Headroom Daily: Killing Single Points of Failure in Last30Days
The last30days-skill repo shipped a cluster of reliability fixes today, all converging on one problem: single vendor or source failures shouldn't take down an entire research run. A four-state doctor audit and version 3.15.0 shipped alongside them.
Duration: PT2M48S
Episode overview
This episode is a short developer briefing from Headroom Daily.
It explains recent repository work in plain language.
- Show: Headroom Daily
- Published: 2026-07-15T22:00:04Z
- Audio duration: PT2M48S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's July 15th, 2026, and this is Headroom Daily.
The clear signal today comes out of the last30days-skill repository, where five separate pull requests all attack the same failure mode: one broken data source shouldn't poison the whole run.
Start with PR 830. A single Instagram transcript returning an HTTP 400 error was getting recorded at the wrong layer — before the exception handler could catch it — so the entire batch reported zero items and an error outcome, even when the search itself succeeded. That's now isolated, tied to issue 829.
PR 825, from Mekiwi, found two compounding bugs: DuckDuckGo blocking datacenter IPs entirely, and Reddit enrichment data leaking into and poisoning unrelated web search results. Together they meant keyless web search silently returned nothing on self-hosted deployments.
PR 827 removes a paid-API dependency altogether — YouTube comments now fetch through yt-dlp, which was already in use for search and transcripts, with ScrapeCreators demoted to a fallback. That matters because a prior run showed ScrapeCreators credit exhaustion taking down YouTube, TikTok, and Instagram data…
And PR 824 fixes a crash where long natural-language…