Route TradingAgents through Claude Code and Codex CLIs

[NagyVikt]

Summary

Adds experimental local CLI-backed LLM providers so TradingAgents can route model calls through an already authenticated Claude Code or Codex subscription instead of requiring direct Anthropic/OpenAI API keys.

What changed

• Adds claude-code provider backed by the local claude -p CLI.
• Adds codex provider backed by the local codex exec CLI in read-only sandbox mode.
• Wires both providers into the LLM factory, default config, model catalog, validators, API-key detection, and interactive CLI provider picker.
• Keeps TradingAgents tool execution in LangGraph by translating local-agent JSON tool-call requests into LangChain AIMessage.tool_calls.
• Supports wrapper commands such as npx claude / npx codex via shlex.split.
• Reports stdout as a fallback when local CLI failures do not write stderr.
• Serializes in-process codex exec calls to reduce Codex subscription refresh races during multi-agent runs.
• Detects Codex refresh-token / websocket 401 failures and returns a clear codex logout / codex login recovery message.
• Uses TRADINGAGENTS_CLAUDE_CODE_EXTRA_ARGS and TRADINGAGENTS_CODEX_EXTRA_ARGS for adapter-specific extra CLI flags.
• Documents environment variables, example setup, stale installed-command troubleshooting, and Codex login refresh recovery in the README.
• Adds focused unit coverage for provider creation, subprocess invocation, no-API-key behavior, custom model validation, tool-call JSON conversion, wrapper commands, stdout-only errors, Codex auth guidance, Codex serialization, and prefixed extra args.

Usage

Claude Code:

export TRADINGAGENTS_LLM_PROVIDER=claude-code
export TRADINGAGENTS_DEEP_THINK_LLM=opus
export TRADINGAGENTS_QUICK_THINK_LLM=sonnet

Codex:

export TRADINGAGENTS_LLM_PROVIDER=codex
export TRADINGAGENTS_DEEP_THINK_LLM=gpt-5.5
export TRADINGAGENTS_QUICK_THINK_LLM=gpt-5.4-mini

Optional overrides:

export TRADINGAGENTS_CLAUDE_CODE_COMMAND=claude
export TRADINGAGENTS_CLAUDE_CODE_TIMEOUT=600
export TRADINGAGENTS_CLAUDE_CODE_EXTRA_ARGS="--verbose"
export TRADINGAGENTS_CODEX_COMMAND=codex
export TRADINGAGENTS_CODEX_TIMEOUT=600
export TRADINGAGENTS_CODEX_EXTRA_ARGS="--profile trading"

If an installed tradingagents command keeps showing old Codex flags such as --ask-for-approval, refresh the installed command from the checkout:

python3 -m pip install --user -e .
python3 -c "import tradingagents, tradingagents.llm_clients.codex_client as c; print(tradingagents.__file__); print(c.__file__)"

Both paths should point at the current checkout, not stale site-packages files.

If Codex reports that its access token could not be refreshed, refresh the local subscription login:

codex logout
codex login

Validation

Focused regression suite:

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 /home/deadpool/.nix-profile/bin/uv run --python 3.13 --with pytest python -m pytest tests/test_codex_client.py tests/test_claude_code_client.py tests/test_api_key_env.py tests/test_model_validation.py -q

Result: 44 passed, 74 subtests passed.

Runtime-path verification after stale install report:

python3.10 -m pip install --user -e .
python3.10 -c "import tradingagents, cli.main, tradingagents.llm_clients.codex_client as c; print(tradingagents.__file__); print(cli.main.__file__); print(c.__file__)"

Result: imports resolve to /home/deadpool/Documents/TradingAgents/..., and Codex subprocess args no longer include --ask-for-approval.

Auth-error verification:

• The installed editable package resolves tradingagents, cli.main, and codex_client.py to this checkout.
• The installed Codex adapter exposes _CODEX_EXEC_LOCK and _codex_error_message.
• Simulated Codex refresh-token failure returns codex logout and codex login guidance.

Local CLI presence checked:

• claude 2.1.138 (Claude Code)
• codex-cli 0.130.0

GitHub reports no checks configured for main, so local focused tests are the validation evidence for this PR.

Notes and limits

This is intentionally a subprocess adapter, not a replacement SDK. It constrains Codex to read-only sandbox mode and Claude Code to print mode with Claude Code tools disabled. TradingAgents data tools still execute through the existing LangGraph tool path.

Native structured output is not supported by these local CLI providers; existing free-text fallback paths handle structured-agent calls. If the Codex CLI refresh token is already invalid, TradingAgents cannot repair it; the user must refresh local Codex login state.

Latest auth-retry finish:

• Added interactive Codex auth refresh retry: on refresh-token/auth failure, TradingAgents pauses in a TTY, lets the user run codex logout / codex login in another terminal, then retries the failed Codex call once.
• Added TRADINGAGENTS_CODEX_AUTH_RETRY=0 to disable the pause for unattended runs.
• Validation: python -m py_compile tradingagents/llm_clients/codex_client.py tests/test_codex_client.py passed.
• Validation: PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 UV_CACHE_DIR=/tmp/uv-cache UV_PROJECT_ENVIRONMENT=/tmp/tradingagents-uv-venv-313 /home/deadpool/.nix-profile/bin/uv run --python /home/deadpool/.nix-profile/bin/python3.13 --with pytest pytest tests/test_codex_client.py tests/test_claude_code_client.py tests/test_model_validation.py tests/test_api_key_env.py passed: 46 passed in 1.08s.

[gemini-code-assist]

Code Review

This pull request introduces experimental support for claude-code and codex LLM providers, allowing the system to route calls through local CLI tools via subprocess adapters. The changes include new client implementations, configuration options, documentation, and comprehensive tests. Feedback focuses on ensuring robust subprocess execution by using shlex.split for multi-word commands, enhancing error reporting by capturing both stdout and stderr on failure, and maintaining consistent environment variable naming while removing redundant configuration logic.