diff --git a/.memory/worklog.json b/.memory/worklog.json index 5b3d04d..44ec84c 100644 --- a/.memory/worklog.json +++ b/.memory/worklog.json @@ -1,32 +1,5 @@ { "entries": [ - { - "files_changed": 3, - "hash": "ecc5894", - "message": "auto-save 2026-05-15 16:27 (~3)", - "ts": "2026-05-15T16:27:46+08:00", - "type": "commit" - }, - { - "files_changed": 4, - "hash": "1336fc6", - "message": "auto-save 2026-05-15 16:33 (~4)", - "ts": "2026-05-15T16:33:17+08:00", - "type": "commit" - }, - { - "files_changed": 1, - "message": "Codex 会话活跃 · 最近命令：codex · 1 项未提交变更 · 最近提交：auto-save 2026-05-15 16:33 (~4)", - "ts": "2026-05-15T08:34:47Z", - "type": "session-heartbeat" - }, - { - "files_changed": 1, - "hash": "2ad7884", - "message": "auto-save 2026-05-15 16:38 (~1)", - "ts": "2026-05-15T16:38:48+08:00", - "type": "commit" - }, { "files_changed": 2, "hash": "6fdcb47", @@ -3255,6 +3228,32 @@ "type": "session-heartbeat", "message": "Codex 会话活跃 · 最近命令：codex · 分支 main · 1 项未提交变更 · 最近提交：fix: force gpt image model", "files_changed": 1 + }, + { + "ts": "2026-05-18T06:33:24+08:00", + "type": "commit", + "message": "auto-save 2026-05-18 06:33 (~5)", + "hash": "18fd8c9", + "files_changed": 5 + }, + { + "ts": "2026-05-17T22:33:44Z", + "type": "session-heartbeat", + "message": "Codex 会话活跃 · 最近命令：codex · 分支 main · 1 项未提交变更 · 最近提交：auto-save 2026-05-18 06:33 (~5)", + "files_changed": 1 + }, + { + "ts": "2026-05-18T06:38:01+08:00", + "type": "commit", + "message": "feat: manage subject view thumbnails", + "hash": "58c3e5c", + "files_changed": 1 + }, + { + "ts": "2026-05-17T22:43:44Z", + "type": "session-heartbeat", + "message": "Codex 会话活跃 · 最近命令：codex · 分支 main · 2 项未提交变更 · 最近提交：feat: manage subject view thumbnails", + "files_changed": 2 } ] } diff --git a/api/main.py b/api/main.py index 51debf3..5116a84 100644 --- a/api/main.py +++ b/api/main.py @@ -2540,6 +2540,42 @@ def start_audio_processing(job_id: str, manage_job_status: bool = True) -> bool: return True +def _image_is_capacity_error(status_code: int, body: str) -> bool: + lower = body.lower() + return ( + status_code == 429 + or ( + status_code in (500, 502, 503, 504) + and any(token in lower for token in ("saturated", "rate", "quota", "capacity", "overload", "timeout", "繁忙", "饱和", "过载")) + ) + ) + + +def _image_retry_delay(attempt: int, status_code: int = 0, body: str = "", retry_after: str | None = None) -> float: + if retry_after: + try: + return max(1.0, min(60.0, float(retry_after))) + except ValueError: + pass + if _image_is_capacity_error(status_code, body): + return [6.0, 14.0, 30.0, 45.0][min(attempt, 3)] + return [1.0, 2.0, 4.0, 8.0][min(attempt, 3)] + + +def _image_failure_message(kind: str, attempts: int, last_err: str, capacity_seen: bool) -> str: + if capacity_seen: + return ( + f"{kind} failed after {attempts} attempts: gpt-image-2 上游负载饱和，" + f"已自动退避重试仍失败，请稍后点重试。