fix: Думать - the model is reminded in the question itself, the label no longer lands in the arrow, and the thinking with its time is kept to show again later

internal/server/ai.go

@@ -41,6 +41,8 @@ type aiConfig struct{ Type, URL, Key, Model string }
 type aiMsg struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
+	// ThinkMS: how long the model thought before this answer (kept to show it).
+	ThinkMS int64 `json:"think_ms,omitempty"`
 }
 
 func (s *Server) aiConfig() aiConfig {
@@ -440,6 +442,11 @@ func (s *Server) aiDeleteChat(owner, chat int64) {
 
 // aiAppendTo keeps a message of owner's chat; the first question names it.
 func (s *Server) aiAppendTo(owner, chat int64, role, content string) {
+	s.aiKeepMsg(owner, chat, aiMsg{Role: role, Content: content})
+}
+
+func (s *Server) aiKeepMsg(owner, chat int64, m aiMsg) {
+	role, content := m.Role, m.Content
 	if content == "" || !s.aiOwns(owner, chat) {
 		return
 	}
@@ -447,7 +454,7 @@ func (s *Server) aiAppendTo(owner, chat int64, role, content string) {
 		content = utf8Safe(content[:64<<10])
 	}
 	now := time.Now().Unix()
-	_, _ = s.db.sql.Exec(`INSERT INTO ai_messages (admin_id, chat_id, role, content, at) VALUES (?, ?, ?, ?, ?)`, owner, chat, role, content, now)
+	_, _ = s.db.sql.Exec(`INSERT INTO ai_messages (admin_id, chat_id, role, content, at, think_ms) VALUES (?, ?, ?, ?, ?, ?)`, owner, chat, role, content, now, m.ThinkMS)
 	_, _ = s.db.sql.Exec(`DELETE FROM ai_messages WHERE chat_id = ? AND id <= (SELECT id FROM ai_messages WHERE chat_id = ? ORDER BY id DESC LIMIT 1 OFFSET ?)`, chat, chat, aiKeep)
 	_, _ = s.db.sql.Exec(`UPDATE ai_chats SET updated = ? WHERE id = ?`, now, chat)
 	if role == "user" {
@@ -461,7 +468,7 @@ func (s *Server) aiAppendTo(owner, chat int64, role, content string) {
 
 // aiHistoryOf is a chat's talk, oldest first, the last limit messages.
 func (s *Server) aiHistoryOf(owner, chat int64, limit int) []aiMsg {
-	rows, err := s.db.sql.Query(`SELECT role, content FROM ai_messages WHERE admin_id = ? AND chat_id = ? ORDER BY id DESC LIMIT ?`, owner, chat, limit)
+	rows, err := s.db.sql.Query(`SELECT role, content, think_ms FROM ai_messages WHERE admin_id = ? AND chat_id = ? ORDER BY id DESC LIMIT ?`, owner, chat, limit)
 	out := []aiMsg{}
 	if err != nil {
 		return out
@@ -469,7 +476,7 @@ func (s *Server) aiHistoryOf(owner, chat int64, limit int) []aiMsg {
 	defer rows.Close()
 	for rows.Next() {
 		var m aiMsg
-		if rows.Scan(&m.Role, &m.Content) == nil {
+		if rows.Scan(&m.Role, &m.Content, &m.ThinkMS) == nil {
 			out = append(out, m)
 		}
 	}
@@ -541,24 +548,44 @@ func (s *Server) handleAIChat(w http.ResponseWriter, r *http.Request) {
 	}
 	if in.Think {
 		system += aiThink
+		// Models heed the last words most: the reminder goes there too (not
+		// kept with the talk).
+		if n := len(msgs); n > 0 && msgs[n-1].Role == "user" {
+			msgs[n-1].Content += "\n\n(Сначала подумай: первая строка ответа - ::думаю::, дальше рассуждение, потом строка ::ответ:: и сам ответ.)"
+		}
 	}
 	w.Header().Set("Content-Type", "text/plain; charset=utf-8")
 	w.Header().Set("Cache-Control", "no-store")
 	w.Header().Set("X-Accel-Buffering", "no")
 	fl, _ := w.(http.Flusher)
 	var answer strings.Builder
+	var thinkFrom, thinkTo time.Time // when the thinking began and ended
 	emit := func(t string) {
 		if t == "" {
 			return
 		}
 		answer.WriteString(t)
+		if a := answer.String(); thinkFrom.IsZero() && (strings.Contains(a, "::думаю::") || strings.Contains(a, "<think>")) {
+			thinkFrom = time.Now()
+		} else if !thinkFrom.IsZero() && thinkTo.IsZero() && (strings.Contains(a, "::ответ::") || strings.Contains(a, "</think>")) {
+			thinkTo = time.Now()
+		}
 		_, _ = io.WriteString(w, t)
 		if fl != nil {
 			fl.Flush()
 		}
 	}
 	err := aiStream(r.Context(), s.aiConfig(), system+"\n\n--- Панель сейчас ---\n"+in.Context, msgs, emit)
-	s.aiAppendTo(a.ID, in.Chat, "assistant", stripThink(answer.String()))
+	// The answer is kept whole, thinking too (to show it again); the model
+	// gets it without (aiWindow).
+	ans := aiMsg{Role: "assistant", Content: answer.String()}
+	if !thinkFrom.IsZero() {
+		if thinkTo.IsZero() {
+			thinkTo = time.Now()
+		}
+		ans.ThinkMS = thinkTo.Sub(thinkFrom).Milliseconds()
+	}
+	s.aiKeepMsg(a.ID, in.Chat, ans)
 	if err != nil && r.Context().Err() == nil {
 		if answer.Len() == 0 {
 			w.WriteHeader(http.StatusBadGateway)
@@ -679,7 +706,9 @@ func aiWindow(rows []aiMsg, budget int) []aiMsg {
 			}
 			m = aiMsg{Role: "user", Content: "Результат действия:\n" + cut(m.Content, max)}
 		case m.Role == "assistant" && i < len(rows)-2:
-			m.Content = cut(m.Content, 3000)
+			m.Content = cut(stripThink(m.Content), 3000)
+		case m.Role == "assistant":
+			m.Content = stripThink(m.Content)
 		case m.Role != "assistant":
 			m.Role = "user"
 		}

