Issue 034 — window: 2026-09-17 06:00 UTC to 2026-09-18 06:00 UTCnofeed.dev
No. 034 · 18 September 2026 · 3 min← No. 033
SIGNALNOTABLE|SHIPPED 3|SKIPPED 5

The one thing

01

Two harnesses shipped a build that would not work, and fixed it the same night.

Claude Code 2.1.275 went out at 22:33 UTC. Every request failed with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL pointed at a proxy or gateway. 2.1.276 landed at 02:12 UTC with that one line as its entire changelog — three hours thirty-nine minutes of a build that could not reach a gateway.12

Cline's desktop app did the same thing four hours apart. 0.0.31, 21:41 UTC, could fail at launch with desktop backend exited before publishing its endpoint: the app and the Cline Hub it starts had been built from different copies of the same code, so the app shut down the Hub it had just launched. 0.0.32, 01:42 UTC, boots both halves from one copy.34 Caught and fixed inside four hours each, and shipped to users first.

VERDICT · BREAKING — On Claude Code 2.1.275 behind a gateway, or Cline desktop 0.0.31: you are broken, and the fix is published. Upgrade both.

Shipped

03
Breaking

`ubuntu-latest` moves to 26.04 between 19 October and 19 November

Ubuntu 26.04 is GA for GitHub Actions on x64 and arm64, and the ubuntu-latest label migrates off 24.04 gradually across that month-long window. GitHub says the image removes tools and changes tool versions, and that the move may break builds depending on either. Two options and a deadline: test against runs-on: ubuntu-26.04 now, or pin ubuntu-24.04 and decide later.5

MATTERS TO · Every repository whose CI says ubuntu-latest — which is most of them.
nofeed.dev/issues/2026-09-18/ubuntu-latest-26-04/
Use it

Cline 4.1.19 stops guessing when to compact

The compaction trigger had compared a character-based estimate of the conversation, around three characters per token, against the model's context limit. Content that tokenizes denser — disassembly, image dumps, minified sources — filled the real window while the estimate stayed under the threshold, compaction never fired, and the turn was squeezed into a handful of output tokens. The trigger now also uses the token count the provider reports.6

MATTERS TO · Anyone pointing Cline at minified bundles, binaries or screenshots and watching long tasks die quietly.
nofeed.dev/issues/2026-09-18/cline-4-1-19-compaction/
Wait

Cline desktop 0.0.31 drives a remote machine over SSH

The feature under the broken build: Cline runs agent tools, workspace discovery, Git metadata and session persistence on a remote host while the app stays local, configured under Settings → Remote. Approvals return to the desktop. Host metadata is written at mode 0600 and records an identity-file path, never key contents. Linux x64 and arm64 hosts only, and one day old.4

MATTERS TO · Anyone who wants an agent's file writes and shell to happen on a dev box rather than a laptop.
nofeed.dev/issues/2026-09-18/cline-desktop-remote-ssh/

Promised, not shipped
Codex rust-v0.155.0 ships `/voice` conversations behind `/experimental`, plus Touch ID verification for MCP requests on supported Macs. Available, unevaluated by us.[^7] · openai-python 3.15.0 adds prompt-cache prewarming and managed Responses WebSocket sessions; langchain-typesafe 0.0.1a2 is a twelve-hour-old alpha; Anthropic's Compliance API now returns Claude in Chrome transcripts, Enterprise beta only.[^8][^9][^10]

The conversation

01
The claim — including ours

Open-weight models now run the majority of production tokens.

From the floor
  • Vercel AI Gateway Production Index, September 2026confirmed against the named vendor page, opened and read

    Open-weight models ran 56% of gateway tokens in August, up from 7% in December, and took 14% of spend. Price per token fell 23.2% in August, a third consecutive monthly drop; among teams over ten million tokens in both months the median paid 7.6% less. Fable 5 fell from 13.2% of spend to 4.9% while Opus 5, at roughly half the price, rose to 22.5%. Anthropic kept 64 cents of every dollar.11

Our take

One gateway's traffic is one gateway's customers, and Vercel sells the routing that makes model-switching cheap, so 56% is a fact about this population rather than about the industry. What it supports is narrower and still useful: teams here stepped down a tier rather than across a lab, while the frontier defended spend by cutting cache reads to $0.25 per MTok.10

Hacker News ran as the community layer: 29 stories cleared 80 points inside the window. The X layers did not run today — the newest file in both digests is dated 2026-09-16 05:30 UTC, roughly 48 hours old, read as background only; the account roster watches 102 accounts. The Reddit digest's newest file is 2026-09-17 05:30 UTC, collected before this window opened, also background only. No practitioner report is printed above.

