[Bug] Native Codex session config normalization incorrectly clamps max and ultra reasoning effort to xhigh

[pg-adm1n]

Describe the bug

When starting a native Codex session (codex-native harness, e.g. via omnigent codex or sub-agent dispatches), Omnigent creates a private session CODEX_HOME and copies the user's config.toml via _populate_codex_home_config.

During this copy, _normalize_copied_codex_effort() in omnigent/inner/codex_executor.py normalizes top-level model_reasoning_effort. However, it checks against CODEX_EFFORTS (the in-process OpenAI Responses API ladder, which tops out at xhigh), rather than CODEX_NATIVE_EFFORTS (which includes max and ultra).

As a result, any native Codex session with model_reasoning_effort = "max" or "ultra" in ~/.codex/config.toml (or configured on native models such as gpt-5.6-luna or sol) has its reasoning effort silently downgraded to xhigh in the per-session config.toml.

Reproduction steps

1. In ~/.codex/config.toml, configure model_reasoning_effort = "max" (or "ultra").
2. Launch a native Codex session (codex-native harness), for example with gpt-5.6-luna.
3. Inspect the per-session config in ~/.omnigent/codex-native/<session>/codex-home/config.toml.
4. Observe that model_reasoning_effort has been rewritten to "xhigh".

Expected behavior

Native Codex sessions should preserve max and ultra reasoning effort in their session config.toml because codex-native drives the real Codex CLI process which validates and supports CODEX_NATIVE_EFFORTS. In-process Responses API sessions can continue normalizing against CODEX_EFFORTS.

Environment

• Platform: macOS / Linux
• Omnigent version: 0.11.0.dev0 (main)
• Harness: codex-native

[github-actions]

Thanks for reporting this. #2696 may be related, and has already been fixed — so the fix may have shipped after the build you're on.

Could you check whether you're on a version that includes it? If you are and this still happens, say so here — that makes it a regression rather than a duplicate, and we'll keep this open.

[github-actions]

🤖 Automated triage

• Bot assessment: Medium impact
• Priority: P2-medium
• Why: Native Codex users selecting max or ultra reasoning have their session silently downgraded to xhigh, reducing result quality but not blocking session execution; manually correcting the copied per-session config is a workaround.

This automated assessment uses the issue content and repository signals. Maintainers can override the priority label.

[pg-adm1n]

This is confirmed on the latest main build (v0.11.0.dev0, past #2696).

#2696 added normalization to prevent OpenAI Responses API 400 errors for the in-process codex SDK executor. However, codex-native (native CLI harness) was later integrated and shares _populate_codex_home_config(). Because _normalize_copied_codex_effort() only checked against CODEX_EFFORTS (Responses API ladder) rather than CODEX_NATIVE_EFFORTS, native Codex sessions (which drive the real Codex binary supporting max and ultra) still had their effort clamped to xhigh.

PR #5361 addresses this by parameterizing supported_efforts so codex-native passes CODEX_NATIVE_EFFORTS while in-process SDK runs continue normalizing against CODEX_EFFORTS.