internal/server/ai_test.go

@@ -10,6 +10,7 @@ import (
 	"strconv"
 	"strings"
 	"testing"
+	"time"
 )
 
 func collect(t *testing.T, cfg aiConfig) (string, error) {
@@ -235,3 +236,33 @@ func TestAIChats(t *testing.T) {
 		t.Fatal("another user wrote into the talk")
 	}
 }
+
+// The thinking is kept with the answer (and how long it took), so it shows
+// again later; the model never gets it back.
+func TestAIThinkingKept(t *testing.T) {
+	s, c := newTestServer(t)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		fl := w.(http.Flusher)
+		fmt.Fprint(w, "data: {\"choices\":[{\"delta\":{\"content\":\"::думаю::\\nхм\\n\"}}]}\n\n")
+		fl.Flush()
+		time.Sleep(1100 * time.Millisecond)
+		fmt.Fprint(w, "data: {\"choices\":[{\"delta\":{\"content\":\"::ответ::\\n391\"}}]}\n\ndata: [DONE]\n\n")
+	}))
+	defer srv.Close()
+	_ = s.db.SetConfig(cfgAIType, "openai")
+	_ = s.db.SetConfig(cfgAIURL, srv.URL)
+	a, _ := s.db.AdminByUsername("adm")
+	_ = s.db.SetConfig(aiUserKey(a.ID), "1")
+	req := httptest.NewRequest(http.MethodPost, "/api/panel/ai/chat", strings.NewReader(`{"message":"17*23","think":true}`))
+	req.AddCookie(c)
+	w := httptest.NewRecorder()
+	s.Handler().ServeHTTP(w, req)
+	chat, _ := strconv.ParseInt(w.Header().Get("X-Chat-Id"), 10, 64)
+	h := s.aiHistoryOf(a.ID, chat, 10)
+	if len(h) != 2 || !strings.Contains(h[1].Content, "::думаю::") || h[1].ThinkMS < 1000 {
+		t.Fatalf("kept: %+v", h)
+	}
+	if win := aiWindow(h, 10000); strings.Contains(win[len(win)-1].Content, "хм") {
+		t.Fatalf("the model gets the thinking back: %+v", win)
+	}
+}

internal/server/db.go

@@ -129,6 +129,7 @@ CREATE TABLE IF NOT EXISTS forward_rules (
 		    role TEXT NOT NULL, content TEXT NOT NULL, at INTEGER NOT NULL)`,
 		`CREATE INDEX IF NOT EXISTS ai_messages_admin ON ai_messages(admin_id, id)`,
 		`ALTER TABLE ai_messages ADD COLUMN chat_id INTEGER NOT NULL DEFAULT 0`,
+		`ALTER TABLE ai_messages ADD COLUMN think_ms INTEGER NOT NULL DEFAULT 0`,
 		`CREATE TABLE IF NOT EXISTS ai_chats (
 		    id INTEGER PRIMARY KEY AUTOINCREMENT,
 		    admin_id INTEGER NOT NULL REFERENCES admins(id) ON DELETE CASCADE,

