klau1011
diff --git a/‎docs/improvements-plan.md‎
Lines changed: 14 additions & 74 deletions b/‎docs/improvements-plan.md‎
Lines changed: 14 additions & 74 deletions
diff --git a/‎tests/test_memory_log.py‎
Lines changed: 133 additions & 6 deletions b/‎tests/test_memory_log.py‎
Lines changed: 133 additions & 6 deletions
diff --git a/‎tests/test_model_routing.py‎
Lines changed: 109 additions & 0 deletions b/‎tests/test_model_routing.py‎
Lines changed: 109 additions & 0 deletions
@@ -17,8 +17,17 @@ shortest working diff, reuse existing utilities, no new dependencies.
   `<results_dir>/web_runs/`, API-key preflight (400), sanitized `ErrorEvent`,
   opaque `report_folder` references + "View report" link, History staleness
   fixes, rehydrated-run handling in `useRunStream`.
-
-**Remaining: PRs 3–6 below.** Each is a self-contained change; suggested
+- **PR 4 — Model routing + adaptive risk debate** (`0517829`):
+  `agent_llm_map` per-stage quick/deep assignment (validated in
+  `GraphSetup`); `adaptive_extra_rounds` extends the risk debate past the
+  base cap only while `direction(parse_rating(...))` disagrees between the
+  Research Manager's and Trader's plans. Both default to today's behavior.
+- **PR 5 — Memory & reflection** (`4b0ef75`): cross-ticker
+  `get_past_context` selection now top-3 by |alpha| (`_parse_pct`,
+  tie-break recency, unparseable last); guarded past-lessons block injected
+  into bull/bear researchers, trader, and the three risk debators.
+
+**Remaining: PRs 3 and 6 below.** Each is a self-contained change; suggested
 order preserved. Per-PR verification: `pytest -m "unit or smoke"` (conftest
 stubs API keys), plus `npm run build` in `web/frontend` when frontend files
 change.
@@ -56,78 +65,6 @@ Verify: `event_to_dict(StatsEvent)` round-trip + stub-graph stats-emission
 test in `tests/test_runner_events.py` (`_StubGraph` pattern);
 `python -c "import cli.main"` smoke.
 
-## PR 4 — Model routing + adaptive risk debate
-
-Two opt-in config knobs, both defaulting to today's exact behavior.
-
-**Per-stage model assignment**
-
-1. `tradingagents/default_config.py`: add `"agent_llm_map": {}`. Stages:
-   `analysts`, `researchers`, `research_manager`, `trader`, `risk_analysts`,
-   `portfolio_manager`, `investor_briefing`; values `"quick"` / `"deep"`.
-   Empty dict = current behavior (analysts + trader on quick, managers on
-   deep). No env override (the `_coerce` machinery doesn't do dicts).
-2. `tradingagents/graph/setup.py`: `GraphSetup.__init__` gains
-   `llm_map: dict | None = None` plus a `_llm(stage, default)` helper
-   (raise `ValueError` on unknown stage/value); replace the hardcoded
-   quick/deep picks (~lines 61–78).
-3. `tradingagents/graph/trading_graph.py` (~line 116): pass
-   `self.config.get("agent_llm_map")`.
-
-No UI exposure; reachable via config / `RunnerConfig.extra_config`.
-
-**Adaptive risk debate**
-
-Cheap deterministic disagreement signal, verified present in state: by the
-time `should_continue_risk_analysis` runs, `state["investment_plan"]`
-(Research Manager) and `state["trader_investment_plan"]` (Trader) both carry
-parseable ratings (`parse_rating` in `agents/utils/rating.py`; `direction()`
-landed in PR 1).
-
-1. `tradingagents/graph/conditional_logic.py`: `__init__` gains
-   `adaptive_extra_rounds=0`. In `should_continue_risk_analysis`, past the
-   base cap but under `3 * (max_risk_discuss_rounds + adaptive_extra_rounds)`,
-   continue only if
-   `direction(parse_rating(investment_plan)) != direction(parse_rating(trader_investment_plan))`.
-2. `default_config.py`: `"adaptive_extra_rounds": 0`;
-   `trading_graph.py` (~lines 112–115) passes it through.
