diff --git a/docs/【编辑器】画布Agent对话面板-2026-07-03.md b/docs/【编辑器】画布Agent对话面板-2026-07-03.md index 81ea47f96..d86c9d27b 100644 --- a/docs/【编辑器】画布Agent对话面板-2026-07-03.md +++ b/docs/【编辑器】画布Agent对话面板-2026-07-03.md @@ -26,7 +26,7 @@ ## 当前分支落地状态 -- 已落地：会话元数据、OSS 消息文档、会话 CRUD、SSE 消息流、后端 LLM / 启发式工具规划、右侧对话面板、会话历史、新建 / 软删会话、停止当前 SSE 回合、附件从画布资源 / 账号素材库选择，以及五类图片工具对既有生成入口的复用。 +- 已落地：会话元数据、OSS 消息文档、会话 CRUD、SSE 消息流、后端 LLM 工具规划、右侧对话面板、会话历史、新建 / 软删会话、停止当前 SSE 回合、附件从画布资源 / 账号素材库选择，以及五类图片工具对既有生成入口的复用。 - 已落地：`tool_started` / `tool_completed` 事件携带 `status`；工具失败时也会写入失败 generation record，并随后发送 `stage=failed` 与 `error`，前端应保留消息内失败条目。 - 未落地：附件弹窗末尾上传格、跨刷新异步生成恢复、external generation task 轮询回填。未落地前，对话消息状态只表示本次 SSE 回合记录，不作为后台任务队列真相。 @@ -42,7 +42,8 @@ ## 生成结果落画板（对现有占位规则的例外） - 对话入口触发的生成**不创建"即将生成"画布占位**（区别于其余生成面板）；生成中状态由对话消息流承载。 -- 生成完成后：结果图按统一 placement 避让模型（视口中心就近、避开现有图层、32px 间距）落画板为新图层，同时在对话消息内显示缩略图（点击可让画布聚焦到对应图层；如果当前画布尚无该资源图层，前端先刷新工程快照再聚焦）。 +- 生成完成后：结果图按统一 placement 避让模型（视口中心就近、避开现有图层、32px 间距）落画板为新图层，同时登记到默认项目素材库，并在对话消息内显示纯缩略图；前端收到 `generation_result` 后立即刷新工程快照与素材库，缩略图本身不显示名称也不承担图层跳转。 +- 消息内生成结果缩略图必须携带并优先使用 `objectKey` / `assetObjectId`，前端通过 `ResolvedAssetImage` / `/api/assets/read-url` 换签后渲染，不能把裸 `/generated-*` 私有路径直接交给 ``。 - 当前第一阶段通过既有编辑器生成 BFF 的 `canvasCompletion` 写回工程快照；刷新后异步任务恢复和轮询回填属于后续能力，不在本阶段声明为已完成。 - 该例外已同步登记在《生成类面板Lovart统一改造方案-2026-06-17》「画布占位落点」节。 @@ -52,6 +53,7 @@ - 对话框与左侧素材 / 图层侧栏也互斥：打开画布 Agent 时收起左侧栏；再次打开素材、图层或任务侧栏时收起 Agent 面板。 - 桌面端对话框固定宽约 360–400px；移动端抽屉式全宽覆盖；收起态为胶囊/圆形入口按钮。 - 会话管理入口在对话框头部：当前会话标题 + 历史会话下拉（按更新时间倒序）+ 新建对话按钮，全部包在对话框内。 +- 收起对话框只是隐藏面板，不卸载当前会话 hook；流式回复、`生成中` 阶段和停止按钮状态必须在收起 / 重新打开之间保持一致。 ## 附件 @@ -65,7 +67,9 @@ ## LLM 与计费 -- 编排复用 `creative_agent_gpt5_client` 的 LLM 接入配置（同 provider/env，独立用途标识），function-calling 注册五类工具。 +- 编排复用 `creative_agent_gpt5_client` 的 LLM 接入配置（同 provider/env，独立用途标识），但画布 Agent 的轻量规划请求显式使用 `gpt-4o`，避免把简单 JSON 意图识别交给容易长 reasoning 的 gpt-5；function-calling 注册五类工具。 +- 每个用户回合必须由 LLM 返回结构化计划；LLM 未配置、请求失败或返回格式不可解析时，后端写入明确错误消息，不使用本地关键词或“收到：...”回显兜底。 +- 画布 Agent 规划请求使用 Responses API、1024 `max_output_tokens` 和 60 秒 Agent 专用请求超时；生成图片/编辑图片仍走对应生成工具和模型计费。 - **对话回合免费**（聊天、分析回复不扣泥点），仅 Agent 实际触发生成工具时按对应模型定价扣泥点。 - 工具调用前后端校验泥点余额；不足时该次生成失败并在对话中以明确错误气泡告知，对话本身可继续。 @@ -76,7 +80,7 @@ 1. 助手文本 SSE 流式输出； 2. 阶段提示行（思考中 → 思考完成 → 生成中 → 完成/失败）； 3. 工具/模型标注行（生成时显示模型名 + 图标）； -4. 消息内生成结果缩略图（点击 → 画布聚焦对应图层）； +4. 消息内生成结果缩略图（纯预览，不显示名称，不点击聚焦图层）； 5. 生成中的进行中动画； 6. 错误气泡（失败/余额不足，带原因）； 7. 发送中断：进行中时发送按钮变「停止」，可中断当前回合（已提交的生成任务不追回，照常落画板）。 @@ -120,4 +124,5 @@ - 打开画布 Agent 后，任务侧栏和左侧素材 / 图层面板应关闭；再次打开任务侧栏或素材 / 图层面板时，Agent 面板应关闭。 - 发送消息时先本地追加用户消息，再消费 SSE 增量；停止按钮只中断当前 SSE 回合，不追回已经提交的生成工具调用。 -- 点击 Agent 消息内生成结果缩略图时，若当前画布已有该资源图层则直接聚焦；若尚无该资源图层，应先刷新工程快照，再聚焦新图层。 +- Agent 消息内生成结果缩略图只用于预览，不显示名称，也不点击跳转图层；收到 `generation_result` 时统一刷新工程快照和素材库。 +- 对话内容可被用户选中复制；用户从输入框或对话内容点击回画布图层 / 生成器时，焦点应回到画布对象，Backspace / Delete 等画布快捷键继续生效。 diff --git a/packages/shared/src/contracts/editorAgent.ts b/packages/shared/src/contracts/editorAgent.ts index 6d6c5cf7f..3476bfbc8 100644 --- a/packages/shared/src/contracts/editorAgent.ts +++ b/packages/shared/src/contracts/editorAgent.ts @@ -46,6 +46,8 @@ export interface EditorAgentAttachmentRef { export interface EditorAgentGeneratedImage { resourceId: string | null; + objectKey?: string | null; + assetObjectId?: string | null; imageSrc: string; thumbnailSrc: string | null; width: number | null; diff --git a/server-rs/crates/api-server/src/editor_agent.rs b/server-rs/crates/api-server/src/editor_agent.rs index 088b4e68e..4e747e8a7 100644 --- a/server-rs/crates/api-server/src/editor_agent.rs +++ b/server-rs/crates/api-server/src/editor_agent.rs @@ -18,7 +18,7 @@ use module_editor_agent::{ EDITOR_AGENT_MESSAGE_ID_PREFIX, derive_conversation_title, editor_agent_messages_object_key, validate_user_message, }; -use platform_llm::{LlmMessage, LlmMessageContentPart, LlmTextRequest}; +use platform_llm::{LlmError, LlmErrorKind, LlmMessage, LlmMessageContentPart, LlmTextRequest}; use platform_oss::{ LegacyAssetPrefix, OssObjectAccess, OssPutObjectRequest, OssSignedGetObjectUrlRequest, }; @@ -65,6 +65,9 @@ const EDITOR_AGENT_MESSAGES_DOCUMENT_MAX_BYTES: usize = 2 * 1024 * 1024; const EDITOR_AGENT_MESSAGES_READ_EXPIRE_SECONDS: u64 = 60; const EDITOR_AGENT_LLM_ATTACHMENT_URL_EXPIRE_SECONDS: u64 = 300; const EDITOR_AGENT_LLM_MAX_HISTORY_MESSAGES: usize = 12; +const EDITOR_AGENT_LLM_PLANNING_MODEL: &str = "gpt-4o"; +const EDITOR_AGENT_LLM_MAX_OUTPUT_TOKENS: u32 = 1024; +const EDITOR_AGENT_LLM_REQUEST_TIMEOUT_MS: u64 = 60_000; const EDITOR_AGENT_TOOL_CALL_ID_PREFIX: &str = "editor-agent-tool"; const EDITOR_AGENT_CANVAS_RESULT_GAP: f64 = 32.0; type EditorAgentConversationLockMap = Mutex>>>; @@ -279,23 +282,35 @@ pub async fn stream_editor_agent_message( }, )); - let turn_plan = plan_editor_agent_turn( + let mut assistant_status = EditorAgentMessageStatus::Completed; + let mut generation_records = Vec::new(); + let mut planning_failed = false; + + let turn_plan = match plan_editor_agent_turn( &stream_state, &stream_conversation, &document, &stream_user_message, ) - .await; - let mut assistant_text = turn_plan.reply_text; - if assistant_text.trim().is_empty() { - assistant_text = heuristic_editor_agent_turn_plan( - stream_user_message.text.as_str(), - stream_user_message.attachments.as_slice(), - ) - .reply_text; - } - let mut assistant_status = EditorAgentMessageStatus::Completed; - let mut generation_records = Vec::new(); + .await { + Ok(turn_plan) => turn_plan, + Err(error) => { + let app_error = error.into_app_error(); + let error_message = app_error.body_text(); + assistant_status = EditorAgentMessageStatus::Failed; + planning_failed = true; + EditorAgentTurnPlan { + reply_text: error_message, + tool_call: None, + } + } + }; + let assistant_text = turn_plan.reply_text; + let assistant_kind = if assistant_status == EditorAgentMessageStatus::Failed { + EditorAgentMessageKind::Error + } else { + EditorAgentMessageKind::Chat + }; yield Ok::(editor_agent_sse_json_event_or_error( "stage", @@ -310,10 +325,19 @@ pub async fn stream_editor_agent_message( conversation_id: stream_conversation.conversation_id.clone(), message_id: stream_assistant_message_id.clone(), role: EditorAgentMessageRole::Assistant, - kind: EditorAgentMessageKind::Chat, + kind: assistant_kind, text_delta: assistant_text.clone(), }, )); + if planning_failed { + yield Ok::(editor_agent_sse_json_event_or_error( + "stage", + EditorAgentStageEvent { + conversation_id: stream_conversation.conversation_id.clone(), + stage: EditorAgentStage::Failed, + }, + )); + } if let Some(tool_call) = turn_plan.tool_call { let tool_call_id = build_prefixed_uuid_id(EDITOR_AGENT_TOOL_CALL_ID_PREFIX); @@ -951,30 +975,65 @@ struct EditorAgentToolExecutionError { error: AppError, } +#[derive(Debug)] +enum EditorAgentPlanningError { + LlmUnavailable, + LlmRequestFailed(LlmError), + InvalidLlmResponse, +} + +impl EditorAgentPlanningError { + fn into_app_error(self) -> AppError { + match self { + Self::LlmUnavailable => AppError::from_status(StatusCode::SERVICE_UNAVAILABLE) + .with_details(json!