[Bug] Duplicate backtest input reruns the same cell within one sweep

Version

0.5.1 (main at 35543d0)

How you run it

Backtest (run_backtest or tradingagents backtest)

LLM provider and models

Any; source-level finding, no live LLM call

Ticker and analysis date

AAPL, 2026-09-01

What happened, and what you expected

In run_backtest, done is initialized once from entries already in the decision log. After a successful graph.propagate(ticker, date, ...), the code increments cells_run but never adds (ticker, date) to done (backtest.py lines 151–165).

If either input list contains a duplicate, the same cell is analyzed again within that sweep, including LLM and vendor calls. The decision-log idempotency guard prevents a duplicate entry only at the end of the second analysis; it cannot prevent the extra work or the inflated cells_run count.

Expected: each unique ticker/date cell is propagated once per successful sweep, and later duplicates count as skipped.

Suggested fix: add done.add((ticker, date)) immediately after successful graph.propagate. Keep it outside the exception path if a failed duplicate is meant to be retried. Add a regression test with duplicate input that asserts one propagation and one skipped cell, plus a failure case documenting retry behavior. This preserves the existing resume-from-log behavior.

This is a deterministic source-level finding on version 0.5.1; I did not run a paid LLM-backed backtest.

Steps or code to reproduce

With an empty run log, call:


run_backtest(
    tickers=["AAPL", "AAPL"],
    dates=["2026-09-01"],
    config=config,
)


Or mock `TradingAgentsGraph.propagate` and count calls without API cost. Expected: 1 propagation, 1 skipped cell. Current loop: 2 propagations and `cells_run == 2`. Duplicating the date instead has the same effect.

Error output

No exception is raised. The symptom is repeated model/vendor work and an inflated cells_run count.