ISSUE #8, AUG 24 2026, 574 CHANGES GRADED

What changedin your stackthis week.

A weekly digest of what changed in 30 developer tools, classified by severity, so you catch the breaking changes before your next deploy.

live triagereading changelogs
v0.40.0v0.40.0
v0.22.53v0.22.53
@ai-sdk/deepgram@3.1.0@ai-sdk/deepgram@3.1.0
v16.4.0-canary.2v16.4.0-canary.2
v8.0.0-rc.5v8.0.0-rc.5
@ai-sdk/xai@4.0.43@ai-sdk/xai@4.0.43
v0.40.0v0.40.0
v0.22.53v0.22.53
@ai-sdk/deepgram@3.1.0@ai-sdk/deepgram@3.1.0
v16.4.0-canary.2v16.4.0-canary.2
v8.0.0-rc.5v8.0.0-rc.5
@ai-sdk/xai@4.0.43@ai-sdk/xai@4.0.43
574 releases in4 crit · 13 notable

Three tiers.
One rubric.

Severity is always color plus a shape plus the word, never color alone. Open a tier to see the real thing from Issue #8.

CRITICAL4 IN ISSUE #8+

A breaking change in a stable release, a security fix, or a forced migration, proven by a verbatim quote from the release notes. The only tier with a "Do this:" action box. Read it before your next deploy.

DIRECTUSv12.3.0

A bug in @liveblocks/client could cause data loss by overwriting room storage when large WebSocket messages are sent.

Do this: update to v12.3.0 immediately.

NOTABLE13 IN ISSUE #8+

Worth knowing, new capabilities, meaningful features, or non-forced migrations you will want on your radar. Nothing that breaks on deploy.

v0.40.0, pocketbase v0.40.0 changes command exit-code propagation so failures now correctly exit non-zero, which can break shell chains relying on the old always-zero behavior. it also bumps the minimum go version to 1.27.0 and migrates to encoding/json/v2, which isn't fully backward compatible
@ai-sdk/deepgram@3.1.0, this minor release of the deepgram provider fixes several transcription options that were silently dropped, and changes the default for speaker diarization from on to off since it's a paid add-on. it also adds voice/language composition for speech, usage metadata, speed passthrough, and better error parsing
@ai-sdk/harness@1.0.81, the ai sdk harness package makes the `destroy` method mandatory on `harnessv1networksandboxsession` implementations, a type-level api change. custom sandbox session implementations that omit `destroy` will fail to type-check
FYI557 IN ISSUE #8+

Patches, dependency bumps, and prereleases, collapsed into a single count so the signal stays clean. Prereleases are capped here in code and cannot be inflated.

+557 patches, dependency bumps, and prereleases in Issue #8, folded into one line so you can skip them with a clear conscience.

HOW IT WORKS

Read all of it,
so you don't.

01/02/03
01
FETCH
Every release, 30 repos.

The pipeline pulls new releases from 30 GitHub repos every day, patches to major versions, nothing skipped.

02
CLASSIFY
Graded by a public rubric.

Claude grades each release CRITICAL / NOTABLE / FYI. Every CRITICAL needs a verbatim quote from the notes; prereleases cap at FYI.

03
REVIEW
Human-supervised QA.

A second AI editor re-reviews every classification each week; corrections are validated and applied, and the rubric is amended when a pattern repeats.

THE PROOF, ISSUE #8
30
TOOLS TRACKED
574
CHANGES · ISSUE #8
4
CRITICAL SURFACED

30 tools, one signal.

StripeSupabaseNeonTursoVercelRailwayCloudflare WorkersPrismaDrizzle ORMKyselyClerkAuth.jsBetter AuthResendTrigger.devInngestVercel AI SDKOpenAI SDKAnthropic SDKNext.jstRPCshadcn/uiRadix UITailwind CSSBunDenoPayload CMSDirectusPocketBaseLiveblocksStripeSupabaseNeonTursoVercelRailwayCloudflare WorkersPrismaDrizzle ORMKyselyClerkAuth.jsBetter AuthResendTrigger.devInngestVercel AI SDKOpenAI SDKAnthropic SDKNext.jstRPCshadcn/uiRadix UITailwind CSSBunDenoPayload CMSDirectusPocketBaseLiveblocks

Payments · databases · auth · AI SDKs · frontend · runtimes · infra. Added to monthly by request.

FAQ

Why not just watch GitHub releases directly?+

Watching 30 repos is 30 noise streams with no triage, 8 Clerk packages bumping the same day, 6 Cloudflare Workers SDK releases in a week, and no signal about which ones actually break your app. StackTrace reads all of them and grades each against a public rubric, so you get the 1 thing that matters, not 100 things that don't.

How do you stop everything from being marked critical?+

The rubric is deterministic and conservative: CRITICAL requires a breaking change in a stable release, a security fix, or a forced migration, and a verbatim quote from the release notes proving it. Prereleases are capped at FYI in code, and any issue with more than 4 CRITICALs is automatically held for review. Ties go to the lower tier. We never promote to fill a slot.

Isn't this just an LLM summarizing changelogs?+

The classification uses Claude, but it runs inside code-enforced guardrails: ungrounded CRITICALs are rejected before an issue can ship, prereleases can't be inflated, and a second AI editor re-reviews every classification each week. The rubric is public, you can read the exact instructions the model gets.

Which tools do you cover, and can you add mine?+

30 tools today across payments, databases, auth, the AI SDKs, frontend, runtimes, and infra. The list is reviewed monthly and grows based on what subscribers use, reply to any issue to request one.

How often does it send, and what does it cost?+

One email a week, free. No account, one-click unsubscribe. Built to be read in two minutes: criticals first, notable next, everything else collapsed.