update

backend/config.py

@@ -114,9 +114,17 @@ FRAGMENT_MIN_COS = 0.15
 # 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
+# Per-topic higher than the process cap: it is SHARED between process and API tier — at 12 a
+# single-topic run (the normal case) would never benefit from the cheap API tier.
+MAX_CONCURRENT_AGENTS_PER_TOPIC = int(os.getenv("MAX_CONCURRENT_AGENTS_PER_TOPIC", "24"))  # per topic
+# Direct-API text calls (agents._run_text_api): ~0 RAM, only network — own, higher global cap.
+MAX_CONCURRENT_API_AGENTS = int(os.getenv("MAX_CONCURRENT_API_AGENTS", "28"))
 MAX_CONCURRENT_INTERACTIVE = 8
 
+# RAM guard for opencode spawns: below this free share (MemAvailable/MemTotal in %) new
+# processes wait instead of starting (agents._ram_gate). 0 = off.
+RAM_MIN_FREE_PCT = int(os.getenv("RAM_MIN_FREE_PCT", "20"))
+
 # Grace window of the consensus races (blocks, guide, OnePager): after the first
 # valid result the remaining agents may still become done for this many seconds
 # (kill only once the minimum is already in).
@@ -186,6 +194,12 @@ ARTEFACT_CHUNK_SUBS = 25   # flashcards/examples bulk chunk
 FACTS_CHECK_PANEL = 3      # judges per facts-check chunk (majority)
 CONSOLIDATION_PANEL = 3    # mapping judges per chunk
 SUBBLOCK_PANEL = 3         # judges in the subblock clarification
+# Board 2, verschmolzene Calls (block_calls.py): Panel-Größen der neuen Struktur.
+# Konsens braucht ≥2 unabhängige Nennungen bzw. Einstimmigkeit — 2 ist das Minimum,
+# 3 kauft Robustheit für +50 % Tokens auf dem jeweiligen Segment.
+GEN_PANEL = 2              # unabhängige Generator-Calls pro Block
+VERIFY_PANEL = 2           # unabhängige Prüfer-Calls pro Block (+ Ersatz bei 1 Ausfall)
+ART_SPLIT_SUBS = 20        # Artefakt-Generator splittet ab so vielen Subs in 2 parallele Calls
 FILTER_RECHECK_PANEL = 3   # judges in the survivor-recheck
 MAX_WRITER_ROUNDS = 2      # guide coverage→writer loop cap
 GATE_FIX_MIN = 3           # fact-gate: unbelegt-claims below this → log only (falsch fixt immer)
@@ -225,6 +239,12 @@ TIMEOUTS = {
     "relevance_check":    (150, 8),  # judge decides contested relevance in the chunk
     "question_pattern":      (300, 15),  # question patterns per block (subblocks × types)
     "question_pattern_check": (150, 8),  # critic cleans up the pattern table per block
+    # Board 2, verschmolzene Calls: größere Outputs pro Call, dafür wenige Segmente
+    "generate":       (450, 0),   # Subs+Facts+Level in einem (Sub-Zahl vorab unbekannt)
+    "verify":         (300, 10),  # Audit über alle Subs (n = Subs), key points gekappt
+    "fix":            (300, 15),  # Korrekturen + Lücken (n = Aufträge)
+    "artefakt":       (450, 15),  # Fragen+Karten+Beispiele (n = Subs)
+    "artefakt_check": (200, 8),   # Beispiel-Verifikation + Fragen-Kritik (n = Subs)
     "writer":        (450, 60),   # per section — split keeps sections ≤30 subs
     "lese_check":    (300, 10),   # per section in the package
     # guide board (per card = one block)