GenarrativeAI/Genarrative
5
0
Fork 0
Code Issues Pull Requests Actions 5 Packages Projects Releases Wiki Activity

Compare commits

base: GenarrativeAI:a18f4db4bb26715de3610a8f3bdccd9d8442169b
Branches Tags
GenarrativeAI:master GenarrativeAI:release GenarrativeAI:codex/wechat GenarrativeAI:hermes/hermes-1e775b03 GenarrativeAI:hermes/hermes-3337436a
...
compare: GenarrativeAI:d39ac86c27474b6ec2044a351dfab3a0055f7e21
Branches Tags
GenarrativeAI:master GenarrativeAI:release GenarrativeAI:codex/wechat GenarrativeAI:hermes/hermes-1e775b03 GenarrativeAI:hermes/hermes-3337436a

2 Commits
a18f4db4bb ... d39ac86c27

Author SHA1 Message Date
高物
d39ac86c27 Merge branch 'master' of http://82.157.175.59:3000/GenarrativeAI/Genarrative
Some checks failed
CI / verify (push) Has been cancelled
Details
2026-05-03 00:18:09 +08:00
高物
801d1d534a 1 2026-05-03 00:17:50 +08:00
16 changed files with 1337 additions and 449 deletions
Expand all files Collapse all files

1
.gitignore vendored
View File

@@ -29,3 +29,4 @@ temp*build*/
/public/generated-characters
/.codex-temp
/target/
/logs

3
docs/technical/README.md
View File

@@ -5,6 +5,9 @@
## 文档列表
- [PUZZLE_RUNTIME_FRONTEND_LOGIC_REHOME_2026-05-02.md](./PUZZLE_RUNTIME_FRONTEND_LOGIC_REHOME_2026-05-02.md)：记录拼图正式平台入口移动、交换、合并、拆分和通关裁决收回前端即时运行态，排行榜、下一关和游玩记录继续由后端持久化处理。
- [RPG_FOUNDATION_DRAFT_ROLE_DOSSIER_TIMEOUT_FALLBACK_2026-05-02.md](./RPG_FOUNDATION_DRAFT_ROLE_DOSSIER_TIMEOUT_FALLBACK_2026-05-02.md)：记录 `agent-foundation-*-dossier-batch-*` 无搜索 Responses 请求超时后的本地养成档案兜底，避免底稿主链被尾部角色润色阶段阻断。
- [RPG_IMAGE_GENERATION_GPT_IMAGE_2_MIGRATION_2026-05-02.md](./RPG_IMAGE_GENERATION_GPT_IMAGE_2_MIGRATION_2026-05-02.md)：记录 RPG 角色主图与场景幕背景图统一迁移到 APIMart OpenAI 兼容 `gpt-image-2` 生图入口的边界、配置和验收口径。
- [RPG_FOUNDATION_DRAFT_LANDMARK_SEED_BATCH_TIMEOUT_FIX_2026-05-02.md](./RPG_FOUNDATION_DRAFT_LANDMARK_SEED_BATCH_TIMEOUT_FIX_2026-05-02.md)：记录 `agent-foundation-landmark-seed-batch-1` 无搜索 Responses 请求超时的根因，并将场景骨架批次收敛为单场景生成。
- [PROFILE_MAIN_ROUTE_VITE_PROXY_FIX_2026-05-02.md](./PROFILE_MAIN_ROUTE_VITE_PROXY_FIX_2026-05-02.md)：记录“我的”和“存档”页面在本地把 `/api/profile/*` 请求落到 Vite SPA fallback、导致 HTML 被当 JSON 解析的根因，以及 `/api/profile` 代理补齐与回归测试。
- [SERVER_RS_DDD_WP_DEL_CLEANUP_2026-05-01.md](./SERVER_RS_DDD_WP_DEL_CLEANUP_2026-05-01.md)：记录 `WP-DEL 删除旧层与命名收口`，物理删除旧 runtime story HTTP DTO、前端 `Rpg*` alias、旧 `/api/custom-world/*` 非 runtime 前缀、Puzzle `local-next-level` 入口和 `/generated-*` 资产直读代理；生成资产读取统一走 OSS read-url 链路。
- [SERVER_RS_DDD_WP_API_BFF_CLOSURE_2026-05-01.md](./SERVER_RS_DDD_WP_API_BFF_CLOSURE_2026-05-01.md)：记录 `WP-API api-server BFF` 收尾，补齐 `/api/llm/chat/completions` 的 `stream=true` SSE 代理，明确手机号/微信配置门控和角色动画资产占位不阻塞本次 BFF 关闭。

40
docs/technical/RPG_FOUNDATION_DRAFT_LANDMARK_SEED_BATCH_TIMEOUT_FIX_2026-05-02.md Normal file
View File

@@ -0,0 +1,40 @@
# RPG foundation draft 场景骨架批次超时修正（2026-05-02）
## 背景
现场底稿生成失败信息：
```text
agent-foundation-landmark-seed-batch-1 LLM 请求失败：LLM 请求超时，累计尝试 2 次
```
本次 `logs/llm-raw` 对应请求体没有 `tools` / `web_search` 字段，说明 2026-05-01 已落地的联网搜索降级并不是本次失败根因。失败发生在无搜索 Responses 请求自身超时。
## 根因
`agent-foundation-landmark-seed-batch-1` 原本一次要求模型生成 2 个场景。每个场景又必须包含：
1. 场景基础字段。
2. `sceneTaskDescription`。
3. 3 条 `actBackgroundPromptTexts`。
4. 3 条 `actEventDescriptions`。
5. 3 个 `actNPCNames`。
6. 相连场景和进入钩子。
这个批次的输出密度明显高于角色 outline 批次。深海题材现场输入下，单次 prompt 已要求开局场景和普通关键场景同时生成，Responses 请求在默认 30 秒超时窗口内连续两次未完成，最终导致 operation 失败。
## 落地策略
1. 保持 `FOUNDATION_DRAFT_LANDMARK_COUNT = 2` 不变，仍生成 1 个开局场景和 1 个普通关键场景。
2. 将 `FOUNDATION_LANDMARK_BATCH_SIZE` 从 `2` 收敛为 `1`。
3. 第一批只生成开局场景，并继续作为 `camp` 写入。
4. 第二批只生成普通关键场景，带上已生成场景名作为 forbidden names，避免重复开局场景。
5. 单场景开局 prompt 不再写“一次性生成开局场景和普通关键场景”，避免模型在 batch_count=1 时被旧文案诱导多产。
## 验收标准
1. `generate_custom_world_foundation_draft_uses_seed_text_and_normalizes_fields` 中必须捕获 2 个 `场景框架名单` 请求。
2. 第 1 个场景请求必须明确“本批场景必须是玩家进入世界时所在的开局场景”。
3. 第 2 个场景请求必须明确“本批只生成普通关键场景”，并带上已生成开局场景名的禁止重复约束。
4. 编译后的 `camp` 仍来自第 1 个生成场景，`landmarks` 仍只保留后续普通关键场景。
5. 运行 `cargo test -p api-server custom_world_foundation_draft --manifest-path server-rs/Cargo.toml` 通过。

43
docs/technical/RPG_FOUNDATION_DRAFT_ROLE_DOSSIER_TIMEOUT_FALLBACK_2026-05-02.md Normal file
View File

@@ -0,0 +1,43 @@
# RPG foundation draft 角色养成档案超时兜底（2026-05-02）
## 背景
场景骨架批次拆小后，现场底稿生成继续失败在角色养成档案阶段：
```text
agent-foundation-story-dossier-batch-1 LLM 请求失败：LLM 请求超时，累计尝试 2 次
```
同一轮 `logs/llm-raw` 还出现过：
```text
agent-foundation-playable-dossier-batch-1 LLM 请求失败：LLM 请求超时，累计尝试 2 次
```
这些请求体都没有 `tools` / `web_search` 字段，说明本次不是联网搜索降级问题，而是无搜索 Responses 请求在角色 `dossier` 阶段自身超时。
## 根因
`dossier` 阶段要求模型为角色补齐 `backstoryReveal`、`skills` 和 `initialItems`。这部分属于结果页和运行时可继续编辑的养成档案润色，不是底稿主链必须依赖 LLM 才能成立的业务真相。
在深海题材现场输入下，即使单个可扮演角色的养成档案请求也发生超时；继续只靠减小 batch size 不能保证主链稳定。
## 落地策略
1. `narrative` 阶段仍保持 LLM 生成，因为它补的是角色背景、性格、动机和行动风格。
2. `dossier` 阶段改为 LLM 优先。
3. 若 `dossier` 阶段发生请求超时、连接失败、上游失败、空响应或 JSON 解析失败，不再让底稿 operation 失败。
4. 兜底使用当前角色对象本地生成：
- `backstoryReveal.publicSummary`
- 4 个固定好感档位章节，档位为 `15 / 30 / 60 / 90`
- 3 个技能
- 3 个初始物品
5. 本地兜底只补缺字段，不覆盖模型已经成功生成的完整字段。
6. SpacetimeDB 仍只接收 api-server 生成后的确定性 `draftProfile`；不新增外部 I/O。
## 验收标准
1. `story-dossier` 超时后，底稿生成继续完成，角色仍包含 `backstoryReveal / skills / initialItems`。
2. fallback 后的角色名必须与输入角色名一致，不能增删改名。
3. `backstoryReveal.chapters` 必须恰好 4 个，`affinityRequired` 固定为 `15 / 30 / 60 / 90`。
4. `cargo test -p api-server custom_world_foundation_draft --manifest-path server-rs/Cargo.toml` 通过。

76
docs/technical/RPG_IMAGE_GENERATION_GPT_IMAGE_2_MIGRATION_2026-05-02.md Normal file
View File

@@ -0,0 +1,76 @@
# RPG 图片生成 gpt-image-2 迁移 2026-05-02
## 背景
RPG 创作链路里有两类正式图片资产需要统一模型：
1. 角色主图候选生成。
2. 场景幕背景图生成。
旧实现中角色主图默认使用 `wan2.7-image-pro`，场景图根据是否有参考图分别使用 DashScope 文生图与图生图模型。拼图链路已经接入 APIMart 的 OpenAI 兼容 `/images/generations`，并以 `gpt-image-2` 作为默认图片模型，因此本次 RPG 图片迁移复用同一类服务端配置与请求口径。
## 落地范围
1. `POST /api/assets/character-visual/generate`
- 前端默认 `imageModel` 改为 `gpt-image-2`。
- 后端把空值、历史 `wan2.7-image-pro`、`wan2.7-image` 统一归一为 `gpt-image-2`。
- 继续保留角色主图 prompt、负向 prompt、审核失败后原创安全 prompt 兜底、PNG 去绿幕/去白底、OSS 草稿与发布链路。
2. `POST /api/runtime/custom-world/scene-image`
- 文生图与参考图生图统一走 `gpt-image-2`。
- 参考图继续只支持 Data URL 与 `/generated-*` 旧路径，经服务端回读后传给上游。
- 继续保留场景 prompt 编译、负向 prompt、OSS、`asset_object` 与 `asset_entity_binding` 链路。
3. 自动草稿资产生成中的角色主图与幕背景图分别复用上述后端函数，因此同步使用 `gpt-image-2`。
## 上游协议
服务端使用：
```text
POST {APIMART_BASE_URL}/images/generations
Authorization: Bearer {APIMART_API_KEY}
model = gpt-image-2
```
请求体统一包含：
1. `model`
2. `prompt`
3. `n`
4. `size`
5. 有参考图时增加 `image_urls`
尺寸归一规则：
1. `1024*1024`、`1024x1024`、`1:1` -> `1:1`
2. `1280*720`、`1600*900`、`16:9` -> `16:9`
响应解析兼容同步 `data[].url`、`data[].b64_json` 与异步 `task_id` / `GET /tasks/{task_id}` 结构。
## 非范围
1. 不迁移角色动作图片序列帧或视频模型。
2. 不迁移 Custom World 封面图生成。
3. 不改变 SpacetimeDB 表结构、migration 或 bindings。
4. 不改变前端 UI 面板文案。
## 配置
本次复用已有 APIMart 配置：
```text
APIMART_BASE_URL=https://api.apimart.ai/v1
APIMART_API_KEY=...
APIMART_IMAGE_REQUEST_TIMEOUT_MS=180000
```
`APIMART_API_KEY` 缺失时，角色主图与场景图返回 `SERVICE_UNAVAILABLE`，`details.provider = "apimart"`。
## 验收
1. 角色主图生成请求上游 `model` 为 `gpt-image-2`。
2. 场景图生成请求上游 `model` 为 `gpt-image-2`。
3. 旧前端或历史草稿传 `wan2.7-image-pro` 时不会回退旧模型。
4. 场景参考图生成仍能把参考图 Data URL 放入 `image_urls`。
5. 角色主图生成后仍执行原有 PNG 透明背景处理与 OSS 写入。
6. `cargo test -p api-server character_visual --manifest-path server-rs/Cargo.toml` 通过。
7. `cargo test -p api-server custom_world_ai --manifest-path server-rs/Cargo.toml` 通过。

