Skip to content

ClawBox v3.1.11 — Codex on ChatGPT accounts

Choose a tag to compare

View all tags
@yalexx yalexx released this 28 Jul 18:37
· 301 commits to main since this release
v3.1.11
8cb6a43
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Fixes Codex for every ClawBox signed in with a ChatGPT account, and stops the update path from leaving a box in a broken state.

Codex

  • Route codex turns through the app-server harness (#280). Without agentRuntime, core posted to /backend-api/responses — a browser endpoint Cloudflare challenges — so every turn died with "the provider returned an HTML error page".
  • Migrate ChatGPT logins into the store the gateway reads (#279). Logins landed in the legacy auth-profiles.json while core reads sqlite, so a box could show the provider connected while every turn 401'd with profile=-.
  • Stop mirroring the rotating refresh token into every codex-home (#278). A single-use refresh token could be spent twice and burned.
  • Default ChatGPT auth to gpt-5.5 — available on every tier — and surface GPT-5.6 Sol/Terra/Luna for entitled accounts (#275, #276).

Gateway and update

  • Update and boot can no longer brick the box; health and recovery waits are parameterised.

Other

  • Semantic memory works without an API key via local embeddings.
  • Docs and site references point at clawbox.com; Wi-Fi and local-model performance figures replaced with measured values.

How to get it

Open the ClawBox UI and take the update. Boxes that can't update through the UI can self-heal with:

bash <(curl -fsSL https://raw.githubusercontent.com/id-robots/clawbox/main/scripts/force-update.sh)

Verification

Tested on two boxes on 2026-07-28: the HTML/Cloudflare failure and the 401 profile=- failure were each reproduced on the old build and fixed by the shipped updater; forced OAuth token rotation survived twice on each box; both 3.1.10 → 3.1.11 and 3.1.5 → 3.1.11 upgrades passed end to end including a full reboot (~4 minutes). CI: 1499 unit tests, E2E, and E2E-install in a real systemd container.

Known issue — not a regression

The Codex model picker offers gpt-5.4 / gpt-5.4-mini, which free-tier ChatGPT accounts cannot run (upstream returns a 400 with no failover). This is present in 3.1.10 as well. The default is gpt-5.5, which every tier runs, so it only affects someone who picks an older model by hand. Fix queued for 3.1.12.

Assets 2
Loading