diff --git a/backend/crates/providers/openai/src/compat/request.rs b/backend/crates/providers/openai/src/compat/request.rs index 7734c7a47..48f87f45f 100644 --- a/backend/crates/providers/openai/src/compat/request.rs +++ b/backend/crates/providers/openai/src/compat/request.rs @@ -5,6 +5,7 @@ use serde_json::{Map, Value, json}; /// 非官方客户端及跨客户端历史回填的兼容入口;调用方必须已选定 Codex/OAuth 上游 /// 只处理已验证被拒绝的字段,不根据客户端品牌推断整份历史是否合法 pub(super) fn normalize_non_codex_request_body(body: &mut Map) { + normalize_codex_oauth_options(body); let Some(input) = body.get_mut("input").and_then(Value::as_array_mut) else { return; }; @@ -33,6 +34,55 @@ pub(super) fn normalize_non_codex_request_body(body: &mut Map) { } } +fn normalize_codex_oauth_options(body: &mut Map) { + // Codex OAuth 拒绝这些客户端选项,即使采样参数取默认值 + for field in [ + "top_p", + "frequency_penalty", + "presence_penalty", + "chat_template_kwargs", + "user", + "metadata", + "safety_identifier", + ] { + body.shift_remove(field); + } + // Responses 自带用量;保留官方流式推理摘要等其他选项 + if let Some(options) = body + .get_mut("stream_options") + .and_then(Value::as_object_mut) + { + options.shift_remove("include_usage"); + } + // 只省略没有实际效果的配置,不静默取消客户端要求的截断、停止条件或缓存策略 + if body.get("truncation").and_then(Value::as_str) == Some("disabled") { + body.shift_remove("truncation"); + } + if body + .get("stop_sequences") + .and_then(Value::as_array) + .is_some_and(Vec::is_empty) + { + body.shift_remove("stop_sequences"); + } + if body + .get("prompt_cache_options") + .and_then(Value::as_object) + .is_some_and(Map::is_empty) + { + body.shift_remove("prompt_cache_options"); + } + // 原生上限优先,Chat 别名只补缺,非法值仍交给上游校验 + for alias in ["max_completion_tokens", "max_tokens"] { + if let Some(limit) = body.shift_remove(alias) + && !limit.is_null() + && body.get("max_output_tokens").is_none_or(Value::is_null) + { + body.insert("max_output_tokens".to_owned(), limit); + } + } +} + /// 补齐 Codex 请求缺省字段并适配已确认不兼容的请求形状,不递归清洗业务正文 /// /// 兼容基准是 Codex Core/Desktop 的模型请求,不是公开 OpenAI Responses API diff --git a/backend/crates/providers/openai/tests/provider/contract/passthrough.rs b/backend/crates/providers/openai/tests/provider/contract/passthrough.rs index 7c5b7c6dd..61e59f68c 100644 --- a/backend/crates/providers/openai/tests/provider/contract/passthrough.rs +++ b/backend/crates/providers/openai/tests/provider/contract/passthrough.rs @@ -2,6 +2,75 @@ use super::*; +#[tokio::test] +async fn responses_options_follow_the_selected_upstream_on_both_transports() { + let oauth_omissions = json!({ + "frequency_penalty": 0, "presence_penalty": 0, "top_p": 1, + "chat_template_kwargs": {"enable_thinking": false}, "user": "example", + "metadata": {"source": "example"}, "safety_identifier": "example", + "truncation": "disabled", "stop_sequences": [], "prompt_cache_options": {} + }); + let mut body = json!({ + "model": "gpt-5.4", "input": "hello", "max_tokens": 16, + "stream_options": {"include_usage": true, "reasoning_summary_delivery": "sequential_cutoff"}, + "client_metadata": {"user": "keep"}, + "tools": [{"type": "function", "name": "sample", "parameters": { + "type": "object", "properties": {"top_p": {"type": "number"}} + }}] + }); + let omissions = oauth_omissions.as_object().unwrap(); + body.as_object_mut().unwrap().extend(omissions.clone()); + for api_key in [false, true] { + for websocket in [false, true] { + let sent = capture_downstream_request(body.clone(), api_key, websocket, false).await; + for (field, value) in omissions { + assert_eq!