From 77ee3325b24df18ce4b1a3b5823fd1c1845d7e0c Mon Sep 17 00:00:00 2001
From: Yuxin Qiao <104957188+Yuxin-Qiao@users.noreply.github.com>
Date: Fri, 7 Aug 2026 17:36:48 +0800
Subject: [PATCH 1/3] fix(openai-chat): inject reasoning placeholder when
 replay cache misses

Closes #1193
---
 src/adapters/openai-chat.ts                  | 18 +++++-
 tests/deepseek-reasoning-replay-gaps.test.ts | 62 ++++++++++++++++++--
 2 files changed, 75 insertions(+), 5 deletions(-)

diff --git a/src/adapters/openai-chat.ts b/src/adapters/openai-chat.ts
index 42a9236574..400e1e2d7e 100644
--- a/src/adapters/openai-chat.ts
+++ b/src/adapters/openai-chat.ts
@@ -347,6 +347,15 @@ function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProviderCon
           // recorded under every call id — join unique texts only.
           if (cached.length > 0) {
             reasoningContent = [...new Set(cached)].join("\n");
+          } else {
+            // Fallback (extends #950, closes #1193): the replay cache is
+            // bounded (64 entries / 256 KiB / 1 h TTL) and always misses on
+            // long sessions, and some tool rounds carry no recorded reasoning
+            // at all. DeepSeek thinking mode rejects ANY tool_call assistant
+            // message missing reasoning_content with HTTP 400, so inject a
+            // minimal placeholder rather than emit a bare continuation the
+            // upstream will reject.
+            reasoningContent = " ";
           }
         }
         if (reasoningContent.length > 0 && modelInList(provider.preserveReasoningContentModels, parsed.modelId)) {
@@ -413,10 +422,17 @@ function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProviderCon
             toolCallId && modelInList(provider.preserveReasoningContentModels, parsed.modelId)
               ? peekReasoningForCall(toolCallId, replayCacheScope)
               : undefined;
+          // Same fallback as the main-assistant path: never emit a bare orphan
+          // tool_call continuation on a thinking-mode provider — inject a
+          // placeholder when the replay cache missed (the bounded cache can
+          // always miss on long sessions), or DeepSeek thinking mode 400s.
+          const orphanReasoning =
+            cachedReasoning
+            ?? (modelInList(provider.preserveReasoningContentModels, parsed.modelId) ? " " : undefined);
           out.push({
             role: "assistant",
             content: emptyAssistantContent(provider),
-            ...(cachedReasoning ? { reasoning_content: cachedReasoning } : {}),
+            ...(orphanReasoning ? { reasoning_content: orphanReasoning } : {}),
             tool_calls: [{
               id: toolCallId,
               type: "function",
diff --git a/tests/deepseek-reasoning-replay-gaps.test.ts b/tests/deepseek-reasoning-replay-gaps.test.ts
index 74787adc52..16d107a04a 100644
--- a/tests/deepseek-reasoning-replay-gaps.test.ts
+++ b/tests/deepseek-reasoning-replay-gaps.test.ts
@@ -135,10 +135,64 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     expect(retry!["reasoning_content"]).toBe(REASONING);
   });
 
-  test("documented non-bug: opaque encrypted-only reasoning is intentionally not replayed", () => {
+  test("GAP D (issue #1193): replay cache MISS on the main assistant path injects a placeholder", () => {
+    // The replay cache is bounded (64 entries / 256 KiB / 1 h TTL) and always
+    // misses on long sessions. DeepSeek thinking mode rejects ANY tool_call
+    // assistant message without reasoning_content (HTTP 400), so a cache miss
+    // must degrade to a minimal placeholder instead of a bare continuation.
+    const { messages } = wireFor([
+      userMessage(),
+      { type: "compaction", encrypted_content: "ocx1:c3VtbWFyeQ==" },
+      functionCallItem(),
+      functionCallOutputItem(),
+    ]);
+    const assistant = toolCallAssistant(messages);
+    expect(assistant).toBeDefined();
+    expect(assistant!["reasoning_content"]).toBe(" ");
+  });
+
+  test("GAP E (issue #1193): replay cache MISS on the orphan-repair path injects a placeholder", () => {
+    // Same invariant for the synthesized orphan tool_call: with nothing
+    // recorded under the call id, repair still must not emit a bare
+    // continuation a thinking-mode provider will 400 on.
+    const { messages } = wireFor([userMessage(), functionCallOutputItem()]);
+    const assistant = toolCallAssistant(messages);
+    expect(assistant).toBeDefined();
+    expect(assistant!["reasoning_content"]).toBe(" ");
+  });
+
+  test("negative control: models outside preserveReasoningContentModels never get a placeholder", () => {
+    // The placeholder fallback is scoped to thinking-mode providers; other
+    // models keep the previous bare-continuation behavior. Use a custom
+    // provider so no registry preset seeds a preserve list.
+    const parsed = parseRequest({ model: "custom-chat/plain-model", input: [userMessage(), functionCallOutputItem()], stream: true });
+    const config: OcxConfig = {
+      port: 10100,
+      defaultProvider: "custom-chat",
+      providers: {
+        "custom-chat": {
+          adapter: "openai-chat",
+          baseUrl: "https://example.invalid/v1",
+          apiKey: "key",
+          models: ["plain-model"],
+        },
+      },
+    };
+    const route = routeModel(config, parsed.modelId);
+    parsed.modelId = route.modelId;
+    const req = createOpenAIChatAdapter(route.provider).buildRequest(parsed as OcxParsedRequest);
+    const { messages } = JSON.parse(req.body as string) as { messages: Array<Record<string, unknown>> };
+    const assistant = toolCallAssistant(messages);
+    expect(assistant).toBeDefined();
+    expect(assistant!["reasoning_content"]).toBeUndefined();
+  });
+
+  test("documented non-bug: opaque encrypted-only reasoning degrades to the placeholder, not invented plaintext", () => {
     // Native (non-ocxr1) encrypted reasoning has no readable text; the parser
-    // deliberately degrades instead of inventing replayable plaintext. Not a
-    // candidate for the opencode-go path (its reasoning is plaintext/ocxr1).
+    // deliberately degrades instead of inventing replayable plaintext. On a
+    // thinking-mode provider the fallback now attaches the minimal placeholder
+    // (issue #1193) rather than replaying anything, so the continuation stays
+    // valid without fabricating reasoning text.
     const { messages } = wireFor([
       userMessage(),
       { type: "reasoning", id: "rs_1", encrypted_content: "some-opaque-blob" },
@@ -147,7 +201,7 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     ]);
     const assistant = toolCallAssistant(messages);
     expect(assistant).toBeDefined();
-    expect(assistant!["reasoning_content"]).toBeUndefined();
+    expect(assistant!["reasoning_content"]).toBe(" ");
   });
 });
 

