Each is a one-row entry in the OpenAI-compatible provider registry (base_url,
key env, CLI option); the model is user-specified since they serve many models.
The OpenAI-compatible family (openai, xAI, DeepSeek, Qwen, GLM, MiniMax,
OpenRouter, Ollama) all speak the same Chat Completions API and differ only by
base_url, key, and two narrow wire-format quirks already isolated in subclasses.
Replace the scattered base-URL dict, key handling, and client-class branches with
one ProviderSpec registry that get_llm and the factory drive off; provider quirks
stay in their subclasses. Add a generic "openai_compatible" provider for any
OpenAI-compatible server (vLLM, LM Studio, llama.cpp, relays) via backend_url +
optional key — adding a provider is now one registry row. Native Anthropic/Google
keep their own clients (genuinely different APIs). Also fixes the env backend URL
being ignored when the provider was chosen interactively (#978).
Adds a cross-provider temperature config (and TRADINGAGENTS_TEMPERATURE),
forwarded to every LLM client when set, so runs can be made less variable
on models that honor it. Adds a README "Reproducibility" section that
separates the sources of run-to-run variation, what users can control
(temperature, non-reasoning model, pinned date), and what is inherent to
LLM-driven analysis, and notes that the identity and verified-data fixes
already removed the "different companies / fabricated prices" variance.
#178#168
Haiku 4.5 rejects the effort parameter with 400. AnthropicClient.get_llm()
now drops effort when the model isn't in the supported set (Opus 4.5+,
Sonnet 4.5+, mythos-preview). Forward-compat regex catches future
claude-{opus,sonnet}-X-Y releases automatically; Haiku and unknown
models stay excluded conservatively.
14 tests cover Haiku exclusion, current Opus/Sonnet inclusion, future-
version inheritance via pattern, mythos-preview, unknown-default
exclusion, and other passthrough kwargs surviving the effort-skip path.
MinimaxChatOpenAI unconditionally set reasoning_split=True, but the
kwarg is only valid on M2.x reasoning models. The openai SDK's strict
kwarg validation raised TypeError for Coding Plan and any other non-
reasoning MiniMax model.
Adds requires_reasoning_split to ModelCapabilities, gates the payload
injection on it, and only sets True for _MINIMAX_THINKING (M2.x exact
IDs and the ^MiniMax-M\d forward-compat pattern). Same shape as the
existing supports_tool_choice gate.
Regression tests cover both halves: M2.x models still receive the flag,
non-reasoning MiniMax models do not.
Users with other models pulled via `ollama pull` (beyond the three
suggested defaults) can now select "Custom model ID" and type any
model name. Matches the same pattern used for DeepSeek, GLM, Qwen,
and MiniMax — the existing _prompt_custom_model_id flow handles the
"custom" value generically, so this is a one-row catalog addition
plus regression coverage.
OLLAMA_BASE_URL now flows through both the CLI dropdown and the
programmatic client (call-time evaluation so tests behave). After
provider selection, the CLI prints the resolved endpoint and marks
when it came from the env var, plus a soft warning when the URL is
missing a scheme or non-default port. Drops the stale "(local)"
suffix from Ollama model labels since the endpoint is now dynamic.
Adds a canonical PROVIDER_API_KEY_ENV mapping (14 providers including
the three dual-region pairs) and an ensure_api_key() helper. When the
selected provider's key is absent from the environment, the CLI prompts
via questionary.password, writes the value to .env via python-dotenv's
set_key (preserves existing lines), and exports it into os.environ so
the run continues without restart. Wired into cli/main.py right after
the region prompts so qwen-cn, glm-cn, and minimax-cn each check their
own region-specific key. openai_client refactored to consult the same
mapping, eliminating its private duplicate of provider→env-var data.
Zhipu serves GLM under two brands with separate accounts (Z.AI
international vs BigModel China); the CLI URL pointed at one while
the openai_client default pointed at the other. Split into glm +
glm-cn with secondary region prompt (same UX as Qwen + MiniMax).
Catalog adds glm-5-turbo and glm-4.5-air per docs.z.ai.
Qwen and MiniMax each had two main-dropdown entries (intl + CN);
consolidate to one entry per provider and prompt for region as a
secondary step. Internal provider keys (qwen-cn, minimax-cn) and
endpoint routing unchanged. Add qwen3.6-flash to the Qwen catalog
and drop the version-less aliases (qwen-flash, qwen-plus) that
auto-shift their backing model per Alibaba's docs.
#758
xAI's official docs lead with grok-4.20-reasoning and
grok-4.20-non-reasoning across all SDK examples. Replace the prior
grok-4-1-fast-* entries (hyphens where docs use dots, no literal
code example) with the verified grok-4.20 family. Keep grok-4-0709
and grok-4-fast variants that are still referenced.
gemini-3.1-flash-lite is now GA per ai.google.dev. Use the stable
version (fewer rate limits, stronger compat guarantees) instead of
the -preview suffix. Labels mark preview vs GA explicitly.
Opus 4.7 is the current frontier per platform.claude.com (frontier
category, listed first). Demote Opus 4.6 to second deep-tier slot.
Polish quick-tier labels to match official wording; effort docstring
includes 4.7.
