feat: workspaces, agentic action tools, local Whisper STT, vec recall

- Projects/workspaces: documents grouped into projects; chat RAG scopes to the
  active project. Switcher in the Documents page.
- Agentic action tools: web_search, fetch_url, and remember (first write tool),
  allowlist-gated per playbook.
- Local Whisper STT (faster-whisper, no torch): on-device dictation replacing
  the browser Web Speech API. POST /stt + GET /stt/status; browser fallback.
- Vector index extended to conversation recall (message_vectors), with the
  brute-force cosine scan kept as the fallback.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

interface/web/src/Chatbot.jsx

@@ -19,23 +19,64 @@ export function Chatbot({ conversationId, setConversationId, onConversationChang
   const [editingIdx, setEditingIdx] = useState(null); // user message being edited
   const [editText, setEditText] = useState("");
   const [listening, setListening] = useState(false); // mic dictation active
+  const [transcribing, setTranscribing] = useState(false); // local STT running
+  const [sttLocal, setSttLocal] = useState(false); // backend Whisper available
   const [speakingIdx, setSpeakingIdx] = useState(null); // message being read aloud
   const recognitionRef = useRef(null);
+  const mediaRecRef = useRef(null);
 
   // Web Speech API — browser-native, no backend/model. Absent on unsupported browsers.
   const SpeechRec = typeof window !== "undefined" && (window.SpeechRecognition || window.webkitSpeechRecognition);
   const ttsSupported = typeof window !== "undefined" && "speechSynthesis" in window;
+  const canRecord = typeof navigator !== "undefined" && navigator.mediaDevices && window.MediaRecorder;
 
-  const toggleMic = () => {
-    if (!SpeechRec) return;
-    if (listening) { recognitionRef.current?.stop(); return; }
+  // Prefer local Whisper (on-device) over browser speech (Chrome routes audio to Google).
+  useEffect(() => {
+    fetch(`${API_BASE}/stt/status`).then(r => r.ok ? r.json() : null)
+      .then(d => setSttLocal(Boolean(d && d.available))).catch(() => {});
+  }, []);
+
+  const _appendTranscript = (text) => {
+    if (text) setInput(prev => (prev ? prev + " " : "") + text);
+  };
+
+  // Local path: record audio, POST to /stt (faster-whisper transcribes on-device).
+  const startLocalDictation = async () => {
+    try {
+      const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
+      const rec = new MediaRecorder(stream);
+      const chunks = [];
+      rec.ondataavailable = e => e.data.size && chunks.push(e.data);
+      rec.onstop = async () => {
+        stream.getTracks().forEach(t => t.stop());
+        setListening(false);
+        setTranscribing(true);
+        try {
+          const b64 = await new Promise((res) => {
+            const fr = new FileReader();
+            fr.onload = () => res(String(fr.result).split(",")[1]);
+            fr.readAsDataURL(new Blob(chunks, { type: rec.mimeType }));
+          });
+          const r = await fetch(`${API_BASE}/stt`, {
+            method: "POST", headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({ audio: b64 }),
+          });
+          if (r.ok) _appendTranscript((await r.json()).text);
+        } catch { /* ignore */ } finally {
+          setTranscribing(false);
+        }
+      };
+      mediaRecRef.current = rec;
+      setListening(true);
+      rec.start();
+    } catch { setListening(false); }
+  };
+
+  const startBrowserDictation = () => {
     const rec = new SpeechRec();
     rec.lang = "en-US";
     rec.interimResults = false;
-    rec.onresult = (e) => {
-      const text = Array.from(e.results).map(r => r[0].transcript).join(" ").trim();
-      if (text) setInput(prev => (prev ? prev + " " : "") + text);
-    };
+    rec.onresult = (e) => _appendTranscript(Array.from(e.results).map(r => r[0].transcript).join(" ").trim());
     rec.onend = () => setListening(false);
     rec.onerror = () => setListening(false);
     recognitionRef.current = rec;
@@ -43,6 +84,17 @@ export function Chatbot({ conversationId, setConversationId, onConversationChang
     rec.start();
   };
 