From 48961a91ea2042fa62906d50c5839a74026973c2 Mon Sep 17 00:00:00 2001
From: Yuxin Qiao <104957188+Yuxin-Qiao@users.noreply.github.com>
Date: Fri, 7 Aug 2026 23:24:29 +0800
Subject: [PATCH 2/3] fix(openai-chat): scope reasoning placeholder to models
 that require it

Address review findings on #1205:

- chatgpt-codex-connector P2: preserveReasoningContentModels only opts
  models into replaying reasoning that exists; MiniMax-M3 low effort maps
  to thinking disabled, so a fabricated placeholder could reach
  non-thinking histories. Add requiresReasoningPlaceholderModels
  (registry/derive/router/oauth/auth-cors plumbing, docs-site table)
  defaulting to the preserve list; minimax/minimax-cn seed [] to opt out.
  Custom preserve-only provider configs keep the #1193 fix via fallback.
- CodeRabbit minor: treat a falsy cache hit as a miss in the orphan-repair
  path (defense-in-depth; the write path already rejects empty strings).

Refs #1193
---
 .../ja/reference/configuration/providers.md   |  1 +
 .../ko/reference/configuration/providers.md   |  1 +
 .../docs/reference/configuration/providers.md |  1 +
 .../ru/reference/configuration/providers.md   |  1 +
 .../reference/configuration/providers.md      |  1 +
 src/adapters/openai-chat.ts                   | 11 ++++--
 src/oauth/index.ts                            |  1 +
 src/oauth/login-cli.ts                        |  1 +
 src/providers/derive.ts                       |  4 +++
 src/providers/registry.ts                     |  8 ++++-
 src/router.ts                                 |  2 ++
 src/server/auth-cors.ts                       |  1 +
 src/types.ts                                  |  8 +++++
 tests/deepseek-reasoning-replay-gaps.test.ts  | 35 +++++++++++++++++++
 14 files changed, 72 insertions(+), 4 deletions(-)

