update

backend/config.py

@@ -1,3 +1,4 @@
+import os
 from pathlib import Path
 
 PROJECT_ROOT = Path(__file__).resolve().parent.parent
@@ -8,6 +9,26 @@ DB_PATH = STORAGE_DIR / "creator.db"
 PROJECTS_DIR = PROJECT_ROOT / "projects"
 UNI_DIR = PROJECT_ROOT / "uni"
 
+
+def _load_env(path: Path) -> None:
+    """Mini .env loader (no dependency): KEY=VALUE lines; existing env always wins
+    (`make dev` already exports .env — this covers bare `uvicorn`/pytest starts)."""
+    try:
+        text = path.read_text(encoding="utf-8")
+    except OSError:
+        return
+    for line in text.splitlines():
+        line = line.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, _, value = line.partition("=")
+        key, value = key.strip(), value.strip().strip('"').strip("'")
+        if key and key not in os.environ:
+            os.environ[key] = value
+
+
+_load_env(PROJECT_ROOT / ".env")
+
 MAX_CONCURRENT_GENERATIONS = 10
 
 # Readability gate: deterministic checker (small German complexity model,
@@ -21,12 +42,9 @@ READABILITY_MAX = 3.5          # section too hard when the sentence average is a
 READABILITY_HARD = 4.0         # an individual sentence is "hard" from here on
 READABILITY_HARD_SHARE = 0.30  # … OR when this share of sentences is hard
 
-# Block consolidation: semantic embedding clustering instead of an LLM list merge.
-# A small multilingual sentence embedding (mean-pool) builds the candidate clusters
-# GLOBALLY (no chunk loss) via cosine + union-find. Title variants of the same concept
-# ("Vertex Cover" / "Vertex Cover Definition") merge; the consensus then counts the
-# real readers per cluster (≥2 = consensus). If transformers/torch are missing or the model
-# won't load → embedding silently off, `_consolidate` falls back to the old panel-judge path.
+# Kanban clustering: semantic embeddings drive the online title clustering and the
+# candidate pairs of the pair check. If transformers/torch are missing or the model
+# won't load → embedding silently off (all pairs go to the judge, clusters stay singletons).
 EMBEDDING_AKTIV = True
 EMBEDDING_MODELL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"  # CPU, multilingual, ~470 MB
 # Stronger (larger) CPU alternative if needed: "BAAI/bge-m3".
@@ -66,10 +84,12 @@ GROUP_RECONCILE_FLOOR = 0.75
 # chain (random-pair baseline), set BELOW the legitimate heterogeneous minimum so it never kills a real model.
 GROUP_MIN_COS_FLOOR = 0.15
 
-# Cap for concurrent CLI agent processes (across all generations).
-# Own lane for interactive calls (chat, elements) so they don't hang behind
-# running writers in the queue.
-MAX_CONCURRENT_AGENTS = 10
+# Caps for concurrent CLI agent processes (env-overridable). Two nested limits, both always active:
+# a per-topic cap and a global cap across all topics. Defaults 10/10 = previous behavior (global
+# dominates). Locally raise the global cap to actually parallelize across topics (per-topic stays 10).
+# Own lane for interactive calls (chat, elements) so they don't hang behind running writers.
+MAX_CONCURRENT_AGENTS = int(os.getenv("MAX_CONCURRENT_AGENTS", "16"))                      # global, all topics
+MAX_CONCURRENT_AGENTS_PER_TOPIC = int(os.getenv("MAX_CONCURRENT_AGENTS_PER_TOPIC", "12"))  # per topic
 MAX_CONCURRENT_INTERACTIVE = 8
 
 # Grace window of the consensus races (blocks, guide, OnePager): after the first
@@ -123,6 +143,10 @@ TIMEOUTS = {
     "question_pattern_check": (300, 10),  # critic cleans up the pattern table per block
     "writer":        (600, 120),  # per section in the chunk
     "lese_check":    (300, 10),   # per section in the package
+    # guide board (per card = one block)
+    "lernziele":     (300, 5),    # backward-design objectives per block
+    "fakten_gate":   (600, 5),    # CoVe claim check per block
+    "coverage":      (300, 5),    # objective↔section mapping per block
 }
 
 # Purpose per format — flows into the outline judge (what the guide should achieve).
@@ -169,3 +193,27 @@ PROVIDERS = {
         "check_url": "http://localhost:11434/api/tags",  # Ollama reachable?
     },
 }
+
+# Role routing ACROSS provider stacks: generation (quick/guide) and judging (judge)
+# may run on different providers within ONE run — judge model ≠ generator model
+# (research-backed: cross-model judging avoids self-preference bias).
+# Value: "" = provider of the run; "minimax" = that stack's role model;
+# "provider:model" = explicit model override.
+ROLE_ROUTING = {
+    "quick": os.getenv("ROLE_QUICK", "minimax"),
+    "judge": os.getenv("ROLE_JUDGE", "claude"),
+    "guide": os.getenv("ROLE_GUIDE", "minimax"),
+    "fast":  os.getenv("ROLE_FAST", ""),
+}
+
+
+def resolve_role(run_provider: str, role: str) -> tuple[str, str]:
+    """→ (provider, model) for one agent call. Pure routing, no availability check —
+    the caller (agents.run_agent) falls back to run_provider if the target is unavailable."""
+    target = ROLE_ROUTING.get(role, "") or run_provider
+    provider, _, model = target.partition(":")
+    if provider not in PROVIDERS:
+        provider, model = run_provider, ""
+    if not model:
+        model = PROVIDERS.get(provider, {}).get(role, "")
+    return provider, model