TradingAgents/tradingagents/agents/analysts/_claude_agent_runner.py

202 lines
7.1 KiB
Python

"""SDK-native analyst runner.
When the configured LLM is :class:`ChatClaudeAgent`, the analyst node delegates
the whole tool-calling loop to ``claude-agent-sdk``. The SDK owns the loop:
Claude iteratively invokes the translated MCP tools and returns a final text
report. No LangGraph ToolNode involvement — the analyst returns a terminal
AIMessage with zero tool_calls, so the existing conditional edges route
straight to the message-clear node.
Debug logging: set ``TRADINGAGENTS_CLAUDE_AGENT_DEBUG=1`` to log SDK activity
to ``/tmp/tradingagents_claude_agent.log`` (or set
``TRADINGAGENTS_CLAUDE_AGENT_DEBUG=/path/to/file`` for a custom path). Tail it
in a second terminal to watch progress in real time:
tail -f /tmp/tradingagents_claude_agent.log
"""
import asyncio
import os
import time
from datetime import datetime
from typing import Any, Dict, List
from langchain_core.messages import AIMessage, HumanMessage
from tradingagents.llm_clients.claude_agent_client import (
ChatClaudeAgent,
extract_usage,
fire_llm_callbacks,
)
from tradingagents.llm_clients.mcp_tool_adapter import build_mcp_server
def _debug_path() -> str | None:
val = os.environ.get("TRADINGAGENTS_CLAUDE_AGENT_DEBUG")
if not val:
return None
if val in ("1", "true", "yes", "on"):
return "/tmp/tradingagents_claude_agent.log"
return val
def _log(msg: str) -> None:
path = _debug_path()
if not path:
return
ts = datetime.now().strftime("%H:%M:%S.%f")[:-3]
try:
with open(path, "a") as f:
f.write(f"[{ts}] {msg}\n")
except OSError:
pass
def _describe_message(msg: Any) -> str:
"""One-line summary of an SDK message for the debug log."""
try:
name = type(msg).__name__
content = getattr(msg, "content", None)
if content is None:
return f"{name} (no content)"
if isinstance(content, list):
block_summary = []
for block in content:
bname = type(block).__name__
if hasattr(block, "text"):
text = str(block.text)
snippet = text[:80].replace("\n", " ")
block_summary.append(f"{bname}[{len(text)} chars]: {snippet!r}")
elif hasattr(block, "name"):
block_summary.append(f"{bname}(name={block.name!r})")
else:
block_summary.append(bname)
return f"{name} with {len(content)} blocks: " + " | ".join(block_summary)
return f"{name}: {str(content)[:200]!r}"
except Exception as e:
return f"(failed to describe: {e!r})"
def _build_user_prompt(state: Dict[str, Any]) -> str:
"""Construct a concrete user request from graph state.
The initial graph state is ``messages = [("human", ticker)]`` — too terse
for Claude to act on unambiguously, which can leave the SDK session idle
waiting for clarification. Build an explicit request from
``company_of_interest`` + ``trade_date`` so Claude always knows what to do.
Any additional human-authored content in the message stream is appended.
"""
ticker = state.get("company_of_interest", "")
trade_date = state.get("trade_date", "")
base = (
f"Produce the requested report for {ticker} as of {trade_date}. "
"Use the available tools to gather the data you need, then write the "
"final report. Do not ask clarifying questions — proceed directly."
).strip()
extra: List[str] = []
for msg in state.get("messages", []):
content = getattr(msg, "content", None)
if isinstance(msg, HumanMessage) and isinstance(content, str):
c = content.strip()
if c and c != ticker:
extra.append(c)
if extra:
return base + "\n\nAdditional context:\n" + "\n".join(extra)
return base
async def _run(
system_prompt: str,
user_prompt: str,
lc_tools: List[Any],
server_name: str,
model: str,
callbacks: List[Any],
) -> tuple[str, Dict[str, int]]:
from claude_agent_sdk import (
AssistantMessage,
ClaudeAgentOptions,
TextBlock,
query,
)
_log(f"[{server_name}] building MCP server with {len(lc_tools)} tools: "
f"{[t.name for t in lc_tools]}")
server, allowed = build_mcp_server(server_name, lc_tools, callbacks=callbacks)
_log(f"[{server_name}] allowed_tools={allowed}")
options = ClaudeAgentOptions(
model=model,
system_prompt=system_prompt,
mcp_servers={server_name: server},
allowed_tools=allowed,
# Block the Claude Code built-ins; only our MCP tools should run.
disallowed_tools=[
"Bash", "Read", "Write", "Edit", "MultiEdit",
"Glob", "Grep", "WebFetch", "WebSearch",
"Task", "TodoWrite", "NotebookEdit",
],
permission_mode="bypassPermissions",
)
_log(f"[{server_name}] starting query(model={model!r}, prompt={user_prompt[:120]!r}...)")
start = time.monotonic()
text_parts: List[str] = []
final_usage: Dict[str, int] = {}
msg_count = 0
async for msg in query(prompt=user_prompt, options=options):
msg_count += 1
elapsed = time.monotonic() - start
_log(f"[{server_name}] +{elapsed:.1f}s msg #{msg_count}: {_describe_message(msg)}")
if isinstance(msg, AssistantMessage):
for block in msg.content:
if isinstance(block, TextBlock):
text_parts.append(block.text)
sdk_usage = getattr(msg, "usage", None)
if isinstance(sdk_usage, dict) and sdk_usage:
final_usage = extract_usage(sdk_usage)
elapsed = time.monotonic() - start
_log(f"[{server_name}] query complete after {elapsed:.1f}s, "
f"{msg_count} messages, {sum(len(t) for t in text_parts)} chars, "
f"usage={final_usage}")
return "\n".join(text_parts).strip(), final_usage
def run_sdk_analyst(
llm: ChatClaudeAgent,
state: Dict[str, Any],
system_prompt: str,
lc_tools: List[Any],
server_name: str,
report_field: str,
) -> Dict[str, Any]:
"""Run an analyst through the Claude Agent SDK tool loop and build the node output."""
user_prompt = _build_user_prompt(state)
_log(f"=== run_sdk_analyst start: server={server_name} report_field={report_field} "
f"ticker={state.get('company_of_interest')!r} date={state.get('trade_date')!r} ===")
try:
report, usage = asyncio.run(
_run(
system_prompt=system_prompt,
user_prompt=user_prompt,
lc_tools=lc_tools,
server_name=server_name,
model=llm.model,
callbacks=llm.callbacks,
)
)
except Exception as e:
_log(f"[{server_name}] EXCEPTION: {type(e).__name__}: {e}")
raise
_log(f"=== run_sdk_analyst done: {report_field}={len(report)} chars usage={usage} ===")
message = AIMessage(content=report, usage_metadata=usage or None)
fire_llm_callbacks(llm.callbacks, message, user_prompt)
return {
"messages": [message],
report_field: report,
}