Development Skill | Agent Skills

Development

Technology Stack

Blender addon (Python)
Code in WSL; Blender runs on Windows

Development Guidelines

Use the language server and project linters where available.

Deploying Changes

Command: ./dev_deploy.sh | grep ".py"

After making changes:

Auto-run the command to deploy to Blender.
Delete old logs to reduce bloat.
Inform the user to reload the addon.

Debugging Across Windows/WSL Filesystem Boundaries

Blender runs on Windows, code workspace is in WSL.

Write Diagnostics from Windows into WSL

In Blender Python, write diagnostics to the workspace via the WSL network path so the agent can read them on Linux. Use \\wsl$\Ubuntu-24.04\.

# Example: write to project .cursor/debug.log from Blender on Windows
log_path = r"\\wsl$\Ubuntu-24.04\home\nyaarium\blender-nyaatools\.cursor\debug.log"
with open(log_path, "a", encoding="utf-8") as f:
    f.write(...)

Debugging Approach

When debugging issues, follow this systematic approach to avoid drowning the codebase in unnecessary logging.

1. Hypothesize First

Before adding any instrumentation, state 3-5 concrete, testable hypotheses. Good ones are specific and testable (e.g. "the auth token is null at checkout" or "the loop uses count < 10 instead of count <= 10"). Avoid vague statements ("something is broken in the payment flow").

2. Instrument to Test Hypotheses

Add write calls from Blender Python to .cursor/debug.log (via the WSL network path below) to confirm or reject each hypothesis (typically 2-6 logs total). Log entry/exit, key values at decision points, which branch was taken, and important return values. Don't log every line, redundant data, or things you already know are correct.

3. Gather Evidence

Run the code and examine the debug log. For each hypothesis, decide:

CONFIRMED - The logs prove this is the issue
REJECTED - The logs prove this is NOT the issue
INCONCLUSIVE - Need different instrumentation to test this

Only fix issues when you have clear runtime evidence pointing to the cause. Don't guess.

4. Fix and Verify

Keep instrumentation in place after a fix, run a verification test, and only remove debug logs once the fix is confirmed. That avoids "fixed one thing, broke another."

Writing to the Debug Log from Blender

Blender runs on Windows; the workspace is in WSL. Write diagnostics to the workspace via the WSL network path so the agent can read .cursor/debug.log on Linux.

Path (from Blender on Windows): \\wsl$\Ubuntu-24.04\home\nyaarium\blender-nyaatools\.cursor\debug.log

Format: One JSON object per line (NDJSON). Each line must include id, timestamp (ms since Unix epoch), location, message, data (object). Optional: hypothesisId, runId. You must supply id yourself (e.g. log_{timestamp}_{random}).

Example

import json
import time
import random
import string

DEBUG_LOG = r"\\wsl$\Ubuntu-24.04\home\nyaarium\blender-nyaatools\.cursor\debug.log"

def _log_id():
    return "log_%d_%s" % (int(time.time() * 1000), "".join(random.choices(string.ascii_lowercase + string.digits, k=8)))

# #region agent log
with open(DEBUG_LOG, "a", encoding="utf-8") as f:
    f.write(json.dumps({
        "id": _log_id(),
        "timestamp": int(time.time() * 1000),
        "location": "operators/export_vrm.py:142",
        "message": "Checking mesh before export",
        "hypothesisId": "A",
        "data": {
            "mesh_count": len(meshes),
            "has_armature": armature is not None,
        },
    }, ensure_ascii=False) + "\n")
# #endregion

GitHub Workflow Tips

If you need to check a Github workflow runner job's log, you can use the gh CLI:

Ask for one of either:
- URL. Route must contain: */actions/runs/*
- Run ID or Job ID. If the ID is 5 char or less, assume it's wrong; There is a shorter human friendly ID on the website <=5 chars.
View it with:
- gh run view $RUN_ID --log
- gh run view --log-failed --job=$JOB_ID

Agent Skills: Development

Install this agent skill to your local

Skill Files