18
server-rs/crates/api-server/logs/llm-raw/1777733782126-16788-000002-upstream_status_failed.input.json Normal file
View File

@@ -0,0 +1,18 @@
{
"provider": "ark",
"protocol": "responses",
"model": "deepseek-v3-2-251201",
"stream": false,
"attempt": 1,
"maxTokens": null,
"messages": [
{
"role": "system",
"content": "你是严格的世界草稿 JSON 生成器。\n只输出一个 JSON 对象，不要输出 Markdown、代码块、解释或额外文字。"
},
{
"role": "user",
"content": "请为下面这一批场景角色补全养成档案。\n你必须只输出一个能被 JSON.parse 直接解析的 JSON 对象，不要输出 Markdown、代码块、注释或解释。\n世界核心信息：\n世界：雾港归航\n副标题：失灯旧案\n世界概述：守灯人与群岛议会围绕沉船旧案对峙。\n世界基调：海雾悬疑\n玩家核心目标：查清父亲沉船真相\n主要势力：群岛议会、灯塔署\n核心冲突：守灯塔的旧档案被人改写。\n开局归处：旧灯塔归舍（海雾边缘的守灯人旧居。）\n关键场景：旧灯塔（雾中仍亮着错位灯火）、沉船湾（退潮后露出旧船骨）\n本批角色：\n- 议长甲 / 群岛议长\n身份：遮掩者\n框架描述：压住旧档的人\n预设好感：-10\n关系切入口：旧档案\n标签：议会\n- 潮医乙 / 潮汐医师\n身份：证人\n框架描述：知道沉船伤痕\n预设好感：20\n关系切入口：救治记录\n标签：证人\n出现场景：沉船湾\n输出 JSON 模板：\n{\n \"storyNpcs\": [\n {\n \"name\": \"角色名称\",\n \"backstoryReveal\": { \"publicSummary\": \"公开摘要\", \"chapters\": [{ \"affinityRequired\": 15, \"title\": \"羁绊章节\", \"summary\": \"章节摘要\" }] },\n \"skills\": [{ \"name\": \"技能名\", \"summary\": \"技能摘要\", \"style\": \"风格\" }],\n \"initialItems\": [{ \"name\": \"物品名\", \"category\": \"道具\", \"quantity\": 1, \"rarity\": \"common\", \"description\": \"描述\", \"tags\": [\"标签\"] }]\n }\n ]\n}\n要求：\n- 必须只补全本批角色，name 必须与本批角色完全一致，不得增删改名。\n- 每个角色必须包含：name、backstoryReveal、skills、initialItems。\n- backstoryReveal 必须包含 publicSummary 和 4 个 chapters，chapters.affinityRequired 固定为 15、30、60、90。\n- skills 默认 3 个；initialItems 默认 3 个；不要输出 backstory、personality、motivation、combatStyle。\n- 所有生成文本都必须使用中文。\n- 返回前自检：必须是一个能被 JSON.parse 直接解析的单个 JSON 对象。"
}
]
}

1
server-rs/crates/api-server/logs/llm-raw/1777733782126-16788-000002-upstream_status_failed.output.txt Normal file
View File

@@ -0,0 +1 @@
{"error":{"message":"story dossier timeout"}}

18
server-rs/crates/api-server/logs/llm-raw/1777733782128-16788-000003-upstream_status_failed.input.json Normal file
View File

@@ -0,0 +1,18 @@
{
"provider": "ark",
"protocol": "responses",
"model": "deepseek-v3-2-251201",
"stream": false,
"attempt": 1,
"maxTokens": null,
"messages": [
{
"role": "system",
"content": "你是严格的世界草稿 JSON 生成器。\n只输出一个 JSON 对象，不要输出 Markdown、代码块、解释或额外文字。"
},
{
"role": "user",
"content": "请为下面这一批场景角色补全养成档案。\n你必须只输出一个能被 JSON.parse 直接解析的 JSON 对象，不要输出 Markdown、代码块、注释或解释。\n世界核心信息：\n世界：雾港归航\n副标题：失灯旧案\n世界概述：守灯人与群岛议会围绕沉船旧案对峙。\n世界基调：海雾悬疑\n玩家核心目标：查清父亲沉船真相\n主要势力：群岛议会、灯塔署\n核心冲突：守灯塔的旧档案被人改写。\n开局归处：旧灯塔归舍（海雾边缘的守灯人旧居。）\n关键场景：旧灯塔（雾中仍亮着错位灯火）、沉船湾（退潮后露出旧船骨）\n本批角色：\n- 雾商丙 / 雾港商人\n身份：中间人\n框架描述：贩卖航线的人\n预设好感：5\n关系切入口：伪造海图\n标签：商人\n- 灯童丁 / 灯塔学徒\n身份：目击者\n框架描述：听见夜钟的人\n预设好感：30\n关系切入口：夜钟\n标签：学徒\n出现场景：旧灯塔\n输出 JSON 模板：\n{\n \"storyNpcs\": [\n {\n \"name\": \"角色名称\",\n \"backstoryReveal\": { \"publicSummary\": \"公开摘要\", \"chapters\": [{ \"affinityRequired\": 15, \"title\": \"羁绊章节\", \"summary\": \"章节摘要\" }] },\n \"skills\": [{ \"name\": \"技能名\", \"summary\": \"技能摘要\", \"style\": \"风格\" }],\n \"initialItems\": [{ \"name\": \"物品名\", \"category\": \"道具\", \"quantity\": 1, \"rarity\": \"common\", \"description\": \"描述\", \"tags\": [\"标签\"] }]\n }\n ]\n}\n要求：\n- 必须只补全本批角色，name 必须与本批角色完全一致，不得增删改名。\n- 每个角色必须包含：name、backstoryReveal、skills、initialItems。\n- backstoryReveal 必须包含 publicSummary 和 4 个 chapters，chapters.affinityRequired 固定为 15、30、60、90。\n- skills 默认 3 个；initialItems 默认 3 个；不要输出 backstory、personality、motivation、combatStyle。\n- 所有生成文本都必须使用中文。\n- 返回前自检：必须是一个能被 JSON.parse 直接解析的单个 JSON 对象。"
}
]
}

1
server-rs/crates/api-server/logs/llm-raw/1777733782128-16788-000003-upstream_status_failed.output.txt Normal file
View File

