MuyueWorkspace/internal/api/browsertest.go

package api

// Browser-test feature: an out-of-process page (the user's target tab)
// connects to Muyue via WebSocket using a short-lived token, and exposes a
// thin RPC: Studio's AI can list clickable elements, click them, evaluate JS,
// read the recent console buffer, and observe what changes after each action.
//
// Threat model: an injected snippet runs in the user's chosen page only, with
// the same origin as that page; the WS endpoint is bound to localhost and
// gated by a 5-minute token issued by the local Muyue server.

import (
	"context"
	"crypto/rand"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"net/http"
	"strings"
	"sync"
	"time"

	"github.com/gorilla/websocket"
	"github.com/muyue/muyue/internal/agent"
)

const (
	browserTestTokenTTL    = 5 * time.Minute
	browserTestCommandTTL  = 30 * time.Second
	browserTestConsoleMax  = 200
	browserTestSessionsMax = 16
)

// BrowserTestSession represents one connected browser tab.
type BrowserTestSession struct {
	ID         string
	URL        string
	Title      string
	conn       *websocket.Conn
	mu         sync.Mutex
	console    []ConsoleEntry
	pending    map[string]chan json.RawMessage
	pendingMu  sync.Mutex
	connectedAt time.Time
	writeMu    sync.Mutex
}

// ConsoleEntry is a captured console message from the connected page.
type ConsoleEntry struct {
	Level   string `json:"level"` // log, info, warn, error, debug
	Message string `json:"message"`
	Time    string `json:"time"`
}

// BrowserTestStore manages active sessions + pending one-shot connect tokens.
type BrowserTestStore struct {
	mu        sync.RWMutex
	sessions  map[string]*BrowserTestSession
	tokens    map[string]time.Time
	tokensMu  sync.Mutex
}

func NewBrowserTestStore() *BrowserTestStore {
	return &BrowserTestStore{
		sessions: map[string]*BrowserTestSession{},
		tokens:   map[string]time.Time{},
	}
}

// IssueToken creates a single-use token used by the snippet to authenticate.
func (s *BrowserTestStore) IssueToken() string {
	buf := make([]byte, 16)
	if _, err := rand.Read(buf); err != nil {
		return fmt.Sprintf("fallback-%d", time.Now().UnixNano())
	}
	tok := hex.EncodeToString(buf)
	s.tokensMu.Lock()
	now := time.Now()
	for k, v := range s.tokens {
		if now.Sub(v) > browserTestTokenTTL {
			delete(s.tokens, k)
		}
	}
	s.tokens[tok] = now
	s.tokensMu.Unlock()
	return tok
}

// ConsumeToken validates and removes a token in one step.
func (s *BrowserTestStore) ConsumeToken(tok string) bool {
	s.tokensMu.Lock()
	defer s.tokensMu.Unlock()
	t, ok := s.tokens[tok]
	if !ok {
		return false
	}
	delete(s.tokens, tok)
	return time.Since(t) <= browserTestTokenTTL
}

// Register inserts a new session, evicting the oldest if at capacity.
func (s *BrowserTestStore) Register(session *BrowserTestSession) {
	s.mu.Lock()
	defer s.mu.Unlock()
	if len(s.sessions) >= browserTestSessionsMax {
		var oldestID string
		var oldest time.Time
		for id, sess := range s.sessions {
			if oldestID == "" || sess.connectedAt.Before(oldest) {
				oldestID = id
				oldest = sess.connectedAt
			}
		}
		if old, ok := s.sessions[oldestID]; ok {
			old.conn.Close()
			delete(s.sessions, oldestID)
		}
	}
	s.sessions[session.ID] = session
}

func (s *BrowserTestStore) Remove(id string) {
	s.mu.Lock()
	defer s.mu.Unlock()
	if sess, ok := s.sessions[id]; ok {
		sess.conn.Close()
		delete(s.sessions, id)
	}
}

func (s *BrowserTestStore) Get(id string) *BrowserTestSession {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.sessions[id]
}

