feat: harden agent tooling compatibility
This commit is contained in:
lutc5
43
README.md
43
README.md
@@ -8,7 +8,7 @@ The project is designed for tools such as Claude Code, Cline, Continue, OpenCode
|
||||
|
||||
## Current Version
|
||||
|
||||
The current desktop line is `v1.2.2`.
|
||||
The current desktop line is `v1.3.0`.
|
||||
|
||||
Release builds are produced by GitHub Actions for:
|
||||
|
||||
@@ -52,7 +52,10 @@ Narrow window layout:
|
||||
| --- | --- | --- |
|
||||
| Health | `GET /` and `GET /health` | supported |
|
||||
| Models | `GET /v1/models` | supported |
|
||||
| Capability Discovery | `GET /capabilities`, `GET /v1/capabilities` | supported |
|
||||
| LM Studio / Ollama Discovery | `GET /api/v1/models`, `GET /api/tags`, `GET /props` | supported |
|
||||
| OpenAI Chat Completions | `POST /v1/chat/completions` | streaming and non-streaming |
|
||||
| OpenAI Chat Alias | `POST /api/v1/chat/completions` | supported |
|
||||
| Anthropic Messages | `POST /v1/messages` | streaming and non-streaming |
|
||||
|
||||
## What This Fork Adds
|
||||
@@ -61,7 +64,10 @@ Compared with the original protocol proof of concept, this repository focuses on
|
||||
|
||||
- **Function Calling / Tools** for both OpenAI and Anthropic clients.
|
||||
- **Tool result continuation** for multi-step agent loops.
|
||||
- **Tool stability hardening** with proxy-side routing hints, core tool examples, missed-tool retry, and common alias mapping such as `Bash` to `terminal` and `Read` to `read_file`.
|
||||
- **Image input** for OpenAI `image_url` and Anthropic image blocks.
|
||||
- **Local and remote image normalization** for data URLs, HTTP URLs, `file://` URLs, and absolute local paths, with automatic JPEG downscaling for large images.
|
||||
- **Request log image redaction** so large base64 payloads are visible as image markers instead of breaking the desktop log view.
|
||||
- **More request parameter compatibility** so stricter clients can connect without custom patches.
|
||||
- **Full request and response recording** in the desktop app for debugging 400/500 errors.
|
||||
- **macOS and Windows desktop app** with start/stop/restart, settings, logs, model discovery, themes, and window lifecycle handling.
|
||||
@@ -77,7 +83,7 @@ The proxy accepts common OpenAI request fields:
|
||||
- `presence_penalty`, `frequency_penalty`
|
||||
- `tools`, `tool_choice`, `parallel_tool_calls`
|
||||
- `response_format`, `seed`, `user`, `reasoning_effort`
|
||||
- image input through `image_url` data URLs or HTTP URLs
|
||||
- image input through `image_url` data URLs, HTTP URLs, `file://` URLs, and absolute local paths
|
||||
|
||||
### Anthropic Compatibility
|
||||
|
||||
@@ -133,7 +139,7 @@ If auto detection fails, set the path manually in the desktop Settings page or p
|
||||
|
||||
```bash
|
||||
lingma-ipc-proxy --transport websocket --ws-url ws://127.0.0.1:36510 --port 8095
|
||||
lingma-ipc-proxy --transport pipe --pipe-name '\\.\pipe\lingma-ipc'
|
||||
lingma-ipc-proxy --transport pipe --pipe '\\.\pipe\lingma-ipc'
|
||||
```
|
||||
|
||||
## Quick Start
|
||||
@@ -175,7 +181,7 @@ export ANTHROPIC_API_KEY="any"
|
||||
Then select a model in Claude Code:
|
||||
|
||||
```text
|
||||
/model Qwen3-Coder
|
||||
/model MiniMax-M2.7
|
||||
```
|
||||
|
||||
### Cline
|
||||
@@ -183,7 +189,7 @@ Then select a model in Claude Code:
|
||||
- Provider: `OpenAI Compatible`
|
||||
- Base URL: `http://127.0.0.1:8095/v1`
|
||||
- API Key: `any`
|
||||
- Model ID: `Qwen3-Coder`
|
||||
- Model ID: `MiniMax-M2.7`
|
||||
|
||||
### Continue
|
||||
|
||||
@@ -193,7 +199,7 @@ Then select a model in Claude Code:
|
||||
{
|
||||
"title": "Lingma Proxy",
|
||||
"provider": "openai",
|
||||
"model": "Qwen3-Coder",
|
||||
"model": "MiniMax-M2.7",
|
||||
"apiKey": "any",
|
||||
"apiBase": "http://127.0.0.1:8095/v1"
|
||||
}
|
||||
@@ -215,7 +221,19 @@ Observed model IDs include:
|
||||
- `Qwen3-Thinking`
|
||||
- `Qwen3.6-Plus`
|
||||
|
||||
For tool-heavy coding workflows, `Qwen3-Coder` is the recommended first choice.
