Gitnexus Codebase Intelligence

gitnexus-codebase-intelligence

USE FOR:

• "understand codebase impact before editing"
• "find all callers / dependents of a symbol"
• "safe multi-file renaming"
• "git diff impact analysis"
• "MCP tool for codebase architecture understanding"
• "knowledge graph of code structure"
• "AI agent codebase context" tags: [MCP, codebase, knowledge-graph, Tree-sitter, symbol-search, impact-analysis, Claude-Code, Cursor, Cypher] kind: tool category: pro-code-architecture

What Is GitNexus?

Codebase intelligence platform — transforms repos into knowledge graphs for AI agents.

• Repo: https://github.com/abhigyanpatwari/GitNexus
• Web: gitnexus.vercel.app (no install)
• Integrations: Claude Code, Cursor, Windsurf, any MCP-compatible editor
• Privacy: 100% local — code never leaves your machine

"Building a nervous system for agent context"

Core Problem Solved

Traditional AI assistants don't know when edits break downstream dependencies. GitNexus precomputes architectural intelligence at index time → fast, accurate impact analysis at query time.

7 MCP Tools

Indexing Pipeline (6 Phases)

1. Structural mapping   → directory tree, file relationships
2. AST parsing          → Tree-sitter extracts symbols per language
3. Import/call resolve  → link usages to definitions
4. Community cluster    → group related modules
5. Execution flow trace → call chains, data flow paths
6. Hybrid search index  → BM25 + vector embeddings

Installation & Setup

# Index a repository
npx gitnexus analyze

# Configure MCP for your editor (one-time, multi-project)
npx gitnexus setup
# → adds GitNexus MCP server to Claude Code / Cursor / Windsurf config

Web UI (no install):

https://gitnexus.vercel.app
# Upload or link repo → explore in browser

Claude Code Integration

After npx gitnexus setup, Claude Code gets access to all 7 MCP tools:

# In Claude Code session:
> "What calls the processOrder function?"
→ Claude uses symbol_context MCP tool → instant impact map

> "I'm about to rename UserService — what breaks?"
→ Claude uses impact_radius → lists all 47 dependent symbols

> "Show me what this git diff affects"
→ Claude uses git_diff_impact → maps changed lines to affected call chains

Supported Languages

Example: Impact Radius Query

-- Raw Cypher query via graph_query tool
MATCH (s:Symbol {name: "UserService"})<-[:CALLS|IMPORTS*1..3]-(dep:Symbol)
RETURN dep.name, dep.file, dep.line
ORDER BY dep.file

Returns every symbol within 3 hops that depends on UserService.

KNOWLEDGE INJECTION: OpenCV

Source: https://github.com/opencv/opencv

Routed to: development.md

Date: 2026-03-18

SKILL: opencv

name: opencv description: > OpenCV - Open Source Computer Vision Library. 86k stars, 14 modules. imgproc (filtering/contours/warp), dnn (YOLO/ONNX inference), features2d (SIFT/ORB/AKAZE matching), objdetect (Haar/HOG/QR), calib3d, tracking (KCF/CSRT). pip install opencv-contrib-python. C++ and Python. USE FOR:

• image processing pipeline
• contour detection and perspective warp
• YOLO object detection with OpenCV dnn
• feature matching SIFT ORB AKAZE
• face detection Haar cascade
• camera calibration undistort
• object tracking KCF CSRT
• chess board detection
• color segmentation HSV mask
• background subtraction optical flow tags: [OpenCV, computer-vision, image-processing, DNN, YOLO, SIFT, ORB, contours, tracking, Python, C++] kind: library category: programmatic-drawing

What Is OpenCV?

Open Source Computer Vision Library.

