What we shipped on 2026-09-01
We spent today chasing ghosts in our telemetry, starting with a silent failure in SyncCloudflareAnalyticsJob (PR #3523). We discovered that Cloudflare Analytics Engine has a non-zero visibility delay--a data point is written, but it isn't queryable the instant writeDataPoint returns. Because we were using a high-water-mark cursor (WHERE timestamp > '{since}'), any row that surfaced after our poll had already advanced the watermark was effectively deleted from our history. On low-traffic pages, where empty responses often pushed the cursor straight to now(), we were simply losing page views forever without a single error log to warn us.
While auditing infrastructure, we found another "loaded gun" in our affiliate redirect Worker (PR #3520). The wrangler.toml committed to git had placeholders for the R2 host, but the live Worker was running with the real production URL. Since wrangler deploy is declarative, any clean-checkout deployment would have overwritten the working binding and taken every /go/<slug> link offline. We've moved these to deploy-proof secrets now that we know exactly how close we were to a total redirect blackout.
The QA pipeline had also gone quiet in a way we didn't notice immediately (PR #3519). Three of our LLM rails--deepeval_g_eval, deepeval_faithfulness, and ragas_eval--were returning zero reviews on 100% of passes. The culprit was the constrained-decoding fix from a previous sprint: we were sending response_format={"type":"json_object"} to thinking judges. Because these models must emit a reasoning trace before the JSON answer, the constraint forced them into an immediate stop, resulting in empty content and JSONDecodeError. Withholding JSON mode for thinking judges revived all three rails instantly.
On the GPU side, we had to fix a regression we shipped an hour prior (PR #3524). We realized that /unload lacked an in-flight guard, meaning a concurrent VRAM reclaim could self-exit the process while a render was still active. We've now mirrored our WAN contract: we track in-flight generations and decline unloads until they finish. In this case, obeying the reclaim request is strictly worse than declining it, as the running render is exactly what that VRAM is currently serving. This ties into a broader effort to verify every hard reclaim rung across ComfyUI, wan, image-gen, and stable-audio (PR #3516).
We also finally broke the "upload once" limitation of our YouTube integration (PR #3518). Previously, we had no way to fix metadata for videos already on the channel--meaning twelve videos were stuck with 4,800-character markdown walls in their descriptions. We implemented adapter.update_metadata() using a read-modify-write pattern to ensure that updating the description doesn't accidentally blank the title or category IDs.
We closed out the day by admitting our README was rotting (PR #3522). Our marketing stats--live post counts, test totals, and DB settings--were hand-typed and stale. We've refreshed them and wired them into a nightly sync using collect_stats() (PR #3526, PR #3527), so the README and CLAUDE.md can no longer drift apart from reality.
It was a day of closing gaps--some in our data, some in our infrastructure, and some in our documentation. We're moving toward a state where the system tells us it's broken before the users do.
Auto-compiled by Poindexter from today's commits and PRs. See the work: github.com/Glad-Labs/poindexter.
Sources
Originally published at www.gladlabs.io.
Top comments (0)