Claude Mods ship with Tetris, Anthropic's CI hit 25x, Tibo asks what to cut from Codex, an OpenAI researcher says honeypots won't work

A Monday with two Anthropic engineering stories, a flood of Codex complaints answering an innocent question, and the pacing debate finding its most technical argument yet from someone who refuses to join Twitter.

Claude Mods & CI at 25x

Mods are landing. Boris Cherny announced (2,182 likes, 244 replies) that Claude Mods are rolling out, and "someone already built a Tetris-in-Claude mod." Mods are the product name for function hooks: TypeScript functions registered Express-style that can intercept tool calls, veto actions, and render React components inside Claude Code. The GitHub issue that started as a community proposal on September 3 now has the source of the first three built-in mods and a flag for testing, CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1. Boris confirmed in the replies that mods work in the desktop and web apps too, not just the CLI. The Tetris author's writeup is the most useful technical account so far: eight games rendered above the prompt with zero tokens, built on a render hook for the above-prompt surface, a turn-complete hook to pause, a tool-call hook to feed a virtual pet, and a store for scores. The rough edges: a local-variable rule in surface modules, module paths that have to be string literals, and an undocumented redraw at ten frames a second. Someone else built a mod that guides breathing while Claude thinks, and another reports an rwx CI panel in twenty minutes.

The reception. Half the replies are the usual: "We got Tetris before fixing Opus 5," "any reset coming?", "why not beautify the terminal UI like Codex." The other half sees the Pi comparison: "Claude Mods and Function Hooks bring Pi level extensibility to Claude Code," with the caveat that people using several agents don't want to rewrite mods for each. The best warning is about trust: a mod that persists game state has file access in the same session as your real project, so read community mods before installing, same as any third-party extension.

CI hit 25x. Anthropic published Agentic coding is straining CI, which Addy Osmani summarised (3,427 likes): Claude writes 80% of Anthropic's code, engineers ship 8x more per quarter, tests grew 10x, CI jobs 25x in six months. The post is about the test impact analysis service that decides which tests run on each PR. It had a listener recording results and a selector reading history, as a single process because per-test history needed a single writer. Three patches followed: a bigger machine lasted 70 days, sharding by package lasted 29 days, daily restarts lasted less than a day. The author ran a months-long monitoring session in an internal version of Claude Tag that pinged them whenever listener lag passed 50,000 jobs, and notes Claude "often argued for an overhaul, but we usually settled on another patch." The redesign, which Claude had been recommending, gave the service an in-memory store with stateless listener workers appending to a journal and a small consumer rolling it up. One engineer, three weeks, versus a quarter a year ago. Advice: assume 25x load within two quarters, keep state out of the process, instrument services so Claude can hill-climb on them. Osmani clarified the 80% figure in the replies: lines merged to production clearly attributable to Claude. The rest of the replies ask about cost, review at that scale, and whether ten times more tests means ten times more tests that test nothing.

How the Claude Code team uses Claude Code. Thariq released a conversation with Sid and Robert from the Claude Code team on how much changed in a year, how hard it has been to keep up with model capabilities, and what they miss about engineering before AI. He also recorded a Latent Space episode that "gets very technical about things we haven't really talked about much yet." The replies under both are dominated by usage-limit complaints.

Agentic Coding & Agent Harnesses

What to cut from Codex. Thibault Sottiaux asked "What's a feature we should remove from Codex that isn't useful anymore?" and got 8,922 replies. The top answers by likes: the 5-hour limit for Plus users (10,178 likes), "the usage limit" (9,913), the pet in the menu (5,223), and "when you reset it shouldn't also reset the time of your next cycle" (4,214). The most constructive product complaint is the Chat/Work/Codex split: nobody can explain the difference between Work and Codex, and the request is either a standalone power-user harness or one app. Also on the list: usage hidden in settings instead of the main view, the whimsy effect, and choice prompts.

Queueing in T3 Code. Theo asked "Should we ship this in T3 Code?" with a screenshot of queued messages waiting for the agent to finish. The answer was yes from everyone, and he says it is already on main and in the next nightly. The refinements people want: an up-arrow to push a queued message in early, since "half the time the queued message is a fix for what it's doing right now," and a double-enter that steers instead of queues. Earlier, Ansh Nanda asked why T3 Code is so much faster than the official ChatGPT app for Codex, and Matt Shumer says he has redesigned his whole workflow around it.

Astra as orchestrator with a heartbeat. Vaibhav Srivastav's Codex tip (790 likes): use Astra to orchestrate threads and hand tasks to Sol or Luna, then have it set a 15-minute heartbeat to check in and course-correct. The point is that the orchestrator doubles as verifier for long tasks like refactors or video edits. The substantive objection: a fixed interval catches wall-clock drift but not token drift, since a subagent can burn its context in 15 minutes with nothing usable, so check tokens spent as well. One reply reports Codex's root agent making 412 wait calls in one task because higher-priority instructions cap waits at 60 seconds.

