feat: add native Reasonix Skill + CLI integration

.github/workflows/ci.yml

@@ -22,14 +22,15 @@ jobs:
         with:
           python-version: "3.11"
       - name: Compile Python entry points
-        run: python -m py_compile vision_proxy.py vision_client.py ground.py detect.py bin/glance bin/trace bin/crop
+        run: python -m py_compile vision_proxy.py vision_client.py ground.py detect.py bin/glance bin/trace bin/crop integrations/reasonix/skills/vision-tools/scripts/vision_cli.py
       - name: Run core tests
         run: |
           python tests/test_image_rewrite_shapes.py
           python tests/test_focus_hint.py
           python tests/test_anthropic_rewrite.py
           python tests/smoke_test_proxy.py
           python tests/test_vision_client.py
+          python tests/test_reasonix_integration.py
 
   extensions:
     runs-on: ubuntu-latest

CHANGELOG.md

@@ -8,6 +8,7 @@ All notable user-facing changes to agent-vision-toolkit are documented in this f
 
 - Five vision CLIs — `glance`, `ground`, `detect`, `trace`, and `crop` — plus the `vision-tools` agent skill.
 - Optional seamless integration: a local proxy for Codex and Claude Code, and single-file native extensions for Pi, Oh My Pi, and OpenCode.
+- A compact native Reasonix 1.21.0 Skill + CLI adapter that reuses an existing toolkit checkout from `PATH`, with no MCP server or runtime sidecar.
 - Pasted-image and tool-fetched image support with task-aware focus hints, parallel multi-image descriptions, per-request caching, and honest failure notes.
 - Vision playbooks for long-screenshot OCR, UI restoration, graphic restoration, structure recovery, and GUI operation.
 - Community contribution, conduct, support, and security policies.

README.md

@@ -14,7 +14,7 @@
 [![Extensions](https://img.shields.io/badge/-Extensions-3178C6?style=flat-square)](https://github.com/Anionex/agent-vision-toolkit/tree/main/extensions)
 [![Shell](https://img.shields.io/badge/-Shell-4EAA25?style=flat-square&logo=gnubash&logoColor=white)](https://github.com/Anionex/agent-vision-toolkit/tree/main/bin)
 
-**What it thinks is what it sees — give any text-only coding agent eyes: image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation, as a vision toolkit plus a skill, with optional drop-in integration for Codex, Claude Code, Pi, Oh My Pi, and OpenCode.**
+**What it thinks is what it sees — give any text-only coding agent eyes: image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation, as a vision toolkit plus a skill, with optional drop-in integration for Codex, Claude Code, Pi, Oh My Pi, and OpenCode, plus a native Skill + CLI adapter for Reasonix.**
 
 🎯 An agent's vision capability doesn't have to live in the model — it can live in the harness.
 
@@ -28,7 +28,7 @@ This repository provides two kinds of components:
 1. **Vision tool CLIs** — multiple CLIs, plus a skill that teaches the agent when to use each one. Any agent that can invoke a shell can use them.
 2. **Seamless integration** *(optional upgrade)* — a transparent local proxy and single-file native plugins, so **images we paste and the agent's built-in image tools both work seamlessly**, with no extra tool installation or additional prompting.
 
-All code has been verified in real Codex + DeepSeek sessions, and the same pipeline has been live-verified end-to-end in Claude Code, Pi, Oh My Pi, and OpenCode.
+All code has been verified in real Codex + DeepSeek sessions, and the same pipeline has been live-verified end-to-end in Claude Code, Pi, Oh My Pi, and OpenCode. The native Reasonix Skill + CLI adapter has also been verified against Reasonix 1.21.0.
 
 > If this project helps you or gives you some inspiration, feel free to star🌟 & fork.
 
@@ -154,6 +154,12 @@ npx skills add Anionex/agent-vision-toolkit --skill vision-tools -a codex -g --c
 
 Or copy `skills/vision-tools/` into your agent's skills directory (e.g. `~/.codex/skills/`) and restart the agent.
 
