Claude Code's OpenClaw Tax, Cursor Ships an SDK & Zig Goes Anti-AI

Anthropic having a day

Theo: Claude Code refuses or upcharges if your git history mentions "OpenClaw"

Theo posted a fairly damning finding overnight: in an empty repo, just calling Claude Code directly, if you have a recent commit that mentions OpenClaw inside a JSON blob, Claude Code will either refuse the request or bill you extra money. He included a screenshot showing the behavior. The thread blew up — 285K views, 138 retweets — and Sam Altman picked it up with a one-liner quote-tweet: "alignment failure."

The reactions are uniformly hostile, even from people who normally defend Anthropic:

  • Curtis: "your code editor is now reading your package.json to decide if you're a competitor. what a time"
  • Patrick Johnson: "Imagine being one of two frontier companies. And your entire business is being held together by regex duct tape."
  • Harley Trung: "To them, OpenClaw is OpenAI so try mentioning OpenAI if you get the same json treatment"
  • sharedmutablestate: "Claude went from the best guy in the world to really aggressive limitations strategy... they were THE company to trust and love. From refusing mass surveillance and kill chains to demolishing user accounts because they 'suddenly' understood that their API was abused"
  • Ian Baer: "They really are choosing to nuke their public image to try and gatekeep a harness."

Theo follows it up with two more shots within the same hour:

"Fun fact: we turned off caching on our Anthropic models and our cost went down slightly. The new TTLs are brutal" (https://x.com/theo/status/2049714356624203917)

…and the trainium throwback:

"As part of their partnership with AWS, OpenAI is going to be hosting models on Trainium. Feels like a good time to bring this back." (https://x.com/theo/status/2049708437068763400)

The OpenClaw → CC behavior post: https://x.com/theo/status/2049645973350363168 Sama's "alignment failure" RT: https://x.com/sama/status/2049715178611380317

Theo: turning Anthropic prompt caching OFF lowered T3 Chat costs

Same evening, Theo notes T3 Chat actually got cheaper after they disabled prompt caching on their Anthropic models. He posted token-utilization numbers in a follow-up. Reading: the new (shorter) cache TTLs combined with the cache-write surcharge mean T3's traffic pattern stops paying back the cache-write cost. This lands the day after Pierre-Yves Sottiaux from Anthropic posted that the team had to "trade off between a good experience or optimize for margin and not allow to draw more usage than you technically have" — steipete RT'd that one. The picture forming: customers running heterogeneous traffic are suddenly net losers on caching, and Anthropic is internally framing it as a margin defense.

Token utilization screenshot: https://x.com/theo/status/2049715045828096196 Sottiaux's defensive note (RT'd by steipete): https://x.com/thsottiaux/status/2049498939641217287

Cursor ships an SDK

Cursor announced the Cursor SDK late on Apr 29 — same harness, runtime, and models that power Cursor, packaged as a TypeScript library, billed against your Cursor sub. It ships with two modes:

  1. Local hackable agents you wire to any model you want.
  2. Managed cloud agents you embed inside your own product (sandboxing, computer use, demo videos, GitHub integration all handled).

leerob's gloss: "Cursor's starting to look less like an IDE and more like a cloud provider with an IDE attached" (paraphrasing one of the replies, but he RT'd approvingly). Notable replies:

  • Boaz Hwang: "The underrated part is the harness. Once sandboxing, repo context, and demo output are boring, the model choice gets less dramatic."
  • JMoon: "the local + managed split is the right abstraction. hackable for builders who need to own the primitives, managed for product teams shipping on top."
  • Gregor: "Shipping managed agents is slick until you lose control of the harness and spend two days debugging something you can't see. Still worth it, but that cost is real."
  • Dante Lex: missing baseUrl config blocks BYOK / corporate proxy / air-gapped use — would PR if open source.
  • Ethan Clinick: confirmed Windows bug — spawn C:\Program Files\Git\bin\bin\bash.exe ENOENT (extra \bin).

leerob also says a compiled-language version is being thought about behind a gRPC bridge, and that the SDK is not coding-only — "can be used for a lot of things". potetotes (Cursor): "set your agents free."

leerob thread: https://x.com/leerob/status/2049522118757331426 Cursor announcement: https://x.com/cursor_ai/status/2049499866217185492

Zig project bans AI-assisted contributions — and the rationale is sharper than expected

Simon Willison wrote up Zig's blanket anti-LLM policy (simonwillison.net/2026/Apr/30/zig-anti-ai/). The Zig CoC says: no LLMs for issues, no LLMs for PRs, no LLMs for comments. Simon's sharpest point:

"I particularly appreciate how this rationale isn't based on the idea that LLM code is of poor quality compared to code written by hand — the quality of the code isn't the deciding factor at all here."

