Aperture v0.1.0

Aperture v0.1.0 — A skill for AI agents that curate better daily reports

Aperture is a deterministic, auditable news-curation skill for AI agents. It replaces stateless RSS readers and one-shot LLM summarizers with a four-stage pipeline that remembers yesterday, learns from feedback, and can explain every decision.

What you get

• Scan the front page — diff each source's list page against yesterday's snapshot. Robust where timestamps are unreliable.
• Tape every decision — every source snapshot, item, rejection reason, profile version, and report is append-only JSONL.
• Prescreen with taste — weighted keywords, category bonuses, and negative terms; cheap and deterministic.
• LLM review — optional second pass for news verification, on-topic scoring, and summarization.
• Deduplicate by meaning — URL normalization + title simhash clustering collapse the same event across feeds.
• Learn from feedback, reversibly — user feedback becomes versioned profile operations; roll back any change.
• Rule-only dry mode — run the entire pipeline without an LLM API key.
• Agent-native — SKILL.md is the spec; python -m engine.pipeline is the deterministic reference run loop.

Install

git clone https://github.com/lukethecat/aperture.git
cd aperture
python -m engine.pipeline --dry --vertical tech --config config/example_vertical.toml

Or install it as a skill into OpenClaw, Raft, Hermes, Claude Code, or Codex — see docs/installing-for-agents.md.

Explore

• Read the skill spec: SKILL.md
• See a real daily issue: docs/sample-issue.md
• Replay any decision: python scripts/replay.py --item <id>

Assets

• Source code (tar.gz)
• Source code (zip)

Full changelog

See CHANGELOG.md.