@@ -0,0 +1 @@
{"error":{"message":"story dossier timeout"}}

462
server-rs/crates/api-server/src/character_visual_assets.rs
View File

@@ -1,7 +1,4 @@
use std::{
collections::BTreeMap,
time::{Duration, Instant},
};
use std::collections::BTreeMap;
use axum::{
Json,
@@ -38,17 +35,20 @@ use crate::{
build_fallback_moderation_safe_character_visual_prompt,
},
http_error::AppError,
openai_image_generation::{
DownloadedOpenAiImage, GPT_IMAGE_2_MODEL, OpenAiImageSettings,
build_openai_image_http_client, create_openai_image_generation,
require_openai_image_settings,
},
platform_errors::map_oss_error,
request_context::RequestContext,
state::AppState,
};
use tokio::time::sleep;
const CHARACTER_VISUAL_MODEL: &str = "wan2.7-image-pro";
const CHARACTER_VISUAL_MODEL: &str = GPT_IMAGE_2_MODEL;
const CHARACTER_VISUAL_ASSET_KIND: &str = "character_visual";
const CHARACTER_VISUAL_ENTITY_KIND: &str = "character";
const CHARACTER_VISUAL_SLOT: &str = "primary_visual";
const CHARACTER_VISUAL_TASK_POLL_INTERVAL_MS: u64 = 2_500;
const CHARACTER_VISUAL_MODERATION_FALLBACK_MAX_ATTEMPTS: u8 = 2;
#[derive(Clone, Debug)]
@@ -79,7 +79,7 @@ pub async fn generate_character_visual(
let fallback_prompt =
build_fallback_moderation_safe_character_visual_prompt(payload.prompt_text.as_str());
let character_id = normalize_required_text(payload.character_id.as_str(), "character");
let model = normalize_required_text(payload.image_model.as_str(), CHARACTER_VISUAL_MODEL);
let model = resolve_character_visual_model(payload.image_model.as_str());
let size = normalize_required_text(payload.size.as_str(), "1024*1024");
let candidate_count = payload.candidate_count.clamp(1, 4);
@@ -94,8 +94,8 @@ pub async fn generate_character_visual(
.map_err(|error| character_visual_error_response(&request_context, error))?;
let result = async {
let settings = require_dashscope_settings(&state)?;
let http_client = build_dashscope_http_client(&settings)?;
let settings = require_openai_image_settings(&state)?;
let http_client = build_openai_image_http_client(&settings)?;
state
.ai_task_service()
@@ -121,6 +121,7 @@ pub async fn generate_character_visual(
"sourceMode": payload.source_mode,
"size": size,
"referenceImageCount": payload.reference_image_data_urls.len(),
"provider": "apimart",
})
.to_string(),
),
@@ -192,7 +193,7 @@ pub async fn generate_character_visual(
),
structured_payload_json: Some(
json!({
"provider": "dashscope",
"provider": "apimart",
"taskId": generated.task_id,
"model": model,
"imageCount": generated.images.len(),
@@ -307,7 +308,7 @@ pub(crate) async fn generate_character_primary_visual_for_profile(
let fallback_prompt =
build_fallback_moderation_safe_character_visual_prompt(payload.prompt_text.as_str());
let character_id = normalize_required_text(payload.character_id.as_str(), "character");
let model = normalize_required_text(payload.image_model.as_str(), CHARACTER_VISUAL_MODEL);
let model = resolve_character_visual_model(payload.image_model.as_str());
let size = normalize_required_text(payload.size.as_str(), "1024*1024");
create_visual_task(
state,
@@ -317,8 +318,8 @@ pub(crate) async fn generate_character_primary_visual_for_profile(
&model,
&prompt,
)?;
let settings = require_dashscope_settings(state)?;
let http_client = build_dashscope_http_client(&settings)?;
let settings = require_openai_image_settings(state)?;
let http_client = build_openai_image_http_client(&settings)?;
state
.ai_task_service()
.start_task(task_id.as_str(), current_utc_micros())
@@ -768,47 +769,17 @@ fn build_character_visual_job_payload(task: AiTaskSnapshot) -> CharacterAssetJob
}
}
fn require_dashscope_settings(state: &AppState) -> Result<DashScopeSettings, AppError> {
// Stage 2 的真实图片生成统一走 DashScope，这里先把配置缺失拦在业务入口前。
let base_url = state.config.dashscope_base_url.trim().trim_end_matches('/');
if base_url.is_empty() {
return Err(
AppError::from_status(StatusCode::SERVICE_UNAVAILABLE).with_details(json!({
"provider": "dashscope",
"reason": "DASHSCOPE_BASE_URL 未配置",
})),
fn resolve_character_visual_model(value: &str) -> String {
// 中文注释：旧前端和历史草稿可能仍传 wan2.7-image-pro；RPG 主图当前统一归一到 gpt-image-2。
let trimmed = value.trim();
if !trimmed.is_empty() && trimmed != CHARACTER_VISUAL_MODEL {
tracing::warn!(
requested_model = trimmed,
effective_model = CHARACTER_VISUAL_MODEL,
"角色主形象图片模型已归一到 gpt-image-2"
);
}
let api_key = state
.config
.dashscope_api_key
.as_deref()
.map(str::trim)
.filter(|value| !value.is_empty())
.ok_or_else(|| {
AppError::from_status(StatusCode::SERVICE_UNAVAILABLE).with_details(json!({
"provider": "dashscope",
"reason": "DASHSCOPE_API_KEY 未配置",
}))
})?;
Ok(DashScopeSettings {
base_url: base_url.to_string(),
api_key: api_key.to_string(),
request_timeout_ms: state.config.dashscope_image_request_timeout_ms.max(1),
})
}
fn build_dashscope_http_client(settings: &DashScopeSettings) -> Result<reqwest::Client, AppError> {
reqwest::Client::builder()
.timeout(Duration::from_millis(settings.request_timeout_ms))
.build()
.map_err(|error| {
AppError::from_status(StatusCode::INTERNAL_SERVER_ERROR).with_details(json!({
"provider": "dashscope",
"message": format!("构造 DashScope HTTP 客户端失败：{error}"),
}))
})
CHARACTER_VISUAL_MODEL.to_string()
}
async fn resolve_reference_image_as_data_url(
@@ -868,9 +839,7 @@ async fn resolve_reference_image_as_data_url(
.get(signed.signed_url)
.send()
.await
.map_err(|error| {
map_dashscope_request_error(format!("读取角色主形象参考图失败：{error}"))
})?;
.map_err(|error| map_image_request_error(format!("读取角色主形象参考图失败：{error}")))?;
let status = response.status();
let content_type = response
.headers()
@@ -879,7 +848,7 @@ async fn resolve_reference_image_as_data_url(
.unwrap_or("image/png")
.to_string();
let body = response.bytes().await.map_err(|error| {
map_dashscope_request_error(format!("读取角色主形象参考图内容失败：{error}"))
map_image_request_error(format!("读取角色主形象参考图内容失败：{error}"))
})?;
if !status.is_success() {
return Err(
@@ -911,7 +880,7 @@ async fn resolve_reference_image_as_data_url(
async fn create_character_visual_generation(
http_client: &reqwest::Client,
settings: &DashScopeSettings,
settings: &OpenAiImageSettings,
model: &str,
prompt: &str,
fallback_prompt: &str,
@@ -922,12 +891,13 @@ async fn create_character_visual_generation(
let mut active_prompt = prompt;
let mut moderation_fallback_applied = false;
let mut last_moderation_error = String::new();
let model = resolve_character_visual_model(model);
for attempt_index in 0..CHARACTER_VISUAL_MODERATION_FALLBACK_MAX_ATTEMPTS {
match create_character_visual_generation_once(
http_client,
settings,
model,
model.as_str(),
active_prompt,
size,
candidate_count,
@@ -944,7 +914,7 @@ async fn create_character_visual_generation(
if attempt_index == 0
&& !fallback_prompt.trim().is_empty()
&& fallback_prompt.trim() != prompt.trim()
&& is_dashscope_moderation_error(&error) =>
&& is_image_moderation_error(&error) =>
{
last_moderation_error = error.body_text();
active_prompt = fallback_prompt;
@@ -954,7 +924,7 @@ async fn create_character_visual_generation(
}
}
Err(map_dashscope_request_error(format!(
Err(map_image_request_error(format!(
"角色主形象安全兜底重试未返回结果：{}",
last_moderation_error.if_empty_then("上游内容审核仍未通过。")
)))
@@ -962,183 +932,44 @@ async fn create_character_visual_generation(
async fn create_character_visual_generation_once(
http_client: &reqwest::Client,
settings: &DashScopeSettings,
model: &str,
settings: &OpenAiImageSettings,
_model: &str,
prompt: &str,
size: &str,
candidate_count: u32,
reference_images: &[String],
) -> Result<GeneratedCharacterVisuals, AppError> {
let mut content = vec![json!({ "text": prompt })];
for image in reference_images {
content.push(json!({ "image": image }));
}
let response = http_client
.post(format!(
"{}/services/aigc/image-generation/generation",
settings.base_url
))
.header(
reqwest::header::AUTHORIZATION,
format!("Bearer {}", settings.api_key),
)
.header(reqwest::header::CONTENT_TYPE, "application/json")
.header("X-DashScope-Async", "enable")
.json(&json!({
"model": model,
"input": {
"messages": [
{
"role": "user",
"content": content,
}
],
},
"parameters": {
"n": candidate_count,
"size": size,
"negative_prompt": build_character_visual_negative_prompt(),
"prompt_extend": true,
"watermark": false,
},
}))
.send()
.await
.map_err(|error| map_dashscope_request_error(format!("创建角色主形象任务失败：{error}")))?;
let response_status = response.status();
let response_text = response.text().await.map_err(|error| {
map_dashscope_request_error(format!("读取角色主形象任务响应失败：{error}"))
})?;
if !response_status.is_success() {
return Err(map_dashscope_upstream_error(
response_text.as_str(),
"创建角色主形象任务失败。",
));
}
let response_json = parse_json_payload(response_text.as_str(), "创建角色主形象任务失败。")?;
let task_id = extract_task_id(&response_json.payload).ok_or_else(|| {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": "角色主形象任务未返回 task_id",
}))
})?;
let deadline = Instant::now() + Duration::from_millis(settings.request_timeout_ms);
while Instant::now() < deadline {
let poll_response = http_client