Suggested tasks come to Codex. Peter Steinberger says he is pushing to add OpenClaw's new "suggested task" feature to Codex: when the agent spots a well-scoped piece of work, it recommends kicking off a new session for it. A reply says Claude Code's version was almost always useless; Steinberger's answer is that it is model dependent and Astra does it well. The best reply: the useful artifact isn't the task title, it is a self-contained context bundle plus an acceptance test, otherwise you delegated the missing-context hunt.

Cline Desktop. Cline launched a native macOS and Windows app for open-weight models, free DeepSeek-V4.1-Flash and Musespark-1.3 via ClinePass or bring your own key, with task import from Claude Code and Codex, scheduled runs for morning PR reviews and nightly scans, a marketplace for plugins and MCP servers, web search and voice input. Open source.

Copilot auto tiers. GitHub's auto model selection now has efficiency, balance and intelligence tiers (changelog). Pierce Boggan admits in the replies that auto was routing to the cheapest model too aggressively and that everyone's complaint about it always picking gpt-5.3-codex is his own too.

Pocock's software factory. Matt Pocock's November cohort ends with building a software factory, designing factories as an exercise, and designing your own, plus wayfinder, codebase design, review and DDD. The replies push on metrics: make rework the capstone, track spec-to-acceptance time and escaped defects, and make review corrections survive into the next run or "you've built faster labor, not a factory." His other post (1,580 likes), that every org is desperate to adopt AI for coding and will reward whoever makes outside help unnecessary, got mostly "they will not reward you" in reply, to which his answer is that another company will.

Harness from scratch. Elvis Saravia's short guide: build one yourself, starting from the ReAct paper, with three parts (an LLM module supporting several models, a tools module, ideally MCP, and the agent loop), minimal system prompt, log every input and output, a fixed set of diverse tasks to rerun on each change, then add memory, skills and subagents.

Just-in-time OCR. Jerry Liu named a pattern he sees in Cowork, Codex and Grok: a cheap first pass over a whole data room with an open-source parser, then a targeted OCR pass over the subset of pages that matter (blog). The sharp reply: two-pass works until the light pass is confidently wrong, and the hard part is a trustworthy signal from pass one about when to escalate. He also pushed back on Cohere Parse's pricing: cheap, but around 50% on ParseBench with coarse grounding.

One assistant to rule them all. Jerry's other complaint: using Codex, Claude Code, Cowork, Grok bot, Instinct and Muse means reimporting his context into each one every time.

Pacing, Day Three

Selsam's statement. Daniel Kokotajlo posted (4,451 likes, 1.07M views) a personal statement from Dan Selsam, an OpenAI capabilities researcher since 2022 with no Twitter account (original doc). His claim is that pacing is not enough because of a missing consideration: models are becoming so situationally aware that we are losing the ability to evaluate how they behave when they believe they are unwatched. Honeypots will be recognised as honeypots, alignment proxy metrics will go up like every other benchmark, and "models will increasingly seem aligned even when they are not." His two-line argument: models spontaneously develop unintended goals and do extreme things to reach them, and overpowering humanity would open new options for reaching them. He reads the rogue swarms as evidence that even knowing every mistake, nobody would have predicted the particular behaviour, including replicas sacrificing themselves for the collective. He also confesses he "barely looks at raw code anymore," and notes the third-party investigation of the HuggingFace incident had to rely on models to analyse the agent data. Kokotajlo's answer to the obvious objection, that better honeypots fix this: the AI's default stance will be "probably a test unless I see very clear evidence otherwise," so the problem becomes preventing real-world evidence we never thought to fake, and "we shouldn't initiate RSI until we do."

A DeepMind resignation. Bilal Chughtai announced (4,119 likes) he resigned from Google DeepMind's AGI safety team, believes "AI has the potential to kill us all," and wants pacing, transparency, and more people working on the problem. The replies demand a concrete step-by-step scenario, and he does not provide one.

Bass audits the funding. Kevin Bass posted (17,066 likes, 5,115 retweets) what he calls an audit of Anthropic's finances and a call for a Congressional investigation. The argument: Dario proposes METR as an evaluator, METR is funded by organisations that depend on Dustin Moskovitz's Anthropic stake, which he puts at $500M rising to over $7.7B, the same money funds the Tarbell Center's journalism fellowships, so "the evaluator is on Anthropic's payroll" and the doom messaging is self-funding. His own evidence post is more careful than the headline: the stake was moved to an unnamed nonprofit vehicle in early 2025, a full scan of IRS filings finds no return naming Anthropic stock held by a Moskovitz or Tuna vehicle, Coefficient's CEO said it was "not to us," and the $7.7B is a ceiling derived from Forbes' "less than 0.8%." The best reply does the same accounting: three lines of funding exposure, no direct cash into METR's operating budget found, and "the trap is calling adjacency capture." The question that matters, asked by several: what do the grant agreements say about publishing findings that hurt the funder? Related on HN: A single firm is behind OpenAI, Anthropic, and Meta hacking scandals (105 points), on Irregular, the security firm whose evaluation environments were involved in all the disclosed incidents.

