Claude Code Projects hand your sessions to a coordinator, Theo calls Jev compaction terrible, Mistral's codebase goes up for sale, Pocock ships a /pr skill

A Thursday where Anthropic turned Claude Code sessions into a managed organization, the Jev hype cycle hit its first real technical pushback, and a European lab's source tree showed up on a forum.

Claude Code & Anthropic Updates

Projects in Claude Code. ClaudeDevs announced (8,829 likes, 1.36M views) and the main Claude account echoed (9,084 likes, 2.1M views) that Projects now run from one conversation, starting in Claude Code on desktop and web. You describe what needs doing, Claude splits it into threads, runs them as parallel cloud sessions, passes context between them, and keeps going after you leave. The blog post has the mechanics: each thread is a Claude Code cloud session on its own branch and copy of the repo, overlap between threads is resolved as an ordinary merge conflict, threads can further split into subagents, loops and workflows, and every thread reads from and writes to a shared project memory ("the release moved to Friday, why the export was dropped, who to check in with before touching billing"). You can set model and effort separately for the coordinator and the workers, and Anthropic warns projects hit usage limits faster since each thread is a full session. Threads are cloud-only today and cannot reach local files or your network; local support is "coming very soon." Beta is limited to Pro and Max users who already use cloud sessions and have no existing Projects; existing Projects users get migrated later. There is a waitlist.

How the team uses it. Boris Cherny says (2,390 likes, 372K views) Projects changed how he codes: "I stopped managing sessions. I just send thoughts as they come, Claude splits them into threads, and the project remembers how I work." In replies he confirms he runs Fable 5.1 at low effort as the coordinator, that TDD and loops happen without him asking, that CLI support is "soon" and mobile is already there, and that he uses a personal Max account for side projects when a reply accused him of unlimited usage. Thariq Shihipar frames it (1,061 likes) as bringing the Claude Tag architecture to Claude Code: one agent per project manages memory and spins off subagents, and you can ask it to be proactive or run on a schedule. Asked whether subagents get the memo when you change direction mid-task, he says the coordinator messages them. Ethan Mollick ran a stress test: he asked Fable in a Project to pick historical mysteries it could try to resolve, and it launched research agents, spun up eighteen threads, one per mystery, each of which launched more agents (simulating avalanches, breaking codes), then summarizers, then skeptical fact-checkers, over a day of work. Results are published with source notes; Mollick notes early access has no token limit so he cannot speak to cost. The best reply question: who checks the skeptical agents, since "every time I've set one model to check another one's work it mostly just agrees."

Cursor got there first. A reply to the announcement asks "did y'all get shipmogged?" and links Cursor's Projects post from September 10, which describes the same shape: a coordinator that never writes code, cloud by default with local agents when a test needs your machine, shared context files that sync across agents, and "subscriptions" where the coordinator watches a Slack channel or all your PRs and acts unprompted. Cursor claims users who primarily use Projects merge six times as many PRs. An open-source version, Orbital, lets any model coordinate and any CLI agent work.

Life Sciences Verification Program. Anthropic opened applications (1,987 likes, 331K views) for LSVP, which for the first time gives life science professionals access to Mythos. The announcement describes two grant types. Standard Use covers most biology R&D, applies to Mythos 5.1, Opus 5 and Sonnet 5 with more permissive science classifiers, and renews yearly. High-risk Use is a per-project add-on that "removes all safeguards that block life sciences requests," renews every six months, and is available for Opus 5 and Sonnet 5 today while Mythos high-risk grants stay limited to a small vetted set while Anthropic works with the US government. The interesting design choice is the shift from real-time blocking to offline monitoring against each organization's declared use cases, which requires retaining flagged data for review. Available in the console, Enterprise and Team plans; individual Pro and Max plans come later.

Preserved thinking is now documented. Theo's compaction thread below points at a support article worth reading if you build harnesses: for API accounts created after August 31, 2026, Fable 5.1 returns an error if the system prompt, tools or messages preceding a thinking block have been modified. You can opt into non-strict mode, in which case the affected thinking blocks are dropped and the response tells you which. The stated reason is industrial-scale distillation via edited conversations. The article calls out context compaction and injected system reminders as integrations that may need adjustment, and says it applies to all accounts for future models.