.get(format!("{}/tasks/{}", settings.base_url, task_id))
.header(
reqwest::header::AUTHORIZATION,
format!("Bearer {}", settings.api_key),
)
.send()
.await
.map_err(|error| {
map_dashscope_request_error(format!("查询角色主形象任务失败：{error}"))
})?;
let poll_status = poll_response.status();
let poll_text = poll_response.text().await.map_err(|error| {
map_dashscope_request_error(format!("读取角色主形象任务状态失败：{error}"))
})?;
if !poll_status.is_success() {
return Err(map_dashscope_upstream_error(
poll_text.as_str(),
"查询角色主形象任务失败。",
));
}
let poll_json = parse_json_payload(poll_text.as_str(), "查询角色主形象任务失败。")?;
let task_status = find_first_string_by_key(&poll_json.payload, "task_status")
.unwrap_or_default()
.trim()
.to_string();
if task_status == "SUCCEEDED" {
let image_urls = extract_image_urls(&poll_json.payload);
if image_urls.is_empty() {
return Err(
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": "角色主形象生成成功，但没有返回可下载图片。",
})),
);
}
let mut images = Vec::with_capacity(image_urls.len());
for image_url in image_urls {
images.push(
download_generated_image(
let generated = create_openai_image_generation(
http_client,
image_url.as_str(),
"下载角色主形象候选图失败。",
settings,
prompt,
Some(build_character_visual_negative_prompt().as_str()),
size,
candidate_count,
reference_images,
"角色主形象生成失败",
)
.await?,
);
}
.await?;
return Ok(GeneratedCharacterVisuals {
task_id,
actual_prompt: find_first_string_by_key(&poll_json.payload, "actual_prompt"),
Ok(GeneratedCharacterVisuals {
task_id: generated.task_id,
actual_prompt: generated.actual_prompt,
submitted_prompt: prompt.to_string(),
moderation_fallback_applied: false,
images,
});
}
if matches!(task_status.as_str(), "FAILED" | "UNKNOWN" | "CANCELED") {
return Err(map_dashscope_upstream_error(
poll_text.as_str(),
"角色主形象任务执行失败。",
));
}
sleep(Duration::from_millis(
CHARACTER_VISUAL_TASK_POLL_INTERVAL_MS,
))
.await;
}
Err(
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": "角色主形象任务执行超时，请稍后重试。",
})),
)
images: generated
.images
.into_iter()
.map(downloaded_openai_to_character_visual_image)
.collect(),
})
}
async fn download_generated_image(
http_client: &reqwest::Client,
image_url: &str,
fallback_message: &str,
) -> Result<DownloadedGeneratedImage, AppError> {
let response = http_client
.get(image_url)
.send()
.await
.map_err(|error| map_dashscope_request_error(format!("{fallback_message}：{error}")))?;
let status = response.status();
let content_type = response
.headers()
.get(reqwest::header::CONTENT_TYPE)
.and_then(|value| value.to_str().ok())
.unwrap_or("image/jpeg")
.to_string();
let body = response
.bytes()
.await
.map_err(|error| map_dashscope_request_error(format!("{fallback_message}：{error}")))?;
if !status.is_success() {
return Err(
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": fallback_message,
"status": status.as_u16(),
})),
);
}
let normalized_mime_type = normalize_downloaded_image_mime_type(content_type.as_str());
let mut bytes = body.to_vec();
let mut extension = mime_to_extension(normalized_mime_type.as_str()).to_string();
let mut mime_type = normalized_mime_type;
fn downloaded_openai_to_character_visual_image(
image: DownloadedOpenAiImage,
) -> DownloadedGeneratedImage {
let mut bytes = image.bytes;
let mut extension = image.extension;
let mut mime_type = image.mime_type;
if mime_type == "image/png"
&& let Some(optimized) = try_apply_background_alpha_to_png(bytes.as_slice())
@@ -1148,11 +979,11 @@ async fn download_generated_image(
mime_type = "image/png".to_string();
}
Ok(DownloadedGeneratedImage {
DownloadedGeneratedImage {
bytes,
mime_type,
extension,
})
}
}
/// 统一的 PNG 透明背景后处理入口。
@@ -1339,79 +1170,32 @@ fn map_character_visual_oss_error(error: platform_oss::OssError) -> AppError {
map_oss_error(error, "aliyun-oss")
}
fn parse_json_payload(
raw_text: &str,
fallback_message: &str,
) -> Result<ParsedJsonPayload, AppError> {
serde_json::from_str::<Value>(raw_text)
.map(|payload| ParsedJsonPayload { payload })
.map_err(|error| {
fn map_image_request_error(message: String) -> AppError {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": format!("{fallback_message}：解析响应失败：{error}"),
}))
})
}
fn parse_api_error_message(raw_text: &str, fallback_message: &str) -> String {
if raw_text.trim().is_empty() {
return fallback_message.to_string();
}
if let Ok(parsed) = serde_json::from_str::<Value>(raw_text) {
if let Some(message) = parsed
.pointer("/error/message")
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return message.to_string();
}
if let Some(message) = parsed
.get("message")
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return message.to_string();
}
if let Some(code) = parsed
.pointer("/error/code")
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return format!("{fallback_message}（{code}）");
}
if let Some(code) = parsed
.get("code")
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return format!("{fallback_message}（{code}）");
}
}
raw_text.trim().to_string()
}
fn map_dashscope_request_error(message: String) -> AppError {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"provider": "apimart",
"message": message,
}))
}
fn map_dashscope_upstream_error(raw_text: &str, fallback_message: &str) -> AppError {
#[cfg(test)]
fn map_image_upstream_error(raw_text: &str, fallback_message: &str) -> AppError {
let message = match raw_text.trim() {
"" => fallback_message.to_string(),
value => value.to_string(),
};
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "dashscope",
"message": parse_api_error_message(raw_text, fallback_message),
"provider": "apimart",
"message": message,
"raw": raw_text.trim(),
}))
}
fn is_dashscope_moderation_error(error: &AppError) -> bool {
#[cfg(test)]
fn is_image_test_moderation_error(error: &AppError) -> bool {
is_image_moderation_error(error)
}
fn is_image_moderation_error(error: &AppError) -> bool {
let text = error.body_text();
let normalized = text.to_ascii_lowercase();
normalized.contains("ipinfringementsuspect")
@@ -1424,77 +1208,6 @@ fn is_dashscope_moderation_error(error: &AppError) -> bool {
|| text.contains("知识产权")
}
fn collect_strings_by_key(value: &Value, target_key: &str, results: &mut Vec<String>) {
match value {
Value::Array(entries) => {
for entry in entries {
collect_strings_by_key(entry, target_key, results);
}
}
Value::Object(object) => {
for (key, nested_value) in object {
if key == target_key
&& let Some(text) = nested_value
.as_str()
.map(str::trim)
.filter(|value| !value.is_empty())
{
results.push(text.to_string());
continue;
}
collect_strings_by_key(nested_value, target_key, results);
}
}
_ => {}
}
}
fn find_first_string_by_key(value: &Value, target_key: &str) -> Option<String> {
let mut results = Vec::new();
collect_strings_by_key(value, target_key, &mut results);
results.into_iter().next()
}
fn extract_task_id(payload: &Value) -> Option<String> {
find_first_string_by_key(payload, "task_id")
}
fn extract_image_urls(payload: &Value) -> Vec<String> {
let mut urls = Vec::new();
collect_strings_by_key(payload, "image", &mut urls);
collect_strings_by_key(payload, "url", &mut urls);
let mut deduped = Vec::new();
for url in urls {
if !deduped.contains(&url) {
deduped.push(url);
}
}
deduped
}
fn normalize_downloaded_image_mime_type(content_type: &str) -> String {
let mime_type = content_type
.split(';')
.next()
.map(str::trim)
.unwrap_or("image/jpeg");
match mime_type {
"image/png" | "image/webp" | "image/jpeg" | "image/jpg" | "image/gif" => {
mime_type.to_string()
}
_ => "image/jpeg".to_string(),
}
}
fn mime_to_extension(mime_type: &str) -> &str {
match mime_type {
"image/png" => "png",
"image/webp" => "webp",
"image/gif" => "gif",
_ => "jpg",
}
}
fn parse_image_data_url(value: &str) -> Option<ParsedImageDataUrl> {
let body = value.trim().strip_prefix("data:")?;
let (mime_type, data) = body.split_once(";base64,")?;
@@ -2012,12 +1725,6 @@ impl EmptyFallback for String {
}
}
struct DashScopeSettings {
base_url: String,
api_key: String,
request_timeout_ms: u64,
}
struct GeneratedCharacterVisuals {
task_id: String,
actual_prompt: Option<String>,
@@ -2032,10 +1739,6 @@ struct DownloadedGeneratedImage {
extension: String,
}
struct ParsedJsonPayload {
payload: Value,
}
struct ParsedImageDataUrl {
mime_type: String,
bytes: Vec<u8>,
@@ -2066,13 +1769,22 @@ mod tests {
}
#[test]
fn dashscope_ip_infringement_error_uses_moderation_fallback() {
let error = map_dashscope_upstream_error(
fn legacy_character_visual_model_normalizes_to_gpt_image_2() {
assert_eq!(
resolve_character_visual_model("wan2.7-image-pro"),
"gpt-image-2"
);
assert_eq!(resolve_character_visual_model(""), "gpt-image-2");
}
#[test]
fn image_ip_infringement_error_uses_moderation_fallback() {
let error = map_image_upstream_error(
r#"{"request_id":"a18fb05d","output":{"task_id":"cb768c95","task_status":"FAILED","code":"IPInfringementSuspect","message":"Input data is suspected of being involved in IP infringement."}}"#,
"角色主形象任务执行失败。",
);
assert!(is_dashscope_moderation_error(&error));
assert!(is_image_test_moderation_error(&error));
}
#[test]

