Codecast

Live-stream coding agent sessions to Discord. No black box — see every tool call, file write, and bash command as it happens. Zero AI tokens burned.

How It Works

┌──────────┐  stream-json  ┌──────────────┐  platform  ┌──────────┐
│ Claude   │ ────────────→ │ parse-stream │ ────────→ │ Discord  │
│ Code -p  │               │ .py          │           │ #channel │
└──────────┘               └──────────────┘           └──────────┘

• Claude Code runs in -p (print) mode with --output-format stream-json --verbose
• parse-stream.py reads JSON lines, posts formatted messages via a platform adapter
• Platform adapters (currently Discord) handle message delivery and threading
• unbuffer (from expect) provides PTY to prevent stdout buffering
• Non-Claude agents fall back to ANSI-stripped raw output relay
• Rate limiting (25 posts/60s) with automatic batching prevents webhook throttling

First-Time Setup

Run these steps once after installing the skill:

1. Make scripts executable

chmod +x <skill-dir>/scripts/dev-relay.sh <skill-dir>/scripts/parse-stream.py

2. Create a Discord webhook

Create a webhook in the target Discord channel via the Discord API or Server Settings → Integrations → Webhooks.

To create via API (if the bot has MANAGE_WEBHOOKS):

curl -s -X POST "https://discord.com/api/v10/channels/<CHANNEL_ID>/webhooks" \
  -H "Authorization: Bot <BOT_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"name":"Codecast"}'

Store the webhook URL:

echo "https://discord.com/api/webhooks/<ID>/<TOKEN>" > <skill-dir>/scripts/.webhook-url
chmod 600 <skill-dir>/scripts/.webhook-url

3. Skip the permissions prompt (Claude Code only)

Create ~/.claude/settings.json if it doesn't exist:

{
  "permissions": {
    "defaultMode": "bypassPermissions",
    "allow": ["*"]
  }
}

4. Install unbuffer (required)

brew install expect    # macOS
apt install expect     # Linux

Invocation

After installing, run chmod +x on the scripts once:

chmod +x <skill-dir>/scripts/dev-relay.sh <skill-dir>/scripts/parse-stream.py

From OpenClaw (recommended)

exec background:true command:"<skill-dir>/scripts/dev-relay.sh -w ~/projects/myapp -- claude -p --dangerously-skip-permissions --output-format stream-json --verbose 'Build a REST API for todos'"

Direct

bash <skill-dir>/scripts/dev-relay.sh -w ~/projects/myapp -- claude -p --dangerously-skip-permissions --output-format stream-json --verbose 'Build auth module'

Options

Thread Mode

Post all messages into a single Discord thread for cleaner channel history:

bash <skill-dir>/scripts/dev-relay.sh --thread -w ~/projects/myapp -- claude -p --dangerously-skip-permissions --output-format stream-json --verbose 'Refactor auth'

Session Resume

Replay a previous session's events (e.g., to a different channel or after a webhook change):

bash <skill-dir>/scripts/dev-relay.sh --resume /tmp/dev-relay.XXXXXX

The relay dir path is printed at session start (📂 Relay: /tmp/dev-relay.XXXXXX).

What Discord Sees

For Claude Code (stream-json mode):

• ⚙️ Model info and permission mode
• 📝 File writes with line count and smart content preview
• ✏️ File edits
• 🖥️ Bash commands
• 📤 Bash command output (truncated to 800 chars)
• 👁️ File reads (hide with --skip-reads)
• 🔍 Web searches
• 💬 Assistant messages
• ✅/❌ Completion summary with turns, duration, cost, and session stats

For other agents (raw mode):

• Output in code blocks with ANSI stripping
• Hang detection warnings
• Completion/error status

End Summary

Every session ends with a summary block showing:

• Files created and edited (with counts)
• Bash commands run
• Tool usage breakdown
• Total cost

Architecture

scripts/
├── dev-relay.sh          # Shell entry point, flag parsing, process management
├── parse-stream.py       # JSON stream parser, rate limiter, event loop
├── .webhook-url          # Discord webhook URL (gitignored)
└── platforms/
    ├── __init__.py       # Platform adapter loader
    └── discord.py        # Discord webhook + thread support

Agent Support

Interactive Input

During an active session, forward input to the agent:

• From OpenClaw: process:submit sessionId:<id> data:"your message"
• Session info stored at /tmp/dev-relay-session.json

Completion Notification

On finish, the relay calls openclaw gateway wake to notify OpenClaw immediately.