diff --git a/app/ai.py b/app/ai.py index daaf899..3179161 100644 --- a/app/ai.py +++ b/app/ai.py @@ -136,12 +136,16 @@ def validate_direct_generation_config( errors = [] for category, label in required: try: - _role_model( + model = _role_model( category, ai_cfg.get("default_%s_model" % category), models_path, models=models, ) + if category == "image": + compatibility_error = appconfig.image_model_config_error(model) + if compatibility_error: + raise AIError(compatibility_error) except Exception as exc: errors.append("%s模型%s" % (label, str(exc))) if errors: @@ -345,45 +349,28 @@ def gen_cover( quality = _jpg_quality(jpg_quality if jpg_quality is not None else ai_cfg.get("jpg_quality", 90)) attempts = _attempt_count(ai_cfg, retry) - api_type = model.get("api_type", "auto") _notify_step(on_step, "cover_build_request") - if api_type == "images_edits": - body, content_type = _image_edit_body(model, cover_prompt, old_cover_path, resolution) - _notify_step(on_step, "cover_request") - data = _call_with_retry( + compatibility_error = appconfig.image_model_config_error(model) + if compatibility_error: + raise AIError(compatibility_error) + body, content_type = _image_edit_body(model, cover_prompt, [old_cover_path], resolution) + _notify_step(on_step, "cover_request") + data = _call_with_retry( + model, + body, + cfg, + attempts, + request_kind="multipart", + content_type=content_type, + on_retry=lambda attempt, total_attempts, exc: _notify_retry( + on_step, + "cover_request", + attempt, + total_attempts, + exc, model, - body, - cfg, - attempts, - request_kind="multipart", - content_type=content_type, - on_retry=lambda attempt, total_attempts, exc: _notify_retry( - on_step, - "cover_request", - attempt, - total_attempts, - exc, - model, - ), - ) - else: - payload = _image_chat_payload(model, cover_prompt, old_cover_path, resolution) - _notify_step(on_step, "cover_request") - data = _call_with_retry( - model, - payload, - cfg, - attempts, - request_kind="json", - on_retry=lambda attempt, total_attempts, exc: _notify_retry( - on_step, - "cover_request", - attempt, - total_attempts, - exc, - model, - ), - ) + ), + ) _notify_step(on_step, "cover_parse_response") image_bytes = _extract_image_bytes(data, model, cfg) @@ -1624,7 +1611,7 @@ def _debug_cmhub_image_url_enabled(): def _extract_cmhub_image_url(data, base_url): - candidate = _find_image_ref(data) + candidate = _find_cmhub_image_ref(data) if not candidate: return "" return _normalize_cmhub_image_url(candidate, base_url) @@ -2554,32 +2541,34 @@ def _chat_payload(model, messages): return payload -def _image_chat_payload(model, cover_prompt, old_cover_path, resolution): - prompt = "%s\n\n目标分辨率：%s。" % (str(cover_prompt or "").strip(), resolution) - content = [ - {"type": "text", "text": prompt.strip()}, - { - "type": "image_url", - "image_url": {"url": _image_data_url(old_cover_path)}, - }, - ] - return _chat_payload(model, [{"role": "user", "content": content}]) - - -def _image_edit_body(model, cover_prompt, old_cover_path, resolution): - fields = { +def _image_edit_body(model, cover_prompt, image_paths, resolution): + if isinstance(image_paths, (str, bytes, os.PathLike)): + image_paths = [image_paths] + image_paths = list(image_paths or []) + if not image_paths: + raise AIError("图片编辑请求至少需要一张本地参考图") + fields = copy.deepcopy(model.get("extra_body", {})) + fields.update({ "model": model["model"], "prompt": str(cover_prompt or ""), "size": _resolution_size_text(resolution), - } - fields.update(copy.deepcopy(model.get("extra_body", {}))) - files = { - "image": ( - os.path.basename(old_cover_path), - open(old_cover_path, "rb").read(), - mimetypes.guess_type(old_cover_path)[0] or "application/octet-stream", + "n": "1", + }) + files = [] + for image_path in image_paths: + path = os.fspath(image_path) + with open(path, "rb") as fh: + data = fh.read() + files.append( + ( + "image[]", + ( + os.path.basename(path), + data, + mimetypes.guess_type(path)[0] or "application/octet-stream", + ), + ) ) - } return _multipart_body(fields, files) @@ -2595,7 +2584,8 @@ def _multipart_body(fields, files): b"\r\n", ] ) - for name, file_info in files.items(): + file_items = files.items() if