126
server-rs/crates/api-server/src/custom_world_ai.rs
View File

@@ -36,6 +36,10 @@ use crate::{
},
http_error::AppError,
llm_model_routing::CREATION_TEMPLATE_LLM_MODEL,
openai_image_generation::{
DownloadedOpenAiImage, GPT_IMAGE_2_MODEL, build_openai_image_http_client,
create_openai_image_generation, require_openai_image_settings,
},
platform_errors::map_oss_error,
prompt::scene_background::{
DEFAULT_CUSTOM_WORLD_SCENE_IMAGE_NEGATIVE_PROMPT, SceneImagePromptLandmark,
@@ -312,6 +316,8 @@ struct DownloadedRemoteImage {
bytes: Vec<u8>,
}
const RPG_SCENE_IMAGE_MODEL: &str = GPT_IMAGE_2_MODEL;
struct CoverPromptContext {
opening_act_title: String,
opening_act_summary: String,
@@ -443,7 +449,7 @@ pub async fn generate_custom_world_scene_image(
let owner_user_id = authenticated.claims().user_id().to_string();
let normalized = normalize_scene_image_request(payload)
.map_err(|error| custom_world_ai_error_response(&request_context, error))?;
require_dashscope_settings(&state)
require_openai_image_settings(&state)
.map_err(|error| custom_world_ai_error_response(&request_context, error))?;
let asset_id = format!("custom-scene-{}", current_utc_millis());
let asset = execute_billable_asset_operation(
@@ -452,8 +458,8 @@ pub async fn generate_custom_world_scene_image(
"scene_image",
asset_id.as_str(),
async {
let settings = require_dashscope_settings(&state)?;
let http_client = build_dashscope_http_client(&settings)?;
let settings = require_openai_image_settings(&state)?;
let http_client = build_openai_image_http_client(&settings)?;
let reference_image =
if let Some(reference_image_src) = normalized.reference_image_src.as_deref() {
Some(
@@ -468,46 +474,32 @@ pub async fn generate_custom_world_scene_image(
} else {
None
};
let generated = if let Some(reference_image) = reference_image.as_deref() {
create_reference_image_generation(
let reference_images = reference_image
.as_ref()
.map(|value| vec![value.clone()])
.unwrap_or_default();
let generated = create_openai_image_generation(
&http_client,
&settings,
state.config.dashscope_reference_image_model.as_str(),
normalized.prompt.as_str(),
normalized.size.as_str(),
&[reference_image.to_string()],
Some(normalized.negative_prompt.as_str()),
"创建参考图场景编辑任务失败",
"参考图场景编辑未返回图片地址",
"scene-edit",
)
.await
} else {
create_text_to_image_generation(
&http_client,
&settings,
state.config.dashscope_scene_image_model.as_str(),
normalized.prompt.as_str(),
Some(normalized.negative_prompt.as_str()),
normalized.size.as_str(),
"创建场景图片生成任务失败",
"查询场景图片任务失败",
"场景图片生成任务失败",
"场景图片生成超时或未返回图片地址",
)
.await
}?;
let scene_model = if reference_image.is_some() {
state.config.dashscope_reference_image_model.clone()
} else {
state.config.dashscope_scene_image_model.clone()
};
let downloaded = download_remote_image(
&http_client,
generated.image_url.as_str(),
"下载生成图片失败",
1,
&reference_images,
"场景图片生成失败",
)
.await?;
let downloaded = generated
.images
.into_iter()
.next()
.map(downloaded_openai_to_custom_world_image)
.ok_or_else(|| {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": "场景图片生成成功但未返回图片。",
}))
})?;
let upload = PreparedAssetUpload {
prefix: LegacyAssetPrefix::CustomWorldScenes,
path_segments: vec![
@@ -539,7 +531,7 @@ pub async fn generate_custom_world_scene_image(
image_src: String::new(),
asset_id: asset_id.clone(),
source_type: "generated".to_string(),
model: Some(scene_model),
model: Some(RPG_SCENE_IMAGE_MODEL.to_string()),
size: Some(normalized.size),
task_id: Some(generated.task_id),
prompt: Some(normalized.prompt),
@@ -588,27 +580,30 @@ pub(crate) async fn generate_custom_world_scene_image_for_profile(
}),
};
let normalized = normalize_scene_image_request(payload)?;
let settings = require_dashscope_settings(state)?;
let http_client = build_dashscope_http_client(&settings)?;
let generated = create_text_to_image_generation(
let settings = require_openai_image_settings(state)?;
let http_client = build_openai_image_http_client(&settings)?;
let generated = create_openai_image_generation(
&http_client,
&settings,
state.config.dashscope_scene_image_model.as_str(),
normalized.prompt.as_str(),
Some(normalized.negative_prompt.as_str()),
normalized.size.as_str(),
"创建场景图片生成任务失败",
"查询场景图片任务失败",
"场景图片生成任务失败",
"场景图片生成超时或未返回图片地址",
)
.await?;
let downloaded = download_remote_image(
&http_client,
generated.image_url.as_str(),
"下载生成图片失败",
1,
&[],
"场景图片生成失败",
)
.await?;
let downloaded = generated
.images
.into_iter()
.next()
.map(downloaded_openai_to_custom_world_image)
.ok_or_else(|| {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": "场景图片生成成功但未返回图片。",
}))
})?;
let asset_id = format!("custom-scene-{}", current_utc_millis());
let upload = PreparedAssetUpload {
prefix: LegacyAssetPrefix::CustomWorldScenes,
@@ -633,7 +628,7 @@ pub(crate) async fn generate_custom_world_scene_image_for_profile(
slot: "scene_image",
source_job_id: Some(generated.task_id.clone()),
};
let model = state.config.dashscope_scene_image_model.clone();
let model = RPG_SCENE_IMAGE_MODEL.to_string();
let prompt = normalized.prompt.clone();
let asset = persist_custom_world_asset(
state,
@@ -1634,6 +1629,14 @@ async fn download_remote_image(
})
}
fn downloaded_openai_to_custom_world_image(image: DownloadedOpenAiImage) -> DownloadedRemoteImage {
DownloadedRemoteImage {
extension: image.extension,
mime_type: image.mime_type,
bytes: image.bytes,
}
}
fn optimize_uploaded_cover_image(
parsed_data_url: &ParsedImageDataUrl,
crop_rect: &CustomWorldCoverCropRect,
@@ -2451,6 +2454,12 @@ mod tests {
serde_json::from_slice(&body).expect("body should be valid json")
}
fn build_state_without_apimart_key() -> AppState {
let mut config = AppConfig::default();
config.apimart_api_key = None;
AppState::new(config).expect("state should build")
}
fn build_state_without_dashscope_key() -> AppState {
let mut config = AppConfig::default();
config.dashscope_api_key = None;
@@ -2458,8 +2467,8 @@ mod tests {
}
#[tokio::test]
async fn scene_image_returns_service_unavailable_when_dashscope_missing() {
let state = build_state_without_dashscope_key();
async fn scene_image_returns_service_unavailable_when_apimart_missing() {
let state = build_state_without_apimart_key();
let request_context = build_request_context("POST /api/runtime/custom-world/scene-image");
let authenticated = build_authenticated(&state);
@@ -2482,7 +2491,7 @@ mod tests {
})),
)
.await
.expect_err("missing dashscope should fail");
.expect_err("missing apimart should fail");
let payload = read_error_response(response).await;
assert_eq!(
@@ -2491,7 +2500,7 @@ mod tests {
);
assert_eq!(
payload["error"]["details"]["provider"],
Value::String("dashscope".to_string())
Value::String("apimart".to_string())
);
}
@@ -2612,6 +2621,11 @@ mod tests {
assert_eq!(normalized.prompt, manual_prompt);
}
#[test]
fn scene_image_response_model_is_gpt_image_2() {
assert_eq!(RPG_SCENE_IMAGE_MODEL, "gpt-image-2");
}
#[tokio::test]
async fn cover_image_returns_service_unavailable_when_dashscope_missing() {
let state = build_state_without_dashscope_key();