({ + "provider": "editor-agent-llm", + "message": "画布 Agent 的 LLM 未配置，无法处理这句话。", + })), + Self::LlmRequestFailed(error) => { + let status = match error.kind() { + LlmErrorKind::Timeout => StatusCode::GATEWAY_TIMEOUT, + LlmErrorKind::InvalidConfig | LlmErrorKind::InvalidRequest => { + StatusCode::SERVICE_UNAVAILABLE + } + LlmErrorKind::Connectivity + | LlmErrorKind::Upstream + | LlmErrorKind::StreamUnavailable + | LlmErrorKind::EmptyResponse + | LlmErrorKind::Transport + | LlmErrorKind::Deserialize => StatusCode::BAD_GATEWAY, + }; + AppError::from_status(status).with_details(json!({ + "provider": "editor-agent-llm", + "message": format!("画布 Agent 调用 LLM 失败：{error}"), + })) + } + Self::InvalidLlmResponse => AppError::from_status(StatusCode::BAD_GATEWAY) + .with_details(json!({ + "provider": "editor-agent-llm", + "message": "画布 Agent 的 LLM 返回格式错误，未能解析回复。", + })), + } + } +} + async fn plan_editor_agent_turn( state: &AppState, conversation: &EditorAgentConversationRecord, document: &EditorAgentConversationMessagesDocument, user_message: &EditorAgentMessage, -) -> EditorAgentTurnPlan { - let fallback = heuristic_editor_agent_turn_plan( - user_message.text.as_str(), - user_message.attachments.as_slice(), - ); +) -> Result { let Some(llm_client) = state.creative_agent_gpt5_client() else { - return fallback; + return Err(EditorAgentPlanningError::LlmUnavailable); }; let request = build_editor_agent_llm_request(state, conversation, document, user_message).await; - match llm_client.request_text(request).await { - Ok(response) => parse_editor_agent_turn_plan( - response.content.as_str(), - fallback.reply_text.as_str(), - user_message.text.as_str(), - ) - .unwrap_or(fallback), - Err(_) => fallback, - } + let response = llm_client + .request_text(request) + .await + .map_err(EditorAgentPlanningError::LlmRequestFailed)?; + parse_editor_agent_turn_plan(response.content.as_str()) + .ok_or(EditorAgentPlanningError::InvalidLlmResponse) } async fn build_editor_agent_llm_request( @@ -997,7 +1056,9 @@ async fn build_editor_agent_llm_request( LlmMessage::user_multimodal(user_parts), ]) .with_responses_api() - .with_max_tokens(900) + .with_model(EDITOR_AGENT_LLM_PLANNING_MODEL) + .with_max_tokens(EDITOR_AGENT_LLM_MAX_OUTPUT_TOKENS) + .with_request_timeout_ms(EDITOR_AGENT_LLM_REQUEST_TIMEOUT_MS) } fn editor_agent_llm_system_prompt() -> String { @@ -1080,34 +1141,30 @@ fn sign_editor_agent_attachment_image_url( .map(|signed| signed.signed_url) } -fn parse_editor_agent_turn_plan( - raw_text: &str, - fallback_reply_text: &str, - fallback_prompt: &str, -) -> Option { +fn parse_editor_agent_turn_plan(raw_text: &str) -> Option { let json_text = extract_editor_agent_json_object(raw_text)?; let raw: EditorAgentRawTurnPlan = serde_json::from_str(json_text).ok()?; - let reply_text = normalize_optional_string(raw.reply_text) - .or_else(|| normalize_optional_string(Some(fallback_reply_text.to_string()))) - .unwrap_or_else(|| build_echo_assistant_text(fallback_prompt, 0)); - let tool_call = raw.tool_call.and_then(|tool| { - let prompt = normalize_optional_string(tool.prompt) - .or_else(|| normalize_optional_string(Some(fallback_prompt.to_string())))?; - Some(EditorAgentToolCallPlan { - tool_name: tool.tool_name, - prompt, - summary: normalize_optional_string(tool.summary), - model: normalize_optional_string(tool.model), - aspect_ratio: normalize_optional_string(tool.aspect_ratio), - image_size: normalize_optional_string(tool.image_size), - icon_descriptions: tool - .icon_descriptions - .unwrap_or_default() - .into_iter() - .filter_map(|description| normalize_optional_string(Some(description))) - .collect(), - }) - }); + let reply_text = normalize_optional_string(raw.reply_text)?; + let tool_call = match raw.tool_call { + Some(tool) => { + let prompt = normalize_optional_string(tool.prompt)?; + Some(EditorAgentToolCallPlan { + tool_name: tool.tool_name, + prompt, + summary: normalize_optional_string(tool.summary), + model: normalize_optional_string(tool.model), + aspect_ratio: normalize_optional_string(tool.aspect_ratio), + image_size: normalize_optional_string(tool.image_size), + icon_descriptions: tool + .icon_descriptions + .unwrap_or_default() + .into_iter() + .filter_map(|description| normalize_optional_string(Some(description))) + .collect(), + }) + } + None => None, + }; Some(EditorAgentTurnPlan { reply_text, @@ -1129,6 +1186,7 @@ fn extract_editor_agent_json_object(raw_text: &str) -> Option<&str> { (start <= end).then_some(&without_fence[start..=end]) } +#[cfg(test)] fn heuristic_editor_agent_turn_plan( text: &str, attachments: &[EditorAgentAttachmentRef], @@ -1202,6 +1260,7 @@ fn heuristic_editor_agent_turn_plan( } } +#[cfg(test)] fn contains_any(text: &str, needles: &[&str]) -> bool { needles.iter().any(|needle| text.contains(needle)) } @@ -1316,7 +1375,7 @@ async fn execute_editor_agent_tool_call( project_id: Some(conversation.project_id.clone()), asset_kind: Some(editor_agent_tool_asset_kind(tool_name).to_string()), generation_inputs, - asset_folder_id: None, + asset_folder_id: Some(editor_agent_default_asset_folder_id()), asset_label: tool_call.summary.clone(), source_resource_id: user_message .attachments @@ -1351,7 +1410,7 @@ async fn execute_editor_agent_tool_call( project_id: Some(conversation.project_id.clone()), asset_kind: Some(editor_agent_tool_asset_kind(tool_name).to_string()), generation_inputs, - asset_folder_id: None, + asset_folder_id: Some(editor_agent_default_asset_folder_id()), asset_label: tool_call.summary.clone(), source_resource_id: user_message .attachments @@ -1390,7 +1449,7 @@ async fn execute_editor_agent_tool_call( image_size: tool_call.image_size.clone(), project_id: Some(conversation.project_id.clone()), generation_inputs, - asset_folder_id: None, + asset_folder_id: Some(editor_agent_default_asset_folder_id()), canvas_completion: Some(completion), }, ) @@ -1508,6 +1567,8 @@ fn editor_agent_generated_images_from_response( .map(|image_src| { vec![EditorAgentGeneratedImage { resource_id: None, + object_key: None, + asset_object_id: None, image_src: image_src.to_string(), thumbnail_src: None, width: data @@ -1535,6 +1596,14 @@ fn editor_agent_generated_image_from_resource( .get("resourceId") .and_then(Value::as_str) .map(str::to_string), + object_key: resource + .get("objectKey") + .and_then(Value::as_str) + .map(str::to_string), + asset_object_id: resource + .get("assetObjectId") + .and_then(Value::as_str) + .map(str::to_string), image_src: resource .get("imageSrc") .and_then(Value::as_str) @@ -1589,6 +1658,10 @@ fn editor_agent_tool_asset_kind(tool_name: EditorAgentToolName) -> &'static str } } +fn editor_agent_default_asset_folder_id() -> String { + "project".to_string() +} + fn editor_agent_tool_default_summary(tool_name: EditorAgentToolName) -> String { match tool_name { EditorAgentToolName::GenerateImage => "生成图片", @@ -1658,6 +1731,7 @@ fn next_editor_agent_canvas_position(layers: Value, _width: f64, _height: f64) - ) } +#[cfg(test)] fn build_echo_assistant_text(text: &str, attachment_count: usize) -> String { let trimmed = text.trim(); if trimmed.is_empty() { @@ -1838,6 +1912,41 @@ mod tests { assert_eq!(normalized.height, Some(128)); } + #[test] + fn generated_image_from_resource_keeps_private_asset_keys() { + let image = editor_agent_generated_image_from_resource( + &json!