isinstance(files, dict) else files + for name, file_info in file_items: filename, data, content_type = file_info chunks.extend( [ @@ -2656,15 +2646,38 @@ def _content_text(content): def _extract_image_bytes(data, model, config): image_ref = _find_image_ref(data) if not image_ref: - raise AIError("AI 返回中没有图片数据") + raise AIError("AI 图片响应不符合 OpenAI 图片编辑接口") if image_ref.startswith("data:"): return _decode_data_url(image_ref) if _looks_base64(image_ref): return base64.b64decode(image_ref) + parts = urllib.parse.urlsplit(image_ref) + if parts.scheme not in {"http", "https"} or not parts.netloc: + raise AIError("AI 图片地址只允许 http/https") return _download_image(image_ref, model, config) def _find_image_ref(value): + """Read only the fixed OpenAI Images API response fields for direct calls.""" + + if not isinstance(value, dict): + return None + data = value.get("data") + if not isinstance(data, list): + return None + for item in data: + if not isinstance(item, dict): + continue + for key in ("b64_json", "url"): + candidate = item.get(key) + if isinstance(candidate, str) and candidate.strip(): + return candidate.strip() + return None + + +def _find_cmhub_image_ref(value): + """Read the default gateway's documented and legacy image response shapes.""" + if isinstance(value, dict): for key in ("b64_json", "base64", "image_base64", "image", "url"): candidate = value.get(key) @@ -2678,7 +2691,7 @@ def _find_image_ref(value): if isinstance(candidate, str): return candidate if isinstance(image_url, list): - candidate = _find_image_ref_from_list(image_url) + candidate = _find_cmhub_image_ref_from_list(image_url) if candidate: return candidate for key in ("image_urls", "urls"): @@ -2686,28 +2699,28 @@ def _find_image_ref(value): if isinstance(candidate, str) and candidate.strip(): return candidate.strip() if isinstance(candidate, list): - found = _find_image_ref_from_list(candidate) + found = _find_cmhub_image_ref_from_list(candidate) if found: return found for key in ("result", "data", "choices", "output", "content", "images", "files"): - candidate = _find_image_ref(value.get(key)) + candidate = _find_cmhub_image_ref(value.get(key)) if candidate: return candidate message = value.get("message") if message is not None: - candidate = _find_image_ref(message) + candidate = _find_cmhub_image_ref(message) if candidate: return candidate elif isinstance(value, list): - return _find_image_ref_from_list(value) + return _find_cmhub_image_ref_from_list(value) return None -def _find_image_ref_from_list(values): +def _find_cmhub_image_ref_from_list(values): for item in values: if isinstance(item, str) and item.strip(): return item.strip() - candidate = _find_image_ref(item) + candidate = _find_cmhub_image_ref(item) if candidate: return candidate return None diff --git a/app/appconfig.py b/app/appconfig.py index 764c674..67a73ae 100644 --- a/app/appconfig.py +++ b/app/appconfig.py @@ -142,10 +142,10 @@ DEFAULT_AI_MODELS_CONFIG = { "name": "Nano Banana 2", "category": "image", "enabled": True, - "url": "https://api.vectorengine.ai/v1/chat/completions", + "url": "https://api.vectorengine.ai/v1", "model": "gemini-3.1-flash-image-preview", "api_key": "", - "api_type": "auto", + "api_type": "images_edits", "connect_timeout_seconds": 30, "timeout_seconds": 0, "extra_body": {}, @@ -1018,6 +1018,56 @@ def get_model(name, path=AI_MODELS_PATH) -> dict: return copy.deepcopy(models[_model_index(models, name)]) +def is_image_edit_model(model) -> bool: + """Return whether a model satisfies the direct image-edit contract.""" + + return ( + isinstance(model, dict) + and model.get("category") == "image" + and model.get("api_type") == "images_edits" + ) + + +def image_model_config_error(model) -> str: + """Return a Chinese actionable error for a direct image model, if any.""" + + if not isinstance(model, dict) or model.get("category") != "image": + return "当前模型不是图片模型" + if model.get("api_type") != "images_edits": + return "图片模型仅支持 OpenAI 图片编辑接口，请在设置中选择该接口类型" + missing = [ + field + for field in ("url", "model", "api_key") + if not str(model.get(field, "") or "").strip() + ] + if missing: + return "图片模型缺少必要配置：" + "、".join(missing) + url = str(model.get("url") or "").strip() + parts = urllib.parse.urlsplit(url) + if parts.scheme not in {"http", "https"} or not parts.netloc: + return "图片模型网址必须使用 http 或 https" + try: + connect_timeout = int(model.get("connect_timeout_seconds", 0) or 0) + except (TypeError, ValueError): + connect_timeout = 0 + if connect_timeout <= 0: + return "图片模型连接超时必须大于 0 秒" + return "" + + +def check_image_model_config(name, path=AI_MODELS_PATH) -> dict: + """Validate one image model locally without making a billable request.""" + + try: + model = get_model(name, path=path) + error = image_model_config_error(model) + if error: + return {"ok": False, "check_only": True, "error": error} + return {"ok": True, "check_only": True} + except Exception as exc: + return {"ok": False, "check_only": True, "error": str(exc)} + + def model_request_url(model) -> str: """Return the HTTP endpoint used for a configured AI model.""" diff --git a/app/gui/tabs/settings.py b/app/gui/tabs/settings.py index d094a3f..0ed862f 100644 --- a/app/gui/tabs/settings.py +++ b/app/gui/tabs/settings.py @@ -36,6 +36,7 @@ class SettingsTab(QWidget): BACKEND_ITEMS = [("默认网关", "cmhub"), ("自定义网关", "direct")] CATEGORY_ITEMS = [("文本", "text"), ("图像", "image")] API_TYPE_ITEMS = [("chat", "chat"), ("images_edits", "images_edits"), ("auto", "auto")] + IMAGE_API_TYPE_ITEMS = [("OpenAI 图片编辑接口", "images_edits")] RESOLUTION_ITEMS = ["512", "1k", "2k", "4k"] def __init__( @@ -431,6 +432,7 @@ class SettingsTab(QWidget): self.delete_model_button.clicked.connect(self.delete_model) self.save_model_button.clicked.connect(self.save_model) self.test_connection_button.clicked.connect(self.test_connection) + self.category_combo.currentIndexChanged.connect(self._on_model_category_changed) self.gateway_default_button.toggled.connect(self._on_gateway_source_toggled) self.gateway_custom_button.toggled.connect(self._on_gateway_source_toggled) self.cmhub_refresh_button.clicked.connect(self.refresh_cmhub_models) @@ -667,6 +669,8 @@ class SettingsTab(QWidget): label = f"{model['name']} · {self._category_label(model['category'])}" if not model.get("enabled", True): label += " · 已停用" + if model.get("category") == "image" and not appconfig.is_image_edit_model(model): + label += " · 当前图片模型不支持 OpenAI 图片编辑接口" self.model_combo.addItem(label, model["name"]) index = self.model_combo.findData(current) self.model_combo.setCurrentIndex(index if index >= 0 else (0 if self.models else -1)) @@ -765,6 +769,7 @@ class SettingsTab(QWidget): ai_models_path=self.ai_models_path, db_path=_database_path(config=self.config), diagnostic_log_dir=diagnostics.DEFAULT_LOG_DIR, + check_image_config=model.get("category") == "image", ) worker.finished.connect(self._on_test_finished) worker.failed.connect(self._on_test_failed) @@ -773,8 +778,12 @@ class SettingsTab(QWidget): self.test_worker = worker self.test_thread = thread self._set_test_running(True) - self.test_result_label.setText("正在测试连接...") - self._set_status(f"正在测试 AI 模型连接：{model['name']}") + if model.get("category") == "image": + self.test_result_label.setText("正在检查图片配置...") + self._set_status(f"正在检查图片模型配置：{model['name']}") + else: + self.test_result_label.setText("正在测试连接...") + self._set_status(f"正在测试 AI 模型连接：{model['name']}") thread.start() def save_app_settings(self, checked=False): @@ -1010,7 +1019,10 @@ class SettingsTab(QWidget): combo.clear() for model in self.models: if model.get("category") == category and model.get("enabled", True): - combo.addItem(model.get("name", ""), model.get("name", "")) + label = model.get("name", "") + if category == "image" and not appconfig.is_image_edit_model(model): + label += " · 当前图片模型不支持 OpenAI 图片编辑接口" + combo.addItem(label, model.get("name", "")) if combo.count() == 0: combo.addItem("无可用模型", None) index = combo.findData(selected) @@ -1058,6 +1070,32 @@ class SettingsTab(QWidget): "extra_body": dict(extra_body), } + def _on_model_category_changed(self, index=None): + category = self.category_combo.currentData() or "text" + current_type = self.api_type_combo.currentData() + if category == "image": + self._set_api_type_options(category, selected="images_edits") + else: + self._set_api_type_options(category, selected=current_type or "chat") + self._update_button_state() + + def _set_api_type_options(self, category, selected=None): + """Render category-specific API choices without mutating legacy models.""" + + self.api_type_combo.blockSignals(True) + self.api_type_combo.clear() + if category == "image": + # Keep a stored legacy choice visible so users can correct it explicitly. + if selected and selected != "images_edits": + self.api_type_combo.addItem("当前不支持（%s）" % selected, selected) + items = self.IMAGE_API_TYPE_ITEMS + else: + items = self.API_TYPE_ITEMS + for label, value in items: + self.api_type_combo.addItem(label, value) + self._set_combo_by_data(self.api_type_combo, selected) + self.api_type_combo.blockSignals(False) + def _populate_form(self, model): widgets = [ self.enabled_checkbox, @@ -1075,7 +1113,7 @@ class SettingsTab(QWidget): self.enabled_checkbox.setChecked(False) self.name_edit.clear() self.category_combo.setCurrentIndex(0) - self.api_type_combo.setCurrentIndex(0) + self._set_api_type_options("text", selected="chat") self.model_id_edit.clear() self.url_edit.clear() self.api_key_edit.clear() @@ -1084,7 +1122,10 @@ class SettingsTab(QWidget): self.enabled_checkbox.setChecked(bool(model.get("enabled", True))) self.name_edit.setText(model.get("name", "")) self._set_combo_by_data(self.category_combo, model.get("category", "text")) - self._set_combo_by_data(self.api_type_combo, model.get("api_type", "auto")) + self._set_api_type_options( + model.get("category", "text"), + selected=model.get("api_type", "auto"), + ) self.model_id_edit.setText(model.get("model", "")) self.url_edit.setText(model.get("url", "")) self.api_key_edit.setText(model.get("api_key", "")) @@ -1118,6 +1159,13 @@ class SettingsTab(QWidget): has_model and not testing and self._can_delete_model(self._current_model()) ) self.test_connection_button.setEnabled(has_model and not testing) + if not testing: + model = self._current_model() + self.test_connection_button.setText( + "检查图片配置" + if model is not None and model.get("category") == "image" + else "测试连接" + ) def _set_test_running(self, running): self._update_button_state() @@ -1132,7 +1180,11 @@ class SettingsTab(QWidget): self._set_test_running(False) def _on_test_finished(self, payload): - if payload.get("ok"): + if payload.get("check_only") and payload.get("ok"): + message = f"图片配置检查通过：{payload.get('name')}" + elif payload.get("check_only"): + message = "图片配置检查失败：%s" % (payload.get("error") or "配置不完整") + elif payload.get("ok"): status = payload.get("status") suffix = f"（HTTP {status}）" if status else "" message = f"测试连接成功：{payload.get('name')}{suffix}" diff --git a/app/gui/workers.py b/app/gui/workers.py index 365fcb1..8f7a37a 100644 --- a/app/gui/workers.py +++ b/app/gui/workers.py @@ -3261,24 +3261,41 @@ class CMHubSettingsWorker(BaseWorker): return _elapsed_ms(started) class AIModelTestWorker(BaseWorker): - """Test one AI model connection without blocking the GUI thread.""" + """Test text models or validate image-model configuration off the GUI thread.""" - def __init__(self, model_name, ai_models_path=None, db_path=None, diagnostic_log_dir=None): + def __init__( + self, + model_name, + ai_models_path=None, + db_path=None, + diagnostic_log_dir=None, + check_image_config=False, + ): super().__init__() self.model_name = model_name self.ai_models_path = ai_models_path or appconfig.AI_MODELS_PATH self.db_path = db_path self.diagnostic_log_dir = diagnostic_log_dir + self.check_image_config = bool(check_image_config) self._run_id = None def execute(self): self._run_id = self._create_run_log() started = time.monotonic() self._log_run_event( - f"step=test_connection result=start detail=AI模型 {self.model_name}" + "step={step} result=start detail=AI模型 {name}".format( + step="check_image_config" if self.check_image_config else "test_connection", + name=self.model_name, + ) ) try: - result = appconfig.test_ai_model(self.model_name, path=self.ai_models_path) + if self.check_image_config: + result = appconfig.check_image_model_config( + self.model_name, + path=self.ai_models_path, + ) + else: + result = appconfig.test_ai_model(self.model_name, path=self.ai_models_path) except Exception as exc: error = diagnostics.redact_log_text(str(exc) or exc.