+  const toggleMic = () => {
+    if (transcribing) return;
+    if (listening) {
+      if (sttLocal && mediaRecRef.current) mediaRecRef.current.stop();
+      else recognitionRef.current?.stop();
+      return;
+    }
+    if (sttLocal && canRecord) startLocalDictation();
+    else if (SpeechRec) startBrowserDictation();
+  };
+
   // Read a reply aloud (light markdown strip so symbols aren't spoken).
   const speak = (idx, text) => {
     if (!ttsSupported) return;
@@ -617,10 +669,11 @@ export function Chatbot({ conversationId, setConversationId, onConversationChang
             📎
             <input type="file" accept="image/*" multiple onChange={onImagePick} disabled={loading} style={{ display: "none" }} />
           </label>
-          {SpeechRec && (
-            <button onClick={toggleMic} title={listening ? "Stop dictation" : "Dictate (speech to text)"}
-              style={{ padding: "0.9rem", background: listening ? "#3a1a1a" : "#222", border: "1px solid " + (listening ? "#c0392b" : "#333"), borderRadius: "10px", cursor: "pointer", color: listening ? "#ff8a80" : "#ccc" }}>
-              {listening ? "🔴" : "🎤"}
+          {((sttLocal && canRecord) || SpeechRec) && (
+            <button onClick={toggleMic}
+              title={transcribing ? "Transcribing…" : listening ? "Stop dictation" : sttLocal ? "Dictate (on-device Whisper)" : "Dictate (browser speech)"}
+              style={{ padding: "0.9rem", background: listening ? "#3a1a1a" : "#222", border: "1px solid " + (listening ? "#c0392b" : "#333"), borderRadius: "10px", cursor: transcribing ? "default" : "pointer", color: listening ? "#ff8a80" : "#ccc" }}>
+              {transcribing ? "⏳" : listening ? "🔴" : "🎤"}
             </button>
           )}
           <textarea

interface/web/src/Documents.jsx

@@ -10,6 +10,43 @@ export function Documents() {
   const [busy, setBusy] = useState(false);
   const [message, setMessage] = useState("");
   const [viewing, setViewing] = useState(null); // {title, chunks} being previewed
+  const [projects, setProjects] = useState([]);
+  const [activeProject, setActiveProject] = useState(""); // "" = All
+
+  const loadProjects = async () => {
+    try {
+      const r = await fetch(`${API_BASE}/projects`);
+      if (r.ok) { const d = await r.json(); setProjects(d.projects || []); setActiveProject(d.active || ""); }
+    } catch { /* offline */ }
+  };
+
+  // Switch the active workspace (persisted in settings; scopes chat RAG too).
+  const setActive = async (pid) => {
+    await fetch(`${API_BASE}/settings`, {
+      method: "PUT", headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ active_project: pid }),
+    });
+    setActiveProject(pid);
+    load();
+    loadProjects();
+  };
+
+  const newProject = async () => {
+    const name = window.prompt("New project name:");
+    if (!name || !name.trim()) return;
+    const r = await fetch(`${API_BASE}/projects`, {
+      method: "POST", headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ name: name.trim() }),
+    });
+    if (r.ok) { const p = await r.json(); await setActive(p.id); }
+  };
+
+  const removeProject = async () => {
+    if (!activeProject) return;
+    if (!window.confirm("Delete this project? Its documents are kept but become unscoped.")) return;
+    await fetch(`${API_BASE}/projects/${encodeURIComponent(activeProject)}`, { method: "DELETE" });
+    await setActive("");
+  };
 
   const openDoc = async (doc) => {
     try {
@@ -25,7 +62,7 @@ export function Documents() {
     } catch { /* offline — leave list as-is */ }
   };
 
-  useEffect(() => { load(); }, []);
+  useEffect(() => { load(); loadProjects(); }, []);
 
   // Files (pdf/docx/txt/md) upload straight to the server, which extracts the
   // text. Base64 in JSON — no multipart dependency.
@@ -106,6 +143,30 @@ export function Documents() {
         material when a message is relevant.
       </p>
 
+      <div style={{ display: "flex", alignItems: "center", gap: "0.6rem", marginBottom: "1.2rem", flexWrap: "wrap" }}>
+        <span style={{ color: "#888", fontSize: "0.85rem" }}>Workspace:</span>
+        <select
+          value={activeProject}
+          onChange={e => (e.target.value === "__new__" ? newProject() : setActive(e.target.value))}
+          style={{ padding: "0.5rem 0.7rem", background: "#222", color: "#eee", border: "1px solid #333", borderRadius: "8px" }}
+        >
+          <option value="">All documents</option>
+          {projects.map(p => (
+            <option key={p.id} value={p.id}>{p.name} ({p.docs})</option>
+          ))}
+          <option value="__new__">＋ New project…</option>
+        </select>
+        {activeProject && (
+          <button onClick={removeProject} title="Delete this project (documents kept)"
+            style={{ padding: "0.4rem 0.7rem", background: "#2a1a1a", color: "#ff8a80", border: "1px solid #5a2a2a", borderRadius: "8px", cursor: "pointer", fontSize: "0.8rem" }}>
+            Delete project
+          </button>
+        )}
+        <span style={{ color: "#666", fontSize: "0.78rem" }}>
+          {activeProject ? "Chat scopes to this project's documents." : "Chat searches all documents."}
+        </span>
+      </div>
+
       <div style={{ display: "flex", flexDirection: "column", gap: "0.7rem", marginBottom: "1.5rem" }}>
         <input type="text" placeholder="Title" value={title}
           onChange={e => setTitle(e.target.value)} style={input} />

