Capture descriptions: support any OpenAI-compatible vision endpoint
维护者通常 1 天内回复
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 30/100
- Issue 类型
- 功能
- 描述清晰度
- 描述清楚
- 活跃度
- 活跃
- 技术栈
- typescript
调研方向
Start in packages/cli/src/capture/contentExtractor.ts, focusing on captionImagesWithGemini and the existing OpenRouter request branch. Trace postExtractionPhase to understand credential detection and degraded reporting. Done means custom endpoint credentials are handled as specified, OpenRouter behavior remains unchanged, and the no-credentials guidance mentions the custom endpoint.
由索引模型根据 Issue 内容生成。
描述
Problem
hyperframes capture describes downloaded images only through OpenRouter, Vertex or the Gemini API (captionImagesWithGemini in packages/cli/src/capture/contentExtractor.ts). None of them is reliably reachable from mainland China, and none can point at a self-hosted vision model. Users in that position get the catalog-fallback asset-descriptions.md on every capture, even when they already pay for an OpenAI-compatible vision endpoint or run one locally.
Proposed solution
Bring back the three variables from #1809 on current main: HYPERFRAMES_VISION_BASE_URL, HYPERFRAMES_VISION_API_KEY and HYPERFRAMES_VISION_MODEL.
- All three set: caption through
<base>/chat/completionswith the OpenAI-style request the OpenRouter branch already sends, ahead of OpenRouter, Vertex and Gemini. - Some but not all set: warn naming the missing variables (never echoing values), skip captioning, and report the phase as degraded. Never fall back to another provider.
- OpenRouter shares the same request helper and keeps its URL, auth header and error text unchanged.
postExtractionPhasecounts a complete custom endpoint as vision credentials, and the no-credentials hint mentions it.
I have this ready and will open a PR that links here.
Alternatives considered
A named preset per regional provider. Every one of them speaks the same chat-completions wire format, so a single base-URL setting covers them all and also covers local Ollama and vLLM.
Additional context
The #1809 review called the base-URL approach "the right shape" and asked for one fix: setting only some of the variables fell through silently and could caption through Gemini, the provider the user is trying to avoid. That PR was closed for merge conflicts and inactivity, before Vertex support (#3561) landed in the same function.
To check that the request shape works outside OpenRouter, I sent exactly what the OpenRouter branch builds today (one image_url data URI, the storyboard prompt, max_tokens: 500) to iFlytek Astron's Token Plan, an OpenAI-compatible endpoint (https://maas-token-api.cn-huabei-1.xf-yun.com/v2). The image was docs/images/image-thumbnail-evidence/after.png.
| Model | HTTP | finish | Caption |
|---|---|---|---|
xopkimik26 (Kimi-K2.6) |
200 | stop | A dark gray webpage displays white text reading "Timeline JPEG thumbnail resource check" with a horizontal strip of rainbow-colored vertical bars beneath it. |
xopqwen35397b (Qwen3.5-397B) |
200 | stop | A dark background displays a horizontal strip of repeating colorful vertical bars labeled "Synthetic JPEG chart" … |
xopqwen36v35b (Qwen3.6-35B) |
200 | stop | A dark-background interface displays a narrow, horizontally stretched bar of repeating rainbow-colored vertical stripes … |
spark-x2.5 (text-only) |
400 | invalid request |
None of the vision models spent the 500-token budget on reasoning. The text-only row is why the docs should say to choose a model that accepts image input.
- 主要语言
- TypeScript
- 星标
- 54.1k
- 派生
- 4.9k
- 平均合并
- 7 小时 18 分钟
- 30 天内合并 PR
- 784
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
heygen-com/hyperframes 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 85/100
heygen-com/hyperframes#5027 ·
维护者通常 1 天内回复
-
fix(producer): propagate useGpu to HDR layered streaming encoder可能已有人在做 @Monster-GM 于 1 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 87/100
heygen-com/hyperframes#5002 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
heygen-com/hyperframes#4702 · 1 条评论 · 1 个 reaction ·
维护者通常 1 天内回复
-
Studio catalog prompt editor has no accessible name可能已有人在做 @lorenzozanee 于 12 天前认领。 未关闭bug difficulty/easy triage/ready
难度 2/5 1-3 小时 新手友好度 78/100
heygen-com/hyperframes#4384 ·
维护者通常 1 天内回复
-
lint: validate composition variables declared on supported root elements可能重新可做 关联的 PR 已关闭且未合并。 未关闭bug difficulty/easy triage/ready
难度 2/5 1-3 小时 新手友好度 88/100
heygen-com/hyperframes#4383 ·
维护者通常 1 天内回复
查看 heygen-com/hyperframes 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 83/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 75/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 68/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 85/100
platformatic/mcp#208 ·
维护者通常 1 天内回复
-
🐛 bug
难度 2/5 1-3 小时 新手友好度 66/100
margelo/react-native-vision-camera#4211 ·
维护者通常 4 天内回复