334
server-rs/crates/api-server/src/custom_world_foundation_draft.rs
View File

@@ -165,7 +165,8 @@ const FOUNDATION_DRAFT_PLAYABLE_COUNT: usize = 1;
const FOUNDATION_DRAFT_STORY_COUNT: usize = 8;
const FOUNDATION_DRAFT_LANDMARK_COUNT: usize = 2;
const FOUNDATION_ROLE_OUTLINE_BATCH_SIZE: usize = 2;
const FOUNDATION_LANDMARK_BATCH_SIZE: usize = 2;
// 中文注释：单个场景已经包含三幕事件、三幕背景图 prompt 和 NPC 分配；按 1 个场景拆批，避免 landmark seed 大 JSON 在 Responses 请求中超时。
const FOUNDATION_LANDMARK_BATCH_SIZE: usize = 1;
const FOUNDATION_ROLE_DETAIL_BATCH_SIZE: usize = 2;
const WORLD_ATTRIBUTE_SLOT_IDS: [&str; 6] =
["axis_a", "axis_b", "axis_c", "axis_d", "axis_e", "axis_f"];
@@ -586,7 +587,7 @@ async fn expand_foundation_role_entries(
.as_str(),
to_batch_progress(progress_range, processed_count, base_entries.len()),
);
let raw = request_foundation_json_stage(
let raw_result = request_foundation_json_stage(
llm_client,
build_custom_world_role_batch_prompt(framework, role_type, batch, stage),
format!(
@@ -610,8 +611,20 @@ async fn expand_foundation_role_entries(
"角色档案补全阶段没有返回有效内容。",
enable_web_search,
)
.await?;
merged_entries.extend(array_field(&raw, role_key(role_type)));
.await;
match raw_result {
Ok(raw) => merged_entries.extend(array_field(&raw, role_key(role_type))),
Err(error) if stage == "dossier" => {
warn!(
error = %error,
role_type,
batch_index = batch_index + 1,
"foundation draft 角色养成档案 LLM 补全失败，使用本地结构化兜底"
);
merged_entries.extend(build_fallback_role_dossier_entries(batch));
}
Err(error) => return Err(error),
}
processed_count = processed_count
.saturating_add(batch.len())
.min(base_entries.len());
@@ -635,6 +648,103 @@ async fn expand_foundation_role_entries(
Ok(merge_entries_by_name(base_entries, &merged_entries))
}
fn build_fallback_role_dossier_entries(entries: &[JsonValue]) -> Vec<JsonValue> {
entries
.iter()
.enumerate()
.map(|(index, entry)| build_fallback_role_dossier_entry(entry, index))
.collect()
}
fn build_fallback_role_dossier_entry(entry: &JsonValue, index: usize) -> JsonValue {
let name = json_text(entry, "name").unwrap_or_else(|| format!("角色{}", index + 1));
let title = json_text(entry, "title").unwrap_or_default();
let role = json_text(entry, "role").unwrap_or_else(|| "关键角色".to_string());
let description = json_text(entry, "description").unwrap_or_else(|| role.clone());
let backstory = json_text(entry, "backstory").unwrap_or_else(|| description.clone());
let motivation = json_text(entry, "motivation").unwrap_or_else(|| description.clone());
let tag = json_string_array(entry, "tags")
.and_then(|items| items.first().cloned())
.unwrap_or_else(|| role.clone());
let item_prefix = if title.trim().is_empty() {
name.clone()
} else {
title.clone()
};
json!({
"name": name.clone(),
"backstoryReveal": {
"publicSummary": format!("{name}的公开档案围绕“{description}”展开。"),
"chapters": [
{
"affinityRequired": 15,
"title": "初识",
"summary": format!("{name}以{role}身份进入玩家视野，留下与“{tag}”有关的第一条线索。"),
},
{
"affinityRequired": 30,
"title": "试探",
"summary": format!("{name}开始透露“{backstory}”背后的压力，但仍保留关键隐情。"),
},
{
"affinityRequired": 60,
"title": "共同行动",
"summary": format!("{name}围绕“{motivation}”与玩家形成更明确的合作或冲突。"),
},
{
"affinityRequired": 90,
"title": "真相",
"summary": format!("{name}交出与“{description}”相关的核心选择，关系走向定型。"),
},
],
},
"skills": [
{
"name": format!("{tag}洞察"),
"summary": format!("围绕“{description}”判断局势与隐藏线索。"),
"style": "侦查",
},
{
"name": format!("{item_prefix}协助"),
"summary": format!("以{role}身份为玩家提供行动支援。"),
"style": "支援",
},
{
"name": "临场应变",
"summary": format!("在压力升级时根据“{motivation}”调整行动。"),
"style": "应变",
},
],
"initialItems": [
{
"name": format!("{item_prefix}记录"),
"category": "道具",
"quantity": 1,
"rarity": "common",
"description": format!("记录{name}与“{description}”相关的线索。"),
"tags": [tag.clone()],
},
{
"name": format!("{tag}信物"),
"category": "道具",
"quantity": 1,
"rarity": "common",
"description": format!("能证明{name}身份和立场的随身物。"),
"tags": [role.clone()],
},
{
"name": "备用补给",
"category": "消耗品",
"quantity": 1,
"rarity": "common",
"description": format!("{name}在关键行动前准备的基础补给。"),
"tags": ["补给"],
},
],
})
}
fn emit_foundation_draft_progress(
on_progress: &mut (impl FnMut(CustomWorldFoundationDraftProgress) + Send),
phase_label: &str,
@@ -2570,6 +2680,60 @@ mod tests {
);
}
#[test]
fn role_dossier_fallback_keeps_names_and_required_fields() {
let entries = vec![json!({
"name": "埃琳娜·沃克",
"title": "深渊学者",
"role": "深海科研联盟成员",
"description": "执着研究深海生物发光现象的年轻科学家",
"backstory": "她长期追踪发光生物与古代遗迹之间的联系。",
"motivation": "用氧气补给换取玩家的目击信息",
"tags": ["科研人员", "偏执学者"]
})];
let fallback = build_fallback_role_dossier_entries(&entries);
let first = fallback.first().expect("fallback entry should exist");
assert_eq!(first.get("name"), Some(&json!("埃琳娜·沃克")));
assert_eq!(
first
.get("backstoryReveal")
.and_then(|value| value.get("chapters"))
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(4)
);
assert_eq!(
first
.get("backstoryReveal")
.and_then(|value| value.get("chapters"))
.and_then(JsonValue::as_array)
.map(|chapters| {
chapters
.iter()
.filter_map(|chapter| chapter.get("affinityRequired"))
.cloned()
.collect::<Vec<_>>()
}),
Some(vec![json!(15), json!(30), json!(60), json!(90)])
);
assert_eq!(
first
.get("skills")
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(3)
);
assert_eq!(
first
.get("initialItems")
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(3)
);
}
#[tokio::test]
async fn generate_custom_world_foundation_draft_uses_seed_text_and_normalizes_fields() {
let request_capture = Arc::new(Mutex::new(Vec::new()));
@@ -2595,7 +2759,10 @@ mod tests {
r#"{"storyNpcs":[{"name":"档吏庚","title":"旧档吏","role":"保管者","description":"藏起原始卷宗","visualDescription":"褐色旧档袍袖口磨白，背着沉重文书匣，眼镜后目光闪躲。","actionDescription":"翻找卷宗时动作极快，被追问便把文书匣抱紧后退。","sceneVisualDescription":"他常守在潮湿档案室深处，旧柜标签被盐雾泡卷。","initialAffinity":10,"relationshipHooks":["原始卷宗"],"tags":["档案"]},{"name":"潮女辛","title":"听潮女","role":"引路人","description":"听懂海雾低语","visualDescription":"银灰长发被贝壳绳束起，披轻薄潮纹披肩，赤足沾水。","actionDescription":"侧耳听潮后抬手指向雾中路径，步伐像避开暗流。","sceneVisualDescription":"她常站在礁石浅水间，海雾绕过脚踝，远处灯火错位。","initialAffinity":35,"relationshipHooks":["海雾低语"],"tags":["引路"]}]}"#,
),
llm_response(
r#"{"landmarks":[{"name":"旧灯塔","description":"雾中仍亮着错位灯火","visualDescription":"旧灯塔立在雾港高礁上，灯室漏出错位光束，石阶和回廊留出可站立空间。","sceneTaskDescription":"首次进入旧灯塔时，追查被篡改的灯火航线记录。","actBackgroundPromptTexts":["雾港高礁上的旧灯塔亮起错位灯火，灯童丁抱灯站在螺旋楼梯口。","潮湿档案室里灯火忽明忽暗，档吏庚抱紧文书匣，海图在桌面卷起。","灯室玻璃被海风震响，灯童丁指向错位航线，远处沉船湾雾光浮现。"],"actEventDescriptions":["灯童丁听见夜钟后发现灯火记录被人动过。","档吏庚试图带走原始卷宗，冲突在灯塔档案室升级。","灯童丁交出旧钥匙，玩家必须决定是否立刻追向沉船湾。"],"actNPCNames":["灯童丁","档吏庚","灯童丁"],"connectedLandmarkNames":["沉船湾"],"entryHook":"灯火按被篡改的航线闪烁。"},{"name":"沉船湾","description":"退潮后露出旧船骨","visualDescription":"退潮泥滩露出黑色旧船骨，破帆挂在礁石间，临时诊台灯影摇晃。","sceneTaskDescription":"首次进入沉船湾时，辨认旧船骨里残留的沉船真相。","actBackgroundPromptTexts":["沉船湾退潮泥滩露出旧船骨，船魂戊浮在黑色肋骨般的船梁旁。","湿木棚下潮医乙翻看伤痕记录，海水漫过脚边，巡海灯逼近湾口。","旧船骨深处传出暗号，船魂戊指向被封住的货舱，雾中灯塔光线错位。"],"actEventDescriptions":["船魂戊在退潮声里显形，指认父亲留下的暗号。","潮医乙发现伤痕与官方记录不符，巡海封锁让局势升级。","船魂戊带玩家接近旧货舱，必须在追捕前取走关键证物。"],"actNPCNames":["船魂戊","潮医乙","船魂戊"],"connectedLandmarkNames":["旧灯塔"],"entryHook":"旧船骨里传出父亲留下的暗号。"}]}"#,
r#"{"landmarks":[{"name":"旧灯塔","description":"雾中仍亮着错位灯火","visualDescription":"旧灯塔立在雾港高礁上，灯室漏出错位光束，石阶和回廊留出可站立空间。","sceneTaskDescription":"首次进入旧灯塔时，追查被篡改的灯火航线记录。","actBackgroundPromptTexts":["雾港高礁上的旧灯塔亮起错位灯火，灯童丁抱灯站在螺旋楼梯口。","潮湿档案室里灯火忽明忽暗，档吏庚抱紧文书匣，海图在桌面卷起。","灯室玻璃被海风震响，灯童丁指向错位航线，远处沉船湾雾光浮现。"],"actEventDescriptions":["灯童丁听见夜钟后发现灯火记录被人动过。","档吏庚试图带走原始卷宗，冲突在灯塔档案室升级。","灯童丁交出旧钥匙，玩家必须决定是否立刻追向沉船湾。"],"actNPCNames":["灯童丁","档吏庚","灯童丁"],"connectedLandmarkNames":["沉船湾"],"entryHook":"灯火按被篡改的航线闪烁。"}]}"#,
),
llm_response(
r#"{"landmarks":[{"name":"沉船湾","description":"退潮后露出旧船骨","visualDescription":"退潮泥滩露出黑色旧船骨，破帆挂在礁石间，临时诊台灯影摇晃。","sceneTaskDescription":"首次进入沉船湾时，辨认旧船骨里残留的沉船真相。","actBackgroundPromptTexts":["沉船湾退潮泥滩露出旧船骨，船魂戊浮在黑色肋骨般的船梁旁。","湿木棚下潮医乙翻看伤痕记录，海水漫过脚边，巡海灯逼近湾口。","旧船骨深处传出暗号，船魂戊指向被封住的货舱，雾中灯塔光线错位。"],"actEventDescriptions":["船魂戊在退潮声里显形，指认父亲留下的暗号。","潮医乙发现伤痕与官方记录不符，巡海封锁让局势升级。","船魂戊带玩家接近旧货舱，必须在追捕前取走关键证物。"],"actNPCNames":["船魂戊","潮医乙","船魂戊"],"connectedLandmarkNames":["旧灯塔"],"entryHook":"旧船骨里传出父亲留下的暗号。"}]}"#,
),
llm_response(
r#"{"playableNpcs":[{"name":"岑灯","backstory":"被停职的守灯人返乡后发现父亲沉船案被改写。","personality":"克制执拗","motivation":"查清父亲沉船真相","combatStyle":"借灯火与海图周旋"}]}"#,
@@ -2624,15 +2791,24 @@ mod tests {
.expect("request capture should lock")
.clone();
let request_text = captured_requests.join("\n---request---\n");
let landmark_seed_requests = captured_requests
.iter()
.filter(|request| request.contains("场景框架名单"))
.collect::<Vec<_>>();
assert!(captured_requests.len() >= 17);
assert!(captured_requests.len() >= 18);
assert!(request_text.contains("在失真的海图上追查一场被篡改的沉船事故。"));
assert!(request_text.contains("世界核心骨架"));
assert!(request_text.contains("attributeSchema"));
assert!(request_text.contains("可扮演角色框架名单"));
assert!(request_text.contains("场景角色框架名单"));
assert!(request_text.contains("场景框架名单"));
assert!(request_text.contains("第一条场景必须是玩家进入世界时所在的开局场景"));
assert_eq!(landmark_seed_requests.len(), 2);
assert!(landmark_seed_requests[0].contains("本批场景必须是玩家进入世界时所在的开局场景"));
assert!(landmark_seed_requests[0].contains("必须生成恰好 1 个场景"));
assert!(landmark_seed_requests[1].contains("本批只生成普通关键场景"));
assert!(landmark_seed_requests[1].contains("这些场景已经生成，禁止重复：旧灯塔"));
assert!(!landmark_seed_requests[0].contains("一次性生成开局场景和普通关键场景"));
assert!(request_text.contains("camp 只表示玩家开局时的落脚处占位"));
assert!(!request_text.contains("camp.sceneTaskDescription"));
assert!(!request_text.contains("camp.actBackgroundPromptTexts"));
@@ -2845,6 +3021,150 @@ mod tests {
);
}
#[tokio::test]
async fn role_dossier_timeout_uses_local_fallback_and_keeps_generation_alive() {
let request_capture = Arc::new(Mutex::new(Vec::new()));
let server_url = spawn_mock_server_with_statuses(
request_capture.clone(),
vec![
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"name":"雾港归航","subtitle":"失灯旧案","summary":"守灯人与群岛议会围绕沉船旧案对峙。","tone":"海雾悬疑","playerGoal":"查清父亲沉船真相","templateWorldType":"WUXIA","majorFactions":["群岛议会","灯塔署"],"coreConflicts":["守灯塔的旧档案被人改写。"],"attributeSchema":{"slots":[{"name":"灯骨"},{"name":"潮步"},{"name":"灯识"},{"name":"雾魄"},{"name":"旧约"},{"name":"回澜"}]},"camp":{"name":"旧灯塔归舍","description":"海雾边缘的守灯人旧居。"}}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"playableNpcs":[{"name":"岑灯","title":"返乡守灯人","role":"主角代理","description":"追查旧案的人","visualDescription":"灰蓝旧灯披风压着海盐痕，腰侧挂旧海图筒和短灯杖。","actionDescription":"举灯照海图，短杖点地辨认潮声。","sceneVisualDescription":"旧灯塔回廊被海雾压低，墙上挂满潮湿航线图。","initialAffinity":24,"relationshipHooks":["旧案牵连"],"tags":["守灯人"]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"议长甲","title":"群岛议长","role":"遮掩者","description":"压住旧档的人","visualDescription":"深色议会长袍垂到靴边，银扣像封蜡，手里总夹着旧档袋。","actionDescription":"抬手下令封锁，动作缓慢却压迫感强。","sceneVisualDescription":"他常出现在议会石厅高处，旧档柜阴影切过半张脸。","initialAffinity":-10,"relationshipHooks":["旧档案"],"tags":["议会"]},{"name":"潮医乙","title":"潮汐医师","role":"证人","description":"知道沉船伤痕","visualDescription":"浅灰防潮医袍挽到肘部，药箱铜扣发暗，袖口沾着海盐。","actionDescription":"俯身检查伤痕并快速记录潮汐症状，动作谨慎而利落。","sceneVisualDescription":"他常在沉船湾临时诊台前，背后是湿木棚和摇晃药灯。","initialAffinity":20,"relationshipHooks":["救治记录"],"tags":["证人"]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"雾商丙","title":"雾港商人","role":"中间人","description":"贩卖航线的人","visualDescription":"暗绿长外套挂满防水口袋，帽檐压低，腰间藏着卷曲海图。","actionDescription":"摊开假海图低声议价，手指总按着袖中短刃。","sceneVisualDescription":"他常站在雾港货棚阴影里，周围堆着封蜡货箱和潮湿灯牌。","initialAffinity":5,"relationshipHooks":["伪造海图"],"tags":["商人"]},{"name":"灯童丁","title":"灯塔学徒","role":"目击者","description":"听见夜钟的人","visualDescription":"瘦小学徒披着过大的灯塔制服，怀里抱黄铜小灯和旧钥匙。","actionDescription":"抱灯快步穿过回廊，听见夜钟时会突然停住回头。","sceneVisualDescription":"他常出现在灯塔螺旋楼梯间，雾光从窄窗切进灰墙。","initialAffinity":30,"relationshipHooks":["夜钟"],"tags":["学徒"]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"船魂戊","title":"沉船残魂","role":"异类","description":"困在潮声里","visualDescription":"半透明水渍轮廓披着破碎船员衣，胸口嵌着发暗船钉。","actionDescription":"随潮声漂移抬手指路，情绪激烈时水雾会拉长身影。","sceneVisualDescription":"它常浮在沉船湾退潮泥滩上，身后旧船骨像黑色肋骨。","initialAffinity":-20,"relationshipHooks":["沉船真相"],"tags":["异类"]},{"name":"巡海己","title":"巡海队长","role":"追捕者","description":"封锁海岸线","visualDescription":"深蓝巡海甲衣覆着雨水，肩章锋利，手握带灯长枪。","actionDescription":"举枪封路并用灯束扫过海岸，步伐整齐带压迫感。","sceneVisualDescription":"他常立在封锁栈桥尽头，巡海灯和铁链把退路切断。","initialAffinity":-15,"relationshipHooks":["封锁令"],"tags":["巡海"]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"档吏庚","title":"旧档吏","role":"保管者","description":"藏起原始卷宗","visualDescription":"褐色旧档袍袖口磨白，背着沉重文书匣，眼镜后目光闪躲。","actionDescription":"翻找卷宗时动作极快，被追问便把文书匣抱紧后退。","sceneVisualDescription":"他常守在潮湿档案室深处，旧柜标签被盐雾泡卷。","initialAffinity":10,"relationshipHooks":["原始卷宗"],"tags":["档案"]},{"name":"潮女辛","title":"听潮女","role":"引路人","description":"听懂海雾低语","visualDescription":"银灰长发被贝壳绳束起，披轻薄潮纹披肩，赤足沾水。","actionDescription":"侧耳听潮后抬手指向雾中路径，步伐像避开暗流。","sceneVisualDescription":"她常站在礁石浅水间，海雾绕过脚踝，远处灯火错位。","initialAffinity":35,"relationshipHooks":["海雾低语"],"tags":["引路"]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"landmarks":[{"name":"旧灯塔","description":"雾中仍亮着错位灯火","visualDescription":"旧灯塔立在雾港高礁上，灯室漏出错位光束，石阶和回廊留出可站立空间。","sceneTaskDescription":"首次进入旧灯塔时，追查被篡改的灯火航线记录。","actBackgroundPromptTexts":["雾港高礁上的旧灯塔亮起错位灯火，灯童丁抱灯站在螺旋楼梯口。","潮湿档案室里灯火忽明忽暗，档吏庚抱紧文书匣，海图在桌面卷起。","灯室玻璃被海风震响，灯童丁指向错位航线，远处沉船湾雾光浮现。"],"actEventDescriptions":["灯童丁听见夜钟后发现灯火记录被人动过。","档吏庚试图带走原始卷宗，冲突在灯塔档案室升级。","灯童丁交出旧钥匙，玩家必须决定是否立刻追向沉船湾。"],"actNPCNames":["灯童丁","档吏庚","灯童丁"],"connectedLandmarkNames":["沉船湾"],"entryHook":"灯火按被篡改的航线闪烁。"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"landmarks":[{"name":"沉船湾","description":"退潮后露出旧船骨","visualDescription":"退潮泥滩露出黑色旧船骨，破帆挂在礁石间，临时诊台灯影摇晃。","sceneTaskDescription":"首次进入沉船湾时，辨认旧船骨里残留的沉船真相。","actBackgroundPromptTexts":["沉船湾退潮泥滩露出旧船骨，船魂戊浮在黑色肋骨般的船梁旁。","湿木棚下潮医乙翻看伤痕记录，海水漫过脚边，巡海灯逼近湾口。","旧船骨深处传出暗号，船魂戊指向被封住的货舱，雾中灯塔光线错位。"],"actEventDescriptions":["船魂戊在退潮声里显形，指认父亲留下的暗号。","潮医乙发现伤痕与官方记录不符，巡海封锁让局势升级。","船魂戊带玩家接近旧货舱，必须在追捕前取走关键证物。"],"actNPCNames":["船魂戊","潮医乙","船魂戊"],"connectedLandmarkNames":["旧灯塔"],"entryHook":"旧船骨里传出父亲留下的暗号。"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"playableNpcs":[{"name":"岑灯","backstory":"被停职的守灯人返乡后发现父亲沉船案被改写。","personality":"克制执拗","motivation":"查清父亲沉船真相","combatStyle":"借灯火与海图周旋"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"playableNpcs":[{"name":"岑灯","backstoryReveal":{"publicSummary":"返乡守灯人的旧案羁绊。","chapters":[{"affinityRequired":15,"title":"返乡","summary":"回到旧灯塔。"},{"affinityRequired":30,"title":"旧档","summary":"发现档案错页。"},{"affinityRequired":60,"title":"沉船","summary":"接近沉船湾。"},{"affinityRequired":90,"title":"真相","summary":"直面议会遮掩。"}]},"skills":[{"name":"读灯","summary":"辨认灯火暗号","style":"侦查"}],"initialItems":[{"name":"旧海图","category":"道具","quantity":1,"rarity":"common","description":"父亲留下的海图。","tags":["线索"]}]}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"议长甲","backstory":"长期维持群岛议会体面并遮掩沉船旧案。","personality":"冷硬周密","motivation":"压住旧案","combatStyle":"以权令封锁线索"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"潮医乙","backstory":"他保存着沉船伤痕和潮汐症状的旧记录。","personality":"谨慎利落","motivation":"保住证据","combatStyle":"以医疗知识支援判断"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"雾商丙","backstory":"他长期倒卖雾港航线和假海图。","personality":"圆滑警惕","motivation":"从旧案里脱身","combatStyle":"以情报和交易周旋"}]}"#,
),
},
MockHttpResponse {
status_code: 200,
body: llm_response(
r#"{"storyNpcs":[{"name":"船魂戊","backstory":"它被沉船旧案困在潮声和船骨之间。","personality":"激烈执拗","motivation":"让真相重新浮上海面","combatStyle":"借潮声与残影指路"}]}"#,
),
},
MockHttpResponse {
status_code: 504,
body: r#"{"error":{"message":"story dossier timeout"}}"#.to_string(),
},
MockHttpResponse {
status_code: 504,
body: r#"{"error":{"message":"story dossier timeout"}}"#.to_string(),
},
],
);
let llm_client = build_test_llm_client(server_url);
let session = build_test_session();
let result = generate_custom_world_foundation_draft(&llm_client, &session, false, |_| {})
.await
.expect("dossier fallback should keep draft generation alive");
let draft_profile = serde_json::from_str::<JsonValue>(&result.draft_profile_json)
.expect("draft profile should parse");
let first_story = draft_profile
.get("storyNpcs")
.and_then(JsonValue::as_array)
.and_then(|entries| entries.first())
.expect("first story role should exist");
assert_eq!(first_story.get("name"), Some(&json!("议长甲")));
assert_eq!(
first_story
.get("backstoryReveal")
.and_then(|value| value.get("chapters"))
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(4)
);
assert_eq!(
first_story
.get("skills")
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(3)
);
assert_eq!(
first_story
.get("initialItems")
.and_then(JsonValue::as_array)
.map(Vec::len),
Some(3)
);
let request_text = request_capture
.lock()
.expect("request capture should lock")
.join("\n---request---\n");
assert!(request_text.contains("请为下面这一批场景角色补全养成档案"));
}
#[test]
fn generated_scene_batch_first_entry_becomes_opening_camp() {
let fallback_camp = json!({