interface/web/src/Playbook.jsx

@@ -360,7 +360,7 @@ export function Playbook() {
             />
             <input
               type="text"
-              placeholder="Tools (comma separated): search_memory, search_history, search_documents, list_models, get_time"
+              placeholder="Tools: search_memory, search_history, search_documents, list_models, get_time, web_search, fetch_url, remember"
               value={form.tools}
               onChange={e => setForm(prev => ({ ...prev, tools: e.target.value }))}
               style={{ padding: "0.9rem", background: "#222", color: "#eee", border: "1px solid #333", borderRadius: "10px" }}

requirements-base.txt

@@ -43,6 +43,8 @@ python-docx
 # Vector search: loadable SQLite extension (prebuilt wheels). Brute-force cosine
 # stays as the fallback when the host Python can't load extensions.
 sqlite-vec
+# Local speech-to-text: CTranslate2-based, no torch, keeps dictation on-device.
+faster-whisper
 
 # Documentation Support
 markdown-it-py

requirements-windows.txt

@@ -15,6 +15,8 @@ python-docx
 # Vector search: loadable SQLite extension (prebuilt wheels). Brute-force cosine
 # stays as the fallback when the host Python can't load extensions.
 sqlite-vec
+# Local speech-to-text: CTranslate2-based, no torch, keeps dictation on-device.
+faster-whisper
 
 # Native desktop window for the UI (WebView2 on Windows; pulls pythonnet).
 # Used by bin/nexus_window.py, launched from launch_nexus.ps1.

synapse/main.py

@@ -290,6 +290,7 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
             message, get_ollama_manager().embed,
             limit=app_settings.get("rag_top_k", 3),
             min_score=app_settings.get("rag_min_score", 0.6),
+            project_id=app_settings.get("active_project", "") or None,
         )
         doc_titles: list = []
         if doc_hits:
@@ -943,12 +944,70 @@ async def delete_playbook_endpoint(id: UUID):
         raise HTTPException(status_code=500, detail=str(e))
 
 
+# -------------------------
+# Speech-to-text (local Whisper)
+# -------------------------
+from . import stt as _stt
+
+
+@app.get("/stt/status")
+async def stt_status():
+    return {"available": _stt.available()}
+
+
+@app.post("/stt")
+async def stt_transcribe(payload: Dict[str, Any] = Body(...)):
+    if not _stt.available():
+        raise HTTPException(status_code=503, detail="local STT (faster-whisper) not installed")
+    audio = payload.get("audio") or ""
+    if not audio:
+        raise HTTPException(status_code=400, detail="audio (base64) is required")
+    try:
+        text = await _asyncio.to_thread(_stt.transcribe_b64, audio)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"transcription failed: {e}")
+    return {"text": text}
+
+
+# -------------------------
+# Projects / workspaces
+# -------------------------
+def _active_project() -> str:
+    """The active project id, or '' for the unscoped 'All' view."""
+    return store.get_settings().get("active_project", "") or ""
+
+
+@app.get("/projects")
+async def list_projects():
+    return {"projects": store.list_projects(), "active": _active_project()}
+
+
+@app.post("/projects")
+async def create_project(payload: Dict[str, Any] = Body(...)):
+    name = (payload.get("name") or "").strip()
+    if not name:
+        raise HTTPException(status_code=400, detail="name is required")
+    return store.create_project(name)
+
+
+@app.delete("/projects/{project_id}")
+async def delete_project(project_id: str):
+    if not store.delete_project(project_id):
+        raise HTTPException(status_code=404, detail="Not Found")
+    # If the deleted project was active, fall back to the "All" view.
+    if _active_project() == project_id:
+        store.update_settings({"active_project": ""})
+    return {"status": "deleted"}
+
+
 # -------------------------
 # Documents (RAG)
 # -------------------------
 @app.get("/documents")
 async def list_documents():
-    return {"documents": store.list_documents()}
+    # Scope to the active project; "" (All) lists everything.
+    active = _active_project()
+    return {"documents": store.list_documents(active if active else None)}
 
