Summary

• telegram.zig — fire-and-forget Telegram sender (send/sendDraft/sendFmt) using oracle_watchdog HTTP pattern
• ralph_hook.zig — standalone Zig binary called by Claude Code hooks; reads stdin JSON, formats and sends to Telegram
• claude_runner.zig — --continue flag for native session resume (replaces HANDOVER.md), mcp__telegram__SEND_MESSAGE in allowedTools
• agent_loop.zig — WAKE/SLEEP Telegram reporting; hooks handle all per-tool reporting
• hooks.json — PostToolUse, PostToolUseFailure, Stop events call ralph-hook binary
• .mcp.json — @iqai/mcp-telegram MCP server for Claude self-reporting

Architecture

ralph-agent (Zig daemon)
├── WAKE → telegram.send("WAKE #N")
├── Spawn claude -p <prompt> --continue --max-turns N
│   ├── PostToolUse hook → ralph-hook → TG: "tool done"
│   ├── PostToolUseFailure hook → ralph-hook → TG: "tool FAILED"
│   ├── Stop hook → ralph-hook → TG: "Session finished"
│   └── Claude MCP → mcp__telegram__SEND_MESSAGE (self-report)
└── SLEEP → telegram.send("Sleeping Ns...")

364 insertions, 38 deletions across 8 files (2 new, 6 modified)

Test plan

• zig build — both ralph-agent (4.5MB) and ralph-hook (4.3MB) compile
• echo '{"hook_event_name":"PostToolUse","tool_name":"Bash"}' | ralph-hook — sends to Telegram API
• ralph-agent --single-shot — shows v2.0.0, Telegram enabled/disabled
• Run with real TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID to verify end-to-end reporting

🤖 Generated with Claude Code