Vibe Coding with the Agent

Vibe coding is a workflow where you describe what you want in plain English, the agent generates or edits the code, you review it in the built-in editor, and refine with feedback. Think of it as a conversation that produces code.

This works for: Python, YAML, JSON, TOML, Bash, Rust, JavaScript, TypeScript, ComfyUI workflow JSON, .thumper.yaml manifests, and .model.yaml/.pack.yaml files.

How it differs from traditional coding: you describe the outcome, not the implementation. The agent reads existing code, understands the structure, and makes targeted changes.

Prerequisites: Thumper-Run installed with at least one app (ComfyUI recommended for this tutorial).

Select the code-helper template from the dropdown at the agent launch screen. This adds a system prompt: "Focus on code-related tasks. Prefer search_code, file_read, file_edit. Show code snippets."

Recommended Models for Code

You can also attach reference files at launch (image/*, .pdf, .txt, .md, .json, .csv) to give the agent initial context without extra tool calls.

Your model and template choices are persisted via CRDT so they're remembered next time.

Example 1: Create a ComfyUI Workflow

Type a specific request:

Create a ComfyUI API workflow JSON that generates a 1024x1024 image using
the SDXL base checkpoint with 25 KSampler steps, CFG scale 7, euler sampler,
normal scheduler. Include positive and negative CLIP text encode nodes.
Save it to ~/comfyui-workflows/portrait.json

The agent reasons through the node structure it needs, then writes the file:

After you approve, the file appears in the code editor panel:

{
  "3": {
    "class_type": "KSampler",
    "inputs": {
      "seed": 42,
      "steps": 25,
      "cfg": 7,
      "sampler_name": "euler",
      "scheduler": "normal",
      "denoise": 1.0
    }
  }
}

Example 2: Edit a .thumper.yaml Manifest

Read the ComfyUI .thumper.yaml and add a ROCm platform patch that
sets HSA_OVERRIDE_GFX_VERSION=11.0.0 for AMD gfx1100 GPUs

The agent first reads the file, then makes a targeted edit:

Example 3: Generate a Python Script

Write a Python script that converts all PNG images in a directory to
WebP format using Pillow. Add argparse for input/output directories
and quality setting.

The agent generates a complete script with imports, argument parsing, and error handling:

import argparse
from pathlib import Path
from PIL import Image
 
def convert_images(input_dir, output_dir, quality):
    input_path = Path(input_dir)
    output_path = Path(output_dir)
    output_path.mkdir(parents=True, exist_ok=True)
    for png in input_path.glob("*.png"):
        img = Image.open(png)
        img.save(output_path / f"{png.stem}.webp", "webp", quality=quality)

When the agent creates or edits files, a code editor panel appears on the right (668px width). This is your review workspace.

File Tabs

Each generated file gets its own tab:

• Active tab — bg-gray-800, pink text
• Inactive tabs — border outline, gray text
• Modified indicator — ● dot next to filename
• Close button — X icon on each tab

Syntax Highlighting

The editor highlights code with a dark theme:

• Tags/keywords — blue
• Attributes/functions — green
• Values/strings — amber
• Text/identifiers — light gray

Supported languages: Python, YAML, JSON, TOML, Bash, Rust, JavaScript, TypeScript, HTML, CSS, and more.

Diff View

When file_edit modifies an existing file, you see a before/after diff:

• Green lines — added
• Red lines — removed

Changes must be approved before they're saved to disk. Think of this as your safety net — a checkpoint you can review before anything changes.

Editor Toolbar

The Refinement Loop

1. Describe what you want (code mode)
2. Review the generated code in the editor
3. Tell the agent what to fix (natural language)
4. Agent reads current state, applies targeted edits
5. Repeat until satisfied

Example: Iterating on a Workflow

Error Recovery

When things go wrong, the agent handles it:

• Tool fails: Agent sees the error in ToolResult and can retry with different parameters.
• file_edit fails (old_string not found): Agent re-reads the file with file_read, adjusts old_string to match current content.
• run_command times out (60s limit): Agent reports the timeout and suggests alternatives.

Multi-File Scaffolding

Ask the agent to create an entire project structure:

"Create a ComfyUI custom node called ThumperResize"

The agent generates __init__.py, resize_node.py, and requirements.txt — each as a separate file_write call with its own editor tab.

Search Then Edit

For bulk changes across files:

"Find all files referencing 'sd_xl_base_1.0.safetensors' and update to 'juggernautXL_v9.safetensors'"

The agent runs search_code to find all matches, then file_edit on each one.

Context Attachment

Attach .json, .yaml, or other reference files at launch to give the agent initial context. This saves tool calls and context tokens.

agent_knowledge.md

Create a persistent knowledge file at ~/.local/share/tr-desktop/agent_knowledge.md (max ~2000 chars). This is loaded at priority 3 into every agent session:

This project uses ComfyUI with SDXL.
Models are in /data/models/.
Use BF16 on AMD GPUs — never FP16.
Preferred sampler: euler_ancestral.

Filesystem Tools Reference