v5 Phase 2: aktuelle Modelle, GGUF-Klartext, Tools, HF-Token

- recipes.py: Juni-2026-Modelle (Qwen3-Coder-30B-A3B, Qwen3-8B/30B, Qwen2.5-VL-7B, Qwen3-4B);
  nur Repo gespeichert, GGUF-Datei wird beim Installieren dynamisch aufgeloest (_pick_gguf) +
  UPGRADES-Map fuer Phase 3.
- cookbook: 'kein GGUF' neutral statt rot + GGUF-Erklaerung (infoDot); ctx-infoDot.
- connect.js: 'Empfohlene Tools (Juni 2026)' (OpenCode/Cline/Continue) + MCP-Hinweis.
- HF-Token in Einstellungen -> als HF_TOKEN an Downloads/Recipe-Install durchgereicht.

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

routers/cookbook.py

@@ -32,6 +32,7 @@ class EvaluateRequest(BaseModel):
 
 class InstallRecipeReq(BaseModel):
     recipe_id: str
+    hf_token: str | None = None
 
 def extract_params_b(repo_id: str) -> float:
     """Extrahiert die Parametergröße (in Milliarden) aus dem Repo-Namen."""
@@ -138,20 +139,41 @@ def install_recipe(req: InstallRecipeReq):
     if not recipe:
         raise HTTPException(404, "Setup nicht gefunden.")
     ram_gb = psutil.virtual_memory().total / (1024 ** 3)
+    env = {"HF_XET_HIGH_PERFORMANCE": "1"}
+    if req.hf_token:
+        env["HF_TOKEN"] = req.hf_token
     cfg = read_config()
     job_ids = []
     for m in recipe["models"]:
+        file = _pick_gguf(m["repo"], m.get("quant", "Q4_K_M"))
+        if not file:
+            continue  # kein GGUF im Repo gefunden -> Modell ueberspringen (Rest installiert trotzdem)
         target = MODELS_DIR / m["repo"].split("/")[-1]
         target.mkdir(parents=True, exist_ok=True)
-        args = ["hf", "download", m["repo"], m["file"], "--local-dir", str(target)]
-        jid = start_job(args, f"download {m['name']}", env={"HF_XET_HIGH_PERFORMANCE": "1"})
-        JOBS[jid]["result_path"] = str(target / m["file"])
+        args = ["hf", "download", m["repo"], file, "--local-dir", str(target)]
+        jid = start_job(args, f"download {m['name']}", env=env)
+        JOBS[jid]["result_path"] = str(target / file)
         job_ids.append(jid)
         # Eintrag jetzt schon schreiben — optimaler Kontext, aber gedeckelt fuer schnellen Erststart.
         ctx = min(max_ctx_for(m["params_b"], m["quant"], ram_gb), 32768)
-        path = str(target / m["file"])
+        path = str(target / file)
         cmd = CMD_TEMPLATE.replace("{model}", path).replace("{ctx}", str(ctx))
         cfg["models"][m["role"]] = {"cmd": LiteralScalarString(cmd + "\n"), "ttl": DEFAULT_TTL}
     write_config(cfg)
-    return {"job_ids": job_ids, "count": len(recipe["models"])}
+    return {"job_ids": job_ids, "count": len(job_ids)}
+
+
+def _pick_gguf(repo: str, quant: str = "Q4_K_M") -> str | None:
+    """Beste GGUF-Datei eines Repos auflösen: bevorzugt gewünschten Quant, keine Split-Teile."""
+    try:
+        with httpx.Client(timeout=10.0) as c:
+            tree = c.get(f"https://huggingface.co/api/models/{repo}/tree/main").json()
+    except Exception:  # noqa: BLE001
+        return None
+    ggufs = [f["path"] for f in tree if isinstance(f, dict) and str(f.get("path", "")).endswith(".gguf")]
+    if not ggufs:
+        return None
+    pref = [g for g in ggufs if quant.lower() in g.lower() and "-of-" not in g]
+    nosplit = [g for g in ggufs if "-of-" not in g]
+    return (pref or nosplit or ggufs)[0]