Problem
Codex v0.117.0 introduced a new `tui_app_server` feature (marked stable, default-on) that performs an `account/read` API call to `chatgpt.com/backend-api/account` during TUI bootstrap. GCP VM IPs are blocked by Cloudflare with a 403 challenge page, so this call fails immediately and Codex exits with code 1 — the TUI never starts.
Symptom: activity-monitor launches Codex in TUI mode inside tmux, Codex exits code 1 on every restart (crash loop). The bot is completely unavailable.
Confirmed on `cocoai-daidaijun2021-1b` running Codex v0.117.0.
Root cause evidence:
- `codex --dangerously-bypass-approvals-and-sandbox` → exit 1 in tmux (even with no prompt)
- `curl -H "Authorization: Bearer " https://chatgpt.com/backend-api/account\` → 403 Cloudflare JS challenge
- `codex --dangerously-bypass-approvals-and-sandbox --disable tui_app_server` → exit 0, TUI starts normally
- Adding `tui_app_server = false` under `[features]` in config.toml → exit 0, TUI starts normally
Expected Behavior
`writeCodexConfig()` in `cli/lib/runtime-setup.js` should include `tui_app_server = false` under `[features]` so all newly provisioned VMs are not affected by this Cloudflare block.
Required change (in `writeCodexConfig()`):
```toml
[features]
multi_agent = true
tui_app_server = false # ← add this line
```
Impact
Any VM provisioned with Codex ≥ v0.117.0 will enter a crash loop immediately after bootstrap. All existing VMs that upgrade to v0.117.0+ and regenerate config.toml via `zylos init` / `zylos runtime codex` will also be affected.
Immediate workaround applied to affected VMs: manual `sed` patch to config.toml.
Related
- This feature appears to have been introduced in Codex v0.117.0 — versions ≤ v0.116.0 did not exhibit this behavior
- The `tui_app_server` feature is listed as `stable` stage (default-on) in v0.117.0
- Zylos VMs run headless/non-interactive; the TUI app server feature provides no value and only introduces this network dependency
Problem
Codex v0.117.0 introduced a new `tui_app_server` feature (marked stable, default-on) that performs an `account/read` API call to `chatgpt.com/backend-api/account` during TUI bootstrap. GCP VM IPs are blocked by Cloudflare with a 403 challenge page, so this call fails immediately and Codex exits with code 1 — the TUI never starts.
Symptom: activity-monitor launches Codex in TUI mode inside tmux, Codex exits code 1 on every restart (crash loop). The bot is completely unavailable.
Confirmed on `cocoai-daidaijun2021-1b` running Codex v0.117.0.
Root cause evidence:
Expected Behavior
`writeCodexConfig()` in `cli/lib/runtime-setup.js` should include `tui_app_server = false` under `[features]` so all newly provisioned VMs are not affected by this Cloudflare block.
Required change (in `writeCodexConfig()`):
```toml
[features]
multi_agent = true
tui_app_server = false # ← add this line
```
Impact
Any VM provisioned with Codex ≥ v0.117.0 will enter a crash loop immediately after bootstrap. All existing VMs that upgrade to v0.117.0+ and regenerate config.toml via `zylos init` / `zylos runtime codex` will also be affected.
Immediate workaround applied to affected VMs: manual `sed` patch to config.toml.
Related