CLI remote compaction loops with 400 invalid_enum for `context_compaction`, consuming quota across sessions

What version of Codex CLI is running?

codex-cli 0.130.0

What subscription do you have?

Team. I also observed the same failure across a second Personal Pro account.

Which model were you using?

gpt-5.5 with xhigh reasoning

What platform is your computer?

Darwin 25.4.0 arm64 arm

What terminal emulator and version are you using (if applicable)?

Ghostty, zsh, TERM=xterm-256color

Codex doctor report

not available {

{
  "available": false,
  "codex_version": "codex-cli 0.130.0",
  "fresh_session": true,
  "codex_update": "Warning: Not upgrading codex, the latest version is already installed. Update ran successfully! Please restart Codex.",
  "help_doctor": "error: unrecognized subcommand 'doctor'",
  "doctor_json": "error: unexpected argument '--json' found",
  "doctor": "Error: stdin is not a terminal"
}

What issue are you seeing?

Remote compaction repeatedly fails because Codex sends or injects an item type named `context_compaction`, but the backend rejects it. The backend error says accepted values include `compaction` and `compaction_trigger`, but not `context_compaction`.

This affected multiple threads/windows, including at least four separate Codex windows, across two different accounts. I am not including account/user IDs in this public issue, but OpenAI should be able to correlate the uploaded feedback thread ID or contact me directly for further diagnostics. 

019e3003-2bda-77e2-a709-d641e65ff317

This materially consumed quota/usage while sessions repeatedly failed.

epresentative error:

Error running remote compact task: {
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "[ItemParam] [input[226]] [invalid_enum_value] Invalid value: 'context_compaction'. Supported values are: 'apply_patch_call', 'apply_patch_call_output', 'code_interpreter_call', 'compaction', 'compaction_trigger', 'computer_call', 'computer_call_output', 'custom_tool_call', 'custom_tool_call_output', 'file_search_call', 'function_call', 'function_call_output', 'image_generation_call', 'item_reference', 'local_shell_call', 'local_shell_call_output', 'mcp_approval_request', 'mcp_approval_response', 'mcp_call', 'mcp_list_tools', 'message', 'reasoning', 'shell_call', 'shell_call_output', 'tool_search_call', 'tool_search_output', and 'web_search_call'."
  },
  "status": 400
}

I also saw repeated websocket failures on the same remote compaction path:

Error running remote compact task: unexpected status 403 Forbidden: Unknown error,
url: wss://chatgpt.com/backend-api/codex/responses

What steps can reproduce the bug?

Uploaded feedback thread:
019e3003-2bda-77e2-a709-d641e65ff317

1. Run Codex CLI 0.130.0 using gpt-5.5 with a long-running session.
2. Let the session grow until automatic remote compaction is triggered.
3. Observe remote compaction failing with invalid enum `context_compaction`.
4. Continue or resume the session; the failure can repeat and continue consuming quota.

Affected local session/thread IDs observed in local logs:

- 019e2e07-10b8-7b72-9979-8ee724045622
  - 35,588 line-level codex_core::compact_remote failure log entries
  - first observed 2026-05-16T02:39:31Z; last observed 2026-05-16T08:58:41Z
  - transcript: codex/sessions/2026/05/16/rollout-2026-05-16T00-44-35-019e2e07-10b8-7b72-9979-8ee724045622.jsonl

- 019e2e85-62c8-7ce1-b0e3-09297bb7e458
  - 13,588 line-level codex_core::compact_remote failure log entries, including 13,581 403 Forbidden failures
  - first observed 2026-05-16T02:20:02Z; last observed 2026-05-16T08:59:42Z
  - transcript: codex/sessions/2026/05/16/rollout-2026-05-16T03-02-34-019e2e85-62c8-7ce1-b0e3-09297bb7e458.jsonl

- 019e2e83-f8d7-7b83-832b-dce6b01ff340
  - 12,555 line-level codex_core::compact_remote failure log entries, including 12,547 403 Forbidden failures
  - first observed 2026-05-16T02:23:47Z; last observed 2026-05-16T08:59:42Z
  - transcript: codex/sessions/2026/05/16/rollout-2026-05-16T03-01-01-019e2e83-f8d7-7b83-832b-dce6b01ff340.jsonl

- 019e2e55-2d10-76f3-8b01-a65314d16bca
  - 14 line-level codex_core::compact_remote failure log entries, all 403 Forbidden failures
  - first observed 2026-05-16T01:52:01Z; last observed 2026-05-16T01:55:03Z
  - transcript: codex/sessions/2026/05/16/rollout-2026-05-16T02-09-54-019e2e55-2d10-76f3-8b01-a65314d16bca.jsonl

This affected multiple threads/windows, including at least four separate Codex windows, across two different accounts.

What is the expected behavior?

Remote compaction should only send item types accepted by the backend schema, such as `compaction` or `compaction_trigger`.

If compaction fails, Codex should stop retrying in a way that consumes quota and should provide a safe recovery path.

Additional information

Isearched existing openai/codex issues before filing. There are many broad remote-compaction issues, but I found no exact match for the context_compaction invalid enum mismatch.

Closest related issues:

• https://github.com/openai/codex/issues/22335
• https://github.com/openai/codex/issues/23003
• https://github.com/openai/codex/issues/14860

Local mitigation applied:
remote_compaction_v2 = false

This avoids the failing path in new sessions, but already-running sessions do not reload that config in place.

The local Codex SQLite/log evidence contained 58 rows matching:
Invalid value: 'context_compaction'

Experimental feature analysis:

• Direct suspect: remote_compaction_v2 was enabled during the incident and is the feature that selects the failing remote compaction path. It is now disabled locally as a mitigation.
• Adjacent suspect: responses_websocket_response_processed is enabled and under development. The failing path is wss://chatgpt.com/backend-api/codex/responses, and the 0.130.0 binary contains both compaction and context_compaction response item variants.
• Amplifiers rather than root causes: goals, multi_agent_v2, enable_fanout, and prevent_idle_sleep can keep long-running sessions active enough to trigger compaction repeatedly, but they should not create an invalid backend enum by themselves.
• Probably unrelated to this specific enum mismatch: MCP/app flags such as enable_mcp_apps, builtin_mcp, apps_mcp_path_override, remote_plugin, plugin_hooks, and tool_search_always_defer_mcp_tools.

Please investigate whether this is a Codex CLI/backend schema mismatch and restore or credit quota consumed by repeated failed remote compaction attempts.