1
server-rs/crates/api-server/src/main.rs
View File

@@ -40,6 +40,7 @@ mod login_options;
mod logout;
mod logout_all;
mod match3d;
mod openai_image_generation;
mod password_entry;
mod password_management;
mod phone_auth;

632
server-rs/crates/api-server/src/openai_image_generation.rs Normal file
View File

@@ -0,0 +1,632 @@
use std::time::{Duration, Instant};
use axum::http::StatusCode;
use base64::{Engine as _, engine::general_purpose::STANDARD as BASE64_STANDARD};
use reqwest::header;
use serde_json::{Map, Value, json};
use tokio::time::sleep;
use crate::{http_error::AppError, state::AppState};
pub(crate) const GPT_IMAGE_2_MODEL: &str = "gpt-image-2";
#[derive(Clone, Debug)]
pub(crate) struct OpenAiImageSettings {
pub base_url: String,
pub api_key: String,
pub request_timeout_ms: u64,
}
#[derive(Clone, Debug)]
pub(crate) struct OpenAiGeneratedImages {
pub task_id: String,
pub actual_prompt: Option<String>,
pub images: Vec<DownloadedOpenAiImage>,
}
#[derive(Clone, Debug)]
pub(crate) struct DownloadedOpenAiImage {
pub bytes: Vec<u8>,
pub mime_type: String,
pub extension: String,
}
// 中文注释：RPG 图片资产与拼图一样走 APIMart 的 OpenAI 兼容图片入口，避免把密钥或供应商协议暴露到前端。
pub(crate) fn require_openai_image_settings(
state: &AppState,
) -> Result<OpenAiImageSettings, AppError> {
let base_url = state.config.apimart_base_url.trim().trim_end_matches('/');
if base_url.is_empty() {
return Err(
AppError::from_status(StatusCode::SERVICE_UNAVAILABLE).with_details(json!({
"provider": "apimart",
"reason": "APIMART_BASE_URL 未配置",
})),
);
}
let api_key = state
.config
.apimart_api_key
.as_deref()
.map(str::trim)
.filter(|value| !value.is_empty())
.ok_or_else(|| {
AppError::from_status(StatusCode::SERVICE_UNAVAILABLE).with_details(json!({
"provider": "apimart",
"reason": "APIMART_API_KEY 未配置",
}))
})?;
Ok(OpenAiImageSettings {
base_url: base_url.to_string(),
api_key: api_key.to_string(),
request_timeout_ms: state.config.apimart_image_request_timeout_ms.max(1),
})
}
pub(crate) fn build_openai_image_http_client(
settings: &OpenAiImageSettings,
) -> Result<reqwest::Client, AppError> {
reqwest::Client::builder()
.timeout(Duration::from_millis(settings.request_timeout_ms))
.build()
.map_err(|error| {
AppError::from_status(StatusCode::INTERNAL_SERVER_ERROR).with_details(json!({
"provider": "apimart",
"message": format!("构造 APIMart 图片生成 HTTP 客户端失败：{error}"),
}))
})
}
pub(crate) async fn create_openai_image_generation(
http_client: &reqwest::Client,
settings: &OpenAiImageSettings,
prompt: &str,
negative_prompt: Option<&str>,
size: &str,
candidate_count: u32,
reference_images: &[String],
failure_context: &str,
) -> Result<OpenAiGeneratedImages, AppError> {
let request_body = build_openai_image_request_body(
prompt,
negative_prompt,
size,
candidate_count,
reference_images,
);
let response = http_client
.post(format!("{}/images/generations", settings.base_url))
.header(
header::AUTHORIZATION,
format!("Bearer {}", settings.api_key),
)
.header(header::CONTENT_TYPE, "application/json")
.json(&request_body)
.send()
.await
.map_err(|error| {
map_openai_image_request_error(format!(
"{failure_context}：创建图片生成任务失败：{error}"
))
})?;
let response_status = response.status();
let response_text = response.text().await.map_err(|error| {
map_openai_image_request_error(format!("{failure_context}：读取图片生成响应失败：{error}"))
})?;
if !response_status.is_success() {
return Err(map_openai_image_upstream_error(
response_status.as_u16(),
response_text.as_str(),
failure_context,
));
}
let response_json = parse_json_payload(response_text.as_str(), failure_context)?;
let image_urls = extract_image_urls(&response_json.payload);
if !image_urls.is_empty() {
return download_images_from_urls(
http_client,
format!("apimart-{}", current_utc_micros()),
image_urls,
candidate_count,
)
.await;
}
let b64_images = extract_b64_images(&response_json.payload);
if !b64_images.is_empty() {
return Ok(images_from_base64(
format!("apimart-{}", current_utc_micros()),
b64_images,
candidate_count,
));
}
let task_id = extract_task_id(&response_json.payload).ok_or_else(|| {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": format!("{failure_context}：上游未返回 task_id 或图片"),
}))
})?;
wait_openai_generated_images(
http_client,
settings,
task_id.as_str(),
candidate_count,
failure_context,
)
.await
}
pub(crate) fn build_openai_image_request_body(
prompt: &str,
negative_prompt: Option<&str>,
size: &str,
candidate_count: u32,
reference_images: &[String],
) -> Value {
let mut body = Map::from_iter([
(
"model".to_string(),
Value::String(GPT_IMAGE_2_MODEL.to_string()),
),
(
"prompt".to_string(),
Value::String(build_prompt_with_negative(prompt, negative_prompt)),
),
("n".to_string(), json!(candidate_count.clamp(1, 4))),
(
"size".to_string(),
Value::String(normalize_image_size(size)),
),
]);
if !reference_images.is_empty() {
body.insert("image_urls".to_string(), json!(reference_images));
}
Value::Object(body)
}
fn build_prompt_with_negative(prompt: &str, negative_prompt: Option<&str>) -> String {
let prompt = prompt.trim();
let Some(negative_prompt) = negative_prompt
.map(str::trim)
.filter(|value| !value.is_empty())
else {
return prompt.to_string();
};
format!("{prompt}\n避免：{negative_prompt}")
}
fn normalize_image_size(size: &str) -> String {
match size.trim() {
"1024*1024" | "1024x1024" | "1:1" => "1:1",
"1280*720" | "1280x720" | "1600*900" | "1600x900" | "16:9" => "16:9",
value if !value.is_empty() => value,
_ => "1:1",
}
.to_string()
}
async fn wait_openai_generated_images(
http_client: &reqwest::Client,
settings: &OpenAiImageSettings,
task_id: &str,
candidate_count: u32,
failure_context: &str,
) -> Result<OpenAiGeneratedImages, AppError> {
let deadline = Instant::now() + Duration::from_millis(settings.request_timeout_ms);
sleep(Duration::from_secs(10)).await;
while Instant::now() < deadline {
let poll_response = http_client
.get(format!("{}/tasks/{}", settings.base_url, task_id))
.header(
header::AUTHORIZATION,
format!("Bearer {}", settings.api_key),
)
.send()
.await
.map_err(|error| {
map_openai_image_request_error(format!(
"{failure_context}：查询图片生成任务失败：{error}"
))
})?;
let poll_status = poll_response.status();
let poll_text = poll_response.text().await.map_err(|error| {
map_openai_image_request_error(format!(
"{failure_context}：读取图片生成任务响应失败：{error}"
))
})?;
if !poll_status.is_success() {
return Err(map_openai_image_upstream_error(
poll_status.as_u16(),
poll_text.as_str(),
failure_context,
));
}
let poll_json = parse_json_payload(poll_text.as_str(), failure_context)?;
let task_status = find_first_string_by_key(&poll_json.payload, "status")
.or_else(|| find_first_string_by_key(&poll_json.payload, "task_status"))
.unwrap_or_default()
.trim()
.to_ascii_lowercase();
if matches!(task_status.as_str(), "completed" | "succeeded" | "success") {
let image_urls = extract_image_urls(&poll_json.payload);
if image_urls.is_empty() {
let b64_images = extract_b64_images(&poll_json.payload);
if b64_images.is_empty() {
return Err(AppError::from_status(StatusCode::BAD_GATEWAY).with_details(
json!({
"provider": "apimart",
"message": format!("{failure_context}：任务成功但未返回图片"),
}),
));
}
let mut generated =
images_from_base64(task_id.to_string(), b64_images, candidate_count);
generated.actual_prompt =
find_first_string_by_key(&poll_json.payload, "actual_prompt");
return Ok(generated);
}
let mut generated = download_images_from_urls(
http_client,
task_id.to_string(),
image_urls,
candidate_count,
)
.await?;
generated.actual_prompt = find_first_string_by_key(&poll_json.payload, "actual_prompt");
return Ok(generated);
}
if matches!(
task_status.as_str(),
"failed" | "error" | "canceled" | "cancelled" | "unknown"
) {
return Err(map_openai_image_upstream_error(
poll_status.as_u16(),
poll_text.as_str(),
failure_context,
));
}
sleep(Duration::from_secs(3)).await;
}
Err(
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": format!("{failure_context}：图片生成超时或未返回图片地址"),
})),
)
}
async fn download_images_from_urls(
http_client: &reqwest::Client,
task_id: String,
image_urls: Vec<String>,
candidate_count: u32,
) -> Result<OpenAiGeneratedImages, AppError> {
let mut images = Vec::with_capacity(candidate_count.clamp(1, 4) as usize);
for image_url in image_urls
.into_iter()
.take(candidate_count.clamp(1, 4) as usize)
{
images.push(download_remote_image(http_client, image_url.as_str()).await?);
}
Ok(OpenAiGeneratedImages {
task_id,
actual_prompt: None,
images,
})
}
fn images_from_base64(
task_id: String,
b64_images: Vec<String>,
candidate_count: u32,
) -> OpenAiGeneratedImages {
let images = b64_images
.into_iter()
.take(candidate_count.clamp(1, 4) as usize)
.filter_map(|raw| decode_generated_image_base64(raw.as_str()))
.collect();
OpenAiGeneratedImages {
task_id,
actual_prompt: None,
images,
}
}
fn decode_generated_image_base64(raw: &str) -> Option<DownloadedOpenAiImage> {
let bytes = BASE64_STANDARD.decode(raw.trim()).ok()?;
let mime_type = infer_image_mime_type(bytes.as_slice());
Some(DownloadedOpenAiImage {
extension: mime_to_extension(mime_type.as_str()).to_string(),
mime_type,
bytes,
})
}
pub(crate) async fn download_remote_image(
http_client: &reqwest::Client,
image_url: &str,
) -> Result<DownloadedOpenAiImage, AppError> {
let response =
http_client.get(image_url).send().await.map_err(|error| {
map_openai_image_request_error(format!("下载生成图片失败：{error}"))
})?;
let status = response.status();
let content_type = response
.headers()
.get(header::CONTENT_TYPE)
.and_then(|value| value.to_str().ok())
.unwrap_or("image/jpeg")
.to_string();
let body = response.bytes().await.map_err(|error| {
map_openai_image_request_error(format!("读取生成图片内容失败：{error}"))
})?;
if !status.is_success() {
return Err(
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": "下载生成图片失败",
"status": status.as_u16(),
})),
);
}
let normalized_mime_type = normalize_downloaded_image_mime_type(content_type.as_str());
Ok(DownloadedOpenAiImage {
extension: mime_to_extension(normalized_mime_type.as_str()).to_string(),
mime_type: normalized_mime_type,
bytes: body.to_vec(),
})
}
fn parse_json_payload(
raw_text: &str,
failure_context: &str,
) -> Result<ParsedJsonPayload, AppError> {
serde_json::from_str::<Value>(raw_text)
.map(|payload| ParsedJsonPayload { payload })
.map_err(|error| {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": format!("{failure_context}：解析响应失败：{error}"),
"rawExcerpt": truncate_raw(raw_text),
}))
})
}
fn map_openai_image_request_error(message: String) -> AppError {
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": message,
}))
}
fn map_openai_image_upstream_error(
upstream_status: u16,
raw_text: &str,
failure_context: &str,
) -> AppError {
let message = parse_api_error_message(raw_text, failure_context);
tracing::warn!(
provider = "apimart",
upstream_status,
raw_excerpt = %truncate_raw(raw_text),
message,
"APIMart 图片生成上游错误"
);
AppError::from_status(StatusCode::BAD_GATEWAY).with_details(json!({
"provider": "apimart",
"message": message,
"upstreamStatus": upstream_status,
"rawExcerpt": truncate_raw(raw_text),
}))
}
fn parse_api_error_message(raw_text: &str, fallback_message: &str) -> String {
if raw_text.trim().is_empty() {
return fallback_message.to_string();
}
if let Ok(parsed) = serde_json::from_str::<Value>(raw_text) {
for pointer in [
"/error/message",
"/message",
"/output/message",
"/data/message",
] {
if let Some(message) = parsed
.pointer(pointer)
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return message.to_string();
}
}
for pointer in ["/error/code", "/code", "/output/code", "/data/code"] {
if let Some(code) = parsed
.pointer(pointer)
.and_then(Value::as_str)
.map(str::trim)
.filter(|value| !value.is_empty())
{
return format!("{fallback_message}（{code}）");
}
}
}
raw_text.trim().to_string()
}
fn collect_strings_by_key(value: &Value, target_key: &str, results: &mut Vec<String>) {
match value {
Value::Array(entries) => {
for entry in entries {
collect_strings_by_key(entry, target_key, results);
}
}
Value::Object(object) => {
for (key, nested_value) in object {
if key == target_key {
match nested_value {
Value::String(text) => {
let text = text.trim();
if !text.is_empty() {
results.push(text.to_string());
continue;
}
}
Value::Array(entries) => {
for entry in entries {
if let Some(text) = entry
.as_str()
.map(str::trim)
.filter(|value| !value.is_empty())
{
results.push(text.to_string());
}
}
}
_ => {}
}
}
collect_strings_by_key(nested_value, target_key, results);
}
}
_ => {}
}
}
fn find_first_string_by_key(value: &Value, target_key: &str) -> Option<String> {
let mut results = Vec::new();
collect_strings_by_key(value, target_key, &mut results);
results.into_iter().next()
}
fn extract_task_id(payload: &Value) -> Option<String> {
find_first_string_by_key(payload, "task_id")
.or_else(|| find_first_string_by_key(payload, "taskId"))
.or_else(|| find_first_string_by_key(payload, "id"))
}
fn extract_image_urls(payload: &Value) -> Vec<String> {
let mut urls = Vec::new();
collect_strings_by_key(payload, "url", &mut urls);
collect_strings_by_key(payload, "image", &mut urls);
collect_strings_by_key(payload, "image_url", &mut urls);
let mut deduped = Vec::new();
for url in urls {
if (url.starts_with("http://") || url.starts_with("https://")) && !deduped.contains(&url) {
deduped.push(url);
}
}
deduped
}
fn extract_b64_images(payload: &Value) -> Vec<String> {
let mut values = Vec::new();
collect_strings_by_key(payload, "b64_json", &mut values);
values
}
fn normalize_downloaded_image_mime_type(content_type: &str) -> String {
let mime_type = content_type
.split(';')
.next()
.map(str::trim)
.unwrap_or("image/jpeg");
match mime_type {
"image/png" | "image/webp" | "image/jpeg" | "image/jpg" | "image/gif" => {
mime_type.to_string()
}
_ => "image/jpeg".to_string(),
}
}
fn mime_to_extension(mime_type: &str) -> &str {
match mime_type {
"image/png" => "png",
"image/webp" => "webp",
"image/gif" => "gif",
_ => "jpg",
}
}
fn infer_image_mime_type(bytes: &[u8]) -> String {
if bytes.starts_with(b"\x89PNG\r\n\x1A\n") {
return "image/png".to_string();
}
if bytes.starts_with(b"\xFF\xD8\xFF") {
return "image/jpeg".to_string();
}
if bytes.starts_with(b"RIFF") && bytes.get(8..12) == Some(b"WEBP") {
return "image/webp".to_string();
}
if bytes.starts_with(b"GIF87a") || bytes.starts_with(b"GIF89a") {
return "image/gif".to_string();
}
"image/png".to_string()
}
fn truncate_raw(raw_text: &str) -> String {
raw_text.chars().take(800).collect()
}
fn current_utc_micros() -> i64 {
use std::time::{SystemTime, UNIX_EPOCH};
let duration = SystemTime::now()
.duration_since(UNIX_EPOCH)
.expect("system time should be after unix epoch");
i64::try_from(duration.as_micros()).expect("current unix micros should fit in i64")
}
struct ParsedJsonPayload {
payload: Value,
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn gpt_image_2_request_normalizes_legacy_sizes_and_reference_images() {
let body = build_openai_image_request_body(
"雾海神殿",
Some("文字，水印"),
"1280*720",
2,
&["data:image/png;base64,abcd".to_string()],
);
assert_eq!(body["model"], GPT_IMAGE_2_MODEL);
assert_eq!(body["size"], "16:9");
assert_eq!(body["n"], 2);
assert_eq!(body["image_urls"][0], "data:image/png;base64,abcd");
assert!(body["prompt"].as_str().unwrap_or_default().contains("避免"));
}
#[test]
fn b64_json_response_decodes_png_image() {
let images = images_from_base64(
"task-1".to_string(),
vec![BASE64_STANDARD.encode(b"\x89PNG\r\n\x1A\nrest")],
1,
);
assert_eq!(images.images.len(), 1);
assert_eq!(images.images[0].mime_type, "image/png");
assert_eq!(images.images[0].extension, "png");
}
}

8
server-rs/crates/api-server/src/prompt/rpg/foundation_draft.rs
View File

@@ -153,7 +153,11 @@ pub(crate) fn build_custom_world_landmark_seed_batch_prompt(
[
"请根据下面的世界核心信息，批量生成场景框架名单。".to_string(),
if is_opening_batch {
if batch_count == 1 {
"这一步只生成开局场景的场景骨架、默认生图描述、逐幕背景描述、幕 NPC 分配和相连场景信息。".to_string()
} else {
"这一步必须一次性生成开局场景和普通关键场景的场景骨架、默认生图描述、逐幕背景描述、幕 NPC 分配和相连场景信息。".to_string()
}
} else {
"这一步必须一次性生成普通关键场景的场景骨架、默认生图描述、逐幕背景描述、幕 NPC 分配和相连场景信息。".to_string()
},
@@ -162,7 +166,11 @@ pub(crate) fn build_custom_world_landmark_seed_batch_prompt(
build_framework_summary_text(framework, 0),
if story_npc_names.is_empty() { "".to_string() } else { format!("可用场景角色名单：{}", story_npc_names.join("、")) },
if is_opening_batch {
if batch_count == 1 {
"本批场景必须是玩家进入世界时所在的开局场景。".to_string()
} else {
"第一条场景必须是玩家进入世界时所在的开局场景，后续条目才是普通关键场景。".to_string()
}
} else {
"本批只生成普通关键场景，不要再生成开局场景。".to_string()
},

2
src/components/rpg-creation-asset-studio/useRoleVisualCandidateWorkflow.ts
View File

@@ -24,7 +24,7 @@ export function useRoleVisualCandidateWorkflow() {
promptText,
referenceImageDataUrls,
candidateCount: 1,
imageModel: 'wan2.7-image-pro',
imageModel: 'gpt-image-2',
size: '1024*1024',
});
};