diff --git a/apps/api/internal/aiprovider/catalog.go b/apps/api/internal/aiprovider/catalog.go
index 243ffa6..ac362ad 100644
--- a/apps/api/internal/aiprovider/catalog.go
+++ b/apps/api/internal/aiprovider/catalog.go
@@ -30,8 +30,17 @@ var PopularCatalog = []PopularProvider{
Name: "openai",
Label: "OpenAI",
BaseURL: "https://api.openai.com/v1",
- DefaultModel: "gpt-4o-mini",
- Models: []string{"gpt-4o-mini", "gpt-4o", "gpt-4.1-mini", "gpt-4.1"},
+ DefaultModel: "gpt-5.6-luna",
+ Models: []string{
+ "gpt-5.6-luna",
+ "gpt-5.6-terra",
+ "gpt-5.6-sol",
+ "gpt-5.6",
+ "gpt-4o-mini",
+ "gpt-4o",
+ "gpt-4.1-mini",
+ "gpt-4.1",
+ },
},
{
Name: "google",
@@ -65,8 +74,8 @@ var PopularCatalog = []PopularProvider{
Name: "openrouter",
Label: "OpenRouter",
BaseURL: "https://openrouter.ai/api/v1",
- DefaultModel: "openai/gpt-4o-mini",
- Models: []string{"openai/gpt-4o-mini", "anthropic/claude-sonnet-4", "google/gemini-2.0-flash-001"},
+ DefaultModel: "openai/gpt-5.6-luna",
+ Models: []string{"openai/gpt-5.6-luna", "openai/gpt-5.6-terra", "openai/gpt-4o-mini", "anthropic/claude-sonnet-4", "google/gemini-2.0-flash-001"},
},
}
diff --git a/apps/api/internal/processing/llm_json.go b/apps/api/internal/processing/llm_json.go
index e83c927..a978cf7 100644
--- a/apps/api/internal/processing/llm_json.go
+++ b/apps/api/internal/processing/llm_json.go
@@ -15,24 +15,33 @@ const (
// 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
+ // GPT-5.6 max_completion_tokens includes hidden reasoning tokens — keep
+ // this high so multi-section HTML descriptions are not truncated.
+ MaxTokensEnhance = 24576
// MaxTokensEnhanceRetry is the one-shot length-cap bump ceiling used by
// OpenAIClient when finish_reason=length yields empty or unparseable JSON.
- MaxTokensEnhanceRetry = 32768
+ MaxTokensEnhanceRetry = 40960
MaxTokensSEO = 180
MaxTokensCampaign = 650
- MaxProductDescRunes = 400
- MaxAttrKeys = 10
- MaxAttrValueRunes = 60
- MaxBrandInjectRunes = 500
- MaxCampaignProducts = 8
- MaxCampaignNameRunes = 80
+ // MaxProductDescRunes caps source description text fed into enhance prompts.
+ // Kept large enough for existing long/HTML descriptions so the model can
+ // rewrite factually instead of inventing from a truncated stub.
+ MaxProductDescRunes = 3500
+ MaxAttrKeys = 10
+ MaxAttrValueRunes = 60
+ MaxBrandInjectRunes = 500
+ MaxCampaignProducts = 8
+ MaxCampaignNameRunes = 80
)
// CompleteOptions tunes a single chat completion for structured tasks.
type CompleteOptions struct {
MaxTokens int
- Temperature float64 // 0 → client default (≤0.3 for structured)
+ Temperature float64 // 0 → client default (≤0.3 for structured); omitted on GPT-5/o-series
+ // ReasoningEffort is Chat Completions reasoning_effort for GPT-5 / o-series.
+ // Empty → auto: "none" for tiny/probe budgets, "low" for content-heavy calls
+ // (long formula HTML) so visible tokens are not eaten by deep reasoning.
+ ReasoningEffort string
}
// CompleterWithOptions is optional; OpenAIClient implements it.
@@ -205,7 +214,9 @@ func CompactBrandPrompt(block string) string {
return truncateRunes(SanitizeText(block), MaxBrandInjectRunes)
}
-// ProductEnhanceUser builds a short user prompt for title/description enhance.