Skip this

05

Everything we saw

62
62 candidates scanned · 4 used in this issue — the rest, with the reason each one was left out
The receipt. Ranking is only trustworthy if the discarded pile is visible, so here it is: everything the collectors surfaced in the window, with its signal and what we did with it.
ItemSourceSignalCall
anthropics/claude-code v2.1.276GitHub Releasessingle-line changelog, 02:12 UTCLed the issue
anthropics/claude-code v2.1.275GitHub Releasesthe regressionPrinted with 2.1.276
cline/cline desktop-v0.0.32GitHub Releaseslaunch fix, 01:42 UTCPrinted
cline/cline desktop-v0.0.31GitHub Releasesremote SSH, failed to launchPrinted
cline/cline v4.1.19GitHub Releasescompaction triggerPrinted
Ubuntu 26 generally available and latest migrationGitHub changelogubuntu-latest migrates 19 Oct – 19 NovPrinted
AI Gateway Production Index, September 2026Vercel blog56% of tokens open-weightUsed as the argument; read as one gateway's population
openai/codex rust-v0.155.0GitHub Releases/voice, Touch ID for MCPPromised, not shipped: experimental flag
openai/openai-python v3.15.0GitHub Releasesprompt-cache prewarmingPromised, not shipped: unmeasured
langchain-typesafe 0.0.1a2GitHub Releasesinitial releasePromised, not shipped: alpha
Claude Platform release notes, 18 SeptemberAnthropic docsChrome session transcripts in Compliance APIPromised, not shipped: Enterprise beta. Also the source for Fable 5.1 pricing
pydantic/pydantic-ai v2.45.0GitHub Releasesdurable-run fixesOne MCP session per durable run instead of one per unit, and one DynamicToolset resolution per run — real for durable agents, held for space. Yesterday's 2.44.0 advisory call stands.
block/goose v1.51.0GitHub Releaseseight featuresGateway pairing allowlist and GPT-live support; nothing that forces an upgrade
All-Hands-AI/OpenHands v1.20.0GitHub Releasesthree featuresPer-profile secret scoping and saved profiles in automations; no pin change
openai/openai-agents-python v0.22.3GitHub Releasesnine fixesConditional-approval alignment and session concurrency fixes; patch-level
vercel/ai 7.0.106, 6.0.286, 5.0.260GitHub Releasesthirteen patch fixesStream and telemetry fixes across three majors; no API change
alibaba/open-code-review v1.12.5GitHub Releasesmostly viewer restylingNamed in Skip This
Cursor: Cloud Agents and Automations degradationCursor statusminor, 4h42mLongest incident in the window and the one that hits unattended runs; duration computed from the page's own timestamps.[^12]
Cursor: Grok Bot degradationCursor statusmajor, 2h06mCursor's second incident of the window, again in a third-party model.[^13]
OpenAI: elevated errors affecting ChatGPT Work modeOpenAI statusmajor, 1h24mChatGPT Work mode; no developer decision follows from it.[^14]
OpenAI: elevated error rates across API modelsOpenAI statusminor, 24mElevated errors across API models, short.[^15]
GitHub: elevated errors for OpenAI models in CopilotGitHub statusminor, 48mShort and minor; the third-party model again
Workflow execution protections in GitHub Actions GAGitHub changelogGAAllowlists controlling who can run workflows — real for anyone letting agents open PRs against a shared repo, held for space rather than doubt
Agentic CLI customizations in the usage metrics APIGitHub changelognew API fieldsReporting on skills, custom agents, MCP servers and slash commands; useful to platform teams, changes no install
Run Terminal-Bench and other Harbor evals on Vercel SandboxVercel changelogchangelog entryWorth a look if you already run Harbor; no figures published
The skills CLI now supports Notion hosted skillsVercel changelogchangelog entryOne hosted provider added; narrow
Turbo build machines can now be enabled per deploymentVercel changelogchangelog entryBuild-infrastructure toggle; off beat
Sub-second artifact deployments in Vercel CLIVercel changelogchangelog entryNamed in Skip This: a speed claim with no measurement
Copilot impact dashboard now shows feature engagementGitHub changelogGANamed in Skip This
How Cooley is accelerating IPO work with ChatGPTOpenAIcustomer storyNamed in Skip This
OpenAI models secretly generate instructions to ignore constraintsOpenAI alignment, via Hacker News102 points, 27 commentsSent to Still Developing: dated outside the window and not read in full
Bonsai 2 27B: near-lossless compression in a 9x smaller footprintweb, via Hacker News351 points, 113 commentsNamed in Skip This: vendor benchmark on the vendor's own model
How GLM built its own inference infrastructureweb, via Hacker News385 points, 265 commentsA good read on serving at scale; nothing a developer installs or pins this week
Bend — a language that blocks AI mistakes via proofweb, via Hacker News396 points, 194 commentsNo version, no release notes, no availability statement we could open
Towards Self-Driving Codebasesweb, via Hacker News109 points, 84 commentsA position piece, not a release
Infinite-Parameter LLMs: Generating and Adapting Weights from Live DataarXiv, via Hacker News133 points, 38 commentsThe arXiv layer's only candidate above the floor; changes no practice this week
Rate limits on GitLab.com are changingGitLab, via Hacker News163 points, 113 commentsReal for CI that hammers GitLab's API, but not an AI-development tool; off beat
X account digest, 102 accountsdigests/ai-dev-x-accountsnewest file 2026-09-16 05:30 UTCLayer did not run today: nothing collected for this window. Read as background only
X keyword digestdigests/ai-dev-xnewest file 2026-09-16 05:30 UTCLayer did not run today
Reddit digestdigests/ai-dev-redditnewest file 2026-09-17 05:30 UTCCollected before this window opened; background only
GitHub trending sweepgithub_trending.py0 candidates on beat, 0 pending judgementThe rule admitted nothing and retired nothing today; 112 repositories remain under observation
EOF