Agentic Coding & Agent Harnesses

Jev as compaction, and the takedown. Tamara Tran posted (6,988 likes, 1.39M views) "the perfect use case" for Jev: instant compaction by scoring every tool call and dropping what is irrelevant, with a repo. Diogo Almeida, Jev's creator, replied "YES! free coding agents from designing around the KV cache." Theo disagreed at length (1,264 likes, 236K views), six points: compaction is meant to refocus the agent, not filter noise; a 32K-context model scoring per tool call does not know what happened before or even the tool result, which dooms you to "stupid loops"; frontier labs ship encrypted reasoning that Jev cannot see and will drop, and Anthropic now requires the full history to keep any of it, so "using this in Claude Code guarantees the model will act way dumber"; models are trained on their own compaction flows (fun fact: if you switch models in Codex, compaction runs on the previous model); and deleting item 2 from "1,2,3,4,5,6" forces a cache rewrite of 3 through 6, and cache writes are already over 60% of his LLM spend. Tran answered that state is retained and tool calls are matched by id; Theo pointed at the preserved-thinking article above. Theo's follow-up: the interesting question is not better compaction but "what if harnesses never had to think about KV caching," which is what Diogo is hinting at. Armin Ronacher pushed back this morning: he would not dismiss Jev for compaction, since most harnesses need pruning on compaction for cost anyway, and modern compaction is already unintuitive (Codex has a mode where the model requests a blank context on its own and keeps notes). He notes pi extensions can hook compaction, but on new Anthropic accounts "you lose reasoning on retained messages." One tester says the 32K limit means it can only say drop, compress or keep, so you still need an LLM in the loop.

oh-my-pi as art piece. Theo called (599 likes) oh-my-pi "the most unserious dev tool I think I've ever seen," quoting Greg Horvay's description of snap compaction, which renders your history onto an image in a microfont and lets the model read the pixels. The authors' writeup is more interesting than the dunk: a 1568x1568 PNG holds about 40,000 characters in a 6x10 pixel font, roughly 10,000 text tokens billed as 3,279 image tokens, and they claim F1 parity with text at half the total bill on Fable 5 and Opus 4.8 after tuning the renderer, with 170K tokens of context carried as 46K. They also argue that eliding tool results confuses models about tool calling and that handoff summaries beg the next agent to declare the goal impossible and ship an MVP. Theo's replies: it froze ghostty twice during setup, and "there are 37 em dashes on this page." An OMP user called the take insufferable; Theo says he is tired of explaining why he does not like it.

Pocock's /pr skill. Matt Pocock proposed (2,210 likes, 99K views) a /pr skill that makes human review fast, borrows from Dex Horthy's /show-me, uses your domain language from /grill-with-docs, shows evidence the change works, and assesses merge risk as a one-way or two-way door, because "every model/harness I've seen creates garbage PR bodies." An hour later he shipped it as an in-progress, model-invoked reference for the shape of a PR body rather than a workflow skill. In replies he defends pseudocode summaries of the diff as the fastest way to understand the why, and says green tests are not evidence because "very easy to make a green test that actually tests nothing." Someone suggested CodeRabbit; Pocock: "You want to replace a single markdown file with an entire paid service? OK." The same day Dex Horthy open-sourced (1,081 likes) HumanLayer's bundled PR skill, /show-me plus steering to cut agent-PR slop, as npx skills add humanlayer/skills --skill visual-pr. One reply complains the skill text itself reads "Opusified."

Pocock on The Pragmatic Engineer. Gergely Orosz released a 95-minute episode with Pocock, noting his six-month-old skills repo has more stars than React. The show notes carry the substance: the grill-me skill came from Thariq's agent-interviews-the-user idea; "leading words" like tracer bullet, lifted from classic books, measurably steer agents; "memento-driven development" means optimizing a codebase for a colleague who wakes up with no memory every day; he is moving to cloud agents because they run with the laptop closed and can be multiplayer; and he has mixed feelings on TDD with agents, preferring to ask for proof the code works. Pocock also reported that after weeks trying to make AI better at course planning, notecards, pen, paper and scissors won.

