"Vibe-Trading: Your Personal Trading Agent"
Python ai-agentalgorithmic-tradingbacktestingfintechllmmcpmulti-agentpythonquantitative-financetrading
"""Shared runtime limits for tool results.
The cap on a tool result used to live in :mod:`src.agent.loop` and was copied as
a bare literal into :mod:`src.swarm.worker`, so the two could drift. It lives
here instead: a leaf module with no imports of its own, which the agent loop,
the swarm worker and individual tools can all read without any of them taking a
dependency on the others.
"""
from __future__ import annotations
#: Maximum characters of a single tool result handed back to the model.
TOOL_RESULT_LIMIT = 10_000
_TRUNCATION_NOTICE = (
"\n\n[TRUNCATED: {shown} of {total} characters delivered. The remainder was "
"not sent. Do not treat this as the complete result — narrow the request, "
"or call the tool again with its paging parameter if it has one.]"
)
def truncate_tool_result(result: str, limit: int = TOOL_RESULT_LIMIT) -> str:
"""Cut an oversized tool result to ``limit`` characters and say so.
A silent cut is indistinguishable from a complete answer, so the model
reports a prefix as if it were the whole thing. Measured on live data,
``get_financial_statements("AAPL.US", period="quarter")`` returns 28,498
characters for 40 periods; the silent cut delivered roughly 12 of them.
Args:
result: The full tool result.
limit: Maximum characters to deliver, notice included.
Returns:
``result`` unchanged when it fits, otherwise a prefix of exactly
``limit`` characters or fewer, ending in a notice that states how much
was withheld.
"""
total = len(result)
if total <= limit:
return result
notice = _TRUNCATION_NOTICE.format(shown=limit, total=total)
body = limit - len(notice)
if body <= 0:
# Pathologically small limit: the notice alone is more useful than a
# fragment, so deliver it rather than silently cutting to nothing.
return notice[:limit]
return result[:body] + _TRUNCATION_NOTICE.format(shown=body, total=total)