fetchwright 0.2.0

0.2.0 · 2026-10-16

• A 429 that names one model's own quota (the free tier gives gemini-3.8-flash 20 requests a day) moves to the next model, which is then used for the rest of the run; the run stops with exit 3 only when every model's quota is spent, or at any other 429. 39 offline tests.

fetchwright 0.1.0

First release. Reads public web pages that refuse scripts, Reddit threads among them, through Gemini's URL context tool.

• fetchwright.py check: whether a Gemini API key is set and where it was found (never the key itself), and which models answer. --no-call sends nothing.
• fetchwright.py read URL ...: a Reddit or Hacker News thread comes back as its title, post and comments, word for word, with the poster's replies marked. Any other page comes back as a faithful section-by-section digest, because Gemini blocks long word-for-word copies as recitation. --verbatim asks for the exact text and falls back to a digest when blocked. --json, --out, --max-comments, --max-chars, --save-raw.
• Models: gemini-3.8-flash, then gemini-3.5-flash-lite on a 5xx, a 404 or a dropped connection. The run stops at a 429, and at 401, 402 or 403.
• Safeguards:
  • requests carry the URL context tool only, and the code refuses Google Search;
  • text without a successful retrieval is discarded;
  • removed-comment placeholders and lone usernames are dropped;
  • store: false, so the exchange is not kept;
  • the key comes from the environment and is never printed.
• Skills: fetchwright (when and how to read, how to read the result) and fetchwright-setup (the key, tiers, quotas, costs and free-tier privacy, with sources dated 2026-10-15).
• 37 offline tests, a live smoke test that skips without a key, a size budget checked in CI (all green), CI on Windows, macOS and Linux with Python 3.10 and 3.14.