-
-Not doing: adaptive bull/bear debate (they disagree by construction),
-LLM-judged disagreement scoring.
-
-Verify: pure-state `ConditionalLogic` unit tests (no LLM): agreement → stop
-at base cap; disagreement → exactly N extra cycles then stop; default 0 →
-identical to today.
-
-## PR 5 — Memory & reflection
-
-**Relevance-based retrieval.** `tradingagents/agents/utils/memory.py`
-`get_past_context` (~line 70): same-ticker selection stays recency (recent
-context on the same name is genuinely most relevant); cross-ticker selection
-changes from "last 3" to "top 3 by |alpha|" (biggest realized wins/losses
-carry the most instructive reflections), tie-break recency. Needs a small
-`_parse_pct("+3.2%") -> float | None` helper since entries store alpha as a
-formatted string (written in `update_with_outcome`, ~line 122). This is a
-magnitude heuristic; the upgrade path is tag/embedding matching. Not doing:
-embeddings, LLM relevance scoring, regime tags (on-disk format change).
-
-**Wider injection.** Copy the guarded block pattern from
-`agents/managers/portfolio_manager.py:35-40` ("Past lessons" section,
-rendered only when `past_context` is non-empty) into:
-`researchers/bull_researcher.py`, `researchers/bear_researcher.py`,
-`trader/trader.py`, `risk_mgmt/aggressive_debator.py`,
-`risk_mgmt/conservative_debator.py`, `risk_mgmt/neutral_debator.py`.
-`past_context` is already in graph state (`graph/propagation.py:40`); all six
-nodes already receive `state`. Accepted token cost: the string is bounded by
-`n_same`/`n_cross` and the 300-char cross-ticker truncation.
-
-Verify: `tests/test_memory_log.py` — |alpha| cross-ticker ordering and
-`_parse_pct`; per-agent prompt-capture tests with a fake `llm.invoke`
-(researchers/debators call `llm.invoke(prompt)` directly; trader via the
-fake-structured-LLM pattern in `tests/test_structured_agents.py`). Assert
-the block is present iff `past_context` is set. Measure the behavior change
-with the PR 1 harness before/after.
-
 ## PR 6 — Sizing outputs
 
 Make the Portfolio Manager's sizing real instead of dead prose fields.
@@ -168,3 +105,6 @@ pass-through test in `tests/test_reports_module.py`.
   real use case.
 - Web run persistence is JSON-per-run with no locking (single-process
   server); move to sqlite if multi-worker ever happens.
+- PR 5's behavior change (|alpha| retrieval + wider injection) has not been
+  measured with the PR 1 eval harness yet — run a before/after backtest when
+  spending eval budget next.
@@ -5,9 +5,16 @@
 import pandas as pd
 import pytest
 
+from tradingagents.agents import (
+    create_aggressive_debator,
+    create_bear_researcher,
+    create_bull_researcher,
+    create_conservative_debator,
+    create_neutral_debator,
+)
 from tradingagents.agents.managers.portfolio_manager import create_portfolio_manager
 from tradingagents.agents.schemas import PortfolioDecision, PortfolioRating
-from tradingagents.agents.utils.memory import TradingMemoryLog
+from tradingagents.agents.utils.memory import TradingMemoryLog, _parse_pct
 from tradingagents.graph.propagation import Propagator
 from tradingagents.graph.reflection import Reflector
 from tradingagents.graph.trading_graph import TradingAgentsGraph
@@ -36,10 +43,10 @@ def make_log(tmp_path, filename="trading_memory.md"):
     return TradingMemoryLog(config)
 
 
-def _seed_completed(tmp_path, ticker, date, decision_text, reflection_text, filename="trading_memory.md"):
+def _seed_completed(tmp_path, ticker, date, decision_text, reflection_text, filename="trading_memory.md", alpha="+0.5%"):
     """Write a completed entry directly to file, bypassing the API."""