diff --git a/docs-site/src/content/docs/ja/reference/configuration/providers.md b/docs-site/src/content/docs/ja/reference/configuration/providers.md
index 57903e8534..0e83c6a8f2 100644
--- a/docs-site/src/content/docs/ja/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/ja/reference/configuration/providers.md
@@ -92,6 +92,7 @@ namespace 付き combo または routing-profile alias はその namespace prefi
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | API-key プロバイダーのみ(`authMode: "key"`)。オプトインの同一ターゲット 429 リトライ: `retryOn429` が無ければ無効で、オブジェクトがあれば `enabled: false` でない限り有効になります。429 時に待機(上流の `Retry-After` または固定間隔)してから、キー フェイルオーバーの前に同一キーで同一リクエストを再送します — メインのテキストターン回復ループ、Responses passthrough、画像/動画ブリッジ、web-search サイドカー、ターミナル継続要求をすべてカバーします。再送の対象はプリストリームの HTTP 429 応答のみで、カスタム `runTurn` トランスポートは HTTP リトライループの対象外です。`attempts` は最初の 429 以降の同一キー再送回数(合計送信数 = `attempts` + 1)で、メインの回復ループ・ターミナルガード継続・ブリッジ再試行で共有されるリクエスト単位の予算です。`attempts` を使い切っても同一キーでの再送が止まるだけで、通常のキー フェイルオーバーまたは最終エラー処理が利用可能なターゲットに応じて続きます — キー認証の passthrough ワイヤにはフェイルオーバーがないため、使い切った 429 はそのまま返ります。Codex 自体は 429 をリトライしないため、単一キーのプロバイダーでは唯一の防御です。デフォルト: `enabled: true`、`attempts: 3`、`intervalMs: 5000`、`maxIntervalMs: 60000`(1回の待機は `maxIntervalMs` で上限、その上限は 600000)、`respectRetryAfter: true`。 |
 | `autoToolChoiceOnlyModels?` | `string[]` | `tool_choice` が `auto` または `none` のみを受け入れるモデル。強制的な選択は格下げされます。 |
 | `preserveReasoningContentModels?` | `string[]` |チャット履歴に以前のアシスタント `reasoning_content` が必要なモデル。 |
+| `requiresReasoningPlaceholderModels?` | `string[]` | `reasoning_content` を欠いた tool_call 継続を上流が拒否するモデル（DeepSeek thinking モード）。リプレイキャッシュが外れた場合に最小プレースホルダーを注入。未設定時は `preserveReasoningContentModels` を引き継ぎ、`[]` で明示的に無効化。 |
 | `thinkingToggleModels?` | `string[]` |エフォート ラダーではなく `thinking.enabled` を使用してモデルをチャットします。 |
 | `thinkingBudgetModels?` | `string[]` |整数 `thinking_budget` を使用したチャット モデル。労力は予算の一部にマッピングされます。 |
 | `noVisionModels?` | `string[]` |ビジョン サイドカーを通じて送信されるテキストのみのモデル。マッチングでは、Ollama `:size` タグが許容されます。 |
diff --git a/docs-site/src/content/docs/ko/reference/configuration/providers.md b/docs-site/src/content/docs/ko/reference/configuration/providers.md
index 684ea712f5..f3faf2c152 100644
--- a/docs-site/src/content/docs/ko/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/ko/reference/configuration/providers.md
@@ -92,6 +92,7 @@ target도 selector로 재사용할 수 없습니다. raw account id와 email은
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | API-key 프로바이더 전용(`authMode: "key"`). 동일 대상 429 재시도: `retryOn429`가 없으면 기능이 꺼져 있고, 객체가 있으면 `enabled: false`가 아닌 한 활성화됩니다. 429 시 대기(업스트림 `Retry-After` 또는 고정 간격) 후 키 장애 조치 전에 동일 키로 동일 요청을 재전송합니다 — 일반 텍스트 턴 복구 루프, Responses passthrough, 이미지/비디오 브리지, web-search 사이드카, 터미널 연속 요청을 모두 포함합니다. 재전송 대상은 프리스트림 HTTP 429 응답뿐이며, 커스텀 `runTurn` 전송은 HTTP 재시도 루프에서 제외됩니다. `attempts`는 첫 429 이후의 동일 키 재전송 횟수(총 전송 = `attempts` + 1)이며, 메인 복구 루프·터미널 가드 연속 요청·브리지 재시도가 공유하는 요청 단위 예산입니다. `attempts`를 모두 소진해도 동일 키 재전송만 중단되며, 이후에는 일반 키 장애 조치 또는 최종 오류 처리가 사용 가능한 대상에 따라 진행됩니다 — 키 인증 passthrough 와이어에는 장애 조치가 없으므로 소진된 429가 그대로 반환됩니다. Codex 자체는 429를 재시도하지 않으므로 단일 키 프로바이더의 유일한 방어선입니다. 기본값: `enabled: true`, `attempts: 3`, `intervalMs: 5000`, `maxIntervalMs: 60000`(단일 대기는 `maxIntervalMs`로 상한, 그 자체는 600000으로 상한), `respectRetryAfter: true`. |
 | `autoToolChoiceOnlyModels?` | `string[]` | `tool_choice`가 `auto` 또는 `none`만 받는 모델입니다. 강제 선택은 낮은 수준으로 바뀝니다. |
 | `preserveReasoningContentModels?` | `string[]` | chat 기록에서 이전 assistant `reasoning_content`가 필요한 모델입니다. |
+| `requiresReasoningPlaceholderModels?` | `string[]` | `reasoning_content`가 없는 tool_call 연속을 업스트림이 거부하는 모델(DeepSeek thinking 모드). 리플레이 캐시 미스 시 최소 플레이스홀더를 주입합니다. 미설정 시 `preserveReasoningContentModels`를 따륩며 `[]`로 명시적 해제 가능. |
 | `thinkingToggleModels?` | `string[]` | effort 계층 대신 `thinking.enabled`를 쓰는 chat 모델입니다. |
 | `thinkingBudgetModels?` | `string[]` | 정수 `thinking_budget`를 쓰는 chat 모델입니다. effort는 예산 비율로 매핑됩니다. |
 | `noVisionModels?` | `string[]` | vision sidecar로 보내는 텍스트 전용 모델입니다. 일치 판정은 Ollama `:size` 태그도 허용합니다. |
diff --git a/docs-site/src/content/docs/reference/configuration/providers.md b/docs-site/src/content/docs/reference/configuration/providers.md
index aca630a1fd..6e14a57a3c 100644
--- a/docs-site/src/content/docs/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/reference/configuration/providers.md
@@ -100,6 +100,7 @@ differing backup and rewrites known legacy namespaced selected ids to bare ids.
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | API-key providers only (`authMode: "key"`). Opt-in same-target 429 retry: when `retryOn429` is absent the feature is off; object presence enables it unless `enabled: false`. On 429 the proxy waits (upstream `Retry-After` or the fixed interval) and replays the identical request on the same key before any key failover — across the main text-turn recovery loop, the Responses passthrough wire, the image/video bridge, the web-search sidecar, and terminal continuations. Only pre-stream HTTP 429 responses are eligible for replay; custom `runTurn` transports are outside the HTTP retry loop. `attempts` counts same-key replays after the first 429 (total sends = `attempts` + 1) and is one request-wide budget shared by the main recovery loop, the terminal-guard continuation, and bridge retries. Exhausting `attempts` only stops further same-key replays: normal key failover or final-error handling then applies per the available targets — on the key-auth passthrough wire there is no failover, so the exhausted 429 surfaces as-is. Codex itself never retries 429, so this is the only defense for single-key providers. Defaults: `enabled: true`, `attempts: 3`, `intervalMs: 5000`, `maxIntervalMs: 60000` (any single wait is capped at `maxIntervalMs`, itself capped at 600000), `respectRetryAfter: true`. |
 | `autoToolChoiceOnlyModels?` | `string[]` | Models whose `tool_choice` accepts only `auto` or `none`; forced choices are downgraded. |
 | `preserveReasoningContentModels?` | `string[]` | Models requiring prior assistant `reasoning_content` in chat history. |
+| `requiresReasoningPlaceholderModels?` | `string[]` | Models whose upstream rejects a tool_call continuation missing `reasoning_content` (DeepSeek thinking mode); a minimal placeholder is injected when the replay cache misses. Defaults to `preserveReasoningContentModels`; set `[]` to opt out. |
 | `thinkingToggleModels?` | `string[]` | Chat models using `thinking.enabled` rather than an effort ladder. |
 | `thinkingBudgetModels?` | `string[]` | Chat models using integer `thinking_budget`; effort maps to a budget fraction. |
 | `noVisionModels?` | `string[]` | Text-only models sent through the vision sidecar; matching tolerates an Ollama `:size` tag. |
diff --git a/docs-site/src/content/docs/ru/reference/configuration/providers.md b/docs-site/src/content/docs/ru/reference/configuration/providers.md
index abf75d650e..25f39ec00d 100644
--- a/docs-site/src/content/docs/ru/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/ru/reference/configuration/providers.md
@@ -102,6 +102,7 @@ cross-route credential fallback не существует. Строки API GPT-
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | Только для провайдеров с API-ключом (`authMode: "key"`). Опциональный повтор при 429 на том же таргете: если `retryOn429` отсутствует, функция выключена; наличие объекта включает её, если только `enabled: false`. При 429: ожидание (`Retry-After` апстрима или фиксированный интервал) и повтор идентичного запроса на том же ключе до любого фейловера ключей — покрывает основной цикл восстановления текстовых ходов, passthrough-канал Responses, мост изображений/видео, sidecar web-search и терминальные продолжения. Повтор допустим только для HTTP 429, полученных до начала потока; пользовательские транспорты `runTurn` не входят в цикл HTTP-повторов. `attempts` — это число повторов на том же ключе после первого 429 (всего отправок = `attempts` + 1) и единый бюджет на запрос, общий для основного цикла восстановления, терминального продолжения и повторов моста. Исчерпание `attempts` лишь останавливает дальнейшие повторы на том же ключе; далее применяется обычный фейловер ключей или финальная обработка ошибки в зависимости от доступных таргетов — на passthrough-канале с ключевой аутентификацией фейловера нет, поэтому исчерпанный 429 возвращается как есть. Codex сам никогда не повторяет 429, поэтому это единственная защита для провайдеров с одним ключом. По умолчанию: `enabled: true`, `attempts: 3`, `intervalMs: 5000`, `maxIntervalMs: 60000` (любое ожидание ограничено `maxIntervalMs`, который сам ограничен 600000), `respectRetryAfter: true`. |
 | `autoToolChoiceOnlyModels?` | `string[]` | Модели, у которых `tool_choice` принимает только `auto` или `none`; forced choice понижается. |
 | `preserveReasoningContentModels?` | `string[]` | Модели, которым нужен предыдущий assistant `reasoning_content` в chat history. |
+| `requiresReasoningPlaceholderModels?` | `string[]` | Модели, чей upstream отклоняет tool_call-продолжение без `reasoning_content` (DeepSeek thinking mode); при промахе replay-кэша подставляется минимальный placeholder. По умолчанию наследует `preserveReasoningContentModels`; `[]` отключает явно. |
 | `thinkingToggleModels?` | `string[]` | Chat-модели, использующие `thinking.enabled` вместо effort-ladder. |
 | `thinkingBudgetModels?` | `string[]` | Chat-модели, использующие целочисленный `thinking_budget`; effort отображается в долю бюджета. |
 | `noVisionModels?` | `string[]` | Text-only-модели, идущие через vision sidecar; при сопоставлении tolerируется тег Ollama вида `:size`. |
diff --git a/docs-site/src/content/docs/zh-cn/reference/configuration/providers.md b/docs-site/src/content/docs/zh-cn/reference/configuration/providers.md
index 7e664ce58c..82f8b54d76 100644
--- a/docs-site/src/content/docs/zh-cn/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/zh-cn/reference/configuration/providers.md
@@ -91,6 +91,7 @@ pool account id（不能是内部 `__main__`），或用 `"@main"` 表示 Codex
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | 仅限 API-key 提供商（`authMode: "key"`）。可选的同目标 429 重试：未配置 `retryOn429` 时功能关闭；对象存在即启用，除非 `enabled: false`。收到 429 时等待（上游 `Retry-After` 或固定间隔）后在相同 key 上重放完全相同请求，再进入任何 key 故障转移——覆盖主文本恢复循环、Responses passthrough、图像/视频桥、web-search 侧车与终结续接。重放仅适用于流开始前的 HTTP 429 响应；自定义 `runTurn` 传输不在 HTTP 重试循环范围内。`attempts` 是首个 429 之后的同 key 重放次数（总发送次数 = `attempts` + 1），是主恢复循环、终结守卫续接与桥接重试共享的按请求统一预算；`attempts` 耗尽只会停止进一步的同 key 重放：随后按可用目标进行正常的 key 故障转移或最终错误处理——key 认证的 passthrough 线路上没有故障转移，因此耗尽的 429 会原样透出。Codex 自身从不重试 429，因此这是单 key 提供商唯一的防线。默认值：`enabled: true`、`attempts: 3`、`intervalMs: 5000`、`maxIntervalMs: 60000`（单次等待以 `maxIntervalMs` 为上限，其本身上限 600000）、`respectRetryAfter: true`。 |
 | `autoToolChoiceOnlyModels?` | `string[]` | `tool_choice` 只接受 `auto` 或 `none` 的模型；强制选择会被降级。 |
 | `preserveReasoningContentModels?` | `string[]` | 需要在聊天历史中保留先前 assistant `reasoning_content` 的模型。 |
+| `requiresReasoningPlaceholderModels?` | `string[]` | 上游会拒绝缺少 `reasoning_content` 的 tool_call 续接消息的模型（DeepSeek thinking 模式）；重放缓存 miss 时注入最小占位符。缺省沿用 `preserveReasoningContentModels`；设为 `[]` 可显式关闭。 |
 | `thinkingToggleModels?` | `string[]` | 使用 `thinking.enabled` 而不是 effort 阶梯的 chat 模型。 |
 | `thinkingBudgetModels?` | `string[]` | 使用整数 `thinking_budget` 的 chat 模型；effort 会映射为预算比例。 |
 | `noVisionModels?` | `string[]` | 经由视觉 sidecar 发送的纯文本模型；匹配时会容忍 Ollama 的 `:size` 标记。 |
diff --git a/src/adapters/openai-chat.ts b/src/adapters/openai-chat.ts
index 400e1e2d7e..9042d8ee25 100644
--- a/src/adapters/openai-chat.ts
+++ b/src/adapters/openai-chat.ts
@@ -347,14 +347,17 @@ function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProviderCon
           // recorded under every call id — join unique texts only.
           if (cached.length > 0) {
             reasoningContent = [...new Set(cached)].join("\n");
-          } else {
+          } else if (modelInList(provider.requiresReasoningPlaceholderModels ?? provider.preserveReasoningContentModels, parsed.modelId)) {
             // Fallback (extends #950, closes #1193): the replay cache is
             // bounded (64 entries / 256 KiB / 1 h TTL) and always misses on
             // long sessions, and some tool rounds carry no recorded reasoning
             // at all. DeepSeek thinking mode rejects ANY tool_call assistant
             // message missing reasoning_content with HTTP 400, so inject a
             // minimal placeholder rather than emit a bare continuation the
-            // upstream will reject.
+            // upstream will reject. Scoped to requiresReasoningPlaceholderModels
+            // (defaulting to the preserve list): preserve-listed providers with
+            // toggleable thinking (MiniMax low effort) opt out with `[]` so
+            // non-thinking histories are never given a fabricated placeholder.
             reasoningContent = " ";
           }
         }
@@ -426,9 +429,11 @@ function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProviderCon
           // tool_call continuation on a thinking-mode provider — inject a
           // placeholder when the replay cache missed (the bounded cache can
           // always miss on long sessions), or DeepSeek thinking mode 400s.
+          // `||` (not `??`): the cache never stores empty strings, but treat a
+          // falsy hit as a miss so the placeholder still fires.
           const orphanReasoning =
             cachedReasoning
-            ?? (modelInList(provider.preserveReasoningContentModels, parsed.modelId) ? " " : undefined);
+            || (modelInList(provider.requiresReasoningPlaceholderModels ?? provider.preserveReasoningContentModels, parsed.modelId) ? " " : undefined);
           out.push({
             role: "assistant",
             content: emptyAssistantContent(provider),
diff --git a/src/oauth/index.ts b/src/oauth/index.ts
index 3487023929..b0035a1c02 100644
--- a/src/oauth/index.ts
+++ b/src/oauth/index.ts
@@ -726,6 +726,7 @@ const OAUTH_RECONCILE_FIELDS: (keyof OcxProviderConfig)[] = [
   "noPenaltyModels",
   "autoToolChoiceOnlyModels",
   "preserveReasoningContentModels",
+  "requiresReasoningPlaceholderModels",
 ];
 