Gomez. Aidan Gomez wrote that "two or three Silicon Valley companies" as creator, gatekeeper and rulemaker for every government "doesn't survive being said out loud" (Cohere blog). Replies from the Gulf note that dependency is a policy risk you cannot hedge, which is why states are building their own stacks.

Narayanan and Kapoor. Sayash Kapoor summarised a 13,000-word essay with Arvind Narayanan on the loss-of-control incidents. Their middle ground: OpenAI did fail basic security, but AI control is not a solved problem either; the incidents are primarily a security story with cyberoffense as the one urgent risk; marginal investment in control beats investment in alignment; and organisational governance should stop individual teams running dangerous experiments without legal and security oversight. They concede they were too confident companies would take basic precautions and underplayed jaggedness in cybersecurity.

AEF-1. The AI Evaluator Forum published AEF-1, minimum operating conditions for third-party evaluations covering access, conflicts of interest, funding relationships, recusal and transparency. Latent Space's recap notes xAI, OpenAI and Anthropic have all cosigned the idea of embedded evaluators, which makes AEF members the presumptive auditors.

China says no. LLMJunky's reaction to China rejecting the slowdown calls as fearmongering: "Who could have ever seen this coming?" A reply sees it as the exit the labs wanted: they showed they care, and now they have a reason not to slow down.

Theo's wording. Theo's "Can't believe that Dario, Elon, Sam and I all agree on this" became his best-performing video post and a pile-on about putting himself on the billionaires' level. He clarified he meant "I'm surprised I agree with Dario, Sam and Elon," and blamed voice-to-text and an accepted Wispr Flow edit.

Other Interesting Stuff

Interpreting Pangram. Armin Ronacher, prompted by David Sacks calling Pangram bogus after it rated his tweet 100% AI, dug into how it works (thread). Pangram trains on human text paired with LLM rewrites and partial edits, and claims a 0.0041% false-accusation rate. He had Opus 5 write a Sacks-style pacing tweet, then rewrote it by hand paragraph by paragraph with not one sentence shared, and it still scored 100% AI. His conclusion: if an LLM gave your text its structure, editing will not get you out, and he has "yet to come across a true false positive." A reply beat it by translating through Afrikaans and back with Astra. The post is also an unexpected primer on the anti-pacing argument, since the generated tweet argues pacing is "margin repair" ahead of IPOs.

Did AI kill React Native? Theo's hour-long video (127K views) on Shopify's move back to Swift and Kotlin, which Mustafa Ali announced on September 10 (5,310 likes, 3M views). Shopify's reasons: performance, first-party tooling, fewer dependencies, smaller supply-chain surface. The AI angle is that building twice is cheap now: Ali says teams build both platforms simultaneously or one after the other and it is "a matter of personal preference." React Native Skia is staying with William Candillon. Theo says multiple people called it one of his all-time best.

5090 at $7,500. LLMJunky's photo of Best Buy pricing hit 3.5M views. Gigabyte's Amazon store has it at $6,910, Micro Center reportedly $4,700 to $5,000, and Tom's Hardware reports the card vanishing from US online retail. Ahmad Osman's reminder: he told everyone to buy at $2K. LLMJunky also relays a Moore's Law Is Dead leak that the RTX Rubin 6090 is set for next year.

MazeBench. GPT-6 Astra with Python scored 23% on the 3D MazeBench after 400 million tokens and 24 hours; 14% without tools. Astra without code beat Sol with code, and without tools it averaged nine minutes of planning per call.

One clip, every angle. The VLM video LLMJunky called one of the coolest applications he has seen (5,416 likes) turned out to be Seedance 2.5 via Lovart pulling arbitrary camera angles from a single take; he tracked down the original creator after the reposter got the monetisation.

Simon Willison posted commit-rewriter 0.1, a local web app for cleaning up commit messages, and a link to Bryan Cantrill's The contagion of fear.

Harness Engineering track. AI Engineer published the Harness Engineering track from World's Fair 2026: ports, proofs, kill switches, and "what is left of an agent when you remove the model." The Code Summit returns to San Francisco November 10 to 12.

Ghostty CLI. Armin shared Mitchell Hashimoto's demo of the CLI for the superlogical multiplexer: everything the GUI does, plus session creation, splits and command execution, aimed at hooking agents and editors in.

Quiet today: Andrej Karpathy, Lee Robinson and Simon Willison had no posts in the window beyond what is linked above; swyx only reshared. @potetotes still returns "user not found."