The one thing
01Cline's hub loads agent plugins from your home directory and deliberately refuses to load them from the repository you just opened.
Cline SDK v0.0.83 published at 05:53 UTC. Packages under ~/.agents/plugins/* are discovered and validated from their root plugin.json; valid skills are exposed through the skills tool as plugin-name:skill-name, and stdio, Streamable HTTP and legacy SSE servers declared in mcp.json start without anything being written to cline_mcp_settings.json. Workspace .agents/plugins directories are not scanned, so opening a repository cannot start MCP servers that the repository controls; extra roots require an explicit agentPluginPaths.1 The CLI carrying the same change, v3.0.62, published at 06:04 UTC — four minutes after this window closed, so we are reporting it from the SDK release.2
Three fixes in the same release are about time rather than trust. A model turn that dies mid-stream on a transient provider error is now retried up to three times with exponential backoff; a single forwarded 429 previously ended the run. A turn that has already streamed text, reasoning, media or a tool call is never retried, so nothing duplicates. Streaming had been throttled because the hub proxied every runtime event to client onEvent hooks as a round trip carrying a full session snapshot, with the agent loop awaiting it. And checkpoints re-hashed every untracked file before each model call, which on multi-gigabyte workspaces blocked messages for seconds to minutes.1
Shipped
02Claude Code v2.1.271 · per-command allowed_domains, plugin commands pinned by hash
Published 22:12 UTC 14 Sep; v2.1.272 followed at 00:42 UTC with bug fixes only. Bash, PowerShell and Monitor accept per-command allowed_domains in auto mode with sandboxing: the hosts a command needs are reviewed alongside that command and opened for it alone, and other hosts are refused. claude plugin install and claude plugin update accept --accept-command <sha256>, which approves exactly the command a previous --json run displayed rather than whatever runs now.34
llama.cpp v0.4.1 · a signature changes, and GDN normalisation is fixed
Published 18:27 UTC 14 Sep. llama_sampler_chain_n() returns int32_t rather than int. GDN normalisation moves from max to rsqrt for affected Qwen, Kimi and GLM models, and MTP context KV cache allocation is fixed for DeepSeek2 and GLM-MoE — bad output from those served locally may have been this.5
Promised, not shipped
Copilot seat charging upfront and model cutover — 1–2 Oct 2026 · Cursor OpenAI bundled end — proposed 12 Nov 2026
The conversation
01A 35kB preprompt that worked against a hosted model does not fit in a self-hosted one, and the thread argues about whose fault that is.
- patrickmccanna.net, on Hacker News — 129 points, 69 commentsReader-posted · first-hand · unverified against any vendor page
Moving 35kB preprompts off Opus to self-hosted models, the reported failure is context: prompts sized for a hosted million-token window fail against a local 65K one.7
- Top-level replies in the same threadReader-posted · contested · no method given
The most-upvoted reply argues a 35kB prompt is unfocused on any model and puts the useful attention ceiling near 250K, well below advertised limits. Another reports 5–10 minute prefills on a 128GB local machine. Neither figure arrives with a method.7
- X and RedditLayers did not run for this window
The newest file in both digest folders is 14 Sep 05:30 UTC, which sits outside this window. A live Reddit pulse across six practitioner subreddits returned zero posts in-window at any comment threshold.89
We are printing the claim and refusing the numbers. A context ceiling is measurable and nobody in that thread measured one.
Releases: 28 repos, sweep complete, authenticated, 12 non-prerelease in-window. Vendor feeds: 12 sources, sweep complete, 7 items. Hacker News swept above 40 points, 53 stories in-window. X digests stop at 14 Sep 05:30, so the X layer did not run today. Reddit: no digest, live pulse zero posts. arXiv not swept.
Skip this
04Vercel AI SDK harness layer authenticates through the harness's own subscription Published 21:28 UTC 14 Sep and real: `direct` and the default `auto` fall back to a native subscription on the host, across Claude Code, Cline, Codex, Cursor, Copilot, OpenCode and others.[^6] Held out of Shipped because Vercel publishes no cost comparison against gateway billing, so the saving is unmeasured.
