TradingAgents v0.4.0 Ships Point-in-Time Data Fixes and CLI Resume

The release continues a pattern of incremental hardening that began with v0.2.0 multi-provider support and accelerated after the Trading-R1 report. Each subsequent tag (0.2.2 through 0.4.0) targets a concrete failure mode: Alpha Vantage leakage, graph-router crashes, path traversal, and now temporal leakage in macroeconomic and sentiment streams. The changelog records six distinct data-contract and resume fixes in four months.

Structured output agents and LangGraph checkpoints produce an auditable decision trace that prior single-model baselines lacked. Evidence appears in the verified data-access contract and CI gate introduced at v0.3.0; these gates reject runs that violate point-in-time boundaries. Operational impact is immediate for backtesting: any replay that previously ingested future FRED releases or Reddit posts will now be rejected or masked.

The framework remains research-only. Its multi-agent debate layer (bullish/bearish researchers plus risk team) still depends on temperature and data quality, yet the added retry budget and Bedrock auth expand the set of reproducible environments. Next milestone is integration of persistent memory across sessions, required before any production-grade risk limits can be attached.

Mobile-agent VM patterns are orthogonal; TradingAgents runs server-side. Its checkpoint format could however be ported to on-device inference runtimes once quantized 5.x models fit within mobile memory envelopes.