- a saved report also includes complete_report.html: the report as one self-contained page with its sections listed beside the text, readable in a browser, on a phone and in print
- answering the save question at the prompt also asks about the page and offers to open it in a local browser window; --no-html and save_reports(..., html=False) skip it
- the Tauric Research logo leads the header at the top left, and a line at the foot credits TradingAgents and Tauric Research
- the page loads nothing and runs nothing (inline styles, a content-security policy); report text is rendered without raw HTML or images
- markdown-it-py is declared as a dependency
- quick_think_provider / deep_think_provider and their backend URLs (with TRADINGAGENTS_* variables) choose a tier's provider; unset, a tier uses llm_provider and backend_url
- a tier on another provider gets that provider's own settings and endpoint, and the run settings and report header name each tier's provider
- the CLI checks each tier provider's model variable and key before the run
- a Memory Log step runs beside the analysts, settles every ticker's due decisions and loads the lessons known by the trade date for the Portfolio Manager; the run no longer settles before it starts
- a settlement failure is logged and noted in the report's header instead of ending the run
- the checkpoint signature records the new layout
- requires-python 3.11 and pandas 3.0.6; the CI matrix runs Python 3.11 to 3.14 and ruff targets 3.11
- datetime.UTC in place of timezone.utc, and the rating, action, band, analyst and asset enums are StrEnum, which formats a member as its value
- backtest --analysts accepts sentiment and is checked against --asset-type before the sweep
- without --analysts, a crypto sweep runs the analysts crypto allows, not fundamentals
- run_backtest takes a progress(done, total, ticker, date) callback; tradingagents backtest prints [n/total] per cell and ends with the --run-id that continues or settles the sweep
- record_decision writes the run's state log, so CLI runs save full_states_log like propagate() does
- with both round counts from the environment, the CLI no longer says a chosen research depth was overridden
- an unknown --analysts name lists sentiment, not social
- each analyst is a graph of its own (model and tools on a private message history) that returns only its report; all start together and the research debate waits for every report
- the message-clearing nodes are gone; a checkpoint saved by the sequential layout starts fresh
- TradingAgentsGraph.stream_run streams the analysts' messages for debug mode and the CLI, whose status and timing now track the analysts side by side
- --ticker, --date and --analysts answer their steps, checked as the prompts check them; --save/--no-save and --show/--no-show answer the questions after the run
- with no terminal, the run stops before it starts and names every flag or TRADINGAGENTS_* variable still needed; a missing API key and an announcement no longer wait for input
- README shows an unattended run
- TradingAgentsGraph.run_settings(): version, provider, models, analysts, debate rounds, language and vendors; no endpoints, keys or paths
- complete_report.md opens with the analysis date and those settings; the saved state log carries them as run_settings
- the Research Manager's plan and the Portfolio Manager's decision live only in investment_plan and final_trade_decision; the debate states no longer copy them as judge_decision
- the saved state log writes the Trader's plan as trader_investment_plan, its state key
- the report tree, the final display and the live CLI read the decisions from those fields
- docstrings, comments, messages and the README section use "memory log", matching TradingMemoryLog and memory_log_path
- the write, read and settle paths are named for what they do, not by design phase
- the typed rating is the state's final_rating; propagate returns it, and the memory log tag, the state log and the CLI review check read it
- the decision text is parsed only when the Portfolio Manager answered in free text
- TradingAgentsGraph.process_signal is removed
- selections.py asks what to run; run.py builds the graph, streams it to the live view and saves the report
- main.py keeps the Typer app and its two commands
- display.py holds the message buffer, layout, status tables and report panels, the analyst wall-time tracker (CLI-only, from graph/analyst_execution) and the one Console
- utils.py is renamed prompts.py, which is what it holds; its analyst list is ANALYST_CHOICES, apart from display's ANALYST_ORDER
- get_initial_analyst_node, a one-line helper with one caller, is inlined
- the wall-time tracker tests sit with the other display tests, and tests import cli.prompts as prompts
- tools.py: the analysts' data tools, previously seven modules under utils
- context.py (was agent_utils, without its tool re-exports), state.py, rating.py and structured.py sit beside schemas.py
- every role package has an __init__
- interface -> router; symbol_utils -> symbols, which also takes safe_ticker_component
- utils is split: get_current_date to date_window, the HTTP helpers to net
- dataflows imports are absolute; the NoMarketDataError re-export from symbols is gone
- is_yahoo_safe, has_checkpoint, get_wall_times, the graph's curr_state and the project_dir config key
- the CLI's final_report and current_agent state, its duplicate get_analysis_date and the save_report_to_disk wrapper
- the dotenv import guard (a hard dependency) and a warning filter for langgraph-checkpoint 4.0.3
- backtest takes --run-id, so an interrupted sweep continues
- an inverted date range or an empty ticker list is an error, not a clean zero-cell run
- a setup failure in a sweep reports one line instead of a traceback
- the remembered endpoint URL is offered back
- a round count set in the environment says it overrode the chosen research depth
- the run directory validates the ticker, and the report saves under results_dir
- the run says whether it resumed a saved run or started fresh
- a non-blank message is text, whatever it would mean as a Python literal
- the live layout renders on the alternate screen, so a tall layout does not scroll
- a run with no readable rating says so instead of looking like a normal result
- the state log keeps non-ASCII readable (#1081)
- analysts, depth, provider, models and language prefill; prompts still shown
- values no longer offered by the current catalog are dropped
- environment variables keep skipping their step
- analysis stays the bare invocation; backtest runs a ticker and date grid
- selections name the models as the config does
- the pending note prints only when a cell is pending
- PortfolioContext input, rendered once into state and read by the trader, risk and portfolio agents
- --portfolio JSON file on the CLI
- a run without it is never treated as a flat book
- the checkpoint signature keys on the portfolio
- the key prompt created .env with the default umask, typically readable by
other local users
- create it 0600 and tighten an existing file before writing the key; a
read-only file is still updated
- on resume, the CLI and propagate re-passed the initial state to a thread with
an existing checkpoint; nodes do not re-run, but the message reducer appended
the initial messages again, duplicating them in the resumed state
- feed None on resume (checkpoint_input) so LangGraph continues the interrupted
run, and wrap the CLI stream in try/finally so the checkpointer tears down even
if the stream raises
- correct the _fetch_returns docstring to the 4-tuple return #1249
- checkpoint setup lived only inside propagate(); the CLI streamed the
checkpointer-less graph with no thread_id, so --checkpoint neither saved nor
resumed a run
- extract the lifecycle into reusable begin_checkpoint / end_checkpoint /
clear_checkpoint_on_success (checkpoint_scope wraps them for propagate) and use
them around the CLI stream #1249
- Windows terminals without a console buffer raised NoConsoleScreenBufferError
before the first prompt, surfacing a raw traceback with no guidance
- gate the Windows-only import on sys.platform so a broken prompt_toolkit still
surfaces there, and the handler stays inert on other platforms #1138
The per-section markdown report tree was written only by the CLI, so programmatic
(TradingAgentsGraph) runs produced no saved reports.