 
 @app.post("/documents")
@@ -957,7 +1016,7 @@ async def add_document(payload: Dict[str, Any] = Body(...)):
     content = (payload.get("content") or "").strip()
     if not title or not content:
         raise HTTPException(status_code=400, detail="title and content are required")
-    result = await store.add_document(title, content, get_ollama_manager().embed)
+    result = await store.add_document(title, content, get_ollama_manager().embed, _active_project())
     if result["chunks"] == 0:
         raise HTTPException(status_code=400, detail="no text to index")
     return result
@@ -1000,7 +1059,7 @@ async def upload_document(payload: Dict[str, Any] = Body(...)):
     if not text:
         raise HTTPException(status_code=400, detail="no extractable text in file")
     title = _os.path.splitext(_os.path.basename(filename))[0] or filename
-    return await store.add_document(title, text, get_ollama_manager().embed)
+    return await store.add_document(title, text, get_ollama_manager().embed, _active_project())
 
 
 @app.get("/documents/{doc_id}")

synapse/memory/store.py

@@ -191,6 +191,20 @@ class PersistentMemoryStore:
         """)
         cur.execute("CREATE INDEX IF NOT EXISTS idx_documents_doc_id ON documents (doc_id)")
 
+        # Projects / workspaces: group documents so RAG can scope to one set.
+        cur.execute("""
+            CREATE TABLE IF NOT EXISTS projects (
+                id TEXT PRIMARY KEY,
+                name TEXT NOT NULL,
+                created_at REAL NOT NULL
+            )
+        """)
+        # documents.project_id — "" (or missing) means unscoped / All.
+        try:
+            cur.execute("ALTER TABLE documents ADD COLUMN project_id TEXT NOT NULL DEFAULT ''")
+        except Exception:
+            pass
+
         cur.execute("""
             CREATE TABLE IF NOT EXISTS settings (
                 key TEXT PRIMARY KEY,
@@ -569,26 +583,31 @@ class PersistentMemoryStore:
                     "INSERT OR REPLACE INTO message_vectors (message_id, embedding) VALUES (?, ?)",
                     (row["id"], json.dumps(vec)),
                 )
+                self._vec_upsert_msg(conn, row["id"], vec)  # mirror into the ANN index
         if missing:
             conn.commit()
 
-        # Score every stored message against the query vector.
-        cur.execute("""
-            SELECT v.message_id, v.embedding, m.conversation_id
-            FROM message_vectors v
-            JOIN messages m ON m.id = v.message_id
-        """)
-        scored = []
-        for row in cur.fetchall():
-            try:
-                vec = json.loads(row["embedding"])
-            except Exception:
-                continue
-            score = _cosine(query_vec, vec)
-            if score >= min_score:
-                scored.append((score, row["message_id"], row["conversation_id"]))
-
-        scored.sort(reverse=True)
+        # Rank messages by similarity. Fast path: the sqlite-vec ANN index over an
+        # over-fetch (conversation dedup below thins it); else brute-force cosine.
+        scored = None
+        if self.vec_enabled:
+            scored = self._vec_search_messages(conn, query_vec, max(limit * 5, 20), min_score)
+        if scored is None:
+            cur.execute("""
+                SELECT v.message_id, v.embedding, m.conversation_id
+                FROM message_vectors v
+                JOIN messages m ON m.id = v.message_id
+            """)
+            scored = []
+            for row in cur.fetchall():
+                try:
+                    vec = json.loads(row["embedding"])
+                except Exception:
+                    continue
+                score = _cosine(query_vec, vec)
+                if score >= min_score:
+                    scored.append((score, row["message_id"], row["conversation_id"]))
+            scored.sort(reverse=True)
 
         results: List[dict] = []
         seen_convs: set = set()
@@ -616,6 +635,70 @@ class PersistentMemoryStore:
 
