The one thing
01Claude Code skips AGENTS.md whenever it cannot fetch a remote flag, and nothing on Anthropic's pages says that is fixed.
Claude Code 2.1.277 added AGENTS.md as a fallback when a project has no CLAUDE.md. The loader is a built-in plugin whose availability is decided by a remote feature flag, with false as the fallback. Set DISABLE_TELEMETRY or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC — to any value, including 0 — and the file is never read, with no warning.1
The reporter's canary test also found that clearing the variables in a project's .claude/settings.json has no effect. A second reporter on the issue says the file is skipped in the first session after install even with telemetry on, because the flag is not cached yet — which is every fresh CI runner.2 Bedrock, Vertex and gateway users hit the same path, since the flag cannot resolve there.1
The v2.1.281 release notes, published at 19:19 UTC, list no AGENTS.md change, and issue #95690 was still open when we read it at 06:00.23
@AGENTS.md next to it; @ imports do not depend on the flag.1The gate, and the one override that works
var W = !1; // isOnByDefault: false
var B = () => Oa("tengu_agents_md_mod", W); // isAvailable: remote flag, fallback falseExcerpt from the 2.1.280 bundle as printed by the reporter.1 A session-level claude --settings '{"env":{"DISABLE_TELEMETRY":"","CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC":""}}' works from the second session on, which is no help on an ephemeral runner.
One commenter on the issue notes that DO_NOT_TRACK is the variable that reads as a boolean; the two above treat any non-empty value as on. We have not tested that ourselves.
Shipped
02Node 20 is gone from GitHub Actions runners, and so is the opt-out
Runners now execute JavaScript actions on Node 24, and ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION no longer works.4 If you maintain a JavaScript action, set runs.using: node24 and cut a release; if you consume one, move to a version that supports Node 24. Self-hosted runners on macOS 13.4 or earlier, or on ARM32, are no longer supported.
The GitHub Copilot app can sandbox local sessions per project
Per-project policy covers extra read/write, read-only and denied folders, outbound and local network, and whether Git and GitHub CLI credentials are passed in.5 If the OS cannot enforce the policy, the shell fails instead of running unsandboxed. It is off by default, public preview, and configured separately from Copilot CLI; /sandbox on applies it to a running session.
The conversation
01The Hacker News title says the AGENTS.md gate is fixed. The primary pages do not.
- mpoteat, Hacker Newsreader-posted, unverified; identity not confirmed, and not reflected in the release notes
Wrote that the gate was a rollout artifact kept switchable by feature flag, and that it is "already fixed as part of v2.1.281 releasing today".
- pszypowicz, author of the post and commenter on issue #95690first-hand, reproduced on 2.1.280 with a published method
Canary test in an empty directory, two sessions per setup: with either variable set, AGENTS.md never loaded; a one-line
@AGENTS.mdCLAUDE.md loaded it with telemetry still off.
The fix may well be real and behind the same flag it is meant to remove. Until a release note or a closed issue says so, test it: the canary check takes two commands and one minute on your own runner.
Hacker News was swept in full: 40 stories seen in the window, 15 above 80 points, with comments. The X layer did not run on today's window: the newest ai-dev-x-accounts file (22 September, a roster of about a hundred accounts) holds a placeholder instead of posts, and the newest ai-dev-x keyword digest was collected 23 September 05:30 UTC, about 24.5 hours old, 40 posts kept of 420 seen — read as background only, and it covers the launch day we reported in No. 37. The Reddit digest for 23 September holds one post, and a live pulse across six subreddits returned nothing.
Skip this
05Claude discovers a novel enzyme system 587 points and the busiest thread of the day. A biology result, not something a developer installs or pins.
GPT-6 Astra drives a car (drivingbench.com) 285 points for a benchmark site; the top comment from an openpilot contributor questions whether a cloud-delivered model can meet the latency. Nothing to evaluate for software work.
OpenAI customer stories: Harvey, invideo, Ringg Vendor case studies with vendor figures — "3x" colour grading, "up to 65%" of calls. Not measurements we can check.
Jev in 25 Lines of Python 644 points, and a commenter notes the post ends by calling itself parody, with no error-rate or latency comparison.
Tokens too cheap to meter An essay extrapolating price curves. A forecast, not a release.
Everything we saw
6262 candidates scanned · 3 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| Claude Code reads AGENTS.md only when telemetry is on | Hacker News / blog.szypowi.cz | 463 points, 262 comments | Led the issue; fix claim checked against release notes and issue |
| Node 20 is no longer available in GitHub Actions | GitHub Changelog | Final notice, opt-out removed | Printed as Breaking |
| Local sandboxing in the GitHub Copilot app | GitHub Changelog | Public preview, off by default | Printed as Wait |
| anthropics/claude-code v2.1.281 | GitHub Releases | Two prompt-cache loss fixes; no AGENTS.md entry | Cited in the lede and Still Developing |
| More ways to request and configure Copilot code reviews | GitHub Changelog | Personal settings on every plan, enterprise default effort | Settings surface; changes no decision |
| GitHub: incident across several services | githubstatus.com | Degraded API requests, 18.7h, impact minor, resolved | Long but minor; logged |
| Cursor: Cloud Agent Dockerfile builds | status.cursor.com | New builds blocked, 1.5h, impact minor, resolved | Weather |
| Cursor: Grok 4.7 degradation; OpenAI: two ChatGPT incidents | status pages | 0.1h, 0.4h and 0.7h, all minor, resolved | Weather |
| ggml-org/llama.cpp v0.5.0 | GitHub Releases | Server binds to several addresses; additive API changes | Nothing removed; no decision changes |
| google-gemini/gemini-cli v0.61.0 | GitHub Releases | 1.3k-character notes, no breaking section | Dropped |
| OpenHands v1.23.0, goose v1.52.0, pydantic-ai v2.49.0 | GitHub Releases | Minor releases, no breaking or removal sections | Dropped |
| MCP TypeScript SDK server 2.1.0 / node 2.1.0 / 1.30.1 | GitHub Releases | Minor bumps on both lines | Not read in full; watching |
| cline CLI v3.0.65, SDK v0.0.86; vercel/ai 7.0.113, 6.0.290, 6.0.289; openai-python 3.19.1–3.19.2; litellm v1.99.3, v1.100.2, v1.101.1; langgraph-cli 0.4.32; langchain-anthropic 1.7.4, langchain-openai 1.6.5; agent-native, orca, BrowserSkill builds | GitHub Releases | Patch cadence | No decision changes |
| Vercel: Blob stores unlimited on every plan; Connect supports TanStack AI | Vercel Changelog | Plan limit lifted | Off the AI-development beat |
| Australia: OpenAI agent accessed a government portal; Medicare breach report | Hacker News | 95 and 159 points | Not opened; no primary source in the window; watching |
| Once Claude can measure something, it can make it faster | Hacker News / claude.dev | 190 points | Vendor engineering post; nothing to install |
| Measuring the Serving Stack Instead of the Model; CliffCompaction; A2M | arXiv | Carried in No. 37 | No new paper in the 48h sweep changes practice |
| ai-dev-x-accounts digest, 2026-09-22 | X digest | File present, contents a placeholder | Source unreachable |
| Reddit live pulse across six subreddits | 0 posts for the window | Fallback ran and came back empty |
End of feed. That is everything from the window worth your time.
Explore more news, analysis and practical guides on NO FEED.