- an analyst's TOOLS tuple is both what it is offered and what its tool node runs
- one routing function replaces four copies of should_continue_<analyst>
- the sentiment analyst has no tools, so the unreachable social tool node is gone
- a graph reused across a backtest grid held each run's complete state for its whole life
- the state log is written from the run's own state; the ticker attribute it read is removed
- the Yahoo vendor owns get_company_profile and get_closes; agent_utils and the graph no longer import yfinance
- a test keeps vendor libraries inside dataflows
- every analyst calls its tools through the router, offline; the decision is parsed and logged
- free-text and structured-output paths, and resume from a checkpoint
- remove the unused mock_llm_client fixture
- propagate and settle_pending bind the graph's config for the length of the run
- a graph built later, or running concurrently, no longer changes another graph's vendors
- annual columns are the periods an annual report covers, so 10-Q balances and twelve-month totals no longer read as fiscal years
- a value is still the latest filing of any form, so a later recast counts from its filing date
- point-in-time integrity across every dated path, and a vendor failure reported as a vendor failure
- SEC EDGAR fundamentals served as filed
- backtesting over a ticker and date grid, and the caller's portfolio as run input
- current model lineups across every provider
- yfinance returned its errors as text, which the router counted as an answer, so the chain stopped and the text reached the analyst
- an empty result is checked against the vendor being reachable, so an outage is not reported as a company with no data
- a chain where every vendor is unavailable says so instead of ending the run
- xAI: Grok 4.6 and 4.5; DeepSeek: deepseek-flash (V4.1); Qwen: the 3.8 series; Mistral: the current dated snapshots
- every provider can name a model the picker does not list
- retired IDs stay valid, so existing configs run without a warning
- a price written as a range or a hedge is dropped like any other unusable value
- a field the model did not give is named as not provided, rather than omitted
- the custom model option does not exist for every provider; name any model ID instead
- alpha is measured against the regional benchmark, not always SPY
- the environment overrides a fixed set of config keys
- list the providers the picker offers, and the macro data key
- drop the example call to a method that was removed
- 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
- rows carry the transaction date; a Form 4 is filed up to two business days later
- an indicator that could not be read is unavailable, not a blank value for the day
- gap filling keeps indicators on a continuous series, but put the previous session's open, high and low under an unsettled bar's date
- load_ohlcv takes fill_gaps, and the snapshot reads the frame as reported
- the fundamentals brief reached the model as a Python tuple
- analysts no longer emit a trade call that nothing reads
- a report that was not produced says so instead of appearing as a blank section
- holding_period_days sets it; the price request covers the calendar span those trading days occupy
- reflection states the window it judges, so a short one does not read as a failed thesis
- a Sell that fell counted as a miss and a Sell that rose as a hit
- Hold claims no direction, so it reports mean alpha and no hit rate
- the report names the window the outcomes were measured over
- 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)
- the field list lived only in the schema, which a provider without structured output never sees
- the trader maps the five-tier recommendation onto its three actions
- the labelled rating decides, whatever dash separates it, and a scale the model echoed is not one
- prose naming several ratings is reviewed rather than read as the first word in the text
- an unreadable decision is tagged REVIEW everywhere instead of a tradeable Hold
- unrated decisions are counted apart from the backtest figures
- statements are limited to facts filed by the run's date, at the values filed then
- quarterly figures are picked by span, so a year to date total is not read as a quarter
- a fourth quarter is reported unavailable rather than derived
- opt in with sec_edgar in the fundamental_data chain; a non-filer falls through
- settling a past decision is per entry; a provider error leaves it pending
- an explicitly configured benchmark ticker is normalized like any other symbol (#1075)
Yijia-Xiao
David Arias, CFA
Yijia Xiao