Compare commits
2 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| ba56567631 | |||
| c6f2855745 |
@@ -1,156 +0,0 @@
|
||||
"""Tool fixation detection — break repetitive tool calling loops.
|
||||
|
||||
Detects when the agent latches onto one tool and calls it repeatedly
|
||||
without making progress. Injects a nudge prompt to break the loop.
|
||||
|
||||
Usage:
|
||||
from agent.tool_fixation_detector import ToolFixationDetector
|
||||
detector = ToolFixationDetector()
|
||||
nudge = detector.record("execute_code")
|
||||
if nudge:
|
||||
# Inject nudge into conversation
|
||||
messages.append({"role": "system", "content": nudge})
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Dict, List, Optional
|
||||
|
||||
|
||||
# Default thresholds
|
||||
_DEFAULT_THRESHOLD = int(os.getenv("TOOL_FIXATION_THRESHOLD", "5"))
|
||||
_DEFAULT_WINDOW = int(os.getenv("TOOL_FIXATION_WINDOW", "10"))
|
||||
|
||||
|
||||
@dataclass
|
||||
class FixationEvent:
|
||||
"""Record of a fixation detection."""
|
||||
tool_name: str
|
||||
streak_length: int
|
||||
threshold: int
|
||||
nudge_sent: bool = False
|
||||
|
||||
|
||||
class ToolFixationDetector:
|
||||
"""Detects and breaks tool fixation loops.
|
||||
|
||||
Tracks the sequence of tool calls and detects when the same tool
|
||||
is called N times consecutively. When detected, returns a nudge
|
||||
prompt to inject into the conversation.
|
||||
"""
|
||||
|
||||
def __init__(self, threshold: int = 0, window: int = 0):
|
||||
self.threshold = threshold or _DEFAULT_THRESHOLD
|
||||
self.window = window or _DEFAULT_WINDOW
|
||||
self._history: List[str] = []
|
||||
self._current_streak: str = ""
|
||||
self._streak_count: int = 0
|
||||
self._nudges_sent: int = 0
|
||||
self._events: List[FixationEvent] = []
|
||||
|
||||
@property
|
||||
def nudges_sent(self) -> int:
|
||||
return self._nudges_sent
|
||||
|
||||
@property
|
||||
def events(self) -> List[FixationEvent]:
|
||||
return list(self._events)
|
||||
|
||||
def record(self, tool_name: str) -> Optional[str]:
|
||||
"""Record a tool call and return nudge prompt if fixation detected.
|
||||
|
||||
Args:
|
||||
tool_name: Name of the tool that was called.
|
||||
|
||||
Returns:
|
||||
Nudge prompt string if fixation detected, None otherwise.
|
||||
"""
|
||||
self._history.append(tool_name)
|
||||
|
||||
# Trim history to window
|
||||
if len(self._history) > self.window:
|
||||
self._history = self._history[-self.window:]
|
||||
|
||||
# Update streak
|
||||
if tool_name == self._current_streak:
|
||||
self._streak_count += 1
|
||||
else:
|
||||
self._current_streak = tool_name
|
||||
self._streak_count = 1
|
||||
|
||||
# Check for fixation
|
||||
if self._streak_count >= self.threshold:
|
||||
event = FixationEvent(
|
||||
tool_name=tool_name,
|
||||
streak_length=self._streak_count,
|
||||
threshold=self.threshold,
|
||||
nudge_sent=True,
|
||||
)
|
||||
self._events.append(event)
|
||||
self._nudges_sent += 1
|
||||
|
||||
return self._build_nudge(tool_name, self._streak_count)
|
||||
|
||||
return None
|
||||
|
||||
def _build_nudge(self, tool_name: str, count: int) -> str:
|
||||
"""Build a nudge prompt to break the fixation loop."""
|
||||
return (
|
||||
f"[SYSTEM: You have called `{tool_name}` {count} times in a row "
|
||||
f"without switching tools. This suggests a fixation loop. "
|
||||
f"Consider:\n"
|
||||
f"1. Is the tool returning an error? Read the error carefully.\n"
|
||||
f"2. Is there a different tool that could help?\n"
|
||||
f"3. Should you ask the user for clarification?\n"
|
||||
f"4. Is the task actually complete?\n"
|
||||
f"Break the loop by trying a different approach.]"
