feat(llm): add openai_codex provider using ChatGPT subscription quota

[LeonWTW]

Closes #1194 — that issue carries the full rationale; this is the implementation.

Adds an openai_codex provider so a run can be billed to an existing ChatGPT subscription instead of API credits. Nothing changes for any other provider: every code path here is behind the new provider name.

Why this needs a provider rather than openai_compatible + backend_url

The endpoint breaks three assumptions the shared OpenAI client makes:

Assumption	Reality	Handled in
Auth comes from an env var	OAuth token in ~/.codex/auth.json, refreshed near expiry, refresh token rotates and must be written back or the Codex client logs out	llm_clients/codex_auth.py via ProviderSpec.credentials_fn
stream: false / store: true / temperature are fine	All rejected	CodexChatOpenAI — first two pinned as constructor kwargs (a class-level field default sets the payload but not langchain's streaming dispatch), temperature filtered in get_llm
system role is fine	400 "System messages are not allowed" — and every agent prompt here is a ChatPromptTemplate starting with a system message, so the first analyst node fails on its first call	_get_request_payload relabels those items developer, the Responses-API name for the same role

What changed

• tradingagents/llm_clients/codex_auth.py (new) — resolve, refresh, rotate, persist. Path overridable via TRADINGAGENTS_CODEX_AUTH_PATH.
• tradingagents/llm_clients/openai_client.py — CodexChatOpenAI; auth-error messages stay actionable after the file re-read.
• tradingagents/llm_clients/model_catalog.py, api_key_env.py — registry entry; no API key is read for this provider.
• tradingagents/graph/trading_graph.py — _get_provider_kwargs forwards openai_reasoning_effort for openai_codex as well.
• cli/utils.py, cli/main.py — picker entry, login preflight (so an unauthenticated user gets "sign in with the Codex client first" instead of a 401 mid-run), and the reasoning-effort prompt.
• README.md — provider docs, including the caveats below.

Trying it

codex login          # or sign in with the Codex desktop app
tradingagents        # pick "OpenAI Codex (ChatGPT subscription quota)"

Subscription traffic hits transient "servers are currently overloaded" errors far more often than the paid API, so the README suggests TRADINGAGENTS_LLM_MAX_RETRIES=5 or higher.

Testing

pytest -q → 614 passed, 5 skipped; ruff check . → clean.

New coverage: auth-file resolution/refresh/rotation (test_codex_auth.py), provider registration and payload shaping (test_codex_provider.py), CLI preflight (test_codex_cli.py). test_codex_smoke.py drives the live endpoint with a real ChatPromptTemplate and is opt-in behind -m smoke, so CI never calls out. The smoke test matters for the system→developer fix specifically: a bare-string invoke carries no system message and passed throughout, which is how that bug survived.

Caveats

This endpoint is undocumented and unversioned, so OpenAI can change it without notice, and driving subscription credentials from a third-party tool is not clearly sanctioned by OpenAI's terms. Both points are stated in the README next to the provider. If you would rather gate this further or not carry it at all, say so in #1194 and I will rework or withdraw it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FFuMQCxPhTkWLwrCQ57tTp

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[LeonWTW]

CI is held at action_required here pending maintainer approval, so I ran the same workflow on my fork in the meantime — all six jobs green (py3.10–3.13, ruff strict full-repo, clean-install smoke):

https://github.com/LeonWTW/TradingAgents/actions/runs/30714455163

No rush on my side; flagging it only so the approval decision doesn't have to double as a "does it even build" question.