• Repo: https://github.com/opencv/opencv
• Stars: 86.6k | Forks: 56.6k | Contributors: 1,775+
• Languages: C++ (87%) with Python, Java, JS bindings
• Docs: https://docs.opencv.org/4.x/

Installation

pip install opencv-contrib-python        # recommended (includes SIFT, tracking)
pip install opencv-contrib-python-headless  # no GUI (servers)

import cv2
print(cv2.__version__)   # e.g. 4.9.0

Module Map

Core: Load, Convert, Save

import cv2, numpy as np

img  = cv2.imread("image.jpg")               # BGR uint8
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
hsv  = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
rgb  = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # for matplotlib
h, w, c = img.shape
roi  = img[y1:y2, x1:x2]                     # crop
cv2.imwrite("out.jpg", img)
cv2.imshow("win", img); cv2.waitKey(0)

imgproc: Filters & Edges

blur   = cv2.GaussianBlur(gray, (5,5), 0)
median = cv2.medianBlur(gray, 5)               # salt-and-pepper
bilat  = cv2.bilateralFilter(img, 9, 75, 75)   # edge-preserving

edges  = cv2.Canny(blur, 50, 150)

kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5))
opened = cv2.morphologyEx(binary, cv2.MORPH_OPEN,  kernel)  # remove noise
closed = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)  # fill holes

imgproc: Thresholding

_, binary  = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
_, otsu    = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
adaptive   = cv2.adaptiveThreshold(gray, 255,
               cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2)

# Color range mask
lower = np.array([100, 50, 50])
upper = np.array([130, 255, 255])
mask  = cv2.inRange(hsv, lower, upper)

imgproc: Contours

cnts, hier = cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

for cnt in cnts:
    area = cv2.contourArea(cnt)
    if area < 500: continue

    peri   = cv2.arcLength(cnt, True)
    approx = cv2.approxPolyDP(cnt, 0.02*peri, True)
    x,y,w,h = cv2.boundingRect(cnt)

    M  = cv2.moments(cnt)
    cx = int(M["m10"] / M["m00"])   # centroid
    cy = int(M["m01"] / M["m00"])

    cv2.drawContours(img, [cnt], 0, (0,255,0), 2)

# Shape by vertex count
n = len(approx)
if   n == 3: shape = "triangle"
elif n == 4: shape = "quad/rect"
elif n == 5: shape = "pentagon"
else:        shape = "circle"

imgproc: Perspective Warp

src = np.float32([[tl_x,tl_y],[tr_x,tr_y],[bl_x,bl_y],[br_x,br_y]])
dst = np.float32([[0,0],[W,0],[0,H],[W,H]])
M      = cv2.getPerspectiveTransform(src, dst)
warped = cv2.warpPerspective(img, M, (W, H))

features2d: SIFT / ORB Matching

sift  = cv2.SIFT_create()
orb   = cv2.ORB_create(nfeatures=1500)

kp1, des1 = sift.detectAndCompute(img1, None)
kp2, des2 = sift.detectAndCompute(img2, None)

bf      = cv2.BFMatcher(cv2.NORM_L2)           # L2 for SIFT
# bf   = cv2.BFMatcher(cv2.NORM_HAMMING)        # Hamming for ORB/AKAZE
matches = bf.knnMatch(des1, des2, k=2)
good    = [m for m,n in matches if m.distance < 0.75*n.distance]

if len(good) > 10:
    src_pts = np.float32([kp1[m.queryIdx].pt for m in good]).reshape(-1,1,2)
    dst_pts = np.float32([kp2[m.trainIdx].pt for m in good]).reshape(-1,1,2)
    H, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)

objdetect: Face & QR

face_cascade = cv2.CascadeClassifier(
    cv2.data.haarcascades + "haarcascade_frontalface_alt.xml")

faces = face_cascade.detectMultiScale(gray, 1.1, 5, minSize=(30,30))
for (x,y,w,h) in faces:
    cv2.rectangle(img, (x,y), (x+w,y+h), (255,0,0), 2)

qr = cv2.QRCodeDetector()
data, pts, _ = qr.detectAndDecode(img)

dnn: YOLO / ONNX Inference

net = cv2.dnn.readNetFromONNX("yolov8n.onnx")
# Optional GPU: net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)

blob = cv2.dnn.blobFromImage(img, 1/255, (640,640), swapRB=True)
net.setInput(blob)
outputs = net.forward(net.getUnconnectedOutLayersNames())

for det in outputs[0]:
    scores     = det[5:]
    class_id   = np.argmax(scores)
    confidence = scores[class_id]
    if confidence > 0.5:
        cx,cy,bw,bh = (det[:4] * np.array([W,H,W,H])).astype(int)
        cv2.rectangle(img, (cx-bw//2, cy-bh//2), (cx+bw//2, cy+bh//2), (0,255,0), 2)