|
||||
)
|
||||
|
||||
def reset(self) -> None:
|
||||
"""Reset the detector state."""
|
||||
self._history.clear()
|
||||
self._current_streak = ""
|
||||
self._streak_count = 0
|
||||
|
||||
def get_streak_info(self) -> dict:
|
||||
"""Get current streak information."""
|
||||
return {
|
||||
"current_tool": self._current_streak,
|
||||
"streak_count": self._streak_count,
|
||||
"threshold": self.threshold,
|
||||
"at_threshold": self._streak_count >= self.threshold,
|
||||
"nudges_sent": self._nudges_sent,
|
||||
}
|
||||
|
||||
def format_report(self) -> str:
|
||||
"""Format fixation events as a report."""
|
||||
if not self._events:
|
||||
return "No tool fixation detected."
|
||||
|
||||
lines = [
|
||||
f"Tool Fixation Report ({len(self._events)} events)",
|
||||
"=" * 40,
|
||||
]
|
||||
for e in self._events:
|
||||
lines.append(f" {e.tool_name}: {e.streak_length} consecutive calls (threshold: {e.threshold})")
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
# Singleton
|
||||
_detector: Optional[ToolFixationDetector] = None
|
||||
|
||||
|
||||
def get_fixation_detector() -> ToolFixationDetector:
|
||||
"""Get or create the singleton detector."""
|
||||
global _detector
|
||||
if _detector is None:
|
||||
_detector = ToolFixationDetector()
|
||||
return _detector
|
||||
|
||||
|
||||
def reset_fixation_detector() -> None:
|
||||
"""Reset the singleton."""
|
||||
global _detector
|
||||
_detector = None
|
||||
42
docs/holographic-vector-hybrid.md
Normal file
42
docs/holographic-vector-hybrid.md
Normal file
@@ -0,0 +1,42 @@
|
||||
# Holographic + Vector Hybrid Memory Architecture
|
||||
|
||||
Research issue #879. Combining HRR (holographic) and vector (Qdrant) memory.
|
||||
|
||||
## Architecture
|
||||
|
||||
Three memory backends, each with unique strengths:
|
||||
|
||||
| Backend | Strength | Weakness | Use Case |
|
||||
|---------|----------|----------|----------|
|
||||
| FTS5 | Exact keyword match | No semantic understanding | Precise recall |
|
||||
| Vector (Qdrant) | Semantic similarity | No compositional queries | Topic search |
|
||||
| HRR (Holographic) | Compositional queries | Limited scale | Complex reasoning |
|
||||
|
||||
## Why Hybrid
|
||||
|
||||
- FTS5 alone: misses ~30-40% of semantically relevant content
|
||||
- Vector alone: can't do compositional queries ("what did I discuss about X after doing Y?")
|
||||
- HRR alone: unique capability but no semantic fallback
|
||||
- Hybrid: best of all three, RRF fusion for ranking
|
||||
|
||||
## Implementation: Reciprocal Rank Fusion
|
||||
|
||||
Results from each backend are merged using RRF:
|
||||
- score = sum(weight / (k + rank)) for each backend
|
||||
- k=60 (standard RRF constant)
|
||||
- Weights: FTS5=0.6, Vector=0.4 (configurable)
|
||||
|
||||
## Status
|
||||
|
||||
- FTS5: EXISTS (hermes_state.py)
|
||||
- Vector (Qdrant): implemented (tools/hybrid_search.py)
|
||||
- HRR: EXISTS (plugins/memory/holographic.py)
|
||||
- RRF fusion: implemented (tools/hybrid_search.py)
|
||||
- Ingestion pipeline: partial
|
||||
|
||||
## Next Steps
|
||||
|
||||
1. Wire HRR into hybrid_search.py
|
||||
2. Session-level vector ingestion
|
||||
3. Benchmark: measure R@5 improvement
|
||||
4. Cross-session memory persistence
|
||||
@@ -1,76 +0,0 @@
|
||||
"""Tests for tool fixation detection."""