// Pick returns the requested session by ID, or the most-recently-connected
// session if id is empty. Returns nil if no session matches.
func (s *BrowserTestStore) Pick(id string) *BrowserTestSession {
	s.mu.RLock()
	defer s.mu.RUnlock()
	if id != "" {
		return s.sessions[id]
	}
	var picked *BrowserTestSession
	for _, sess := range s.sessions {
		if picked == nil || sess.connectedAt.After(picked.connectedAt) {
			picked = sess
		}
	}
	return picked
}

func (s *BrowserTestStore) List() []map[string]interface{} {
	s.mu.RLock()
	defer s.mu.RUnlock()
	out := make([]map[string]interface{}, 0, len(s.sessions))
	for _, sess := range s.sessions {
		out = append(out, map[string]interface{}{
			"id":           sess.ID,
			"url":          sess.URL,
			"title":        sess.Title,
			"connected_at": sess.connectedAt.Format(time.RFC3339),
		})
	}
	return out
}

// Send issues an RPC command to the browser session and waits up to TTL for
// the matching reply. Returns the raw payload or an error.
func (sess *BrowserTestSession) Send(action string, params map[string]interface{}) (json.RawMessage, error) {
	cid := newCorrelationID()
	ch := make(chan json.RawMessage, 1)
	sess.pendingMu.Lock()
	sess.pending[cid] = ch
	sess.pendingMu.Unlock()
	defer func() {
		sess.pendingMu.Lock()
		delete(sess.pending, cid)
		sess.pendingMu.Unlock()
	}()

	cmd := map[string]interface{}{
		"id":     cid,
		"action": action,
		"params": params,
	}
	sess.writeMu.Lock()
	err := sess.conn.WriteJSON(cmd)
	sess.writeMu.Unlock()
	if err != nil {
		return nil, fmt.Errorf("write: %w", err)
	}

	select {
	case payload := <-ch:
		return payload, nil
	case <-time.After(browserTestCommandTTL):
		return nil, fmt.Errorf("browser session did not reply within %s", browserTestCommandTTL)
	}
}

// AppendConsole records a console line, trimming to the buffer cap.
func (sess *BrowserTestSession) AppendConsole(level, message string) {
	sess.mu.Lock()
	defer sess.mu.Unlock()
	sess.console = append(sess.console, ConsoleEntry{
		Level:   level,
		Message: message,
		Time:    time.Now().Format(time.RFC3339),
	})
	if len(sess.console) > browserTestConsoleMax {
		sess.console = sess.console[len(sess.console)-browserTestConsoleMax:]
	}
}

// SnapshotConsole returns a copy of the current console buffer.
func (sess *BrowserTestSession) SnapshotConsole() []ConsoleEntry {
	sess.mu.Lock()
	defer sess.mu.Unlock()
	out := make([]ConsoleEntry, len(sess.console))
	copy(out, sess.console)
	return out
}

func newCorrelationID() string {
	buf := make([]byte, 8)
	rand.Read(buf)
	return hex.EncodeToString(buf)
}

// HTTP handlers --------------------------------------------------------------

func (s *Server) handleBrowserTestSnippet(w http.ResponseWriter, r *http.Request) {
	if r.Method != "GET" {
		writeError(w, "GET only", http.StatusMethodNotAllowed)
		return
	}
	tok := s.browserTestStore.IssueToken()
	host := r.Host
	if host == "" {
		host = "127.0.0.1"
	}
	scheme := "ws"
	if r.TLS != nil {
		scheme = "wss"
	}
	wsURL := fmt.Sprintf("%s://%s/api/ws/browser-test?token=%s", scheme, host, tok)
	snippet := buildBrowserTestSnippet(wsURL)
	writeJSON(w, map[string]interface{}{
		"token":      tok,
		"ws_url":     wsURL,
		"snippet":    snippet,
		"expires_in": int(browserTestTokenTTL / time.Second),
	})
}

func (s *Server) handleBrowserTestSessions(w http.ResponseWriter, r *http.Request) {
	if r.Method != "GET" {
		writeError(w, "GET only", http.StatusMethodNotAllowed)
		return
	}
	writeJSON(w, map[string]interface{}{
		"sessions": s.browserTestStore.List(),
	})
}

func (s *Server) handleBrowserTestConsole(w http.ResponseWriter, r *http.Request) {
	if r.Method != "GET" {
		writeError(w, "GET only", http.StatusMethodNotAllowed)
		return
	}
	id := strings.TrimPrefix(r.URL.Path, "/api/test/console/")
	sess := s.browserTestStore.Pick(id)
	if sess == nil {
		writeError(w, "no active browser test session", http.StatusNotFound)
		return
	}
	writeJSON(w, map[string]interface{}{
		"session_id": sess.ID,
		"url":        sess.URL,
		"console":    sess.SnapshotConsole(),
	})
}

