Add image generation support (bridge Tabbit's built-in image gen)

[liwei9745]

Background

Tabbit browser has built-in image generation capability (triggered via task mode), but tabbit2api currently only bridges the text chat interface (/v1/chat/completions) and does not expose image generation capabilities to downstream clients like CherryStudio, Codex, or Claude Code.

Packet Capture Analysis

We conducted a network packet capture using Playwright to reverse-engineer how Tabbit handles image generation internally.

Image Generation API Call Chain

1. POST /api/v1/chat/completion — User sends an image prompt (e.g., "画一只猫")
2. GET /proxy/v0/browser-task/graph — Tabbit polls task mode status
3. POST /proxy/v0/cos/presigned-download-url — Fetches a presigned download URL from Tencent COS
4. Image URL is returned via SSE stream embedded as Markdown: ![image](cos_url)

Key Internal Endpoints

Endpoint	Method	Purpose
/api/v1/chat/completion	POST	Send chat/image generation request
/proxy/v0/browser-task/graph	GET	Poll task execution status
/proxy/v0/cos/presigned-download-url	POST	Get COS presigned download URL

Image Storage

• Provider: Tencent Cloud COS (Singapore region)
• Path pattern: https://tab-sg-1300456063.cos.ap-singapore.myqcloud.com/{user_id}/{session_id}/image_generation/{YYYYMMDD}/{UUID}.png
• Access: Presigned URLs with time-limited signatures

Example Request

POST /api/v1/chat/completion
{
  "messages": [
    {"role": "user", "content": "生成一张猫的图片"}
  ],
  "model": "gpt-4o",
  "stream": true
}

Example Response (SSE)

data: {"choices":[{"delta":{"content":"![image](https://tab-sg-1300456063.cos.ap-singapore.myqcloud.com/3bc8.../image_generation/20260623/32a9...png?q-sign-algorithm=sha1&...)"}}]}

Proposed Solutions

Option A: Extend Existing /v1/chat/completions (Recommended for MVP)

Parse Markdown image syntax from the SSE stream and return image URLs as part of the assistant message content.

Pros: Minimal code changes, backward compatible.
Cons: Clients need to parse Markdown images; no standard /v1/images/generations endpoint.

Option B: Add /v1/images/generations Endpoint (OpenAI Compatible)

Implement the standard OpenAI image generation API:

POST /v1/images/generations
{
  "prompt": "a cat",
  "n": 1,
  "size": "1024x1024"
}

Response:

{
  "created": 1234567890,
  "data": [{ "url": "https://..." }]
}

Pros: Fully compatible with OpenAI API spec; any OpenAI client can use it directly.
Cons: Requires waiting for SSE stream to complete; higher latency.

Option C: Browser Automation Click (Fallback)

Use Playwright to simulate clicking the "image generation" button in the Tabbit UI.

Pros: Does not depend on internal API stability.
Cons: Fragile to UI changes; high overhead.

Recommended Implementation

Phase 1 (Short-term): Extend chat/completions to detect and relay image URLs.
Phase 2 (Medium-term): Add native /v1/images/generations endpoint.

Environment

• tabbit2api v0.1.3
• Tabbit browser 1.41.10
• OS: Windows 10
• Capture date: 2026-06-23

Additional Context

• Models with supports_images: true flag (e.g., GPT-5.5, Claude-Opus-4.7) should work
• Image generation currently requires Tabbit's "task mode" to be activated in the UI
• The existing Playwright-based bridge (tabbit-web-bridge.js) could be extended to pass image prompts

[hwttop5]

感谢整理，信息很完整。这个方向可以作为图片生成支持的 feature request 跟进。

先说明一下当前边界：tabbit2api 目前主要桥接 Tabbit 的文本 chat 能力，并支持图片/文档作为输入附件；这和“通过 Tabbit task mode 生成图片”不是同一个能力。所以现在还没有标准的 /v1/images/generations 实现。