__class__.__name__) elapsed_ms = self._elapsed_ms(started) diff --git a/docs/api.md b/docs/api.md index e9a81b3..87fa090 100644 --- a/docs/api.md +++ b/docs/api.md @@ -73,6 +73,8 @@ update_ai_model(name, **fields) -> None delete_ai_model(name) -> None # 至少各留一个 text+image；删到剩一禁用 model_request_url(model) -> str # url 为 /v1 或 /api/v1 base 时按 api_type 补 /chat/completions 或 /images/edits test_ai_model(name) -> dict # 「测试连接」：用 key/url/model 发最小请求 -> {ok, status?, error?} +check_image_model_config(name) -> dict # 图片模型本地校验，不发网络/计费请求 -> {ok, check_only, error?} +is_image_edit_model(model) -> bool # category=image 且 api_type=images_edits get_model(name) -> dict # 返回模型定义，含 api_key（调用方不得写日志） ``` @@ -330,7 +332,7 @@ fetch_cmhub_models(base_url, api_key, connect_timeout=10, read_timeout=30) -> li 要点： -- `backend=direct`：设置页「自定义网关」的客户端直连兼容路径；标题用 `default_text_model`、封面用 `default_image_model`（`appconfig.get_model` 取定义，含 url/key/api_type）。②在创建 worker 前按本轮生成内容校验所需模型；运行中使用冻结的内存模型快照，不重读 `ai_models.json`。它不支持图片理解或商品套图新建任务。 +- `backend=direct`：设置页「自定义网关」的客户端直连兼容路径；标题用 `default_text_model`，保留 `chat` / `auto`；封面用 `default_image_model`，只接受 `category=image` 且 `api_type=images_edits`（`appconfig.get_model` 取定义，含 url/key/api_type）。封面固定 `POST /v1/images/edits`，multipart 中按顺序重复 `image[]`，首项是主体图，固定 `n=1`，只读取 `data[].b64_json` 或 `data[].url`，仅允许 data URL/base64/HTTP(S) 下载结果。②在创建 worker 前按本轮生成内容校验所需模型；运行中使用冻结的内存模型快照，不重读 `ai_models.json`。历史 `chat` / `auto` 图片模型仍可保存和查看，但不可提交图片生成。它不支持图片理解或商品套图新建任务。 - `backend=cmhub`：设置页「默认网关」及普通产品默认路径；标题调用 `POST /api/v1/generate/title`；②批量封面生成调用 `POST /api/v1/generate/image/tasks` + `GET /api/v1/generate/image/tasks/{task_id}`，模型字段使用 `ai.cmhub.title_alias/image_alias`，Key 来自 `data/config/cmhub.json`。商品套图中的「AI帮写」单独调用 `POST /api/v1/analyze/images`，只使用 `ai.cmhub.vision_alias`，不得回退或混用生文/生图别名；同一商品项目的原图在一条请求内联合理解，用户勾选状态不参与选图。`gen_cover()` 单独调用没有任务/DB 上下文，第一版保留旧同步 `POST /api/v1/generate/image` 兼容路径。 - 标题提示词组装：`gen_title()` 的 direct 与 cmhub 路径共用标题 prompt 规则。若标题提示词包含 `{旧标题}`，生成前替换为该任务旧标题，不再自动追加旧标题块；若不包含 `{旧标题}`，保持旧行为自动追加“旧标题：...”块。两种情况都会追加“请只返回新标题，不要解释。”输出约束；其它 `{...}` 原样保留。 - `fetch_cmhub_models()` 调 `GET /api/v1/models` 返回别名清单，供设置页动态下拉、商品套图 AI帮写和商品套图正式生成的付费前确认共用；Base URL 会先规整为网关根，HTTP 404 映射为 `not_found` 并提示检查 Base URL 或实例是否部署 `/api/v1/models`。GUI 只在进程内按规整网关地址和别名短期缓存公开模型元数据，不缓存 API Key，也不写入配置、SQLite、日志或导出。AI帮写仅在当前 `vision_alias` 命中 `operation_type=vision`、`requires_image=true`、`pricing_status=priced` 且有唯一无条件 `points_cost` 时显示预计扣点；正式套图同样只在当前 `image_alias` 命中唯一无条件的图片生图价格时显示单张和按最终 `specs` 计算的总价。其它情况只提示实际以网关返回为准，预估值不参与扣减或成功判定。 @@ -505,7 +507,7 @@ T-523 后 GUI 已从旧 `app/gui.py` 拆为 `app/gui/` 包：`__init__.py` 负 设置当前要点（T-501）： - `SettingsTab` 使用居中内容区 + 适度左右留白布局，当前留白已从 T-506 初始实现缩短到约 40%；实现上使用最大内容宽度和自适应 margin，避免固定像素导致小屏挤压。各设置组默认响应式 3 列表单：短字段占 1 格，长字段（URL/API Key/路径）跨 2 格或 3 格，窄窗口降为 2 列/1 列。点击「保存设置」成功后，调用 `QMessageBox.information` 弹出“设置已保存”轻量提示框，同时保留状态栏提示。T-531 已完成：`save_app_settings()` 返回 bool，成功写 `data/config.json` + `data/config/cmhub.json` 后清 dirty，失败保留 dirty 并让调用方阻止离开。 -- `SettingsTab` 的生成网关配置：用 `gatewayDefaultButton` / `gatewayCustomButton` 分段选择器表达 `cmhub` / `direct` 来源，选择仅标记待保存，保存后才写 `ai.backend`。默认网关 Base URL 保存/刷新前规整为网关根，API Key 单独读写 `data/config/cmhub.json`，别名下拉来自 `fetch_cmhub_models()`；自定义网关复用 `data/config/ai_models.json` 的模型 CRUD 与文本/图片角色选择。两套配置并存，模型 CRUD 立即保存；`refresh_cmhub_models()` / `test_cmhub_connection()` 使用输入框实时值但不得自动保存。`settingsSaved` 通知主窗口刷新⑥能力状态；`is_dirty()` / `discard_unsaved_changes()` / `_suspend_dirty`（或等价机制）用于保存和放弃来源选择。 +- `SettingsTab` 的生成网关配置：用 `gatewayDefaultButton` / `gatewayCustomButton` 分段选择器表达 `cmhub` / `direct` 来源，选择仅标记待保存，保存后才写 `ai.backend`。默认网关 Base URL 保存/刷新前规整为网关根，API Key 单独读写 `data/config/cmhub.json`，别名下拉来自 `fetch_cmhub_models()`；自定义网关复用 `data/config/ai_models.json` 的模型 CRUD 与文本/图片角色选择。图片模型只展示「OpenAI 图片编辑接口」；历史不兼容配置保留并标示。图片模型的「检查图片配置」只做本地校验，不发送网络图片请求。两套配置并存，模型 CRUD 立即保存；`refresh_cmhub_models()` / `test_cmhub_connection()` 使用输入框实时值但不得自动保存。`settingsSaved` 通知主窗口刷新⑥能力状态；`is_dirty()` / `discard_unsaved_changes()` / `_suspend_dirty`（或等价机制）用于保存和放弃来源选择。 - T-532 要求 `SettingsTab._on_cmhub_finished()` 从 worker payload 的 `balance` / user/account 字段提取 cmhub 账号身份，成功文案优先显示 `cmhub 账号「<账号名>」连接成功：...