feat(cx): add reasoning effort variants for GPT-5.5 and GPT-5.4

[mugnimaestra]

Summary

• Adds a withCodexEffortVariants() helper function (analogous to withCodexReviewModels()) that auto-generates -xhigh, -high, -low, -none model entries for specified Codex models
• Applies it to gpt-5.5 and gpt-5.4 in the cx provider, generating 8 new model entries (4 effort levels × 2 models)
• No changes to codex.js — the executor already handles suffix stripping generically (line 184-192)

Motivation

Consumer tools like opencode use @ai-sdk/openai-compatible (Vercel AI SDK), which validates reasoningEffort against z.enum(["low", "medium", "high"]). This means none and xhigh are rejected by the SDK before the request even reaches 9router.

The model name suffix approach (e.g. gpt-5.5-xhigh) bypasses SDK validation entirely because reasoningEffort never appears in the request body — the effort level is encoded in the model name, which codex.js already parses and strips generically at lines 184-192:

const effortLevels = ['none', 'low', 'medium', 'high', 'xhigh'];
for (const level of effortLevels) {
  if (body.model.endsWith(`-${level}`)) {
    modelEffort = level;
    body.model = body.model.replace(`-${level}`, '');
    break;
  }
}

Without server-side model entries for gpt-5.5-xhigh, gpt-5.5-none, etc., consumers using the Vercel AI SDK cannot access none or xhigh reasoning effort for GPT-5.5 and GPT-5.4.

This is an oversight — gpt-5.3-codex already has all effort variants manually enumerated, but gpt-5.5 and gpt-5.4 were added as quick 1-line entries without them.

What Changed

open-sse/config/providerModels.js

1. Added CODEX_EFFORT_LEVELS constant and withCodexEffortVariants(modelIds) helper function — a higher-order function (like withCodexReviewModels) that expands targeted model entries into effort level variants
2. Composed it with the existing withCodexReviewModels() in the cx provider:

   cx: withCodexReviewModels(withCodexEffortVariants(["gpt-5.5", "gpt-5.4"])([
     { id: "gpt-5.5", name: "GPT 5.5" },
     { id: "gpt-5.4", name: "GPT 5.4" },
     // ... rest unchanged
   ])),

New models generated

Model ID	Display Name
gpt-5.5-xhigh	GPT 5.5 (xHigh)
gpt-5.5-high	GPT 5.5 (High)
gpt-5.5-low	GPT 5.5 (Low)
gpt-5.5-none	GPT 5.5 (None)
gpt-5.4-xhigh	GPT 5.4 (xHigh)
gpt-5.4-high	GPT 5.4 (High)
gpt-5.4-low	GPT 5.4 (Low)
gpt-5.4-none	GPT 5.4 (None)

Each of these also gets a -review variant via withCodexReviewModels() (e.g. gpt-5.5-xhigh-review).

Why a helper function instead of manual entries?

• DRY: Future Codex models just need to be added to the target array (e.g. ["gpt-5.5", "gpt-5.4", "gpt-6.0"])
• Consistent naming: The helper ensures uniform (xHigh), (High), (Low), (None) suffixes
• Composable: Works with withCodexReviewModels() — effort variants also get -review variants automatically
• Non-breaking: Existing manually-defined gpt-5.3-codex-* variants are unaffected (they're not in the target set)

Related Issues

• Bug: Codex OAuth exposes unsupported ChatGPT-account models and misclassifies availability #718 — Codex OAuth exposes unsupported models (account tier issue, not model capability)
• The 'GPT-5.5' model is not supported when using Codex with a ChatGPT account. #787 — GPT-5.5 not supported on ChatGPT accounts
• Error with Codex GPT 5.5 #812 — Parameter stripping bug with GPT 5.5 (already fixed)

Testing

The executor's suffix parsing (codex.js lines 184-192) is model-name-agnostic — it works on any model name ending with -xhigh, -high, -low, -medium, or -none. No executor changes needed.

[mugnimaestra]

Updated this branch after review.

Current scope:

• Adds Codex model aliases for GPT-5.5/GPT-5.4 effort levels: xhigh, high, medium, low, matching the current upstream OpenAI Codex catalog.
• This lets clients that validate reasoningEffort still select xhigh via a model alias such as cx/gpt-5.5-xhigh; 9router already strips the suffix into Codex reasoning effort before dispatch.
• none was intentionally removed from the generated aliases because the upstream Codex catalog for these slugs currently advertises low/medium/high/xhigh, not none.
• Added focused tests for generated aliases, absence of none, review variant composition, upstream ID resolution, and quota family metadata.

Related but not exact duplicates: #748 / #668 are model-catalog cleanup requests; #718 / #717 / #812 cover Codex OAuth/request-contract/account availability; #836 covers review-model quota handling.

[mugnimaestra]

Hi @decolua, when you have a chance, could you please take a look? I updated the PR with upstream Codex-aligned effort aliases and focused tests. Thanks!

[vivekrp]

Any update on this?

[joseph-lewis]

Any update? I am doing this manually myself, everytime I update 9Router I add patches fixes that adds this. But it would be great if it was native in 9Router...