Skip to content

update.md

Latest commit

 

History

History
314 lines (231 loc) · 6.96 KB

update.md

File metadata and controls

314 lines (231 loc) · 6.96 KB

Updating CLIProxyAPI from Upstream

This guide documents how to update your fork from the upstream repository while preserving local changes.

Repository Setup

origin    -> https://github.com/yinkev/CLIProxyAPI.git (your fork)
upstream  -> https://github.com/router-for-me/CLIProxyAPI.git (main repo)

Update Process

1. Fetch Latest Tags and Commits

git fetch --tags upstream

2. Check for Local Changes

git status
git diff --stat

3. Stash Local Changes (if any)

If you have uncommitted changes that would conflict with the merge:

git stash push -m "Local changes before update"

4. Merge Upstream

git merge upstream/main --no-edit

5. Reapply Your Changes

git stash pop

If there are conflicts, resolve them manually, then:

git add <resolved-files>

6. Verify the Update

git log --oneline -3
git describe --tags --abbrev=0

Example: Update to v6.7.9

# Fetch latest
git fetch --tags upstream

# Stash local work
git stash push -m "Local changes before v6.7.9 update"

# Merge upstream
git merge upstream/main --no-edit

# Reapply local changes
git stash pop

# Verify
git describe --tags --abbrev=0  # Should show v6.7.9

What We Updated (v6.7.9)

Date: January 18, 2026

Changes merged:

  • 36 files modified
  • 2,694 lines added / 660 lines removed

Key features:

  • Reasoning state tracking and summary handling improvements
  • Gemini API compatibility fix (enum values to strings)
  • Thinking validation with fromSuffix parameter and budget logic
  • Token filename prefix refactoring for Qwen/iFlow
  • README updates (ZeroLimit project added)

Local changes preserved:

  • internal/api/handlers/management/auth_files.go
  • internal/api/server.go
  • internal/registry/model_definitions.go

Contributing Back to Upstream

IMPORTANT: Do NOT create a PR from your fork's main branch!

Your fork's main branch contains your entire commit history, including local customizations. If you PR from there, you'll submit ALL your changes, not just the feature you want to contribute.

Correct Workflow: Cherry-Pick

# 1. Create a clean branch from upstream
git fetch upstream
git checkout -b feat/my-feature upstream/main

# 2. Make your changes and commit
# ... edit files ...
git add <files>
git commit -m "feat: add my feature"

# 3. Push to your fork
git push -u origin feat/my-feature

# 4. Create PR from this clean branch
gh pr create --repo router-for-me/CLIProxyAPI

If You Already Committed to Your Fork's Main

# 1. Note your commit hash
git log --oneline -5  # Find your commit, e.g., abc1234

# 2. Create clean branch from upstream
git checkout -b feat/my-feature-clean upstream/main

# 3. Cherry-pick ONLY your specific commit
git cherry-pick abc1234

# 4. Push and create PR
git push -u origin feat/my-feature-clean
gh pr create --repo router-for-me/CLIProxyAPI

What We Learned (v6.7.31)

Bad: Created PR #1312 from fork branch → 1824 additions, 14 files (included entire fork history)

Good: Created PR #1313 from cherry-picked branch → 38 additions, 1 file (only our feature)

PRs Submitted

PR Feature Status
#1313 GetAllStaticModels() + LookupStaticModelInfo consistency Pending review
#1317 code_execution + url_context tool passthrough Pending review

New Features Added (v6.7.31+)

Gemini Tools Passthrough

Added support for Gemini's built-in tools:

{
  "tools": [
    {"google_search": {}},
    {"code_execution": {}},
    {"url_context": {}}
  ]
}
  • google_search - Grounding with Google Search (already in upstream)
  • code_execution - Agentic Vision, Python execution (Issue #1318)
  • url_context - Live URL content fetching (Issue #1318)

Using Agentic Vision

Gemini 3 Flash's Agentic Vision uses code execution to analyze images with Python. The model can draw bounding boxes, annotations, count objects, and manipulate images.

Requirements

  1. Model: gemini-3-flash-preview or gemini-2.5-flash
  2. Tool: code_execution enabled
  3. Input: An image (base64 or URL)

Basic Request

curl http://localhost:8317/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-proxy" \
  -d '{
    "model": "gemini-3-flash-preview",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Count the objects and draw bounding boxes around each one."},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,YOUR_BASE64_HERE"}}
      ]
    }],
    "tools": [{"code_execution": {}}],
    "max_tokens": 4000
  }'

Example Use Cases

1. Object Detection with Bounding Boxes

{
  "content": [
    {"type": "text", "text": "Identify all objects and draw bounding boxes with labels."},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
  ],
  "tools": [{"code_execution": {}}]
}

2. Image Occlusion for Studying

{
  "content": [
    {"type": "text", "text": "Create image occlusion by drawing red boxes over key terms and answers for studying."},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
  ],
  "tools": [{"code_execution": {}}]
}

3. Chart/Graph Analysis

{
  "content": [
    {"type": "text", "text": "Extract data from this chart using code execution."},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
  ],
  "tools": [{"code_execution": {}}]
}

Response Format

The model returns Python code that would annotate/analyze the image:

import PIL.Image
import PIL.ImageDraw

img = PIL.Image.open('input_file_0.png')
draw = PIL.ImageDraw.Draw(img)

# Bounding boxes with normalized coordinates [ymin, xmin, ymax, xmax]
boxes = [
    {'box_2d': [100, 80, 200, 300], 'label': 'object1'},
    {'box_2d': [250, 150, 400, 350], 'label': 'object2'},
]

for box in boxes:
    ymin, xmin, ymax, xmax = box['box_2d']
    # Convert normalized (0-1000) to pixel coordinates
    left = xmin * width / 1000
    top = ymin * height / 1000
    right = xmax * width / 1000
    bottom = ymax * height / 1000
    draw.rectangle([left, top, right, bottom], outline='red', width=3)

img.save('transformed_image.png')

Available Python Libraries

  • PIL (Pillow) - Image manipulation
  • numpy - Numerical operations
  • pandas - Data analysis
  • matplotlib - Plotting

Limitations

  • 30 second timeout per code execution
  • Up to 5 executions per request
  • No file I/O beyond input/output images
  • No custom library installation

Combining Tools

You can use multiple tools together:

{
  "tools": [
    {"code_execution": {}},
    {"google_search": {}},
    {"url_context": {}}
  ]
}

URL Context Example

Fetch and analyze web content:

{
  "messages": [{
    "role": "user",
    "content": "Summarize the content at https://example.com/article"
  }],
  "tools": [{"url_context": {}}]
}