feat: implement pre-market planner with Gemini integration (issue #83)
PreMarketPlanner generates DayPlaybook via single Gemini API call per market: - Structured JSON prompt with scan candidates + strategic context - Cross-market context (KR reads US scorecard, US reads KR scorecard) - Robust JSON parser with markdown fence stripping - Unknown stock filtering (only scanner candidates allowed) - MAX_SCENARIOS_PER_STOCK enforcement - Defensive playbook on failure (HOLD + stop-loss) - Empty playbook when no candidates (safe, no trades) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
agentson
408
src/strategy/pre_market_planner.py
Normal file
408
src/strategy/pre_market_planner.py
Normal file
@@ -0,0 +1,408 @@
|
||||
"""Pre-market planner — generates DayPlaybook via Gemini before market open.
|
||||
|
||||
One Gemini API call per market per day. Candidates come from SmartVolatilityScanner.
|
||||
On failure, returns a defensive playbook (all HOLD, no trades).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import logging
|
||||
from datetime import UTC, date, datetime
|
||||
from typing import Any
|
||||
|
||||
from src.analysis.smart_scanner import ScanCandidate
|
||||
from src.brain.context_selector import ContextSelector, DecisionType
|
||||
from src.brain.gemini_client import GeminiClient
|
||||
from src.config import Settings
|
||||
from src.context.store import ContextLayer, ContextStore
|
||||
from src.strategy.models import (
|
||||
CrossMarketContext,
|
||||
DayPlaybook,
|
||||
GlobalRule,
|
||||
MarketOutlook,
|
||||
PlaybookStatus,
|
||||
ScenarioAction,
|
||||
StockCondition,
|
||||
StockPlaybook,
|
||||
StockScenario,
|
||||
)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Mapping from string to MarketOutlook enum
|
||||
_OUTLOOK_MAP: dict[str, MarketOutlook] = {
|
||||
"bullish": MarketOutlook.BULLISH,
|
||||
"neutral_to_bullish": MarketOutlook.NEUTRAL_TO_BULLISH,
|
||||
"neutral": MarketOutlook.NEUTRAL,
|
||||
"neutral_to_bearish": MarketOutlook.NEUTRAL_TO_BEARISH,
|
||||
"bearish": MarketOutlook.BEARISH,
|
||||
}
|
||||
|
||||
_ACTION_MAP: dict[str, ScenarioAction] = {
|
||||
"BUY": ScenarioAction.BUY,
|
||||
"SELL": ScenarioAction.SELL,
|
||||
"HOLD": ScenarioAction.HOLD,
|
||||
"REDUCE_ALL": ScenarioAction.REDUCE_ALL,
|
||||
}
|
||||
|
||||
|
||||
class PreMarketPlanner:
|
||||
"""Generates a DayPlaybook by calling Gemini once before market open.
|
||||
|
||||
Flow:
|
||||
1. Collect strategic context (L5-L7) + cross-market context
|
||||
2. Build a structured prompt with scan candidates
|
||||
3. Call Gemini for JSON scenario generation
|
||||
4. Parse and validate response into DayPlaybook
|
||||
5. On failure → defensive playbook (HOLD everything)
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
gemini_client: GeminiClient,
|
||||
context_store: ContextStore,
|
||||
context_selector: ContextSelector,
|
||||
settings: Settings,
|
||||
) -> None:
|
||||
self._gemini = gemini_client
|
||||
self._context_store = context_store
|
||||
self._context_selector = context_selector
|
||||
self._settings = settings
|
||||
|
||||
async def generate_playbook(
|
||||
self,
|
||||
market: str,
|
||||
candidates: list[ScanCandidate],
|
||||
) -> DayPlaybook:
|
||||
"""Generate a DayPlaybook for a market using Gemini.
|
||||
|
||||
Args:
|
||||
market: Market code ("KR" or "US")
|
||||
candidates: Stock candidates from SmartVolatilityScanner
|
||||
|
||||
Returns:
|
||||
DayPlaybook with scenarios. Empty/defensive if no candidates or failure.
