Every Codex turn fails on gpt-5.6 Sol after 0.148.0

Codex CLI 0.148.0 fails on every turn with gpt-5.6-sol because the client still sends prompt_cache_retention, a parameter the model rejects: the first session on the new version logged 3 turns and 3 errors. The bug report was filed against the Codex repository on Github.

At a glance

• The failure surfaces as an invalid_request_error with code invalid_parameter and the message that prompt_cache_retention is not supported on this model, returned for the param of the same name.
• Rollout logs from the same machine, model, config and auth show 0.147.0 completing 45 turns across 10 sessions with no errors between Aug 14 and Aug 18.
• Downgrading to 0.147.0 restores the model, and gpt-5.5 still accepts the parameter, so the report frames the fix as gating prompt_cache_retention per model instead of always sending it.

Codex ships parameters on behalf of whichever model the config names, and a mismatch between client defaults and a model's accepted parameter set breaks every request rather than degrading gracefully. The failure mode described here is the awkward kind: turns that already invoked tools appear to hang, which likely costs more triage time than a clean error would. It also reads as a maintenance question about how the CLI tracks per-model capabilities.

All three turns of the first 0.148.0 session failed between 02:16Z and 02:41Z

According to the report, the update check finished at 02:12Z, the first 0.148.0 session started two minutes later, and all three of its turns errored at 02:16Z, 02:34Z and 02:41Z. The machine, model, configuration and authentication were unchanged from the previous days.

Reproduction is three steps: install @openai/[email protected] globally from npm, set model to gpt-5.6-sol in config.toml, and send any prompt. The environment is win32-x64 on Windows 11 10.0.26200 with Node v22.22.3, model_reasoning_effort set to xhigh, and ChatGPT account authentication rather than an API key on the default ChatGPT backend.

The only variable the report identifies is the version number: 0.147.0 ran 10 sessions on gpt-5.6-sol without a single failure, while the one session recorded on 0.148.0 failed on all of the turns it attempted.

Turns without tool calls die in four seconds, while turns with tool calls read as stuck

When a turn has already started tool calls, the work still completes but the final message never returns, so the session appears to hang rather than fail. A turn with no tool calls dies in four seconds. The report flags this split as the main obstacle to diagnosis.

The report separates this from the 0.120.0 compaction bug tracked in issues #17809 to #17819, all now closed. Those returned "Unknown parameter" and fired only during compaction, while the first failure here landed at 154k tokens inside a 258k window, with no compaction involved. The new error string also differs, saying the parameter is not supported on this model, and it appears on every normal turn.

Codex's own upgrade doc calls the older prompt_cache_retention shape deprecated

The requested fix is to gate prompt_cache_retention on the model rather than sending it unconditionally with every request. The report cites a file already in the Codex repository, codex-rs/skills/src/assets/samples/openai-docs/references/upgrading-to-gpt-5p6-sol.md, which addresses the same parameter in its guidance for the newer model:

Treat the older prompt_cache_retention shape as deprecated

The same report links the problem to issue #31882, where use_responses_lite and multi_agent_version are hardcoded for the sol, terra and luna models, and argues that a shared capability table would close both cases. Issues #25604 and #35300 are listed as related.

The workarounds listed in the report are a downgrade to 0.147.0 through npm, or switching the configured model to gpt-5.5, which still accepts prompt_cache_retention. Neither changes the client behaviour that sends the parameter, and both keep the reported setup otherwise unchanged.

Whether 0.149.0 carries the gate Version 0.149.0-alpha.1 was cut at 2026-08-19T00:54Z, and the report says no related commit is visible on main, so the author expects the alpha to carry the same behaviour. No fix timeline is attached to the issue. Testing was limited to the default ChatGPT backend, with Azure and custom model_provider setups left unchecked, so the scope beyond that configuration is unstated.