     entry = (
-        f"[{date} | {ticker} | Buy | +1.0% | +0.5% | 5d]\n\n"
+        f"[{date} | {ticker} | Buy | +1.0% | {alpha} | 5d]\n\n"
         f"DECISION:\n{decision_text}\n\n"
         f"REFLECTION:\n{reflection_text}"
         + _SEP
@@ -263,7 +270,7 @@ def test_get_past_context_cross_ticker(self, tmp_path):
         log = make_log(tmp_path)
         _seed_completed(tmp_path, "AAPL", "2026-01-05", "Buy AAPL — Services growth.", "Correct.")
         ctx = log.get_past_context("NVDA")
-        assert "Recent cross-ticker lessons" in ctx
+        assert "Cross-ticker lessons" in ctx
         assert "Past analyses of NVDA" not in ctx
 
     def test_n_same_limit_respected(self, tmp_path):
@@ -748,8 +755,8 @@ def test_same_ticker_prioritised(self, tmp_path):
         _resolve_entry(log, "AAPL", "2026-01-06", DECISION_SELL, "Overvalued.")
         result = log.get_past_context("NVDA")
         assert "Past analyses of NVDA" in result
-        assert "Recent cross-ticker lessons" in result
-        same_block, cross_block = result.split("Recent cross-ticker lessons")
+        assert "Cross-ticker lessons" in result
+        same_block, cross_block = result.split("Cross-ticker lessons")
         assert "NVDA" in same_block
         assert "AAPL" in cross_block
 
@@ -869,3 +876,123 @@ def test_full_pipeline_no_regression(self, tmp_path):
         assert len(entries) == 1
         assert entries[0]["ticker"] == "NVDA"
         assert entries[0]["pending"] is True
+
+
+# ---------------------------------------------------------------------------
+# Cross-ticker relevance: |alpha| ordering (PR 5)
+# ---------------------------------------------------------------------------
+
+@pytest.mark.unit
+class TestCrossTickerRelevance:
+
+    def test_parse_pct(self):
+        assert _parse_pct("+3.2%") == 3.2
+        assert _parse_pct("-1.5%") == -1.5
+        assert _parse_pct("n/a") is None
+        assert _parse_pct(None) is None
+        assert _parse_pct("") is None
+
+    def test_cross_ticker_ordered_by_abs_alpha(self, tmp_path):
+        log = make_log(tmp_path)
+        _seed_completed(tmp_path, "AAPL", "2026-01-05", "d", "AAPL lesson.", alpha="+1.0%")
+        _seed_completed(tmp_path, "MSFT", "2026-01-06", "d", "MSFT lesson.", alpha="-8.0%")
+        _seed_completed(tmp_path, "TSLA", "2026-01-07", "d", "TSLA lesson.", alpha="+3.0%")
+        _seed_completed(tmp_path, "AMZN", "2026-01-08", "d", "AMZN lesson.", alpha="+0.2%")
+        ctx = log.get_past_context("NVDA")
+        assert "AMZN lesson." not in ctx  # smallest |alpha| drops out of top 3
+        assert (
+            ctx.index("MSFT lesson.")
+            < ctx.index("TSLA lesson.")
+            < ctx.index("AAPL lesson.")
+        )
+
+    def test_tie_break_most_recent_first(self, tmp_path):
+        log = make_log(tmp_path)
+        _seed_completed(tmp_path, "AAPL", "2026-01-05", "d", "Older lesson.", alpha="+2.0%")
+        _seed_completed(tmp_path, "MSFT", "2026-01-06", "d", "Newer lesson.", alpha="-2.0%")
+        ctx = log.get_past_context("NVDA")
+        assert ctx.index("Newer lesson.") < ctx.index("Older lesson.")
+
+    def test_unparseable_alpha_sorts_last(self, tmp_path):
+        log = make_log(tmp_path)
+        _seed_completed(tmp_path, "AAPL", "2026-01-06", "d", "No-alpha lesson.", alpha="n/a")
+        _seed_completed(tmp_path, "MSFT", "2026-01-05", "d", "Small-alpha lesson.", alpha="+0.1%")
+        ctx = log.get_past_context("NVDA")
+        assert ctx.index("Small-alpha lesson.") < ctx.index("No-alpha lesson.")
