feat(security): request-size caps, concurrency limits, model-pull allowlist

DoS/quota guardrails for the unauthenticated local APIs:

* Body-size middleware rejects oversized requests (Content-Length) before they
  are buffered/base64-decoded (NEXUS_MAX_REQUEST_MB, default 32).
* Document upload enforces a decoded-byte cap (NEXUS_MAX_UPLOAD_MB, default 20)
  and a PDF page-count cap (NEXUS_MAX_PDF_PAGES, default 500) as backstops for
  chunked bodies and pathological files.
* A counter-based in-flight limiter bounds concurrent chats and document
  ingests (NEXUS_MAX_CONCURRENT_CHATS/UPLOADS), returning 429 when saturated;
  the chat slot is held for the whole SSE stream and released on completion or
  client disconnect.
* /models/pull gains an opt-in allowlist (NEXUS_MODEL_ALLOWLIST); empty by
  default so behaviour is unchanged, otherwise a bare repo name permits all its
  tags.

Co-authored-by: Cursor <cursoragent@cursor.com>

synapse/main.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import asyncio as _asyncio
+import contextlib as _contextlib
 import json as _json
 import secrets as _secrets
 import uuid as _uuid
@@ -13,12 +14,16 @@ import httpx
 import os as _os
 import platform as _platform
 from pathlib import Path
-from fastapi import FastAPI, HTTPException, Body
+from fastapi import FastAPI, HTTPException, Body, Request
 from fastapi.middleware.cors import CORSMiddleware
 from starlette.middleware.trustedhost import TrustedHostMiddleware
-from fastapi.responses import StreamingResponse, FileResponse
+from fastapi.responses import StreamingResponse, FileResponse, JSONResponse
 
-from .nexus_config import settings, VERSION, DEFAULT_CHAT_MODEL, ALLOWED_HOSTS, ALLOWED_ORIGINS
+from .nexus_config import (
+    settings, VERSION, DEFAULT_CHAT_MODEL, ALLOWED_HOSTS, ALLOWED_ORIGINS,
+    MAX_REQUEST_BYTES, MAX_UPLOAD_BYTES, MAX_PDF_PAGES,
+    MAX_CONCURRENT_CHATS, MAX_CONCURRENT_UPLOADS, model_pull_allowed,
+)
 from .chat import generate_chat_response, stream_chat_response, _synapse_trace
 from . import chat as _chat
 from .ollama_manager import initialize_ollama, initialize_ollama_async, get_ollama_manager
@@ -200,6 +205,63 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+
+# --- Request-size cap ---
+# Reject oversized bodies before they are buffered/decoded (a base64 upload or a
+# giant chat payload could otherwise exhaust memory). Header-based: covers the
+# realistic clients (browsers/httpx always send Content-Length for JSON). A
+# chunked request with no length still reaches the handler, where the upload path
+# enforces its own decoded-byte cap as a backstop.
+@app.middleware("http")
+async def _limit_request_body(request: Request, call_next):
+    cl = request.headers.get("content-length")
+    if cl:
+        try:
+            if int(cl) > MAX_REQUEST_BYTES:
+                return JSONResponse(
+                    status_code=413,
+                    content={"detail": f"request body too large (max {MAX_REQUEST_BYTES} bytes)"},
+                )
+        except ValueError:
+            pass
+    return await call_next(request)
+
+
+class _InFlightLimiter:
+    """Bounds concurrent in-flight ops for one expensive entry point so a single
+    caller can't fan out unlimited inference/embedding work. The server runs on
+    one asyncio loop, so a plain counter is safe — there's no await between the
+    check and the increment. Over the limit -> HTTP 429."""
+
+    def __init__(self, limit: int, label: str):
+        self._limit = max(1, int(limit))
+        self._label = label
+        self._n = 0
+
+    def acquire(self) -> None:
+        if self._n >= self._limit:
+            raise HTTPException(
+                status_code=429,
+                detail=f"{self._label} is busy ({self._limit} concurrent max) — retry shortly",
+            )
+        self._n += 1
+
+    def release(self) -> None:
+        if self._n > 0:
+            self._n -= 1
+
+    @_contextlib.contextmanager
+    def slot(self):
+        self.acquire()
+        try:
+            yield
+        finally:
+            self.release()
+
+
+_CHAT_INFLIGHT = _InFlightLimiter(MAX_CONCURRENT_CHATS, "chat")
+_UPLOAD_INFLIGHT = _InFlightLimiter(MAX_CONCURRENT_UPLOADS, "document ingest")
+
 # --- GLOBALS ---
 playbooks = PlaybookManager()
 ollama = None
@@ -245,6 +307,11 @@ async def root():
 # -------------------------
 @app.post("/chat/stream")
 async def chat_stream_endpoint(payload: Dict[str, Any]):
+    # Bound concurrent chats so a flood can't fan out unlimited model inference.
+    # Acquired here (covers the pre-stream RAG/embedding work too) and released
+    # when the SSE stream finishes — see the guarded wrapper below.
+    _CHAT_INFLIGHT.acquire()
+    _chat_slot_held = True
     try:
         message = payload.get("message", "")
         app_settings = store.get_settings()
@@ -498,12 +565,29 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
                 except Exception as e:
                     _synapse_trace(f"\n⚠ memory extraction call failed: {e}\n")
 