Supported: ONNX | TensorFlow .pb | Caffe | Darknet YOLO | OpenVINO IR

tracking: Object Trackers

tracker = cv2.TrackerCSRT_create()    # best accuracy
# tracker = cv2.TrackerKCF_create()   # balanced
# tracker = cv2.TrackerMOSSE_create() # fastest

ok = tracker.init(frame, (x, y, w, h))
while cap.isOpened():
    ok, frame = cap.read()
    ok, bbox  = tracker.update(frame)
    if ok:
        x,y,w,h = [int(v) for v in bbox]
        cv2.rectangle(frame, (x,y), (x+w,y+h), (0,255,0), 2)

calib3d: Camera Calibration

objp = np.zeros((6*9,3), np.float32)
objp[:,:2] = np.mgrid[0:9,0:6].T.reshape(-1,2)

objpts, imgpts = [], []
for img in calib_images:
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    ret, corners = cv2.findChessboardCorners(gray, (9,6))
    if ret:
        objpts.append(objp)
        imgpts.append(cv2.cornerSubPix(gray, corners, (11,11), (-1,-1),
            (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 30, 0.001)))

ret, mtx, dist, rvecs, tvecs = cv2.calibrateCamera(objpts, imgpts, gray.shape[::-1], None, None)
undist = cv2.undistort(frame, mtx, dist)

video: Background Subtraction & Optical Flow

# Background subtraction
fgbg  = cv2.createBackgroundSubtractorMOG2(history=500, detectShadows=True)
fgmask = fgbg.apply(frame)

# Sparse optical flow (Lucas-Kanade)
p0     = cv2.goodFeaturesToTrack(prev_gray, 100, 0.3, 7)
p1, st, _ = cv2.calcOpticalFlowPyrLK(prev_gray, gray, p0, None)
good_new = p1[st==1]

# Dense optical flow
flow   = cv2.calcOpticalFlowFarneback(prev_gray, gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1])

Drawing Functions

cv2.line(img, (x1,y1), (x2,y2), (B,G,R), thickness)
cv2.rectangle(img, (x1,y1), (x2,y2), color, thickness)   # -1 = filled
cv2.circle(img, (cx,cy), radius, color, thickness)
cv2.putText(img, "text", (x,y), cv2.FONT_HERSHEY_SIMPLEX, scale, color, thickness)
cv2.polylines(img, [pts], isClosed=True, color, thickness)
cv2.arrowedLine(img, pt1, pt2, color, thickness)

Chess Board Pipeline (project context)

# Used in chess/ desktop app (board_detector.py)
gray    = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY)
blur    = cv2.GaussianBlur(gray, (5,5), 0)
edges   = cv2.Canny(blur, 50, 150)
cnts, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
board   = max(cnts, key=cv2.contourArea)   # largest contour = board
approx  = cv2.approxPolyDP(board, 0.02*cv2.arcLength(board,True), True)
# approx should have 4 points for the board corners
warped  = cv2.warpPerspective(screen, M, (512, 512))
# Then split warped into 8x8 grid -> classify each square

KNOWLEDGE INJECTION: Orange3

Source: https://github.com/biolab/orange3

Routed to: development.md

Date: 2026-03-18

SKILL: orange3-data-mining

name: orange3-data-mining description: > Orange3 - open-source visual data mining and ML platform by Univ. of Ljubljana. No-code canvas-based workflow: connect widgets for data loading, preprocessing, visualization, classification, clustering, regression. Add-ons: text, time series, bioinformatics, image analytics, geo. pip/conda/winget install. USE FOR:

• visual ML workflow no-code
• data mining and visualization
• interactive scatter plot histogram heatmap
• decision tree random forest clustering
• exploratory data analysis tags: [Orange3, visual-ML, data-mining, no-code, clustering, classification, regression] kind: tool category: pro-code-architecture

What Is Orange3?

