0.120.0 regression: resumed threads fail during remote compaction with "Unknown parameter: 'prompt_cache_retention'"

After upgrading Codex CLI to 0.120.0, resumed long-running threads started failing when context compaction kicked in. Rolling back to 0.119.0 immediately resolved the problem in the same environment.

Environment:

• macOS
• Codex installed globally via npm
• version that fails: 0.120.0
• version that works: 0.119.0
• client environment: Qodex Desktop, but the root error is visible in the Codex app-server/chat log

Steps to reproduce:

1. Install @openai/codex@0.120.0
2. Resume an existing thread with enough history to trigger context compaction
3. Let the turn continue until compaction runs
4. Observe the turn failure

Actual result:

• the resumed turn fails during compaction
• log shows:

Error running remote compact task: {
  "error": {
    "message": "Unknown parameter: 'prompt_cache_retention'.",
    "type": "invalid_request_error",
    "param": "prompt_cache_retention",
    "code": "unknown_parameter"
  }
}

Related runtime lines also include:

remote compaction failed
Failed to run pre-sampling compact

Expected result:

• the thread should compact successfully and continue responding

Regression evidence:

• downgrading to @openai/codex@0.119.0 in the same environment immediately fixed the issue
• after rollback, the same session compacted context and continued responding normally

Rollback command:

npm install -g @openai/codex@0.119.0

Notes:

• This looks like a 0.120.0 regression in the remote compaction path for resumed threads.
• In my app, a separate UI bug initially made the error appear as just }, but that was only a formatting issue. The real runtime error is the prompt_cache_retention rejection above.