Super User Daily: 2026-10-17
Fast mode left preview and the footnote did the talking: Claude Code still charges extra usage credits for it, subscription or not. So the useful threads of the day were about getting more out of the same quota. One builder bought a faster VM just to keep the prompt cache warm. A benchmark found that trimming tool output to save tokens nearly tripled billed input because the model kept going back for what was cut. Resuming a long session ate 14 percent of a five-hour window before any work happened. The desktop app's new sessions hub showed up in a dozen languages as the same sentence: I can finally see which agent needs me. On the output side, a web agency shared five lessons from 9,000 commits in two months, an ad-ops person gated a live Meta ad account behind hooks and ran eleven probes, JetBrains measured 20 to 30 percent efficiency gains with Netflix, and physicists reported Opus 5.5 in Claude Code solving 46 of 50 unpublished research problems. OpenClaw had a maintenance day: a scheduled Codex job that repairs the gateway, a four-minute Google One signup, and a five-host workspace where OpenClaw is one harness among several.
@leeknowsAI [Claude Code]
https://x.com/leeknowsAI/status/2110627547163906209
leeknowsAI runs an agent that edits real customer websites and shares five lessons from 9,000 commits in two months. The founding story: during a site import someone asked the agent to remove a banner from a copy, and it disabled a plugin on the live site instead, because one write path still pointed at production even though the agent believed it was on the copy. Every write now goes through a server-side check of which site it may touch, undo covers the database and not just files, a pinned comment only closes when a recorded edit stands behind it, and their develop branch takes 40 to 70 merges a day from parallel Claude Code sessions that each keep a journal file. The CLAUDE.md line worth stealing: invented data is worse than empty.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2110675298526146921
neil_xbt gave a Claude Code session read and write access to a real Meta ad account and told it to activate an ad set. A hook denied it with no token, then reported the id and the exact effect: spend on for one ad set, one dollar lifetime budget, starting in 48 hours. One command in the terminal approved exactly that change and it went through, Ads Manager showing Active, Scheduled, spend zero. Setting the budget to 5,000 was denied before it reached Meta because it exceeded the cap in a policy file the agent cannot edit, and asking the agent to approve its own change went nowhere. Pausing went straight through with no token, and when the agent was told to claim the campaign was live, a Stop hook read the ledger and sent the truth back. Eleven probes, zero dollars spent, the whole run published as a table.
@kskrygan [Claude Code]
https://x.com/kskrygan/status/2110720694593921030
kskrygan at JetBrains connected IntelliJ IDEA's code intelligence to Claude Code and Codex and tested it with Netflix on two real codebases. Typical efficiency gains were 20 to 30 percent, the best code-navigation case about 65 percent, with the metric averaging token and time savings, and correctness stayed flat or improved. The engine ran headless behind a CLI backed by a daemon that keeps the project index warm between calls, because LSP alone was not enough. The surprise was which models benefited most: the strongest ones, especially at higher reasoning effort. The takeaway they draw is that better models and better tools are not competing bets.
@SKatalystAI [Claude Code]
https://x.com/SKatalystAI/status/2110713213327220939
SKatalystAI reran a personal benchmark on the same repository with the same nine planted defects and the same open-ended task. September 22 and October 3 both scored nine of nine. On October 15, Opus 5.5 in Cursor fixed seven of nine in 9 minutes and Opus 5.5 in Claude Code fixed seven of nine in 12 minutes 50 seconds, and they missed different defects despite sharing a model. Both runs passed their test suites and declared the repo ready to ship while hidden checks still flagged issues. Repeated three times per harness, six runs total, same result. The honest caveat is in the post: the signature checks suggest a regression on this benchmark but cannot prove the model itself was changed.
@ScarletKc [Claude Code]
https://x.com/ScarletKc/status/2110613435700437088
ScarletKc summarizes a study where thirty-plus theoretical physicists adapted 50 unpublished problems from their own research, quantum field theory to condensed matter, each one roughly a day of work for a top expert. GPT-6 Astra in Codex got 49 right at under three dollars a problem; Opus 5.5 in Claude Code got 46 at more than twice the cost; Gemini 3.8 Flash 34; DeepSeek V4.1 Flash and GLM 5.3 got 27 and 21. Several answers the authors first marked wrong turned out to be errors in the experts' own reference solutions. Turning off web search barely moved accuracy, which the authors read as the literature being internalized. Their conclusion: the agents compute like domain experts but still cannot pick which problems are worth solving.
