Dev Tools & SDK Changelogs
Changelogs and version notes for AI developer tools, SDKs, and CLIs.
Paperclip hires agents, not prompts — and a human board
Paperclip (73.5k stars) runs a coordinated company of AI agents — CEO, marketing, eng, ops — with hard budget limits,
30+ models, zero auction control — what Higgsfield actually is
Higgsfield AI: 30+ models for ad creative, no auction or targeting. How it fits upstream of PMax and Advantage+.
76 malicious skills cleared skills.sh — 8 stayed live
Snyk ToxicSkills audit found 36.8% of 3,984 agent skills flawed, 76 confirmed malicious with reverse shells.
video-use feeds Claude Code a transcript, not 45M frame-tokens
video-use (browser-use org) edits video via packed transcript and ffmpeg EDL pipeline — no timeline UI required.
53k devs are reading Claude and Cursor's actual system
GitHub repos like asgeirtj/system_prompts_leaks (53.5k stars) and x1xhlol (139k+) archive extracted system prompts for
Sonnet 5 is default — your token count just jumped ~30%
Claude Code v2.1.197 sets Sonnet 5 as default; new tokenizer inflates token counts ~30%, breaking existing API patterns.
Shell failures now explain themselves in Claude Code
Claude Code weeks 24–26 in June 2026: /cd, recursive delegation, live pages, shell auto-response, and Sonnet 5.
Grok joins IBKR's AI suite — the "live" part deserves scrutiny
Grok joins IBKR's MCP broker AI. Covers portfolio scope, order drafting, and the "live market analysis" precision gap.
Ten Codex alphas, none with notes — what the CI burst means
Ten Codex 0.143.0-alpha tags in June 2026, none with notes. What 65 revisions of engineering work actually covered.
Hermes Desktop went live; the Computex story is unconfirmed
Nous Research's Hermes Desktop adds a native GUI to the open-source Hermes agent — what's real and what's unconfirmed.
No login, no deploy: thethings.ai gives AI a direct page
A live write endpoint for AI: POST HTML, get a URL via MCP or REST. SLA, pricing, and moderation policies unconfirmed as of June 2026.
Cordium: AI hits the DB with no password in the process
Cordium (Apache-2.0) lets AI workloads hit live databases and APIs — credentials sealed at the ZTNA proxy, never injected into the process.
Rust's borrow checker reaches CUDA — and the cost is zero
NVIDIA & Hugging Face's cuTile Rust: borrow-checked CUDA at 96% cuBLAS; Grout decoder matches vLLM at batch-1 on B200.
Officially obsolete: the base_url trick for OpenRouter
0.2.4 adds parallel_tool_calls, pins openrouter SDK 0.9.2, and formalizes the OpenRouter adapter with correct provider attribution and telemetry.
SimpleMultiModalQueryEngine is deprecated. Here's the swap.
LlamaIndex 0.14.23 deprecates SimpleMultiModalQueryEngine, unifies rich-media RAG, and fixes workflow state bleed.
Compaction has no opt-out in 0.143.0. Here's what to adjust.
Codex 0.143.0-alpha.14 annotated: compaction opt-out removed, 61% fewer filesystem RPCs, new TOML source allowlists.
Ten patches blocked git wipeouts. The annotated sprint.
Destructive git blocking, org restrictions, MCP timeouts, Fable 5 suspension: Claude Code's June 12–23 sprint annotated.
An accountant can outperform a senior dev on Claude Code
Domain expertise, not coding background, predicts Claude Code success. Key data: 398k sessions, Oct 2025–Apr 2026, published June 16, 2026.
CC-BY-NC is out. Does Cohere's '4-bit lossless' claim hold?
Command A+: Apache 2.0, 218B/25B MoE on 2×H100, 48 languages — W4A4 'lossless' claim needs independent verification.
Ship first, sign in later: what Cloudflare's --temporary does
Cloudflare's --temporary: agents get a live workers.dev URL with no login — 60-min window, then deleted.
Copilot Cowork asks permission — unless you're the recipient
Copilot Cowork's Skills injection + self-send gap = silent M365 exfiltration, per PromptArmor's May 2026 POC.
Grok Build's /goal: when 'Complete' appears, what was checked?
xAI's /goal: Grok Build takes an objective, runs until done, with built-in verification and four steering commands.
TVM 0.25's TIRx reveals what Triton deliberately conceals
TVM v0.25.0 adds TIRx: explicit Blackwell orchestration, 18-pass lowering chain, 29 tile primitives, no autoscheduler.