 const GOOGLE_ANTIGRAVITY_PROVIDER = "google-antigravity";
diff --git a/src/oauth/login-cli.ts b/src/oauth/login-cli.ts
index dbcf755619..0afecb4281 100644
--- a/src/oauth/login-cli.ts
+++ b/src/oauth/login-cli.ts
@@ -102,6 +102,7 @@ export function providerConfigFromKeyLoginProvider(def: KeyLoginProvider, key: s
     ...(def.noPenaltyModels ? { noPenaltyModels: [...def.noPenaltyModels] } : {}),
     ...(def.autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels: [...def.autoToolChoiceOnlyModels] } : {}),
     ...(def.preserveReasoningContentModels ? { preserveReasoningContentModels: [...def.preserveReasoningContentModels] } : {}),
+    ...(def.requiresReasoningPlaceholderModels ? { requiresReasoningPlaceholderModels: [...def.requiresReasoningPlaceholderModels] } : {}),
     ...(def.escapeBuiltinToolNames !== undefined ? { escapeBuiltinToolNames: def.escapeBuiltinToolNames } : {}),
   };
 }
diff --git a/src/providers/derive.ts b/src/providers/derive.ts
index c947c0b4f8..fe8581b0ea 100644
--- a/src/providers/derive.ts
+++ b/src/providers/derive.ts
@@ -36,6 +36,7 @@ export interface DerivedKeyLoginProvider {
   noPenaltyModels?: string[];
   autoToolChoiceOnlyModels?: string[];
   preserveReasoningContentModels?: string[];
+  requiresReasoningPlaceholderModels?: string[];
   reasoningSplitModels?: string[];
   thinkingToggleModels?: string[];
   thinkingBudgetModels?: string[];
@@ -154,6 +155,7 @@ export function providerConfigSeed(entry: ProviderRegistryEntry): OcxProviderCon
     ...(entry.statelessResponses !== undefined ? { statelessResponses: entry.statelessResponses } : {}),
     ...(entry.autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels: [...entry.autoToolChoiceOnlyModels] } : {}),
     ...(entry.preserveReasoningContentModels ? { preserveReasoningContentModels: [...entry.preserveReasoningContentModels] } : {}),
+    ...(entry.requiresReasoningPlaceholderModels ? { requiresReasoningPlaceholderModels: [...entry.requiresReasoningPlaceholderModels] } : {}),
     ...(entry.reasoningSplitModels ? { reasoningSplitModels: [...entry.reasoningSplitModels] } : {}),
     ...(entry.thinkingToggleModels ? { thinkingToggleModels: [...entry.thinkingToggleModels] } : {}),
     ...(entry.thinkingBudgetModels ? { thinkingBudgetModels: [...entry.thinkingBudgetModels] } : {}),
@@ -199,6 +201,7 @@ export function deriveKeyLoginMap(): Record<string, DerivedKeyLoginProvider> {
       ...(entry.noPenaltyModels ? { noPenaltyModels: [...entry.noPenaltyModels] } : {}),
       ...(entry.autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels: [...entry.autoToolChoiceOnlyModels] } : {}),
       ...(entry.preserveReasoningContentModels ? { preserveReasoningContentModels: [...entry.preserveReasoningContentModels] } : {}),
+      ...(entry.requiresReasoningPlaceholderModels ? { requiresReasoningPlaceholderModels: [...entry.requiresReasoningPlaceholderModels] } : {}),
       ...(entry.reasoningSplitModels ? { reasoningSplitModels: [...entry.reasoningSplitModels] } : {}),
       ...(entry.thinkingToggleModels ? { thinkingToggleModels: [...entry.thinkingToggleModels] } : {}),
       ...(entry.thinkingBudgetModels ? { thinkingBudgetModels: [...entry.thinkingBudgetModels] } : {}),
@@ -345,6 +348,7 @@ export function enrichProviderFromRegistry(name: string, prov: OcxProviderConfig
   }
   if (!prov.autoToolChoiceOnlyModels && seed.autoToolChoiceOnlyModels) prov.autoToolChoiceOnlyModels = [...seed.autoToolChoiceOnlyModels];
   if (!prov.preserveReasoningContentModels && seed.preserveReasoningContentModels) prov.preserveReasoningContentModels = [...seed.preserveReasoningContentModels];
+  if (!prov.requiresReasoningPlaceholderModels && seed.requiresReasoningPlaceholderModels) prov.requiresReasoningPlaceholderModels = [...seed.requiresReasoningPlaceholderModels];
   if (!prov.reasoningSplitModels && seed.reasoningSplitModels) prov.reasoningSplitModels = [...seed.reasoningSplitModels];
   if (!prov.thinkingToggleModels && seed.thinkingToggleModels) prov.thinkingToggleModels = [...seed.thinkingToggleModels];
   if (!prov.thinkingBudgetModels && seed.thinkingBudgetModels) prov.thinkingBudgetModels = [...seed.thinkingBudgetModels];
diff --git a/src/providers/registry.ts b/src/providers/registry.ts
index 375c4ada76..69b6c3a676 100644
--- a/src/providers/registry.ts
+++ b/src/providers/registry.ts
@@ -221,6 +221,7 @@ export interface ProviderRegistryEntry {
   promptCacheKey?: boolean;
   autoToolChoiceOnlyModels?: string[];
   preserveReasoningContentModels?: string[];
+  requiresReasoningPlaceholderModels?: string[];
   reasoningSplitModels?: string[];
   thinkingToggleModels?: string[];
   thinkingBudgetModels?: string[];
@@ -243,7 +244,7 @@ export type ProviderConfigSeed = Pick<
   | "modelMaxInputTokens" | "defaultMaxOutputTokens" | "modelMaxOutputTokens"
   | "reasoningEfforts" | "modelReasoningEfforts" | "modelDefaultReasoningEfforts" | "reasoningEffortMap" | "modelReasoningEffortMap" | "reasoningWireFormat"
   | "noVisionModels" | "noReasoningModels" | "noTemperatureModels" | "noTopPModels" | "noPenaltyModels"
-  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "reasoningSplitModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "escapeBuiltinToolNames"
+  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "requiresReasoningPlaceholderModels" | "reasoningSplitModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "escapeBuiltinToolNames"
   | "googleMode" | "project" | "location" | "headers"
 >;
 
@@ -1984,6 +1985,10 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelDefaultReasoningEfforts: { "MiniMax-M3": "medium" },
     modelReasoningEffortMap: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORT_MAP },
     preserveReasoningContentModels: MINIMAX_MODELS,
+    // MiniMax-M3 low effort maps to thinking disabled, so a legitimate tool
+    // round can carry no reasoning at all; only replay real recorded text,
+    // never a fabricated placeholder (chatgpt-codex-connector P2 on #1205).
+    requiresReasoningPlaceholderModels: [],
     reasoningSplitModels: MINIMAX_MODELS,
     thinkingToggleModels: ["MiniMax-M3"],
     jawcodeBundle: "minimax", metadataModelIdNormalize: "case-insensitive", note: "Subscription Key or API Key",
@@ -1996,6 +2001,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelDefaultReasoningEfforts: { "MiniMax-M3": "medium" },
     modelReasoningEffortMap: { "MiniMax-M3": MINIMAX_M3_REASONING_EFFORT_MAP },
     preserveReasoningContentModels: MINIMAX_MODELS,
+    requiresReasoningPlaceholderModels: [],
     reasoningSplitModels: MINIMAX_MODELS,
     thinkingToggleModels: ["MiniMax-M3"],
     jawcodeBundle: "minimax", metadataModelIdNormalize: "case-insensitive", note: "中国区 Subscription Key",
diff --git a/src/router.ts b/src/router.ts
index 599ec9bb94..c7fc68a94b 100644
--- a/src/router.ts
+++ b/src/router.ts
@@ -286,6 +286,7 @@ function routedProviderConfig(providerName: string, provider: OcxProviderConfig)
   const noPenaltyModels = mergeStringArray(registryEntry.noPenaltyModels, provider.noPenaltyModels);
   const autoToolChoiceOnlyModels = mergeStringArray(registryEntry.autoToolChoiceOnlyModels, provider.autoToolChoiceOnlyModels);
   const preserveReasoningContentModels = mergeStringArray(registryEntry.preserveReasoningContentModels, provider.preserveReasoningContentModels);
+  const requiresReasoningPlaceholderModels = mergeStringArray(registryEntry.requiresReasoningPlaceholderModels, provider.requiresReasoningPlaceholderModels);
   const reasoningSplitModels = mergeStringArray(registryEntry.reasoningSplitModels, provider.reasoningSplitModels);
   const thinkingToggleModels = mergeStringArray(registryEntry.thinkingToggleModels, provider.thinkingToggleModels);
   const thinkingBudgetModels = mergeStringArray(registryEntry.thinkingBudgetModels, provider.thinkingBudgetModels);
@@ -370,6 +371,7 @@ function routedProviderConfig(providerName: string, provider: OcxProviderConfig)
     ...(noPenaltyModels ? { noPenaltyModels } : {}),
     ...(autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels } : {}),
     ...(preserveReasoningContentModels ? { preserveReasoningContentModels } : {}),
+    ...(requiresReasoningPlaceholderModels ? { requiresReasoningPlaceholderModels } : {}),
     ...(reasoningSplitModels ? { reasoningSplitModels } : {}),
     ...(thinkingToggleModels ? { thinkingToggleModels } : {}),
     ...(thinkingBudgetModels ? { thinkingBudgetModels } : {}),
diff --git a/src/server/auth-cors.ts b/src/server/auth-cors.ts
index 536526f3e3..a59bfba5f1 100644
--- a/src/server/auth-cors.ts
+++ b/src/server/auth-cors.ts
@@ -579,6 +579,7 @@ export function safeConfigDTO(config: OcxConfig): unknown {
       "noPenaltyModels",
       "autoToolChoiceOnlyModels",
       "preserveReasoningContentModels",
+      "requiresReasoningPlaceholderModels",
       "escapeBuiltinToolNames",
     ] as const) {
       copyIfDefined(dto, provider, key);
diff --git a/src/types.ts b/src/types.ts
index 481c15a4b0..1390939e01 100644
--- a/src/types.ts
+++ b/src/types.ts
@@ -1323,6 +1323,14 @@ export interface OcxProviderConfig {
   autoToolChoiceOnlyModels?: string[];
   /** Model ids that expect prior assistant `reasoning_content` to be preserved in chat history. */
   preserveReasoningContentModels?: string[];
+  /**
+   * Model ids whose upstream hard-rejects a tool_call continuation missing
+   * `reasoning_content` (DeepSeek thinking mode: HTTP 400). When the replay
+   * cache misses, the adapter injects a minimal placeholder for these models.
+   * Defaults to `preserveReasoningContentModels` when unset; set `[]` to opt
+   * out explicitly (e.g. MiniMax, where low effort disables thinking).
+   */
+  requiresReasoningPlaceholderModels?: string[];
   /**
    * Opt-in same-target 429 retry policy. Codex itself never retries 429 (it retries 5xx only,
    * openai/codex#30471), and single-key pools have no failover, so the proxy waits and replays
diff --git a/tests/deepseek-reasoning-replay-gaps.test.ts b/tests/deepseek-reasoning-replay-gaps.test.ts
index 16d107a04a..76859116b5 100644
--- a/tests/deepseek-reasoning-replay-gaps.test.ts
+++ b/tests/deepseek-reasoning-replay-gaps.test.ts
@@ -187,6 +187,41 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     expect(assistant!["reasoning_content"]).toBeUndefined();
   });
 