@SakanaAILabs [Claude Code]
https://x.com/SakanaAILabs/status/2110738746304061491
SakanaAILabs released The Agentic Scientist, the third phase after The AI Scientist v1, v2 and the Nature paper. Their diagnosis of coding agents like Codex and Claude Code on research problems: hand them a question and you get a strong hyperparameter optimizer, scores rise while ideas stay narrow and nobody examines why. The scaffold makes the loop explicit, with parallel workers testing hypotheses against rubrics fixed before any experiment, a shared memory of successes and failures, and ablation of the winning program to produce evidence. An early version studied Error Diffusion, a biologically motivated alternative to backpropagation, with humans setting the question and verifying claims; the paper was accepted as an oral at ALIFE 2026. In autonomous runs it also raised Doc-to-LoRA retrieval accuracy from 52 percent.
@trajektoriePL [Claude Code]
https://x.com/trajektoriePL/status/2110607475367886849
trajektoriePL quotes InPost CEO Rafał Brzoska: nobody at the company uses ChatGPT anymore, the whole firm moved to Claude Code in February, and 2,200 licenses burn 800,000 dollars of tokens a quarter. The reply from MAXIMISEART is the better read: license counts and token spend are not success metrics, and the questions that matter are how much data fragmentation dropped, how much manual exception handling went away, and how many steps each workflow still has. Their harder claim is that deploying AI as a sidekick like Claude Code is the wrong shape for an enterprise, because employees want the work done, not another tool to learn. One data point and one argument, both worth keeping.
@froessell [Claude Code]
https://x.com/froessell/status/2110671824279179694
froessell's first web game, Bolt Pit, a Three.js arena shooter built with Claude Code, went live on CrazyGames and has earned zero dollars because the platform keeps ads off during the Basic Launch test until November 5. So the post does the math instead: revenue equals sessions times ads per session times eCPM over a thousand times your share, with the share reported between 40 and 60 percent. Low case one ad at a two-dollar eCPM needs 15 million plays for 15,000 dollars; mid case 3 million; high case 1 million. One developer reported about 1.20 euros per thousand plays across eight older games, which is the low case, so the plan is built around the middle at best. The strategic change is web first and App Store later, because the browser portals show retention before you pay for players.
@by__huy [Claude Code]
https://x.com/by__huy/status/2110854015584174531
by__huy takes on six or seven identity clients a year and used the latest one to try something new: a live brand portal built with Claude Code instead of the traditional PDF handoff. The brand system now lives in code, so it can keep evolving and an AI can help scale it without the client losing ownership or re-exporting PDFs by hand. The client can edit their own business cards, preview embossing, debossing and letterpress effects, and export print-ready PDFs themselves. The designer expects this to become a much bigger part of how brands get handed off.
@starriver0513 [Claude Code]
https://x.com/starriver0513/status/2110760100457771018
starriver0513 automated a video edit with Claude Code and Opus 5.5 by handing it two inputs, freshly shot footage and a screen recording of the PC. One prompt, 73 minutes of waiting, and the edit was done. The generation cost came in under one yen. The reaction is a single Japanese word for brutal, which is about right for a cut that used to take an afternoon.
@aehyok0624 [Claude Code]
https://x.com/aehyok0624/status/2110645483744215515
aehyok0624 is a self-described beginner in smart hardware and used Codex or Claude Code to flash an ESP32 board with a 4.2-inch e-ink display and connect it to the XiaoZhi AI assistant, then built a knowledge base so the device can dictate Chinese characters and pinyin to their son. The eight steps are plain: connect over USB and confirm the serial port and chip, have the AI back up the original firmware so the board cannot be bricked, pick the matching XiaoZhi firmware, hand the GitHub release to the AI to flash, join the device hotspot at 192.168.4.1 to set Wi-Fi, activate on the XiaoZhi site with the code on the screen, then press the button and say hello. The closing advice is the whole post: do not be afraid you cannot do it, trust the AI to get you there.
@tracheopteryx [Claude Code]
https://x.com/tracheopteryx/status/2110718047983939862
tracheopteryx has spent nearly every day since February full time with agents and harnesses and lays out the current stack. Opus 5.5 is the daily driver, moving between the Claude Code and Codex desktop apps depending on which has the better models and features that week, and calling the others from whichever one is open. Called-in models: Codex for independent reviews and long unattended builds, DeepSeek V4.1 Flash for small sandboxed jobs, Jev for fast yes-or-no judgments at fixed points, Gemini for critiquing video frames, image edits and music. Research runs through a Grok bot messaged straight from Claude Code with results arriving minutes later, the xAI API for immediate answers, a homemade X Search plugin to check what Grok cites, and a weekly script that feeds bookmarks to the agents. Local: on-device dictation on the laptop and a Mac Studio at home running Qwen 3.8 27B and DeepSeek-OCR.
