Files
Genarrative/docs/technical
suzmii e15e3b455f
Project CI / Repository checks (push) Successful in 3m12s
Project CI / Frontend tests (push) Successful in 3m21s
Project CI / Backend tests (push) Successful in 7m42s
Project CI / Native shell tests (push) Successful in 20m14s
Project CI / Frontend tests (pull_request) Successful in 50m43s
Project CI / Repository checks (pull_request) Successful in 20m21s
Project CI / Backend tests (pull_request) Successful in 6m18s
Project CI / Native shell tests (pull_request) Successful in 18m19s
修复AGC大上下文请求体限制 (#295)
## 背景与问题

AGC(AI 游戏创作智能体 App)Direct Codex 模式在携带图片工具结果等大上下文调用 LLM 时,`/api/llm/responses` 与 `/api/llm/chat/completions` 会命中 Axum 默认 **2 MiB** 请求体上限,直接被 413 拒绝,大上下文任务无法执行;同时 Codex app-server 的 failed turn 未把上游 413 映射为稳定错误分类,前端只能显示笼统的“其他错误”,无法引导用户处理。

## 改动内容

**1. api-server(server-rs)**
- 新增常量 `LLM_REQUEST_MAX_BODY_BYTES = 32 MiB`,作为两个 LLM 代理路由的正式请求体上限;
- 两个路由显式配置 `DefaultBodyLimit::max(32 MiB)`,避免 Axum 默认 2 MiB 提前拒绝;handler 内保留超限检查,超过 32 MiB 仍返回 `413 PAYLOAD_TOO_LARGE`;
- 补充回归测试:>2 MiB 大上下文请求不再命中默认限制;超过 32 MiB 仍返回 413。

**2. AGC 壳(apps/ai-game-creator-shell)**
- `codex_app_server` 错误映射:上游/连接层 HTTP 413、`PAYLOAD_TOO_LARGE`、`provider request too large` 等统一映射为稳定分类 `codex-app-server-error:request-too-large`,不再落入 `other` 或误判为权限/安全策略错误;
- 前端新增用户可见文案“模型请求体过大,请减少参考图或上下文后重试”,并补充单测断言。

**3. 文档**
- 【技术方案】AI游戏创作智能体App实施计划-2026-06-24.md 补充“2026-09-06 AGC LLM 代理请求体合同”:明确 32 MiB 上限、必须显式配置 `DefaultBodyLimit`、413 映射规则与用户文案。

## 验证

- `cargo test -p api-server llm_routes_accept_large_context_bodies_beyond_axum_default` 通过
- `cargo test -p api-server llm_responses_rejects_bodies_above_explicit_limit` 通过
- `cargo test ... ai-game-creator-shell ... codex_app_server_failed_turn_maps_request_too_large_details` 通过
- `npm run test -- apps/ai-game-creator-shell/tests/agentRuntimeModel.test.ts`(27 通过)

全量 workspace 测试与 CI 门禁待推送后由 CI 覆盖。

Reviewed-on: #295
Reviewed-by: 段舒康 <kdletters@qq.com>
Co-authored-by: 董羽秦 <suzmii@qq.com>
Co-committed-by: 董羽秦 <suzmii@qq.com>
2026-09-07 17:05:42 +08:00
..