Skip to content

exposed controls, VLM import, prompt customization and related experiments - #631

Open
CCSSNE wants to merge 244 commits into
a-ghorbani:mainfrom
CCSSNE:main
Open

exposed controls, VLM import, prompt customization and related experiments#631
CCSSNE wants to merge 244 commits into
a-ghorbani:mainfrom
CCSSNE:main

Conversation

@CCSSNE

@CCSSNE CCSSNE commented Mar 14, 2026

Copy link
Copy Markdown

Description

Since the majority of the users of this application are those who wish to gain a deeper understanding of large models, I place greater emphasis on learning and exploration. I try to provide more parameters for users to debug, with easy recovery from incorrect adjustments, while exposing as much of the pipeline as possible so users can clearly see the model’s inputs, outputs, and adjustment methods. Because the chain is fully controllable, when the model encounters issues, configurations can be adjusted directly instead of waiting for code-level changes, making experimentation more flexible.

This branch represents a working implementation that I have been using and testing on my side.

I intentionally squashed this into a single commit, since the original commit history mostly reflected the development process and was quite messy, making it unsuitable for review.

If some of the directions here seem useful or aligned with the project, I can help isolate and clean those parts into proper commits or smaller PRs later.

This fork mainly reflects my own usage needs.

Main changes in this fork include:

  • Added the ability to directly import local VLM models
  • Support for both model imports and visual component imports for MM Projects
  • Fully exposed and customizable prompt template logic
  • Added functionality to share GGUF model files
  • Implemented logging features
  • Exposed several interfaces that previously existed internally but were not visible in the UI
  • Optimized the template matching mechanism in v132. Besides the 10 recommended built-in models, it now prioritizes reading the model's own configuration to reduce mismatching across different model families

One thing I originally intended to add but did not complete:

  • real-time monitoring of RAM / virtual memory usage, CPU utilization, and power consumption for performance debugging. I dropped this for now because the implementation complexity became too high.

The README in my fork also summarizes the general direction of these changes.

Platform Affected

  • [ ok] Android

Checklist

  • Necessary comments have been made.
  • I have tested this change on:
    • iOS Simulator/Device
    • [ok ] Android Emulator/Device
  • [ ok] Unit tests and integration tests pass locally.

CCSSNE added 25 commits March 24, 2026 06:34
1
1
- Android 原生 HTTP/SSE 传输(Kotlin ServerSocket,零新增依赖)
- GET /v1/models 与 POST /v1/chat/completions(流式 SSE + 非流式)
- 工具调用透传 + 老版 functions/function_call 转译,思维链 reasoning_content 输入输出
- MCP 工具注册表与命名空间合并,客户端 MCP 经标准 tool 循环透明支持
- FIFO 排队与 UI 聊天互斥;Bearer/x-api-key 鉴权默认开启
- 侧边栏新增 API 服务页:启停、局域网地址复制、端口/Key/自启动、请求日志
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants