v1.3.199

v1.3.199

Highlights

• 16 new AI agent intelligence detectors: Added 16 deterministic, zero-LLM-cost detectors that identify frontier agentic failure patterns including scope creep, strategy exhaustion, build idempotency violations, premature success claims, goal drift, orphaned files, and more.
• Panic safety hardening: All 53 new detector fields verified for nil-safety with constructor initialization and guarded type assertions.

New Features

AI Agent Intelligence Detectors (16 detectors)

All detectors are deterministic (zero LLM cost), non-blocking (guidance injection only), and based on frontier AI agent research:

Detector	Research Basis	Pattern
History Error Accumulation	Error compounding in agentic loops	Detects escalating error density across tool call history
Scope Overgeneralization	SICA (arXiv:2504.15228)	Flags edits exceeding requested scope boundaries
Search Result Invalidation	Context engineering (ICLR 2026)	Detects continued use of invalidated search results
Truncated Output Completeness Fallacy	ACE (ICLR 2026)	Identifies false completion claims on truncated output
Strategy Exhaustion	AgentDebug (arXiv:2509.25370)	Detects repeated failed strategy without adaptation
Build Idempotency Violation	Build system analysis	Flags non-deterministic build results
Green Build Illusion	ICSE 2025	Detects claiming success without actual verification
Premature Success Declaration	FSE 2025	Identifies success claims before verification completes
Todo Contract Drop	Task tracking analysis	Detects abandoned todo items without completion
Working-Tree Invalidation	File freshness research	Warns when external changes invalidate prior reads
Context-Length Goal Drift	RECAP 2026	Detects goal drift correlated with context length growth
Orphaned New File Integration	FSE 2025, SICA	Flags new files not integrated into build/imports
Undo-Edit Blind Continuation	AgentDebug	Detects continuing after undo without re-reading state
Verification Coverage Gap	arXiv:2607.05775	Identifies insufficient test coverage for changes
Delayed Observation Contradiction	DRIFT 2026, Task2Quiz	Detects contradictions from stale observations
Bridging Rationalization	RECAP 2026	Identifies rationalized bridging logic from stale context

Safety & Reliability

• All new detector state fields initialized via constructor functions
• Safe JSON deserialization with nil guards
• Regex group access guarded with length checks
• Type assertions use safe ok pattern throughout