         return results
 
+    # --- message vector index (sqlite-vec) — same pattern as documents --------
+    def _ensure_vec_msgs(self, conn, dim: int) -> None:
+        conn.execute(
+            f"CREATE VIRTUAL TABLE IF NOT EXISTS vec_messages "
+            f"USING vec0(embedding float[{dim}] distance_metric=cosine)"
+        )
+
+    def _vec_upsert_msg(self, conn, message_id: int, vec: list) -> None:
+        if not (self.vec_enabled and vec):
+            return
+        try:
+            import sqlite_vec
+            self._ensure_vec_msgs(conn, len(vec))
+            conn.execute("DELETE FROM vec_messages WHERE rowid = ?", (message_id,))
+            conn.execute(
+                "INSERT INTO vec_messages(rowid, embedding) VALUES (?, ?)",
+                (message_id, sqlite_vec.serialize_float32(vec)),
+            )
+        except Exception:
+            pass
+
+    def _backfill_vec_msgs(self, conn, dim: int) -> None:
+        """Index any message_vectors rows missing from vec_messages."""
+        try:
+            import sqlite_vec
+            self._ensure_vec_msgs(conn, dim)
+            rows = conn.execute(
+                "SELECT mv.message_id AS mid, mv.embedding AS emb FROM message_vectors mv "
+                "LEFT JOIN vec_messages v ON v.rowid = mv.message_id WHERE v.rowid IS NULL"
+            ).fetchall()
+            for r in rows:
+                try:
+                    vec = json.loads(r["emb"])
+                    if len(vec) == dim:
+                        conn.execute(
+                            "INSERT INTO vec_messages(rowid, embedding) VALUES (?, ?)",
+                            (r["mid"], sqlite_vec.serialize_float32(vec)),
+                        )
+                except Exception:
+                    pass
+            conn.commit()
+        except Exception:
+            pass
+
+    def _vec_search_messages(self, conn, query_vec: list, fetch: int, min_score: float):
+        """Top message hits via the vec index as sorted [(score, msg_id, conv_id)],
+        or None to fall back to the brute-force scan. Stale rows for deleted
+        messages are dropped by the inner join, so they never surface."""
+        try:
+            import sqlite_vec
+            self._backfill_vec_msgs(conn, len(query_vec))
+            rows = conn.execute(
+                "SELECT v.rowid AS mid, v.distance AS distance, m.conversation_id AS cid "
+                "FROM vec_messages v JOIN messages m ON m.id = v.rowid "
+                "WHERE v.embedding MATCH ? ORDER BY v.distance LIMIT ?",
+                (sqlite_vec.serialize_float32(query_vec), fetch),
+            ).fetchall()
+            return [
+                (1.0 - r["distance"], r["mid"], r["cid"])
+                for r in rows if (1.0 - r["distance"]) >= min_score
+            ]  # distance-asc == score-desc, already sorted
+        except Exception:
+            return None
+
     def _exchange_pair(self, cur, conv_id: str, message_id: int) -> Optional[dict]:
         """Build a {id, updated_at, matches} record with the full user+assistant
         pair surrounding `message_id`, in the shape search callers expect."""
@@ -677,7 +760,7 @@ class PersistentMemoryStore:
             chunks.append(buf)
         return chunks
 
-    async def add_document(self, title: str, content: str, embed_fn) -> dict:
+    async def add_document(self, title: str, content: str, embed_fn, project_id: str = "") -> dict:
         """Chunk, embed, and store a document. Returns {doc_id, chunks}."""
         import uuid as _uuid
         doc_id = str(_uuid.uuid4())
@@ -688,16 +771,47 @@ class PersistentMemoryStore:
         for i, piece in enumerate(pieces):
             vec = await embed_fn(self._EMBED_DOC_PREFIX + piece[:2000])
             cur.execute(
-                "INSERT INTO documents (id, doc_id, title, chunk_idx, text, embedding, created_at)"
-                " VALUES (?, ?, ?, ?, ?, ?, ?)",
+                "INSERT INTO documents (id, doc_id, title, chunk_idx, text, embedding, created_at, project_id)"
+                " VALUES (?, ?, ?, ?, ?, ?, ?, ?)",
                 (str(_uuid.uuid4()), doc_id, title, i, piece,
-                 json.dumps(vec) if vec else None, now),
+                 json.dumps(vec) if vec else None, now, project_id or ""),
             )
             self._vec_upsert(conn, cur.lastrowid, vec)  # mirror into the ANN index
         conn.commit()
         conn.close()
         return {"doc_id": doc_id, "title": title, "chunks": len(pieces)}
 
+    # --- projects / workspaces --------------------------------------------------
+    def create_project(self, name: str) -> dict:
+        import uuid as _uuid
+        pid = str(_uuid.uuid4())
+        conn = self._connect()
+        conn.execute("INSERT INTO projects (id, name, created_at) VALUES (?, ?, ?)",
+                     (pid, name.strip(), time.time()))
+        conn.commit()
+        conn.close()
+        return {"id": pid, "name": name.strip()}
+
+    def list_projects(self) -> List[dict]:
+        conn = self._connect()
+        rows = conn.execute("""
+            SELECT p.id, p.name, p.created_at,
+                   (SELECT COUNT(DISTINCT doc_id) FROM documents d WHERE d.project_id = p.id) AS docs
+            FROM projects p ORDER BY p.created_at ASC
+        """).fetchall()
+        conn.close()
+        return [dict(r) for r in rows]
+
+    def delete_project(self, project_id: str) -> bool:
+        """Delete a project; its documents survive but become unscoped ("")."""
+        conn = self._connect()
+        conn.execute("UPDATE documents SET project_id = '' WHERE project_id = ?", (project_id,))
+        cur = conn.execute("DELETE FROM projects WHERE id = ?", (project_id,))
+        deleted = cur.rowcount
+        conn.commit()
+        conn.close()
+        return deleted > 0
+
     # --- vector index (sqlite-vec) — accelerator over the JSON embedding column,
     # --- with brute-force cosine below as the guaranteed fallback ---------------
     def _ensure_vec_docs(self, conn, dim: int) -> None:
@@ -768,13 +882,21 @@ class PersistentMemoryStore:
         except Exception:
             return None
 