+// ProductEnhanceUser builds the compact user prompt for title/description enhance.
+// Source description is capped at MaxProductDescRunes so long/HTML inputs stay available
+// for factual rewrite (not invented from a short stub).
func ProductEnhanceUser(category, name, description string, attrs map[string]any) string {
var b strings.Builder
b.WriteString("Category: ")
diff --git a/apps/api/internal/processing/llm_json_test.go b/apps/api/internal/processing/llm_json_test.go
index 41d1e65..cddaca2 100644
--- a/apps/api/internal/processing/llm_json_test.go
+++ b/apps/api/internal/processing/llm_json_test.go
@@ -109,11 +109,23 @@ func TestCompleteJSON_returnsRetryError(t *testing.T) {
}
func TestProductEnhanceUser_truncates(t *testing.T) {
- long := strings.Repeat("x", 2000)
+ long := strings.Repeat("x", MaxProductDescRunes+500)
u := ProductEnhanceUser("Cat", "Name", long, map[string]any{"brand": "B"})
- if len([]rune(u)) > 900 {
+ // Category/Name/Attrs overhead + MaxProductDescRunes of description.
+ if len([]rune(u)) > MaxProductDescRunes+200 {
t.Fatalf("user too long: %d", len([]rune(u)))
}
+ descIdx := strings.Index(u, "Description: ")
+ if descIdx < 0 {
+ t.Fatalf("missing Description: in %q", u)
+ }
+ descPart := u[descIdx+len("Description: "):]
+ if i := strings.Index(descPart, "\n"); i >= 0 {
+ descPart = descPart[:i]
+ }
+ if len([]rune(descPart)) > MaxProductDescRunes {
+ t.Fatalf("description runes=%d want <= %d", len([]rune(descPart)), MaxProductDescRunes)
+ }
if !strings.Contains(u, "brand") {
t.Fatalf("%s", u)
}
diff --git a/apps/api/internal/processing/openai.go b/apps/api/internal/processing/openai.go
index 5137ee5..4f44baf 100644
--- a/apps/api/internal/processing/openai.go
+++ b/apps/api/internal/processing/openai.go
@@ -223,13 +223,6 @@ func classifyOpenAITransportErr(model string, err error) (error, bool) {
return err, true
}
-type chatRequest struct {
- Model string `json:"model"`
- Messages []chatMessage `json:"messages"`
- Temperature float64 `json:"temperature"`
- MaxTokens int `json:"max_tokens,omitempty"`
-}
-
type chatMessage struct {
Role string `json:"role"`
Content string `json:"content"`
@@ -283,6 +276,7 @@ func (c *OpenAIClient) CompleteWithOptions(ctx context.Context, system, user str
temp = 0.3
}
maxTok := opts.MaxTokens
+ effort := opts.ReasoningEffort
var lastErr error
for attempt := 0; attempt <= c.MaxRetries; attempt++ {
@@ -298,7 +292,7 @@ func (c *OpenAIClient) CompleteWithOptions(ctx context.Context, system, user str
if err := c.waitRate(ctx); err != nil {
return Completion{}, err
}
- comp, retryable, err := c.doComplete(ctx, system, user, temp, maxTok)
+ comp, retryable, err := c.doComplete(ctx, system, user, temp, maxTok, effort)
if err == nil {
return comp, nil
}
@@ -471,16 +465,11 @@ func (c *OpenAIClient) doEmbed(ctx context.Context, texts []string) ([][]float32
return out, false, nil
}
-func (c *OpenAIClient) doComplete(ctx context.Context, system, user string, temperature float64, maxTokens int) (Completion, bool, error) {
- reqBody := chatRequest{
- Model: c.Model,
- Messages: []chatMessage{
- {Role: "system", Content: system},
- {Role: "user", Content: user},
- },
- Temperature: temperature,
- MaxTokens: maxTokens,
- }
+func (c *OpenAIClient) doComplete(ctx context.Context, system, user string, temperature float64, maxTokens int, reasoningEffort string) (Completion, bool, error) {
+ reqBody := buildChatCompletionBody(c.Model, []chatMessage{
+ {Role: "system", Content: system},
+ {Role: "user", Content: user},
+ }, temperature, maxTokens, reasoningEffort)
body, err := json.Marshal(reqBody)
if err != nil {
return Completion{}, false, err
diff --git a/apps/api/internal/processing/openai_chat_params.go b/apps/api/internal/processing/openai_chat_params.go
new file mode 100644
index 0000000..ea8bb16
--- /dev/null
+++ b/apps/api/internal/processing/openai_chat_params.go
@@ -0,0 +1,97 @@
+package processing
+
+import (
+ "strings"
+)
+
+// openAIChatUsesMaxCompletionTokens reports models that reject legacy max_tokens
+// on Chat Completions (o-series + GPT-5 family including GPT-5.6).