// browserTestUpgrader accepts any origin: the connection is gated by a
// short-lived token issued to the local UI, not by Origin checking.
var browserTestUpgrader = websocket.Upgrader{
	CheckOrigin: func(r *http.Request) bool { return true },
}

func (s *Server) handleBrowserTestWS(w http.ResponseWriter, r *http.Request) {
	tok := r.URL.Query().Get("token")
	if tok == "" || !s.browserTestStore.ConsumeToken(tok) {
		writeError(w, "invalid or expired token", http.StatusUnauthorized)
		return
	}
	conn, err := browserTestUpgrader.Upgrade(w, r, nil)
	if err != nil {
		return
	}
	conn.SetReadLimit(2 << 20)

	// Read the hello message: page sends {"type":"hello","url":"...","title":"..."}.
	conn.SetReadDeadline(time.Now().Add(10 * time.Second))
	var hello struct {
		Type  string `json:"type"`
		URL   string `json:"url"`
		Title string `json:"title"`
	}
	if err := conn.ReadJSON(&hello); err != nil || hello.Type != "hello" {
		conn.WriteJSON(map[string]string{"type": "error", "message": "expected hello"})
		conn.Close()
		return
	}
	conn.SetReadDeadline(time.Time{})

	id := newCorrelationID()
	sess := &BrowserTestSession{
		ID:          id,
		URL:         hello.URL,
		Title:       hello.Title,
		conn:        conn,
		pending:     map[string]chan json.RawMessage{},
		connectedAt: time.Now(),
	}
	s.browserTestStore.Register(sess)
	defer s.browserTestStore.Remove(id)

	// Acknowledge with the assigned session ID.
	sess.writeMu.Lock()
	conn.WriteJSON(map[string]string{"type": "registered", "session_id": id})
	sess.writeMu.Unlock()

	for {
		_, raw, err := conn.ReadMessage()
		if err != nil {
			return
		}
		var msg struct {
			Type  string          `json:"type"`
			ID    string          `json:"id,omitempty"`
			Level string          `json:"level,omitempty"`
			Text  string          `json:"text,omitempty"`
			URL   string          `json:"url,omitempty"`
			Data  json.RawMessage `json:"data,omitempty"`
		}
		if err := json.Unmarshal(raw, &msg); err != nil {
			continue
		}
		switch msg.Type {
		case "console":
			sess.AppendConsole(msg.Level, msg.Text)
		case "url_change":
			sess.mu.Lock()
			sess.URL = msg.URL
			sess.mu.Unlock()
		case "reply":
			sess.pendingMu.Lock()
			ch, ok := sess.pending[msg.ID]
			sess.pendingMu.Unlock()
			if ok {
				select {
				case ch <- msg.Data:
				default:
				}
			}
		case "ping":
			sess.writeMu.Lock()
			conn.WriteJSON(map[string]string{"type": "pong"})
			sess.writeMu.Unlock()
		}
	}
}

// Agent tool -----------------------------------------------------------------

// BrowserTestParams is the schema exposed to the AI for the browser_test tool.
type BrowserTestParams struct {
	Action    string `json:"action" description:"One of: list_clickables, click, eval, console, current_url, wait, type, summary"`
	SessionID string `json:"session_id,omitempty" description:"Browser session id (optional, defaults to most recent)"`
	Selector  string `json:"selector,omitempty" description:"CSS selector for click/type actions"`
	Index     int    `json:"index,omitempty" description:"Alternative to selector: index into the last list_clickables result (0-based)"`
	Expr      string `json:"expr,omitempty" description:"JS expression to evaluate (eval action only)"`
	Text      string `json:"text,omitempty" description:"Text to type (type action only)"`
	WaitMs    int    `json:"wait_ms,omitempty" description:"Milliseconds to wait (wait action only, max 5000)"`
	Tail      int    `json:"tail,omitempty" description:"Console action: how many recent lines to return (default 50, max 200)"`
}

