/codex:adversarial-review fails with default gpt-5.5 on codex-cli 0.125.0, but succeeds with --model gpt-5.4

[mudsg24]

Summary

/codex:adversarial-review fails when the plugin inherits a top-level Codex config default of model = "gpt-5.5", even though the local Codex CLI is already on the latest release I can install (codex-cli 0.125.0). The same adversarial review succeeds when I explicitly pass --model gpt-5.4.

This makes the error message misleading: it says GPT-5.5 requires a newer Codex CLI, but the CLI is already current, and the failure appears specific to the plugin's adversarial-review structured-output path rather than general GPT-5.5 access.

Environment

• OS: macOS 26.4.1 / Darwin 25.4.0 arm64
• Node: v25.8.0
• Codex CLI: codex-cli 0.125.0
• npm global package: @openai/codex@0.125.0
• codex-plugin-cc: 1.0.4
• Auth: ChatGPT login
• Claude Code plugin command: /codex:adversarial-review

Relevant ~/.codex/config.toml settings:

model = "gpt-5.5"
model_context_window = 1000000
model_auto_compact_token_limit = 800000
model_reasoning_effort = "xhigh"

Reproduction

1. Configure Codex with top-level default model gpt-5.5.
2. In Claude Code, run:

/codex:adversarial-review --wait

3. The review fails with the following error captured in the plugin job log / result:

{"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model requires a newer version of Codex. Please upgrade to the latest app or CLI and try again."}}

The rendered result is:

# Codex Adversarial Review

Codex did not return valid structured JSON.

- Parse error: {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.5' model requires a newer version of Codex. Please upgrade to the latest app or CLI and try again."}}

Control checks

The installed CLI is current:

codex --version
# codex-cli 0.125.0

npm list -g --depth=0 | rg '@openai/codex|codex'
# @openai/codex@0.125.0

A plain plugin task with the default gpt-5.5 model succeeds from the same plugin runtime:

node ~/.claude/plugins/cache/openai-codex/codex/1.0.4/scripts/codex-companion.mjs task --json "Return exactly OK."

Result:

{
  "status": 0,
  "rawOutput": "OK",
  "touchedFiles": [],
  "reasoningSummary": []
}

The same adversarial review succeeds when the model is explicitly downgraded:

/codex:adversarial-review --model gpt-5.4 --wait

Result from the plugin runtime included valid structured JSON and parseError: null.

Expected behavior

One of:

1. /codex:adversarial-review should work with inherited model = "gpt-5.5" when the installed Codex CLI is 0.125.0, or
2. the plugin should surface a clearer diagnostic if GPT-5.5 is not supported for this specific turn/start + outputSchema / structured JSON path, or
3. the plugin should document or apply a temporary workaround such as --model gpt-5.4 for adversarial review until the upstream app-server path supports GPT-5.5 fully.

Actual behavior

The plugin reports a structured-output parse failure whose underlying 400 says:

The 'gpt-5.5' model requires a newer version of Codex. Please upgrade to the latest app or CLI and try again.

This is misleading because codex-cli 0.125.0 is already installed and basic GPT-5.5 plugin tasks work.

Suspected scope

This looks specific to the adversarial-review path, likely the plugin's runAppServerTurn(... outputSchema: readOutputSchema(REVIEW_SCHEMA) ...) flow, not general GPT-5.5 model availability.

Related but not exact duplicates:

• Remote SSH Desktop reuses stale app-server and masks model_not_found as Reconnecting codex#19300: stale app-server / GPT-5.5 model catalog mismatch in Desktop remote SSH
• [P0 / Blocker] Remote compact task fails 100% with "tools.defer_loading requires tools.tool_search" — GPT-5.5 unusable at context limit, no client-side workaround codex#19486: GPT-5.5 app/server-side schema/tool request issue during compaction
• GPT-5.5 Codex session hits unrecoverable compaction failure around ~220k tokens despite advertised 400k context codex#19386: GPT-5.5 compaction failure in long-running Codex sessions
• feat: allow selecting a Codex profile for companion-launched jobs #251: plugin-launched jobs currently inherit global config and do not support selecting a Codex profile

[mudsg24]

Follow-up after more testing:

After fully restarting Claude Code and reloading the conversation, /codex:adversarial-review no longer fails immediately with the GPT-5.5 "requires a newer version of Codex" error.

The successful run was launched through the same plugin cache version (codex-plugin-cc 1.0.4), but this time against a branch/base diff rather than the default working-tree fallback:

/codex:adversarial-review

Claude Code ultimately invoked:

node ~/.claude/plugins/cache/openai-codex/codex/1.0.4/scripts/codex-companion.mjs adversarial-review "--base 7393069"

The job started in the background, completed successfully, and did not hit the earlier GPT-5.5 version error.

This makes the issue look more like a stale app-server / stale plugin runtime / stale model capability state problem than a permanently unsupported GPT-5.5 path. The original failure still seems worth tracking because:

1. codex-cli 0.125.0 was already installed when the error occurred.
2. The error told the user to upgrade Codex, which was not actionable in this environment.
3. Restarting Claude Code / reloading the conversation appears to be a viable recovery step, but the plugin did not surface that possibility.
4. If an old app-server/runtime is reused after a model rollout or CLI update, the plugin probably needs either a compatibility check or a clearer diagnostic.

So the refined report is: GPT-5.5 did work after runtime refresh, but the plugin can get into a stale state where adversarial-review returns a misleading "requires a newer version of Codex" error even though the installed CLI is current.

[tornikegomareli]

Confirming on a different environment (macOS 26.1 arm64 / Node 25.2.1 / codex-cli 0.125.0 / @openai/codex@0.125.0 / ChatGPT login). Same gpt-5.5 requires a newer version of Codex error, three additional data points worth flagging:

1. /codex:review has no --model escape hatch.

The --model gpt-5.4 workaround only works for adversarial-review and task. The built-in review subcommand doesn't accept it:

node ~/.claude/plugins/cache/openai-codex/codex/1.0.4/scripts/codex-companion.mjs review --help
# `/codex:review` now maps directly to the built-in reviewer ...

I am hitting this on /codex:review first have no flag-level workaround, I have to know to run /codex:adversarial-review --model gpt-5.4 instead.

2. Killing the shared-session broker is not sufficient.

After kill <app-server pid> and rm -rf /var/folders/*/T/cxc-* (broker socket dirs), the next companion call still produced the same error. So whatever cached state triggers the gpt-5.5 request lives somewhere other than the local broker, possibly persisted plugin job metadata or a server-side capability claim. The earlier comment about full Claude Code restart helping is consistent with that it's plugin-runtime state, not broker state.