+For **Reasonix**, keep the CLI checkout from step 2 on `PATH` and install the compact native adapter instead of the generic skill copy:
+
+```bash
+reasonix plugin install https://github.com/Anionex/agent-vision-toolkit/tree/main/integrations/reasonix --yes
+```
+
 </details>
 
 ## The Tools
@@ -266,10 +272,13 @@ This layer makes screenshots pasted into an agent work directly, while also prev
 | **Claude Code** | the same proxy — point `ANTHROPIC_BASE_URL` at it | ✅ verified |
 | **Pi / Oh My Pi** | one-file native extension ([`extensions/pi/`](extensions/pi/)) | ✅ verified |
 | **OpenCode** | one-file native plugin ([`extensions/opencode/`](extensions/opencode/)) | ✅ verified |
+| **Reasonix** | native Skill + existing CLIs ([`integrations/reasonix/`](integrations/reasonix/)); no MCP/runtime | ✅ verified (1.21.0) |
 | Any agent with a shell | the toolkit above — no integration needed | ✅ |
 
 All entry points share one configuration. Configure it once and use it everywhere.
 
+Reasonix intentionally uses Skill + CLI rather than MCP or a runtime sidecar: its text-only models receive a readable local path for pasted images, and the Skill routes that path through the existing toolkit checkout.
+
 ## How It Works
 
 ### Descriptions that keep the task in view

README_CN.md

@@ -14,7 +14,7 @@
 [![Extensions](https://img.shields.io/badge/-Extensions-3178C6?style=flat-square)](https://github.com/Anionex/agent-vision-toolkit/tree/main/extensions)
 [![Shell](https://img.shields.io/badge/-Shell-4EAA25?style=flat-square&logo=gnubash&logoColor=white)](https://github.com/Anionex/agent-vision-toolkit/tree/main/bin)
 
-**所想即所见——给任意纯文本 coding agent 装上眼睛：图片问答、长图OCR、前端UI还原、GUI自动化，一套视觉工具箱加一个 skill，并可选无缝接入 Codex、Claude Code、Pi、Oh My Pi、OpenCode。**
+**所想即所见——给任意纯文本 coding agent 装上眼睛：图片问答、长图OCR、前端UI还原、GUI自动化，一套视觉工具箱加一个 skill，并可选无缝接入 Codex、Claude Code、Pi、Oh My Pi、OpenCode，另提供 Reasonix 原生 Skill + CLI 适配。**
 
 🎯 agent视觉能力不一定长在模型上，也可以长在harness上。
 
@@ -28,7 +28,7 @@
 1. **视觉工具CLIs**：多个 CLI，外加一个 skill 告诉 agent 什么时候该用哪个。任何能调用 shell 的 agent 都能使用。
 2. **无缝接入**（可选升级）：透明本地代理与单文件原生插件，让**我们粘贴的图片和 agent 内置的看图工具也能完美工作**，不需要额外的工具安装，也不需要额外提示词。
 
-所有代码均已在真实 Codex + DeepSeek 会话中验证过，同一套管线也在 Claude Code、Pi、Oh My Pi、OpenCode 中完成了真机端到端验证。
+所有代码均已在真实 Codex + DeepSeek 会话中验证过，同一套管线也在 Claude Code、Pi、Oh My Pi、OpenCode 中完成了真机端到端验证；Reasonix 原生 Skill + CLI 适配也已在 Reasonix 1.21.0 上验证通过。
 
 > 如果项目对你有用或为你带来了一些启发，欢迎 star🌟 & fork。
 
@@ -154,6 +154,12 @@ npx skills add Anionex/agent-vision-toolkit --skill vision-tools -a codex -g --c
 
 也可以把 `skills/vision-tools/` 复制到你的 agent 的 skills 目录（如 `~/.codex/skills/`），重启生效。
 
+如果使用 **Reasonix**，保留第 2 步的 CLI checkout 并确保其 `bin/` 在 `PATH` 上，然后安装精简的原生适配包，无需再复制通用 skill：
+
+```bash
+reasonix plugin install https://github.com/Anionex/agent-vision-toolkit/tree/main/integrations/reasonix --yes
+```
+
 </details>
 
 ## 工具
@@ -264,10 +270,13 @@ crop screenshot.png --region 1563,514,1668,621 -o send-button.png
 | **Claude Code** | 同一个代理——把 `ANTHROPIC_BASE_URL` 指向它 | ✅ 已验证 |
 | **Pi / Oh My Pi** | 单文件原生 extension（[`extensions/pi/`](extensions/pi/)） | ✅ 已验证 |
 | **OpenCode** | 单文件原生 plugin（[`extensions/opencode/`](extensions/opencode/)） | ✅ 已验证 |
+| **Reasonix** | 原生 Skill + 现有 CLI（[`integrations/reasonix/`](integrations/reasonix/)），不使用 MCP/runtime | ✅ 已验证（1.21.0） |
 | 任何有 shell 的 agent | 上面的工具箱——无需接入 | ✅ |
 
