autoresearch — Nightly Research Pipeline

A zero-cost nightly research aggregator that rotates through 3 topic tracks, pulls from 4 independent sources, synthesises a structured markdown report, and prints a 3-line Telegram teaser to stdout.

Quick Start

# Dry run — fetches real data, prints teaser, no file writes
cd ~/.openclaw/workspace
uv run --with httpx python skills/autoresearch/scripts/run.py --dry-run

# Full run — writes to memory/ and advances state
uv run --with httpx python skills/autoresearch/scripts/run.py

# Force a specific track
uv run --with httpx python skills/autoresearch/scripts/run.py --track crypto

# Verbose output (debug logs to stderr)
uv run --with httpx python skills/autoresearch/scripts/run.py --dry-run --verbose

Tracks

The pipeline rotates through 3 tracks in order (persisted in state.json):

Sources

Output Files

CLI Flags

Exit Codes

Configuration

Edit config.json to customise per-track queries:

{
  "tracks": {
    "ai": {
      "arxiv_categories": ["cs.AI", "cs.MA", "cs.CL", "cs.LG"],
      "arxiv_keywords": ["agent", "LLM", ...],
      "github_languages": ["python", "rust", "typescript"],
      "github_topics": ["ai-agent", "llm", ...],
      "hn_keywords": ["AI", "GPT", "Claude", ...],
      "web_queries": ["AI agent framework news 2026", ...]
    }
  }
}

Cron Integration

# Add to OpenClaw cron: 1 AM Sydney (14:00 UTC previous day)
# The cron wrapper captures stdout and sends to Telegram
0 14 * * * cd ~/.openclaw/workspace && uv run --with httpx python skills/autoresearch/scripts/run.py 2>>~/.openclaw/workspace/memory/autoresearch-errors.log

The script prints a 3-line teaser to stdout:

🔬 **Nightly Research: AI & Agents**
• Top paper: Scaling Laws for Agent Reasoning… — We study how reasoning…
• Trending: microsoft/autogen ⭐342 | HN: Show HN: I built…

The cron agent captures stdout and sends it to Telegram via the message tool.

State

State is persisted in state.json:

{
  "current_track_index": 1,
  "last_run": "2026-03-15T14:02:31.123456+00:00",
  "last_tracks": ["ai"]
}

State only advances on a successful run (exit 0). If all sources fail, state stays at the same track so tomorrow retries the same track.

Dependencies

• httpx — all HTTP (via uv run --with httpx)
• xml.etree.ElementTree — arXiv Atom XML parsing (stdlib)
• json, re, asyncio, argparse, pathlib — stdlib

No additional dependencies needed. No pyproject.toml required in the skill dir.

Integration with Book Draft

Other cron jobs or agents can read the latest report directly:

cat ~/.openclaw/workspace/memory/autoresearch-latest.md

Or in Python:

from pathlib import Path
report = Path.home() / ".openclaw/workspace/memory/autoresearch-latest.md"
content = report.read_text()

File Structure

skills/autoresearch/
├── SKILL.md          # This file
├── PLAN.md           # Architecture and spec
├── config.json       # Track definitions + source config
├── state.json        # Runtime state (auto-managed)
└── scripts/
    ├── run.py        # CLI entrypoint (main pipeline)
    ├── sources.py    # Data fetchers (arXiv, GitHub, HN, web)
    ├── synthesise.py # Report builder (markdown synthesis)
    └── state.py      # Track rotation state machine