Pylon Support Bot

AI-powered support draft bot for Pylon. Receives Pylon webhooks, generates draft replies using Claude (via Kilo Gateway), and posts them as internal notes on the Pylon issue for human review.

Architecture

Rendering mermaid diagram...

Setup

1. Val Town Secrets

Add these environment variables in the Val Town project settings:

Key	Description	Where to get it
`KILO_GATEWAY_KEY`	Kilo Gateway API key	app.kilo.ai/profile
`PYLON_API_KEY`	Pylon API token (Bearer auth)	Pylon → Settings → API → Create token
`PYLON_WEBHOOK_SECRET`	Webhook destination secret	Pylon → Settings → API → Webhook destination secret

2. Pylon Webhook Destination

Go to Pylon → Settings → API → Webhooks.
Create a new webhook destination with URL: https://YOUR-VAL-ENDPOINT.web.val.run/
Select the event types you want to trigger on (e.g., issue.message.created).
Save the secret — add it as PYLON_WEBHOOK_SECRET in Val Town.

3. Pylon Trigger Configuration

Create a Pylon Trigger that fires on new customer messages and sends a webhook with the issue ID:

{
  "id": "{{issue.id}}"
}

The bot will fetch the full issue context (metadata and conversation history) from the Pylon API.

Files

File	Purpose
`main.ts`	HTTP handler — webhook receiver, signature verification, background processing
`pylon.ts`	Pylon API client — webhook verification, context fetching, note posting
`pipeline.ts`	Core pipeline — fetch context → generate LLM reply → post note
`prompt.md`	System prompt for the AI model with response guidelines
`scripts/test-pipeline.ts`	End-to-end test script
`README.md`	This file

How It Works

Webhook received at POST / — the handler verifies the Pylon HMAC-SHA256 signature.
Early return — responds 200 OK immediately, then fires a background request to POST /process.
Fetch context — retrieves full issue metadata and conversation history from Pylon API.
AI inference — loads prompt.md as the system prompt, sends the issue context to Claude Opus 4.6 via Kilo Gateway.
Internal note — posts the AI-generated draft as an internal-only note on the Pylon issue.
Error handling — on failure, sends an error notification email to steve@val.town.

Customization

Change the AI Model

Edit pipeline.ts to use a different model:

// Current: Claude Opus 4.6 via Kilo Gateway
const kilo = createOpenAI({
  baseURL: "https://api.kilo.ai/api/gateway",
  apiKey: Deno.env.get("KILO_GATEWAY_KEY"),
});

const { text: draftReply } = await generateText({
  model: kilo("anthropic/claude-opus-4-6"),
  system: systemPrompt,
  prompt: userPrompt,
});

// Example: Use a different model through Kilo Gateway
const { text: draftReply } = await generateText({
  model: kilo("openai/gpt-4-turbo"),
  system: systemPrompt,
  prompt: userPrompt,
});

Customize the System Prompt

Edit prompt.md to add:

Product-specific knowledge
FAQ pairs from your support history
Links to your documentation
Custom tone and style guidelines
Troubleshooting trees for common issues

Testing

Run the test script to verify the pipeline end-to-end:

// Edit scripts/test-pipeline.ts with a real issue ID
const TEST_ISSUE_ID = "your-pylon-issue-id";

// Run the script from Val Town

Webhook Signature Verification

The bot verifies Pylon webhook signatures using HMAC-SHA256:

Pylon sends X-Pylon-Signature header with hex-encoded digest
Bot computes HMAC-SHA256 of raw request body using PYLON_WEBHOOK_SECRET
Constant-time comparison prevents timing attacks
Invalid signatures return 401 Unauthorized

See Pylon webhook signature docs for more details.

API Usage

Pylon API Endpoints Used

GET /issues/{id} — Fetch issue metadata
GET /issues/{id}/messages — Fetch conversation messages (with pagination)
POST /issues/{id}/threads — Create a new thread
POST /issues/{id}/note — Post an internal note to a thread

See Pylon API Reference for more details.

Deployment

This bot is deployed on Val Town, a serverless platform for TypeScript/JavaScript functions.

To deploy your own:

Fork this val or create a new val in your Val Town account
Add the required environment variables (see Setup section)
Configure the Pylon webhook destination to point to your val's endpoint
Test with a real issue using scripts/test-pipeline.ts