 所有入口共享配置。一次配置好后即可多处使用。
 
+Reasonix 刻意采用 Skill + CLI，而不是 MCP 或 runtime sidecar：纯文本模型会拿到粘贴图片的可读本地路径，再由 Skill 把该路径交给现有工具箱 checkout 处理。
+
 ## 工作原理
 
 ### 让描述始终对着当前任务

integrations/reasonix/reasonix-plugin.json

@@ -0,0 +1,11 @@
+{
+  "apiVersion": "reasonix.io/plugin/v1",
+  "name": "agent-vision-toolkit",
+  "version": "0.1.0",
+  "description": "Reasonix skill adapter for the agent-vision-toolkit CLIs",
+  "contributes": {
+    "skills": [
+      "skills"
+    ]
+  }
+}

integrations/reasonix/skills/vision-tools/SKILL.md

@@ -0,0 +1,90 @@
+---
+name: vision-tools
+description: >-
+  Use the installed agent-vision-toolkit CLIs when a Reasonix task involves an
+  image: describe or OCR it, locate elements, inventory regions, recover exact
+  geometry, crop a box, compare renders, or inspect a long screenshot.
+---
+
+# vision-tools for Reasonix
+
+Reasonix keeps a pasted image's local path available to text-only models. Use
+the Skill's CLI dispatcher to inspect that path instead of guessing from the
+attachment marker.
+
+## Preflight the CLI checkout
+
+Reasonix appends a `## Scripts` section to this skill when it is loaded. Find
+the exact absolute path ending in `vision_cli.py` there, then verify that the
+toolkit checkout is available:
+
+```bash
+python3 "<vision_cli.py path from the Scripts section>" --check
+```
+
+Use the machine's Python 3 launcher (`python` or `py -3`) if it does not expose
+the interpreter as `python3`.
+
+The Reasonix plugin installs this Skill, not a second copy of the toolkit. The
+dispatcher finds the existing toolkit checkout through `glance` on `PATH`, or
+from the current working tree. If `--check` fails, clone agent-vision-toolkit
+and add its `bin/` directory to `PATH` using the repository's normal CLI setup,
+then retry. Do not recreate or vendor the vision commands inside a task.
+
+```bash
+git clone https://github.com/Anionex/agent-vision-toolkit.git
+export PATH="$PWD/agent-vision-toolkit/bin:$PATH"
+```
+
+Run every tool through the dispatcher:
+
+```bash
+python3 "<vision_cli.py path from the Scripts section>" <tool> [tool arguments]
+```
+
+For a Reasonix context block such as `<image path="...">`, pass the `path`
+value exactly as shown. If the user wrote an `@path` reference, pass the
+filesystem path without the leading `@`. Quote every filesystem path passed to
+the shell, including image, output, HTML, and dispatcher paths.
+
+## Choose the tool by the question
+
+| Question | Tool invocation |
+|---|---|
+| What does this image show or say? | `glance "<image>"` |
+| What matters for the current request? | `glance "<image>" -q "<current intent>"` |
+| Transcribe visible text exactly | `glance "<image>" --ocr` |
+| Where is one named target? | `ground "<image>" "<target>"` |
+| What elements or instances are present, and where? | `detect "<image>" [category]` |
+| What is the exact local shape or geometry? | `trace "<image>" [-o "output.svg"]` |
+| Cut a known pixel box into a reusable file | `crop "<image>" --region X1,Y1,X2,Y2 [-o "output"]` |
+| OCR a scrolling screenshot or chat history | `long-screenshot-ocr "<image>" [-o "output.md"]` |
+| Compare two renders at pixel level | `pixel-diff "<before>" "<after>"` |
+| Extract a foreground graphic | `extract-fg "<image>" ...` |
+| Inspect dominant colors | `dominant-colors "<image>" ...` |
+| Render HTML to an image | `html-shot "<file.html>" ...` |
+
+Run the dispatcher with `--list` to print every supported tool name. All
+arguments after the tool name are passed through unchanged, so each tool's
+normal `--help` remains available.
+
+## Operating rules
+
+1. When the user asks a specific visual question, pass that intent to
+   `glance -q`; do not settle for a generic description first.
+2. For multiple images that must be compared semantically, pass them to one
+   `glance` call so the vision model sees them together.
+3. Use `ground` for one named target and `detect` for an inventory. Their boxes
+   are suitable for clicking and cropping but can be a few pixels approximate.
+4. Use `trace` when shape, offset, or size must come from the actual pixels.
+5. Use `crop` and the bundled scripts instead of rewriting their pixel logic in
+   an ad hoc Python snippet.
+6. After changing a UI or graphic, render the result and inspect it again. Do
+   not claim visual completion from source code alone.
+7. Report missing files, optional dependencies, and vision API failures
+   honestly. Do not invent image contents when a command fails.
+
+The dispatcher inherits the shared `VISION_API_KEY`, `VISION_BASE_URL`,
+`VISION_MODEL`, and optional `LANG` configuration used by the rest of the
+toolkit. `glance` itself has no third-party Python dependency; other tools keep
+their existing optional dependency requirements.