-        return StreamingResponse(event_stream(), media_type="text/event-stream")
+        # Hold the concurrency slot for the stream's lifetime, then release it
+        # exactly once when the generator is exhausted or closed (client
+        # disconnect). Ownership passes to the wrapper, so the endpoint's own
+        # finally must not also release.
+        _inner = event_stream()
+
+        async def _guarded_stream() -> AsyncGenerator[str, None]:
+            try:
+                async for _chunk in _inner:
+                    yield _chunk
+            finally:
+                _CHAT_INFLIGHT.release()
+
+        _chat_slot_held = False
+        return StreamingResponse(_guarded_stream(), media_type="text/event-stream")
 
     except HTTPException:
         raise
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        if _chat_slot_held:
+            _CHAT_INFLIGHT.release()
 
 
 @app.post("/chat/approve")
@@ -684,6 +768,11 @@ async def pull_model(payload: Dict[str, Any] = Body(...)):
     name = payload.get("name", "").strip()
     if not name:
         raise HTTPException(status_code=400, detail="Missing model name")
+    if not model_pull_allowed(name):
+        raise HTTPException(
+            status_code=403,
+            detail=f"model '{name}' is not in NEXUS_MODEL_ALLOWLIST",
+        )
 
     async def _stream():
         try:
@@ -1120,7 +1209,13 @@ def _extract_text(filename: str, data: bytes) -> str:
     if name.endswith(".pdf"):
         from pypdf import PdfReader
         reader = PdfReader(io.BytesIO(data))
-        return "\n\n".join((page.extract_text() or "") for page in reader.pages)
+        pages = reader.pages
+        if len(pages) > MAX_PDF_PAGES:
+            raise HTTPException(
+                status_code=413,
+                detail=f"PDF has {len(pages)} pages (max {MAX_PDF_PAGES})",
+            )
+        return "\n\n".join((page.extract_text() or "") for page in pages)
     if name.endswith(".docx"):
         import docx
         doc = docx.Document(io.BytesIO(data))
@@ -1142,8 +1237,20 @@ async def upload_document(payload: Dict[str, Any] = Body(...)):
         raw = base64.b64decode(b64)
     except Exception:
         raise HTTPException(status_code=400, detail="data must be base64")
+    # Decoded-byte backstop: the body-size middleware caps the encoded payload,
+    # but base64 inflates ~33% and a chunked request has no Content-Length, so
+    # enforce the real decoded ceiling here too.
+    if len(raw) > MAX_UPLOAD_BYTES:
+        raise HTTPException(
+            status_code=413,
+            detail=f"file too large: {len(raw)} bytes (max {MAX_UPLOAD_BYTES})",
+        )
+    # Bound concurrent ingests: extract + chunk + embed is CPU/model-heavy.
+    with _UPLOAD_INFLIGHT.slot():
         try:
             text = _extract_text(filename, raw).strip()
+        except HTTPException:
+            raise
         except Exception as e:
             raise HTTPException(status_code=400, detail=f"could not read {filename}: {e}")
         if not text:

