feat(multimodal): 增强附件 UID、图文混排与参考图生图 by 69gg · Pull Request #53 · 69gg/Undefined

69gg · 2026-04-02T02:05:16Z

变更概览

新增持久化附件 UID 注册层，为普通消息与合并转发里的图片/文件建立当前会话可复用的 pic_* / file_* 标识。
打通图文混排发送链路，支持在回复中使用 <pic uid="..."/> 内嵌图片，并在历史记录与 WebUI 中保持可追溯的影子文本。
扩展 ai_draw_one：默认返回可嵌入图片 UID，新增参考图生图能力，支持通过 reference_image_uids 走 images/edits，并新增平级配置 [models.image_edit]。
为生图请求增加基于 agent 模型的提示词审核；同时保留上游错误透传，便于定位模型或服务端能力问题。
侧写与认知记忆补充“改名收敛”：当用户或群名称变化时，自动刷新现有侧写的展示名与向量元数据，不改变实体归属与摘要正文。

主要内容

多模态与文件分析链路现在优先使用内部附件 UID，不再依赖临时 file_id -> url 解析作为主路径。
file_analysis_agent / download_file 已支持 UID、URL 和 legacy file_id 三种输入来源。
WebUI 会话也接入了附件注册与本地图展示，file:// 图片在运行时聊天中可正确显示。
参考图生图配置采用与 models_image_gen 平级的 [models.image_edit]，字段与生成模型配置保持一致，能力由上游模型自行决定。

测试

uv run pytest tests/

chatgpt-codex-connector · 2026-04-02T02:05:24Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

Add delivery parameter (embed/send, default embed) to render_markdown, render_latex, render_html, get_picture, and minecraft_skin so they register images via AttachmentRegistry and return <pic uid="..."/> tags. Update wenchang_dijun to register sign images by UID instead of returning raw URLs. Add new fetch_image_uid tool for converting arbitrary image URLs into attachment UIDs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add async binary reads for reference images and cover related prompt-format regressions. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Apply Ruff formatting to the new async read regression test. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Update package versions, lockfiles, and changelog for the 3.2.8 release. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Avoid syncing placeholder poke names into cognitive profiles, align WebUI tool timeout handling with actual agent schemas, and load attachment registry asynchronously. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Apply Ruff formatting to the new review regression tests. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Make WebUI session scope explicit in runtime chat context and add regression tests for WebUI-scoped image embeds. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Reorder prompt construction so stable instruction blocks stay ahead of frequently changing context, improving prompt-cache friendliness without changing prompt content. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Restore bounded WebUI tool invoke timeouts for non-agent tools, prune and flush attachment registry state, deduplicate profile name refresh scheduling, pass history_message in group auto replies, and simplify ai_draw_one request param flow. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

Stop AttachmentRegistry from reading disk during construction and keep initial registry loading explicit and async. Co-authored-by: GPT-5.4 xhigh <noreply@openai.com>

69gg added 10 commits March 31, 2026 23:59

fix(agent): remove invoke timeout caps

ec04412

fix(image-gen): support base64 responses and preserve size

1c3061f

fix(skills): register dynamic modules before exec

49ea667

fix(image-gen): expose upstream errors and lock model

b8477b9

fix(image-gen): default responses to base64

03ce712

feat(multimodal): add attachment uid registry and pic embeds

a4f402d

feat(image-gen): add agent-based prompt moderation

dad2970

feat(multimodal): support forwarded attachment uids

7cf6bef

feat(image-gen): support reference image edits

9102ee2

fix(cognitive): refresh profile display names on rename

73f3801