What version of the Codex App are you using (From “About Codex” dialog)?
26.803.61601
What subscription do you have?
Pro 5x
What platform is your computer?
Darwin 24.6.0 arm64 arm
What issue are you seeing?
I am experiencing a recurring issue with GPT-5.6 Sol in the Codex App.
Model: GPT-5.6 Sol
Reasoning effort: High
Codex App version: 26.803.61601
The model initially works normally. It can analyze the task, generate responses, and begin working on the project successfully.
However, the error frequently occurs after Codex reaches a Run Command stage.
A typical sequence is:
- I send a task to GPT-5.6 Sol High.
- The model starts working normally and returns some content.
- Codex invokes Run Command to execute a command.
- During or after this stage, the following error appears:
prompt_cache_retention is not supported on this model
- The agent run stops and cannot continue the task.
This does not necessarily happen on the first model request. The initial response usually succeeds. The failure frequently occurs later in the agent loop around Run Command/tool execution.
The issue occurs often enough that GPT-5.6 Sol High becomes unreliable for tasks involving command execution.
Expected behavior
After Run Command finishes, GPT-5.6 Sol should receive the command result and continue working on the task normally.
Actual behavior
Codex frequently reports:
prompt_cache_retention is not supported on this model
The current agent run is then interrupted and cannot continue.
Because the initial model request works correctly and the failure tends to occur during a later stage of the agent loop, I suspect that prompt_cache_retention may be added to a subsequent request, such as a tool-result continuation, context caching, or context compaction request.
Please check whether Codex App 26.803.61601 is sending prompt_cache_retention to GPT-5.6 Sol during subsequent requests in the Run Command/tool execution loop.
What steps can reproduce the bug?
- Open Codex App 26.803.61601.
- Select GPT-5.6 Sol/Luna/Terra with any reasoning effort.
- Start a normal coding task that requires the agent to use tools and execute shell commands.
- Let the agent work normally for some time. The model initially responds correctly and can successfully perform multiple mixed tool calls and Run Command operations.
- After a period of normal agent activity involving multiple tool/command calls, the following error frequently appears:
prompt_cache_retention is not supported on this model
- Once the error appears, the current agent run stops and cannot continue.
There does not appear to be one specific shell command that triggers the problem. It can occur during otherwise normal coding tasks after the agent has been working for some time and has performed a mixture of tool calls and command executions.
The problem is frequent, but the exact Run Command/tool call on which it occurs is not deterministic.
In my experience, simply giving GPT-5.6 Sol High a coding task that requires continued tool and command usage is often sufficient to eventually reproduce the issue.
I do not currently have a minimal code snippet that deterministically triggers the bug, because it appears to be related to the ongoing agent/tool execution loop rather than the contents of a particular command.
What is the expected behavior?
GPT-5.6 should be able to continue the agent loop normally after executing commands and using tools.
After a Run Command or other tool call completes, Codex should pass the result back to the model and allow it to continue reasoning, executing additional commands, and completing the task without failing because of an unsupported internal caching parameter.
Long-running coding tasks involving multiple mixed tool calls and command executions should continue normally until the task is completed or the user stops the run.
Additional information
No response
What version of the Codex App are you using (From “About Codex” dialog)?
26.803.61601
What subscription do you have?
Pro 5x
What platform is your computer?
Darwin 24.6.0 arm64 arm
What issue are you seeing?
I am experiencing a recurring issue with GPT-5.6 Sol in the Codex App.
Model: GPT-5.6 Sol
Reasoning effort: High
Codex App version: 26.803.61601
The model initially works normally. It can analyze the task, generate responses, and begin working on the project successfully.
However, the error frequently occurs after Codex reaches a Run Command stage.
A typical sequence is:
prompt_cache_retention is not supported on this modelThis does not necessarily happen on the first model request. The initial response usually succeeds. The failure frequently occurs later in the agent loop around Run Command/tool execution.
The issue occurs often enough that GPT-5.6 Sol High becomes unreliable for tasks involving command execution.
Expected behavior
After Run Command finishes, GPT-5.6 Sol should receive the command result and continue working on the task normally.
Actual behavior
Codex frequently reports:
prompt_cache_retention is not supported on this modelThe current agent run is then interrupted and cannot continue.
Because the initial model request works correctly and the failure tends to occur during a later stage of the agent loop, I suspect that
prompt_cache_retentionmay be added to a subsequent request, such as a tool-result continuation, context caching, or context compaction request.Please check whether Codex App 26.803.61601 is sending
prompt_cache_retentionto GPT-5.6 Sol during subsequent requests in the Run Command/tool execution loop.What steps can reproduce the bug?
prompt_cache_retention is not supported on this modelThere does not appear to be one specific shell command that triggers the problem. It can occur during otherwise normal coding tasks after the agent has been working for some time and has performed a mixture of tool calls and command executions.
The problem is frequent, but the exact Run Command/tool call on which it occurs is not deterministic.
In my experience, simply giving GPT-5.6 Sol High a coding task that requires continued tool and command usage is often sufficient to eventually reproduce the issue.
I do not currently have a minimal code snippet that deterministically triggers the bug, because it appears to be related to the ongoing agent/tool execution loop rather than the contents of a particular command.
What is the expected behavior?
GPT-5.6 should be able to continue the agent loop normally after executing commands and using tools.
After a Run Command or other tool call completes, Codex should pass the result back to the model and allow it to continue reasoning, executing additional commands, and completing the task without failing because of an unsupported internal caching parameter.
Long-running coding tasks involving multiple mixed tool calls and command executions should continue normally until the task is completed or the user stops the run.
Additional information
No response