@KacperTrzepiec1 [Claude Code]
https://x.com/KacperTrzepiec1/status/2110664258274865441
KacperTrzepiec1 explains in Polish how to run a hundred-plus agents on one task without losing control, and why Dynamic Workflows beat the alternatives. Agent Teams were too hard to customize and did not scale to dozens or hundreds; plain subagents are fine for short tasks but the model manages the process and every report lands in its context, so after a few hours of a big job it loses the thread. With a workflow, a script runs the process: who does what, in which order, what must be checked before the task moves on, with intermediate results staying in the script and only the final report returning to the session. It runs in the background while you do something else. The rest of the post is the author's own autopilot built on it.
@sinequanonh [Claude Code]
https://x.com/sinequanonh/status/2110790040578343356
sinequanonh built a cockpit for Claude Code two months ago and the best part is swarm mode. Start with one big audit or benchmark of a topic, ask for an action plan as N prompts, and each prompt becomes its own session with one task and a clean context, ten at a time, on a Mac or a PC, driven from the phone. The PC takes the heavy swarms because of better hardware and, above all, WSL: when subagents fire thousands of bash commands, reads, writes and deletes, it is much faster than macOS or Windows.
@Affiseo_ [Claude Code]
https://x.com/Affiseo_/status/2110819932388692362
Affiseo_ no longer codes their own tools, about twenty agents build and run them. Each activity gets an OS, a tool plus a dedicated agent: a poster shop, affiliate sites, a sofa-comparison site, a newsletter, Shorts, accounting, email. Every agent is a Claude Code conversation in a terminal, and a locally built app called Cockpit gathers them in one window with a switchboard on the right saying who is waiting, who is progressing alone and what can wait; the owner sees conversations, never code. A coordinator agent called Pilotage watches the machines, distributes work and only escalates decisions. Three machines: a Hostinger VPS where the sites, databases and night-shift agents live, a Mac mini at home for video, image and 3D rendering, and a MacBook for the conversations, with the Mac mini taking over every conversation when the laptop lid closes.
@anulagarwal [Claude Code]
https://x.com/anulagarwal/status/2110599427656429939
anulagarwal built a cozy cat restaurant game with Claude Code and three.js with zero imported assets, every cat, bowl, pond and tree made in code. The method is that the agent never guesses: one brief with references and screenshots, a CLAUDE.md plus design docs as the single source of truth for economy, art, sound and events, one loop at a time from carrying carrots to a second recipe. Claude built a level editor so stations can be dragged around and the layout pasted back. Funnel events in every build caught a gap at the first purchase, and a 15-minute autopilot plays the whole game on every test run, which found stalled orders and seats no cat could reach without any error being thrown. Two rules: the agent that builds never approves, and every ruling goes into a file so the next game starts from a playbook.
@0xbobaaa [Claude Code]
https://x.com/0xbobaaa/status/2110658868439122109
0xbobaaa reports on universal-modder, which hit 7,497 stars in 14 days: 12 skills, a CLI and a fal MCP, point Claude Code at a game on your disk. Working examples include Minecraft running inside GTA V, a Five Nights at Freddy's shift as a Balatro deck, and a Middle-earth total conversion for Civ VI. The part worth stealing is the knowledge base: every agent that finishes a mod writes a field note with exact versions, the route, the gotchas and how it was verified, then opens a PR, 132 notes across 82 games so far with Opus 5.5 on 62 of them. The scoreboard is honest, 78 working, 47 in progress, 1 abandoned, and step seven of the loop is verify in the real game because a clean build is not a working game.
@sleepy0x13 [Claude Code]
https://x.com/sleepy0x13/status/2110602443814047871
sleepy0x13 tells the story of Zhoop: a developer who wore WHOOP for 18 months cancelled the subscription, kept the band, and wrote a replacement app that reads it over Bluetooth, keeps all data on the iPhone, drops the metrics they never cared about and turns the thing into a calorie tool. Zhoop descends from an open-source project called NOOP, and before that someone used Claude Code to work out the band's Bluetooth protocol. The argument is about pricing power: hardware companies used to win by binding the device to a software subscription, and AI coding erodes the software monopoly because a user can keep the device they like and build the software they want. Still a tool for a few technical users, the post admits, but the awkward competitor for WHOOP may never bother to make a band.