synapse/nexus_config.py

@@ -166,6 +166,42 @@ _LOCAL_ORIGINS = [
 ALLOWED_HOSTS = _csv_env("NEXUS_ALLOWED_HOSTS", _LOCAL_HOSTS)
 ALLOWED_ORIGINS = _csv_env("NEXUS_ALLOWED_ORIGINS", _LOCAL_ORIGINS)
 
+
+# --- resource limits (DoS guardrails for the unauthenticated local APIs) ---
+# Even local-only, an unbounded base64 upload or a flood of concurrent inference
+# requests can exhaust RAM/CPU. These are generous defaults for single-user use,
+# all env-overridable.
+def _int_env(name: str, default: int) -> int:
+    try:
+        return int((os.getenv(name) or "").strip() or default)
+    except ValueError:
+        return default
+
+MAX_REQUEST_BYTES = _int_env("NEXUS_MAX_REQUEST_MB", 32) * 1024 * 1024
+MAX_UPLOAD_BYTES = _int_env("NEXUS_MAX_UPLOAD_MB", 20) * 1024 * 1024
+MAX_PDF_PAGES = _int_env("NEXUS_MAX_PDF_PAGES", 500)
+MAX_CONCURRENT_CHATS = _int_env("NEXUS_MAX_CONCURRENT_CHATS", 4)
+MAX_CONCURRENT_UPLOADS = _int_env("NEXUS_MAX_CONCURRENT_UPLOADS", 2)
+
+# Opt-in allowlist for /models/pull. Empty (default) = unrestricted, preserving
+# current behaviour. Set NEXUS_MODEL_ALLOWLIST=mistral,llama3 to bound which
+# models can be downloaded; a bare repo name (before the ':tag') matches all of
+# its tags, so "mistral" permits "mistral:latest", "mistral:7b", etc.
+MODEL_ALLOWLIST = _csv_env("NEXUS_MODEL_ALLOWLIST", [])
+
+
+def model_pull_allowed(name: str) -> bool:
+    """True if `name` may be pulled: always when no allowlist is configured,
+    otherwise when the full name or its repo part (before the first ':') is
+    listed. Case-insensitive."""
+    if not MODEL_ALLOWLIST:
+        return True
+    n = (name or "").strip().lower()
+    if not n:
+        return False
+    allow = {a.lower() for a in MODEL_ALLOWLIST}
+    return n in allow or n.split(":", 1)[0] in allow
+
 # exported instance
 settings = Settings()
 
@@ -176,7 +212,10 @@ __all__ = ["Settings", "settings", "path", "VERSION",
            "MEMORY_DIR", "LOGS_DIR", "PLAYBOOK_DIR", "UPLOADS_DIR",
            "EXPORTS_DIR", "MEMORY_DB",
            "BACKEND_LOG", "OLLAMA_LOG", "CHAT_LOG",
-           "ALLOWED_HOSTS", "ALLOWED_ORIGINS"]
+           "ALLOWED_HOSTS", "ALLOWED_ORIGINS",
+           "MAX_REQUEST_BYTES", "MAX_UPLOAD_BYTES", "MAX_PDF_PAGES",
+           "MAX_CONCURRENT_CHATS", "MAX_CONCURRENT_UPLOADS",
+           "MODEL_ALLOWLIST", "model_pull_allowed"]
            
 # --- quick runtime sanity check when run directly (no side effects on import) ---
 if __name__ == "__main__":