`；当前 `/balance` 结构兼容 `{ "user": "cmhub_user", "points_balance": 88, "account": { "username": "cmhub_user", "display_name": "主账号" } }`，显示名优先 `account.display_name`，再兜底 `account.username` / `user` / 顶层常见字段；账号字段缺失时保持 `cmhub 连接成功：...`。显示名必须脱敏处理邮箱，且不得把 API Key、token 或完整敏感响应写入 GUI、run log 或诊断日志。 - `MainWindow` 已负责设置页离开守卫：切 Tab 与 `closeEvent` 发现 `SettingsTab.is_dirty()` 时弹保存/放弃/取消；保存成功后继续，保存失败或取消时回到设置。由于 `QTabWidget.currentChanged` 是切换后信号，需维护上一个 index，并用 `_reverting_tab_change` 或等价 guard 防止 `setCurrentIndex()` 递归。 - 模型详情字段按 3 个组件一组排列：启用、类别、api_type、连接超时等短字段一格；服务商名、模型 ID 视宽度占一格或两格；网址、密钥跨整行或跨 2/3 列。 diff --git a/docs/cmhub-integration-design.md b/docs/cmhub-integration-design.md index fb15fc6..5d71feb 100644 --- a/docs/cmhub-integration-design.md +++ b/docs/cmhub-integration-design.md @@ -15,6 +15,7 @@ > **v3.6 修订（2026-07-08，T-564）**：cmhub 已新增异步生图任务接口，②批量生图改为 `POST /api/v1/generate/image/tasks` submit + `GET /api/v1/generate/image/tasks/{task_id}` poll；cmshopee 持久化 `tasks.image_task_id/image_task_key`，支持停止/超时/重启后续查，避免 900 秒同步长连接和读超时重复扣点。旧同步 `POST /api/v1/generate/image` 仅保留给单独 `gen_cover()` 兼容/回滚路径。 > **v3.7 修订（2026-07-17，T-645）**：⑥「商品套图」的 AI帮写接入独立图片理解能力：使用 `vision_alias` 调 `POST /api/v1/analyze/images`，不复用标题接口或 `title_alias`。请求最多8张有序本地原图，单图不超过10MiB、总计不超过32MiB；读取等待120秒且读超时不自动重发，避免结果未确认时重复扣点。 > **v3.8 修订（2026-07-20，T-677）**：T-529 的普通 UI 单来源收口有限放开为「默认网关 / 自定义网关」分段选择器，来源仅在保存设置后切换。自定义网关是客户端直连的过渡兼容路径，只覆盖②生文、生图；模型清单与默认网关配置并存。⑥商品套图和图片理解仍只允许默认网关新建请求，但此前已由默认网关接受、且同时带 `generation_source=cmhub`、`provider=cmhub` 与 `task_id` 的套图 job 可继续轮询、下载、保存，避免已扣点结果丢失。运行中的 worker 冻结来源、模型与密钥的内存快照，后续保存设置不影响本轮；密钥不写入 SQLite、日志或 UI。 +> **v3.9 修订（2026-07-20，T-678）**：自定义网关的图片生成固定采用 OpenAI 图片编辑契约：`POST /v1/images/edits`、multipart 有序重复 `image[]`、第一张为业务主体图、`n=1`，仅从 `data[].b64_json` 或 `data[].url` 读取结果。`category=image` 只有 `api_type=images_edits` 可发起生成；历史 `chat` / `auto` 图片模型保留可查看但在设置和生成预检中明确为不支持，绝不静默迁移。默认网关的 cmhub 请求、异步任务与宽松响应解析不受该收敛影响。⑤对图片模型只做本地配置检查，不发送可能计费的空图片请求；实际可用性通过②或后续⑥的正常生成确认验证。 ## 1. 背景与目标 diff --git a/docs/routes.md b/docs/routes.md index fe29bca..74fd778 100644 --- a/docs/routes.md +++ b/docs/routes.md @@ -173,8 +173,8 @@ - 设置页整体布局：内容区居中，左右留白已从 T-506 初始实现缩短到约 40%；实现上用最大内容宽度 + 自适应 margin，而不是写死窗口像素。所有设置组默认响应式 3 列表单：短字段占 1 格，URL/API Key/路径等长字段跨 2 格或 3 格；窄窗口自动降为 2 列/1 列。点击「保存设置」成功后，状态栏显示“设置已保存”，并弹出轻量提示框。T-531 已完成：保存成功会清除未保存标记；保存失败时保留未保存标记并阻止离开。 - 生成网关（T-677）：用具备选中态的「默认网关 / 自定义网关」来源选择器替代历史「AI 后端」下拉。切换只改变待保存状态并立即切换面板，点击「保存设置」才写入 `ai.backend=cmhub/direct`；默认网关的 `data/config/cmhub.json` 与自定义模型的 `data/config/ai_models.json` 始终并存，模型新增、保存、删除仍立即写盘，不受来源待保存语义影响。 - 默认网关面板展示 Base URL、API Key、生文/生图/图片理解别名、连接超时、批量生成前检查余额、刷新别名和测试连接/查余额；Base URL 输入框旁提示“只填网关根，如 https://host”，保存/刷新前规整掉 `/api`、`/api/v1` 或其它路径。测试连接成功优先显示 `账号「<账号名>」连接默认网关成功`，邮箱脱敏；没有账号信息时显示「默认网关连接成功」。Key 只本地明文保存、UI 打码显示，不进入日志或导出。 - - 自定义网关复用既有模型列表、模型详情、默认文本模型、默认图像模型和模型连接测试；它仅兼容现有 direct 标题/封面请求，`chat/auto` 走聊天请求、`images_edits` 走图片编辑请求，不承诺自动探测或所有 OpenAI 兼容服务可用。说明行明确「图片理解与商品套图仅支持默认网关」。 - - 允许保存不完整配置。②点击生成后、自定义 worker 创建前按本轮标题/封面选择校验所需模型的启用状态、类别、地址、模型 ID、API Key 和接口类型，失败提示去⑤补齐且不发请求、不改任务状态；自定义网关不显示默认网关余额，运行日志和状态栏说明“不计点数，费用由服务商收取”。 + - 自定义网关复用既有模型列表、模型详情、默认文本模型与默认图像模型。文本模型保留 `chat` / `auto`；图片模型固定使用「OpenAI 图片编辑接口」（`images_edits`），②通过 `/v1/images/edits` 以有序重复 `image[]` 上传，第一张是商品主体图、每次只请求 `n=1` 输出。历史 `chat` / `auto` 图片模型不会被改写或隐藏，但会标示「图片接口不支持」，生成前会被拦截。图片响应只接受 `data[].b64_json` / `data[].url`，不适配聊天响应或服务商私有结构。说明行仍明确「图片理解与商品套图仅支持默认网关」。 + - 图片模型按钮显示「检查图片配置」，只在本地检查接口类型、URL、模型 ID、密钥和超时，绝不对 `/images/edits` 发送无图片的测试请求；实际可用性必须以②或后续⑥的正常生成确认判断。允许保存不完整配置。②点击生成后、自定义 worker 创建前按本轮标题/封面选择校验所需模型的启用状态、类别、地址、模型 ID、API Key 和接口类型，失败提示去⑤补齐且不发请求、不改任务状态；自定义网关不显示默认网关余额，运行日志和状态栏说明“不计点数，费用由服务商收取”。 - ⑥商品套图在自定义网关下禁用新建、重新生成和 AI帮写；但已有 `generation_source=cmhub`、`provider=cmhub` 且 `task_id` 非空的可恢复任务仍显示「继续查询已提交图片」，仅轮询、下载和保存，不新建 job 或重复扣点。若默认网关配置已不可用，提示恢复原默认网关配置。所有②/⑥ worker 开始时冻结来源、实际模型和密钥的仅内存快照，保存设置不会改变正在执行的一轮。 - T-531 已完成：设置页任意可编辑控件变更都进入未保存状态，保存按钮旁显示“● 未保存更改”；切换到其它 Tab 或关闭窗口时弹出保存/放弃/取消。放弃会重新从本地配置文件回填控件，避免未保存的 URL/API Key 留在界面上；程序化回填、保存后重载和刷新别名填充下拉不会误触发未保存状态。 - 自定义模型清单继续保存于 `data/config/ai_models.json`；在自定义网关面板中可维护并立即保存，切换回默认网关不会清除这些配置。 