最后错误：{last_err}" + ) + return f"{kind} failed after {attempts} attempts: {last_err}" + + +def _image_error_status(error: Exception) -> int: + msg = str(error) + return 503 if ("上游负载饱和" in msg or "HTTP 429" in msg or "saturated" in msg.lower()) else 500 + + def _image_edit_call( image_path: Path, prompt: str, @@ -2574,9 +2610,6 @@ def _image_edit_call( except Exception: # PIL 失败兜底走原文件 img_bytes_in = image_path.read_bytes() - img_b64 = b64lib.b64encode(img_bytes_in).decode("ascii") - data_uri = f"data:image/jpeg;base64,{img_b64}" - plan: list[str] = ["edit"] * max_attempts if fallback_text: plan.append("text") @@ -2584,18 +2617,22 @@ def _image_edit_call( last_err = "" resp_data: dict = {} effective_mode = "edit" + capacity_seen = False for attempt, current_mode in enumerate(plan): current_model = models_cycle[min(attempt, len(models_cycle) - 1)] + status_code = 0 + body = "" + retry_after: str | None = None try: if current_mode == "edit": with httpx.Client(timeout=120) as client: r = client.post( - f"{IMAGE_BASE_URL}/images/generations", + f"{IMAGE_BASE_URL}/images/edits", headers={ "Authorization": f"Bearer {IMAGE_API_KEY}", - "Content-Type": "application/json", }, - json={"model": current_model, "prompt": prompt, "image": data_uri, "n": 1}, + data={"model": current_model, "prompt": prompt, "n": "1"}, + files={"image": ("reference.jpg", img_bytes_in, "image/jpeg")}, ) r.raise_for_status() resp_data = r.json() @@ -2610,23 +2647,32 @@ def _image_edit_call( last_err = f"empty data · {err_obj.get('code', '')} · {str(err_obj.get('message', ''))[:200]} · model={current_model}" except httpx.HTTPStatusError as e: body = e.response.text - sc = e.response.status_code - fatal = sc in (401, 403) - last_err = f"HTTP {sc}: {body[:200]} · model={current_model}" + status_code = e.response.status_code + retry_after = e.response.headers.get("retry-after") + capacity_seen = capacity_seen or _image_is_capacity_error(status_code, body) + fatal = status_code in (401, 403) + last_err = f"HTTP {status_code}: {body[:200]} · model={current_model}" if fatal: - raise RuntimeError(f"image edit HTTP {sc}: {body[:300]}") + raise RuntimeError(f"image edit HTTP {status_code}: {body[:300]}") except Exception as e: last_err = f"{type(e).__name__}: {e} · model={current_model}" if attempt < len(plan) - 1: tag = f"retry {attempt + 1}/{len(plan)} → {GPT_IMAGE_MODEL}" - print(f"[image edit {tag}] {last_err}", flush=True) - _time.sleep(1.0) + delay = _image_retry_delay(attempt, status_code, body, retry_after) + print(f"[image edit {tag}, sleep {delay:.0f}s] {last_err}", flush=True) + _time.sleep(delay) data_arr = resp_data.get("data", []) if not data_arr: - raise RuntimeError(f"image edit failed after {len(plan)} attempts: {last_err}") - b64 = data_arr[0].get("b64_json") + raise RuntimeError(_image_failure_message("image edit", len(plan), last_err, capacity_seen)) + item = data_arr[0] + b64 = item.get("b64_json") + if not b64 and item.get("url"): + with httpx.Client(timeout=120) as client: + image_resp = client.get(item["url"]) + image_resp.raise_for_status() + return image_resp.content, effective_mode if not b64: raise RuntimeError("image edit returned no b64_json") return b64lib.b64decode(b64), effective_mode @@ -2646,8 +2692,11 @@ def _image_text_call( models_cycle = [GPT_IMAGE_MODEL] last_err = "" resp_data: dict = {} + capacity_seen = False for attempt in range(max_attempts): current_model = models_cycle[min(attempt, len(models_cycle) - 1)] + status_code = 0 + body = "" try: resp = image_llm().images.generate(model=current_model, prompt=prompt, n=1) resp_data = resp.model_dump() if hasattr(resp, "model_dump") else {"data": [{"b64_json": resp.data[0].b64_json}]} @@ -2659,10 +2708,14 @@ def _image_text_call( last_err = f"empty data · {err_obj.get('code', '')} · {str(err_obj.get('message', ''))[:200]} · model={current_model}" except Exception as e: last_err = f"{type(e).