Make reasoning levels enforce real token budgets

2026-09-01 13:55:11 +02:00
parent 5f793020b0
commit b3e86cc7ae
5 changed files with 52 additions and 21 deletions
@@ -194,11 +194,8 @@ services:
      - --jinja
      - --reasoning
      - auto
-      # Bound each individual thinking phase. Long agent jobs can still use
-      # many phases around tool calls, but one degenerate reasoning loop can
-      # no longer consume the complete response budget indefinitely.
-      - --reasoning-budget
-      - "8192"
+      # No fixed --reasoning-budget here: the router supplies a real budget
+      # per request from the client's reasoning_effort selection.
      - --reasoning-preserve
      - --host
      - 0.0.0.0