|
||||
### Model Metadata and Recommendation
|
||||
|
||||
The proxy only reports models actually exposed by your Lingma plugin. The table below combines official model information where available with local proxy testing. If Lingma exposes a model name without public model-card metadata, the README marks it as observed rather than inventing a context length.
|
||||
|
||||
| Model | Best use | Context / capability basis |
|
||||
| --- | --- | --- |
|
||||
| `MiniMax-M2.7` | Default recommendation for third-party agents | NVIDIA's [MiniMax M2.7 model card](https://developer.nvidia.com/blog/minimax-m2-7-advances-scalable-agentic-workflows-on-nvidia-platforms-for-complex-ai-applications/) describes a language MoE model with 200K input context and agentic use cases; local proxy testing passed read/search/terminal/web/patch/vision smoke tests. |
|
||||
| `Kimi-K2.6` | Multimodal and long-context agent work | Kimi's [official API docs](https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart) describe native text/image/video input, a 256K context window, and multi-step tool invocation support. |
|
||||
| `Qwen3-Coder` | Code-specialized fallback | Qwen's [official blog](https://qwenlm.github.io/blog/qwen3-coder/) describes 256K native context, up to 1M with extrapolation, and agentic coding/tool protocols. |
|
||||
| `Qwen3.6-Plus` | General/vision fallback | Exposed by Lingma and passed local smoke tests, but this repository does not have an official Lingma-specific context-length source for it. |
|
||||
| `Qwen3-Max` | Fast general/vision model | Exposed by Lingma and strong in simple tests, but less stable on forced edit/read tool calls in this proxy. |
|
||||
|
||||
Default model when the client omits `model`: `MiniMax-M2.7`.
|
||||
|
||||
## Configuration
|
||||
|
||||
@@ -274,7 +292,14 @@ Lingma does not expose a native public OpenAI/Anthropic tool-call protocol, so t
|
||||
4. Convert parsed actions back into OpenAI `tool_calls` or Anthropic `tool_use`.
|
||||
5. Feed tool results back into Lingma for continuation.
|
||||
|
||||
This is most reliable with `Qwen3-Coder`.
|
||||
Current proxy hardening includes:
|
||||
|
||||
- a generated tool routing table based on the client's actual tool names
|
||||
- dedicated examples for `read_file`, `search_files`, `terminal`, and `web_search`
|
||||
- automatic retry when the model says it cannot access files, terminal, or web despite tools being present
|
||||
- common tool alias normalization such as `Bash` -> `terminal`, `Read` -> `read_file`, `Grep` -> `search_files`, and `Edit` -> `patch`
|
||||
|
||||
In local smoke tests after this hardening, `MiniMax-M2.7`, `Kimi-K2.6`, `Qwen3.6-Plus`, and `Qwen3-Coder` all completed read/search/terminal/web/patch/vision checks, with `MiniMax-M2.7` having the lowest average latency in the tested set.
|
||||
|
||||
## Local Desktop Build
|
||||
|
||||
@@ -306,7 +331,7 @@ The desktop bundle name is always `Lingma IPC Proxy`.
|
||||
|
||||
The release workflow is triggered by:
|
||||
|
||||
- pushing a tag such as `v1.2.2`
|
||||
- pushing a tag such as `v1.3.0`
|
||||
- manually running the `Release` workflow with a tag input
|
||||
|
||||
Planned improvements:
|
||||
|
||||
Reference in New Issue
Block a user