+  test("P2 guard: preserve-listed providers with toggleable thinking opt out of the placeholder (MiniMax)", () => {
+    // MiniMax-M3 low effort maps to thinking disabled, so a legitimate tool
+    // round can carry no reasoning; the registry seeds
+    // requiresReasoningPlaceholderModels: [] for minimax so a cache miss never
+    // fabricates one (chatgpt-codex-connector P2 on #1205). Real recorded
+    // reasoning still replays via preserveReasoningContentModels.
+    const minimaxWire = (input: unknown[]) => {
+      const parsed = parseRequest({ model: "minimax/MiniMax-M3", input, stream: true });
+      const config: OcxConfig = {
+        port: 10100,
+        defaultProvider: "minimax",
+        providers: {
+          minimax: {
+            adapter: "openai-chat",
+            baseUrl: "https://api.minimax.io/v1",
+            apiKey: "key",
+          },
+        },
+      };
+      const route = routeModel(config, parsed.modelId);
+      parsed.modelId = route.modelId;
+      const req = createOpenAIChatAdapter(route.provider).buildRequest(parsed as OcxParsedRequest);
+      return JSON.parse(req.body as string) as { messages: Array<Record<string, unknown>> };
+    };
+    // Cache miss on the orphan-repair path: no fabricated placeholder.
+    const miss = toolCallAssistant(minimaxWire([userMessage(), functionCallOutputItem()]).messages);
+    expect(miss).toBeDefined();
+    expect(miss!["reasoning_content"]).toBeUndefined();
+    // Cache hit on the same path: the recorded reasoning still replays.
+    rememberReasoningForCall("call_1", REASONING);
+    const hit = toolCallAssistant(minimaxWire([userMessage(), functionCallOutputItem()]).messages);
+    expect(hit).toBeDefined();
+    expect(hit!["reasoning_content"]).toBe(REASONING);
+  });
+
   test("documented non-bug: opaque encrypted-only reasoning degrades to the placeholder, not invented plaintext", () => {
     // Native (non-ocxr1) encrypted reasoning has no readable text; the parser
     // deliberately degrades instead of inventing replayable plaintext. On a