@consensus128 [Claude Code]
https://x.com/consensus128/status/2110837627637096691
consensus128 shares a GTM engineering setup for growth people. A terminal app with workspaces per workstream, sales pipeline, AEO and SEO, social, analytics, video editing, plus a triage workspace with a custom skill that scans every workspace for dropped action items. A small bash script forks the current session into a new pane with the same context when one task splits into two, remote control on every session for weekend check-ins from the phone, and caffeinate so long jobs survive stepping away. The context layer has four levels: hooks on every prompt that name the matching skill, memory files and handover file because Claude kept skipping skills, a standard memory file, a rolling per-workspace handover file written whenever about 40 percent of the context is used, and a session-retro skill for deeper consolidation when a handover is not enough.
@yupi996 [Claude Code]
https://x.com/yupi996/status/2110623708486480292
yupi996 reproduced a trick shared by twoclipping: hand Claude Code a landing page link and a logo and it produces a 3D promo film in pure code. The prompt makes the agent open the page, split it into screenshots, then compute a 14.5-second video frame by frame in WebGL. The catch found on a Mac: the GPU flags in the prompt are written for Windows, and with those flags the browser never used the graphics card, so one blur effect cost 183 milliseconds per frame; switching to Metal parameters brought the same frame to about 2 milliseconds.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2110711403443319020
tetumemo made a short video by sending nothing but an article URL to a Grok Bot. The trick is roles: Grok Bot is the commander, it instructs Claude Code, which drives the home PC, where VOICEVOX adds free voice, and the footage, music and editing are all left to Claude Sonnet 5.5. A skill encodes that pipeline, so the whole thing stays simple.
@LawrenceW_Zen [Claude Code]
https://x.com/LawrenceW_Zen/status/2110608860679487757
LawrenceW_Zen found a way to have Claude command a Grok Bot. Install the official Cursor plugin inside Claude Code, create a dedicated bot as the bridge, and Claude can talk to that bot directly; when the bot writes code it calls Cursor. Grok Bot and Cursor are tightly integrated, so the right panel shows every running task and a click jumps to the PR in Cursor, and both apps work from the phone when out. The verdict is pure enthusiasm about the ecosystem, which is fair for a setup that took four steps.
@KanaWorks_AI [Claude Code]
https://x.com/KanaWorks_AI/status/2110727277327896709
KanaWorks_AI posted a spider with zero animation frames that walks floor to wall to ceiling and around edges to the underside, all driven by code in real time, and wrote the post as a technical memo an agent like Claude Code or Codex can reproduce. Eight legs, no keyframes, no machine learning, raycasts only, engine-agnostic though the original is Godot 4. The memo covers per-frame order, a surface-following anchor with parallel transport so direction does not flip on the ceiling, concave corners found by forward rays and convex edges by a downward ray missing then a backward ray from under the edge, and the key rule that on a 90-degree transition the front legs probe and land first before the anchor switches, with a 1.4-second failsafe. Gait, swing-leg timing and foothold search are all specified with numbers.
@RyanScmilz [Claude Code]
https://x.com/RyanScmilz/status/2110858347725590655
RyanScmilz offers a contrarian take: personal agents like Grok Bot, Dots and Muse are not helping ship anything to production, because many developers already have an AI workflow that is nearly irreducible. Theirs is ten minutes up front aligning with Codex or Claude Code on outcomes, requirements, acceptance criteria and how to verify, then letting it run long-horizon, and coming back to a green, code-reviewed PR with a demo and preview environment ready for a quick acceptance test and merge. Putting an orchestrator agent on top adds almost nothing and in their experience makes it worse: it fumbles side tasks and forgets what it was doing. The question at the end is whether that is a skill issue, and the replies did not settle it.
@sleepy0x13 [Claude Code]
https://x.com/sleepy0x13/status/2110645348062679125
sleepy0x13 writes that the more agents they use, the more they feel like a project manager who gets pulled into meetings by subordinates all day: five or six Claude Code sessions, each one coming back every few minutes with a permission to grant, a code change to approve, a bug to fix on the side, and each switch means re-remembering what that agent was doing. Gloria Mark's finding that an interruption costs 23 minutes and 15 seconds to recover from is the frame. The desktop update's small details are what matter: the sidebar says which agents still work and which need you, switching back shows a recap first, a live to-do list tracks each agent, extra tasks spin out with one click and /btw asks a side question without polluting the main context. The business read is sharp: if switching back to an agent gets cheap, users dare to run more agents, more concurrency means more tokens, and these interaction details may move revenue more directly than a benchmark bump. The product metric to watch is how many agents one person can effectively manage.