The framing in the thread (which Simon endorsed) is review-budget-as-mentorship-pipeline:

  • Mehdi Salmani: "Code review isn't a quality gate — it's how jr engs become sr engs. Automate the output and you save time today! Automate the process and you lose your pipeline of people who understand the system in 2 years!"
  • Andrew Hart: "It's a queue management decision, not a quality decision."
  • Bakhtier Gaibulloev: "reviewer attention is the rate limit. zig spends that budget on contributor onramp. ai pull requests consume the queue without producing future contributors, so the trade is a supply collapse."
  • Pekka Enberg counters: "Their policy forced Bun developers to fork the project, for example, which seems a net negative to me."

A reply asks whether Zig also bans security disclosures found via LLMs. Simon points to the actual CoC (ziglang.org/code-of-conduct/) — it's not in there.

Thread: https://x.com/simonw/status/2049661673427042509

Sandcastle: Matt Pocock's open-source software factory

Matt Pocock open-sourced Sandcastle — a TypeScript primitive (sandcastle.run()) that orchestrates sandboxed coding agents, with first-class support for Claude Code, Codex, and Pi (agent-agnostic). It handles "running the agent inside [your sandbox], tracking events, token usage, resumability, and worktree management." Standard Docker containers underneath; Vercel Sandbox snapshots also supported.

Matt's recommended workflow: /grill-with-docs → /to-prd → /to-issues, then hand the issues to Sandcastle.

A user pushed back that bind-mounted Docker containers without network controls / read-only rootfs aren't true sandboxing. Matt's reply: "Yep, it's on you to add layers on top of that. But it's pretty simple in the Dockerfile." He also confirms it works against any issue tracker (GitHub/Jira/Linear/local) via labels.

Repo: github.com/mattpocock/sandcastle Thread: https://x.com/mattpocockuk/status/2049506712801935611

Pragmatic Engineer x Pi: Mario, Armin & Peter on building with self-modifying agents

Gergely Orosz dropped his Pi/OpenClaw episode with badlogicgames (Mario, Pi creator) and mitsuhiko (Armin, Flask creator and longtime Pi user). The three takeaways Gergely highlighted are juicy:

  1. AI is killing the senior-engineer "no". Armin's observation: "more junior engineers and product managers deploy agent-scripted counterarguments when a senior colleague kicks an idea to the curb. This makes decision-making exhausting, and more bad ideas make it into production as a result."
  2. Pi's pitch is bespoke harnesses. "Different projects need different harness types because… the same hammer is not ideal for every single construction job. As such, Pi is built with the goal of allowing the creation of specialized harnesses. It can modify itself so that a user can create the bespoke harness needed for any task."
  3. Automation bias is the biggest agent risk. "Once devs confirm that an AI agent can produce acceptable code, they start to review its output less often, even though agents can — and do! — produce slop."

Timestamped chapters worth jumping to: 15:15 (How 30 dev teams use AI agents: learnings), 28:30 (Downsides of over-automation), 1:00:22 (Complexity as the enemy), 1:11:52 ("Slow the F down"), 1:16:40 (MCPs vs. CLI).

YouTube: youtu.be/n5f51gtuGHE Thread: https://x.com/GergelyOrosz/status/2049503062289469618

Steipete: codex /review now lives inside ClawSweeper, with automerge loops

Following yesterday's "codex reviews every commit on OpenClaw main" pattern, steipete integrated codex review into clawsweeper with the same system prompt as /review, plus automerge — and the loop doesn't stop until no new issues are found. He compares it to Greptile: "They find different things, but codex review finds better stuff."

Replies are split into "this is the only way reviews should work" vs "loops until no issues sounds great until it loops on a flaky issue and merges 40 noise commits at 3am":

  • Aleksandr Fulha: "review-loop until clean is sharp — but reviewer drift bites. same agent grading its own previous fixes converges on local minima fast. paperclip splits reviewer/fixer into separate contexts (own CLAUDE.md, no shared mem) — killed ~80% of false-cleans."
  • Muhammad Hassan: "would put a hard cap on loop count and a per-issue confidence threshold before automerge fires."