// RegisterBrowserTestTool wires the agent tool against a session store.
func RegisterBrowserTestTool(reg *agent.Registry, store *BrowserTestStore) error {
	tool, err := agent.NewTool("browser_test",
		"Drive the user's connected browser tab for end-to-end testing. Available actions: list_clickables (returns indexed clickable elements), click (by selector or index), eval (run a JS expression and return result), console (read recent console output, ideal to spot errors after a click), current_url, wait (sleep ms before next check), type (set value on an input), summary (URL+title+last console entries). Always start with list_clickables; click; then console to verify no errors.",
		func(ctx context.Context, p BrowserTestParams) (agent.ToolResponse, error) {
			sess := store.Pick(p.SessionID)
			if sess == nil {
				return agent.TextErrorResponse("no active browser session — ask the user to paste the snippet from the Tests tab in their target page"), nil
			}

			action := strings.ToLower(strings.TrimSpace(p.Action))
			switch action {
			case "":
				return agent.TextErrorResponse("action is required"), nil
			case "list_clickables", "click", "eval", "current_url", "type":
			case "console", "summary", "wait":
			default:
				return agent.TextErrorResponse("unknown action: " + p.Action), nil
			}

			if action == "console" {
				tail := p.Tail
				if tail <= 0 {
					tail = 50
				}
				if tail > browserTestConsoleMax {
					tail = browserTestConsoleMax
				}
				entries := sess.SnapshotConsole()
				if len(entries) > tail {
					entries = entries[len(entries)-tail:]
				}
				out, _ := json.MarshalIndent(map[string]interface{}{
					"session_id": sess.ID,
					"console":    entries,
				}, "", "  ")
				return agent.TextResponse(string(out)), nil
			}

			if action == "summary" {
				entries := sess.SnapshotConsole()
				if len(entries) > 20 {
					entries = entries[len(entries)-20:]
				}
				out, _ := json.MarshalIndent(map[string]interface{}{
					"session_id":      sess.ID,
					"url":             sess.URL,
					"title":           sess.Title,
					"recent_console":  entries,
				}, "", "  ")
				return agent.TextResponse(string(out)), nil
			}

			if action == "wait" {
				ms := p.WaitMs
				if ms <= 0 {
					ms = 200
				}
				if ms > 5000 {
					ms = 5000
				}
				select {
				case <-ctx.Done():
					return agent.TextErrorResponse("cancelled"), nil
				case <-time.After(time.Duration(ms) * time.Millisecond):
				}
				return agent.TextResponse(fmt.Sprintf("waited %dms", ms)), nil
			}

			// Capture console snapshot length before so we can return only the delta
			// after the action — useful so the AI can spot errors caused by the click.
			pre := len(sess.SnapshotConsole())

			params := map[string]interface{}{}
			if p.Selector != "" {
				params["selector"] = p.Selector
			}
			if p.Index > 0 || (action == "click" && p.Selector == "") {
				params["index"] = p.Index
			}
			if p.Expr != "" {
				params["expr"] = p.Expr
			}
			if p.Text != "" {
				params["text"] = p.Text
			}

			payload, err := sess.Send(action, params)
			if err != nil {
				return agent.TextErrorResponse(err.Error()), nil
			}

			// Console delta: messages logged during this command.
			post := sess.SnapshotConsole()
			var delta []ConsoleEntry
			if len(post) > pre {
				delta = post[pre:]
			}

			result := map[string]interface{}{
				"action":         action,
				"reply":          json.RawMessage(payload),
				"console_delta":  delta,
				"current_url":    sess.URL,
			}
			out, _ := json.MarshalIndent(result, "", "  ")
			return agent.TextResponse(string(out)), nil
		})
	if err != nil {
		return err
	}
	return reg.Register(tool)
}

