Astra in Codex, 872k Context Hack, Orca Setups — AI Daily Sep 5

590 messages · 77 active members

590
messages
77
active members
@anonymous, @jasonakatiff, @jcartu
top contributors

Overview

Saturday was Astra's day. The model started appearing inside Codex for $200-tier ChatGPT users (logout/login required) and builders immediately called it a step-function upgrade over Fable 5.1. Jason ran Astra across a large repo in ~2 hours plus a second pass, then had Grok double-check the output and found essentially nothing to add — setting the tone for a lot of enthusiasm around deep code reviews, integrations, and bug squashing. Weekend ships included a Mario clone and a slot machine POC, and Jason opened builtwithai.com/projects for members to publish repos, decks, and talks. Infra and config talk ran alongside the rollout. @Anonymoushat shared an override (model_context_window = 872000, auto_compact at 800k) that pushes Astra's default 272k window to 872k on the Codex desktop app, though Codex cloud still defaults to Sol and usage burns fast past the soft cap. Members compared Orca setups on Hetzner/Hostinger VPS and Mac minis controlled from MacBook/iPhone over Tailscale, with @ferchonaso promising a full tutorial and @jcartu and @samb69 preferring simpler harnesses like Paseo or plain tmux. Orchestrator stacks were debated too: Codex/GPT-5.5 for Hermes agents, Gemini or Grok for app orchestration, and ChatGPT Pro's normal chat with MCP/GitHub for the deepest planning work. A parallel marketing thread argued AI video (Seedance 2.5 and similar) is ~100x cheaper than $500 influencer videos, with one builder sharing a system that scores influencer profiles by predicted post value plus a TikTok organic-then-boost playbook. A long hardware sub-thread pitted ThinkPad loyalists against recent Mac converts, with @tounano reporting a 3-4 week dip before surpassing his old Windows baseline thanks to Claude-assisted customization.

Topics

Astra Rolls Out in Codex with Deep Code Reviews55 msgs

Astra appeared inside Codex for $200-plan users (logout/login to see it, some regions delayed) and was called a clear jump above Fable 5.1. Jason ran a full repo review in ~2 hours plus a second pass and had Grok verify — finding nothing to add. Users report it excels at integrations, bug squashing, and end-to-end reviews with minimal rework.

Astra 272k → 872k Context Window Override8 msgs

@Anonymoushat shared a config hack (model_context_window = 872000, model_auto_compact_token_limit = 800000) to push Astra past its default 272k window without paying for the 1M API tier. Confirmed working on the Codex desktop app but not the web/cloud version, which still defaults to Sol. @jcartu cautioned that usage burns fast past the soft cap.

Orca Remote Setups on VPS and Mac Minis20 msgs

Members compared running Orca on Hetzner or Hostinger VPS and Mac minis, controlling from MacBook and iPhone via Tailscale. @iannagy reported no meaningful lag on his Hetzner setup, and @ferchonaso offered a full Mac-mini-based tutorial. @jcartu and @samb69 pushed back on Orca's complexity and preferred simpler harnesses like Paseo or plain tmux.

Orchestrator and Hermes Model Picks20 msgs

Builders shared stacks: Codex/GPT-5.5 for Hermes agents, Gemini or Grok for app orchestration, and Orca with remote servers to fan agents across old Macs. For the deepest architecture and planning work, ChatGPT Pro's normal chat (not Codex CLI) with MCP/GitHub was recommended — separate quotas too.

AI Video vs Paid Influencer Economics18 msgs

Members argued Seedance 2.5 and similar models are ~100x cheaper than $500 influencer videos with none of the ego or turnaround pain. One builder shared a system that scores influencer profile metrics to predict post value and baseline fair fees, plus a TikTok playbook: let influencers post organically, then only pay ad codes to boost what the algo already likes.

Words worth knowing

Three terms from the glossary.

New from members

Just connected to what we're building.

  • theleadrouter-ad-studio

    theLeadRouter — Ad Studio v2. Research, create, launch, and manage ad campaigns. Self-host with the Railway one-click installer.

    Pythonstars: 53@jasonakatiff
  • trivium

    A Node Framework is based on the concept that routes, web sockets and chronological events should be trivial.

    JavaScriptstars: 3@geilt

Key Takeaways

  • Astra is rolling out in Codex for $200-tier ChatGPT plans — log out and back in to see it; desktop app only, cloud still defaults to Sol.
  • Override Astra's default 272k window with model_context_window = 872000 and auto_compact at 800k, but watch usage past the soft cap.
  • For deepest planning, use ChatGPT Pro's normal chat with deep research + MCP/GitHub, not Codex CLI — you also get separate quotas.
  • Run Orca on a VPS (Hetzner/Hostinger) or Mac mini and control from MacBook/iPhone over Tailscale for a low-lag remote dev setup.
  • Score influencers on historical metrics before paying; on TikTok, let them post organically and only boost the videos the algo already likes.

Hot Threads

@jasonakatiffstarted

Astra code review found nothing after Grok double-check

18 replies8 participants
@anonymousstarted

ThinkPad vs Mac vs Framework for AI dev machines

30 replies6 participants
@thewildzenostarted

Automating unboxing/product videos and influencer economics

20 replies9 participants

Linked Items

@levelsio on X

x.com
@anonymous

x.com

x.com
@fmill1

beginbot 🃏 on X

x.com
@anonymous

evnsnclr on X

x.com
@Coybh

Genius Thinking on X

x.com
@anonymous