Compare commits

..

1 Commits

Author SHA1 Message Date
Alexander Whitestone
ab968e910c feat: tool fixation detection — break repetitive loops (#886)
Some checks failed
Contributor Attribution Check / check-attribution (pull_request) Failing after 37s
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 43s
Tests / e2e (pull_request) Successful in 1m57s
Tests / test (pull_request) Failing after 18m57s
Marathon sessions show tool fixation: agent latches onto one tool
and calls it repeatedly. Observed streaks of 8-25 identical calls.

New agent/tool_fixation_detector.py:
- ToolFixationDetector: tracks consecutive tool calls
- record(tool_name): returns nudge prompt when threshold reached
- Default threshold: 5 consecutive calls (configurable via
  TOOL_FIXATION_THRESHOLD env var)
- Nudge prompt explains the fixation and suggests alternatives:
  1. Read error carefully
  2. Try different tool
  3. Ask user for clarification
  4. Check if task is complete
- get_streak_info(): current streak state
- format_report(): human-readable fixation events
- Singleton via get_fixation_detector()

Config:
- TOOL_FIXATION_THRESHOLD (default: 5)
- TOOL_FIXATION_WINDOW (default: 10)

Tests: tests/test_tool_fixation_detector.py (9 tests)

Closes #886
2026-04-17 01:57:37 -04:00
4 changed files with 232 additions and 228 deletions

View File

@@ -1,149 +0,0 @@
"""
988 Suicide & Crisis Lifeline Integration (#673).
When crisis is detected, provides immediate access to help:
- Phone: 988 (call or text)
- Text: Text HOME to 988
- Chat: 988lifeline.org/chat
- Spanish: 1-888-628-9454
- Emergency: 911
This module provides the resource data. agent/crisis_protocol.py
handles detection. This module formats the resources for display.
"""
from dataclasses import dataclass
from typing import List
@dataclass
class CrisisResource:
"""A crisis support contact method."""
name: str
contact: str
description: str
url: str = ""
available: str = "24/7"
language: str = "English"
# 988 Suicide & Crisis Lifeline — all channels
LIFELINE_988 = CrisisResource(
name="988 Suicide and Crisis Lifeline",
contact="Call or text 988",
description="Free, confidential support for people in suicidal crisis or emotional distress.",
url="https://988lifeline.org",
available="24/7",
language="English",
)
LIFELINE_988_TEXT = CrisisResource(
name="988 Crisis Text Line",
contact="Text HOME to 988",
description="Free, 24/7 crisis support via text message.",
url="",
available="24/7",
language="English",
)
LIFELINE_988_CHAT = CrisisResource(
name="988 Lifeline Chat",
contact="988lifeline.org/chat",
description="Free, confidential online chat with a trained crisis counselor.",
url="https://988lifeline.org/chat",
available="24/7",
language="English",
)
LIFELINE_988_SPANISH = CrisisResource(
name="988 Lifeline (Spanish)",
contact="1-888-628-9454",
description="Línea de prevención del suicidio en español.",
url="https://988lifeline.org/help-yourself/en-espanol/",
available="24/7",
language="Spanish",
)
CRISIS_TEXT_LINE = CrisisResource(
name="Crisis Text Line",
contact="Text HOME to 741741",
description="Free, 24/7 crisis support via text message.",
url="https://www.crisistextline.org",
available="24/7",
language="English",
)
EMERGENCY_911 = CrisisResource(
name="Emergency Services",
contact="911",
description="Immediate danger — police, fire, ambulance.",
url="",
available="24/7",
language="Any",
)
# All resources in priority order
ALL_RESOURCES: List[CrisisResource] = [
EMERGENCY_911,
LIFELINE_988,
LIFELINE_988_TEXT,
LIFELINE_988_CHAT,
CRISIS_TEXT_LINE,
LIFELINE_988_SPANISH,
]
def get_crisis_resources(language: str = None) -> List[CrisisResource]:
"""Get crisis resources, optionally filtered by language.
Args:
language: Filter by language ("English", "Spanish", or None for all)
Returns:
List of CrisisResource objects
"""
if language:
return [r for r in ALL_RESOURCES if r.language.lower() == language.lower()]
return ALL_RESOURCES
def format_crisis_resources(resources: List[CrisisResource] = None) -> str:
"""Format crisis resources as a user-facing message.
Args:
resources: List of resources to format. Defaults to all resources.
Returns:
Formatted string suitable for displaying to a user in crisis.
"""
if resources is None:
resources = ALL_RESOURCES
lines = ["**Please reach out — help is available right now:**
"]
for r in resources:
if r.url:
lines.append(f"- **{r.name}:** {r.contact} ({r.url})")
else:
lines.append(f"- **{r.name}:** {r.contact}")
lines.append("")
lines.append("All services are free, confidential, and available 24/7.")
lines.append("You are not alone.")
return "
".join(lines)
def get_immediate_help_message() -> str:
"""Get the most urgent crisis help message.
Used when crisis is detected at CRITICAL level.
"""
return (
"If you are in immediate danger, call **911** right now.
"
+ format_crisis_resources()
)

View File

@@ -0,0 +1,156 @@
"""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

View File

@@ -1,79 +0,0 @@
"""Tests for 988 Crisis Lifeline integration (#673)."""
import pytest
from agent.crisis_resources import (
LIFELINE_988,
LIFELINE_988_TEXT,
LIFELINE_988_CHAT,
LIFELINE_988_SPANISH,
CRISIS_TEXT_LINE,
EMERGENCY_911,
ALL_RESOURCES,
get_crisis_resources,
format_crisis_resources,
get_immediate_help_message,
CrisisResource,
)
class TestCrisisResources:
def test_988_phone(self):
assert "988" in LIFELINE_988.contact
assert "24/7" in LIFELINE_988.available
def test_988_text(self):
assert "HOME" in LIFELINE_988_TEXT.contact
assert "988" in LIFELINE_988_TEXT.contact
def test_988_chat(self):
assert "988lifeline.org/chat" in LIFELINE_988_CHAT.url
def test_988_spanish(self):
assert "1-888-628-9454" in LIFELINE_988_SPANISH.contact
assert LIFELINE_988_SPANISH.language == "Spanish"
def test_crisis_text_line(self):
assert "741741" in CRISIS_TEXT_LINE.contact
def test_911(self):
assert "911" in EMERGENCY_911.contact
def test_all_resources_not_empty(self):
assert len(ALL_RESOURCES) >= 5
class TestGetResources:
def test_returns_all_by_default(self):
assert len(get_crisis_resources()) == len(ALL_RESOURCES)
def test_filter_english(self):
english = get_crisis_resources("English")
assert all(r.language == "English" for r in english)
assert len(english) > 0
def test_filter_spanish(self):
spanish = get_crisis_resources("Spanish")
assert len(spanish) >= 1
assert all(r.language == "Spanish" for r in spanish)
class TestFormatting:
def test_format_includes_988(self):
msg = format_crisis_resources()
assert "988" in msg
def test_format_includes_741741(self):
msg = format_crisis_resources()
assert "741741" in msg
def test_format_includes_911(self):
msg = format_crisis_resources()
assert "911" in msg
def test_immediate_help_includes_911_first(self):
msg = get_immediate_help_message()
assert msg.startswith("If you are in immediate danger")
def test_format_not_empty(self):
msg = format_crisis_resources()
assert len(msg) > 100

View File

@@ -0,0 +1,76 @@
"""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