// Snippet generator ----------------------------------------------------------

func buildBrowserTestSnippet(wsURL string) string {
	// Note: this is the JS injected into the user's target page. It opens the
	// WS, hooks console, and dispatches commands. Kept terse on purpose.
	return `(function(){
  if (window.__muyueTestRunner) { console.log('[Muyue] runner already attached'); return; }
  var WS_URL = ` + jsString(wsURL) + `;
  var ws = new WebSocket(WS_URL);
  var lastList = [];
  function send(obj){ try{ ws.send(JSON.stringify(obj)); }catch(e){} }
  function reply(id, data){ send({type:'reply', id:id, data:data}); }
  function safeText(el){
    var t = (el.innerText || el.textContent || '').trim();
    if (t.length > 80) t = t.slice(0,80)+'…';
    return t;
  }
  function describe(el){
    var sel = el.id ? '#'+el.id : el.tagName.toLowerCase();
    if (!el.id && el.className && typeof el.className === 'string') {
      sel += '.' + el.className.trim().split(/\s+/).slice(0,2).join('.');
    }
    var label = el.getAttribute('aria-label') || el.getAttribute('title') || el.getAttribute('name') || '';
    return { tag: el.tagName.toLowerCase(), selector: sel, text: safeText(el), label: label, type: el.getAttribute('type')||'', disabled: !!el.disabled };
  }
  function list(){
    var els = Array.from(document.querySelectorAll('button, a[href], input[type=submit], input[type=button], [role=button], [onclick]'));
    lastList = els.filter(function(e){ var r=e.getBoundingClientRect(); return r.width>0 && r.height>0; });
    return lastList.map(describe).map(function(d,i){ d.index = i; return d; });
  }
  function clickEl(el){
    if (!el) return { ok:false, error:'element not found' };
    if (el.disabled) return { ok:false, error:'element is disabled' };
    try { el.scrollIntoView({block:'center'}); el.click(); return { ok:true }; }
    catch(e){ return { ok:false, error:String(e) }; }
  }
  function dispatch(msg){
    var p = msg.params || {};
    switch(msg.action){
      case 'list_clickables': return list();
      case 'click': {
        var el;
        if (p.selector) el = document.querySelector(p.selector);
        else if (typeof p.index === 'number') el = lastList[p.index];
        return clickEl(el);
      }
      case 'eval': {
        try { var r = (0,eval)(p.expr); return { ok:true, value: serialize(r) }; }
        catch(e){ return { ok:false, error:String(e) }; }
      }
      case 'current_url': return { url: location.href, title: document.title };
      case 'type': {
        var el = p.selector ? document.querySelector(p.selector) : (lastList[p.index]);
        if (!el) return { ok:false, error:'element not found' };
        var proto = Object.getPrototypeOf(el);
        var setter = Object.getOwnPropertyDescriptor(proto, 'value');
        try { setter && setter.set ? setter.set.call(el, p.text||'') : (el.value = p.text||''); }
        catch(e){ el.value = p.text||''; }
        el.dispatchEvent(new Event('input', {bubbles:true}));
        el.dispatchEvent(new Event('change', {bubbles:true}));
        return { ok:true };
      }
    }
    return { ok:false, error:'unknown action' };
  }
  function serialize(v){
    if (v === undefined) return 'undefined';
    try { return JSON.parse(JSON.stringify(v)); }
    catch(e){ return String(v); }
  }
  ['log','info','warn','error','debug'].forEach(function(lvl){
    var orig = console[lvl];
    console[lvl] = function(){
      try {
        var parts = Array.from(arguments).map(function(a){
          if (typeof a === 'string') return a;
          try { return JSON.stringify(a); } catch(e){ return String(a); }
        });
        send({type:'console', level: lvl, text: parts.join(' ')});
      } catch(e){}
      return orig.apply(console, arguments);
    };
  });
  window.addEventListener('error', function(e){
    send({type:'console', level:'error', text:'window.onerror: '+(e.message||e.error||'unknown')});
  });
  window.addEventListener('unhandledrejection', function(e){
    send({type:'console', level:'error', text:'unhandledrejection: '+String(e.reason)});
  });
  var lastUrl = location.href;
  setInterval(function(){
    if (location.href !== lastUrl){ lastUrl = location.href; send({type:'url_change', url: lastUrl}); }
  }, 500);
  ws.onopen = function(){ send({type:'hello', url: location.href, title: document.title}); };
  ws.onmessage = function(ev){
    try { var msg = JSON.parse(ev.data); }
    catch(e){ return; }
    if (msg.type === 'registered') { console.log('[Muyue] connected — session', msg.session_id); return; }
    if (msg.action) reply(msg.id, dispatch(msg));
  };
  ws.onclose = function(){ console.log('[Muyue] runner disconnected'); window.__muyueTestRunner = null; };
  window.__muyueTestRunner = { ws: ws, list: list };
})();`
}

func jsString(s string) string {
	b, _ := json.Marshal(s)
	return string(b)
}