111 lines
8.3 KiB
Markdown
111 lines
8.3 KiB
Markdown
<!-- Last verified: 2026-03-23 -->
|
|
|
|
# Conventions
|
|
|
|
## Configuration
|
|
|
|
- Env var override pattern: `TRADINGAGENTS_<UPPERCASE_KEY>=value` — empty/unset preserves default. (`default_config.py`)
|
|
- Per-tier overrides: each tier has `{tier}_llm_provider` and `{tier}_backend_url`, falling back to top-level `llm_provider` and `backend_url`. (`default_config.py`)
|
|
- `load_dotenv()` runs at module level in `default_config.py` — import-order-independent. Check actual env var values when debugging auth. (`default_config.py`)
|
|
- `llm_provider` and `backend_url` must always exist at top level — `scanner_graph.py` and `trading_graph.py` use them as fallbacks. (ADR 006)
|
|
- `mid_think_llm` defaults to `None`, meaning mid-tier falls back to `quick_think_llm`. (`default_config.py`)
|
|
|
|
## Agent Creation
|
|
|
|
- Factory pattern: `create_X(llm)` returns a closure `_node(state)`. Some factories take extra params: `create_bull_researcher(llm, memory)`, `create_trader(llm, memory)`. (`tradingagents/agents/`)
|
|
- When `bind_tools()` is used, there MUST be a tool execution path — either `ToolNode` in graph or `run_tool_loop()` inline. (ADR 004)
|
|
|
|
## Tool Execution
|
|
|
|
- Trading graph: analysts use `ToolNode` in the LangGraph graph with conditional routing (`should_continue_X`). (`graph/setup.py`)
|
|
- Scanner agents: use `run_tool_loop()` inline — no `ToolNode`, tools execute inside the agent node. (`agents/utils/tool_runner.py`)
|
|
- `MAX_TOOL_ROUNDS = 5` — max iterations of tool calling before returning. (`tool_runner.py`)
|
|
- `MIN_REPORT_LENGTH = 2000` — if first response is shorter and has no tool calls, a nudge message is appended asking the LLM to call tools. Fires at most once. (`tool_runner.py`)
|
|
|
|
## Vendor Routing
|
|
|
|
- Fail-fast by default (ADR 011). Only methods in `FALLBACK_ALLOWED` get cross-vendor fallback:
|
|
- `get_stock_data`
|
|
- `get_market_indices`
|
|
- `get_sector_performance`
|
|
- `get_market_movers`
|
|
- `get_industry_performance`
|
|
- Never add news, indicator, or financial-statement tools to `FALLBACK_ALLOWED` — data contracts differ across vendors. (ADR 011)
|
|
- Functions inside `route_to_vendor` must RAISE on failure, not embed errors in return values. (`interface.py`)
|
|
- Catch `(AlphaVantageError, FinnhubError, ConnectionError, TimeoutError)`, not just `RateLimitError`. (`interface.py`)
|
|
- Exception chaining required: `raise RuntimeError(...) from last_error`. (ADR 011)
|
|
- 2-level routing: category-level (`data_vendors` config dict) + tool-level override (`tool_vendors` config dict). (`interface.py`)
|
|
|
|
## yfinance Gotchas
|
|
|
|
- `top_companies` has ticker as the DataFrame INDEX, not a column. Access via `.index`, not a column name. (ADR 003)
|
|
- `Sector.overview` has NO performance data. Use ETF proxies (SPDR sector ETFs) for sector performance. (ADR 003)
|
|
- Always use `.head(10)` for both download and display in industry performance. (ADR 009)
|
|
|
|
## LangGraph State
|
|
|
|
- Any state field written by parallel nodes MUST have a reducer (`Annotated[str, reducer_fn]`). (ADR 005)
|
|
- `ScannerState` uses `_last_value` reducer (keeps newest value) for all report fields. (`scanner_states.py`)
|
|
- State classes: `AgentState` (trading), `InvestDebateState` (debate sub-state), `RiskDebateState` (risk sub-state), `ScannerState` (scanner). (`agent_states.py`, `scanner_states.py`)
|
|
|
|
## Threading & Rate Limiting
|
|
|
|
- Never hold a lock during `sleep()` or IO. Pattern: release lock, sleep outside, re-acquire. (ADR 007)
|
|
- Alpha Vantage: 75 calls/min (premium). (`alpha_vantage_common.py`)
|
|
- Finnhub: 60 calls/min (free tier). (`finnhub_common.py`)
|
|
- Finnhub paid-tier endpoints (`/stock/candle`, `/financials-reported`, `/indicator`) must never be called on free key. (ADR 010)
|
|
|
|
## Ollama
|
|
|
|
- Never hardcode `localhost:11434`. Use configured `base_url` from config. (ADR 001)
|
|
|
|
## CLI Patterns
|
|
|
|
- Typer for command definitions, Rich for live UI. (`cli/main.py`)
|
|
- `MessageBuffer` — deque-based singleton tracking agent statuses, reports, tool calls. Fixed agents grouped by team (`FIXED_AGENTS`), analysts selectable. (`cli/main.py`)
|
|
- `StatsCallbackHandler` — token and timing statistics for display. (`cli/stats_handler.py`)
|
|
- All reports go under `reports/daily/{date}/` — use helpers from `report_paths.py`: `get_market_dir(date)` for scan results, `get_ticker_dir(date, ticker)` for per-ticker analysis, `get_eval_dir(date, ticker)` for eval logs. Never hardcode report paths. (`report_paths.py`)
|
|
|
|
## Pipeline Patterns
|
|
|
|
- `MacroBridge` is the facade class for scan → filter → per-ticker analysis. (`pipeline/macro_bridge.py`)
|
|
- `ConvictionLevel = Literal["high", "medium", "low"]`; `CONVICTION_RANK = {"high": 3, "medium": 2, "low": 1}`. (`macro_bridge.py`)
|
|
- `extract_json()` handles DeepSeek R1 `<think>` blocks, markdown fences, and raw JSON. (`json_utils.py`)
|
|
|
|
## Testing
|
|
|
|
- Run tests: `conda activate tradingagents && pytest tests/ -v`
|
|
- Skip integration tests: `pytest tests/ -v -m "not integration"`
|
|
- Skip paid-tier tests: `pytest tests/ -v -m "not paid_tier"`
|
|
- Mocking vendor methods: patch `VENDOR_METHODS` dict entries directly (it stores function refs), not module attributes. (`interface.py`)
|
|
- Env isolation: always mock env vars before `importlib.reload()` — `load_dotenv()` leaks real `.env` values otherwise.
