fix

apps/api/internal/processing/llm_json.go

@@ -8,21 +8,25 @@ import (
 	"strings"
 )
 
-// Local / weak-model defaults (8k-class context). See docs/local-llm-tuning.md.
+// Local / weak-model defaults. See docs/local-llm-tuning.md.
 const (
 	DefaultStructuredTemp = 0.2
-	// Reasoning-capable OpenAI-compatible models (e.g. code-fast / Qwen3) spend
-	// hundreds–thousands of tokens in reasoning_content before writing message.content.
-	// 350 capped mid-thought → empty content → "empty model response".
-	MaxTokensEnhance     = 4096
-	MaxTokensSEO         = 180
-	MaxTokensCampaign    = 650
-	MaxProductDescRunes  = 400
-	MaxAttrKeys          = 10
-	MaxAttrValueRunes    = 60
-	MaxBrandInjectRunes  = 500
-	MaxCampaignProducts  = 8
-	MaxCampaignNameRunes = 80
+	// Reasoning-capable OpenAI-compatible models (e.g. OverloadedBot code-fast)
+	// spend thousands of tokens in reasoning_content before message.content.
+	// Formula HTML JSON often needs a large completion budget; 4096 hit
+	// finish_reason=length with empty/truncated content in live enhance.
+	MaxTokensEnhance = 16384
+	// MaxTokensEnhanceRetry is the one-shot length-cap bump ceiling used by
+	// OpenAIClient when finish_reason=length yields empty or unparseable JSON.
+	MaxTokensEnhanceRetry = 32768
+	MaxTokensSEO          = 180
+	MaxTokensCampaign     = 650
+	MaxProductDescRunes   = 400
+	MaxAttrKeys           = 10
+	MaxAttrValueRunes     = 60
+	MaxBrandInjectRunes   = 500
+	MaxCampaignProducts   = 8
+	MaxCampaignNameRunes  = 80
 )
 
 // CompleteOptions tunes a single chat completion for structured tasks.