diff --git a/docs/tasks/T-678.md b/docs/tasks/T-678.md index c69711f..7ba0a9c 100644 --- a/docs/tasks/T-678.md +++ b/docs/tasks/T-678.md @@ -3,7 +3,7 @@ id: T-678 title: 自定义网关图片模型锁定 OpenAI 图片编辑接口并支持多图输入 phase: 7 deps: [T-677] -status: TODO +status: DONE created: 2026-07-20 --- @@ -88,4 +88,8 @@ git diff --check ## 执行记录 -（做完在这里写：改了什么文件、跑了什么验证命令及结果、官方接口契约核实来源与日期、遇到的阻塞和关键决策。） +- 2026-07-20 完成：保留全局 `API_TYPES`，新增按图片类别的 OpenAI 图片编辑接口校验；存量 `chat` / `auto` 图片模型仍可载入、查看和保存，但设置页标示为不支持，②预检及 direct 生图会在发请求前用中文拦截。 +- direct 封面请求统一使用 `/v1/images/edits` multipart；`_image_edit_body()` 支持有序多图，重复 `image[]` 字段的第一张为主体图，并固定 `n=1`。响应解析收敛到 `data[].b64_json` / `data[].url`；cmhub 保持独立的宽松解析函数，默认网关行为未改变。 +- 图片模型按钮改为「检查图片配置」，worker 只调用本地字段、接口类型、URL 和超时检查，不发送空图片请求；文字模型继续走既有连接测试。 +- 文档同步：`docs/cmhub-integration-design.md`、`docs/routes.md`、`docs/api.md`。接口契约于 2026-07-20 参考 [OpenAI Images API 文档](https://platform.openai.com/docs/guides/image-generation)：图片编辑使用 multipart 图片输入，图片结果使用 base64 或 URL 表达；本项目固定采用 `image[]`、`n=1` 和标准 `data` 结果字段，不适配服务商私有响应。 +- 验证通过：`py -3.10 -m unittest tests.test_ai tests.test_appconfig tests.test_gui`（287 passed）；`py -3.10 -m unittest discover -s tests`（643 passed）；`py -3.10 -m ruff check app tests main.py`；`py -3.10 -m compileall app main.py`；`git diff --check`。本任务未改 CDP，未作线上蝦皮实跑。 diff --git a/tests/test_ai.py b/tests/test_ai.py index f1fd628..6ff5b29 100644 --- a/tests/test_ai.py +++ b/tests/test_ai.py @@ -69,10 +69,10 @@ class AITests(TempDirMixin, unittest.TestCase): "name": "Image", "category": "image", "enabled": True, - "url": "https://example.invalid/v1/chat/completions", + "url": "https://example.invalid/v1", "model": "image-model", "api_key": "sk-image-secret", - "api_type": "auto", + "api_type": "images_edits", "connect_timeout_seconds": 1, "timeout_seconds": 1, "extra_body": {}, @@ -378,9 +378,13 @@ class AITests(TempDirMixin, unittest.TestCase): b64_image = base64.b64encode(generated.getvalue()).decode("ascii") def fake_urlopen(request, timeout=None): - body = json.loads(request.data.decode("utf-8")) - self.assertEqual("image-model", body["model"]) - self.assertIn("目标分辨率：512", body["messages"][0]["content"][0]["text"]) + body = request.data + self.assertIn(b'name="model"', body) + self.assertIn(b"image-model", body) + self.assertIn(b'name="image[]"', body) + self.assertIn(b'name="n"', body) + self.assertIn(b"\r\n1\r\n", body) + self.assertIn("multipart/form-data", request.headers["Content-type"]) return _Response({"data": [{"b64_json": b64_image}]}) with mock.patch("app.ai.urllib.request.urlopen", side_effect=fake_urlopen): @@ -399,6 +403,44 @@ class AITests(TempDirMixin, unittest.TestCase): self.assert_removed(temp_dir) + def test_image_edit_body_keeps_repeated_image_fields_in_order(self): + with self.make_temp_dir() as temp_dir: + first = os.path.join(temp_dir, "primary.jpg") + second = os.path.join(temp_dir, "reference.png") + with open(first, "wb") as fh: + fh.write(b"first-image") + with open(second, "wb") as fh: + fh.write(b"second-image") + + body, content_type = ai._image_edit_body( + {"model": "image-model", "extra_body": {"quality": "high"}}, + "生成商品图", + [first, second], + "1k", + ) + + self.assertIn("multipart/form-data", content_type) + self.assertEqual(2, body.count(b'name="image[]"')) + self.assertLess(body.index(b"primary.jpg"), body.index(b"reference.png")) + self.assertIn(b'name="n"', body) + self.assertIn(b"\r\n1\r\n", body) + + self.assert_removed(temp_dir) + + def test_direct_image_response_only_accepts_openai_data_fields(self): + encoded = base64.b64encode(b"image-bytes").decode("ascii") + + self.assertEqual(encoded, ai._find_image_ref({"data": [{"b64_json": encoded}]})) + self.assertEqual( + "https://images.example.com/new.png", + ai._find_image_ref({"data": [{"url": "https://images.example.com/new.png"}]}), + ) + self.assertIsNone(ai._find_image_ref({"choices": [{"image_url": "x"}]})) + with self.assertRaisesRegex(ai.AIError, "OpenAI 图片编辑接口"): + ai._extract_image_bytes({"message": {"url": "https://bad.example/x"}}, {}, {}) + with self.assertRaisesRegex(ai.AIError, "只允许 http/https"): + ai._extract_image_bytes({"data": [{"url": "file:///tmp/image.png"}]}, {}, {}) + def test_cmhub_gen_title_uses_alias_and_emits_metadata(self): with self.make_temp_dir() as temp_dir: cfg, key_path = self._cmhub_config(temp_dir) diff --git a/tests/test_appconfig.py b/tests/test_appconfig.py index cb0ebcb..33294ed 100644 --- a/tests/test_appconfig.py +++ b/tests/test_appconfig.py @@ -635,6 +635,40 @@ class AppConfigTests(TempDirMixin, unittest.TestCase): ), ) + def test_image_model_config_check_keeps_legacy_models_but_blocks_them_for_generation(self): + with self.make_temp_dir() as temp_dir: + models_path = os.path.join(temp_dir, "ai_models.json") + config = appconfig.default_ai_models_config() + legacy_image = config["models"][1] + legacy_image.update( + { + "url": "https://legacy.example.com/v1/chat/completions", + "model": "legacy-image", + "api_key": "sk-legacy-secret", + "api_type": "auto", + } + ) + appconfig.save_ai_models_config(config, path=models_path) + + loaded = appconfig.get_model("Nano Banana 2", path=models_path) + self.assertEqual("auto", loaded["api_type"]) + result = appconfig.check_image_model_config("Nano Banana 2", path=models_path) + self.assertFalse(result["ok"]) + self.assertTrue(result["check_only"]) + self.assertIn("OpenAI 图片编辑接口", result["error"]) + self.assertNotIn("sk-legacy-secret", result["error"]) + + appconfig.update_ai_model( + "Nano Banana 2", + path=models_path, + api_type="images_edits", + ) + self.assertTrue( + appconfig.check_image_model_config("Nano Banana 2", path=models_path)["ok"] + ) + + self.assert_removed(temp_dir) + def test_ai_model_test_uses_resolved_base_url(self): with self.make_temp_dir() as temp_dir: models_path = os.path.join(temp_dir, "ai_models.json") diff --git a/tests/test_gui.py b/tests/test_gui.py index d90e7e1..086e86e 100644 --- a/tests/test_gui.py +++ b/tests/test_gui.py @@ -2232,6 +2232,7 @@ class GuiTests(TempDirMixin, unittest.TestCase): with self.make_temp_dir() as temp_dir: cfg = self.make_config(temp_dir) models_path = cfg["ai_models_path"] + statuses = [] appconfig.save_ai_models_config( { "models": [ @@ -2264,7 +2265,11 @@ class GuiTests(TempDirMixin, unittest.TestCase): path=models_path, ) - tab = SettingsTab(config=cfg, ai_models_path=models_path) + tab = SettingsTab( + config=cfg, + ai_models_path=models_path, + status_callback=statuses.append, + ) self.addCleanup(tab.close) self.assertEqual(2, tab.model_combo.count()) @@ -2523,6 +2528,74 @@ class GuiTests(TempDirMixin, unittest.TestCase): self.assert_removed(temp_dir) + def test_settings_tab_image_model_only_offers_image_edit_and_checks_locally(self): + with self.make_temp_dir() as temp_dir: + cfg = self.make_config(temp_dir) + models_path = cfg["ai_models_path"] + statuses = [] + appconfig.save_ai_models_config( + { + "models": [ + { + "name": "Text A", + "category": "text", + "enabled": True, + "url": "https://example.invalid/v1", + "model": "text-model", + "api_key": "sk-text", + "api_type": "chat", + "connect_timeout_seconds": 30, + "timeout_seconds": 0, + "extra_body": {}, + }, + { + "name": "Legacy Image", + "category": "image", + "enabled": True, + "url": "https://example.invalid/v1/chat/completions", + "model": "image-model", + "api_key": "sk-image", + "api_type": "auto", + "connect_timeout_seconds": 30, + "timeout_seconds": 0, + "extra_body": {}, + }, + ] + }, + path=models_path, + ) + tab = SettingsTab( + config=cfg, + ai_models_path=models_path, + status_callback=statuses.append, + ) + self.addCleanup(tab.close) + tab.model_combo.setCurrentIndex(tab.model_combo.findData("Legacy Image")) + + self.assertIn("当前图片模型不支持 OpenAI 图片编辑接口", tab.model_combo.currentText()) + self.assertEqual("auto", tab.api_type_combo.currentData()) + self.assertEqual("检查图片配置", tab.test_connection_button.text()) + self.assertIn("当前不支持", tab.api_type_combo.currentText()) + + class FakeSignal: + def connect(self, callback): + self.callback = callback + + class FakeThread: + def __init__(self): + self.finished = FakeSignal() + + def start(self): + pass + + with mock.patch("app.gui.run_worker", return_value=FakeThread()): + tab.test_connection() + + self.assertTrue(tab.test_worker.check_image_config) + self.assertIn("正在检查图片模型配置", statuses[-1]) + + self.assert_removed(temp_dir) + def test_settings_tab_saves_role_generation_path_and_port_config(self): with self.make_temp_dir() as temp_dir: cfg = self.make_config(temp_dir)