|
|
- `callable()` returns False on LangChain `@tool` objects — use `hasattr(x, "invoke")` instead.
|
|
|
|
## Observability
|
|
|
|
- Create one `RunLogger` per CLI command (analyze/scan/pipeline). Attach `logger.callback` to LLM constructors. (`observability.py`)
|
|
- Call `set_run_logger(logger)` at run start so vendor/tool layers can access it via `get_run_logger()`. (`observability.py`)
|
|
- Vendor calls: `log_vendor_call(method, vendor, success, duration_ms)` — called inside `route_to_vendor`. (`observability.py`, `interface.py`)
|
|
- Tool calls: `log_tool_call(tool_name, args_summary, success, duration_ms)` — called inside `run_tool_loop`. (`observability.py`, `tool_runner.py`)
|
|
- Write the run log at the end of each command: `logger.write_log(report_dir / "run_log.jsonl")`. (`observability.py`)
|
|
|
|
## Error Handling
|
|
|
|
- Fail-fast by default — no silent fallback unless method is in `FALLBACK_ALLOWED`. (ADR 011)
|
|
- Alpha Vantage hierarchy: `AlphaVantageError` → `APIKeyInvalidError`, `RateLimitError`, `ThirdPartyError`, `ThirdPartyTimeoutError`, `ThirdPartyParseError`. (`alpha_vantage_common.py`)
|
|
- Finnhub hierarchy: `FinnhubError` → `APIKeyInvalidError`, `RateLimitError`, `ThirdPartyError`, `ThirdPartyTimeoutError`, `ThirdPartyParseError`. (`finnhub_common.py`)
|
|
|
|
## AgentOS Patterns
|
|
|
|
- **REST endpoints only queue runs** — WebSocket is the sole executor. POST `/api/run/{type}` writes to in-memory store, WS `/ws/stream/{run_id}` picks up and executes. (`runs.py`, `websocket.py`)
|
|
- **Event mapping is crash-proof** — `_map_langgraph_event()` wraps each event type branch in try/except. `_safe_dict()` helper converts non-dict metadata to empty dict. (`langgraph_engine.py`)
|
|
- **Model name extraction** uses 3 fallbacks: `invocation_params` → serialized kwargs → `metadata.ls_model_name`. (`langgraph_engine.py`)
|
|
- **Prompt extraction** tries 5 locations: `data.messages` → `data.input.messages` → `data.input` → `data.kwargs.messages` → raw dump. (`langgraph_engine.py`)
|
|
- **ReactFlow nodes are incremental** — never rebuilt from scratch. `useRef(Set)` deduplication prevents duplicates. (`AgentGraph.tsx`)
|
|
- **useAgentStream uses statusRef** to avoid stale closures in WebSocket callbacks. Status is not a useCallback dependency. (`useAgentStream.ts`)
|
|
- **Pipeline recursion limit** — `run_pipeline()` must pass `config={"recursion_limit": propagator.max_recur_limit}` to `astream_events()`. Default LangGraph limit of 25 is too low for debate+risk cycles. (`langgraph_engine.py`)
|
|
- **Portfolio field mapping** — `/latest` endpoint maps backend model fields to frontend shape: `shares` → `quantity`, `portfolio_id` → `id`, `cash` → `cash_balance`, `trade_date` → `executed_at`. Computed fields (`market_value`, `unrealized_pnl`) included from runtime properties. (`portfolios.py`)
|
|
- **Dashboard drawer has 2 modes** — `'event'` (single event detail from terminal click) and `'node'` (all events for a graph node from ReactFlow click). (`Dashboard.tsx`)
|
|
- **Run buttons track activeRunType** — only the triggered button spins, others disabled during run. (`Dashboard.tsx`)
|
|
- **Collapsible param panel** — date/ticker/portfolio_id with per-run-type validation. (`Dashboard.tsx`)
|