( + sent.get(field), + api_key.then_some(value), + "{field}, api_key={api_key}, websocket={websocket}" + ); + } + if api_key { + assert_eq!(sent["stream_options"], body["stream_options"]); + assert_eq!(sent["max_tokens"], 16); + assert!(sent.get("max_output_tokens").is_none()); + } else { + assert_eq!( + sent["stream_options"], + json!({"reasoning_summary_delivery": "sequential_cutoff"}) + ); + assert_eq!(sent["max_output_tokens"], 16); + assert!(sent.get("max_tokens").is_none()); + } + assert_eq!(sent["client_metadata"]["user"], "keep"); + assert_eq!(sent["tools"], body["tools"]); + } + } +} + +#[tokio::test] +async fn codex_oauth_keeps_explicit_limits_and_context_options() { + for (native, completion, expected) in [ + (json!(16), json!(32), json!(16)), + (Value::Null, json!(32), json!(32)), + (Value::Null, Value::Null, json!(64)), + ] { + let body = json!({ + "model": "gpt-5.4", "input": "hello", "max_output_tokens": native, + "max_completion_tokens": completion, "max_tokens": 64, + "truncation": "auto", "stop_sequences": ["end"], + "prompt_cache_options": {"future_option": true} + }); + let sent = capture_downstream_request(body.clone(), false, false, false).await; + assert_eq!(sent["max_output_tokens"], expected); + assert!(sent.get("max_completion_tokens").is_none()); + assert!(sent.get("max_tokens").is_none()); + for key in ["truncation", "stop_sequences", "prompt_cache_options"] { + assert_eq!(sent[key], body[key]); + } + } +} + #[tokio::test] async fn effort_values_reach_upstream_before_and_after_middleware() { let store = Arc::new(MemoryAccountStore::default()); diff --git a/docs/api.md b/docs/api.md index f0f18e238..e9cd4cdd1 100644 --- a/docs/api.md +++ b/docs/api.md @@ -325,8 +325,8 @@ Chat 沿用 [Client Key 鉴权](#openai-数据面客户端接口)、[请求期 [请求头与身份](#请求头与身份)及[隐私策略](#隐私策略)。隐私策略的 `request_body` 面向转换后的 Responses 正文, 例如 `$.input`,不需要另外维护一套针对 `messages` 的规则 -转换后的请求按选中 Provider 的正文规则处理。OpenAI 的[正文兼容](#正文兼容)会移除顶层 `temperature` 和 -`prompt_cache_retention`。输出上限原值交给上游校验,低于模型下限时会收到上游错误;`top_p` 是否可用也由上游判断 +转换后的请求按选中的 Provider 执行。OpenAI 的字段处理见[正文兼容](#正文兼容); +输出上限原值交给上游校验,低于模型下限时会收到上游错误 ### Images 请求 @@ -454,6 +454,15 @@ OpenAI Provider 的 Responses 上游编码会移除 Codex 不接受的顶层 `te 不删除工具参数 schema、输入内容或 `client_metadata` 内的同名业务字段;其他未知字段继续透传。 `prompt_cache_key` 不用于补造会话或线程请求头;客户端未提供缓存键、会话或线程身份时保持缺省 +发往 Codex/OAuth 的 Responses 请求另按下表处理;API Key 上游保留这些字段。规则只匹配表中列出的位置 + +| 请求字段 | Codex/OAuth 出站行为 | +| --- | --- | +| `top_p`、`frequency_penalty`、`presence_penalty`、`chat_template_kwargs`、`user`、`metadata`、`safety_identifier` | 省略;采样参数取默认值时也省略 | +| `stream_options.include_usage` | 省略该成员,保留其他流式选项 | +| `truncation`、`stop_sequences`、`prompt_cache_options` | 分别仅在值为 `"disabled"`、`[]`、`{}` 时省略;显式策略交给上游校验 | +| `max_completion_tokens`、`max_tokens` | 转为 `max_output_tokens`。原生字段优先,其次按左列顺序取首个非 `null` 值 | + Codex/OAuth 上游的历史回填按字段形状兼容,不以 User-Agent 品牌区分:显式 `type: "reasoning"` 的 `input` 项移除顶层 `status`;该项具有非空字符串 `encrypted_content` 时,还会移除非空数组 `content`。其他字段及顺序保持不变,缺少非空加密载荷的明文历史由上游判定。