tg-channel-analysis/CLAUDE.md

43 lines
1.9 KiB
Markdown

# CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
## Overview
Telegram Channel Analyzer Bot — fetches public channel history via Telethon, analyzes with Claude AI (chunked summarization pipeline), delivers reports via aiogram bot.
## Stack
- **Python 3.12**, async throughout
- **aiogram 3** — Telegram bot interface (commands, inline keyboards, progress messages)
- **Telethon** — userbot client for reading public channel history
- **anthropic** (AsyncAnthropic) — Claude API with streaming
- **pydantic-settings** — config from environment variables
## Architecture
- Both aiogram and Telethon share one asyncio loop (no threads)
- Telethon client is attached to the bot instance in `__main__.py`
- Analysis pipeline: Fetch → Chunk (token-bounded) → Summarize each chunk → Synthesize final report
- Rate limit handling: semaphore(1), 60s cooldown between chunks, retry with backoff on 429
- Output: Markdown→HTML conversion, split on section boundaries at 4000 chars
## Key Files
- `bot/config.py` — all settings from env vars, `CLAUDE_MODEL` selects the model
- `bot/services/analyzer.py` — Claude API calls, adaptive thinking only for opus-4-6
- `bot/services/chunker.py``MAX_TOKENS_PER_CHUNK` and `CHARS_PER_TOKEN` control chunking
- `bot/prompts/` — prompt templates per report type (chunk_summary.py, synthesis.py)
## Running
- Container-based: `Containerfile` + `compose.yml`
- `--login` flag for interactive Telethon session creation
- Session persists in `data/` volume
- `.env` file must not have inline comments (Podman/Docker limitation)
## Common Tasks
- To change chunk size: edit `MAX_TOKENS_PER_CHUNK` in `bot/services/chunker.py`
- To add a report type: add to `ReportType` enum, add prompts in both `prompts/` files
- To change model: set `CLAUDE_MODEL` env var; thinking params auto-adapt in `analyzer.py`