fix(codex): send reasoning.context=all_turns on Responses-Lite

changelog/unreleased/codex-lite-reasoning-context.md

@@ -0,0 +1,7 @@
+### English
+
+- Codex Responses-Lite models (such as GPT-5.6) now send `reasoning.context: "all_turns"`, fixing the upstream 400 "requires reasoning.context to be all_turns".
+
+### 中文
+
+- Codex Responses-Lite 模型（如 GPT-5.6）现在会发送 `reasoning.context: "all_turns"`，修复上游返回 400 “requires reasoning.context to be all_turns” 的问题。

internal/providers/codex/responses_lite_test.go

@@ -0,0 +1,44 @@
+package codex
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+// Regression for #300: requests sent under the Responses-Lite header must
+// carry reasoning.context=all_turns, or the backend rejects them with 400.
+func TestNormalizeCodexUpstreamLiteForcesReasoningContext(t *testing.T) {
+	for name, raw := range map[string]string{
+		"missing":  "",
+		"effort":   `{"effort":"high","summary":"auto"}`,
+		"override": `{"context":"current_turn"}`,
+	} {
+		t.Run(name, func(t *testing.T) {
+			fields := map[string]json.RawMessage{"input": json.RawMessage(`[]`)}
+			if raw != "" {
+				fields["reasoning"] = json.RawMessage(raw)
+			}
+			if err := normalizeCodexUpstream(fields, "gpt-5.6-sol", true, ""); err != nil {
+				t.Fatal(err)
+			}
+			var reasoning map[string]string
+			if err := json.Unmarshal(fields["reasoning"], &reasoning); err != nil {
+				t.Fatalf("reasoning %s: %v", fields["reasoning"], err)
+			}
+			if reasoning["context"] != "all_turns" || string(fields["parallel_tool_calls"]) != "false" {
+				t.Fatalf("reasoning=%s parallel_tool_calls=%s", fields["reasoning"], fields["parallel_tool_calls"])
+			}
+			if name == "effort" && (reasoning["effort"] != "high" || reasoning["summary"] != "auto") {
+				t.Fatalf("existing members lost: %s", fields["reasoning"])
+			}
+		})
+	}
+
+	fields := map[string]json.RawMessage{"input": json.RawMessage(`[]`)}
+	if err := normalizeCodexUpstream(fields, "gpt-5.5", false, ""); err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := fields["reasoning"]; ok {
+		t.Fatalf("non-lite request gained reasoning: %s", fields["reasoning"])
+	}
+}

internal/providers/codex/upstream.go

@@ -110,6 +110,12 @@ func normalizeCodexUpstream(fields map[string]json.RawMessage, model string, lit
 	// except on lite where the backend still expects the explicit false.
 	if lite {
 		fields["parallel_tool_calls"] = json.RawMessage(`false`)
+		// The lite header also requires reasoning.context=all_turns (#300).
+		reasoning, err := withReasoningContextAllTurns(fields["reasoning"])
+		if err != nil {
+			return err
+		}
+		fields["reasoning"] = reasoning
 	} else if !emptyJSONArray(fields["tools"]) {
 		fields["parallel_tool_calls"] = json.RawMessage(`true`)
 	} else {
@@ -122,6 +128,19 @@ func normalizeCodexUpstream(fields map[string]json.RawMessage, model string, lit
 	return nil
 }
 
+// withReasoningContextAllTurns sets reasoning.context to all_turns and keeps
+// every other reasoning member as sent.
+func withReasoningContextAllTurns(raw json.RawMessage) (json.RawMessage, error) {
+	members := map[string]json.RawMessage{}
+	if !isNullJSON(raw) {
+		if err := json.Unmarshal(raw, &members); err != nil {
+			return nil, err
+		}
+	}
+	members["context"] = json.RawMessage(`"all_turns"`)
+	return json.Marshal(members)
+}
+
 // normalizeServiceTier keeps only the tiers the ChatGPT Codex backend accepts.
 // "fast" is the client spelling of "priority" and is rewritten; anything else
 // is removed rather than forwarded.