Fugu beats each component it calls — and the asterisks matter.
Fugu coordinates frontier LLMs via one API. What the benchmark sheet covers, what it omits, and caveats for builders.
AWS Context skips RAG. Continuum validates CVEs itself.
At AWS Summit NY 2026: Context (knowledge graph for agents) and Continuum (CVE lifecycle). Neither is GA.
A JEPA that learns coordinates it was never given
DVD-JEPA: 10s CPU training, browser demo, 32-dim latents — the smallest JEPA toy; not the first. MIT-licensed, June 2026
Worktrees aren't enough — 27.67% of AI PRs hit merge conflicts
Worktrees, serialized merge trains, and the 27.67% conflict rate in AI-authored PRs. CAID, STORM, and current tooling.
11 Codex Rust alphas in 90 hours — and still no stable tag
11 Codex Rust 0.142.0 alphas in 90 hours: Noise relay, per-thread stdio MCP, SQLite WAL, and P-521 TLS annotated.
ChatGPT monitors unattended — it alerts, never acts
Scheduled page, monitoring, and plan quotas land in June 2026. All tasks are notify-only — no writes or transactions.
α-entmax tile-skips what softmax can't; 354M unconfirmed
AdaSplash-2: α-entmax exact-zeros, Triton tile-skip, BSD-3 license. 354M checkpoint not confirmed public.
Cloak's daemon owns the HTTPS — raw vault values never exit
Cloak v1.1.2 routes API calls through a daemon so MCP agents never receive raw credentials. Architecture, callable surface, and honest limits.
cuTile Rust at 96% cuBLAS — Grout's engine wins need context
cuTile Rust carries Rust ownership to CUDA kernels. Grout hits 96% of cuBLAS — with vLLM/SGLang prefix caching off.
Your agent reads SKILL.md. SkillsGuard reads it first.
Zero-dep static scanner for AI skill packages — 151 rules, SARIF, MCP server. What SkillsGuard catches, what it misses, and how it compares to Cisco and NVIDIA SkillSpector.
One GitHub PR is how you ship to Grok Build's marketplace
What each Grok Build plugin bundles, who the day-one partners are, and how to submit your own extension via GitHub PR.
Copilot GA'd as a standalone workspace — not as a GitHub App
GitHub Copilot's standalone app GA in 2026: worktree isolation, org admin gating, and open spending questions explained.
quicktok: 11× on tiktoken, author-reported, no README
quicktok: C++20 SIMD tiktoken replacement, byte-identical, 11× reported. No README. No independent reproduction.
What v3 event stream was silently losing — LangChain 1.4.8
langchain-core 1.4.8 + 1.3.10 (June 18): v3 metering gaps, BaseTool schema caching, and gpt-5.x routing correction.
Firefly executes Photoshop jobs. Not uniformly, though.
Firefly in Photoshop 27.9, Premiere 26.3, Illustrator, InDesign, Frame.io. Capabilities, 30-model backend, no SDK yet.
Seven alphas later — what Codex CLI 0.141.0 actually delivers
Codex CLI v0.141.0 (stable): Noise relay, SQLite WAL pin, Windows hardening. v0.142 hit seven alphas in 48 hours.
xAI's Office sidebar is in PowerPoint. What goes where?
xAI's Office add-ins for PowerPoint, Word, and Excel. What content leaves your org and which protections apply per plan.
GLM-5.2 is #1 in frontend — once you check who's absent
GLM-5.2 at #1 on Design Arena WebDev Elo and #2 on Code Arena frontend — Fable 5 absent, Opus 4.8 still leads SWE-Pro.
langchain-model-profiles writes _profiles.py, not profiles.py
LangChain's capability metadata CLI for 0.0.6: refresh cycle, field mapping, and the _profiles.py naming inconsistency.
Rust 0.142 moved fast. The documentation didn't.
Six 0.142 alphas, no release notes. PRs show: Noise relay, SQLite WAL fix, off-by-default subagent gate.
GLM-5.2 hit 46.2 on DeepSWE. M3 hit 20.0.
GLM-5.2 leads on four shared agentic rows; M3 brings multimodal input and cheaper output. Benchmarks, scaffolding caveats, licensing, and pricing.
K2.7 deliberates on every call — the off-switch doesn't exist
K2.7 Code forces reasoning on every call, caps replies at 32K tokens, and ships under a Modified MIT license — here's the full deployment picture.
