Add connectivity tests and fix Hermes handle + venv setup

- test_connectivity.py: connectivity tests for all four endpoint types
  (anthropic, openai, lm_studio, hermes, agent0) — treats no-credits as success
- test_hermes.py: raw WebSocket frame logger used to reverse-engineer protocol
- Fix handle_hermes: skip prompt.submit ack frame, read full text from
  message.complete payload.text, always raise on status==error
- Fix requirements.txt: use >= pins (fastapi/uvicorn versions didn't exist)
- Fix dev.sh: prefer python3.12 for venv (mcp>=1.9.0 requires 3.10+)
- Remove ANTHROPIC_KEY env var dependency from server.py (keys come from DB)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

server.py

@@ -451,21 +451,43 @@ async def handle_hermes(
                 log.info(f"[{agent.name}] created Hermes session {hermes_session_id}")
 
             # ── 2. Submit prompt ──────────────────────────────────────────
-            await rpc("prompt.submit", {"session_id": hermes_session_id, "text": req.message})
+            # Send prompt.submit but do NOT wait for its ack via rpc() —
+            # the ack arrives interleaved with events (after message.start).
+            # We drain all frames below until message.complete arrives.
+            submit_id = f"h{req_id + 1}"
+            req_id += 1
+            await ws.send(json.dumps({
+                "jsonrpc": "2.0", "id": submit_id,
+                "method": "prompt.submit",
+                "params": {"session_id": hermes_session_id, "text": req.message},
+            }))
 
             # ── 3. Stream events until message.complete ───────────────────
             full_text = ""
             while True:
                 raw = await asyncio.wait_for(ws.recv(), timeout=120.0)
                 msg = json.loads(raw)
+
+                # RPC ack for prompt.submit — ignore, keep reading
+                if msg.get("id") == submit_id:
+                    continue
+
                 if msg.get("method") != "event":
                     continue
+
                 params = msg.get("params") or {}
                 etype = params.get("type", "")
                 payload = params.get("payload") or {}
+
                 if etype == "message.delta":
                     full_text += payload.get("text", "")
                 elif etype == "message.complete":
+                    # payload.text holds the full assembled response
+                    complete_text = payload.get("text", "")
+                    if complete_text:
+                        full_text = complete_text
+                    if payload.get("status") == "error":
+                        raise RuntimeError(full_text or "Hermes returned an error response")
                     break
                 elif etype == "error":
                     raise RuntimeError(payload.get("message", "Hermes error"))