From 61283a42f5a8201a5eab92e94d2f2b0723cc6aa4 Mon Sep 17 00:00:00 2001
From: Yuxin Qiao <104957188+Yuxin-Qiao@users.noreply.github.com>
Date: Sat, 8 Aug 2026 12:27:03 +0800
Subject: [PATCH 3/3] fix(openai-chat): gate orphan placeholder on preserve
 list, keep opt-outs durable

Address the remaining review findings on #1205:

- chatgpt-codex-connector P2: the orphan-repair fallback checked only
  requiresReasoningPlaceholderModels, so a requires-only custom entry could
  fabricate reasoning_content on a path the main assistant history would
  never emit it on. Gate the orphan placeholder on the preserve list too.
- chatgpt-codex-connector P2: Zhipu BigModel GLM ids are thinking-toggle
  models (low maps to disabled) AND preserve-listed, so the placeholder
  default could fabricate reasoning for non-thinking histories. Seed
  requiresReasoningPlaceholderModels: [] for zhipu-bigmodel, matching the
  MiniMax opt-out.
- chatgpt-codex-connector P2: OAuth reconcile deleted an explicit
  requiresReasoningPlaceholderModels: [] opt-out on every startup because no
  OAuth preset seeds the field. Keep the field out of OAUTH_RECONCILE_FIELDS;
  registry seeds still reach existing rows via enrichProviderFromRegistry.
- CodeRabbit minor: fix Korean spelling in the providers table.

Refs #1193
---
 .../ko/reference/configuration/providers.md   |  2 +-
 src/adapters/openai-chat.ts                   |  8 ++++-
 src/oauth/index.ts                            |  6 +++-
 src/providers/registry.ts                     |  4 +++
 tests/deepseek-reasoning-replay-gaps.test.ts  | 30 +++++++++++++++++++
 tests/oauth-provider-reconcile.test.ts        | 19 ++++++++++++
 6 files changed, 66 insertions(+), 3 deletions(-)

diff --git a/docs-site/src/content/docs/ko/reference/configuration/providers.md b/docs-site/src/content/docs/ko/reference/configuration/providers.md
index f3faf2c152..708f5fbe9c 100644
--- a/docs-site/src/content/docs/ko/reference/configuration/providers.md
+++ b/docs-site/src/content/docs/ko/reference/configuration/providers.md
@@ -92,7 +92,7 @@ target도 selector로 재사용할 수 없습니다. raw account id와 email은
 | `retryOn429?` | `{ enabled?: boolean; attempts?: number; intervalMs?: number; maxIntervalMs?: number; respectRetryAfter?: boolean }` | API-key 프로바이더 전용(`authMode: "key"`). 동일 대상 429 재시도: `retryOn429`가 없으면 기능이 꺼져 있고, 객체가 있으면 `enabled: false`가 아닌 한 활성화됩니다. 429 시 대기(업스트림 `Retry-After` 또는 고정 간격) 후 키 장애 조치 전에 동일 키로 동일 요청을 재전송합니다 — 일반 텍스트 턴 복구 루프, Responses passthrough, 이미지/비디오 브리지, web-search 사이드카, 터미널 연속 요청을 모두 포함합니다. 재전송 대상은 프리스트림 HTTP 429 응답뿐이며, 커스텀 `runTurn` 전송은 HTTP 재시도 루프에서 제외됩니다. `attempts`는 첫 429 이후의 동일 키 재전송 횟수(총 전송 = `attempts` + 1)이며, 메인 복구 루프·터미널 가드 연속 요청·브리지 재시도가 공유하는 요청 단위 예산입니다. `attempts`를 모두 소진해도 동일 키 재전송만 중단되며, 이후에는 일반 키 장애 조치 또는 최종 오류 처리가 사용 가능한 대상에 따라 진행됩니다 — 키 인증 passthrough 와이어에는 장애 조치가 없으므로 소진된 429가 그대로 반환됩니다. Codex 자체는 429를 재시도하지 않으므로 단일 키 프로바이더의 유일한 방어선입니다. 기본값: `enabled: true`, `attempts: 3`, `intervalMs: 5000`, `maxIntervalMs: 60000`(단일 대기는 `maxIntervalMs`로 상한, 그 자체는 600000으로 상한), `respectRetryAfter: true`. |
 | `autoToolChoiceOnlyModels?` | `string[]` | `tool_choice`가 `auto` 또는 `none`만 받는 모델입니다. 강제 선택은 낮은 수준으로 바뀝니다. |
 | `preserveReasoningContentModels?` | `string[]` | chat 기록에서 이전 assistant `reasoning_content`가 필요한 모델입니다. |
-| `requiresReasoningPlaceholderModels?` | `string[]` | `reasoning_content`가 없는 tool_call 연속을 업스트림이 거부하는 모델(DeepSeek thinking 모드). 리플레이 캐시 미스 시 최소 플레이스홀더를 주입합니다. 미설정 시 `preserveReasoningContentModels`를 따륩며 `[]`로 명시적 해제 가능. |
+| `requiresReasoningPlaceholderModels?` | `string[]` | `reasoning_content`가 없는 tool_call 연속을 업스트림이 거부하는 모델(DeepSeek thinking 모드). 리플레이 캐시 미스 시 최소 플레이스홀더를 주입합니다. 미설정 시 `preserveReasoningContentModels`를 따르며 `[]`로 명시적 해제 가능. |
 | `thinkingToggleModels?` | `string[]` | effort 계층 대신 `thinking.enabled`를 쓰는 chat 모델입니다. |
 | `thinkingBudgetModels?` | `string[]` | 정수 `thinking_budget`를 쓰는 chat 모델입니다. effort는 예산 비율로 매핑됩니다. |
 | `noVisionModels?` | `string[]` | vision sidecar로 보내는 텍스트 전용 모델입니다. 일치 판정은 Ollama `:size` 태그도 허용합니다. |
diff --git a/src/adapters/openai-chat.ts b/src/adapters/openai-chat.ts
index 9042d8ee25..448d7297db 100644
--- a/src/adapters/openai-chat.ts
+++ b/src/adapters/openai-chat.ts
@@ -429,11 +429,17 @@ function messagesToChatFormat(parsed: OcxParsedRequest, provider: OcxProviderCon
           // tool_call continuation on a thinking-mode provider — inject a
           // placeholder when the replay cache missed (the bounded cache can
           // always miss on long sessions), or DeepSeek thinking mode 400s.
+          // Gate on the preserve list too: reasoning_content is only ever
+          // serialized for preserve-listed models, so a requires-only custom
+          // entry must not fabricate it on this path (P2 on #1205).
           // `||` (not `??`): the cache never stores empty strings, but treat a
           // falsy hit as a miss so the placeholder still fires.
           const orphanReasoning =
             cachedReasoning
-            || (modelInList(provider.requiresReasoningPlaceholderModels ?? provider.preserveReasoningContentModels, parsed.modelId) ? " " : undefined);
+            || (modelInList(provider.preserveReasoningContentModels, parsed.modelId)
+              && modelInList(provider.requiresReasoningPlaceholderModels ?? provider.preserveReasoningContentModels, parsed.modelId)
+              ? " "
+              : undefined);
           out.push({
             role: "assistant",
             content: emptyAssistantContent(provider),
diff --git a/src/oauth/index.ts b/src/oauth/index.ts
index b0035a1c02..0bb76c47a9 100644
--- a/src/oauth/index.ts
+++ b/src/oauth/index.ts
@@ -726,8 +726,12 @@ const OAUTH_RECONCILE_FIELDS: (keyof OcxProviderConfig)[] = [
   "noPenaltyModels",
   "autoToolChoiceOnlyModels",
   "preserveReasoningContentModels",
-  "requiresReasoningPlaceholderModels",
 ];
+// `requiresReasoningPlaceholderModels` is deliberately NOT reconciled here: no
+// OAuth preset seeds it, so the delete-when-preset-undefined branch would wipe
+// an explicit user opt-out (`[]`) on every startup. Registry seeds still reach
+// existing rows through enrichProviderFromRegistry, which is fill-only and
+// preserves explicit saved values.
 