Visual data mining and ML toolbox — no programming required.

• Repo: https://github.com/biolab/orange3
• By: University of Ljubljana Biolab
• Approach: Canvas-based drag-and-drop widget workflow

Installation

winget install --id UniversityofLjubljana.Orange
conda install orange3
pip install orange3   # requires PyQt6 first

Widget Categories

• Data: File, SQL, URL, Data Table, Select Rows, Feature Constructor
• Visualize: Scatter Plot, Box Plot, Distributions, Heatmap, Mosaic
• Model: Decision Tree, Random Forest, SVM, Naive Bayes, Neural Network, Logistic Regression
• Evaluate: Test & Score, Confusion Matrix, ROC, Calibration Plot
• Unsupervised: K-Means, Hierarchical Clustering, PCA, t-SNE, DBSCAN
• Text: Corpus, Preprocess Text, Bag of Words, Topic Modelling
• Time Series: As Timeseries, Moving Transform, Seasonal Adjustment

Workflow Pattern

File → Preprocess → Select Features
                         → Train/Test Split → Random Forest → Evaluate
                         → PCA → Scatter Plot (colored by class)

KNOWLEDGE INJECTION: xonsh

Source: https://github.com/xonsh/xonsh

Routed to: development.md

Date: 2026-03-18

SKILL: xonsh-python-shell

name: xonsh-python-shell description: > xonsh - Python-powered cross-platform shell. Superset of Python 3: mix shell commands and Python code in the same session. Direct manipulation of command outputs as Python objects. Extension system (xontribs/plugins). AI-friendly. Used by conda, mamba, Snakemake, Jupyter. pip install xonsh. USE FOR:

• Python + shell hybrid scripting
• manipulate command output as Python objects
• automate shell tasks with Python logic
• cross-platform shell scripting
• conda mamba workflow automation tags: [xonsh, Python-shell, hybrid, cross-platform, scripting, automation, xontribs] kind: tool category: pro-code-architecture

What Is xonsh?

Python-powered shell — superset of Python 3, cross-platform, AI-friendly.

• Repo: https://github.com/xonsh/xonsh

Install

pip install xonsh
# or via conda:
conda install -c conda-forge xonsh

Syntax Examples

# Standard shell commands
cd $HOME
cat /etc/passwd | grep root

# Python variables in commands
name = "snail"
echo @(name) > /tmp/@(name)

# Python expressions
var = "hello".upper()
echo @(var)

# Command output as Python object
result = $(ls -la).split("\n")
print(len(result))

# Mixed pipeline
len($(curl -L https://xon.sh))

# Subprocess mode with Python control flow
for f in $(find . -name "*.py").split():
    print(f)

xontribs (Extensions)

xpip install xontrib-vox          # virtual environment manager
xpip install xontrib-prompt-ret   # return code in prompt
xpip install xontrib-autojump     # autojump integration

Use Cases

• conda/mamba package management scripts
• Snakemake bioinformatics workflows
• Jupyter-based interactive computing
• DevOps automation with Python logic
• Data pipeline scripting

KNOWLEDGE INJECTION: Orange3

Source: https://github.com/biolab/orange3

Routed to: development.md

Date: 2026-03-18

SKILL: orange3-data-mining

name: orange3-data-mining description: > Orange3 - open-source visual data mining and ML platform by Univ. of Ljubljana. No-code canvas-based workflow: connect widgets for data loading, preprocessing, visualization, classification, clustering, regression. Add-ons: text, time series, bioinformatics, image analytics, geo. pip/conda/winget install. USE FOR:

• visual ML workflow no-code
• data mining and visualization
• interactive scatter plot histogram heatmap
• decision tree random forest clustering
• exploratory data analysis tags: [Orange3, visual-ML, data-mining, no-code, clustering, classification, regression] kind: tool category: pro-code-architecture

What Is Orange3?