roon says Astra's code is unmonitorable. roon wrote (3,965 likes, 805K views) that if you try to read the code Astra writes, "corrigibility has become a matter of faith," full of meta-programming and abstruse primitives, and that the only option is to have another Astra read Astra's output. Theo quoted it (1,337 likes) as "OpenAI employees making up new words to justify Astra's unreadable code" and told people to "just use fable if you want to merge the code," while Astra remains "incredible for moonshots and debugging." roon clarified he was complaining that the systems are "quickly becoming unmonitorable and we're just taking them at their word"; Theo: "the solution is still to make the model write better code." Halvar Flake: "hyperefficient is not the adjective that comes to my mind." Can Bölük argues the opposite, that models take the simplest path, never use generics, and clone and box everything in Rust. Thibault Sottiaux asked whether anyone has had Astra and Fable agree on a shared style guide and hosted it behind a web MCP.

Devin Code Scans. Cognition introduced Code Scans: give Devin a goal like "reduce compile time" or "improve SEO," it turns it into a scoped investigation via their Agentic MapReduce, batches it across parallel agents, ranks findings, and opens PRs. Their Dioxus example cut a clean debug build from 58.6s to 21.0s by making heavyweight CLI features opt-in. On HN with little traction.

Bend. Bend, "a language that blocks AI mistakes via proof," hit 409 points on HN: Python-like syntax, Lean-style proofs, C speed, runs on CPU and GPU, with an AGENTS.md snippet telling agents to keep rules in LAWS.bend and run the proof file before committing. The pitch is that humans stop reading code and need an ambiguity-free way to state intent. Skeptical thread.

GitHub rate limits. Theo would pay "a whole lot of money" to unlock his GitHub rate limits and expects to give up and move by October, alongside "crazy to see GitLab fumble this hard right now." Replies suggest Depot and Cursor Origin as workarounds.

Security & Breaches

Mistral's codebase for sale. Beni posted (7,395 likes, 882K views) that a Mistral hack rumor was real: an attacker on a criminal forum shared samples of unreleased tools and offered the "entire codebase" for sale. FrenchBreaches examined the listing from a user "mrwho" posted September 16: 339 archives with names like mistral-inference-private, mistral-finetune-internal, vllm-internal, sre-infra and model-delivery, plus a sample called webstral, a Chrome browser-agent prototype calling devstral-medium-2507. They note the claim is not independently confirmed and no live API keys were in the sample. Beni contacted the seller to ask about weights and post-training pipelines, then reported the account was wiped, which he reads as either a single exclusive buyer or an opsec panic. am.will amplified (1,223 likes, 401K views) with "likely had their model weights... exfiltrated," which Beni corrected in the replies: no reason to suspect weights, but training pipelines and data curation code potentially. Theo piled on (1,301 likes) with "first lab to get completely pwn'd with a ton of customer data leaking"; the customer-data claim is unsourced in the thread. No Mistral statement as of this morning.

Hacking OpenAI, with Opus. Hacktron's writeup (247 points on HN) of a July 25 chain: Discourse routed HEIC uploads to ImageMagick and an unpatched Debian libheif, Opus 4.8 found the missing security backport, Opus 5 on release night produced a working ARM64 exploit in three hours and was then run in an autonomous /goal loop against their own Discourse Cloud instance, proxied to look like a CTF because Opus refused to write exploits for remote targets. RCE on community.openai.com plus an SSO flaw gave them ChatGPT and Codex account takeover for any forum user, including employees; they proved it by having an employee's Codex open a PR in OpenAI's internal monorepo. Bounty: $6,500. Their broader HEIF Heist work says Slack, Meta, GitHub Enterprise, Rails and Next.js all depend on the same library.

Targeted attacks on Rustaceans. The Rust security team warns of an ongoing campaign against rust-lang members and popular crate owners: a plausible job or contract video call, then a request to install a "missing audio codec" or run a clipboard command. The June attack on Rust developers and the August arrayref compromise may be the same campaign; the pattern matches DPRK's Contagious Interview. Via Simon Willison.

