What shipped in Supabase on August 22, 2026?

Supabase's eval judge is getting smarter, but the RLS guide fix is the one that will keep your docs green. In total, 41 commits, 46 pull requests, and 5 releases landed.

Who contributed to Supabase on August 22, 2026?

12 developers shipped this update, including Miranda Limonczenko, supabase-evals-releaser, claude[bot], charislam, kemaldotearth, SaxonF, claude, and filipecabaco, and 4 more.

What were the notable Supabase updates?

fix(core): default the judge to gpt-5.6-sol at medium reasoning effort (#222), fix(evals): soften the pgTAP coverage rubric in the RLS guide eval (#224), and chore(deps): consolidate TypeScript and update Effect (#6289).

@supabase

The open-source Firebase alternative powering thousands of startups

github ↗

Pick a date

Topics: JavaScript Full archive →

The Wire · Showcase

JUDGE UPGRADE: evals now reasons at medium effort

By RepoJournal · Filed 06:57 UTC on August 22, 2026 · About Supabase

12 people shipped this

Miranda Limonczenko @czenko 2 cited

Julien Goux @jgoux 2 cited

claude[bot] @claude[bot] 1 cited

charislam @charislam 1 cited

kemaldotearth @kemaldotearth 1 cited

SaxonF @SaxonF 1 cited

filipecabaco @filipecabaco 1 cited

iambriccardo @iambriccardo 1 cited

supabase-evals-releaser @supabase-evals-releaser

claude @claude

Jordan McQueen @jmqd

dependabot @dependabot

Supabase's eval judge is getting smarter, but the RLS guide fix is the one that will keep your docs green.

The evals repo's default judge moved from gpt-5.5 at low reasoning to gpt-5.6-sol at medium effort [1], a shift that directly addresses the weakest link in your scoring pipeline. Separately, the RLS guide eval softened its pgTAP coverage rubric after a bonus test-quality check kept turning the whole doc red while all 40 RLS checks passed [2]. Over in the CLI, the monorepo consolidated on TypeScript 7 and bumped the entire Effect family to RC.111, removing stale beta pins and parallel aliases [3]. The stack runtime is now Effect-native across Node and Bun, with on-demand slim resource preparation via `off`, `lazy`, and `eager` policies [4]. Studio shipped three quality-of-life wins: permission presets for scoped PATs [5], preserved assistant tool previews after completion [6], and a restored notebook diff preview for already-applied updates [7]. ETL now handles flexible BigQuery column names across initial copy, CDC, schema changes, partitioning, and clustering [8], and Realtime surfaces the Postgres Changes connection pool in settings [9]. The MCP server instructions are now intent-based, fixing broken tool references on hosted platforms [10].

Action items

→ Pull the latest evals to get the gpt-5.6-sol judge and the softened RLS rubric supabase/evals [immediate]
→ Review the CLI Effect RC.111 upgrade for flag and output changes before adopting supabase/cli [plan]
→ Watch BigQuery flexible name support for any CDC or partitioning edge cases supabase/etl [monitor]
→ Check the Realtime dashboard for the new Postgres Changes pool setting supabase/supabase [monitor]

References

[1] fix(core): default the judge to gpt-5.6-sol at medium reasoning effort (#222) supabase/evals
[2] fix(evals): soften the pgTAP coverage rubric in the RLS guide eval (#224) supabase/evals
[3] chore(deps): consolidate TypeScript and update Effect (#6289) supabase/cli
[4] feat(stack): prepare slim resources on demand (#6250) supabase/cli
[5] feat(studio): add permission presets to scoped pat creation form ↗ supabase/supabase
[6] feat(studio): preserve assistant tool previews after completion ↗ supabase/supabase
[7] Restore notebook diff preview for completed updates ↗ supabase/supabase
[8] feat(bigquery): Add support for flexible names ↗ supabase/etl
[9] fix(realtime): add pg changes pool to realtime settings ↗ supabase/supabase
[10] fix: make server instructions intent-based instead of naming tools (#372) supabase/mcp

Quick answers

What shipped in Supabase on August 22, 2026?: Supabase's eval judge is getting smarter, but the RLS guide fix is the one that will keep your docs green. In total, 41 commits, 46 pull requests, and 5 releases landed.
Who contributed to Supabase on August 22, 2026?: 12 developers shipped this update, including Miranda Limonczenko, supabase-evals-releaser, claude[bot], charislam, kemaldotearth, SaxonF, claude, and filipecabaco, and 4 more.
What were the notable Supabase updates?: fix(core): default the judge to gpt-5.6-sol at medium reasoning effort (#222), fix(evals): soften the pgTAP coverage rubric in the RLS guide eval (#224), and chore(deps): consolidate TypeScript and update Effect (#6289).

MCP SERVERS MOVE TO SDK V2

Supabase's MCP packages cut over to the v2 wire protocol today, and if you consume them, the next upgrade is not optional.

+11

javascript 126 shipped

@supabase Aug 20

ROW LEVEL SECURITY DOCS GET A REWRITE

Supabase restructured its RLS documentation overnight, splitting the guide into clear sections so developers can find what they need without wading through reference material.

javascript 56 shipped

@supabase Aug 19

SUPABASE CLI v2.115.0 BREAKS YOUR CI

If you run schema diffing or supabase test db in CI, this release is about to change what those commands do.

javascript 79 shipped

@supabase Aug 18

SUPABASE HARDENS CORS AND SHIPS PROMPT-INJECTION EVALS

A config hardening for self-hosted envoy plus four new security evals signal Supabase is closing attack surfaces on both the infrastructure and AI agent fronts.

javascript 85 shipped

Elsewhere on the wire

AI Agents Aug 22

Cline hubs get drain and replay as Langchain plumb gates

Cline just shipped rollover-grade hub commands while Langchain exposes gateway internals, and OpenHands quietly broke its own monitoring contract.

ai-agents 42 shipped

Local LLMs Aug 22

SGLang ships v0.5.18 with 710 PRs and DeepSeek V4 FP4 auto-routing

The biggest release of the week just landed: SGLang v0.5.18 adds new models, fixes a silent disagg bug, and defaults DeepSeek V4 FP4 checkpoints to the right MoE runner.