Steipete shards coding-agent sessions across about 5 Macs

• The local ceiling showed up as a full Codex task board; steipete's Codex screenshot says the Mac Studio was at about as many sessions as it could take.
• The setup is already distributed, with steipete on Jump Desktop saying he shards work over about five machines.
• The machine in the screenshot is a 512GB M3 Max, and steipete's hardware reply confirms the spec.
• The safety harness is autoreview plus testing, with steipete on autoreview and tests giving the short version and steipete's crabbox reply adding his autoreview skill plus crabbox.
• The throughput choice is desktop Codex on Max, since steipete on desktop says the CLI is no longer worth it and steipete on Max vs Ultra says Ultra can make sessions 2-3x longer.

A Jump Desktop support page says Jump can add multiple computers for remote access, which matches the five-machine setup. OpenAI's Auto-review docs say the feature routes eligible sandbox-boundary approvals to a separate reviewer agent, while its Codex pricing page says usage varies by model, context, reasoning, tool use, retrieval, and caching. The screenshot's actual work is telling: dead-export burn-down, CI cleanup, i18n pinning, Discord and Meet voice fixes, and web hardening, all bounded chores after steipete's crabbox reply says he defines the need up front.

steipete framed the screenshot as a hardware ceiling: that was about as many sessions as the Mac Studio could take.

The scaling trick was elsewhere: steipete on Jump Desktop said he shards work over about five machines. Jump Desktop's support page describes adding additional remote computers to a Jump setup.

A reader asked about the box, and steipete's hardware reply answered: 512GB M3 Max. That puts the workflow closer to a local agent rack than a single IDE session.

The OpenClaw queue in steipete's Codex screenshot breaks into maintenance buckets:

• Dead exports: plugin exports, agent exports, UI exports, SDK exports.
• Dependency cleanup: remove dotenv and glob, adopt p-limit and p-map, adopt execa.
• TypeScript cleanup: refactor hot TS files, canonicalize Result types.
• Product hardening: harden OpenClaw Web, speed up web UI, keep reconnect errors visible.
• Voice work: review xAI voice and TTS, triage Discord voice PRs, fix Discord and Meet voice issues.
• Release plumbing: fix PR land, fix CI and speed tests, audit env vars, fix the stale Apple i18n pin.

Asked why he could push so much at once, steipete said refactors are easier. One item had cleanup behind it: steipete's CI reply says a session already covered the work, but the CI around it had broken.

His compressed safety recipe was two parts: autoreview and clever testing.

OpenAI's Auto-review docs define that layer narrowly: the main agent stays in the same sandbox, and a separate reviewer agent handles eligible approval requests. steipete's longer explanation adds the harness detail for this run: his crabbox reply says he defines the need first, then relies on an autoreview skill plus crabbox to catch issues.

LLMJunky pushed the same pattern in two replies. LLMJunky's auto-review reply called auto review the same experience but safer, and LLMJunky's backup reply paired auto review with regular backups.

His answer to “desktop or CLI” was one word before the verdict: desktop. The CLI, he said, is not worth it anymore.

OpenAI's Codex changelog says Codex is now part of the ChatGPT desktop app on macOS and Windows, with PR review in the sidebar and Codex Remote for connected hosts. That explains a second comment in the thread: steipete on the second layer said the app can move things around Spaces and handle menu-item work that other layers cannot.

steipete did not pick the slowest reasoning setting for the pile. He told one reader the run was on Max rather than Ultra because Ultra often makes sessions take 2-3x longer.

Another reply cut it shorter: steipete on Ultra called Ultra so slow that it is rarely worth it.

Quota was in the air too: steipete's limit reply asked whether the limit had been raised. OpenAI's Codex pricing page says local messages, cloud tasks, and code reviews are counted in five-hour windows, with usage varying by model, context, reasoning, tool use, retrieval, and caching.