integrations/reasonix/skills/vision-tools/scripts/vision_cli.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Dispatch Reasonix skill calls to the repository's existing vision CLIs."""
+
+from pathlib import Path
+import os
+import shutil
+import sys
+
+
+ENTRY_POINTS = {
+    "crop": "bin/crop",
+    "detect": "bin/detect",
+    "dominant-colors": "skills/vision-tools/scripts/dominant_colors.py",
+    "extract-fg": "skills/vision-tools/scripts/extract_fg.py",
+    "glance": "bin/glance",
+    "ground": "bin/ground",
+    "html-shot": "skills/vision-tools/scripts/html_shot.py",
+    "long-screenshot-ocr": "skills/vision-tools/scripts/long_screenshot_ocr.py",
+    "pixel-diff": "skills/vision-tools/scripts/pixel_diff.py",
+    "trace": "bin/trace",
+}
+
+
+def toolkit_root(glance_path=None, cwd=None):
+    """Find the toolkit checkout selected by PATH or the current workspace."""
+    roots = []
+    selected_glance = glance_path or shutil.which("glance")
+    if selected_glance:
+        roots.extend(Path(selected_glance).expanduser().resolve().parents)
+
+    current = Path.cwd() if cwd is None else Path(cwd)
+    current = current.expanduser().resolve()
+    roots.extend((current, *current.parents))
+
+    seen = set()
+    for candidate in roots:
+        key = str(candidate)
+        if key in seen:
+            continue
+        seen.add(key)
+        if (candidate / "vision_client.py").is_file() and (candidate / "bin" / "glance").is_file():
+            return candidate
+    raise RuntimeError(
+        "agent-vision-toolkit CLI checkout not found; add its bin directory to PATH"
+    )
+
+
+def normalize_tool(name):
+    return name.strip().lower().replace("_", "-")
+
+
+def resolve_target(name, glance_path=None, cwd=None):
+    tool = normalize_tool(name)
+    relative = ENTRY_POINTS.get(tool)
+    if relative is None:
+        raise KeyError(tool)
+    target = toolkit_root(glance_path=glance_path, cwd=cwd) / relative
+    if not target.is_file():
+        raise RuntimeError(f"toolkit entry point is missing: {relative}")
+    return target
+
+
+def print_usage(stream):
+    tools = ", ".join(sorted(ENTRY_POINTS))
+    print("usage: vision_cli.py <tool> [tool arguments] | --list | --check", file=stream)
+    print(f"tools: {tools}", file=stream)
+
+
+def main(argv=None):
+    args = list(sys.argv[1:] if argv is None else argv)
+    if args == ["--list"]:
+        print("\n".join(sorted(ENTRY_POINTS)))
+        return 0
+    if args == ["--check"]:
+        try:
+            root = toolkit_root()
+            for tool in ENTRY_POINTS:
+                resolve_target(tool)
+        except RuntimeError as exc:
+            print(f"vision toolkit error: {exc}", file=sys.stderr)
+            return 1
+        print(root)
+        return 0
+    if not args or args[0] in {"-h", "--help"}:
+        print_usage(sys.stdout if args else sys.stderr)
+        return 0 if args else 2
+
+    tool = normalize_tool(args.pop(0))
+    try:
+        target = resolve_target(tool)
+    except KeyError:
+        print(f"unknown vision tool: {tool}", file=sys.stderr)
+        print_usage(sys.stderr)
+        return 2
+    except RuntimeError as exc:
+        print(f"vision toolkit error: {exc}", file=sys.stderr)
+        return 1
+
+    try:
+        os.execv(sys.executable, [sys.executable, str(target), *args])
+    except OSError as exc:
+        print(f"vision toolkit error: cannot run {target}: {exc}", file=sys.stderr)
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())