|
||||
"""
|
||||
today = date.today()
|
||||
|
||||
if not candidates:
|
||||
logger.info("No candidates for %s — returning empty playbook", market)
|
||||
return self._empty_playbook(today, market)
|
||||
|
||||
try:
|
||||
# 1. Gather context
|
||||
context_data = self._gather_context()
|
||||
cross_market = self.build_cross_market_context(market)
|
||||
|
||||
# 2. Build prompt
|
||||
prompt = self._build_prompt(market, candidates, context_data, cross_market)
|
||||
|
||||
# 3. Call Gemini
|
||||
market_data = {
|
||||
"stock_code": "PLANNER",
|
||||
"current_price": 0,
|
||||
"prompt_override": prompt,
|
||||
}
|
||||
decision = await self._gemini.decide(market_data)
|
||||
|
||||
# 4. Parse response
|
||||
playbook = self._parse_response(
|
||||
decision.rationale, today, market, candidates, cross_market
|
||||
)
|
||||
playbook_with_tokens = playbook.model_copy(
|
||||
update={"token_count": decision.token_count}
|
||||
)
|
||||
logger.info(
|
||||
"Generated playbook for %s: %d stocks, %d scenarios, %d tokens",
|
||||
market,
|
||||
playbook_with_tokens.stock_count,
|
||||
playbook_with_tokens.scenario_count,
|
||||
playbook_with_tokens.token_count,
|
||||
)
|
||||
return playbook_with_tokens
|
||||
|
||||
except Exception:
|
||||
logger.exception("Playbook generation failed for %s", market)
|
||||
if self._settings.DEFENSIVE_PLAYBOOK_ON_FAILURE:
|
||||
return self._defensive_playbook(today, market, candidates)
|
||||
return self._empty_playbook(today, market)
|
||||
|
||||
def build_cross_market_context(self, target_market: str) -> CrossMarketContext | None:
|
||||
"""Build cross-market context from the other market's L6 data.
|
||||
|
||||
KR planner → reads US scorecard from previous night.
|
||||
US planner → reads KR scorecard from today.
|
||||
"""
|
||||
other_market = "US" if target_market == "KR" else "KR"
|
||||
today = date.today()
|
||||
timeframe = today.isoformat()
|
||||
|
||||
scorecard_key = f"scorecard_{other_market}"
|
||||
scorecard_data = self._context_store.get_context(
|
||||
ContextLayer.L6_DAILY, timeframe, scorecard_key
|
||||
)
|
||||
|
||||
if scorecard_data is None:
|
||||
logger.debug("No cross-market scorecard found for %s", other_market)
|
||||
return None
|
||||
|
||||
if isinstance(scorecard_data, str):
|
||||
try:
|
||||
scorecard_data = json.loads(scorecard_data)
|
||||
except (json.JSONDecodeError, TypeError):
|
||||
return None
|
||||
|
||||
if not isinstance(scorecard_data, dict):
|
||||
return None
|
||||
|
||||
return CrossMarketContext(
|
||||
market=other_market,
|
||||
date=timeframe,
|
||||
total_pnl=float(scorecard_data.get("total_pnl", 0.0)),
|
||||
win_rate=float(scorecard_data.get("win_rate", 0.0)),
|
||||
index_change_pct=float(scorecard_data.get("index_change_pct", 0.0)),
|
||||
key_events=scorecard_data.get("key_events", []),
|
||||
lessons=scorecard_data.get("lessons", []),
|
||||
)
|
||||
|
||||
def _gather_context(self) -> dict[str, Any]:
|
||||
"""Gather strategic context using ContextSelector."""
|
||||
layers = self._context_selector.select_layers(
|
||||
decision_type=DecisionType.STRATEGIC,
|
||||
include_realtime=True,
|
||||
)
|
||||
return self._context_selector.get_context_data(layers, max_items_per_layer=10)
|
||||
|
||||
def _build_prompt(
|
||||
self,
|
||||
market: str,
|
||||
candidates: list[ScanCandidate],
|
||||
context_data: dict[str, Any],
|
||||
cross_market: CrossMarketContext | None,
|
||||
) -> str:
|
||||
"""Build a structured prompt for Gemini to generate scenario JSON."""