Visual data mining and ML toolbox — no programming required.

• Repo: https://github.com/biolab/orange3
• By: University of Ljubljana Biolab
• Approach: Canvas-based drag-and-drop widget workflow

Installation

winget install --id UniversityofLjubljana.Orange
conda install orange3
pip install orange3   # requires PyQt6 first

Widget Categories

• Data: File, SQL, URL, Data Table, Select Rows, Feature Constructor
• Visualize: Scatter Plot, Box Plot, Distributions, Heatmap, Mosaic
• Model: Decision Tree, Random Forest, SVM, Naive Bayes, Neural Network, Logistic Regression
• Evaluate: Test & Score, Confusion Matrix, ROC, Calibration Plot
• Unsupervised: K-Means, Hierarchical Clustering, PCA, t-SNE, DBSCAN
• Text: Corpus, Preprocess Text, Bag of Words, Topic Modelling
• Time Series: As Timeseries, Moving Transform, Seasonal Adjustment

Workflow Pattern

File → Preprocess → Select Features
                         → Train/Test Split → Random Forest → Evaluate
                         → PCA → Scatter Plot (colored by class)

KNOWLEDGE INJECTION: OpenCV

Source: https://github.com/opencv/opencv

Routed to: development.md

Date: 2026-03-18

SKILL: opencv

name: opencv description: > OpenCV - Open Source Computer Vision Library. 86k stars, 14 modules. imgproc (filtering/contours/warp), dnn (YOLO/ONNX inference), features2d (SIFT/ORB/AKAZE matching), objdetect (Haar/HOG/QR), calib3d, tracking (KCF/CSRT). pip install opencv-contrib-python. C++ and Python. USE FOR:

• image processing pipeline
• contour detection and perspective warp
• YOLO object detection with OpenCV dnn
• feature matching SIFT ORB AKAZE
• face detection Haar cascade
• camera calibration undistort
• object tracking KCF CSRT
• chess board detection
• color segmentation HSV mask
• background subtraction optical flow tags: [OpenCV, computer-vision, image-processing, DNN, YOLO, SIFT, ORB, contours, tracking, Python, C++] kind: library category: programmatic-drawing

What Is OpenCV?

Open Source Computer Vision Library.

• Repo: https://github.com/opencv/opencv
• Stars: 86.6k | Forks: 56.6k | Contributors: 1,775+
• Languages: C++ (87%) with Python, Java, JS bindings
• Docs: https://docs.opencv.org/4.x/

Installation

pip install opencv-contrib-python        # recommended (includes SIFT, tracking)
pip install opencv-contrib-python-headless  # no GUI (servers)

import cv2
print(cv2.__version__)   # e.g. 4.9.0

Module Map

Core: Load, Convert, Save

import cv2, numpy as np

img  = cv2.imread("image.jpg")               # BGR uint8
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
hsv  = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
rgb  = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # for matplotlib
h, w, c = img.shape
roi  = img[y1:y2, x1:x2]                     # crop
cv2.imwrite("out.jpg", img)
cv2.imshow("win", img); cv2.waitKey(0)

imgproc: Filters & Edges

blur   = cv2.GaussianBlur(gray, (5,5), 0)
median = cv2.medianBlur(gray, 5)               # salt-and-pepper
bilat  = cv2.bilateralFilter(img, 9, 75, 75)   # edge-preserving

edges  = cv2.Canny(blur, 50, 150)

kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5))
opened = cv2.morphologyEx(binary, cv2.MORPH_OPEN,  kernel)  # remove noise
closed = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)  # fill holes

imgproc: Thresholding

_, binary  = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
_, otsu    = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
adaptive   = cv2.adaptiveThreshold(gray, 255,
               cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2)

# Color range mask
lower = np.array([100, 50, 50])
upper = np.array([130, 255, 255])
mask  = cv2.inRange(hsv, lower, upper)

imgproc: Contours

cnts, hier = cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

for cnt in cnts:
    area = cv2.contourArea(cnt)
    if area < 500: continue

    peri   = cv2.arcLength(cnt, True)
    approx = cv2.approxPolyDP(cnt, 0.02*peri, True)
    x,y,w,h = cv2.boundingRect(cnt)