({ + "resourceId": "resource-1", + "imageSrc": "/generated-editor-assets/result.png", + "thumbnailSrc": "/generated-editor-assets/result-thumb.png", + "objectKey": "generated-editor-assets/result.png", + "assetObjectId": "asset-object-result", + "width": 1024, + "height": 768 + }), + None, + None, + None, + ); + + assert_eq!(image.resource_id.as_deref(), Some("resource-1")); + assert_eq!( + image.object_key.as_deref(), + Some("generated-editor-assets/result.png") + ); + assert_eq!( + image.asset_object_id.as_deref(), + Some("asset-object-result") + ); + assert_eq!(image.width, Some(1024)); + assert_eq!(image.height, Some(768)); + } + + #[test] + fn editor_agent_generations_target_default_asset_library_folder() { + assert_eq!(editor_agent_default_asset_folder_id(), "project"); + } + #[test] fn editor_agent_turn_plan_parses_fenced_json_tool_call() { let plan = parse_editor_agent_turn_plan( @@ -1852,8 +1961,6 @@ mod tests { } } ```"#, - "fallback", - "像素风银发游侠", ) .expect("fenced JSON should parse"); @@ -1865,6 +1972,23 @@ mod tests { assert_eq!(tool_call.image_size.as_deref(), Some("1K")); } + #[test] + fn editor_agent_turn_plan_rejects_invalid_llm_response() { + assert!(parse_editor_agent_turn_plan("收到：画一棵树").is_none()); + assert!( + parse_editor_agent_turn_plan( + r#"{"toolCall":{"toolName":"generate_image","prompt":"画一棵树"}}"#, + ) + .is_none() + ); + assert!( + parse_editor_agent_turn_plan( + r#"{"replyText":"我来生成图片。","toolCall":{"toolName":"generate_image"}}"#, + ) + .is_none() + ); + } + #[test] fn editor_agent_heuristic_routes_tools_from_user_text() { let character_plan = diff --git a/server-rs/crates/api-server/src/llm.rs b/server-rs/crates/api-server/src/llm.rs index 072c7fad9..b312b77a3 100644 --- a/server-rs/crates/api-server/src/llm.rs +++ b/server-rs/crates/api-server/src/llm.rs @@ -44,6 +44,8 @@ pub async fn proxy_llm_chat_completions( max_tokens: None, enable_web_search: false, request_timeout_ms: None, + response_reasoning_effort: None, + response_text_verbosity: None, }; if payload.stream { diff --git a/server-rs/crates/platform-agent/src/apimart_gpt5_adapter.rs b/server-rs/crates/platform-agent/src/apimart_gpt5_adapter.rs index ee528203d..4a1d7389c 100644 --- a/server-rs/crates/platform-agent/src/apimart_gpt5_adapter.rs +++ b/server-rs/crates/platform-agent/src/apimart_gpt5_adapter.rs @@ -55,5 +55,7 @@ pub fn build_gpt5_multimodal_request( request_timeout_ms: None, enable_web_search: false, protocol: LlmTextProtocol::Responses, + response_reasoning_effort: None, + response_text_verbosity: None, } } diff --git a/server-rs/crates/platform-llm/src/lib.rs b/server-rs/crates/platform-llm/src/lib.rs index 0f51347cb..225d046aa 100644 --- a/server-rs/crates/platform-llm/src/lib.rs +++ b/server-rs/crates/platform-llm/src/lib.rs @@ -81,6 +81,8 @@ pub struct LlmTextRequest { pub enable_web_search: bool, pub protocol: LlmTextProtocol, pub request_timeout_ms: Option, + pub response_reasoning_effort: Option, + pub response_text_verbosity: Option, } // 文本协议必须由业务请求显式选择，避免全局默认模型把不同场景混到同一上游形态。 @@ -90,6 +92,40 @@ pub enum LlmTextProtocol { Responses, } +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum LlmResponseReasoningEffort { + Low, + Medium, + High, +} + +impl LlmResponseReasoningEffort { + fn as_str(self) -> &'static str { + match self { + Self::Low => "low", + Self::Medium => "medium", + Self::High => "high", + } + } +} + +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum LlmResponseTextVerbosity { + Low, + Medium, + High, +} + +impl LlmResponseTextVerbosity { + fn as_str(self) -> &'static str { + match self { + Self::Low => "low", + Self::Medium => "medium", + Self::High => "high", + } + } +} + // 上层在流式消费时拿到的是“累计文本 + 当前增量”，避免每层重新自己拼接。 #[derive(Clone, Debug, PartialEq, Eq)] pub struct LlmStreamDelta { @@ -213,6 +249,10 @@ struct ResponsesRequestBody { max_output_tokens: Option, #[serde(skip_serializing_if = "Option::is_none")] tools: Option>, + #[serde(skip_serializing_if = "Option::is_none")] + reasoning: Option, + #[serde(skip_serializing_if = "Option::is_none")] + text: Option, } #[derive(Serialize)] @@ -235,6 +275,16 @@ struct ResponsesWebSearchTool { max_keyword: u8, } +#[derive(Serialize)] +struct ResponsesReasoningOptions { + effort: &'static str, +} + +#[derive(Serialize)] +struct ResponsesTextOptions { + verbosity: &'static str, +} + #[derive(Serialize)] #[serde(rename_all = "camelCase")] struct LlmRawFailureInputLog<'a> { @@ -519,6 +569,8 @@ impl LlmTextRequest { enable_web_search: false, protocol: LlmTextProtocol::ChatCompletions, request_timeout_ms: None, + response_reasoning_effort: None, + response_text_verbosity: None, } } @@ -549,6 +601,16 @@ impl LlmTextRequest { self } + pub fn with_response_reasoning_effort(mut self, effort: LlmResponseReasoningEffort) -> Self { + self.response_reasoning_effort = Some(effort); + self + } + + pub fn with_response_text_verbosity(mut self, verbosity: LlmResponseTextVerbosity) -> Self { + self.response_text_verbosity = Some(verbosity); + self + } + pub fn with_request_timeout_ms(mut self, request_timeout_ms: u64) -> Self { self.request_timeout_ms = Some(request_timeout_ms); self @@ -1178,6 +1240,16 @@ fn build_request_body( max_keyword: 3, }] }), + reasoning: request + .response_reasoning_effort + .map(|effort| ResponsesReasoningOptions { + effort: effort.as_str(), + }), + text: request + .response_text_verbosity + .map(|verbosity| ResponsesTextOptions { + verbosity: verbosity.as_str(), + }), }), } } @@ -2040,6 +2112,8 @@ mod tests { .with_model("deepseek-v3-2-251201") .with_responses_api() .with_web_search(true) + .with_response_reasoning_effort(LlmResponseReasoningEffort::Low) + .with_response_text_verbosity(LlmResponseTextVerbosity::Low) .with_max_tokens(128), ) .await @@ -2074,6 +2148,14 @@ mod tests { request_json["tools"], serde_json::json!([{ "type": "web_search", "max_keyword": 3 }]) ); + assert_eq!( + request_json["reasoning"], + serde_json::json!({ "effort": "low" }) + ); + assert_eq!( + request_json["text"], + serde_json::json!({ "verbosity": "low" }) + ); assert!(request_json.get("official_fallback").is_none()); assert_eq!( request_json["input"][0]["content"][0], diff --git a/server-rs/crates/shared-contracts/src/editor_agent.rs b/server-rs/crates/shared-contracts/src/editor_agent.rs index 7fa5a17d2..de93f755c 100644 --- a/server-rs/crates/shared-contracts/src/editor_agent.rs +++ b/server-rs/crates/shared-contracts/src/editor_agent.rs @@ -92,6 +92,10 @@ pub struct EditorAgentAttachmentRef { pub struct EditorAgentGeneratedImage { #[serde(default)] pub resource_id: Option, + #[serde(default)] + pub object_key: Option, + #[serde(default)] + pub asset_object_id: Option, pub image_src: String, #[serde(default)] pub thumbnail_src: Option, @@ -338,6 +342,8 @@ mod tests { model: Some("z-image".to_string()), images: vec![EditorAgentGeneratedImage { resource_id: Some("resource-2".to_string()), + object_key: Some("generated/result.png".to_string()), + asset_object_id: Some("asset-object-2".to_string()), image_src: "https://example.test/generated.png".to_string(), thumbnail_src: None, width: Some(1024), @@ -354,6 +360,10 @@ mod tests { payload["data"]["images"][0]["resourceId"], json!("resource-2") ); + assert_eq!( + payload["data"]["images"][0]["objectKey"], + json!