__name__}: {e} · model={current_model}" + body = str(e) + status_code = 429 if "429" in body or "saturated" in body.lower() or "饱和" in body else 0 + capacity_seen = capacity_seen or _image_is_capacity_error(status_code, body) if attempt < max_attempts - 1: - print(f"[image text retry {attempt + 1}/{max_attempts} → {GPT_IMAGE_MODEL}] {last_err}", flush=True) - _time.sleep(1.0) - raise RuntimeError(f"image text failed after {max_attempts} attempts: {last_err}") + delay = _image_retry_delay(attempt, status_code, body) + print(f"[image text retry {attempt + 1}/{max_attempts} → {GPT_IMAGE_MODEL}, sleep {delay:.0f}s] {last_err}", flush=True) + _time.sleep(delay) + raise RuntimeError(_image_failure_message("image text", max_attempts, last_err, capacity_seen)) # ---------- API 路由 ---------- @@ -3166,9 +3219,9 @@ def generate_image(job_id: str, idx: int, req: GenerateReq) -> Job: import time as _time import httpx - img_b64: str | None = None + img_bytes_in: bytes | None = None if req.mode == "edit": - img_b64 = b64lib.b64encode(reference_path.read_bytes()).decode("ascii") + img_bytes_in = reference_path.read_bytes() # 尝试 i2i 最多 3 次，全失败时降级 text-only 再试 1 次 plan: list[str] = ([req.mode] * 3) if req.mode == "edit" else [req.mode] @@ -3177,24 +3230,23 @@ def generate_image(job_id: str, idx: int, req: GenerateReq) -> Job: resp_data: dict = {} last_err = "" effective_mode = req.mode + capacity_seen = False for attempt, current_mode in enumerate(plan): + status_code = 0 + body = "" + retry_after: str | None = None try: if current_mode == "edit": - data_uri = f"data:image/jpeg;base64,{img_b64}" - # OpenAI SDK 不直接支持 image 参数，用底层 httpx + if img_bytes_in is None: + raise RuntimeError("edit mode reference image missing") with httpx.Client(timeout=120) as client: r = client.post( - f"{IMAGE_BASE_URL}/images/generations", + f"{IMAGE_BASE_URL}/images/edits", headers={ "Authorization": f"Bearer {IMAGE_API_KEY}", - "Content-Type": "application/json", - }, - json={ - "model": model, - "prompt": full_prompt, - "image": data_uri, - "n": 1, }, + data={"model": model, "prompt": full_prompt, "n": "1"}, + files={"image": ("reference.jpg", img_bytes_in, "image/jpeg")}, ) r.raise_for_status() resp_data = r.json() @@ -3210,15 +3262,20 @@ def generate_image(job_id: str, idx: int, req: GenerateReq) -> Job: last_err = f"empty data · {err_obj.get('code', '')} · {str(err_obj.get('message', ''))[:200]}" except httpx.HTTPStatusError as e: body = e.response.text + status_code = e.response.status_code + retry_after = e.response.headers.get("retry-after") + capacity_seen = capacity_seen or _image_is_capacity_error(status_code, body) transient = ( - e.response.status_code >= 500 + status_code == 429 + or status_code >= 500 or "incomplete_generation" in body or "rate_limit" in body or "timeout" in body.lower() + or _image_is_capacity_error(status_code, body) ) - last_err = f"HTTP {e.response.status_code}: {body[:200]}" + last_err = f"HTTP {status_code}: {body[:200]}" if not transient: - raise HTTPException(500, f"image gen HTTP {e.response.status_code}: {body[:300]}") + raise HTTPException(500, f"image gen HTTP {status_code}: {body[:300]}") except Exception as e: last_err = f"{type(e).