|
||||
max_scenarios = self._settings.MAX_SCENARIOS_PER_STOCK
|
||||
|
||||
candidates_text = "\n".join(
|
||||
f" - {c.stock_code} ({c.name}): price={c.price}, "
|
||||
f"RSI={c.rsi:.1f}, volume_ratio={c.volume_ratio:.1f}, "
|
||||
f"signal={c.signal}, score={c.score:.1f}"
|
||||
for c in candidates
|
||||
)
|
||||
|
||||
cross_market_text = ""
|
||||
if cross_market:
|
||||
cross_market_text = (
|
||||
f"\n## Other Market ({cross_market.market}) Summary\n"
|
||||
f"- P&L: {cross_market.total_pnl:+.2f}%\n"
|
||||
f"- Win Rate: {cross_market.win_rate:.0f}%\n"
|
||||
f"- Index Change: {cross_market.index_change_pct:+.2f}%\n"
|
||||
)
|
||||
if cross_market.lessons:
|
||||
cross_market_text += f"- Lessons: {'; '.join(cross_market.lessons[:3])}\n"
|
||||
|
||||
context_text = ""
|
||||
if context_data:
|
||||
context_text = "\n## Strategic Context\n"
|
||||
for layer_name, layer_data in context_data.items():
|
||||
if layer_data:
|
||||
context_text += f"### {layer_name}\n"
|
||||
for key, value in list(layer_data.items())[:5]:
|
||||
context_text += f" - {key}: {value}\n"
|
||||
|
||||
return (
|
||||
f"You are a pre-market trading strategist for the {market} market.\n"
|
||||
f"Generate structured trading scenarios for today.\n\n"
|
||||
f"## Candidates (from volatility scanner)\n{candidates_text}\n"
|
||||
f"{cross_market_text}"
|
||||
f"{context_text}\n"
|
||||
f"## Instructions\n"
|
||||
f"Return a JSON object with this exact structure:\n"
|
||||
f'{{\n'
|
||||
f' "market_outlook": "bullish|neutral_to_bullish|neutral|neutral_to_bearish|bearish",\n'
|
||||
f' "global_rules": [\n'
|
||||
f' {{"condition": "portfolio_pnl_pct < -2.0", "action": "REDUCE_ALL", "rationale": "..."}}\n'
|
||||
f' ],\n'
|
||||
f' "stocks": [\n'
|
||||
f' {{\n'
|
||||
f' "stock_code": "...",\n'
|
||||
f' "scenarios": [\n'
|
||||
f' {{\n'
|
||||
f' "condition": {{"rsi_below": 30, "volume_ratio_above": 2.0}},\n'
|
||||
f' "action": "BUY|SELL|HOLD",\n'
|
||||
f' "confidence": 85,\n'
|
||||
f' "allocation_pct": 10.0,\n'
|
||||
f' "stop_loss_pct": -2.0,\n'
|
||||
f' "take_profit_pct": 3.0,\n'
|
||||
f' "rationale": "..."\n'
|
||||
f' }}\n'
|
||||
f' ]\n'
|
||||
f' }}\n'
|
||||
f' ]\n'
|
||||
f'}}\n\n'
|
||||
f"Rules:\n"
|
||||
f"- Max {max_scenarios} scenarios per stock\n"
|
||||
f"- Only use stocks from the candidates list\n"
|
||||
f"- Confidence 0-100 (80+ for actionable trades)\n"
|
||||
f"- stop_loss_pct must be <= 0, take_profit_pct must be >= 0\n"
|
||||
f"- Return ONLY the JSON, no markdown fences or explanation\n"
|
||||
)
|
||||
|
||||
def _parse_response(
|
||||
self,
|
||||
response_text: str,
|
||||
today: date,
|
||||
market: str,
|
||||
candidates: list[ScanCandidate],
|
||||
cross_market: CrossMarketContext | None,
|
||||
) -> DayPlaybook:
|
||||
"""Parse Gemini's JSON response into a validated DayPlaybook."""
|
||||
cleaned = self._extract_json(response_text)
|
||||
data = json.loads(cleaned)
|
||||
|
||||
valid_codes = {c.stock_code for c in candidates}
|
||||
|
||||
# Parse market outlook
|
||||
outlook_str = data.get("market_outlook", "neutral")
|
||||
market_outlook = _OUTLOOK_MAP.get(outlook_str, MarketOutlook.NEUTRAL)
|
||||
|
||||
# Parse global rules
|
||||
global_rules = []
|
||||
for rule_data in data.get("global_rules", []):
|
||||
action_str = rule_data.get("action", "HOLD")
|
||||
action = _ACTION_MAP.get(action_str, ScenarioAction.HOLD)
|
||||
global_rules.append(
|
||||
GlobalRule(
|
||||
condition=rule_data.get("condition", ""),
|
||||
action=action,
|
||||
rationale=rule_data.get("rationale", ""),
|
||||
)
|
||||
)
|
||||
|
||||
# Parse stock playbooks
|
||||
stock_playbooks = []
|
||||
max_scenarios = self._settings.MAX_SCENARIOS_PER_STOCK
|
||||
for stock_data in data.get("stocks", []):
|
||||
code = stock_data.get("stock_code", "")
|
||||
if code not in valid_codes:
|
||||
logger.warning("Gemini returned unknown stock %s — skipping", code)
|
||||
continue
|
||||
|
||||
scenarios = []
|
||||
for sc_data in stock_data.get("scenarios", [])[:max_scenarios]:
|
||||
scenario = self._parse_scenario(sc_data)
|
||||
if scenario:
|
||||
scenarios.append(scenario)
|
||||
|
||||
if scenarios:
|
||||
stock_playbooks.append(
|
||||
StockPlaybook(
|
||||
stock_code=code,
|
||||
scenarios=scenarios,
|
||||
)
|
||||
)
|
||||
|
||||
return DayPlaybook(
|
||||
date=today,
|
||||
market=market,
|
||||
market_outlook=market_outlook,
|
||||
global_rules=global_rules,
|
||||
stock_playbooks=stock_playbooks,
|
||||
cross_market=cross_market,
|
||||
)
|
||||
|
||||
def _parse_scenario(self, sc_data: dict) -> StockScenario | None:
|
||||
"""Parse a single scenario from JSON data. Returns None if invalid."""
