Bug: TradingAgents hangs with post-7/31 deepseek-v4-flash — long reasoning chains exhaust token budget / trip gateway idle timeout #1204

Bug: TradingAgents hangs indefinitely with DeepSeek V4 Flash (post-7/31 stable) — long reasoning chains exhaust token budget / trip gateway idle timeout

Environment

• Branch: dev (15caeb9, 2026-08-02)
• Provider: openai (OpenAI-compatible) with TRADINGAGENTS_LLM_BACKEND_URL=https://opencode.ai/zen/go/v1
• Models: deepseek-v4-flash (both quick & deep think; the 7/31 stable release)
• Ticker: 159566.SZ (A-share ETF), all 4 analysts + full pipeline

Symptom

The run hangs indefinitely during the analyst/data-collection phase: process sits in wait_woken with near-zero CPU, the log stops updating after the initial data-fetch errors (StockTwits/Reddit/Polymarket are all non-fatal, expected for A-shares), and no report directory is ever created. Reproduced 3/3 times (pro+flash, flash-only, strict .env flash). A similar run on 2026-07-30 with the previous (pre-7/31) v4-flash completed fine.

Root-cause evidence (direct API tests against the same backend)

Analysis

1. The 7/31 stable v4-flash (post-training) produces much longer reasoning chains — 500+ tokens for complex tasks vs ~26 for trivial ones.
2. _PASSTHROUGH_KWARGS in llm_clients/openai_client.py does not include max_tokens, so TradingAgents sends no upper bound.
3. With no bound, the backend lets the reasoning chain run long, and a single generation exceeds the ~3 min idle timeout of the gateway (the same timeout documented in feat(llm): add OpenCode Go provider #782 for opencode-go) → connection dropped → process hangs.
4. With a small bound (wherever a default applies), the reasoning chain consumes the entire budget and content comes back empty → downstream agents get nothing to parse.

Proposed fix

1. Inject a sensible default max_tokens (e.g. 8192) into the OpenAI-compatible client calls, overridable via a new TRADINGAGENTS_MAX_TOKENS env var (mirroring the existing _ENV_OVERRIDES pattern).
2. Consider enabling streaming for the OpenAI-compatible path (as feat(llm): add OpenCode Go provider #782 did for opencode-go), including reasoning_content capture from stream deltas, so long generations keep the socket alive.

Related

• feat(llm): add OpenCode Go provider #782 (opencode-go provider: streaming required, gateway 3-min idle timeout, reasoning_content delta capture)
• fix(llm): map OpenRouter DeepSeek models to DeepSeek capabilities #1199 (DeepSeek thinking models + tool_choice → very slow responses, adjacent class of problem)