+
+    def test_same_ticker_stays_recency_ordered(self, tmp_path):
+        log = make_log(tmp_path)
+        _seed_completed(tmp_path, "NVDA", "2026-01-05", "Old decision.", "r", alpha="+9.0%")
+        _seed_completed(tmp_path, "NVDA", "2026-01-06", "New decision.", "r", alpha="+0.1%")
+        ctx = log.get_past_context("NVDA")
+        assert ctx.index("New decision.") < ctx.index("Old decision.")
+
+
+# ---------------------------------------------------------------------------
+# Past-lessons injection into researchers and risk debators (PR 5)
+# ---------------------------------------------------------------------------
+
+def _researcher_state(past_context=""):
+    return {
+        "company_of_interest": "NVDA",
+        "investment_debate_state": {
+            "history": "", "bull_history": "", "bear_history": "",
+            "current_response": "", "count": 0,
+        },
+        "market_report": "Market report.",
+        "sentiment_report": "Sentiment report.",
+        "news_report": "News report.",
+        "fundamentals_report": "Fundamentals report.",
+        "past_context": past_context,
+    }
+
+
+def _risk_debator_state(past_context=""):
+    return {
+        "company_of_interest": "NVDA",
+        "risk_debate_state": {
+            "history": "", "aggressive_history": "", "conservative_history": "",
+            "neutral_history": "", "current_aggressive_response": "",
+            "current_conservative_response": "", "current_neutral_response": "",
+            "count": 0,
+        },
+        "market_report": "Market report.",
+        "sentiment_report": "Sentiment report.",
+        "news_report": "News report.",
+        "fundamentals_report": "Fundamentals report.",
+        "trader_investment_plan": "Trader plan.",
+        "past_context": past_context,
+    }
+
+
+_LESSONS_HEADER = "Past lessons from prior decisions and outcomes"
+_INJECTION_AGENTS = [
+    (create_bull_researcher, _researcher_state),
+    (create_bear_researcher, _researcher_state),
+    (create_aggressive_debator, _risk_debator_state),
+    (create_conservative_debator, _risk_debator_state),
+    (create_neutral_debator, _risk_debator_state),
+]
+
+
+@pytest.mark.unit
+class TestPastContextInjection:
+
+    @pytest.mark.parametrize(
+        "factory,make_state", _INJECTION_AGENTS,
+        ids=lambda x: getattr(x, "__name__", ""),
+    )
+    def test_lessons_block_present_iff_past_context(self, factory, make_state):
+        llm = MagicMock()
+        llm.invoke.return_value = MagicMock(content="ok")
+        node = factory(llm)
+
+        node(make_state(past_context="[2026-01-05 | AAPL | Buy | +5.0% | +2.0% | 5d]\nGreat call."))
+        prompt = llm.invoke.call_args[0][0]
+        assert _LESSONS_HEADER in prompt
+        assert "Great call." in prompt
+
+        llm.invoke.reset_mock()
+        node(make_state(past_context=""))
+        prompt = llm.invoke.call_args[0][0]
+        assert _LESSONS_HEADER not in prompt
@@ -0,0 +1,109 @@
+"""Tests for PR 4: per-stage model routing and the adaptive risk debate."""
+
+from unittest.mock import MagicMock
+
+import pytest
+
+from tradingagents.graph.conditional_logic import ConditionalLogic
+from tradingagents.graph.setup import LLM_MAP_STAGES, GraphSetup
+
+# ---------------------------------------------------------------------------
+# Per-stage LLM map
+# ---------------------------------------------------------------------------
+
+
+def _make_setup(llm_map=None):
+    quick, deep = MagicMock(name="quick"), MagicMock(name="deep")
+    setup = GraphSetup(quick, deep, {}, ConditionalLogic(), llm_map=llm_map)
+    return setup, quick, deep
+
+
+@pytest.mark.unit
+class TestAgentLlmMap:
+    def test_empty_map_returns_defaults(self):
+        setup, quick, deep = _make_setup()
+        for stage in LLM_MAP_STAGES:
+            assert setup._llm(stage, quick) is quick
+            assert setup._llm(stage, deep) is deep
+
+    def test_map_overrides_single_stage(self):
+        setup, quick, deep = _make_setup({"research_manager": "quick"})
+        assert setup._llm("research_manager", deep) is quick
+        assert setup._llm("portfolio_manager", deep) is deep
+        assert setup._llm("analysts", quick) is quick
+
+    def test_deep_override(self):
+        setup, quick, deep = _make_setup({"risk_analysts": "deep"})
+        assert setup._llm("risk_analysts", quick) is deep
+
+    def test_unknown_stage_raises(self):
+        with pytest.raises(ValueError, match="unknown stage"):
+            _make_setup({"trader_agent": "quick"})
+
+    def test_unknown_value_raises(self):
+        with pytest.raises(ValueError, match="'quick' or 'deep'"):
+            _make_setup({"trader": "gpt-5.5"})
+
+
+# ---------------------------------------------------------------------------
+# Adaptive risk debate
+# ---------------------------------------------------------------------------
+
+AGREE_PLAN = "**Recommendation**: Buy\n\n**Rationale**: Strong setup."