|
||||
try:
|
||||
cond_data = sc_data.get("condition", {})
|
||||
condition = StockCondition(
|
||||
rsi_below=cond_data.get("rsi_below"),
|
||||
rsi_above=cond_data.get("rsi_above"),
|
||||
volume_ratio_above=cond_data.get("volume_ratio_above"),
|
||||
volume_ratio_below=cond_data.get("volume_ratio_below"),
|
||||
price_above=cond_data.get("price_above"),
|
||||
price_below=cond_data.get("price_below"),
|
||||
price_change_pct_above=cond_data.get("price_change_pct_above"),
|
||||
price_change_pct_below=cond_data.get("price_change_pct_below"),
|
||||
)
|
||||
|
||||
if not condition.has_any_condition():
|
||||
logger.warning("Scenario has no conditions — skipping")
|
||||
return None
|
||||
|
||||
action_str = sc_data.get("action", "HOLD")
|
||||
action = _ACTION_MAP.get(action_str, ScenarioAction.HOLD)
|
||||
|
||||
return StockScenario(
|
||||
condition=condition,
|
||||
action=action,
|
||||
confidence=int(sc_data.get("confidence", 50)),
|
||||
allocation_pct=float(sc_data.get("allocation_pct", 10.0)),
|
||||
stop_loss_pct=float(sc_data.get("stop_loss_pct", -2.0)),
|
||||
take_profit_pct=float(sc_data.get("take_profit_pct", 3.0)),
|
||||
rationale=sc_data.get("rationale", ""),
|
||||
)
|
||||
except (ValueError, TypeError) as e:
|
||||
logger.warning("Failed to parse scenario: %s", e)
|
||||
return None
|
||||
|
||||
@staticmethod
|
||||
def _extract_json(text: str) -> str:
|
||||
"""Extract JSON from response, stripping markdown fences if present."""
|
||||
stripped = text.strip()
|
||||
if stripped.startswith("```"):
|
||||
# Remove first line (```json or ```) and last line (```)
|
||||
lines = stripped.split("\n")
|
||||
lines = lines[1:] # Remove opening fence
|
||||
if lines and lines[-1].strip() == "```":
|
||||
lines = lines[:-1]
|
||||
stripped = "\n".join(lines)
|
||||
return stripped.strip()
|
||||
|
||||
@staticmethod
|
||||
def _empty_playbook(today: date, market: str) -> DayPlaybook:
|
||||
"""Return an empty playbook (no stocks, no scenarios)."""
|
||||
return DayPlaybook(
|
||||
date=today,
|
||||
market=market,
|
||||
market_outlook=MarketOutlook.NEUTRAL,
|
||||
stock_playbooks=[],
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def _defensive_playbook(
|
||||
today: date,
|
||||
market: str,
|
||||
candidates: list[ScanCandidate],
|
||||
) -> DayPlaybook:
|
||||
"""Return a defensive playbook — HOLD everything with stop-loss ready."""
|
||||
stock_playbooks = [
|
||||
StockPlaybook(
|
||||
stock_code=c.stock_code,
|
||||
scenarios=[
|
||||
StockScenario(
|
||||
condition=StockCondition(price_change_pct_below=-3.0),
|
||||
action=ScenarioAction.SELL,
|
||||
confidence=90,
|
||||
stop_loss_pct=-3.0,
|
||||
rationale="Defensive stop-loss (planner failure)",
|
||||
),
|
||||
],
|
||||
)
|
||||
for c in candidates
|
||||
]
|
||||
return DayPlaybook(
|
||||
date=today,
|
||||
market=market,
|
||||
market_outlook=MarketOutlook.NEUTRAL_TO_BEARISH,
|
||||
default_action=ScenarioAction.HOLD,
|
||||
stock_playbooks=stock_playbooks,
|
||||
global_rules=[
|
||||
GlobalRule(
|
||||
condition="portfolio_pnl_pct < -2.0",
|
||||
action=ScenarioAction.REDUCE_ALL,
|
||||
rationale="Defensive: reduce on loss threshold",
|
||||
),
|
||||
],
|
||||
)
|
||||
Reference in New Issue
Block a user