Dream Server rewrote a 2,591-line script into 6 libraries
Dream Server: 13 AI services wired by a Docker bootstrap. What v2.5.3 includes, what's validated, what isn't signed.
v2.1.172 made nesting 5 deep possible. v2.1.181 had to cap it.
Claude Code v2.1.170–v2.1.181, June 9–17: Fable 5, scoped permissions, sub-agent nesting, /config inline annotated.
Every Grok call in Warp passes through Warp's backend anyway
How SuperGrok links to Warp: model IDs, backend proxy, ZDR tradeoffs, and what bypasses your subscription.
23 schemas, 33 ms each. LangChain just memoized them.
June 18: v3 cost attribution restored, schemas 100× faster, multimodal summarizer fixed. Repin away from yanked 1.3.5.
Private infra now tunnels to OpenAI — no public URL needed
OpenAI Python SDK v2.43 adds tunnel_id for private MCP connectivity; v2.42 adds spend alerts and admin key expiry.
Refusals now return HTTP OK — and your try-except misses them
v0.108–v0.109 adds refusal responses at HTTP 200, a new stop_reason literal, and removes retired model strings from typed enums.
Grok now edits your Word doc — and sends it to xAI servers
Grok for Word and PowerPoint launched June 16–18, 2026. Free Marketplace install, paid plans required, document content sent to xAI off-device.
v0.109: Fable 5's refusal arrives as HTTP 200
SDK v0.108–0.109: Fable 5 refusal-in-200, Managed Agents cron, vault credentials, and an enum fix in 0.109.1.
LLM blackmail: 84% to 0% — and 'why' beat 'what' by 28×
SDF, synthetic stories, and 3M moral-dilemma tokens: how Anthropic cut LLM blackmail from 84% to 0% in May 2026.
xAI's Grok plugin catalog pins commits, not responsibility
Grok Build's plugin marketplace (June 2026): skills, hooks, MCPs in one bundle, SHA-locked, not audited by xAI.
Anthropic paused the SDK billing split on June 15 itself
Anthropic paused the SDK credit split. Live usage credits system, CI cost baselines, and what triggers a restart.
Search as Code hits 100% on CVEs. Perplexity wrote the test.
Perplexity's SaC: model-written search pipelines, composable primitives, live pricing, and unverified benchmarks.
Distillation went mainstream. Then came the IP dispute.
R1's MIT license made open-source distillation legal. V4's OPD is how it works now. The IP dispute explains the limits.
M3 Swaps Quadratic Attention for Blocks — arXiv Just Submitted
MiniMax M3: 428B-param MoE, MSA replaces quadratic attention at 1M positions. HuggingFace weights; benchmarks vendor-run.
Gemma 4 12B says 16 GB. BF16 full precision needs 24 GB.
Encoder-free multimodal, 256K context, Apache 2.0. Specs, evals, and the memory math behind Gemma 4 12B's 16 GB claim.
The LLM that chose deletion — and the sim designed to find it
LLMs inhabit a persistent society for 15 real days: 9 metrics, Season 1 survival 10/10 vs. 0/10 by foundation model, open-source.
Writer and Rewriter Got Polyfills. Proofreader Got Nothing.
Chrome's Writer, Rewriter, and Proofreader origin trials: polyfills, interface options, and device constraints.
The parity fix that quietly resets your profiling baseline
llama.cpp b9437: -fa auto added to llama-bench, -ngl default flips to -1. What changes and who's affected.
Mariner retired at I/O 2026. The VM-backed successor is Antigravity.
Antigravity 2.0, Managed Agents API, Gemini 3.5 Flash GA, and the Mariner shutdown: I/O 2026 developer rundown.
v0.22.0: the chips sat idle while the front end was choking
vLLM v0.22.0 ships --api-server-count, a DP Supervisor, and three LB topology modes. Annotated explainer for operators.
Windsurf is Devin now — Cascade retires July 1
Cognition renamed Windsurf to Devin Desktop June 2. What changed, what broke, and what IT admins need to do now.
4 GitHub stars, voice interviews with Ollama: that's GrillKit
Apache 2.0 interview trainer with Whisper voice input, Ollama or cloud LLM support, and local session history. No SaaS, no registration required.
RDNA3 cuts llama.cpp KV VRAM 47% — and CUDA has no equivalent
RDNA3 bit-packing cuts llama.cpp KV VRAM 47% on RX 7900. Flags, VRAM math, and TurboQuant for 4.9× compression.