+// See https://developers.openai.com/api/reference/resources/chat
+func openAIChatUsesMaxCompletionTokens(model string) bool {
+ return openAIReasoningChatModel(model)
+}
+
+// openAIChatOmitsCustomTemperature reports models that reject non-default
+// temperature (only the API default of 1 is accepted, if temperature is sent).
+func openAIChatOmitsCustomTemperature(model string) bool {
+ return openAIReasoningChatModel(model)
+}
+
+// openAIChatSupportsReasoningEffort reports Chat Completions models that accept
+// top-level reasoning_effort (Responses API uses reasoning.effort instead).
+func openAIChatSupportsReasoningEffort(model string) bool {
+ return openAIReasoningChatModel(model)
+}
+
+// openAIReasoningChatModel matches OpenAI reasoning / GPT-5(+).5/5.6 chat models
+// and common snapshot aliases (e.g. gpt-5.6-luna-2026-…).
+func openAIReasoningChatModel(model string) bool {
+ m := strings.ToLower(strings.TrimSpace(model))
+ if m == "" {
+ return false
+ }
+ // Strip provider prefixes used by OpenRouter-style gateways.
+ if i := strings.LastIndex(m, "/"); i >= 0 {
+ m = m[i+1:]
+ }
+ switch {
+ case strings.HasPrefix(m, "o1"),
+ strings.HasPrefix(m, "o3"),
+ strings.HasPrefix(m, "o4"),
+ strings.HasPrefix(m, "gpt-5"),
+ strings.HasPrefix(m, "chatgpt-o"):
+ return true
+ default:
+ return false
+ }
+}
+
+// defaultReasoningEffort picks Chat Completions reasoning_effort when unset.
+// Long formula-HTML enhance uses "low" so max_completion_tokens prefers visible
+// HTML over deep hidden reasoning. Tiny probes use "none" for speed/cost.
+func defaultReasoningEffort(model string, maxTokens int, explicit string) string {
+ explicit = strings.ToLower(strings.TrimSpace(explicit))
+ if explicit != "" {
+ return explicit
+ }
+ if !openAIChatSupportsReasoningEffort(model) {
+ return ""
+ }
+ if maxTokens <= 0 || maxTokens < 512 {
+ return "none"
+ }
+ return "low"
+}
+
+// chatCompletionBody is the Chat Completions JSON body. Temperature is a pointer
+// so reasoning models can omit it entirely (omitempty).
+type chatCompletionBody struct {
+ Model string `json:"model"`
+ Messages []chatMessage `json:"messages"`
+ Temperature *float64 `json:"temperature,omitempty"`
+ MaxTokens int `json:"max_tokens,omitempty"`
+ MaxCompletionTokens int `json:"max_completion_tokens,omitempty"`
+ ReasoningEffort string `json:"reasoning_effort,omitempty"`
+}
+
+// buildChatCompletionBody maps CompleteOptions onto the correct Chat Completions
+// fields for classic vs GPT-5 / o-series models.