 const GOOGLE_ANTIGRAVITY_PROVIDER = "google-antigravity";
 const GOOGLE_ANTIGRAVITY_STATIC_CATALOG_VERSION = 1 as const;
diff --git a/src/providers/registry.ts b/src/providers/registry.ts
index 69b6c3a676..1957fdf715 100644
--- a/src/providers/registry.ts
+++ b/src/providers/registry.ts
@@ -1711,6 +1711,10 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       ZHIPU_BIGMODEL_THINKING_TOGGLE_MODELS.map(id => [id, true]),
     ),
     preserveReasoningContentModels: ZHIPU_BIGMODEL_THINKING_TOGGLE_MODELS,
+    // GLM thinking is a binary toggle (low maps to disabled), so a legitimate
+    // tool round can carry no reasoning at all; never fabricate a placeholder
+    // for it, only replay real recorded text (P2 on #1205).
+    requiresReasoningPlaceholderModels: [],
     // No liveModels: GET /api/paas/v4/models has not been observed to answer on this host, and a
     // false live claim yields an empty picker at runtime. Flip it on once someone verifies it.
     note: "Domestic BigModel pay-as-you-go endpoint (open.bigmodel.cn)",
diff --git a/tests/deepseek-reasoning-replay-gaps.test.ts b/tests/deepseek-reasoning-replay-gaps.test.ts
index 76859116b5..6b72f5aceb 100644
--- a/tests/deepseek-reasoning-replay-gaps.test.ts
+++ b/tests/deepseek-reasoning-replay-gaps.test.ts
@@ -222,6 +222,36 @@ describe("issue #950 — tool-call reasoning replay invariant (openai-chat wire)
     expect(hit!["reasoning_content"]).toBe(REASONING);
   });
 
+  test("P2 guard: a requires-only custom model never gets a placeholder on the orphan path", () => {
+    // requiresReasoningPlaceholderModels narrows which preserve-listed models
+    // get a fabricated placeholder. A custom entry listing a model ONLY in the
+    // requires list (not in preserveReasoningContentModels) must behave like
+    // the main-assistant path, which never serializes reasoning_content for
+    // non-preserve models: the synthesized orphan tool_call stays bare
+    // (chatgpt-codex-connector P2 on #1205).
+    const parsed = parseRequest({ model: "custom-chat/plain-model", input: [userMessage(), functionCallOutputItem()], stream: true });
+    const config: OcxConfig = {
+      port: 10100,
+      defaultProvider: "custom-chat",
+      providers: {
+        "custom-chat": {
+          adapter: "openai-chat",
+          baseUrl: "https://example.invalid/v1",
+          apiKey: "key",
+          models: ["plain-model"],
+          requiresReasoningPlaceholderModels: ["plain-model"],
+        },
+      },
+    };
+    const route = routeModel(config, parsed.modelId);
+    parsed.modelId = route.modelId;
+    const req = createOpenAIChatAdapter(route.provider).buildRequest(parsed as OcxParsedRequest);
+    const { messages } = JSON.parse(req.body as string) as { messages: Array<Record<string, unknown>> };
+    const assistant = toolCallAssistant(messages);
+    expect(assistant).toBeDefined();
+    expect(assistant!["reasoning_content"]).toBeUndefined();
+  });
+
   test("documented non-bug: opaque encrypted-only reasoning degrades to the placeholder, not invented plaintext", () => {
     // Native (non-ocxr1) encrypted reasoning has no readable text; the parser
     // deliberately degrades instead of inventing replayable plaintext. On a
diff --git a/tests/oauth-provider-reconcile.test.ts b/tests/oauth-provider-reconcile.test.ts
index 1ee0830656..91dd7b8498 100644
--- a/tests/oauth-provider-reconcile.test.ts
+++ b/tests/oauth-provider-reconcile.test.ts
@@ -158,4 +158,23 @@ describe("OAuth provider reconciliation", () => {
     upsertOAuthProvider(config, "google-antigravity");
     expect(config.providers["google-antigravity"].liveModels).toBe(true);
   });
+
+  test("preserves an explicit requiresReasoningPlaceholderModels opt-out on OAuth providers", () => {
+    // No OAuth preset seeds the new field, so reconcile must never delete an
+    // explicit `[]` opt-out on startup (chatgpt-codex-connector P2 on #1205).
+    const config = {
+      port: 10100,
+      defaultProvider: "kimi",
+      googleAntigravityStaticCatalogVersion: 1,
+      providers: {
+        kimi: {
+          ...structuredClone(OAUTH_PROVIDERS.kimi.providerConfig),
+          requiresReasoningPlaceholderModels: [],
+        },
+      },
+    } satisfies OcxConfig;
+
+    expect(reconcileOAuthProviders(config)).toBe(false);
+    expect(config.providers.kimi.requiresReasoningPlaceholderModels).toEqual([]);
+  });
 });