Class: Rubino::Security::DoomLoopDetector

Inherits:
Object
  • Object
show all
Defined in:
lib/rubino/security/doom_loop_detector.rb

Overview

Detects when the agent enters a doom loop - repeatedly calling the same tool with identical arguments without progress.

Two dimensions, both config-driven (Hermes tool_guardrails alignment, #414):

- threshold: how many identical consecutive calls trip detection
  (default 5; Hermes grades 5-8). The old default was 3, which hard-
  denied a legitimate 3rd retry of an idempotent read.
- hard_stop: when true, a tripped detector means BLOCK (the policy
  returns :deny). When false (the default) it WARNS but allows — the
  policy surfaces a one-time warning to the model and lets the call run.

Constant Summary collapse

DEFAULT_THRESHOLD = 
5

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(threshold: DEFAULT_THRESHOLD, hard_stop: false) ⇒ DoomLoopDetector

Returns a new instance of DoomLoopDetector.



21
22
23
24
25
 
# File 'lib/rubino/security/doom_loop_detector.rb', line 21

def initialize(threshold: DEFAULT_THRESHOLD, hard_stop: false)
  @threshold = threshold
  @hard_stop = hard_stop
  @history = []
end

Instance Attribute Details

#threshold ⇒ Object (readonly)

Returns the value of attribute threshold.



19
20
21
 
# File 'lib/rubino/security/doom_loop_detector.rb', line 19

def threshold
  @threshold
end

Instance Method Details

#hard_stop? ⇒ Boolean

True when the detector is configured to BLOCK on detection (vs. warn).

Returns:

  • (Boolean) 


28
29
30
 
# File 'lib/rubino/security/doom_loop_detector.rb', line 28

def hard_stop?
  @hard_stop == true
end

#record(tool_name:, arguments:) ⇒ Object

Records a tool call and returns true if a doom loop is detected (the last ‘threshold` calls are identical). Detection is independent of hard_stop — the caller decides whether a hit blocks or only warns.



35
36
37
38
39
40
41
42
43
44
45
46
 
# File 'lib/rubino/security/doom_loop_detector.rb', line 35

def record(tool_name:, arguments:)
  signature = generate_signature(tool_name, arguments)
  @history << signature

  # Check if the last N calls are identical
  if @history.size >= @threshold
    recent = @history.last(@threshold)
    return true if recent.uniq.size == 1
  end

  false
end

#reset! ⇒ Object

Resets the detector (e.g., when user provides new input)



49
50
51
 
# File 'lib/rubino/security/doom_loop_detector.rb', line 49

def reset!
  @history.clear
end