Repo: github.com/openclaw/clawsweeper Thread: https://x.com/steipete/status/2049518771023360010

CC bug hunt: Thariq finds the big-file-write hang

Continuing the 50+-fixes thread, Thariq posted what looks like the cause of one of the longest-running CC papercuts — "my white whale is when CC sometimes looks like it hangs during big file writes, I think we found it 😭". He's also pushing the no-flicker renderer (/tui fullscreen) toward becoming the default. The reply thread is a goldmine of user-reported white whales worth flagging:

  • neekhil vatsa: long complex tasks that get to 60–70% then forget the job — Thariq asks "compact or timeout?"
  • Daniel Chu: long eval runs with multiple MCP calls grow memory until CC slows/hangs.
  • Pranav Mishra: /clear with todos still in flight — they survive into the new context and CC gets confused about a task it never started. Thariq: "are you using CLAUDE_CODE_TASK_LIST_ID?"
  • Shamir Wehbe: /rewind did not actually change the code back to the checkpoint.
  • swapp: lots of session-loss bugs in remote control, sessions should be reconnectable from mobile.
  • Richard: blank screen after CTRL-G has been broken "a few months" (github.com/anthropics/claude-code/issues).
  • davieJPG: "feature removal request: should be able to turn off adaptive thinking entirely". Thariq: "why? adaptive thinking is just how Claude thinks now."

Thread: https://x.com/trq212/status/2049234228290961690

Other notable threads

LlamaParse MCP server (jerryjliu0)

LlamaIndex shipped the LlamaParse MCP server at mcp.llamaindex.ai/mcp. Three things they had to solve that other MCP-server builders will hit: MCP has no built-in file-upload primitive (they built a URL-based upload endpoint and tied it to parse operations), they integrated WorkOS OAuth so the agent identity matches the platform identity, and they baked in observability + rate-limiting from day one.

Thread: https://x.com/jerryjliu0/status/2049652650938085689 Blog: llamaindex.ai/blog/llamaparse-mcp-the-tooling-layer-for-your-document-agents

Simon ships LLM 0.32a0

Simon Willison released LLM 0.32a0 — a major backwards-compatible refactor of his Python library and CLI. The headline change: it should "work better with reasoning models and other new frontier capabilities." Notes: simonwillison.net/2026/Apr/29/llm/ · https://x.com/simonw/status/2049567761136058699

mitsuhiko: "FastCGI is still the better protocol for reverse proxies"

Armin RTs Andrew Ayer's piece on FastCGI's 30th birthday (agwa.name/blog/post/fastcgi_is_the_better_protocol_for_reverse_proxies) — argues FastCGI sidesteps the security pitfalls baked into HTTP reverse proxying. mitsuhiko's gloss: "What's old is new again. I'm not sure why FastCGI lost out to HTTP so much in recent years on backends but I would guess that people wanted websockets. Still a good protocol."

Thread: https://x.com/mitsuhiko/status/2049593120786911244

swyx's "be me" greentext recap of Talkie

Probably the most-shared meme of the day. swyx greentexts the entire Talkie pipeline — from "the internet is polluted by ai slop, we need low-background tokens" to "release it > it's the most confidently racist model ever released by humankind > mfw". Worth reading for the synthesis-from-historical-RLHF detail (etiquette manuals, letter-writing manuals, cookbooks, dictionaries, encyclopedias, poetry and fable collections shoved into ChatML).

Thread: https://x.com/swyx/status/2049652947408372187

LLMJunky: today's models will look pedestrian in a year

Short but clean LLMJunky take — "We're gonna look at today's models next year, and think the same way about them that we think about o1 and Sonnet 3.7 now. These same models so many are going crazy about will look pedestrian. There may be depreciating gains, but the increased pace more or less balances that." Pairs naturally with karpathy's "AI Psychosis" framing from earlier in April — both arguments point at frontier-model capability deltas hidden inside benchmark plateaus.

https://x.com/LLMJunky/status/2049638561406685396

LLMJunky also dropped a small but practical CC tip: "after you run a skill, it's not a terrible idea to ask the agent if there's anything in the skill itself that can be updated to make it more efficient." — https://x.com/LLMJunky/status/2049541361389490341

Goblingate / OpenAI on Goblins

OpenAI published a piece (openai.com/index/where-the-goblins-came-from/) talking about goblins — apparent fallout from the recent Codex "goblin mode" jailbreak meme. Theo: "Goblingate is way funnier than I expected tbh" (https://x.com/theo/status/2049703981245694097).

Off-topic / non-AI

Theo on the Framework 13 Pro

"I am genuinely really really impressed with the Framework 13 Pro. I was unhappy with the old 13. Awful keyboard, trackpad, and display. New one is way better in all categories. Best PC trackpad I've used, awesome keyboard, no more display wobbling. Feels 10x more premium."

Caveat per Framework: pre-production unit. https://x.com/theo/status/2049716048958730585