- Extract the writer into tradingagents/reporting.write_report_tree.
- The CLI's save_report_to_disk delegates to it (no behavior change).
- Add TradingAgentsGraph.save_reports(final_state, ticker) so headless/API callers
get the same report tree, defaulting under results_dir.
The knob was accepted but inert — analysts run strictly sequentially and the
value was never used. Remove it rather than ship a misleading config key.
Parallel analyst execution is tracked for v0.3 (#634/#671/#487).
Interactive selections and flag defaults overrode TRADINGAGENTS_* env vars.
Rule: an explicit env value or CLI flag wins; otherwise the env-applied
default is kept.
- Research depth: skip the prompt when both round-count env vars are set, and
stop overwriting them (#977).
- Checkpoint: --checkpoint/--no-checkpoint is tri-state; omitting it keeps
TRADINGAGENTS_CHECKPOINT_ENABLED (#976).
- Docker ollama: use TRADINGAGENTS_LLM_PROVIDER + OLLAMA_BASE_URL, not a bare
LLM_PROVIDER the overlay never reads (#975).
- Reasoning/thinking knobs: settable via env; the prompt is skipped when set.
- Effort gating: forward effort only to models that accept it (Anthropic
Opus 4.5+/Sonnet 4.6+, OpenAI reasoning models); drop it elsewhere.
- Boolean env values: raise a named error on invalid input instead of
silently becoming False.
Label each OpenRouter model prompt by mode (quick/deep) like the other
providers, so the two consecutive selections are distinguishable. Populate the
dropdown with the newest models from mainstream chat providers rather than the
universal-newest (which surfaced niche/experimental releases); Custom ID still
reaches anything. Cancelled required prompts now exit cleanly instead of
crashing, and the output-language prompt falls back to English.
Clear the deferred full-repo lint backlog so the whole tree passes the strict
ruff select (E,W,F,I,B,UP,C4,SIM). Mechanical fixes dominate: import sorting,
pep585/604 annotations, dropped dead imports, and whitespace. The few semantic
changes are behavior-preserving: declare __all__ on the agent_utils and
alpha_vantage re-export hubs; expand 'from x import *' to explicit names; use
immutable tuple defaults instead of mutable list defaults; contextlib.suppress
for try/except/pass; and narrow an over-broad assertRaises.
Bedrock uses the Converse API (langchain-aws) and the AWS credential chain, so
it has its own client like Anthropic/Google rather than the OpenAI-compatible
registry. langchain-aws is an optional dependency (pip install ".[bedrock]"),
lazy-imported with a clear install hint; importing the package never requires
it. The model name is a Bedrock model ID / inference profile ID.
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).
The CLI validated, normalized, and classified tickers with its own logic that
diverged from the data layer: it rejected '=' symbols like GC=F (#980),
classified BTCUSD as a stock (#981), and accepted unpriceable BTC-USDT (#982).
Route the CLI through normalize_symbol (now mapping USDT/USDC crypto quotes to
Yahoo's -USD pair), so validation, classification, and pricing agree.
Setting the LLM env vars now skips the matching CLI selection step and uses
the value, so OpenAI-compatible endpoints (opencode, LM Studio, etc.) and
unattended runs work without prompting. Unset vars are chosen interactively
as before.
TRADINGAGENTS_LLM_PROVIDER -> skips provider step (still verifies API key)
TRADINGAGENTS_LLM_BACKEND_URL -> custom endpoint (else provider default)
TRADINGAGENTS_DEEP_THINK_LLM / _QUICK_THINK_LLM -> skips model step
TRADINGAGENTS_OUTPUT_LANGUAGE -> skips language step
Builds on the existing TRADINGAGENTS_* config overrides (which already feed
DEFAULT_CONFIG); this wires the CLI to honor them instead of re-prompting.
Agents had no ground-truth ticker→company mapping, so the market analyst
could pattern-match a price chart to the wrong company (e.g. TOTDY read as
"TotalEnergies"), and every downstream agent inherited the bad framing.
Resolve identity once at run start via a cached, fail-open yfinance lookup
and inject company/sector/exchange into the shared instrument context that
all twelve agents consume, with an explicit do-not-substitute instruction.
Resolution runs on both the propagate() and CLI entry points.
Also replaces the bare "Continue" message-clear placeholder, which some
OpenAI-compatible providers interpreted as the user task, with a
context-anchored placeholder carrying the resolved identity and date.
#814#888
Yijia-Xiao
Yijia Xiao