What we shipped on 2026-07-28
We spent the morning fixing a bug that literally erased trillions of dollars from our podcast narratives (PR #2876). Our TTS filename rule was using a regex--re.compile(r"[\w/\\]+\.\w{2,4}(?:\s|$)")--intended to strip file extensions from the text before speaking it. The problem is that nothing required the extension to be alphabetic. When the narrator hit "The $1.65 Trillion Secret," the system parsed 1.65 as a stem and an extension, deleted it, and told the audience there was a "$ Trillion Secret." It turns out any 2-to-4-digit fractional part followed by whitespace was being eaten. We've tightened the rule to stop the bleeding and updated the system to properly accept H2 headlines.
Beyond the media glitches, we shipped release 0.112.0 (PR #2881), which cleaned up several business-critical edges in our affiliate tracking. We added an affiliate keyword CLI and seeded prose-shaped aliases (PR #2874) to make the AI's promotional insertions feel less like bots and more like natural content. To keep the data honest, we also updated the attribution logic to exclude bot clicks and attribute by source page (PR #2868).
There was some necessary housekeeping on the infrastructure side today. We added an alert for mains undervoltage in our observation system (PR #2855) to catch power drops before they hard-crash the host. In the console, we capped the SOCIAL DISTRIBUTION height and added a status filter (PR #2856) because the UI was becoming unmanageable as the volume of drafts grew. Speaking of drafts, we finally started pruning rejected social drafts that were never posted (PR #2867), and officially retired the corsair_csv sensor feed after confirming the iCUE CSV source is gone for good (PR #2875).
We also pushed through a batch of dependency updates to keep the agent's brain current, including bumping langgraph to 1.2.9 (PR #2736) and langchain-core to 1.5.1 (PR #2735), which brings in reasoning_effort as a standard chat model parameter.
Today was a reminder that the "last 5%" of polishing an AI content pipeline is mostly just fighting regexes and monitoring power rails. We're moving toward a state where the system can self-correct its own prose errors, but for now, we'll keep auditing the transcripts to make sure our numbers actually stay in the script.
Auto-compiled by Poindexter from today's commits and PRs. See the work: github.com/Glad-Labs/poindexter.
Top comments (0)