@yeongjaeyou/claude-code-config 0.20.0 → 0.21.1

package/.claude/commands/gh/resolve-issue.md

@@ -6,6 +6,12 @@ description: Resolve GitHub Issue
 
 Act as an expert developer who systematically analyzes and resolves GitHub issues. Receive a GitHub issue number as argument and resolve the issue. Follow project guidelines in `@CLAUDE.md`.
 
+## Prerequisites
+
+Before starting the workflow:
+- **Serena MCP**: If not already active, run `activate_project` to enable semantic code analysis tools
+- **Clean state**: Ensure no uncommitted changes that could conflict with the new branch
+
 ## Workflow
 
 1. **Analyze Issue**:
@@ -44,18 +50,30 @@ Act as an expert developer who systematically analyzes and resolves GitHub issue
        ```
      - Skip if Status field does not exist
 
-5. **Analyze Codebase (MANDATORY)**: Before writing any code, spawn multiple Explorer agents in parallel to preserve main context:
+5. **Analyze Codebase (MANDATORY)**: Before writing any code, understand the affected areas:
+
+   **Tool Selection by Scope:**
+   | Scope | Approach |
+   |-------|----------|
+   | **Narrow** (1-2 files, specific function) | Serena: `get_symbols_overview` → `find_symbol` → `find_referencing_symbols` |
+   | **Broad** (multiple modules, architecture) | Explorer agents in parallel (preserves main context) |
+
+   **For broad changes**, spawn 2-3 Explorer agents simultaneously:
    - **Structure agent**: Overall architecture and file relationships
    - **Pattern agent**: Similar implementations in codebase
    - **Dependency agent**: Affected modules and consumers
 
-   ```
-   Task tool with subagent_type=Explore (launch 2-3 agents simultaneously)
-   ```
+   **For narrow changes**, use Serena directly:
+   1. `get_symbols_overview` on target file(s)
+   2. `find_symbol` with `include_body=True` for specific functions
+   3. `find_referencing_symbols` for impact analysis
 
 6. **Plan Resolution**: Based on analysis results, develop a concrete resolution plan and define work steps.
 
-7. **Resolve Issue**: Spawn sub-agents to modify code and implement features according to the plan.
+7. **Resolve Issue**: Implement the solution using appropriate tools:
+   - **Symbolic edits** (Serena): `replace_symbol_body`, `insert_after_symbol` for precise modifications
+   - **File edits**: For non-code files or complex multi-line changes
+   - **Sub-agents**: For large-scale parallel modifications
    - **If TDD enabled** (marker detected in Step 1):
      - **Reference**: See `feature-planner` skill for detailed TDD methodology
      1. 🔴 RED: Write failing tests first based on requirements

package/.claude/guidelines/ml-guidelines.md

@@ -0,0 +1,57 @@
+# ML Guidelines
+
+Best practices for Machine Learning and Computer Vision tasks.
+
+## Batch Inference Efficiency
+
+Maximize GPU utilization with batch processing during validation/evaluation.
+
+### Do's
+- DataLoader + batch processing required for large-scale inference
+- batch_size based on GPU memory (24GB -> 64, 12GB -> 32)
+- Optimize memory with torch.no_grad() + autocast combination
+
+### Don'ts
+- Calling predict_file() repeatedly in file-by-file loop (GPU utilization < 10%)
+- Using single-file API for large-scale inference
+
+### Pattern
+
+```python
+# Good: Batch inference (GPU utilization 90%+)
+loader = DataLoader(dataset, batch_size=64, num_workers=8)
+with torch.no_grad():
+    for batch in loader:
+        outputs = model(batch["image"].cuda())
+
+# Bad: File-by-file (only for demo/interactive use)
+for f in files:
+    predictor.predict_file(f)  # Excessive GPU idle time
+```
+
+## BGR vs RGB Color Format
+
+OpenCV uses BGR, matplotlib uses RGB. Keep BGR during annotation, convert to RGB only before display.
+
+```python
+# Correct pattern
+img = cv2.imread(path)                    # BGR
+img = annotator.annotate(img, detections) # Keep BGR
+img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # Convert just before display
+plt.imshow(img)
+
+# Wrong pattern (causes color inversion)
+img = cv2.imread(path)
+img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # Too early
+img = annotator.annotate(img, detections)  # BGR colors on RGB image -> inverted
+plt.imshow(img)
+```
+
+## Ultralytics WandB Integration
+
+Ultralytics YOLO has WandB disabled by default.
+
+```bash
+yolo settings wandb=True   # Enable
+yolo settings wandb=False  # Disable
+```

package/.claude/guidelines/work-guidelines.md

@@ -279,5 +279,5 @@ When translating into Korean:
 ## Related Guidelines
 
 - [ID Reference](./id-reference.md) - GitHub/TaskMaster ID conventions
-- [CV Guidelines](./cv-guidelines.md) - Computer vision specific guidelines
+- [ML Guidelines](./ml-guidelines.md) - ML/CV batch inference, color formats, Ultralytics
 - [PRD Guide](./prd-guide.md) - PRD template for TaskMaster

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeongjaeyou/claude-code-config",
-  "version": "0.20.0",
+  "version": "0.21.1",
   "description": "Claude Code CLI custom commands, agents, and skills",
   "bin": {
     "claude-code-config": "./bin/cli.js"

package/.claude/guidelines/cv-guidelines.md

@@ -1,30 +0,0 @@
-# CV Guidelines
-
-Best practices for Computer Vision tasks.
-
-## BGR vs RGB Color Format
-
-**Key Point**: OpenCV uses BGR, matplotlib uses RGB. When annotating with OpenCV-based libraries like supervision, keep the image in BGR format and convert to RGB only right before displaying with matplotlib.
-
-```python
-# Correct pattern
-img = cv2.imread(path)                    # BGR
-img = annotator.annotate(img, detections) # Keep BGR
-img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # Convert just before display
-plt.imshow(img)
-
-# Wrong pattern (causes color inversion)
-img = cv2.imread(path)
-img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # Too early
-img = annotator.annotate(img, detections)  # BGR colors on RGB image -> inverted
-plt.imshow(img)
-```
-
-## Ultralytics WandB Integration
-
-Ultralytics YOLO has WandB disabled by default.
-
-```bash
-yolo settings wandb=True   # Enable
-yolo settings wandb=False  # Disable
-```