fix: graceful LLM 404 handling + per-tier fallback model config (#108)
* fix: per-tier fallback LLM for provider 404/policy errors
- tool_runner: catch status_code==404 from chain.invoke(), re-raise as
RuntimeError with actionable message (OpenRouter privacy URL + env var hint)
- langgraph_engine: wrap astream_events in try/except, detect policy errors
and re-raise with model/provider context
- langgraph_engine: _run_one_ticker distinguishes policy 404s (logger.error,
no traceback) from real bugs (logger.exception with traceback); if fallback
is configured, rebuilds pipeline with fallback model tier and retries
- langgraph_engine: add _is_policy_error() and _build_fallback_config() helpers
- default_config: add quick/mid/deep_think_fallback_llm + _provider keys
(TRADINGAGENTS_QUICK_THINK_FALLBACK_LLM etc.)
- .env.example: document new fallback env vars
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs: ADR 017 LLM policy fallback, correct ADR 016 findings, update CLAUDE.md
- docs/agent/decisions/017: add ADR for per-tier LLM fallback design decision
- docs/agent/decisions/016: correct 3 inaccurate review findings — list_pm_decisions
ObjectId projection, created_at datetime type, and base_dir pointer handling are
all already correctly implemented in PR#106
- CLAUDE.md: add Per-Tier Fallback LLM section and _is_policy_error critical pattern
- CURRENT_STATE.md: update milestone and recent progress for PR#106/107/108 merges
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
2145b04318