Files
hermes-agent/tools/budget_config.py

125 lines
5.5 KiB
Python

"""Configurable budget constants for tool result persistence.
Per-tool resolution: pinned > config overrides > registry > default.
"""
from dataclasses import dataclass, field
from typing import Dict
# Tools whose thresholds must never be overridden.
# read_file=inf prevents infinite persist->read->persist loops.
PINNED_THRESHOLDS: Dict[str, float] = {
"read_file": float("inf"),
}
# Single source of truth for the defaults; tool_result_storage.py imports these.
DEFAULT_RESULT_SIZE_CHARS: int = 100_000
DEFAULT_TURN_BUDGET_CHARS: int = 200_000
DEFAULT_PREVIEW_SIZE_CHARS: int = 1_500
# Tighter per-result default for MCP tools (``mcp_`` prefix): MCP servers
# routinely return un-paginated 20-50K-char payloads that sail under the
# generic 100K threshold and bloat context. 50K matches the strictest general
# competitor caps while spillover (unlike truncation) keeps the full payload on
# disk. Overridable via ``tool_budget.mcp_result_size_chars`` in config.yaml.
DEFAULT_MCP_RESULT_SIZE_CHARS: int = 50_000
# Same prefix the untrusted-content wrapper keys on (agent/tool_dispatch_helpers.py).
MCP_TOOL_PREFIX: str = "mcp_"
def _configured_mcp_result_size() -> int:
"""Read ``tool_budget.mcp_result_size_chars`` via ``load_config_readonly`` (the
sanctioned read path; raw config.yaml parsing outside owner modules is test-guarded).
Any error, missing key or non-positive value returns the built-in default."""
try:
from hermes_cli.config import load_config_readonly
data = load_config_readonly()
block = data.get("tool_budget") if isinstance(data, dict) else None
raw = block.get("mcp_result_size_chars") if isinstance(block, dict) else None
if raw is not None and int(raw) > 0:
return int(raw)
except Exception:
pass
return DEFAULT_MCP_RESULT_SIZE_CHARS
@dataclass(frozen=True)
class BudgetConfig:
"""Immutable budget constants for the 3-layer tool result persistence system.
Layer 2 (per-result): resolve_threshold(tool_name) -> threshold in chars.
Layer 3 (per-turn): turn_budget -> aggregate char budget across one assistant turn.
Preview: preview_size -> inline snippet size after persistence.
"""
default_result_size: int = DEFAULT_RESULT_SIZE_CHARS
turn_budget: int = DEFAULT_TURN_BUDGET_CHARS
preview_size: int = DEFAULT_PREVIEW_SIZE_CHARS
mcp_result_size: int = DEFAULT_MCP_RESULT_SIZE_CHARS
tool_overrides: Dict[str, int] = field(default_factory=dict)
def resolve_threshold(self, tool_name: str) -> int | float:
"""Priority: pinned -> tool_overrides -> mcp_ prefix -> registry per-tool -> default.
MCP tools get ``mcp_result_size`` because they have no per-tool registry
entry to constrain them. Both the MCP value and the registry value are
capped at ``default_result_size`` so a context-scaled budget (small
model) still constrains tools that register a large fixed
``max_result_size_chars`` (web/terminal/x_search all register 100K);
a no-op for the default budget, but for a scaled-down budget it stops a
registry value from re-inflating the cap past the model's window.
"""
if tool_name in PINNED_THRESHOLDS:
return PINNED_THRESHOLDS[tool_name]
if tool_name in self.tool_overrides:
return self.tool_overrides[tool_name]
if tool_name.startswith(MCP_TOOL_PREFIX):
return min(self.mcp_result_size, self.default_result_size)
from tools.registry import registry
registry_value = registry.get_max_result_size(tool_name, default=self.default_result_size)
if registry_value == float("inf"):
return registry_value
return min(registry_value, self.default_result_size)
# Default config -- matches the historical hardcoded behavior exactly.
DEFAULT_BUDGET = BudgetConfig()
# Token<->char ratio for scaling to a context window; same rough 4-chars-per-token
# the estimator uses (agent/model_metadata.py). A smaller divisor would
# UNDER-protect small models.
_CHARS_PER_TOKEN: int = 4
# Window fraction a SINGLE tool result / the WHOLE turn's tool output may occupy.
# System prompt, tool schemas, history and the reply all compete, so well under 1.0.
_PER_RESULT_WINDOW_FRACTION: float = 0.15
_PER_TURN_WINDOW_FRACTION: float = 0.30
# Floors so a tiny model still gets a usable preview/result, never a 0-char budget.
_MIN_RESULT_SIZE_CHARS: int = 8_000
_MIN_TURN_BUDGET_CHARS: int = 16_000
def budget_for_context_window(context_length: int | None) -> BudgetConfig:
"""Return a BudgetConfig scaled to the active model's context window. The fixed
defaults suit 200K+ token models but on a 65K model one result/turn can fill
the window; the proportional value is clamped to the defaults as a CAP (large
models stay byte-identical) and floored so a usable preview always survives."""
mcp_result_size = _configured_mcp_result_size()
if not context_length or context_length <= 0:
if mcp_result_size == DEFAULT_MCP_RESULT_SIZE_CHARS:
return DEFAULT_BUDGET
return BudgetConfig(mcp_result_size=mcp_result_size)
window_chars = context_length * _CHARS_PER_TOKEN
return BudgetConfig(
default_result_size=max(_MIN_RESULT_SIZE_CHARS, min(int(window_chars * _PER_RESULT_WINDOW_FRACTION), DEFAULT_RESULT_SIZE_CHARS)),
turn_budget=max(_MIN_TURN_BUDGET_CHARS, min(int(window_chars * _PER_TURN_WINDOW_FRACTION), DEFAULT_TURN_BUDGET_CHARS)),
preview_size=DEFAULT_PREVIEW_SIZE_CHARS,
mcp_result_size=mcp_result_size,
)