__name__}: {e}" @@ -3226,22 +3283,29 @@ def generate_image(job_id: str, idx: int, req: GenerateReq) -> Job: next_mode = plan[attempt + 1] tag = f"fallback → {next_mode}" if next_mode != current_mode else f"retry {attempt + 1}/{len(plan)}" print(f"[image gen {tag}] {last_err}", flush=True) - _time.sleep(1.5 * (attempt + 1)) + _time.sleep(_image_retry_delay(attempt, status_code, body, retry_after)) data_arr = resp_data.get("data", []) if not data_arr: - raise HTTPException(500, f"image gen failed after {len(plan)} attempts: {last_err}") + raise HTTPException(503 if capacity_seen else 500, _image_failure_message("image gen", len(plan), last_err, capacity_seen)) item = data_arr[0] b64 = item.get("b64_json") - if not b64: + if b64: + out_bytes = b64lib.b64decode(b64) + elif item.get("url"): + with httpx.Client(timeout=120) as client: + image_resp = client.get(item["url"]) + image_resp.raise_for_status() + out_bytes = image_resp.content + else: raise HTTPException(500, "image gen returned no b64_json") # 保存到本地 jobs//gen/_.jpg gen_dir = job_dir(job_id) / "gen" gen_dir.mkdir(parents=True, exist_ok=True) out_path = gen_dir / f"{idx:03d}_{gen_id}.jpg" - out_path.write_bytes(b64lib.b64decode(b64)) + out_path.write_bytes(out_bytes) new_gen = GeneratedImage( id=gen_id, @@ -4073,7 +4137,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat try: img_bytes, _mode = _image_edit_call(model_src, prompt, models=models, fallback_text=False, max_attempts=3, max_side=1280) except RuntimeError as e: - raise HTTPException(500, f"subject asset {view} failed: {e}") + raise HTTPException(_image_error_status(e), f"subject asset {view} failed: {e}") asset_id = f"subject_{idx:03d}_{element_id}_{view}_{uuid.uuid4().hex[:8]}" out_path = job_dir(job_id) / "assets" / f"{asset_id}.jpg" @@ -5073,9 +5137,9 @@ def generate_product_angle_asset(job_id: str, req: GenerateProductAngleAssetReq) ) models = [GPT_IMAGE_MODEL] try: - img_bytes, _mode = _image_edit_call(source_path, prompt, models=models, fallback_text=False, max_attempts=3, max_side=1280) + img_bytes, _mode = _image_edit_call(source_path, prompt, models=models, fallback_text=False, max_attempts=5, max_side=1280) except RuntimeError as e: - raise HTTPException(500, f"product angle generation failed: {e}") + raise HTTPException(_image_error_status(e), f"product angle generation failed: {e}") asset_id = f"product_angle_{uuid.uuid4().hex[:10]}" out_path = job_dir(job_id) / "assets" / f"{asset_id}.jpg" _normalize_asset_image(img_bytes, out_path, source_path, "1024", "white", square=True, fill_subject=True) diff --git a/web/lib/api.ts b/web/lib/api.ts index c4b345b..c51065f 100644 --- a/web/lib/api.ts +++ b/web/lib/api.ts @@ -1,5 +1,16 @@ const API_BASE = process.env.NEXT_PUBLIC_API_BASE ?? "http://localhost:4291" +function apiError(prefix: string, status: number, text: string) { + let detail = text + try { + const parsed = JSON.parse(text) + detail = parsed?.detail || text + } catch { + detail = text + } + return new Error(`${prefix} ${status} ${String(detail).slice(0, 300)}`) +} + export type JobStatus = | "created" | "downloading"