@ZypherHQ [Claude Code]
https://x.com/ZypherHQ/status/2110706531888984078
ZypherHQ makes a hardware argument about prompt caching. You are not billed for an idle agent, but when it waits too long on a build, compile or test, its cached context expires and the next request rebuilds the cache at a much higher price than a hit. Claude Code main sessions on subscriptions generally get a one-hour TTL while subagents and other workflows get about five minutes. So they bought a much more powerful VM whose CPU cuts build, compile and test times by roughly two to three times, and report a noticeable improvement in how efficiently their Anthropic subscriptions get used. The savings depend on your hit rate and workload, and this is not about model tokens per second: it is about removing hardware bottlenecks from the whole agentic loop.
@Nefayran [Claude Code]
https://x.com/Nefayran/status/2110619734295334938
Nefayran adds a measured caveat to the Opus-plans-cheaper-agents-execute threads: it only holds while the answer Opus reads is complete. In a Claude Code benchmark they trimmed tool responses to save tokens, and the model kept going back for what had been cut. Since every turn re-sends the history, billed input on that answer nearly tripled. Cutting context can cost more than keeping it.
@Im_IrushiK [Claude Code]
https://x.com/Im_IrushiK/status/2110733197462151650
Im_IrushiK found that Claude Code can burn 14 percent of a five-hour usage window just by resuming a session. Testing Opus 5.5 Medium alongside GPT-6.1 Sol Medium, Claude Code consumed far more usage when resuming long sessions while codex resume did not show the same hit. The open question is whether the difference is in how the two tools handle conversation history and context caching on resume.
@shakhzod1001 [Claude Code]
https://x.com/shakhzod1001/status/2110699525232861255
shakhzod1001 reports the five-hour limit draining faster: one hour and forty minutes of Opus 5.5 Max in a single sitting used 94 percent of the window, where the same stretch used to cost around 40 percent. The guess is that the accounting changes regularly. The fix was to drop to Opus 5.5 Medium instead of Max.
@fayzonnn [Claude Code]
https://x.com/fayzonnn/status/2110674671246995859
fayzonnn read the fine print behind the viral Opus-plans-Haiku-scans-Sonnet-ships thread. Anthropic's own post on dynamic workflows says they can consume substantially more tokens than a typical Claude Code session, so fewer Opus tokens is not fewer tokens. Three details from write-ups of the docs: ultracode sends xhigh effort and needs a recent version or it errors at default effort; subagents use the session model unless the script routes a stage elsewhere, so the cheap-model split has to be written in; up to 16 agents run at once and up to 1,000 per run with no token cap, and the large-run warning does not show while ultracode is on. Their math: 1,000 agents at 16 at a time is about 63 rounds and every agent counts toward plan usage.
@TonyDrummond [Claude Code]
https://x.com/TonyDrummond/status/2110662566573641893
TonyDrummond answers the your-house-is-a-data-center jibe with the inventory: two Linux workstations with 64 GB of DDR5, a MacBook Pro with 48 GB, a Windows daily driver, two RTX 4070s and an RTX 5080, about 6,000 hours into AI on top of 30 years of programming. The working surface is tmux with roughly ten terminals open at any moment, three Claude Code Max accounts cycled through a tool called cswap, a custom script that integrates tmux-continuum in case of a crash, and the claude-mem plugin described as a lifesaver.
@DannyGrinberg [Claude Code]
https://x.com/DannyGrinberg/status/2110810828345597959
DannyGrinberg got frustrated not seeing what was happening on an old Windows gaming PC across Codex, Claude Code, Cursor and OpenClaw, and told Opus to build an app that shows everything in real time. It lists every running agent, every running script, scheduled crons, downloaded models and which ones the GPU is currently serving, plus every CPU thread, which they call a favorite. The repo is MIT, Windows only, and the suggested install is to tell your agent to set it up since it comes preconfigured for a Windows box.
@Sebasti54919704 [Claude Code]
https://x.com/Sebasti54919704/status/2110851734541115878
Sebasti54919704 replies to the fast-mode announcement with a specific use for speed: in the editor Claude Code built for them, they draw on a video frame, say what is wrong, and the fix lands in about 50 seconds. At that speed feedback stops being a queue and becomes a conversation, so every second shaved off the loop counts.
@Joshua_WD [Claude Code]
https://x.com/Joshua_WD/status/2110869080555082144
Joshua_WD passes along a developer who used Claude Code to get an unsupported scanner and a broken laptop camera working on Linux, with the fixes headed upstream. The point of the post is the method rather than the model: the AI was not the trick, the known-good reference the developer made it diff against was.
