Spaces:
Running
Running
File size: 6,746 Bytes
69e310f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 | """Data agent — decides what to compute, then consumes MCP tools to get it.
Runs in parallel with the news agent. Every number it contributes to state came
out of the MCP tool server; the agent's own contribution is the *plan*.
"""
from __future__ import annotations
import logging
from typing import Any
from app.core.budget import BudgetExceededError
from app.core.claude import AgentRole, LLMRequest, PromptHint
from app.core.events import EventKind
from app.graph.context import RunContext, current_context
from app.graph.llm import call_model
from app.graph.prompts import DATA_AGENT_SYSTEM, PLAN_MARKET_DATA_TOOL
from app.graph.state import RunState
from app.mcp_server.client import McpToolError
from app.mcp_server.providers import MAX_HISTORY_DAYS
from app.models.market import Fundamentals, Metrics, PriceHistory
from app.models.run import RunError, RunStatus, Stage
logger = logging.getLogger(__name__)
AGENT_NAME: Stage = "data_agent"
async def _plan_targets(
ctx: RunContext, state: RunState, targets: list[str]
) -> tuple[list[str], int, dict[str, Any]]:
"""Ask the model which tickers to pull and over what window."""
default_days = int(state.get("days", ctx.settings.price_history_days))
request = LLMRequest(
role=AgentRole.DATA,
hint=PromptHint.DATA_ANALYSE,
system=DATA_AGENT_SYSTEM,
messages=[
{
"role": "user",
"content": (
f"Assigned tickers: {', '.join(targets)}.\n"
f"Default history window: {default_days} calendar days.\n"
"Call plan_market_data with the tickers to fetch and the window to use."
),
}
],
tools=[PLAN_MARKET_DATA_TOOL],
forced_tool=PLAN_MARKET_DATA_TOOL.name,
context={"tickers": targets, "days": default_days},
)
outcome = await call_model(ctx, request)
plan = outcome.result.first_tool(PLAN_MARKET_DATA_TOOL.name) or {}
# Scope containment: the model may narrow the assignment but never widen it.
# A prompt-injected "also fetch XYZ" therefore cannot reach the provider.
assigned = set(targets)
requested = [
str(t).strip().upper()
for t in plan.get("tickers", [])
if str(t).strip().upper() in assigned
]
chosen = requested or targets
days = plan.get("days", default_days)
try:
days_int = max(2, min(int(days), MAX_HISTORY_DAYS))
except (TypeError, ValueError):
days_int = default_days
spend = outcome.spend.model_dump()
return chosen, days_int, spend
async def data_agent_node(state: RunState) -> dict[str, Any]:
"""Graph node: fetch prices, fundamentals and metrics for the assigned tickers."""
ctx = current_context()
targets = list((state.get("plan") or {}).get("market_tickers", []))
if not targets:
return {"agents_completed": [AGENT_NAME]}
await ctx.emit(
EventKind.AGENT_STARTED,
f"data_agent working {len(targets)} ticker(s)",
{"agent": AGENT_NAME, "tickers": targets},
)
prices: dict[str, PriceHistory] = {}
fundamentals: dict[str, Fundamentals] = {}
metrics: dict[str, Metrics] = {}
errors: list[RunError] = []
attempts: dict[str, int] = {f"market:{ticker}": 1 for ticker in targets}
spend_delta: dict[str, Any] = {}
with (
ctx.mcp.collect() as records,
ctx.tracer.step("data_agent", run_id=ctx.run_id, input_data={"tickers": targets}) as span,
):
try:
chosen, days, spend_delta = await _plan_targets(ctx, state, targets)
except BudgetExceededError as exc:
return {
"agents_completed": [AGENT_NAME],
"attempts": attempts,
"status": RunStatus.BUDGET_ABORT,
"abort_reason": str(exc),
"errors": [RunError(stage=AGENT_NAME, message=str(exc))],
}
for ticker in chosen:
try:
history = await ctx.mcp.get_price_history(ticker, days)
prices[ticker] = history
company = await ctx.mcp.get_fundamentals(ticker)
fundamentals[ticker] = company
computed = await ctx.mcp.compute_metrics(
ticker, list(history.bars), company.pe_ratio
)
metrics[ticker] = computed
if history.error:
errors.append(RunError(stage=AGENT_NAME, ticker=ticker, message=history.error))
if company.error:
errors.append(
RunError(
stage=AGENT_NAME,
ticker=ticker,
message=company.error,
severity="warning",
)
)
if computed.error:
errors.append(RunError(stage=AGENT_NAME, ticker=ticker, message=computed.error))
except McpToolError as exc:
logger.warning("data agent MCP failure for %s: %s", ticker, exc)
errors.append(
RunError(
stage=AGENT_NAME,
ticker=ticker,
message=f"MCP tool unavailable: {exc}",
)
)
except Exception as exc:
logger.exception("unexpected data agent failure for %s", ticker)
errors.append(
RunError(
stage=AGENT_NAME,
ticker=ticker,
message=f"unexpected failure: {type(exc).__name__}",
)
)
span.update(
output={
"tickers_completed": sorted(metrics),
"errors": len(errors),
"tool_calls": len(records),
}
)
tool_call_rows = [record.to_dict() for record in records]
ok_count = sum(1 for metric in metrics.values() if metric.ok)
await ctx.emit(
EventKind.AGENT_COMPLETED,
f"data_agent finished — {ok_count}/{len(targets)} ticker(s) with metrics",
{"agent": AGENT_NAME, "ok": ok_count, "errors": len(errors)},
)
update: dict[str, Any] = {
"prices": prices,
"fundamentals": fundamentals,
"metrics": metrics,
"errors": errors,
"attempts": attempts,
"tool_calls": tool_call_rows,
"agents_completed": [AGENT_NAME],
}
if spend_delta:
from app.models.run import TokenSpend
update["token_spend"] = TokenSpend.model_validate(spend_delta)
return update
|