›_ Codex Issues
Codex Issues / #30439

Bug: model field in automation.toml is ignored; automation threads always use gpt-5.5

Open 💬 2 comments Opened Jun 28, 2026 by yangjinhui-lab
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

Bug: automation.toml 中的 model 字段被忽略，自动化线程始终使用 gpt-5.5

问题描述

在 Codex Desktop 中配置的自动化任务（automation.toml），即使明确指定了 model = "agnes-2.0-flash" 和 reasoning_effort = "high"，实际创建的线程仍然使用 gpt-5.5 + medium。

环境信息

  • Codex Desktop 版本: 0.142.3 (Build 4514)
  • Codex CLI 版本: 0.142.3
  • 操作系统: macOS
  • 自定义模型提供商: http://127.0.0.1:57321/v1

配置

~/.codex/config.toml

model = "agnes-2.0-flash"
model_reasoning_effort = "xhigh"
model_provider = "custom"

~/.codex/automations/automation/automation.toml

version = 1
id = "automation"
kind = "cron"
name = "技能审计"
status = "ACTIVE"
model = "agnes-2.0-flash"
reasoning_effort = "high"
rrule = "RRULE:FREQ=DAILY;INTERVAL=1;BYHOUR=0,8,16;BYMINUTE=0;BYSECOND=0"

证据

从 ~/.codex/state_5.sqlite 查询结果：

| 线程ID前8位 | 模型 | 推理强度 | 来源 | 创建时间 |
|------------|------|---------|------|---------|
| 019f0d62 | gpt-5.5 | medium | automation | 2026-06-28 16:40 |
| 019f0d40 | gpt-5.5 | medium | automation | 2026-06-28 16:02 |
| 019f0b88 | gpt-5.5 | medium | automation | 2026-06-28 08:01 |
| 019f09d0 | gpt-5.5 | medium | automation | 2026-06-28 00:01 |
| 019f0818 | gpt-5.5 | medium | automation | 2026-06-27 16:00 |

所有 thread_source = 'automation' 的线程都使用 gpt-5.5 + medium。

作为对比，手动创建的线程（thread_source = 'user'）正确使用了 agnes-2.0-flash + xhigh。

预期行为

自动化调度器（Jitter）应该读取 automation.toml 中的 model 和 reasoning_effort 字段，并使用指定的模型创建线程。

实际行为

自动化调度器忽略了 TOML 配置，硬编码使用 gpt-5.5 + medium。

影响

  • 自动化任务无法使用用户自定义的模型
  • 即使用户已配置自定义模型提供商，自动化仍回退到默认模型
  • 可能导致自定义模型不可用时自动化失败（如 503 错误）

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 2 months ago

English translation:

Bug: model field in automation.toml is ignored; automation threads always use gpt-5.5

Problem Description

For automation tasks configured in Codex Desktop (automation.toml), even when model = "agnes-2.0-flash" and reasoning_effort = "high" are explicitly specified, the threads that are actually created still use gpt-5.5 + medium.

Environment Information

  • Codex Desktop version: 0.142.3 (Build 4514)
  • Codex CLI version: 0.142.3
  • Operating system: macOS
  • Custom model provider: http://127.0.0.1:57321/v1

Configuration

~/.codex/config.toml

model = "agnes-2.0-flash"
model_reasoning_effort = "xhigh"
model_provider = "custom"

~/.codex/automations/automation/automation.toml

version = 1
id = "automation"
kind = "cron"
name = "Skill Audit"
status = "ACTIVE"
model = "agnes-2.0-flash"
reasoning_effort = "high"
rrule = "RRULE:FREQ=DAILY;INTERVAL=1;BYHOUR=0,8,16;BYMINUTE=0;BYSECOND=0"

Evidence

Query results from ~/.codex/state_5.sqlite:

| First 8 chars of thread ID | Model | Reasoning effort | Source | Created at |
|------------|------|---------|------|---------|
| 019f0d62 | gpt-5.5 | medium | automation | 2026-06-28 16:40 |
| 019f0d40 | gpt-5.5 | medium | automation | 2026-06-28 16:02 |
| 019f0b88 | gpt-5.5 | medium | automation | 2026-06-28 08:01 |
| 019f09d0 | gpt-5.5 | medium | automation | 2026-06-28 00:01 |
| 019f0818 | gpt-5.5 | medium | automation | 2026-06-27 16:00 |

All threads with thread_source = 'automation' use gpt-5.5 + medium.

By comparison, manually created threads (thread_source = 'user') correctly use agnes-2.0-flash + xhigh.

Expected Behavior

The automation scheduler (Jitter) should read the model and reasoning_effort fields in automation.toml and create threads using the specified model.

Actual Behavior

The automation scheduler ignores the TOML configuration and is hardcoded to use gpt-5.5 + medium.

Impact

  • Automation tasks cannot use the user's custom model
  • Even if the user has configured a custom model provider, automation still falls back to the default model
  • This may cause automation to fail when the custom model is unavailable (such as with 503 errors)

<!-- codex-issue-translator -->

czf688-ai · 1 month ago

Subject:
Windows Desktop Codex cannot start local agent (127.0.0.1:57321), while Codex CLI works normally

Hello OpenAI Support,

I'm experiencing an issue with the latest Windows Codex Desktop application.

Environment:

  • Windows 11
  • Codex Desktop version: 26.623.61825
  • Codex CLI version: v0.142.4
  • ChatGPT account login (not API key)

Problem:

The Desktop application always shows:

unexpected status 502 Bad Gateway

URL:
http://127.0.0.1:57321/v1/responses

However, after investigation, I found that the local service is not actually listening on port 57321.

Running:

curl http://127.0.0.1:57321/v1/models

returns:

curl: (7) Failed to connect to 127.0.0.1 port 57321

Also,

netstat -ano | findstr 57321

only shows:

127.0.0.1:6627 -> 127.0.0.1:57321 SYN_SENT

There is no LISTENING state on port 57321.

Things I have already tried:

  • Completely uninstalled Codex Desktop.
  • Removed previous Codex++ installation.
  • Deleted Codex configuration.
  • Reinstalled the latest Desktop from the official website.
  • Restarted Windows.
  • Disabled Clash Verge proxy completely.
  • Added HTTP_PROXY / HTTPS_PROXY / ALL_PROXY environment variables.
  • Verified runtime files were downloaded successfully.
  • Verified node.exe and Codex.exe processes are running.

Interestingly, Codex CLI (v0.142.4) works perfectly using the same ChatGPT account.

I can log in successfully and use GPT-5.5 without any issue.

This suggests that:

  • My account is working.
  • Authentication is working.
  • Network connectivity is working.

The issue appears to be specific to the Windows Desktop application's local agent failing to start.

Could you please help investigate this issue?

Thank you.