fix(logs): classify requested output limits correctly

Author: Ingwannu

README.md

@@ -329,6 +329,10 @@ Every non-empty family has one default. You can drag a route, or use its visible
 a mouse, touch, or keyboard. **Save and apply** writes the profile to Claude Desktop. JSON export
 and import are available for backup or moving the same setup to another machine.
 
+If Claude Desktop's footer picker does not change the model for an already-running 3P
+conversation, use `/model <id>` in that conversation. OpenCodex routes the model id carried by
+each request; **Logs → requestedModel** shows which id Desktop actually sent.
+
 ```bash
 ocx claude desktop [apply]                         # save and apply the current profile
 ocx claude desktop show [--json]                   # inspect routes, families, and defaults

docs-site/src/content/docs/guides/claude-code.md

@@ -151,6 +151,11 @@ capabilities (reasoning-effort ladder, thinking types) in the official ModelInfo
 Desktop's third-party gateway mode can offer its effort selector. Real Anthropic models keep their
 canonical ids. The synthetic 2026 date is an internal slot, not a release date. Legacy hash aliases
 and `claude-ocx-<provider>--<model>` ids from older configs still resolve.
+
+If Claude Desktop's footer picker does not change the model for an already-running 3P
+conversation, use `/model <id>` in that conversation. OpenCodex cannot observe picker state; it
+routes the model id carried by each request. Confirm the result under **Logs → requestedModel**.
+
 Models with an authoritative 1M context window get an extra `…[1m]` picker row: selecting it makes
 Claude Code account a full 1M context for that model (auto-compaction stays on) — the proxy strips
 the marker before routing.

docs-site/src/content/docs/ja/guides/claude-code.md

@@ -77,6 +77,10 @@ Claude Code 2.1.129 以降は `GET /v1/models?limit=1000` でゲートウェイ
 提供します。両系列は継続してデコードできるため、どちらの形式でも `settings.json` に保存したモデルは
 引き続き動作します。
 
+Claude Desktop のフッターピッカーで実行中の 3P 会話のモデルが切り替わらない場合は、その会話で
+`/model <id>` を使用してください。OpenCodex はピッカーの状態を直接参照できず、各リクエストに
+含まれるモデル ID をルーティングします。結果は **Logs → requestedModel** で確認できます。
+
 **エイリアス構文ルール:** provider には `/` や `--` を含められず `native` と同じでもいけません。model には
 `/` を含められません。読みやすい形式で表現できないルートはハッシュエイリアスに置き換えます。モデル
 ID には `--` を含め**られます**(解析時は最初の `--` だけを基準に分割します)。`--` を含む

docs-site/src/content/docs/ko/guides/claude-code.md

@@ -112,6 +112,10 @@ Claude Code 2.1.129 이상은 `GET /v1/models?limit=1000`에서 게이트웨이
 제공해요. 두 계열은 계속 디코딩할 수 있으므로 어느 형식이든 `settings.json`에 저장한 모델이
 계속 작동해요.
 
+Claude Desktop의 하단 선택기로 이미 실행 중인 3P 대화의 모델이 바뀌지 않는다면, 그 대화에서
+`/model <id>`를 사용하세요. OpenCodex는 선택기 상태를 따로 볼 수 없고 각 요청에 실린 모델 ID를
+라우팅해요. 적용 결과는 **Logs → requestedModel**에서 확인할 수 있어요.
+
 **별칭 문법 규칙:** provider에는 `/`나 `--`를 넣을 수 없고 `native`와 같아도 안 돼요. model에는
 `/`를 넣을 수 없어요. 읽기 쉬운 형식으로 표현할 수 없는 라우트는 해시 별칭으로 대체해요. 모델
 ID에는 `--`를 넣을 **수 있어요**(해석할 때 첫 번째 `--`만 기준으로 나눠요). `--`가 포함된

docs-site/src/content/docs/ru/guides/claude-code.md

@@ -82,6 +82,10 @@ user-agent `claude-code/*` получает читаемую CLI-форму, а
 семейства декодируются бессрочно — модель, сохранённая в `settings.json` в любой из форм,
 продолжает работать.
