NotificationsYou must be signed in to change notification settings - Fork 16.2k
CLIIssues related to the Codex CLIIssues related to the Codex CLI
bugSomething isn't workingSomething isn't working
configIssues involving config.toml, config keys, config merging, or config updatesIssues involving config.toml, config keys, config merging, or config updates
windows-osIssues related to Codex on Windows systemsIssues related to Codex on Windows systems
Description #
I updated to 0.148.0 today and every single turn now fails with the below.
{
"error": {
"message": "prompt_cache_retention is not supported on this model",
"type": "invalid_request_error",
"param": "prompt_cache_retention",
"code": "invalid_parameter"
}
}
Codex is still sending prompt_cache_retention on gpt-5.6-sol. That model does not accept it anymore.
I checked my own rollout logs. Same machine, same model, same config, same auth. The only thing that changed is the version.
0.147.0 + gpt-5.6-sol = 10 sessions, 45 turns, 0 errors (Aug 14-18)
0.148.0 + gpt-5.6-sol = 1 session, 3 turns, 3 errors (Aug 19)
On the failing day my update check finished at 02:12Z, the first 0.148.0 session started at 02:14Z, and all 3 of its turns errored at 02:16Z, 02:34Z and 02:41Z.
To reproduce:
1- npm install -g @openai/codex@0.148.0
2- set model = "gpt-5.6-sol" in config.toml
3- send any prompt
Environment: Codex CLI 0.148.0 from npm, win32-x64, Windows 11 10.0.26200, Node v22.22.3, model_reasoning_effort xhigh, ChatGPT account auth and not an API key. Default ChatGPT backend, I did not test Azure or a custom model_provider.
Just to be clear this is NOT the 0.120.0 compaction bug (#17809 to #17819, all closed). Those said "Unknown parameter" and only happened during compaction. Mine says "is not supported on this model" and it happens on every normal turn. My first failure was at 154k tokens in a 258k window so there was no compaction involved.
One thing that makes this confusing to diagnose. When a turn has already started tool calls the work still finishes, but the final message never comes back, so it reads as stuck instead of failed. A turn with no tool calls dies in 4 seconds.
The fix is to gate prompt_cache_retention on the model instead of always sending it. Your own repo already says this in codex-rs/skills/src/assets/samples/openai-docs/references/upgrading-to-gpt-5p6-sol.md:
Treat the older
prompt_cache_retention
shape as deprecated
Same pattern as #31882 where use_responses_lite and multi_agent_version are hardcoded for sol/terra/luna. A shared capability table would close both.
Workaround for anyone else hitting this, downgrade with npm install -g @openai/codex@0.147.0. Switching to gpt-5.5 also works because that model still accepts the parameter.
0.149.0-alpha.1 was cut at 2026-08-19T00:54Z and I do not see a related commit on main, so I expect it is still broken.
Related: #31882, #25604, #35300.
Thank you
Metadata #
Metadata #
Assignees
Labels
CLIIssues related to the Codex CLIIssues related to the Codex CLI
bugSomething isn't workingSomething isn't working
configIssues involving config.toml, config keys, config merging, or config updatesIssues involving config.toml, config keys, config merging, or config updates
windows-osIssues related to Codex on Windows systemsIssues related to Codex on Windows systems