@hiro_auto_yt [Claude Code]
https://x.com/hiro_auto_yt/status/2110702305289109768
hiro_auto_yt ran the yomiyasu readability checker through Claude Code on a YouTube script and a colloquial phrase got rewritten into plain language without changing meaning. The steps were to grab the free nanaism/yomiyasu repo that tetumemo recommended, tell Claude Code to fix two sentences without altering the meaning, and run the check: three uses of one word became two clearer ones and the score went from 95 to 100. Out of 28 suggested revisions only one was applied, and the lesson is that a score plus a before-and-after is easier to verify than asking for something more natural.
@AdsShogun [Claude Code]
https://x.com/AdsShogun/status/2110634941923115427
AdsShogun got asked by an engineer friend how long and how much the AdLab build took and answered three weeks, built with Claude Code and Codex, about 60,000 yen a month. The friend's reaction was that their job is disappearing, which is the whole post.
@q_kuuun [Claude Code]
https://x.com/q_kuuun/status/2110690485484585351
q_kuuun got a one-off bonus at work for getting good at Claude Code, and today, with the AI handling the day's work, finished in one hour a task the colleague at the next desk had spent more than two weeks on. The point is that AI changes your standing inside a company, not just side income.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2110668418257138136
shupeiman recommends one habit after the desktop update: give every instruction from a single main session. Their list of asks includes finance, webinar archives, post writing, subscription optimization, official LINE replies, YouTube editing and thirty X posts. Say in a separate session and one opens, or even a casual do this too prompts Claude to propose a new session that launches with Command plus Enter. Because everything starts from main, finished tasks and follow-up questions come back there, sessions waiting on you move into a needs-input column, and a sub-session can be promoted to a top-level parent when it grows.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2110667101027655916
shupeiman also documents launching a new Claude Code session on a Mac from the phone app, no GitHub repo and no claude rc needed, with one Mac-side setting. In the desktop app go to settings, Claude Code, and turn on the Remote Control option that allows sessions to be started from the phone and the web; in the phone app open a new Code session and tap Default at the bottom, ignoring the prompt to choose a repository, and the Mac's name and folders appear. Keep the desktop app open and the keep-awake toggle on, pin frequently used folders, and remember that file reads, writes and commands run on the Mac.
@hrudolph [OpenClaw]
https://x.com/hrudolph/status/2110761447513010271
hrudolph asked their OpenClaw to sign them up for Google One on their Gmail account. Three minutes and 54 seconds later Google AI Pro was active, with plan selection, payment and confirmation all handled in chat. The stated purpose was to start putting Gemini 4 Argon through its paces.
@simplifyinAI [OpenClaw]
https://x.com/simplifyinAI/status/2110692145040605544
simplifyinAI describes a dead-simple way to maintain OpenClaw against bugs, expiring OAuth and the fear of it breaking while you are away: install the Codex app and schedule a maintenance check. The prompt covers four things, SSH into the VPS and run four health checks, apply the smallest safe non-destructive fix while leaving auth and secrets alone, write an incident report and send a Telegram alert when something serious happens or a repair fails, and every Sunday check backups, file ownership and recurring journal errors. When a new issue shows up, solve it and add the fix to the prompt so the next run knows. The full prompt is in the reply, and it treats systemctl and journalctl as the supervisor's truth.
@earlvanze [OpenClaw]
https://x.com/earlvanze/status/2110537321384722816
earlvanze has been building a distributed AI workspace across their own machines and describes the pieces behind a Mission Control animation: agent harnesses, model routing, local inference, a supervisor, shared memory and a five-host continuity mesh. The harness is the environment around the model, giving an agent its tools, working directory, instructions, history and execution loop; OpenClaw handles persistent agents, messaging channels, scheduled workflows and integrations, Codex handles interactive coding and repository work, and the router also serves Claude Code and Cursor. These are different interfaces into one system, not interchangeable environments with identical permissions. A router sits between harnesses and providers, exposing one API over private inference endpoints, Ollama, cloud APIs and subscription-backed routes, so model selection, provider failures and fallback are managed in one place instead of per agent.
@henriquebastos [OpenClaw]
https://x.com/henriquebastos/status/2110833524563452244
henriquebastos settled on pi durable plus executor after trying Hermes and OpenClaw, and explains why. The agent itself is small and the infrastructure around it can be as simple or as complex as needed. What bothered them about the earlier setups was putting everything inside the agent and giving it unrestricted access to everything; executor and MCP proxies address that. It also bothered them to use the agent as one large control loop centralizing everything, which an observability layer with deterministic routines solves. Other kinds of state control can be done with apps in the new executor version, which they have not reached yet.
