feat(proxy-clients): pydantic-ai / mem0 / graphiti integration layer

[RainGodPRC]

What

Three complementary subsystems wired through a shared proxy-client layer. All opt-in via environment variables — zero regression when disabled (the existing LangGraph pipeline is unchanged unless TRADINGAGENTS_USE_PYDANTIC_AI=1 / TRADINGAGENTS_MEM0_ENABLED=1 are set).

Shared layer (tradingagents/llm_clients/proxy_clients.py)

• resolve_gateway_token() — extracts the cc-switch Bearer token from env or a running Claude Code process (the proxy at 127.0.0.1:15721 rejects x-api-key, requires Authorization: Bearer)
• make_anthropic_client() — auth_token-based (Bearer, not x-api-key)
• Local HuggingFace embedder — the proxy serves no /v1/embeddings; mem0 + graphiti both use this local fallback
• Factory helpers: make_pydantic_ai_agent(), make_mem0_memory(), make_graphiti()

pydantic-ai backend (structured.py)

Optional path in invoke_structured_or_freetext, enabled by TRADINGAGENTS_USE_PYDANTIC_AI=1. Adds a validation-retry loop (the model is prompted to correct malformed output) that the legacy one-shot fallback lacked — preserves structured signals (BUY/SELL/HOLD, rating) that currently get lost when structured output fails and silently degrades to free text.

Trader, Portfolio Manager, Research Manager pass schema= + instructions= so they benefit when the flag is on.

mem0 decision memory (mem0_bridge.py + trading_graph._run_graph)

Mem0Bridge provides remember_decision() / recall_relevant(). Wired into _run_graph: semantic recall of past decisions is appended to past_context before the pipeline runs; the completed decision is stored after. Cross-ticker recall ("I've seen this Fed-pivot pattern before") complements the ticker-matched markdown log. Enabled by TRADINGAGENTS_MEM0_ENABLED=1.

graphiti Winterfell import (scripts/import_winterfell_to_graphiti.py)

Imports Decisions/Collisions/Analysis/Insights Markdown as temporal episodes, using frontmatter date/created as reference_time. Monkey-patches graphiti's search functions to skip Kuzu's broken fulltext/vector indices (Kuzu is deprecated in graphiti for this reason) — ingestion works on Kuzu, but full dedup/temporal invalidation requires Neo4j/FalkorDB.

Tests

• 3 smoke tests (pydantic-ai, mem0, graphiti) — verify proxy connectivity
• test_trader_pydantic_ai_integration.py — Trader node produces structured proposal via pydantic-ai
• test_mem0_bridge_integration.py — cross-ticker semantic recall
• test_e2e_integration.py — full loop: graphiti ingest → pydantic-ai decision → mem0 recall

Verification

All tests pass against the local cc-switch proxy. No regression in the existing 492-test suite (imports verified, flags default off).

Known limitations

• graphiti on Kuzu has no temporal invalidation (fulltext index bug); full capability needs Neo4j/FalkorDB
• Token extraction depends on a running Claude Code process (ccs- gateway token)
• trading_graph.py also carries a pre-existing _fetch_returns signature refinement that was in the working tree; it's logically intertwined with the mem0 bridge in _run_graph

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.