.Sender, perUserCap, b.cfg.DailyUSDCeiling); { case err != nil: // A limiter failure is on our side — don't leave the user wondering. b.log.Error("limiter reserve failed", "sender", ev.Sender, "err", err) b.react(ctx, roomID, ev.EventID, reactError) return case res == reserveDeniedUser: // Per-user cap (anti-abuse, F24): stop answering, but always signal the limit — // no message addressed to the bot is left without feedback. b.log.Info("per-user daily cap reached; reacting", "sender", ev.Sender) b.react(ctx, roomID, ev.EventID, reactRateLimit) return case res == reserveDeniedGlobal: // Global USD ceiling. A reaction is cheap and non-intrusive (unlike the old // once-per-day text notice), so signal every affected message rather than // going silent after the first. b.log.Warn("global daily USD ceiling reached", "room", roomID, "sender", ev.Sender) b.react(ctx, roomID, ev.EventID, reactRateLimit) return } // Show "Vojo AI печатает…" for the whole generation. The keepalive refreshes the // typing notification every 20s (the server expires it after 30s) so the indicator // never lapses on a slow/retried answer, and the deferred stop clears it on exit. stopTyping := b.startTypingKeepalive(ctx, roomID) defer stopTyping() msgs := buildContext(b.cfg.SystemPrompt, history, isDM, mc.Body, b.cfg.MaxCtxEvent, 8000) resp, err := b.xai.Complete(ctx, b.cfg.XAIModel, msgs, b.cfg.MaxOutTok, b.cfg.XAITemp) if err != nil { // at-most-once already retried transient failures inside Complete; refund the // reserved request so an xAI outage doesn't burn the user's daily cap, and // signal the failure (react → no anti-loop, no language). b.log.Error("xai completion failed", "sender", ev.Sender, "err", err) if rerr := b.st.RefundRequest(ev.Sender); rerr != nil { b.log.Error("refund failed", "sender", ev.Sender, "err", rerr) } b.react(ctx, roomID, ev.EventID, reactError) return } // A 2xx from xAI is billed even if the text came back empty — always book the real // cost so both caps see it (an empty 200 must not bypass the per-user cap and the // global ceiling). usd := computeUSD(resp.Usage, b.cfg) if err := b.st.Reconcile(ev.Sender, usd); err != nil { b.log.Error("reconcile spend failed", "sender", ev.Sender, "err", err) } text := resp.Text() if text == "" { // Billed but no usable text (content filter / length cap / empty choices). Never // leave a billed request without feedback — react "couldn't answer". b.log.Warn("xai returned empty completion (billed, reacting)", "sender", ev.Sender, "usd", usd) b.react(ctx, roomID, ev.EventID, reactError) return } b.log.Info("answered", "room", roomID, "sender", ev.Sender, "dm", isDM, "usd", usd, "prompt_tokens", resp.Usage.PromptTokens, "completion_tokens", resp.Usage.CompletionTokens) b.sendReply(ctx, roomID, ev, mc, text) } // computeUSD prices the call from the API-returned token usage (authoritative // counts) and the configured per-1M prices — so the hard ceiling tracks real // usage even if the model/price changes (only the constants need updating). func computeUSD(u xaiUsage, cfg *Config) float64 { cached := u.PromptTokensDetails.CachedTokens nonCached := u.PromptTokens - cached if nonCached < 0 { nonCached = 0 } return float64(nonCached)/1e6*cfg.PriceInputPerM + float64(cached)/1e6*cfg.PriceCachedPerM + float64(u.CompletionTokens)/1e6*cfg.PriceOutputPerM } // react adds an emoji m.reaction to the triggering event — the bot's language-free // way to signal a system state (error / rate limit / encrypted / media) it can't // express as a model-generated answer. Best-effort: a failed reaction is logged, not // retried. Reactions are m.reaction (not m.room.message), so they never re-enter // handleMessage and need no anti-loop tracking. func (b *Bot) react(ctx context.Context, roomID, eventID, emoji string) { content := map[string]any{ "m.relates_to": map[string]any{ "rel_type": "m.annotation", "event_id": eventID, "key": emoji, }, } if _, err := b.mx.SendEvent(ctx, roomID, "m.reaction", content); err != nil { b.log.Error("react failed", "room", roomID, "emoji", emoji, "err", err) } } // reactEncryptedOnce reacts 🔒 to the first message seen in an encrypted room and // records a durable flag so a restart doesn't re-react (F5). Vojo disables E2EE by // default, so this is a near-dead safety path; the reaction is far less intrusive // than the old text notice, but the once-gate keeps it from annotating every message // in the rare encrypted room. func (b *Bot) reactEncryptedOnce(ctx context.Context, roomID, eventID string) { warned, err := b.st.HasWarnedEncrypted(roomID) if err != nil { b.log.Error("warned-flag read failed", "room", roomID, "err", err) return } if warned { return } b.react(ctx, roomID, eventID, reactEncrypted) if err := b.st.SetWarnedEncrypted(roomID); err != nil { b.log.Error("persist warned-flag failed", "room", roomID, "err", err) } } // sendReply sends the model's actual answer and records the completed exchange in the // conversation buffer so the next turn has context. func (b *Bot) sendReply(ctx context.Context, roomID string, trigger *Event, triggerMC *MessageContent, body string) { id := b.sendMessage(ctx, roomID, trigger, triggerMC, body) if id == "" { return } // Record the user trigger AND the assistant answer together, only AFTER the answer // was sent, so a failed or empty generation never leaves a dangling user turn (a // question with no reply) in the buffer — which would skew later completions. // Single-flight guarantees no other turn for this room interleaves between the two. b.appendBuf(roomID, bufferedMsg{sender: trigger.Sender, body: triggerMC.Body, isBot: false}) b.appendBuf(roomID, bufferedMsg{sender: b.cfg.BotMXID, body: body, isBot: true}) } // sendMessage builds and sends an m.notice reply, tracks our own event id, and returns // the new event id ("" on failure). func (b *Bot) sendMessage(ctx context.Context, roomID string, trigger *Event, triggerMC *MessageContent, body string) string { content := buildNoticeContent(trigger.EventID, trigger.Sender, triggerMC.RelatesTo, body) id, err := b.mx.SendEvent(ctx, roomID, "m.room.message", content) if err != nil { b.log.Error("send failed", "room", roomID, "err", err) return "" } // Track our own reply so a future reply-to-it is recognised as addressing us. b.botSent.Add(id) return id } // startTypingKeepalive starts the typing indicator and keeps it alive for the whole // generation (the CS-API server-side typing notification expires after the 30s we // pass, so we refresh every 20s). The returned stop clears the indicator and is safe // to call once via defer. Typing is best-effort UX — failures are non-fatal. func (b *Bot) startTypingKeepalive(ctx context.Context, roomID string) func() { b.setTyping(ctx, roomID, true) done := make(chan struct{}) go func() { t := time.NewTicker(20 * time.Second) defer t.Stop() for { select { case <-ctx.Done(): return case <-done: return case <-t.C: b.setTyping(ctx, roomID, true) } } }() var once sync.Once return func() { once.Do(func() { close(done) b.setTyping(ctx, roomID, false) }) } } // setTyping sets/clears the bot's typing indicator (best-effort UX; failures are // non-fatal). The 30s server-side timeout is refreshed by startTypingKeepalive. func (b *Bot) setTyping(ctx context.Context, roomID string, typing bool) { if err := b.mx.SendTyping(ctx, roomID, typing, 30000); err != nil { b.log.Debug("set typing failed", "room", roomID, "typing", typing, "err", err) } } // buildNoticeContent builds the reply. m.notice (not m.text) so the anti-loop // skip catches our own output. Thread-aware (F27): a trigger from a thread gets a // thread relation so the answer lands in the thread, not the main timeline. func buildNoticeContent(replyTo, sender string, triggerRelates *RelatesTo, body string) map[string]any { relates := map[string]any{} if triggerRelates != nil && triggerRelates.RelType == "m.thread" && triggerRelates.EventID != "" { relates["rel_type"] = "m.thread" relates["event_id"] = triggerRelates.EventID relates["is_falling_back"] = true relates["m.in_reply_to"] = map[string]any{"event_id": replyTo} } else { relates["m.in_reply_to"] = map[string]any{"event_id": replyTo} } content := map[string]any{ "msgtype": "m.notice", "body": body, "m.mentions": map[string]any{"user_ids": []string{sender}}, "m.relates_to": relates, } // The model answers in markdown; render it to org.matrix.custom.html so clients // show formatting instead of raw `**`, `#`, lists, code fences. Only attach // formatted_body when there's actual formatting — a plain answer keeps rendering // from `body` exactly as before. if html, formatted := markdownToHTML(body); formatted { content["format"] = matrixHTMLFormat content["formatted_body"] = html } return content } // --- per-room single-flight ---------------------------------------------------- // tryClaim marks a room as generating and returns true if the caller won the claim // (no generation was already in flight). The loser must drop its message. func (b *Bot) tryClaim(roomID string) bool { b.mu.Lock() defer b.mu.Unlock() if b.inflight[roomID] { return false } b.inflight[roomID] = true return true } func (b *Bot) release(roomID string) { b.mu.Lock() defer b.mu.Unlock() delete(b.inflight, roomID) } // --- per-room metadata helpers (all guarded by b.mu; probes run outside it) ----- // getMetaLocked returns (creating if needed) the room's meta. Caller MUST hold b.mu. func (b *Bot) getMetaLocked(roomID string) *roomMeta { m := b.meta[roomID] if m == nil { m = &roomMeta{} b.meta[roomID] = m } return m } func (b *Bot) invalidateCounts(roomID string) { b.mu.Lock() defer b.mu.Unlock() if m := b.meta[roomID]; m != nil { m.countsKnown = false } } func (b *Bot) setEncrypted(roomID string) { b.mu.Lock() defer b.mu.Unlock() m := b.getMetaLocked(roomID) m.encrypted, m.encKnown = true, true } func (b *Bot) forgetRoom(roomID string) { b.mu.Lock() defer b.mu.Unlock() delete(b.meta, roomID) delete(b.buf, roomID) delete(b.inflight, roomID) } // ensureEncryption returns whether the room is encrypted, probing the CS-API once // (without holding the lock) and caching the result. On probe error it returns false // (treated as not-encrypted this round) and leaves the state unknown for a re-probe. func (b *Bot) ensureEncryption(ctx context.Context, roomID string) bool { b.mu.Lock() if m := b.getMetaLocked(roomID); m.encKnown { enc := m.encrypted b.mu.Unlock() return enc } b.mu.Unlock() enc, err := b.mx.RoomEncrypted(ctx, roomID) if err != nil { b.log.Warn("encryption probe failed", "room", roomID, "err", err) return false // leave unknown; re-probed on the next message } // Re-fetch under the lock instead of writing to the pointer captured before the // unlocked probe: if the room was forgotten (leave/ban) mid-probe its meta was // deleted, and writing to the captured pointer would resurrect a dead room. b.mu.Lock() if m := b.meta[roomID]; m != nil { m.encrypted, m.encKnown = enc, true } b.mu.Unlock() return enc } // ensureCounts returns (countsKnown, isDM, foreign), probing /members once (without // holding the lock) and caching the result. On probe error it returns // (false, false, false): the caller treats an unclassified room conservatively and // logs a visible WARN rather than silently dropping. func (b *Bot) ensureCounts(ctx context.Context, roomID string) (countsKnown, isDM, foreign bool) { b.mu.Lock() known := b.getMetaLocked(roomID).countsKnown b.mu.Unlock() if !known { joined, invited, servers, err := b.mx.RoomMembership(ctx, roomID) if err != nil { b.log.Warn("member probe failed", "room", roomID, "err", err) return false, false, false } isForeign := false for s := range servers { if !b.cfg.AllowedServers[s] { isForeign = true break } } // Re-fetch under the lock rather than writing a pointer captured before the // unlocked /members probe (see ensureEncryption): a leave/ban mid-probe must // not be undone by resurrecting the room's meta. b.mu.Lock() if m := b.meta[roomID]; m != nil { m.joined, m.invited, m.foreign, m.countsKnown = joined, invited, isForeign, true } b.mu.Unlock() } b.mu.Lock() defer b.mu.Unlock() if m := b.meta[roomID]; m != nil { return m.countsKnown, m.isDM(), m.foreign } return false, false, false } func (b *Bot) snapshotBuf(roomID string) []bufferedMsg { b.mu.Lock() defer b.mu.Unlock() src := b.buf[roomID] if len(src) == 0 { return nil } out := make([]bufferedMsg, len(src)) copy(out, src) return out } func (b *Bot) appendBuf(roomID string, msg bufferedMsg) { b.mu.Lock() defer b.mu.Unlock() limit := b.cfg.MaxCtxEvent * 2 if limit < 8 { limit = 8 } buf := append(b.buf[roomID], msg) if len(buf) > limit { buf = buf[len(buf)-limit:] } b.buf[roomID] = buf }