6-Hour Codex Run That Survived a 5-Hour Pause

Codex CLI v0.128.0 was released on April 30, 2026, introducing /goal as a core feature. It persists goal state on the app server, keeping sessions alive through terminal restarts, laptop sleep, and gaps of several hours, and when you return, runtime continuation injects a developer message so work continues without the user having to type anything again.

The new release included persisted goals, app-server APIs, model tools, runtime continuation, and TUI controls for create/pause/resume/clear. The same version added scrollback reflow, codex update, plan mode prompting, keymap configuration, and expanded approval profiles, while --full-auto was deprecated. Plan mode was first added in v0.122.0 (April 20, 2026).

The actual verification took place in a voice interview system within a TypeScript monorepo. With approval_policy set to never and sandbox_mode set to danger-full-access, a /goal prompt of about 600 words was used, including an XML-style block, a list of 10+ files to read first, four concrete success criteria via a done_when contract, and anti-pattern blocking language. Per-session reasoning is adjusted via model_reasoning_effort, and automatic context compaction via model_auto_compact_token_limit.

• The model was gpt-5.5, with reasoning effort set to high.
• It started at 9:19 PM, and the first turn was turn_aborted after 57 seconds.
• After returning about 5 hours 30 minutes later, it automatically resumed.
• Cumulative input was about 6.8 million tokens, output about 10,000 tokens, reasoning tokens about 2,600, and the cache hit rate was about 94%.
• Wall-clock time was 6 hours 44 minutes, actual model compute was about 41 minutes, and the final state was TASK_COMPLETE.

A single context compaction occurred at around 6.7 million tokens. Thanks to the 94% cache hit rate, the actual incremental cost was much lower, and TTS time-to-first-byte remained null because the upstream library did not emit that runtime event. Manual review found no prompt loop, liveness spiral, or premature close, and all 4 end-to-end voice scenarios passed.

The comparison target, Ralph Wiggum Loop, uses Git history and an external shell loop to write a fresh context each time, whereas /goal provides continuity within the session through app-server state and automatic resumption. However, it isn't suited for exploratory tasks with unclear success criteria, security-sensitive paths, tasks with uncertain external dependencies, or short tasks that finish within 10 minutes — the key point is handling autonomous execution through upfront contracts rather than supervision.