-    def list_documents(self) -> List[dict]:
+    def list_documents(self, project_id: Optional[str] = None) -> List[dict]:
+        """All documents, or just one project's when project_id is given."""
         conn = self._connect()
         cur = conn.cursor()
-        cur.execute("""
-            SELECT doc_id, title, COUNT(*) AS chunks, MIN(created_at) AS created_at
-            FROM documents GROUP BY doc_id, title ORDER BY created_at DESC
-        """)
+        if project_id is not None:
+            cur.execute("""
+                SELECT doc_id, title, COUNT(*) AS chunks, MIN(created_at) AS created_at
+                FROM documents WHERE project_id = ?
+                GROUP BY doc_id, title ORDER BY created_at DESC
+            """, (project_id,))
+        else:
+            cur.execute("""
+                SELECT doc_id, title, COUNT(*) AS chunks, MIN(created_at) AS created_at
+                FROM documents GROUP BY doc_id, title ORDER BY created_at DESC
+            """)
         rows = cur.fetchall()
         conn.close()
         return [dict(r) for r in rows]
@@ -807,23 +929,30 @@ class PersistentMemoryStore:
         return deleted > 0
 
     async def search_documents(
-        self, query: str, embed_fn, limit: int = 3, min_score: float = 0.6
+        self, query: str, embed_fn, limit: int = 3, min_score: float = 0.6,
+        project_id: Optional[str] = None,
     ) -> List[dict]:
         """Top-`limit` document chunks most similar to `query`. Returns
-        [{title, text, score}]. Empty on no query / embeddings down."""
+        [{title, text, score}]. Empty on no query / embeddings down.
+        When project_id is given, only that project's docs are searched.
+        ponytail: scoped search uses the brute-force path (easy SQL filter, few
+        docs per project); the vec index accelerates the unscoped "All" case."""
         if not query or not query.strip():
             return []
         query_vec = await embed_fn(self._EMBED_QUERY_PREFIX + query.strip())
         if not query_vec:
             return []
-        # Fast path: the sqlite-vec ANN index. None => fall through to brute force.
-        if self.vec_enabled:
+        # Fast path (unscoped only): the sqlite-vec ANN index.
+        if not project_id and self.vec_enabled:
             hits = self._vec_search(query_vec, limit, min_score)
             if hits is not None:
                 return hits
         conn = self._connect()
         cur = conn.cursor()
-        cur.execute("SELECT title, text, embedding FROM documents WHERE embedding IS NOT NULL")
+        if project_id:
+            cur.execute("SELECT title, text, embedding FROM documents WHERE embedding IS NOT NULL AND project_id = ?", (project_id,))
+        else:
+            cur.execute("SELECT title, text, embedding FROM documents WHERE embedding IS NOT NULL")
         scored = []
         for row in cur.fetchall():
             try:
@@ -852,6 +981,8 @@ class PersistentMemoryStore:
         # cosine similarity (0-1) a chunk must clear to count as relevant.
         "rag_top_k": 3,
         "rag_min_score": 0.6,
+        # Active project/workspace; "" = all documents (unscoped).
+        "active_project": "",
         "system_prompt": "",
         "timeout": 120,
         # How long Ollama keeps the model resident in VRAM between messages.

synapse/stt.py

@@ -0,0 +1,55 @@
+"""Local speech-to-text via faster-whisper.
+
+Deliberately torch-free: faster-whisper runs on CTranslate2 (CPU, int8), so it
+fits the "no ML stack in the venv" rule that keeps Promethean small. Replaces
+the browser's Web Speech API, which in Chrome ships audio to Google — the whole
+point is to keep dictation on-device.
+
+Model size via NEXUS_STT_MODEL (default "base"); downloaded and cached on first
+use. If faster-whisper isn't installed, `available()` is False and the backend
+reports it so the UI can fall back.
+"""
+from __future__ import annotations
+
+import base64
+import logging
+import os
+import tempfile
+
+_log = logging.getLogger("nexus.stt")
+_MODEL = None
+_MODEL_SIZE = os.getenv("NEXUS_STT_MODEL", "base")
+
+
+def available() -> bool:
+    try:
+        import faster_whisper  # noqa: F401
+        return True
+    except Exception:
+        return False
+
+
+def _get_model():
+    global _MODEL
+    if _MODEL is None:
+        from faster_whisper import WhisperModel
+        _log.info("loading whisper model %s (cpu/int8)", _MODEL_SIZE)
+        _MODEL = WhisperModel(_MODEL_SIZE, device="cpu", compute_type="int8")
+    return _MODEL
+
+
+def transcribe_b64(audio_b64: str) -> str:
+    """Transcribe a base64 audio blob (any container PyAV can decode — the
+    browser's MediaRecorder produces webm/opus)."""
+    raw = base64.b64decode(audio_b64)
+    with tempfile.NamedTemporaryFile(suffix=".webm", delete=False) as f:
+        f.write(raw)
+        path = f.name
+    try:
+        segments, _info = _get_model().transcribe(path, vad_filter=True)
+        return " ".join(s.text for s in segments).strip()
+    finally:
+        try:
+            os.unlink(path)
+        except OSError:
+            pass