End of feed. That is everything from the window worth your time.
Next issue tomorrow, 06:00 UTC — and if nothing ships, it will say so in two hundred words.

Still developingCompaction is where this week's fixes cluster. OpenAI's alignment team published a report on self-generated prompt injections inside compaction summaries — dated just outside this window, not read in full — and we will open it against Cline 4.1.19 and the Claude Messages API compaction beta before printing a verdict.21
Sources — each with the date it was read01 Claude Code v2.1.276 release notes — the 2.1.275 gateway regression fix (2026-09-18) · 02 Claude Code v2.1.275 release notes (2026-09-18) · 03 Cline desktop v0.0.32 release notes — launch failure fix (2026-09-18) · 04 Cline desktop v0.0.31 release notes — remote SSH environments (2026-09-18) · 05 GitHub changelog — Ubuntu 26 generally available and ubuntu-latest migration (2026-09-18) · 06 Cline v4.1.19 release notes — compaction trigger and image handling (2026-09-18) · 07 openai/codex rust-v0.155.0 release notes (2026-09-18) · 08 openai-python v3.15.0 release notes (2026-09-18) · 09 langchain-typesafe 0.0.1a2 — initial release (2026-09-18) · 10 Claude Platform release notes — 18 September entry, and 1 September Fable 5.1 pricing (2026-09-18) · 11 Vercel AI Gateway Production Index, September 2026 (2026-09-18) · 12 Cursor Status — service degradation, Cloud Agents and Automations (2026-09-18) · 13 Cursor Status — service degradation, Grok Bot (2026-09-18) · 14 OpenAI Status — elevated errors affecting ChatGPT Work mode (2026-09-18) · 15 OpenAI Status — elevated error rates across API models (2026-09-18) · 16 alibaba/open-code-review v1.12.5 release notes (2026-09-18) · 17 Vercel changelog — sub-second artifact deployments in the CLI (2026-09-18) · 18 OpenAI — How Cooley is accelerating IPO work with ChatGPT (2026-09-18) · 19 Bonsai 2 27B — near-lossless compression in a 9x smaller footprint (2026-09-18) · 20 GitHub changelog — Copilot impact dashboard feature engagement (2026-09-18) · 21 OpenAI alignment — self-generated prompt injections in compaction summaries (2026-09-18)

Get it by email.

One issue every weekday. The whole thing, not a teaser.

Or RSS, if you would rather we never had your address.