Prompt Coder

Generate/replace raw code from prompts written inside comments.
Keeps your @ai markers. Strips Markdown code fences/backticks. Works with Ollama, OpenAI, or a generic HTTP backend.

✨ What it does

• Turn inline @ai comments into real code in your current file.
• Never deletes your directive lines.
• Removes ```/~~~ fences and inline backticks from model output.
• Inserts after the directive (single-line) or replaces only the interior (block).

🧩 Supported directives (case-insensitive)

1) Single-line generate

// @ai generate: function sum(a,b) that returns a+b

➡️ Inserts generated code after this line.

2) Multi-line generate

@ai generate:
React button component with rounded corners and onClick logging "clicked"
@ai end

➡️ Keeps begin/end. Replaces only the lines between them.

3) Replace block

@ai replace begin: write a useState hook for "X"
const [x, setX] = useState('');
@ai replace end

➡️ Keeps begin/end. Replaces only the interior.

📝 Comment parsing (strict rule)

For @ai generate: and @ai replace begin: inside comments:

• If the trimmed line starts with a multi-line opener — <!--, /*, or {/* — we scan forward to the first close (-->, */, */}) and insert after the close.
• Otherwise, treat as single-line: prompt ends at EOL and we insert after the line (no forward scan).

This prevents “jumping” to a stray */ in strings later in the file.

🔧 Install

• From VSIX: Extensions → ⋮ → Install from VSIX…
• Or from Marketplace: search “Prompt Coder” (by your publisher).

Requires VS Code 1.85.0+.

⚙️ Settings (File → Preferences → Settings → “Prompt Coder”)

• promptCoder.provider: ollama | openai | http (default: ollama)
• promptCoder.model: model name/tag (e.g. qwen2.5-coder:7b, gpt-4o-mini)
• promptCoder.endpoint: server URL (Ollama/HTTP)
• promptCoder.apiKey: for OpenAI/HTTP (if required)
• promptCoder.httpPath: path for generic HTTP (default /api/generate)
• promptCoder.maxTokens: default 2048
• promptCoder.temperature: default 0.2
• promptCoder.systemPrompt: “Return only code; no Markdown fences/backticks.”
• promptCoder.runOnSave: process directives on save (off by default)
• promptCoder.wrapSelectionWithCodeFence: affects Generate Here only

Backend quickstarts

Ollama (local, default)

• Install Ollama; run a code model (example):
  ollama pull qwen2.5-coder:7b
• Settings:
  • provider: ollama
  • endpoint: http://localhost:11434
  • model: qwen2.5-coder:7b

OpenAI

• Settings:
  • provider: openai
  • apiKey: sk-...
  • model: e.g. gpt-4o-mini
  • temperature/maxTokens as needed

Generic HTTP

• Settings:
  • provider: http
  • endpoint: http://localhost:3000
  • httpPath: /api/generate
  • (optional) apiKey
  • model/temperature/maxTokens as needed

▶️ How to use

1. Write a directive in your code.
2. Click “Run @ai” above the line (CodeLens) — or use Command Palette:
   • Prompt Coder: Process AI Comment At Line
   • Prompt Coder: Process Current AI Comment Only
3. See the status bar: 🔄 Prompt running… → 🚀 Prompt Coder.
4. Use “Stop” to cancel.

🧪 Examples

JavaScript

// @ai generate: export function greet(name){ returns "Hello, <name>!" }

/* @ai generate:
Create a debounce(fn, wait) utility
@ai end */

// @ai replace begin: create singleton Logger with log()
const placeholder = () => {};
// @ai replace end

React / JSX

{/* @ai generate: useState hook for "X" default '' */}

{/* @ai replace begin: button component "Click me" with rounded corners */}
<Button>TODO</Button>
{/* @ai replace end */}

HTML

<!-- @ai generate: blue rounded button titled "Click Me" with alert on click -->

Python

# @ai generate: def factorial(n): iterative, raises on negative

SQL

-- @ai generate: create table users(id int primary key, name text not null)

🧼 Output sanitization (always on)

• Removes triple fences (``` / ~~~), keeps inner code only.
• If whole output is single-backtick wrapped, unwraps it.
• Trims blank edges and removes common left indentation.
• Inserts exactly one trailing newline.
• No prose or Markdown fences are written to your file.

🧭 Commands

• Process AI Comment At Line (CodeLens “Run @ai”)
• Process Current AI Comment Only
• Process AI Comments (Generate/Replace) — whole file
• Stop Current Generation
• Helpers:
  • Generate Here (Insert/Replace Selection)
  • Generate & Replace Whole File
  • Generate Into New File
• Utilities:
  • Refresh CodeLens
  • Debug – Show Directive Matches

✅ Guarantees

• Edits are anchored to the directive, not the cursor.
• Single-line: insert after the line.
• Block forms: replace interior only.
• Multi-line close scan runs only when the line starts with a block opener (<!--, /*, {/*).
• Directives are never removed or altered.

🛠️ Troubleshooting

• No buttons shown: ensure Editor › Code Lens is enabled; open an editable file (not a diff). Use Prompt Coder: Refresh CodeLens.
• Inserted at wrong place: check if your directive line actually starts with a block opener; otherwise it’s single-line → insert after EOL.
• Output contains fences: set/keep systemPrompt and verify your backend returns pure code.
• Cancel errors: use the Stop button; the extension cancels the HTTP request and doesn’t write partial output.

📄 License

MIT. Contributions welcome.