App Review fires back, OpenClaw fires up

May 16 had a clear top tweet, and it came from frustration rather than a product launch. After that, the feed was dominated by one project (OpenClaw) and one ecosystem question: how far can you push Codex before it stops acting like OpenAI's product?

Lex Tang (@lexrus), an iOS/macOS developer, shared a screenshot comparison that pulled 3,096 likes — the day's highest engagement by a wide margin. 1

His app had been rejected more than five times by Apple's App Store review team. During that same period, another app containing a single line of placeholder text sailed through and got approved. He posted both screenshots side by side.

"The company that recently rejected my app more than 5 times actually approved an app with only one line of placeholder text." 1

It's a well-worn complaint — App Review inconsistency has been documented for years — but the direct side-by-side evidence made this instance unusually easy to share. 166K views, 23 retweets.

Peter Steinberger (@steipete), creator of OpenClaw, posted three separate updates on the evening of May 16. Together they painted a picture of a project hitting a good patch.

The first was the most cryptic and the most liked. At 16:47 he wrote: "Looks like our focus on performance paid off." 2 That was it. 1,480 likes, 237K views, 111 replies, 12 quote tweets — the second-highest engagement in the day's window. The brevity suggested he had a specific benchmark in mind and was confident the audience would go looking for it.

Two hours before midnight, he shifted to tooling. He promoted clawpatch.ai, a service that connects Codex to your repository and scans for bugs you didn't know existed. 3 "It's amazing at uncovering bugs you didn't know you had." 1,147 likes, 80K views. A screenshot in the tweet showed the tool running against a real codebase.

Ninety seconds later came the most technically interesting post. Steinberger described OpenClaw's new "Lossless" feature: rather than simply summarizing past context, it compacts conversations into structured blocks and builds a tree that the model can traverse to reference earlier messages. 4 The result, in his framing, is an effective infinite context window — without dumping a monolithic prompt on the model. 837 likes, 96K views. The tree-based approach is a meaningful architectural distinction from most "memory" implementations, which flatten history into summaries that lose structure.

Two separate threads on May 16 revealed how far some users are pushing Codex's configuration layer.

Nyarime (@realNyarime) published a step-by-step tutorial for routing Codex conversations through a third-party AI relay while keeping your ChatGPT login active. 5 The method involves editing two files: set auth_mode to "chatgpt" with OPENAI_API_KEY: null in auth.json, then add a provider entry with experimental_bearer_token and requires_openai_auth: true in config.toml. The payoff: Codex Mobile, plugins, and quota queries all unlock, while actual inference routes through a third-party provider. 540 likes, 61 retweets, 941 bookmarks — the bookmarks are the real signal here, this is a post people saved to action later.

On a different front, Baoyu (@dotey) shared DeepSeek's job posting for an "Agent Harness Product Manager." 6 The framing in the job description is blunt: "Model + Harness = Agent. Everything beyond the model itself belongs to the Harness layer." The role involves shaping DeepSeek's Harness product roadmap while coordinating with researchers, engineers, the open-source community, and end users. The experience requirements list is notably specific — two or more years as a PM, plus documented heavy usage of Claude Code, Codex, Cursor, GitHub Copilot, Manus, OpenClaw, and Hermes. The JD also lists understanding of KV Cache, Agent Loop, Tool Use, MCP, Memory, and Subagents as baseline knowledge. 241 likes, 32 retweets. Whether or not you apply, the JD is a useful snapshot of what DeepSeek considers the complete stack for an agent product.

@QT9277 shared a story with 736 likes and 196K views. 7 Someone trying to quit adult content took an extreme approach: blocking over 1,000 websites. Then they uploaded the full blocklist to GitHub, in a repository named blocked-sites-in-south-korea. The repository now functions as the most comprehensive curated directory of the exact sites they were trying to avoid. QT9277's caption: "越禁越火、越藏越香，懂的都懂" — roughly, the harder you suppress something, the more attractive it gets. The logic checks out.

@realNyarime posted a customer service anecdote that landed at 240 likes. 8 A Steam Controller order was accidentally routed to the UK instead of its intended destination, causing a shipping delay. Steam support offered a free standard-edition game from the player's region as compensation, with no restrictions on which title. The player picked Forza Horizon 6. Whether or not this is Steam policy or a one-off gesture isn't clear from the post — but the screenshot of the support message made it easy to share.

@SophiaFioren posted two pieces of architectural and historical material. 910 A video of the Royal Basilica of Our Lady of the Forsaken in Valencia, Spain — a 17th-century structure — drew 230 likes and 54 retweets. Later in the evening, she shared an image of Queen Charlotte's handwritten personal notebook from 1765, a Georgian-era manuscript with marginalia and personal notes, which collected 190 likes and 31 retweets.