 
+Если нижний селектор Claude Desktop не переключает модель в уже запущенном 3P-диалоге,
+используйте `/model <id>` внутри этого диалога. OpenCodex не видит состояние селектора и
+маршрутизирует id модели из каждого запроса. Результат можно проверить в **Logs → requestedModel**.
+
 **Правила грамматики алиасов:** provider не может содержать `/` или `--` и не может быть равен
 `native`; model не может содержать `/`. Маршруты, которые невозможно выразить читаемой формой,
 откатываются на хешированный алиас. Id моделей МОГУТ содержать `--` (при разрешении деление

docs-site/src/content/docs/zh-cn/guides/claude-code.md

@@ -84,6 +84,10 @@ opencodex 会将已路由模型公开为稳定且可逆的别名：
 user-agent 会获得易读的 CLI 形式，其他客户端会获得 Desktop 哈希形式。两种别名族都会永久
 保持可解码——以任一形式保存在 `settings.json` 中的模型都能继续工作。
 
+如果 Claude Desktop 底部的选择器没有切换已运行 3P 对话的模型，请在该对话中使用
+`/model <id>`。OpenCodex 无法读取选择器状态，只会路由每个请求实际携带的模型 ID；可在
+**Logs → requestedModel** 中确认结果。
+
 **别名语法规则：**provider 不得包含 `/` 或 `--`，也不得等于 `native`；model 不得包含
 `/`。易读形式无法表达的路由会回退到哈希别名。模型 ID **可以**包含 `--`（解析时只按第一个
 `--` 拆分）；包含 `--` 的原生 slug 会回退到哈希形式。

src/server/request-log.ts

@@ -77,6 +77,8 @@ export interface RequestLogContext {
   upstreamError?: string;
   /** HTTP status derived from a terminal `response.failed` SSE payload (429/401/503/etc.). */
   terminalHttpStatus?: number;
+  /** Structured reason from `response.incomplete`; internal-only input to log classification. */
+  terminalIncompleteReason?: string;
   affinity?: "reused" | "new_bind" | "rebound" | "cleared";
   transportPhase?: "pre_headers" | "mid_stream" | "terminal_sse";
   terminalSource?: "upstream" | "synthetic";
@@ -532,7 +534,7 @@ export function inspectResponseLogSsePayload(logCtx: RequestLogContext, payload:
  * run it through redactSecretString so secrets never reach /api/logs. Pure; safe on any text.
  */
 function captureUpstreamError(logCtx: RequestLogContext, text: string | null): void {
-  if (!text || logCtx.upstreamError) return;
+  if (!text) return;
   try {
     const json = JSON.parse(text) as {
       type?: unknown;
@@ -544,6 +546,14 @@ function captureUpstreamError(logCtx: RequestLogContext, text: string | null): v
       };
     };
     captureTerminalHttpStatus(logCtx, json);
+    const reason = json?.response?.incomplete_details?.reason;
+    if (json.type === "response.incomplete"
+      && logCtx.terminalIncompleteReason === undefined
+      && typeof reason === "string"
+      && reason.trim()) {
+      logCtx.terminalIncompleteReason = reason.trim();
+    }
+    if (logCtx.upstreamError) return;
     const message = json?.error?.message
       ?? json?.last_error?.message
       ?? json?.response?.error?.message;
@@ -555,11 +565,11 @@ function captureUpstreamError(logCtx: RequestLogContext, text: string | null): v
     // the bridge on a stall-timeout or adapter EOF (response.incomplete). Maps the raw reason to a
     // reader-facing label so a generic 502 in /api/logs explains WHY the turn ended, not just the
     // mapped HTTP code.