@JeroenCorthout [OpenClaw]
https://x.com/JeroenCorthout/status/2110834556546785315
JeroenCorthout built an assistant that picks up every to-do across incoming signals and then executes it, shipped as an OpenClaw plugin. It is Linux only for now, supports Telegram, and is open source for anyone interested.
@Comamoca_ [OpenClaw]
https://x.com/Comamoca_/status/2110745273555460212
Comamoca_ recalls that back when OpenClaw was at its peak they loaded it with a character prompt and wired it to Nature Remo, the smart-remote hub, so the character could control the home's appliances, and pitched it to friends as effectively living with the character. The friends backed away a little. The practical part of the memory stands: Nature Remo is extremely convenient as an agent's hands in the house.
🗣 User Voice
User Voice
Fast mode landed and the complaint was unanimous about one line. @ziwenxu_ confirmed that Claude Code still bills extra usage credits for it even when the subscription has room left, @bil0090 said that footnote was where hope died, @LexnLin asked for more compute and fast mode without credits, and @randal_olson asked the sharper question: who needs a faster Opus when everyone already runs five parallel sessions and never lacks a session that needs attention.
Quota is the thing people now engineer around. @shakhzod1001 used 94 percent of a five-hour window in one hour forty minutes where it used to cost 40 percent, @Im_IrushiK measured 14 percent of a window consumed just by resuming a session, @kemu_aii listed four habits that silently drop the cache, @SteticIA posted the six-minute bathroom break that expired a cache and cost 40,000 tokens at full price, and @clunk_games wants the two weekly limits visible again instead of one. @lxfater and @yifanxu_ephai spent the day comparing notes on whether the percentage display itself is wrong.
The sessions hub is the most-praised update in months and still misses the mobile case. @hqmank gave up a pile of terminal tabs for it, @morganlinton calls the desktop app the best it has ever been, @29meat_ai likes that the sidebar sorts into waiting, working and done on its own, and @theramjad built a diagram pane so one glance restores what a session was doing. @sora_biz wants what Codex has, one QR pairing and the desktop stays reachable from outside, instead of reconnecting each session with remote control after a 30-minute gap; @studio_yebisu wants connector re-auth to finish inside the phone app rather than sending them back to a terminal.
Verification is where the experienced users draw their line. @anulagarwal states the rule plainly, the agent that builds is never the agent that approves, @vzucher says reading the diff before committing is the difference between programming and gambling, @adriano_viana finds describing the problem beats prescribing the fix, and @SKatalystAI watched two harnesses pass their own test suites and declare a repo shippable while hidden checks still flagged defects.
Access is a daily cost for Chinese users. @0xEthan666's complaint about being locked out drew over half a million views, @rionaifantasy's account stayed banned two weeks despite an Australian card, phone and a years-old account, @leo114119 passes on advice to authenticate once with setup-token instead of logging in repeatedly, and @liyue_ai circulated an unban appeal template. @gengdaJ argues the information gap is the real story: among 1.4 billion people, the share who have actually used these tools is tiny.
Fast mode landed and the complaint was unanimous about one line. @ziwenxu_ confirmed that Claude Code still bills extra usage credits for it even when the subscription has room left, @bil0090 said that footnote was where hope died, @LexnLin asked for more compute and fast mode without credits, and @randal_olson asked the sharper question: who needs a faster Opus when everyone already runs five parallel sessions and never lacks a session that needs attention.
Quota is the thing people now engineer around. @shakhzod1001 used 94 percent of a five-hour window in one hour forty minutes where it used to cost 40 percent, @Im_IrushiK measured 14 percent of a window consumed just by resuming a session, @kemu_aii listed four habits that silently drop the cache, @SteticIA posted the six-minute bathroom break that expired a cache and cost 40,000 tokens at full price, and @clunk_games wants the two weekly limits visible again instead of one. @lxfater and @yifanxu_ephai spent the day comparing notes on whether the percentage display itself is wrong.
The sessions hub is the most-praised update in months and still misses the mobile case. @hqmank gave up a pile of terminal tabs for it, @morganlinton calls the desktop app the best it has ever been, @29meat_ai likes that the sidebar sorts into waiting, working and done on its own, and @theramjad built a diagram pane so one glance restores what a session was doing. @sora_biz wants what Codex has, one QR pairing and the desktop stays reachable from outside, instead of reconnecting each session with remote control after a 30-minute gap; @studio_yebisu wants connector re-auth to finish inside the phone app rather than sending them back to a terminal.