GitHub Copilot auto model selection gains efficiency, balance and intelligence tiers Rolling out now, but no number moved: you are charged for whichever model auto picks regardless of tier, and paid subscribers keep the same 10% discount.
GPT-5.6 Luna vs GPT-6 Astra for code review — 137 points on HN A code-review vendor benchmarking models at code review. The harness is not published, so the comparison cannot be checked.
GitHub trending: 55 on-beat repositories, 5 admitted Nine cleared the objective gates and got a beat verdict today; five joined the list and four were judged off-beat, among them a voice-cloning app and a prose rewriter. Three more cut no release in ninety days, which a release sweep cannot see.
Everything we saw
7272 candidates scanned · 3 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| cline/cline SDK v0.0.83 | github releases | 05:53 UTC 15 Sep, hub-managed agent plugins | led the issue |
| cline/cline CLI v3.0.62 | github releases | 06:04 UTC 15 Sep, four minutes past the window | carried the same change |
| anthropics/claude-code v2.1.271 | github releases | 22:12 UTC 14 Sep | shipped |
| anthropics/claude-code v2.1.272 | github releases | 00:42 UTC 15 Sep, bug fixes only | shipped |
| ggml-org/llama.cpp v0.4.1 | github releases | 18:27 UTC 14 Sep, one API signature change | shipped, breaking |
| Vercel — AI SDK harness layer native subscription authentication | vercel changelog | 21:28 UTC 14 Sep | shipped, wait |
| BerriAI/litellm v1.101.0 | github releases | 03:41 UTC 15 Sep, timing headers on /v1/messages and /v1/responses | watching, no decision this week |
| openai/openai-python v3.14.0 | github releases | 23:28 UTC 14 Sep, stream errors normalised | watching |
| vercel/ai ai@7.0.101 | github releases | 04:19 UTC 15 Sep, patch: serialise tool output JSON | patch, nothing to decide |
| block/goose v1.50.1 | github releases | 21:00 UTC 14 Sep, reverts model updates and MCP default version selection | a rollback, not a shipment |
| sst/opencode v1.18.31 | github releases | 17:47 UTC 14 Sep, ACP session bugfixes | bugfix release |
| GitHub — Configure cost and quality in Copilot auto model selection | github changelog | 16:05 UTC 14 Sep, rolling out | skipped, no number moved |
| OpenAI — Elevated errors affecting Work Mode in ChatGPT | openai status | to 15:58 UTC 14 Sep, 11.2h, minor | in production |
| OpenAI — Degraded Performance affecting Agents API | openai status | 23:20 UTC 14 Sep, 0.3h, impact none | in production |
| GitHub — Actions larger runner jobs slow to start | github status | 18:40 UTC 14 Sep, 0.9h, minor | in production |
| Cursor — Investigating Grok service degradation | cursor status | 08:14 UTC 14 Sep, 0.9h, minor, Grok 4.6 | short and resolved |
| OpenAI — How Fyxer built an AI executive assistant people trust | openai news | 12:00 UTC 14 Sep | vendor customer story |
| Notes on gotchas migrating 35kB preprompts to self-hosted models | hacker news | 129 points, 69 comments, contested | the conversation |
| GPT-5.6 Luna vs GPT-6 Astra: is a $1.20 model good enough for code review? | hacker news | 137 points, 124 comments | skipped, vendor benchmark |
| Pion, an agent designed to run any company autonomously | hacker news | 357 points, 402 comments | skipped, no availability |
| Apple's Siri AI can be swapped out for Claude, ChatGPT, code shows | hacker news | 221 points, 155 comments | skipped, unannounced |
| OpenAI bots knew about the RubyGems caching vulnerability | hacker news | 422 points, 341 comments | watching, no developer decision yet |
| X accounts digest, newest file | digests | collected 05:30 UTC 14 Sep, no 15 Sep file | layer did not run |
| Reddit live pulse, six practitioner subreddits | scrapecreators | 0 posts in-window | layer empty |
| GitHub trending sweep | discovery | 55 on-beat, 41 pending, 5 admitted | watchlist 28 to 33 |
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.