Release / build (push) Successful in 1m10s
Before the AI (free): - Seen-URL ledger: every article URL is remembered for 180+ days, so an item still listed in a feed after its article was pruned (rejected ones go after 14 days) is never fetched or judged again. - Wire copy: an identical headline from another outlet within 48 h joins the existing story (or is filtered like its twin) with no triage, grouping or summary call, and does not trigger a rewrite. Triage: - Up to 8 articles from one source are judged in one call, so rules, instructions and examples are sent once; the worker waits up to a minute for a fetch's other articles to finish extracting. Articles the model skips are judged singly. - 800 characters of article text instead of 1500. Grouping without embeddings: - Free pre-filter on shared names (4-letter stems, so Trump/Trumpov and Bahrain/Bahrein agree) and numbers; no call when no recent story shares one, and at most 12 candidates instead of 60. Title and description: - Max article characters default 6000 (was 12000); a multi-source rewrite shares that budget. - New "Reasoning effort" setting, sent as reasoning_effort. Measured on production: Haiku 5.5 spends ~590 hidden reasoning tokens per summary. No wasted retries: - 401/403/404 from the chat endpoint (bad key, unknown model, wrong URL) pauses LLM work with a banner instead of failing every article; it resumes on a settings change or after 10 minutes. - Simpler structured-output modes are tried only on 400/422, so a misconfiguration costs one request, not three. Visibility: - Activity: calls and tokens per pipeline step, calls per published story, and work saved without AI. Playground: tokens per model. Tests for each saving and for the pause on misconfiguration. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>