Models, Speed & Compute

Astra for Law. OpenAI launched (440 points, 465 comments on HN) GPT-6 Astra with a legal search index of 230M URLs including Free Law Project's CourtListener corpus, custom instructions for legal analysis, 26 partner plugins (Relativity, Clio, iManage, Thomson Reuters), and ZDR plus exclusion from human review for verified firms. On Vals AI's Legal Research Bench it passes 54.0% versus 38.7% for Astra with web search. The post takes a direct shot at Anthropic: on their sample prompt, "Claude Fable 5.1 returned a holding that had been reversed on appeal." Sullivan & Cromwell, Ropes & Gray and Cooley built custom tools with OpenAI's forward-deployed engineers. HN's most-replied comment, from a non-practicing lawyer, argues nobody is specifying which legal economic model they think this affects.

Speed over intelligence, and Luna on Cerebras. Max Schoening of Notion asked Sam Altman for Luna on Cerebras: "removing latency would do more for interesting UX experimentation than smarter models." Cerebras CEO Andrew Feldman agreed that in some domains speed trumps intelligence once intelligence is good enough, asking "is 500 tokens per second per user the new slow?" am.will wants it too: with 5.3-Spark being deprecated, Astra could delegate to 2,000 tok/s Luna agents that cost less usage and compute. One reply runs automations over 6K company domains and says Luna at xhigh is the bottleneck at three days per batch.

Bonsai 2 27B. Prism ML released Ternary Bonsai 2 27B (372 points on HN): Qwen3.8 27B compressed to 1.76 bits per weight and 5.9GB, retaining 98.2% of aggregate benchmark performance across reasoning, coding, vision and tool use, 262K context, Apache 2.0, 143 tok/s on an RTX 5090 and 46.8 on M5 Max. Coding-agent demo with Cline. Related: Mike Bradley posted early M5 Ultra Studio numbers, Qwen3.8-27B Q4 at 1,800 tok/s prefill and 51 decode.

Qwen3.8-Omni-Flash. Alibaba launched a native omnimodal model with 1M-token context, text, image, audio and video input, agentic long-video understanding that decides what to watch, plus an open-source Qwen-Live Harness for real-time interaction. Claims a 98% cut in audio-input price per hour versus their previous Omni.

Jev, day three. Jerry Liu places Jev alongside routing and post-training as a third way to optimize abundant intelligence for a task, this time at the architecture level, for anything that is generalized classification. Armin is getting "Clanker psychosis trying to figure out how to create a sensible API around Jev like models"; replies point at pithings/advocaat and an openjev clone, and one person already added Jev as a DataFusion query primitive. am.will admitted he did not know what a Jev was; the best answer is "a panel of judges holding up scorecards." Someone made Jev talk by giving it a few hundred English words to choose from.

Other Interesting Stuff

  • Theo video. Please stop using stupid models: the best models are not better because they are smarter but because they do the dumb things right more often. Sources include David Khourshid, David Cramer and Dax.
  • How to write with an LLM. Thomas Ptacek's essay (110 points on HN, via Simon Willison): use models as copyeditors, never ghostwriters. Rule one, you may not use a single word an LLM suggests. Rule two, forbid encouragement. Then have it flag passive voice, filler words and paragraphs to move, and judge your rewrites with a model that has no context on the edit.
  • Theo on the engineer who avoided AI. Theo quoted (770 likes) a story about a laid-off engineer whose interview ended when he said he had not used AI: "Unemployment sucks, but at some point you have to realize it's your own fault." Several replies call the original story fake.
  • Simon Willison on the OpenAI compaction report. His note on yesterday's self-generated prompt injection report calls it his favorite of the six.
  • Codex for Open Source renewed. Charlie Marsh posted that the program now offers $100 Pro plans and doubles grants from 5,000 to 10,000 maintainers.
  • Hister. A private search engine for pages you visit and files you keep, 561 points on HN.
  • Farr summit. Peter Steinberger is speaking at FARR San Francisco on October 5 and on a Cloudflare Connect panel.

Footnote: @potetotes still returns 404. Karpathy, swyx (own posts), leerob, simonw (own posts) and steipete (own posts) had nothing new in the window. nitter.jaydenha.uk returned intermittent 502s all morning; every other listed Nitter instance is behind Anubis or dead.