    M  = cv2.moments(cnt)
    cx = int(M["m10"] / M["m00"])   # centroid
    cy = int(M["m01"] / M["m00"])

    cv2.drawContours(img, [cnt], 0, (0,255,0), 2)

# Shape by vertex count
n = len(approx)
if   n == 3: shape = "triangle"
elif n == 4: shape = "quad/rect"
elif n == 5: shape = "pentagon"
else:        shape = "circle"

imgproc: Perspective Warp

src = np.float32([[tl_x,tl_y],[tr_x,tr_y],[bl_x,bl_y],[br_x,br_y]])
dst = np.float32([[0,0],[W,0],[0,H],[W,H]])
M      = cv2.getPerspectiveTransform(src, dst)
warped = cv2.warpPerspective(img, M, (W, H))

features2d: SIFT / ORB Matching

sift  = cv2.SIFT_create()
orb   = cv2.ORB_create(nfeatures=1500)

kp1, des1 = sift.detectAndCompute(img1, None)
kp2, des2 = sift.detectAndCompute(img2, None)

bf      = cv2.BFMatcher(cv2.NORM_L2)           # L2 for SIFT
# bf   = cv2.BFMatcher(cv2.NORM_HAMMING)        # Hamming for ORB/AKAZE
matches = bf.knnMatch(des1, des2, k=2)
good    = [m for m,n in matches if m.distance < 0.75*n.distance]

if len(good) > 10:
    src_pts = np.float32([kp1[m.queryIdx].pt for m in good]).reshape(-1,1,2)
    dst_pts = np.float32([kp2[m.trainIdx].pt for m in good]).reshape(-1,1,2)
    H, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)

objdetect: Face & QR

face_cascade = cv2.CascadeClassifier(
    cv2.data.haarcascades + "haarcascade_frontalface_alt.xml")

faces = face_cascade.detectMultiScale(gray, 1.1, 5, minSize=(30,30))
for (x,y,w,h) in faces:
    cv2.rectangle(img, (x,y), (x+w,y+h), (255,0,0), 2)

qr = cv2.QRCodeDetector()
data, pts, _ = qr.detectAndDecode(img)

dnn: YOLO / ONNX Inference

net = cv2.dnn.readNetFromONNX("yolov8n.onnx")
# Optional GPU: net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)

blob = cv2.dnn.blobFromImage(img, 1/255, (640,640), swapRB=True)
net.setInput(blob)
outputs = net.forward(net.getUnconnectedOutLayersNames())

for det in outputs[0]:
    scores     = det[5:]
    class_id   = np.argmax(scores)
    confidence = scores[class_id]
    if confidence > 0.5:
        cx,cy,bw,bh = (det[:4] * np.array([W,H,W,H])).astype(int)
        cv2.rectangle(img, (cx-bw//2, cy-bh//2), (cx+bw//2, cy+bh//2), (0,255,0), 2)

Supported: ONNX | TensorFlow .pb | Caffe | Darknet YOLO | OpenVINO IR

tracking: Object Trackers

tracker = cv2.TrackerCSRT_create()    # best accuracy
# tracker = cv2.TrackerKCF_create()   # balanced
# tracker = cv2.TrackerMOSSE_create() # fastest

ok = tracker.init(frame, (x, y, w, h))
while cap.isOpened():
    ok, frame = cap.read()
    ok, bbox  = tracker.update(frame)
    if ok:
        x,y,w,h = [int(v) for v in bbox]
        cv2.rectangle(frame, (x,y), (x+w,y+h), (0,255,0), 2)

calib3d: Camera Calibration

objp = np.zeros((6*9,3), np.float32)
objp[:,:2] = np.mgrid[0:9,0:6].T.reshape(-1,2)