+func buildChatCompletionBody(model string, messages []chatMessage, temperature float64, maxTokens int, reasoningEffort string) chatCompletionBody {
+ body := chatCompletionBody{
+ Model: model,
+ Messages: messages,
+ }
+ if openAIChatUsesMaxCompletionTokens(model) {
+ if maxTokens > 0 {
+ body.MaxCompletionTokens = maxTokens
+ }
+ } else if maxTokens > 0 {
+ body.MaxTokens = maxTokens
+ }
+ if !openAIChatOmitsCustomTemperature(model) {
+ t := temperature
+ body.Temperature = &t
+ }
+ body.ReasoningEffort = defaultReasoningEffort(model, maxTokens, reasoningEffort)
+ return body
+}
diff --git a/apps/api/internal/processing/openai_chat_params_test.go b/apps/api/internal/processing/openai_chat_params_test.go
new file mode 100644
index 0000000..06f1715
--- /dev/null
+++ b/apps/api/internal/processing/openai_chat_params_test.go
@@ -0,0 +1,123 @@
+package processing
+
+import (
+ "encoding/json"
+ "testing"
+)
+
+func TestOpenAIReasoningChatModel(t *testing.T) {
+ t.Parallel()
+ cases := []struct {
+ model string
+ want bool
+ }{
+ {"gpt-5.6-luna", true},
+ {"gpt-5.6-terra", true},
+ {"gpt-5.6-sol", true},
+ {"gpt-5.6", true},
+ {"GPT-5.4", true},
+ {"openai/gpt-5.6-luna", true},
+ {"o1-mini", true},
+ {"o3", true},
+ {"o4-mini", true},
+ {"gpt-4o-mini", false},
+ {"gpt-4.1", false},
+ {"mock-llm", false},
+ {"", false},
+ }
+ for _, tc := range cases {
+ if got := openAIReasoningChatModel(tc.model); got != tc.want {
+ t.Fatalf("openAIReasoningChatModel(%q)=%v want %v", tc.model, got, tc.want)
+ }
+ }
+}
+
+func TestBuildChatCompletionBody_gpt56Luna(t *testing.T) {
+ t.Parallel()
+ body := buildChatCompletionBody("gpt-5.6-luna", []chatMessage{
+ {Role: "system", Content: "sys"},
+ {Role: "user", Content: "ping"},
+ }, 0.2, MaxTokensEnhance, "low")
+ raw, err := json.Marshal(body)
+ if err != nil {
+ t.Fatal(err)
+ }
+ var m map[string]any
+ if err := json.Unmarshal(raw, &m); err != nil {
+ t.Fatal(err)
+ }
+ if _, ok := m["temperature"]; ok {
+ t.Fatalf("temperature must be omitted for gpt-5.6-luna, got %v", m["temperature"])
+ }
+ if _, ok := m["max_tokens"]; ok {
+ t.Fatalf("max_tokens must be omitted for gpt-5.6-luna, got %v", m["max_tokens"])
+ }
+ gotMax, _ := m["max_completion_tokens"].(float64)
+ if int(gotMax) != MaxTokensEnhance {
+ t.Fatalf("max_completion_tokens=%v want %d", gotMax, MaxTokensEnhance)
+ }
+ if m["reasoning_effort"] != "low" {
+ t.Fatalf("reasoning_effort=%v want low", m["reasoning_effort"])
+ }
+}
+
+func TestBuildChatCompletionBody_classicModel(t *testing.T) {
+ t.Parallel()
+ body := buildChatCompletionBody("gpt-4o-mini", []chatMessage{
+ {Role: "user", Content: "hi"},
+ }, 0.2, 350, "")
+ raw, err := json.Marshal(body)
+ if err != nil {
+ t.Fatal(err)
+ }
+ var m map[string]any
+ if err := json.Unmarshal(raw, &m); err != nil {
+ t.Fatal(err)
+ }
+ if _, ok := m["max_completion_tokens"]; ok {