-    const reason = json?.response?.incomplete_details?.reason;
     if (typeof reason === "string" && reason.trim()) {
       logCtx.upstreamError = redactSecretString(incompleteReasonLabel(reason.trim())).slice(0, 500);
     }
   } catch {
+    if (logCtx.upstreamError) return;
     const trimmed = text.trim();
     if (trimmed) {
       logCtx.upstreamError = redactSecretString(trimmed).slice(0, 500);
@@ -570,6 +580,8 @@ function captureUpstreamError(logCtx: RequestLogContext, text: string | null): v
 /** Map a raw `incomplete_details.reason` (emitted by the bridge) to a reader-facing label. */
 function incompleteReasonLabel(reason: string): string {
   switch (reason) {
+    case "max_output_tokens":
+      return `Output reached the requested token limit (${reason})`;
     case "upstream_stall_timeout":
       return `Upstream stalled: no data for the stall-timeout window (${reason})`;
     case "adapter_eof":
@@ -614,6 +626,21 @@ export function httpStatusForRequestLogTerminal(
   status: ResponsesTerminalStatus,
   logCtx?: RequestLogContext,
 ): number {
+  /**
+   * [Decision Log]
+   * - 목적과 의도: Keep request logs aligned with the successful HTTP/SSE contract.
+   * - 기존 구현 및 제약 조건: All incomplete terminals were recorded as 502 even when the
+   *   client-requested output limit was reached normally.
+   * - 검토한 주요 대안: Treat every incomplete as success, or infer the reason from display text.
+   * - 선택한 방식: Only structured max_output_tokens incompletes map to 200.
+   * - 다른 대안 대신 이 방식을 선택한 이유: Stall, EOF, and unknown incompletes must remain
+   *   visible failures, and display text is not a stable classification contract.
+   * - 장점, 단점 및 영향: Logs stop reporting false upstream errors while retaining the
+   *   incomplete terminal detail; native callers without a structured reason keep old behavior.
+   */
+  if (status === "incomplete" && logCtx?.terminalIncompleteReason === "max_output_tokens") {
+    return 200;
+  }
   if (status === "failed" && logCtx?.terminalHttpStatus !== undefined) {
     return logCtx.terminalHttpStatus;
   }

tests/request-log.test.ts

@@ -986,6 +986,42 @@ describe("request log metadata", () => {
     expect(entries[0].upstreamError).toContain("Upstream stalled");
   });
 
+  test("deferred SSE logging treats the requested output limit as a successful terminal", async () => {
+    const entries: RequestLogEntry[] = [];
+    const incompletePayload = JSON.stringify({
+      type: "response.incomplete",
+      response: {
+        status: "incomplete",
+        incomplete_details: { reason: "max_output_tokens" },
+        usage: { input_tokens: 9, output_tokens: 64 },
+      },
+    });
+    const body = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode(`data: ${incompletePayload}\n\n`));
+        controller.close();
+      },
+    });
+    const response = responseWithDeferredRequestLog(
+      new Response(body, { status: 200, headers: { "content-type": "text/event-stream" } }),
+      "ocx-test-requested-output-limit",
+      Date.now(),
+      { model: "anthropic/claude-sonnet-5", provider: "anthropic" },
+      entry => entries.push(entry),
+    );
+
+    await response.text();
+    expect(entries).toHaveLength(1);
+    expect(entries[0]).toMatchObject({
+      terminalStatus: "incomplete",
+      status: 200,
+      usageStatus: "reported",
+      usage: { inputTokens: 9, outputTokens: 64 },
+      upstreamError: "Output reached the requested token limit (max_output_tokens)",
+    });
+    expect(entries[0]).not.toHaveProperty("errorCode");
+  });
+
   test("deferred SSE logging surfaces adapter_eof reason as upstreamError", async () => {
     const entries: RequestLogEntry[] = [];
     const incompletePayload = JSON.stringify({