GPT-5.5 (Apr 2026, 1M ctx, $5/$30 per 1M) replaces GPT-5.4 as the
catalog flagship. GPT-5.5 Pro replaces 5.4 Pro in the most-capable
slot. GPT-5.4 demotes to previous-gen cost-effective option.
M2.x tool_choice is enum-only (none/auto), so route through the
no-tool_choice dispatch. MinimaxChatOpenAI injects reasoning_split
so <think> blocks stay out of content. Catalog rounded out to the
full official M2.x lineup plus forward-compat regex.
- dataflows/config: deepcopy + one-level dict merge so a partial
set_config doesn't clobber sibling defaults
- graph: thread max_recur_limit from config to Propagator
- openai_client: name the missing env var in the API-key error
#788#764#680
Two regional endpoints (global api.minimax.io, China api.minimaxi.com)
with separate API keys. Models M2.7 / M2.5 plus -highspeed variants,
204K context. Follows the existing provider-preset pattern.
#789#609#577#546#395#378
DeepSeek V4 and reasoner reject tool_choice but accept tools.
Route via a per-model capability table that suppresses tool_choice
for thinking-mode models.
#678#689
Resolves#599: thinking-mode models require reasoning_content to be
echoed back across turns; multi-turn agent runs failed with HTTP 400.
The fix isolates DeepSeek's quirks (reasoning_content round-trip and
the deepseek-reasoner no-tool_choice limitation) into a subclass so
the general OpenAI-compatible client stays untouched. Adds DeepSeek
V4 Pro/Flash to the catalog. 9 new tests; rationale documented in
the class docstrings.
Design adapted from #600; #611 closed in favour of this approach.
This release bundles substantial work since v0.2.3:
- Structured-output Research Manager, Trader, and Portfolio Manager
(canonical with_structured_output pattern, single LLM call per agent,
rendered markdown preserves the existing report shape).
- LangGraph checkpoint resume for crash recovery (--checkpoint flag).
- Persistent decision log replacing the per-agent BM25 memory, with
deferred reflection driven by yfinance returns + alpha vs SPY.
- DeepSeek, Qwen, GLM, and Azure OpenAI provider support; dynamic
OpenRouter model selection.
- Docker support; cache and logs moved to ~/.tradingagents/ to fix
Docker permission issues.
- Windows UTF-8 encoding fix on every file I/O site.
- 5-tier rating consistency (Buy / Overweight / Hold / Underweight / Sell)
across Research Manager, Portfolio Manager, signal processor, memory log.
Plus the small quality items in this commit:
1. Suppress noisy Pydantic serializer warnings from OpenAI Responses-API
parse path by defaulting structured-output to method="function_calling"
(root-cause fix, not a warnings filter — same typed result, no warnings).
2. Ship scripts/smoke_structured_output.py so contributors can verify
their provider's structured-output path with one command.
3. Add opt-in memory_log_max_entries config — when set, oldest resolved
memory log entries are pruned once the cap is exceeded; pending
entries (unresolved) are never pruned.
4. backend_url default changed from the OpenAI URL to None so the
per-provider client falls back to its native endpoint instead of
leaking OpenAI's URL into Gemini / other clients.
CHANGELOG.md added with the full v0.2.4 entry. 92 tests pass without API keys.
Apply review suggestions: use concise `or` pattern for API key
resolution, consolidate tests into parameterized subTest, move
import to module level per PEP 8.
GoogleClient now accepts the unified `api_key` parameter used by
OpenAI and Anthropic clients, mapping it to the provider-specific
`google_api_key` that ChatGoogleGenerativeAI expects. Legacy
`google_api_key` still works for backward compatibility.
Resolves TODO.md item #2 (inconsistent parameter handling).
Add effort parameter (high/medium/low) for Claude 4.5+ and 4.6 models,
consistent with OpenAI reasoning_effort and Google thinking_level.
Also add content normalization for Anthropic responses.
- Point requirements.txt to pyproject.toml as single source of truth
- Resolve welcome.txt path relative to module for CLI portability
- Include cli/static files in package build
- Extract shared normalize_content for OpenAI Responses API and
Gemini 3 list-format responses into base_client.py
- Update README install and CLI usage instructions
Enable use_responses_api for native OpenAI provider, which supports
reasoning_effort with function tools across all model families.
Removes the UnifiedChatOpenAI subclass workaround.
Closes#403
- Add http_client and http_async_client parameters to all LLM clients
- OpenAIClient, GoogleClient, AnthropicClient now support custom httpx clients
- Fixes SSL certificate verification errors on Windows Conda environments
- Users can now pass custom httpx.Client with verify=False or custom certs
Fixes#369
- OpenAI: add GPT-5.4, GPT-5.4 Pro; remove o-series and legacy GPT-4o
- Anthropic: add Claude Opus 4.6, Sonnet 4.6; remove legacy 4.1/4.0/3.x
- Google: add Gemini 3.1 Pro, 3.1 Flash Lite; remove deprecated
gemini-3-pro-preview and Gemini 2.0 series
- xAI: clean up model list to match current API
- Simplify UnifiedChatOpenAI GPT-5 temperature handling
- Add missing tradingagents/__init__.py (fixes pip install building)
- Add StatsCallbackHandler for tracking LLM calls, tool calls, and tokens
- Integrate callbacks into TradingAgentsGraph and all LLM clients
- Dynamic agent/report counts based on selected analysts
- Fix report completion counting (tied to agent completion)