internal/server/web/static/ai.js

@@ -207,7 +207,7 @@
       if (chatId !== id) return;
       h.forEach(m => {
         if (m.role === 'result') showResult(m.content);
-        else { const el = add(m.role, m.content); el.querySelectorAll('.ai-act').forEach(b => { b.disabled = true; }); }
+        else { const el = add(m.role, m.content, m.think_ms); el.querySelectorAll('.ai-act').forEach(b => { b.disabled = true; }); }
       });
       if (!h.length) q('.ai-hello').hidden = false;
     } catch (e) { /* stays empty */ }
@@ -365,11 +365,12 @@
   // Markdown; ``` blocks apart, action blocks with their button. The
   // thinking block is made once and only its words change, so it stays open
   // or shut as the user left it while the answer streams.
-  const CHEV = '<svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="m9 6 6 6-6 6"/></svg>';
+  // The arrow of the + menu's groups: right shut, down open.
+  const CHEV = '<span class="chev">›</span>';
   function render(el, text, done) {
     let think = el.querySelector(':scope > .ai-think'), body = el.querySelector(':scope > .ai-body');
     if (!body) {
-      el.innerHTML = '<div class="ai-think" hidden><button type="button" class="ai-think-h">' + CHEV + '<span></span></button>' +
+      el.innerHTML = '<div class="ai-think" hidden><button type="button" class="ai-think-h">' + CHEV + '<span class="lbl"></span></button>' +
         '<div class="ai-think-b"><div class="ai-think-in"></div></div></div><div class="ai-body"></div>';
       think = el.querySelector('.ai-think'); body = el.querySelector('.ai-body');
       think.querySelector('.ai-think-h').onclick = () => think.classList.toggle('open');
@@ -381,8 +382,9 @@
       if (!going && !el._t1) el._t1 = Date.now();
       think.hidden = false;
       think.classList.toggle('going', going);
-      think.querySelector('.ai-think-h span').textContent = going ? 'Думаю…'
-        : 'Думал ' + Math.max(1, Math.round((el._t1 - el._t0) / 1000)) + ' с';
+      // How long: timed here as it streams, or kept with the answer.
+      const ms = el._thinkMs || (el._t1 - el._t0);
+      think.querySelector('.ai-think-h .lbl').textContent = going ? 'Думаю…' : 'Думал ' + Math.max(1, Math.round(ms / 1000)) + ' с';
       think.querySelector('.ai-think-in').innerHTML = WSUtil.md(th[1].trim());
       text = text.slice(th[0].length);
     }
@@ -407,11 +409,11 @@
     body.innerHTML = html;
     el.dataset.text = text;
   }
-  function add(role, text) {
+  function add(role, text, thinkMs) {
     q('.ai-hello').hidden = true;
     const el = document.createElement('div');
     el.className = 'ai-msg ' + role;
-    if (role === 'assistant') render(el, text, true); else el.textContent = text;
+    if (role === 'assistant') { el._thinkMs = thinkMs || 0; render(el, text, true); } else el.textContent = text;
     list.appendChild(el);
     list.scrollTop = list.scrollHeight;
     return el;

internal/server/web/static/app.css

@@ -943,9 +943,9 @@ main#app > .view:not([hidden]) { animation:view-in .22s ease; } /* no fill: a la
 .ai-think[hidden] { display:none; }
 .ai-think-h { display:inline-flex; align-items:center; gap:5px; padding:2px 4px 2px 0; background:none; border:0; color:var(--muted); font-size:12.5px; cursor:pointer; }
 .ai-think-h:hover { color:var(--text); }
-.ai-think-h svg { transition:transform .2s ease; }
-.ai-think.open .ai-think-h svg { transform:rotate(90deg); }
-.ai-think.going .ai-think-h span { background:linear-gradient(90deg, var(--muted) 30%, var(--text) 50%, var(--muted) 70%) 0 0/200% 100%;
+.ai-think-h .chev { display:inline-block; width:10px; text-align:center; font-size:15px; line-height:1; transition:transform .2s ease; }
+.ai-think.open .ai-think-h .chev { transform:rotate(90deg); }
+.ai-think.going .ai-think-h .lbl { background:linear-gradient(90deg, var(--muted) 30%, var(--text) 50%, var(--muted) 70%) 0 0/200% 100%;
   -webkit-background-clip:text; background-clip:text; color:transparent; animation:ai-shine 1.4s linear infinite; }
 @keyframes ai-shine { to { background-position:-200% 0; } }
 .ai-think-b { display:grid; grid-template-rows:0fr; transition:grid-template-rows .22s ease, opacity .2s ease; opacity:0; }