v0.4.0-beta.1 — Codex Third-Party Workers public beta
Pre-release
Codex Third-Party Workers public beta
This beta introduces a reviewed provider-pack layer that lets Codex delegate
bounded tasks to selected third-party model APIs while keeping the main Codex
thread on its OpenAI model, provider, and authentication flow.
Supported delegation is intentionally limited to suitable text, code, research
synthesis, and local-validation tasks. Codex remains responsible for reviewing
and synthesizing the worker result.
Included provider packs
This release includes three built-in provider packs:
- DeepSeek V4 Flash
- MiniMax-M3
- Alibaba Model Studio Qwen3.7-Max
A built-in pack is not automatically a runtime-verified pack.
Runtime-verified providers
MiniMax-M3
Verified with:
- Responses API
- SSE streaming
- two-turn Function Calling
- Codex CLI
- Codex Desktop
minimax_worker - successful task-bridge completion and release
Qwen3.7-Max
Verified with:
- Responses API
- SSE streaming
- automatic Function Calling
- Codex CLI
- Codex Desktop
qwen_worker - successful task-bridge completion and release
Qwen remains text-only in this project. Thinking mode does not accept
tool_choice: required; automatic tool choice was verified.
Public-installer --apply and verify acceptance remain separate pending checks
for a real user environment.
Built-in, runtime validation still pending
DeepSeek V4 Flash
DeepSeek V4 Flash is built in and covered by isolated installer, routing,
catalog, and bridge tests. Public-installer runtime verification is still
pending. DeepSeek V4 Pro is explicitly excluded.
Candidate providers — real API validation required
These providers have relevant official Responses documentation but are not
included or runtime verified:
- StepFun Responses models
- Volcano Ark Responses models
They must pass streaming, Function Calling, Codex subagent execution, failure
fallback, bridge release, and main-thread review before becoming built-in packs.
Partial / gateway candidates
- Baidu Qianfan Responses gateway
- Tencent Cloud TokenHub
Gateway compatibility does not prove that the underlying model vendor's direct
API is compatible.
Currently incompatible direct paths
The following direct provider paths do not currently satisfy this project's
Responses contract:
- Kimi K3 direct API
- Zhipu GLM direct API
- legacy Tencent Hunyuan endpoints
- SiliconFlow direct API
External protocol translators may provide other integration paths, but they are
outside this project's native provider-pack scope.
Verification
37/37isolated local tests passed- GitHub Actions passed
- Responses, SSE, and Function Calling evidence for MiniMax and Qwen
- Codex CLI and Desktop subagent smoke tests
- owner-only single-slot bridge and redacted archive checks
- secret-safety, personal-path, fixed-UID, and credential-value checks
Security boundaries
- API keys remain in macOS Keychain
- the installer never accepts
--api-key - secrets are never committed to the repository
- dry-run is the default; changes require explicit
--apply - the main
~/.codex/config.tomlis not modified - active bridge files use owner-only permissions
- completed and failed task archives redact task messages and working paths
- browser, desktop, image, audio, video, MCP, and Computer Use delegation are
outside the supported scope
Users remain responsible for reviewing each provider's privacy, pricing,
data-retention, billing, and regional policies.
Documentation
This is an unofficial, macOS-only public beta. It is not endorsed by OpenAI,
DeepSeek, MiniMax, or Alibaba Cloud. It does not claim compatibility with every
provider and does not promise a fixed cost-saving percentage.