+AGREE_TRADER = "**Action**: Buy\n\nFINAL TRANSACTION PROPOSAL: **BUY**"
+DISAGREE_TRADER = "**Action**: Sell\n\nFINAL TRANSACTION PROPOSAL: **SELL**"
+
+
+def _risk_state(count, investment_plan=AGREE_PLAN, trader_plan=AGREE_TRADER,
+                latest_speaker="Neutral"):
+    return {
+        "risk_debate_state": {"count": count, "latest_speaker": latest_speaker},
+        "investment_plan": investment_plan,
+        "trader_investment_plan": trader_plan,
+    }
+
+
+@pytest.mark.unit
+class TestAdaptiveRiskDebate:
+    def test_default_zero_stops_at_base_cap_even_on_disagreement(self):
+        logic = ConditionalLogic(max_risk_discuss_rounds=1)
+        state = _risk_state(3, trader_plan=DISAGREE_TRADER)
+        assert logic.should_continue_risk_analysis(state) == "Portfolio Manager"
+
+    def test_agreement_stops_at_base_cap_despite_extra_rounds(self):
+        logic = ConditionalLogic(max_risk_discuss_rounds=1, adaptive_extra_rounds=2)
+        state = _risk_state(3)
+        assert logic.should_continue_risk_analysis(state) == "Portfolio Manager"
+
+    def test_disagreement_runs_exactly_n_extra_cycles(self):
+        logic = ConditionalLogic(max_risk_discuss_rounds=1, adaptive_extra_rounds=2)
+        # Past the base cap (3) but under the hard cap (9): keeps debating.
+        for count in range(3, 9):
+            state = _risk_state(count, trader_plan=DISAGREE_TRADER)
+            assert logic.should_continue_risk_analysis(state) != "Portfolio Manager"
+        # At the hard cap: stops regardless of disagreement.
+        state = _risk_state(9, trader_plan=DISAGREE_TRADER)
+        assert logic.should_continue_risk_analysis(state) == "Portfolio Manager"
+
+    def test_speaker_rotation_below_cap_unchanged(self):
+        logic = ConditionalLogic(max_risk_discuss_rounds=1, adaptive_extra_rounds=2)
+        cases = {
+            "Aggressive": "Conservative Analyst",
+            "Conservative": "Neutral Analyst",
+            "Neutral": "Aggressive Analyst",
+        }
+        for speaker, expected in cases.items():
+            state = _risk_state(1, latest_speaker=speaker)
+            assert logic.should_continue_risk_analysis(state) == expected
+        # Rotation also holds in the adaptive window while disagreeing.
+        state = _risk_state(4, trader_plan=DISAGREE_TRADER, latest_speaker="Aggressive")
+        assert logic.should_continue_risk_analysis(state) == "Conservative Analyst"
+
+    def test_hold_vs_hold_counts_as_agreement(self):
+        logic = ConditionalLogic(max_risk_discuss_rounds=1, adaptive_extra_rounds=2)
+        state = _risk_state(
+            3,
+            investment_plan="**Recommendation**: Hold",
+            trader_plan="**Action**: Hold",
+        )
+        assert logic.should_continue_risk_analysis(state) == "Portfolio Manager"