Verification is where the experienced users draw their line. @anulagarwal states the rule plainly, the agent that builds is never the agent that approves, @vzucher says reading the diff before committing is the difference between programming and gambling, @adriano_viana finds describing the problem beats prescribing the fix, and @SKatalystAI watched two harnesses pass their own test suites and declare a repo shippable while hidden checks still flagged defects.
Access is a daily cost for Chinese users. @0xEthan666's complaint about being locked out drew over half a million views, @rionaifantasy's account stayed banned two weeks despite an Australian card, phone and a years-old account, @leo114119 passes on advice to authenticate once with setup-token instead of logging in repeatedly, and @liyue_ai circulated an unban appeal template. @gengdaJ argues the information gap is the real story: among 1.4 billion people, the share who have actually used these tools is tiny.
📡 Eco Products Radar
Eco Products Radar
Codex (OpenAI): the second seat in almost every setup, now reachable the other way too since Claude Code can drive Codex's Computer Use locally without a Codex login
Opus 5.5: the default main model everywhere, with Sonnet 5.5 and Haiku 5.5 as the cheap tier in every cost thread
REA (Reverse Engineer Anything): the week's loudest repo, with the Touhou rebuild of 697 functions and 696 exact matches quoted in a dozen posts, plus Eleven as its desktop front end
Agent Vault (Self Labs): launched with roughly twenty near-identical posts about one vault for Claude Code, Codex, Grok Bot, OpenClaw, Hermes and Muse, with spending caps and a kill switch
Grok Bot (xAI): the commander layer people put above Claude Code, now reachable from inside a session through the Cursor plugin
Blender and Unity 6: the non-coding canvases of the day, with the open-sourced Scenario plugin driving an eight-role 3D team through MCP and two separate solo game builds
Jev (TypeSafe): the millisecond decision model wired under Opus in the routing templates, now also sold as jev-guard hooks and a lossless compaction plugin
CLI-Anything (HKUDS): 52,000 stars for wrapping desktop apps in command lines agents can drive, with a hub that installs GIMP, Blender, LibreOffice and Calibre
diagram-design: the 44-type diagram skill at 54,000 stars, recommended for making an agent draw its plan before it writes code
Muse Spark Open (Meta): 816B total and 42B active with a 1M context, landing on Ollama the same day and runnable from Claude Code in one command
OpenTelemetry: Claude Code, Managed Agents and the SDK all emit traces now, so token spend, latency and tool failures per subagent can go into Grafana
Ledger Agent Intent: physical confirmation on a device for agent-proposed transactions, integrated with OpenClaw on day one
Orca: the multi-agent workbench that runs Claude Code, Codex and OpenCode in one place with worktrees per task
Termius and Earpiece: two takes on being notified when an agent needs you, one on Android over OSC 7501 and one in the Mac notch
Codex (OpenAI): the second seat in almost every setup, now reachable the other way too since Claude Code can drive Codex's Computer Use locally without a Codex login
Opus 5.5: the default main model everywhere, with Sonnet 5.5 and Haiku 5.5 as the cheap tier in every cost thread
REA (Reverse Engineer Anything): the week's loudest repo, with the Touhou rebuild of 697 functions and 696 exact matches quoted in a dozen posts, plus Eleven as its desktop front end
Agent Vault (Self Labs): launched with roughly twenty near-identical posts about one vault for Claude Code, Codex, Grok Bot, OpenClaw, Hermes and Muse, with spending caps and a kill switch
Grok Bot (xAI): the commander layer people put above Claude Code, now reachable from inside a session through the Cursor plugin
Blender and Unity 6: the non-coding canvases of the day, with the open-sourced Scenario plugin driving an eight-role 3D team through MCP and two separate solo game builds
Jev (TypeSafe): the millisecond decision model wired under Opus in the routing templates, now also sold as jev-guard hooks and a lossless compaction plugin
CLI-Anything (HKUDS): 52,000 stars for wrapping desktop apps in command lines agents can drive, with a hub that installs GIMP, Blender, LibreOffice and Calibre
diagram-design: the 44-type diagram skill at 54,000 stars, recommended for making an agent draw its plan before it writes code
Muse Spark Open (Meta): 816B total and 42B active with a 1M context, landing on Ollama the same day and runnable from Claude Code in one command
OpenTelemetry: Claude Code, Managed Agents and the SDK all emit traces now, so token spend, latency and tool failures per subagent can go into Grafana
Ledger Agent Intent: physical confirmation on a device for agent-proposed transactions, integrated with OpenClaw on day one
Orca: the multi-agent workbench that runs Claude Code, Codex and OpenCode in one place with worktrees per task
Termius and Earpiece: two takes on being notified when an agent needs you, one on Android over OSC 7501 and one in the Mac notch
Comments