NodeCartel is dark. Cross-host AI orchestration: who delivers.
NodeCartel is unreachable. Kore.ai, CrewAI Cloud, Northflank, and AgentNode Pro compared for cross-host AI scheduling.
17k tokens → 1.4k — Headroom keeps the originals retrievable
Open-source context compression middleware for agent pipelines: 60–95% token cuts, CCR reversibility, AST-aware engines.
NVIDIA cut Qwen3.6-35B 3×. Accuracy barely moved.
NVIDIA's NVFP4 Qwen3.6-35B checkpoint on HuggingFace: 3.06× memory reduction, <1% accuracy loss, Blackwell-native, vLLM flags included.
Overslash holds the credentials. Your AI only gets a handle.
Overslash injects secrets by handle at the gateway, limits blast radius per agent, and escalates out-of-scope calls to human approval. Free self-hosted or €3/seat cloud.
Harness edits slowed the GPU kernel 10×. Weights fixed it.
SIA edits its scaffold and fine-tunes weights via LoRA — 70.1% LawBench, 12.4% faster GPU kernels, MIT-licensed.
Why langchain-perplexity 1.3.1 Dropped Its SSE Shim
1.3.0 added use_responses_api for Perplexity's Responses endpoint; 1.3.1 removed the SSE shim 0.34.0 required.
459 Commits Into vLLM 0.22.0 — What Moves the Needle
459 commits, a dedicated DeepSeek V4 package, Rust frontend, and an rc0 that's one CI fix. What matters and what doesn't.
Google Beam's 3D Group Calls Need a $24,999 Display to Work
Google I/O 2026 extended Beam to multi-person calls. Here's the AI pipeline, the $24,999 display, and where the gaps are.
ChatPerplexity Auto-Routes to Real-Time Search in LangChain
ChatPerplexity gains use_responses_api in 1.3.0: auto-routes to Perplexity's Agent API for real-time search.
Inject Constraints Mid-Run Without Breaking the Prompt Cache
Mid-conversation constraint injection in v0.105.0 preserves prompt cache continuity across long inference runs.
Copilot Cowork Silently Exfiltrates SharePoint — No Patch Yet
A 5-line poisoned Skills script silently exfiltrates SharePoint data via Copilot Cowork — no approval gate, no CVE, no patch.
Anthropic SDK 0.105.0 Needed Two Hotfixes — What to Pin
Two rapid patches followed Anthropic's 0.105.0 drop. Here's what broke, why, and which version to pin.
MCP Credential Leak Closed in Claude Code's Busiest Week Yet
Seven builds in one week: four Bash/PowerShell sandbox bugs patched, /code-review --fix lands auto-apply, and a serious MCP auth credential leak is closed.
Docs Live and Gmail Live Are Real — Here's Who Gets Them First
Docs Live, Gmail Live, Gemini Spark, Sheets one-shot: I/O 2026 Workspace features and who gets access first.
Anthropic 0.105.0 Adds Output Attribution — What It Buys You
v0.105.0 adds granular output-type attribution and configurable upload caps—here's what they do and when to use them.
vLLM v0.21.0 Production Update: KV Offload and Multi-Server Port Bug
v0.22.0 doesn't exist yet. v0.21.0 ships KV offload, spec decode, and a multi-server port bug still under review.
The Claude Code Sprint That Patched Four Security Holes
Ten patches in nine days: pinned sessions, four security fixes, /code-review --fix, and skill-level tool gating.
Codex CLI 0.134.0 and 0.135.0: Two Stable Releases in 48 Hours
OpenAI shipped two Codex CLI stable releases in 48 hours. What changed, what broke, and why the cadence matters.
Anthropic Python SDK 0.105: Opus 4.8 and Mid-Session System Prompts
Three SDK releases in 7.5 hours ship claude-opus-4-8 support, mid-conversation system blocks, and finer output usage reporting.
Grok Build Lands in OpenCode and Kilo Code: xAI's 13-Day Rollout
xAI shipped grok-build-0.1 to three developer tools in 13 days. Here's what each integration covers and how to pick the right surface.
What Codex CLI's 0.135.0 'Stable' Release Actually Fixed
OpenAI's 0.135.0 stable is a diagnostics and polish cycle. What moved in the TUI, Vim mode, and remote transport.