objpts, imgpts = [], []
for img in calib_images:
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    ret, corners = cv2.findChessboardCorners(gray, (9,6))
    if ret:
        objpts.append(objp)
        imgpts.append(cv2.cornerSubPix(gray, corners, (11,11), (-1,-1),
            (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 30, 0.001)))

ret, mtx, dist, rvecs, tvecs = cv2.calibrateCamera(objpts, imgpts, gray.shape[::-1], None, None)
undist = cv2.undistort(frame, mtx, dist)

video: Background Subtraction & Optical Flow

# Background subtraction
fgbg  = cv2.createBackgroundSubtractorMOG2(history=500, detectShadows=True)
fgmask = fgbg.apply(frame)

# Sparse optical flow (Lucas-Kanade)
p0     = cv2.goodFeaturesToTrack(prev_gray, 100, 0.3, 7)
p1, st, _ = cv2.calcOpticalFlowPyrLK(prev_gray, gray, p0, None)
good_new = p1[st==1]

# Dense optical flow
flow   = cv2.calcOpticalFlowFarneback(prev_gray, gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1])

Drawing Functions

cv2.line(img, (x1,y1), (x2,y2), (B,G,R), thickness)
cv2.rectangle(img, (x1,y1), (x2,y2), color, thickness)   # -1 = filled
cv2.circle(img, (cx,cy), radius, color, thickness)
cv2.putText(img, "text", (x,y), cv2.FONT_HERSHEY_SIMPLEX, scale, color, thickness)
cv2.polylines(img, [pts], isClosed=True, color, thickness)
cv2.arrowedLine(img, pt1, pt2, color, thickness)

Chess Board Pipeline (project context)

# Used in chess/ desktop app (board_detector.py)
gray    = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY)
blur    = cv2.GaussianBlur(gray, (5,5), 0)
edges   = cv2.Canny(blur, 50, 150)
cnts, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
board   = max(cnts, key=cv2.contourArea)   # largest contour = board
approx  = cv2.approxPolyDP(board, 0.02*cv2.arcLength(board,True), True)
# approx should have 4 points for the board corners
warped  = cv2.warpPerspective(screen, M, (512, 512))
# Then split warped into 8x8 grid -> classify each square

KNOWLEDGE INJECTION: xonsh

Source: https://github.com/xonsh/xonsh

Routed to: development.md

Date: 2026-03-18

SKILL: xonsh-python-shell

name: xonsh-python-shell description: > xonsh - Python-powered cross-platform shell. Superset of Python 3: mix shell commands and Python code in the same session. Direct manipulation of command outputs as Python objects. Extension system (xontribs/plugins). AI-friendly. Used by conda, mamba, Snakemake, Jupyter. pip install xonsh. USE FOR:

• Python + shell hybrid scripting
• manipulate command output as Python objects
• automate shell tasks with Python logic
• cross-platform shell scripting
• conda mamba workflow automation tags: [xonsh, Python-shell, hybrid, cross-platform, scripting, automation, xontribs] kind: tool category: pro-code-architecture

What Is xonsh?

Python-powered shell — superset of Python 3, cross-platform, AI-friendly.

• Repo: https://github.com/xonsh/xonsh

Install

pip install xonsh
# or via conda:
conda install -c conda-forge xonsh

Syntax Examples

# Standard shell commands
cd $HOME
cat /etc/passwd | grep root

# Python variables in commands
name = "snail"
echo @(name) > /tmp/@(name)

# Python expressions
var = "hello".upper()
echo @(var)

# Command output as Python object
result = $(ls -la).split("\n")
print(len(result))

# Mixed pipeline
len($(curl -L https://xon.sh))

# Subprocess mode with Python control flow
for f in $(find . -name "*.py").split():
    print(f)

xontribs (Extensions)

xpip install xontrib-vox          # virtual environment manager
xpip install xontrib-prompt-ret   # return code in prompt
xpip install xontrib-autojump     # autojump integration

Use Cases

• conda/mamba package management scripts
• Snakemake bioinformatics workflows
• Jupyter-based interactive computing
• DevOps automation with Python logic
• Data pipeline scripting