("generated/result.png") + ); let error_envelope = EditorAgentSseEnvelope { event: EditorAgentSseEventType::Error, diff --git a/src/components/image-editor/EditorAgentConversationPanelView.test.tsx b/src/components/image-editor/EditorAgentConversationPanelView.test.tsx index 6bf4024d4..c6b7dc8b1 100644 --- a/src/components/image-editor/EditorAgentConversationPanelView.test.tsx +++ b/src/components/image-editor/EditorAgentConversationPanelView.test.tsx @@ -1,6 +1,7 @@ /* @vitest-environment jsdom */ import { + act, fireEvent, render, screen, @@ -253,6 +254,135 @@ describe('EditorAgentConversationPanelView', () => { }); }); + it('keeps streamed reply and generating state when the panel is collapsed and reopened', async () => { + const client = createClient(); + let finishStream = () => {}; + vi.mocked(client.streamMessage).mockImplementation( + async (conversationId, _payload, options) => { + options.onEvent?.({ + event: 'message_delta', + data: { + conversationId, + messageId: 'assistant-stream', + role: 'assistant', + kind: 'chat', + textDelta: '我来生成图片。', + }, + }); + options.onEvent?.({ + event: 'stage', + data: { conversationId, stage: 'generating' }, + }); + options.onEvent?.({ + event: 'tool_started', + data: { + conversationId, + messageId: 'assistant-stream', + toolCallId: 'tool-call-generating', + toolName: 'generate_image', + taskId: 'task-generating', + model: 'gpt-image-2', + }, + }); + await new Promise((resolve) => { + finishStream = resolve; + }); + }, + ); + + const { rerender } = render( + , + ); + + await waitFor(() => { + expect(screen.getByText('已经看到画布内容')).toBeTruthy(); + }); + + fireEvent.change(screen.getByLabelText('发送给画布 Agent'), { + target: { value: '生成一张图片' }, + }); + fireEvent.click(screen.getByRole('button', { name: '发送' })); + + await waitFor(() => { + expect(screen.getByText('我来生成图片。')).toBeTruthy(); + }); + expect(screen.getByText('生成中')).toBeTruthy(); + expect(screen.getByRole('button', { name: '停止' })).toBeTruthy(); + + rerender( + , + ); + expect(screen.getByRole('button', { name: '打开画布 Agent' })).toBeTruthy(); + + rerender( + , + ); + + expect(screen.getByText('我来生成图片。')).toBeTruthy(); + expect(screen.getByText('生成中')).toBeTruthy(); + expect(screen.getByRole('button', { name: '停止' })).toBeTruthy(); + + await act(async () => { + finishStream(); + }); + }); + + it('keeps wheel scrolling inside the message history and input', async () => { + const client = createClient(); + const parentWheel = vi.fn(); + + render( +
+ +
, + ); + + await waitFor(() => { + expect(screen.getByText('已经看到画布内容')).toBeTruthy(); + }); + + const messageWheel = new WheelEvent('wheel', { + bubbles: true, + cancelable: true, + deltaY: 120, + }); + fireEvent( + screen.getByRole('log', { name: '画布 Agent 消息流' }), + messageWheel, + ); + expect(parentWheel).not.toHaveBeenCalled(); + expect(messageWheel.defaultPrevented).toBe(false); + + const inputWheel = new WheelEvent('wheel', { + bubbles: true, + cancelable: true, + deltaY: 120, + }); + fireEvent(screen.getByLabelText('发送给画布 Agent'), inputWheel); + expect(parentWheel).not.toHaveBeenCalled(); + expect(inputWheel.defaultPrevented).toBe(false); + }); + it('confirms conversation deletion from an independent dialog', async () => { const client = createClient(); render( diff --git a/src/components/image-editor/EditorAgentConversationPanelView.tsx b/src/components/image-editor/EditorAgentConversationPanelView.tsx index 48e288f13..db719bbc6 100644 --- a/src/components/image-editor/EditorAgentConversationPanelView.tsx +++ b/src/components/image-editor/EditorAgentConversationPanelView.tsx @@ -10,11 +10,18 @@ import { Trash2, X, } from 'lucide-react'; -import { type FormEvent, useMemo, useState } from 'react'; +import { + type