What we shipped on 2026-07-26
We spent a good chunk of today chasing why our premium Chatterbox voices had silently stopped working, only to find we'd been fighting a hardware compatibility ghost in "move TTS sidecar to a Blackwell-compatible torch build" (PR #2798). The pipeline had fallen back to Speaches/Kokoro without warning because the RTX 5090's sm_120 capability wasn't supported by the default PyTorch installation. Compounding that, we had a redundant CUDA_VISIBLE_DEVICES filter fighting against device_ids, which effectively asked CUDA for a second, nonexistent GPU and returned False for availability every time.
While auditing the console, we realized our Service Health "Restart" button was pure client-side theater (PR #2797). It flipped a local state and pushed a toast, but never actually hit the backend because the route didn't exist. We fixed this by implementing an operator-triggered restart intent queue via a new service_restart_requests table and routing those intents to the brain using the same docker_restart_container primitive our self-healing firefighter uses.
We also had to admit that our QA scoring was lying to us. In "stop the pre-QA heuristic from outranking the real judge-rail score" (PR #2795), we found that qa.aggregate was promoting a cheap pattern-based heuristic as the final quality_score whenever it happened to be higher than the actual LLM critics. This meant fake numbers were triggering our auto_publish_gate. The real judge-rail score is now authoritative whenever a rail has actually reviewed the content.
The SEO Harvest Loop provided a harder lesson in "mechanically working vs. actually valuable." An audit showed that while 122 seo_refresh runs completed, they were a net-zero return because we were targeting page-2 posts with almost no visibility (PR #2793). We added a demand floor--specifically seo.striking_distance.min_impressions defaulting to 100--to ensure we aren't wasting human approval cycles on posts that have no CTR signal to move. This now feeds into the "Lock-2" graduation mechanism (PR #2789), where a streak of clean human approvals via count_trailing_clean_approvals finally allows these refreshes to graduate to auto-publish.
Finally, we reclaimed some VAE-decode fragmentation headroom on the wan-server (PR #2792). By setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, we stopped 480p hero renders from OOMing when they hit a few hundred MBs of spike margin.
We're moving past the phase where "it runs" is enough; today was about pruning the heuristics and theater to make sure the system actually delivers the value it claims on the dashboard.
Auto-compiled by Poindexter from today's commits and PRs. See the work: github.com/Glad-Labs/poindexter.
Top comments (0)