+ t.Fatalf("max_completion_tokens unexpected for classic model: %v", m)
+ }
+ gotMax, _ := m["max_tokens"].(float64)
+ if int(gotMax) != 350 {
+ t.Fatalf("max_tokens=%v want 350", gotMax)
+ }
+ gotTemp, _ := m["temperature"].(float64)
+ if gotTemp != 0.2 {
+ t.Fatalf("temperature=%v want 0.2", gotTemp)
+ }
+ if _, ok := m["reasoning_effort"]; ok {
+ t.Fatalf("reasoning_effort unexpected for classic model: %v", m)
+ }
+}
+
+func TestBuildChatCompletionBody_probeDefaultsNoneEffort(t *testing.T) {
+ t.Parallel()
+ body := buildChatCompletionBody("gpt-5.6-luna", []chatMessage{
+ {Role: "user", Content: "ping"},
+ }, 0.2, 0, "")
+ if body.ReasoningEffort != "none" {
+ t.Fatalf("reasoning_effort=%q want none for probe", body.ReasoningEffort)
+ }
+ if body.MaxCompletionTokens != 0 || body.MaxTokens != 0 {
+ t.Fatalf("expected no token cap on probe, got %+v", body)
+ }
+ if body.Temperature != nil {
+ t.Fatalf("temperature must be nil for probe on gpt-5.6")
+ }
+}
+
+func TestDefaultReasoningEffort(t *testing.T) {
+ t.Parallel()
+ if got := defaultReasoningEffort("gpt-5.6-luna", MaxTokensEnhance, ""); got != "low" {
+ t.Fatalf("enhance default=%q", got)
+ }
+ if got := defaultReasoningEffort("gpt-5.6-luna", 0, ""); got != "none" {
+ t.Fatalf("probe default=%q", got)
+ }
+ if got := defaultReasoningEffort("gpt-5.6-luna", 100, "medium"); got != "medium" {
+ t.Fatalf("explicit=%q", got)
+ }
+ if got := defaultReasoningEffort("gpt-4o", 1000, ""); got != "" {
+ t.Fatalf("classic=%q", got)
+ }
+}
diff --git a/apps/api/internal/processing/openai_test.go b/apps/api/internal/processing/openai_test.go
index 72c6f73..39529b9 100644
--- a/apps/api/internal/processing/openai_test.go
+++ b/apps/api/internal/processing/openai_test.go
@@ -179,6 +179,52 @@ func TestOpenAIClient_doComplete_emptyContentWithReasoningJSON(t *testing.T) {
}
}
+func TestOpenAIClient_doComplete_gpt56LunaUsesMaxCompletionTokens(t *testing.T) {
+ t.Parallel()
+ srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+ var req map[string]any
+ _ = json.NewDecoder(r.Body).Decode(&req)
+ if _, ok := req["temperature"]; ok {
+ t.Errorf("temperature must be omitted for gpt-5.6-luna")
+ }
+ if _, ok := req["max_tokens"]; ok {
+ t.Errorf("max_tokens must be omitted for gpt-5.6-luna")
+ }
+ gotMax, _ := req["max_completion_tokens"].(float64)
+ if int(gotMax) != MaxTokensEnhance {
+ t.Errorf("max_completion_tokens=%v want %d", gotMax, MaxTokensEnhance)
+ }
+ if req["reasoning_effort"] != "low" {
+ t.Errorf("reasoning_effort=%v want low", req["reasoning_effort"])
+ }
+ _ = json.NewEncoder(w).Encode(map[string]any{
+ "model": "gpt-5.6-luna",
+ "choices": []map[string]any{{
+ "finish_reason": "stop",
+ "message": map[string]any{
+ "role": "assistant",
+ "content": `{"name":"Acme Monitor","description":"Acme Monitor
IPS panel for office use.