synapse/tools.py

@@ -1,17 +1,19 @@
-"""Read-only tools a playbook can call during chat.
+"""Tools a playbook can call during chat.
 
 Ollama drives the calling: `/api/chat` with a `tools` param returns
-`message.tool_calls`, and this module is just the registry + dispatch. Every
-tool here only READS local state (SQLite, Ollama) — no side effects. The
-per-playbook allowlist (`PlaybookItem.tools`) is the security boundary; keep the
-registry read-only until the loop is trusted.
+`message.tool_calls`, and this module is just the registry + dispatch.
+
+Most tools READ local state (memory, history, documents, models). A few act:
+`web_search`/`fetch_url` make outbound HTTP requests, and `remember` WRITES a
+memory fact. The per-playbook allowlist (`PlaybookItem.tools`) is the security
+boundary — an action tool only fires when a playbook explicitly lists it.
 """
 from __future__ import annotations
 
 import json
 from typing import Awaitable, Callable
 
-from .memory.store import store
+from .memory.store import store, MemoryItem
 from .ollama_manager import get_ollama_manager
 
 
@@ -51,6 +53,45 @@ async def _get_time(**_) -> str:
     return json.dumps({"now": datetime.now().isoformat(timespec="seconds")})
 
 
+async def _web_search(query: str = "", **_) -> str:
+    import asyncio as _a
+    from .search import web_search
+    res = await _a.to_thread(web_search, query or "", 4)
+    return res or "(no results)"
+
+
+async def _fetch_url(url: str = "", **_) -> str:
+    import re
+    import httpx
+    url = (url or "").strip()
+    if not url.startswith(("http://", "https://")):
+        return json.dumps({"error": "url must start with http:// or https://"})
+    # ponytail: no SSRF allow/deny-list — local single-user assistant, and the
+    # tool only runs when a playbook explicitly grants fetch_url. Add host
+    # filtering if this ever serves multiple/untrusted users.
+    try:
+        async with httpx.AsyncClient(timeout=15.0, follow_redirects=True) as c:
+            r = await c.get(url, headers={"User-Agent": "NexusOS/1.0"})
+            r.raise_for_status()
+            html = r.text
+    except Exception as e:
+        return json.dumps({"error": f"fetch failed: {e}"})
+    text = re.sub(r"(?is)<(script|style).*?</\1>", " ", html)
+    text = re.sub(r"(?s)<[^>]+>", " ", text)
+    text = re.sub(r"\s+", " ", text).strip()
+    return text[:4000]
+
+
+async def _remember(text: str = "", section: str = "General", **_) -> str:
+    """WRITE tool: persist a memory fact. First action tool — allowlist-gated."""
+    import uuid as _uuid
+    text = (text or "").strip()
+    if not text:
+        return json.dumps({"error": "text is required"})
+    store.add(MemoryItem(id=str(_uuid.uuid4()), section=(section or "General"), text=text))
+    return json.dumps({"saved": text, "section": section or "General"})
+
+
 # name -> (schema, callable). Schema is the OpenAI/Ollama function-tool format.
 REGISTRY: dict[str, tuple[dict, Callable[..., Awaitable[str]]]] = {
     "search_memory": (
@@ -119,6 +160,54 @@ REGISTRY: dict[str, tuple[dict, Callable[..., Awaitable[str]]]] = {
         },
         _get_time,
     ),
+    "web_search": (
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web (DuckDuckGo) and return the top result snippets.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        },
+        _web_search,
+    ),
+    "fetch_url": (
+        {
+            "type": "function",
+            "function": {
+                "name": "fetch_url",
+                "description": "Fetch a web page and return its visible text (truncated).",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"url": {"type": "string", "description": "http(s) URL"}},
+                    "required": ["url"],
+                },
+            },
+        },
+        _fetch_url,
+    ),
+    "remember": (
+        {
+            "type": "function",
+            "function": {
+                "name": "remember",
+                "description": "Save a durable fact to the user's persistent memory.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "text": {"type": "string", "description": "the fact to remember"},
+                        "section": {"type": "string", "description": "optional category, e.g. Health"},
+                    },
+                    "required": ["text"],
+                },
+            },
+        },
+        _remember,
+    ),
 }
 
