v0.4.0
TradingAgents v0.4.0
Added 2
- Add configurable output-token cap via TRADINGAGENTS_MAX_TOKENS environment variable to bound models that emit unbounded reasoning
- Add GPT-5.6 family models (gpt-5.6, gpt-5.6-terra, gpt-5.6-luna) and GLM-5.3 models (glm-5.3, glm-5.3-flash)
Changed 3
- The Trader now receives the technical market report so entry and stop levels anchor to real price structure
- DeepSeek routed through OpenRouter now reuses native DeepSeek capabilities
- Default models are now gpt-5.6 for deep reasoning and gpt-5.6-luna for quick execution
Fixed 7
- FRED macro requests now pin the data vintage to the as-of date so backtests no longer see later revisions
- Social sentiment from StockTwits and Reddit is trimmed to the analysis window instead of showing current chatter
- Decision-log memory records when each outcome became known and only injects lessons resolved by the trade date
- The latest OHLCV bar is no longer silently dropped when its close is NaN and dates are normalized DST and non-US-market safe
- An unparseable Portfolio Manager rating surfaces a REVIEW sentinel instead of a tradeable Hold
- Debate openers no longer fabricate the opponent's argument when none has been made
- Checkpoint flag now actually resumes on the CLI path and a resume no longer duplicates messages
From TradingAgents
Summary
Look-ahead and point-in-time fixes across the data and memory layers, clearer decision signals, working CLI checkpoint resume, and the GPT-5.6 / GLM-5.3 models.
Point-in-time / look-ahead integrity
- FRED macro requests now pin the data vintage to the as-of date, so a backtest no longer sees later revisions (#1275).
- Social sentiment (StockTwits, Reddit) is trimmed to the analysis window instead of showing today's chatter, via one shared UTC window rule reused by news (#1220).
- Decision-log memory records when each outcome became known and only injects lessons resolved by the trade date (#1251); a decision is no longer settled before its holding window fully trades (#1169).
- The latest OHLCV bar is no longer silently dropped when its close is NaN, and dates are normalized DST- and non-US-market safe (#1201).
More honest decision signals
- An unparseable Portfolio Manager rating surfaces a REVIEW sentinel instead of a tradeable Hold (#1170).
- Debate openers no longer fabricate the opponent's argument when none has been made (#1176).
- The Trader now receives the technical market report, so entry/stop levels anchor to real price structure (#1167).
Execution robustness
--checkpointactually resumes on the CLI path, and a resume no longer duplicates messages (#1249).- A configurable output-token cap (
TRADINGAGENTS_MAX_TOKENS) bounds a model that emits unbounded reasoning (#1204). - DeepSeek routed through OpenRouter (
deepseek/<id>) now reuses the native DeepSeek capabilities (#1199).
Models
- Added the GPT-5.6 family (
gpt-5.6/gpt-5.6-terra/gpt-5.6-luna) and GLM-5.3 (glm-5.3,glm-5.3-flash); the defaults are nowgpt-5.6(deep) /gpt-5.6-luna(quick).
See CHANGELOG.md for the full list, including contributor acknowledgments.