Summary
Recent reports highlight issues with the Codex app server's context compaction process, specifically focusing on two significant problems. The first, identified as issue 54897, involves a turn that remains in progress indefinitely due to a failure in the background compaction mechanism. This issue occurs when the background task closes prematurely, leading to a panic that disrupts the process without terminating the server. The recommended solution is to call turn slash interrupt on the stuck turn, allowing users to start a new turn without affecting other threads.
The second issue, issue 54822, reveals a potential data loss scenario when using the thread slash queue slash add function. If context compaction is interrupted, an accepted input can disappear from both the queue and the persisted history. The reporter notes that while the seed message remains intact, subsequent messages may vanish without a trace. Users are advised to check for their client user message ID in the thread slash turn slash list before resubmitting any work to avoid unnecessary duplication.
Additional minor issues include a problem with voice in the CLI disconnecting unexpectedly and guidance on using standalone installers effectively. Users are also reminded that GPT 5.5 has been retired from Codex, and they should update their configurations accordingly. The ongoing developments emphasize the importance of careful management of context compaction and input handling within the Codex framework.
Show Notes
A closed background-compaction channel could leave a Codex app-server turn stuck in progress; learn the reported recovery steps and the fix associated with PR 55581, whose release availability is still unclear. We also cover an open report of queued input disappearing when pre-turn compaction is interrupted, plus notes on CLI voice disconnects, installer download-host selection, and GPT 5.5 retirement.
More from The Codex Changelog
Codex 0.164.0 adds inline arguments to /model and /permissions, letting you switch models, reasoning effort, speed, and permission profiles without navigating pickers. Learn a practical workflow for saving those combinations in /presets, plus caveats around unsupported settings and resumed sessions.
Codex’s new PromptCacheKey policy lets hosts keep session identity separate from prompt-cache routing. Learn how to choose a shared key for parallel subagents, omit the field for compatible gateways, and reapply the policy when resuming threads—plus caveats for cache hits and prompt consistency. Also: updates to fork-prefix preservation, authentication errors, Guardian metrics, and compaction recovery.
Codex 0.164.0 alpha adds a startup check for feature-flag mismatches between the CLI and its background daemon, replacing a silent fallback to embedded mode with a recovery prompt. Learn how to configure feature overrides and reconnect safely, including the warning that restarting a shared daemon can interrupt active subagents; the episode also covers improvements to TUI question rendering, picker ranking, remote reconnects, and Linux launcher security.
Codex’s new PreModelRequest and PostModelRequest hooks expose model-call lifecycle data, including retries, status, and reported token usage. Learn how to register an async logging hook, account for cancelled turns, and check build availability, plus smaller security and workflow updates.
An open bug report for Codex 0.163.0 describes sessions stuck on “Waiting for startup” when an unused provider’s command-based auth table omits cwd and the daemon’s startup directory has been deleted. Learn how to spot the risky config, choose a durable explicit cwd, and recover a stuck session; the episode also covers a prerelease terminal-input stall, a WSL2 resume workaround, and the new /side command.