What future-agi/future-agi shipped
Generated by FoxPlug from public activity — not affiliated with Futureagi.
Week of September 14, 2026
What shipped
- v1.38.4 released with fixes for eval-tasks: gating errored/skipped requeue on terminal watermark, sniffing URL type across chunks, and stopping continuous tasks from re-running converged failures on every poll. Release
- Fixed continuous eval tasks re-running converged failures on every poll, which was causing ~700 GB/day of S3 egress in production. Pull request #2846
- Renamed the 'RL Environment' label to 'Environment' in the harness feature under Simulate. Pull request #2863
- v1.38.3 hotfix released with batching of changed span identities below query size limit and fixes for dense continuous and sparse historical selection. Release
- v1.38.2 released with fixes to apply Dataset and Eval Source filters to eval metric charts. Release
- Fixed system eval binding config to preserve dataset-bound evaluation configuration when reopening and saving existing evaluations. Pull request #2621
- Promoted dev to main for v1.38.4 release with 16 commits including eval alert threshold scaling fixes. Pull request #2866
- Reworked eval alert thresholds to label them as 0-1 fractions instead of 0-100 percent on the backend. Pull request #2497
Why it matters
This week shipped critical stability fixes for eval-tasks that were causing excessive cloud storage egress in production. Multiple releases addressed filtering issues on dashboards and preserved evaluation configuration across saves. The project continues to refine alert thresholds and evaluation handling.
Changelog entry
- v1.38.4: Gate errored/skipped requeue on terminal watermark, sniff URL type across chunks, stop continuous tasks re-running converged failures every poll Release
- v1.38.3: Batch changed span identities below query size limit, bound remaining continuous candidate queries, hotfix dense continuous and sparse historical selection Release
- v1.38.2: Apply Dataset and Eval Source filters to eval metric charts Release
- fix(eval-tasks): Stop continuous tasks re-running converged failures every poll by gating requeue on terminal watermark Pull request #2846
- chore(harness): Rename 'RL Environment' label to 'Environment' in Simulate feature Pull request #2863
- fix(evals): Preserve system eval binding config when reopening and saving existing evaluations Pull request #2621
- fix(alerts): Label eval alert thresholds as 0-1 fractions instead of 0-100 percent Pull request #2497
- fix(dashboards): Apply Dataset and Eval Source filters to eval metric charts Pull request #2724
v1.38.4 is out with fixes for eval-task requeue storms that were costing ~700 GB/day in S3 egress. Also includes dashboard filter fixes and label improvements.
This week we shipped v1.38.4 with critical fixes for continuous eval-tasks that were re-running converged failures on every poll cycle, causing ~700 GB/day of unnecessary S3 egress in production. We also fixed dashboard filters not applying to eval metric charts and improved the harness UI labeling. Multiple hotfixes ensure stability across the evaluation pipeline.