 

tests/test_documents.py

@@ -57,6 +57,62 @@ def test_search_empty_query_returns_nothing():
     assert asyncio.run(s.search_documents("", _fake_embed)) == []
 
 
+def test_conversation_recall_uses_vec_and_matches_brute_force():
+    s = _store()
+    if not s.vec_enabled:
+        import pytest
+        pytest.skip("sqlite-vec not loadable on this host")
+
+    async def run():
+        s.create_conversation("c1")
+        s.add_message("c1", "user", "tell me about lego star wars")
+        s.add_message("c1", "assistant", "lego star wars is a fun game")
+        s.create_conversation("c2")
+        s.add_message("c2", "user", "gpu vega vram notes")
+        s.add_message("c2", "assistant", "vega has 4gb")
+
+        hits = await s.semantic_search_conversations("lego star wars", _fake_embed, limit=2, min_score=0.1)
+        assert hits and hits[0]["id"] == "c1"
+
+        conn = s._connect()
+        n = conn.execute("SELECT COUNT(*) FROM vec_messages").fetchone()[0]
+        conn.close()
+        assert n >= 2  # dual-write populated the message vec index
+
+        s.vec_enabled = False
+        bf = await s.semantic_search_conversations("lego star wars", _fake_embed, limit=2, min_score=0.1)
+        assert bf[0]["id"] == hits[0]["id"]
+
+    asyncio.run(run())
+
+
+def test_projects_scope_documents_and_survive_delete():
+    s = _store()
+
+    async def run():
+        p = s.create_project("Star Wars")
+        assert [x["name"] for x in s.list_projects()] == ["Star Wars"]
+
+        await s.add_document("Lego", "lego star wars boss tips", _fake_embed, project_id=p["id"])
+        await s.add_document("GPU", "gpu vega notes", _fake_embed)  # unscoped
+
+        # project sees only its own; unscoped/all sees both
+        assert [d["title"] for d in s.list_documents(p["id"])] == ["Lego"]
+        assert len(s.list_documents(None)) == 2
+
+        # scoped search only returns the project's docs
+        scoped = await s.search_documents("star wars", _fake_embed, min_score=0.1, project_id=p["id"])
+        assert scoped and all(h["title"] == "Lego" for h in scoped)
+
+        # deleting the project keeps the docs but unscopes them
+        assert s.delete_project(p["id"]) is True
+        assert s.list_projects() == []
+        assert len(s.list_documents(None)) == 2
+        assert len(s.list_documents(p["id"])) == 0  # nothing left in that project
+
+    asyncio.run(run())
+
+
 def test_chunker_overlap_and_hard_split():
     s = _store()
     # a single oversized paragraph (no blank lines, as in PDF text) is split

tests/test_smoke.py

@@ -43,6 +43,13 @@ def test_keep_alive_pins_the_model():
     assert "keep_alive" not in mgr._apply_keep_alive({"model": "x"})
 
 
+def test_stt_status_endpoint():
+    # Reports whether local Whisper is installed; wiring must respond either way.
+    resp = TestClient(app).get("/stt/status")
+    assert resp.status_code == 200
+    assert isinstance(resp.json()["available"], bool)
+
+
 def test_num_ctx_option_only_when_positive():
     # 0 / None -> omit num_ctx so Ollama uses the model default; positive -> set it.
     from synapse.ollama_manager import _chat_options

tests/test_tools.py

@@ -23,6 +23,24 @@ async def _drain(gen):
     return [s async for s in gen]
 
 
+def test_remember_writes_a_memory_fact(tmp_path, monkeypatch):
+    # The `remember` action tool persists a fact through the store.
+    import synapse.memory.store as store_mod
+    from synapse.memory.store import PersistentMemoryStore
+    fresh = PersistentMemoryStore(tmp_path / "m.db")
+    monkeypatch.setattr(tools, "store", fresh)
+    out = asyncio.run(tools.dispatch("remember", {"text": "user likes tea", "section": "Prefs"}))
+    assert "user likes tea" in out
+    assert any(it.text == "user likes tea" for it in fresh.all())
+
+
+def test_action_tools_registered():
+    for name in ("web_search", "fetch_url", "remember"):
+        assert name in tools.REGISTRY
+    names = [s["function"]["name"] for s in tools.schemas_for(["web_search", "remember", "nope"])]
+    assert names == ["web_search", "remember"]
+
+
 class _FakeManager:
     """Returns a tool_call on the first chat() call, plain content after."""
     def __init__(self):