ChineseModels Codex adapter — Windows
Validated 2026-10-22 with Codex CLI 0.162.1 and bundled 0.163.0-alpha.10.2.

Requirements: Windows, PowerShell 5.1+, Node.js 20+, Codex, ChineseModels API key.
No API key, authentication token, or user configuration is included in this archive.

From a PowerShell terminal in this extracted folder:
  powershell -NoProfile -ExecutionPolicy Bypass -File .\install-windows.ps1

Enter your API key at the hidden prompt. It is encrypted with Windows DPAPI.
The installer prepares a profile without changing the main Codex provider.

Start a separate session:
  codex --no-daemon --profile chinesemodels
Then type /model. Select the model and supported effort.

Normal Codex and desktop sessions retain your OpenAI models, including Astra.
Use the separate profile command above for ChineseModels. This setup uses
one provider per session. Merging the profile into the main config replaces
the normal picker, so the default instructions preserve the main provider.

Refresh available models:
  powershell -NoProfile -ExecutionPolicy Bypass -File "$env:USERPROFILE\codex-chinesemodels\refresh-models.ps1"
Restart Codex afterward; the picker reads the local catalog at startup.

The adapter runs only on 127.0.0.1:18473 and authenticates local requests.
Codex starts it on demand using the profile's auth helper. No startup task is installed.
To use a different port, change it in bridge-settings.json AND the profile/main base_url.

Model capabilities:
- /v1/models is the source of account availability, not full capability metadata.
- Image and video generation models are excluded from the text picker.
- Thinking mappings are in capabilities.mjs, based on model-family documentation.
- Some models always think. Some only toggle On/Off. Effort tiers vary by model.
- A model/option appearing in the picker does not certify live upstream tool support.
- "High" for a toggle-only model means Thinking On, not a separate effort tier.
- Provider-default entries do not send a thinking override.
- MiniMax M3 uses adaptive thinking; a simple request may produce no reasoning.
- Start a fresh conversation when changing models with preserved reasoning.
- The context limit is a conservative 32,768-token adapter setting, not a model specification.

Limitations:
Text input only. No image/video generation, vision input, hosted web search,
audio, server-side response history, or /responses/compact implementation.
The adapter consumes upstream streaming but emits completed output in batches;
token-by-token thinking is not displayed. Reasoning is encrypted for replay.
Unsupported features fail explicitly. Long conversations depend on Codex's local
compaction; remote compaction is not supported.
Provider support and aliases can change. Refresh metadata and verify important workloads.
If the bridge auth token is replaced, prior encrypted reasoning cannot be replayed.

To restore your previous provider, restore the backup of config.toml made during
activation and restart Codex. Do not overwrite later unrelated config changes blindly.