|
||||
|
||||
import pytest
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||
|
||||
from agent.tool_fixation_detector import ToolFixationDetector, get_fixation_detector
|
||||
|
||||
|
||||
class TestFixationDetection:
|
||||
def test_no_fixation_below_threshold(self):
|
||||
d = ToolFixationDetector(threshold=5)
|
||||
for i in range(4):
|
||||
assert d.record("execute_code") is None
|
||||
|
||||
def test_fixation_at_threshold(self):
|
||||
d = ToolFixationDetector(threshold=3)
|
||||
d.record("execute_code")
|
||||
d.record("execute_code")
|
||||
nudge = d.record("execute_code")
|
||||
assert nudge is not None
|
||||
assert "execute_code" in nudge
|
||||
assert "3 times" in nudge
|
||||
|
||||
def test_fixation_above_threshold(self):
|
||||
d = ToolFixationDetector(threshold=3)
|
||||
d.record("execute_code")
|
||||
d.record("execute_code")
|
||||
d.record("execute_code") # threshold hit
|
||||
nudge = d.record("execute_code") # still nudging
|
||||
assert nudge is not None
|
||||
|
||||
def test_streak_resets_on_different_tool(self):
|
||||
d = ToolFixationDetector(threshold=3)
|
||||
d.record("execute_code")
|
||||
d.record("execute_code")
|
||||
d.record("terminal") # breaks streak
|
||||
assert d._streak_count == 1
|
||||
assert d._current_streak == "terminal"
|
||||
|
||||
def test_nudges_sent_counter(self):
|
||||
d = ToolFixationDetector(threshold=2)
|
||||
d.record("a")
|
||||
d.record("a") # nudge 1
|
||||
d.record("a") # nudge 2
|
||||
assert d.nudges_sent == 2
|
||||
|
||||
def test_events_recorded(self):
|
||||
d = ToolFixationDetector(threshold=2)
|
||||
d.record("x")
|
||||
d.record("x")
|
||||
assert len(d.events) == 1
|
||||
assert d.events[0].tool_name == "x"
|
||||
assert d.events[0].streak_length == 2
|
||||
|
||||
def test_report(self):
|
||||
d = ToolFixationDetector(threshold=2)
|
||||
d.record("x")
|
||||
d.record("x")
|
||||
report = d.format_report()
|
||||
assert "x" in report
|
||||
|
||||
def test_reset(self):
|
||||
d = ToolFixationDetector(threshold=2)
|
||||
d.record("x")
|
||||
d.record("x")
|
||||
d.reset()
|
||||
assert d._streak_count == 0
|
||||
assert d._current_streak == ""
|
||||
|
||||
def test_singleton(self):
|
||||
d1 = get_fixation_detector()
|
||||
d2 = get_fixation_detector()
|
||||
assert d1 is d2
|
||||
@@ -44,6 +44,34 @@ from typing import Dict, Any, Optional, Tuple
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
def _format_error(
|
||||
message: str,
|
||||
skill_name: str = None,
|
||||
file_path: str = None,
|
||||
suggestion: str = None,
|
||||
context: dict = None,
|
||||
) -> Dict[str, Any]:
|
||||
"""Format an error with rich context for better debugging."""
|
||||
parts = [message]
|
||||
if skill_name:
|
||||
parts.append(f"Skill: {skill_name}")
|
||||
if file_path:
|
||||
parts.append(f"File: {file_path}")
|
||||
if suggestion:
|
||||
parts.append(f"Suggestion: {suggestion}")
|
||||
if context:
|
||||
for key, value in context.items():
|
||||
parts.append(f"{key}: {value}")
|
||||
return {
|
||||
"success": False,
|
||||
"error": " | ".join(parts),
|
||||
"skill_name": skill_name,
|
||||
"file_path": file_path,
|
||||
"suggestion": suggestion,
|
||||
}
|
||||
|
||||
|
||||
# Import security scanner — agent-created skills get the same scrutiny as
|
||||
# community hub installs.
|
||||
try:
|
||||
|
||||
Reference in New Issue
Block a user