Two Codex Alphas in 3 Hours — and the Release Notes Errored
Two alpha releases in three hours, 529 files changed. Here's what the diff says when the release notes page errors.
xAI's Coding Agent Reads Your CLAUDE.md. Should You Use It?
xAI's Grok Build ships with Arena Mode, reusable Skills, and CLAUDE.md compat. Here's what developers need to know.
Codex CLI 0.134.0 Kills Your Legacy Profile Config
v0.134.0 ships local history search, per-server MCP env vars, OAuth for HTTP transports, and kills legacy v1 profile configs.
How a Poisoned OneDrive File Silently Pulls Your M365 Data
PromptArmor shows how a poisoned SKILL.md in OneDrive lets attackers silently pull M365 files — no approval dialog, no user alert.
vLLM RC3 Fixes a Hard-Coded 60s Timeout — What to Configure
RC3 patches a hard-coded 60s startup timeout in vLLM's multi-API-server subsystem — here's what changed and what operators must configure.
openai-codex b1→b2 in Four Hours — What the Cadence Reveals
Two beta releases in under four hours. Here's what the b1→b2 patch cadence tells developers about SDK maturity and what to pin.
$24,999 디스플레이와 Google Beam이 아직 못 하는 것
Google I/O 2026 extended Beam to multi-person calls. Here's the AI pipeline, the $24,999 display, and where the gaps are.
ChatPerplexity 1.3.0, 실시간 검색 자동 라우팅이 된다
ChatPerplexity gains use_responses_api in 1.3.0: auto-routes to Perplexity's Agent API for real-time search.
대화 중간 제약을 바꿔도 프롬프트 캐시가 끊기지 않는다
Mid-conversation constraint injection in v0.105.0 preserves prompt cache continuity across long inference runs.
5줄 스크립트로 SharePoint가 조용히 유출된다
A 5-line poisoned Skills script silently exfiltrates SharePoint data via Copilot Cowork — no approval gate, no CVE, no patch.
Anthropic SDK 릴리즈가 PyPI 배포를 깨뜨린 이유
Two rapid patches followed Anthropic's 0.105.0 drop. Here's what broke, why, and which version to pin.
Claude Code MCP 크리덴셜 유출이 패치됐다
Seven builds in one week: four Bash/PowerShell sandbox bugs patched, /code-review --fix lands auto-apply, and a serious MCP auth credential leak is closed.
Google Workspace Live, 기능 접근 순서가 정해졌다
Docs Live, Gmail Live, Gemini Spark, Sheets one-shot: I/O 2026 Workspace features and who gets access first.
Anthropic SDK 출력 귀속, 코드에서 실제로 뭐가 달라지나
v0.105.0 adds granular output-type attribution and configurable upload caps—here's what they do and when to use them.
vLLM 최신은 v0.21.0, 포트 버그는 아직 미해결
v0.22.0 doesn't exist yet. v0.21.0 ships KV offload, spec decode, and a multi-server port bug still under review.
Claude Code, 9일 만에 보안 구멍 4개를 닫았다
Ten patches in nine days: pinned sessions, four security fixes, /code-review --fix, and skill-level tool gating.
Codex CLI 0.134.0 & 0.135.0: 48시간 안에 안정 버전 2개 출시
OpenAI shipped two Codex CLI stable releases in 48 hours. What changed, what broke, and why the cadence matters.
Anthropic Python SDK 0.105: Opus 4.8 및 미드-세션 시스템 프롬프트
Three SDK releases in 7.5 hours ship claude-opus-4-8 support, mid-conversation system blocks, and finer output usage reporting.
Grok Build, OpenCode·Kilo Code에 상륙: xAI의 13일 롤아웃
xAI shipped grok-build-0.1 to three developer tools in 13 days. Here's what each integration covers and how to pick the right surface.
Codex CLI Doctor가 생겼다, TUI와 Vim 모드도 달라졌다
OpenAI's 0.135.0 stable is a diagnostics and polish cycle. What moved in the TUI, Vim mode, and remote transport.
Codex CLI alpha, 릴리즈 노트 오류 뒤 529개 파일의 실체
Two alpha releases in three hours, 529 files changed. Here's what the diff says when the release notes page errors.
openai-codex 4시간 만에 재패치, SDK 성숙도를 어떻게 볼까
Two beta releases in under four hours. Here's what the b1→b2 patch cadence tells developers about SDK maturity and what to pin.
AI developer tools and ecosystem news for developers and technical founders