"}`,
+ },
+ }},
+ "usage": map[string]int{"prompt_tokens": 20, "completion_tokens": 40, "total_tokens": 60},
+ })
+ }))
+ defer srv.Close()
+ c := NewOpenAIClient("test-key", srv.URL, "gpt-5.6-luna", 0, 1)
+ c.HTTPClient = srv.Client()
+ comp, err := c.CompleteWithOptions(context.Background(), "sys", "user", CompleteOptions{
+ MaxTokens: MaxTokensEnhance,
+ Temperature: DefaultStructuredTemp,
+ ReasoningEffort: "low",
+ })
+ if err != nil {
+ t.Fatal(err)
+ }
+ if !strings.Contains(comp.Text, "Acme Monitor") {
+ t.Fatalf("text=%q", comp.Text)
+ }
+}
+
func TestOpenAIClient_doComplete_emptyContentLengthRetriesWithBudget(t *testing.T) {
t.Parallel()
calls := 0
diff --git a/apps/api/internal/processing/sanitize.go b/apps/api/internal/processing/sanitize.go
index 9d1b404..c9747a6 100644
--- a/apps/api/internal/processing/sanitize.go
+++ b/apps/api/internal/processing/sanitize.go
@@ -11,7 +11,7 @@ import (
const maxPromptFieldRunes = 4000
// maxOutputFieldRunes bounds stored / polled model text (titles + multi-section
-// formula HTML descriptions). Must stay well above MaxTokensEnhance (~16k tokens)
+// formula HTML descriptions). Must stay well above MaxTokensEnhance (~24k tokens)
// so SanitizeOutput does not chop JSON completion bodies or A1 HTML mid-string.
const maxOutputFieldRunes = 120000
diff --git a/apps/api/internal/processing/steps.go b/apps/api/internal/processing/steps.go
index 7add4b8..8bbb976 100644
--- a/apps/api/internal/processing/steps.go
+++ b/apps/api/internal/processing/steps.go
@@ -800,8 +800,9 @@ func (e *Engine) enhance(ctx context.Context, in ProductInput, category string,
logEnhancePrompt(in, catUID, catName, system, user)
started := time.Now()
comp, obj, err := CompleteJSON(ctx, e.Completer, system, user, CompleteOptions{
- MaxTokens: MaxTokensEnhance,
- Temperature: DefaultStructuredTemp,
+ MaxTokens: MaxTokensEnhance,
+ Temperature: DefaultStructuredTemp,
+ ReasoningEffort: "low",
})
elapsed := time.Since(started)
if err != nil {
@@ -891,8 +892,9 @@ func (e *Engine) enhance(ctx context.Context, in ProductInput, category string,
retryUser := user + descriptionFormulaRetrySuffix
retryStarted := time.Now()
comp2, obj2, err2 := CompleteJSON(ctx, e.Completer, system, retryUser, CompleteOptions{
- MaxTokens: MaxTokensEnhance,
- Temperature: DefaultStructuredTemp,
+ MaxTokens: MaxTokensEnhance,
+ Temperature: DefaultStructuredTemp,
+ ReasoningEffort: "low",
})
elapsed += time.Since(retryStarted)
comp.PromptTokens += comp2.PromptTokens
diff --git a/apps/web/src/lib/admin-ai-roles.ts b/apps/web/src/lib/admin-ai-roles.ts
index d378cf8..83f3fb0 100644
--- a/apps/web/src/lib/admin-ai-roles.ts
+++ b/apps/web/src/lib/admin-ai-roles.ts
@@ -31,7 +31,7 @@ export const AI_ROLE_META: Record<
processing: {
label: "Processing",
description: "Product pipeline chat/completions (titles, descriptions, enhance).",
- modelPlaceholder: "gpt-4o-mini"
+ modelPlaceholder: "gpt-5.6-luna"
},
vectorization: {
label: "Vectorization",
@@ -42,13 +42,13 @@ export const AI_ROLE_META: Record<
label: "Docs / API",
description:
"Future docs/API assistant slot — /docs Ask stays rule-based and must never call this role.",
- modelPlaceholder: "gpt-4o-mini"
+ modelPlaceholder: "gpt-5.6-luna"
},
support: {
label: "Support",
description:
"Ticket auto-reply AI fallback — configure provider/key/model here; enable delivery in Support knowledge → Auto-reply.",
- modelPlaceholder: "gpt-4o-mini"
+ modelPlaceholder: "gpt-5.6-luna"
}
};
diff --git a/docs/local-llm-tuning.md b/docs/local-llm-tuning.md
index 2046adc..6166ad9 100644
--- a/docs/local-llm-tuning.md
+++ b/docs/local-llm-tuning.md
@@ -16,27 +16,29 @@ Restart **both** `api` and `worker` after changing `OPENAI_*`. Worker logs shoul
## Product enhance model choice
-Admin AI role for **product enhance** should prefer a **fast non-reasoning chat model** that emits `message.content` (JSON) under budget.
