The earlier Phase 1-4 recovery left one unique worker-1 slice unrecovered: provenance extraction logic was still duplicated in the runner and the full-state log path still dropped the structured research fields. This change centralizes provenance extraction in agent state helpers, reuses it from the LLM runner, and writes the same structured fields into TradingAgents full-state logs with focused regression tests.\n\nConstraint: Preserve the existing debate-string output shape while making provenance reuse consistent across runner and state-log surfaces\nRejected: Cherry-pick worker-1 auto-checkpoint wholesale | it mixed duplicate A/B files and uv.lock churn with the useful provenance helper changes\nConfidence: high\nScope-risk: narrow\nDirective: Keep research provenance extraction centralized; new consumers should call the helper instead of re-listing field names by hand\nTested: python -m pytest -q tradingagents/tests/test_research_guard.py orchestrator/tests/test_trading_graph_config.py orchestrator/tests/test_llm_runner.py orchestrator/tests/test_profile_stage_chain.py orchestrator/tests/test_profile_ab.py orchestrator/tests/test_contract_v1alpha1.py orchestrator/tests/test_live_mode.py\nTested: python -m compileall tradingagents/agents/utils/agent_states.py tradingagents/graph/trading_graph.py orchestrator/llm_runner.py orchestrator/tests/test_trading_graph_config.py tradingagents/tests/test_research_guard.py\nNot-tested: Live-provider end-to-end analysis run that emits a new full_states_log file
Research provenance now rides with the debate state, cache metadata, live payloads, and trace dumps so degraded research no longer masquerades as a normal sample. Bull/Bear/Manager nodes also return explicit guarded fallbacks on timeout or exception, which gives the graph a real node budget boundary without rewriting the bull/bear output shape or removing debate.\n\nConstraint: Must preserve bull/bear debate structure and output shape while adding provenance and node guards\nRejected: Skip bull/bear debate in compact mode | would trade away analysis quality before A/B evidence exists\nConfidence: high\nScope-risk: moderate\nReversibility: clean\nDirective: Treat research_status and data_quality as rollout gates; do not collapse degraded research back into normal success samples\nTested: python -m pytest tradingagents/tests/test_research_guard.py orchestrator/tests/test_llm_runner.py orchestrator/tests/test_live_mode.py web_dashboard/backend/tests/test_executors.py web_dashboard/backend/tests/test_services_migration.py web_dashboard/backend/tests/test_api_smoke.py -q; python -m compileall tradingagents/graph/setup.py tradingagents/agents/utils/agent_states.py tradingagents/graph/propagation.py orchestrator/llm_runner.py orchestrator/live_mode.py orchestrator/profile_stage_chain.py; python orchestrator/profile_stage_chain.py --ticker 600519.SS --date 2026-04-10 --provider anthropic --model MiniMax-M2.7-highspeed --base-url https://api.minimaxi.com/anthropic --selected-analysts market --analysis-prompt-style compact --timeout 45 --max-retries 0 --overall-timeout 120 --dump-raw-on-failure\nNot-tested: Full successful live-provider completion through Portfolio Manager after the post-research connection failure
This change set introduces a versioned result contract, shared config schema/loading, provider/data adapter seams, and a no-strategy application-service skeleton so the current research graph, orchestrator layer, and dashboard backend stop drifting further apart. It also keeps the earlier MiniMax compatibility and compact-prompt work aligned with the new contract shape and extends regression coverage so degradation, fallback, and service migration remain testable during the next phases.
Constraint: Must preserve existing FastAPI entrypoints and fallback behavior while introducing an application-service seam
Constraint: Must not turn application service into a new strategy or learning layer
Rejected: Full backend rewrite to service-only execution now | too risky before contract and fallback paths stabilize
Rejected: Leave provider/data/config logic distributed across scripts and endpoints | continues boundary drift and weakens verification
Confidence: high
Scope-risk: broad
Directive: Keep future application-service changes orchestration-only; move any scoring, signal fusion, or learning logic to orchestrator or tradingagents instead
Tested: python -m compileall orchestrator tradingagents web_dashboard/backend
Tested: python -m pytest orchestrator/tests/test_signals.py orchestrator/tests/test_llm_runner.py orchestrator/tests/test_quant_runner.py orchestrator/tests/test_contract_v1alpha1.py orchestrator/tests/test_application_service.py orchestrator/tests/test_provider_adapter.py web_dashboard/backend/tests/test_main_api.py web_dashboard/backend/tests/test_portfolio_api.py web_dashboard/backend/tests/test_api_smoke.py web_dashboard/backend/tests/test_services_migration.py -q
Not-tested: live MiniMax/provider execution against external services
Not-tested: full dashboard/manual websocket flow against a running frontend
Not-tested: omx team runtime end-to-end in the primary workspace
Add effort parameter (high/medium/low) for Claude 4.5+ and 4.6 models,
consistent with OpenAI reasoning_effort and Google thinking_level.
Also add content normalization for Anthropic responses.
Prevents UnicodeEncodeError on Windows where the default encoding
(cp1252/gbk) cannot handle Unicode characters in LLM output.
Closes#77, closes#114, closes#126, closes#215, closes#332
- Add StatsCallbackHandler for tracking LLM calls, tool calls, and tokens
- Integrate callbacks into TradingAgentsGraph and all LLM clients
- Dynamic agent/report counts based on selected analysts
- Fix report completion counting (tied to agent completion)
陈少杰
Yijia-Xiao
makk9
luohy15
mirza-samad-ahmed-baig
Edward Sun
maxer137