FormEvent, + type WheelEvent as ReactWheelEvent, + useEffect, + useMemo, + useState, +} from 'react'; import { EDITOR_AGENT_MAX_ATTACHMENTS, type EditorAgentAttachmentRef, + type EditorAgentGenerationResultEvent, type EditorAgentGenerationRecord, type EditorAgentMessage, type EditorAgentStage, @@ -22,6 +29,7 @@ import { import { PlatformActionButton } from '../common/PlatformActionButton'; import { PlatformDangerConfirmDialog } from '../common/PlatformDangerConfirmDialog'; import { UnifiedModal } from '../common/UnifiedModal'; +import { ResolvedAssetImage } from '../ResolvedAssetImage'; import type { CanvasLayer, EditorAsset } from './ImageCanvasEditorTypes'; import { type EditorAgentConversationClient, @@ -42,10 +50,14 @@ type EditorAgentConversationPanelViewProps = { onToggleOpen: () => void; layers?: CanvasLayer[]; assets?: EditorAsset[]; - onFocusResource?: (resourceId: string) => void; + onGenerationResult?: (event: EditorAgentGenerationResultEvent) => void; client?: EditorAgentConversationClient; }; +function stopAgentPanelWheel(event: ReactWheelEvent) { + event.stopPropagation(); +} + function attachmentKey(attachment: EditorAgentAttachmentRef) { return `${attachment.source}:${attachment.referenceId}`; } @@ -169,8 +181,10 @@ function AttachmentChip({ ) : null} {attachment.thumbnailSrc || attachment.imageSrc ? ( - @@ -182,10 +196,8 @@ function AttachmentChip({ function GenerationRecordsView({ generations, - onFocusResource, }: { generations: EditorAgentGenerationRecord[]; - onFocusResource?: (resourceId: string) => void; }) { if (!generations.length) { return null; @@ -197,42 +209,42 @@ function GenerationRecordsView({ key={generation.toolCallId} className="rounded-2xl border border-slate-200 bg-white/80 p-2 text-xs text-slate-600" > -
- {generation.status === 'generating' ? ( -
+ {generation.status === 'generating' || generation.error ? ( +
+ {generation.status === 'generating' ? ( +
+ ) : null} {generation.error ? (
{generation.error}
) : null} {generation.images.length ? (
{generation.images.map((image, index) => ( - +
))} ) : null} @@ -244,10 +256,8 @@ function GenerationRecordsView({ function MessageBubble({ message, - onFocusResource, }: { message: EditorAgentMessage; - onFocusResource?: (resourceId: string) => void; }) { const isUser = message.role === 'user'; return ( @@ -279,7 +289,6 @@ function MessageBubble({ ) : null} @@ -367,10 +376,12 @@ function AttachmentPickerModal({ key={option.key} className="flex cursor-pointer flex-col gap-2 rounded-2xl border border-slate-200 bg-white p-2 text-sm text-slate-700 shadow-sm" > - @@ -398,10 +409,16 @@ export function EditorAgentConversationPanelView({ onToggleOpen, layers = [], assets = [], - onFocusResource, + onGenerationResult, client, }: EditorAgentConversationPanelViewProps) { - const effectiveProjectId = open ? projectId : null; + const [hasConversationMounted, setHasConversationMounted] = useState(open); + useEffect(() => { + if (open) { + setHasConversationMounted(true); + } + }, [open]); + const effectiveProjectId = hasConversationMounted ? projectId : null; const { conversations, activeConversationId, @@ -418,7 +435,11 @@ export function EditorAgentConversationPanelView({ sendMessage, stopCurrentTurn, deleteActiveConversation, - } = useEditorAgentConversation({ projectId: effectiveProjectId, client }); + } = useEditorAgentConversation({ + projectId: effectiveProjectId, + client, + onGenerationResult, + }); const [draftText, setDraftText] = useState(''); const [attachments, setAttachments] = useState( [], @@ -517,7 +538,7 @@ export function EditorAgentConversationPanelView({ return (