+Admin AI role for **product enhance** should prefer a model that emits `message.content` (JSON) under budget. Recommended OpenAI: **`gpt-5.6-luna`** (high-volume / cost-efficient GPT-5.6 tier).
-Reasoning / “code-fast” style models often spend the entire `max_tokens` budget in `reasoning_content`, then return `finish_reason=length` with empty or truncated JSON. Descrybe retries once at `MaxTokensEnhanceRetry` and may synthesize formula HTML as salvage — that is **not** live LLM formula compliance.
+Reasoning / “code-fast” style models often spend the entire completion budget in `reasoning_content`, then return `finish_reason=length` with empty or truncated JSON. Descrybe retries once at `MaxTokensEnhanceRetry` and may synthesize formula HTML as salvage — that is **not** live LLM formula compliance.
-If enhance logs show `finish_reason=length` with empty/truncated content even after the retry, switch Admin enhance to a non-reasoning model.
+For GPT-5 / o-series Chat Completions, the client sends `max_completion_tokens` (not `max_tokens`), omits custom `temperature`, and sets `reasoning_effort` (`none` for probes, `low` for long formula-HTML enhance) so visible HTML is not eaten by deep reasoning.
+
+If enhance logs show `finish_reason=length` with empty/truncated content even after the retry, switch Admin enhance to a lower-reasoning or classic chat model (e.g. `gpt-4o-mini`).
## Hardening defaults (code)
| Constant | Value | Where |
|----------|------:|-------|
-| `DefaultStructuredTemp` | `0.2` (capped ≤0.3) | `processing/llm_json.go`, `openai.go` |
-| `MaxTokensEnhance` | `16384` | product title/description (reasoning + formula HTML JSON) |
-| `MaxTokensEnhanceRetry` | `32768` | one-shot bump on `finish_reason=length` empty/truncated JSON |
+| `DefaultStructuredTemp` | `0.2` (capped ≤0.3; omitted on GPT-5/o-series) | `processing/llm_json.go`, `openai.go` |
+| `MaxTokensEnhance` | `24576` | product title + multi-section HTML description JSON |
+| `MaxTokensEnhanceRetry` | `40960` | one-shot bump on `finish_reason=length` empty/truncated JSON |
| `MaxTokensSEO` | `180` | meta title/description |
| `MaxTokensCampaign` | `650` | email JSON |
-| `MaxProductDescRunes` | `400` | user context |
+| `MaxProductDescRunes` | `3500` | source description in enhance user context |
| `MaxAttrKeys` / `MaxAttrValueRunes` | `10` / `60` | attrs in enhance prompt |
| `MaxBrandInjectRunes` | `500` | brand kit block |
| `MaxCampaignProducts` | `8` | campaign product list |
-OpenAI-compatible chat completions here use a single `max_tokens` completion budget (no separate reasoning vs content split on typical gateways). Length-cap handling lives in `OpenAIClient.doComplete` / `CompleteWithOptions`.
+Classic OpenAI-compatible chat completions use `max_tokens`. GPT-5 / o-series use `max_completion_tokens` (budget includes hidden reasoning tokens). Length-cap handling lives in `OpenAIClient.doComplete` / `CompleteWithOptions`.
## Prompt style
@@ -49,7 +51,7 @@ OpenAI-compatible chat completions here use a single `max_tokens` completion bud
`processing.CompleteJSON`:
-1. Call Completer with `max_tokens` + low temperature
+1. Call Completer with completion budget + low temperature (classic) / GPT-5.6 param mapping
2. `StripJSONFences` / isolate `{…}`
3. On parse fail → **one retry** with `INVALID. Reply with ONLY one JSON object…`
4. Call sites fall back to originals/templates instead of storing garbage prose as titles