Agent Intelligence: Metacognitive Monitoring

[samugit83]

Description

Add a metacognitive monitoring system that runs alongside the think node — detecting reasoning loops, assessing confidence, identifying knowledge gaps, and automatically recommending strategy changes before the agent wastes iterations.

Why this matters

The agent's current failure detection is primitive: "3+ consecutive failures triggers a warning in the prompt." This misses most stuck patterns:

1. Oscillation loops: the agent alternates between nmap and nuclei endlessly (A→B→A→B) — each individual tool "succeeds" (returns output) but makes no progress. The current detector doesn't catch this because there are no consecutive failures.
2. Confidence collapse: after 5 failed exploit attempts, the agent keeps trying variants of the same approach with decreasing likelihood of success. It doesn't recognize that its confidence should be near-zero and it should switch strategies entirely.
3. Knowledge gaps masquerading as tool failures: the agent runs search CVE-2024-XXXXX in Metasploit, gets "no results," and tries again with slightly different syntax. The real problem is the CVE has no Metasploit module — the agent should web_search for alternative exploit code or try a different approach. Without knowledge gap detection, it just retries.
4. Progress stalls: the agent has been in the "informational" phase for 20 iterations. It keeps gathering more recon data instead of transitioning to exploitation. Nothing is "failing" — it's just not making meaningful forward progress.

What needs to be built

• MetacognitiveMonitor class with four detection methods:
  • _detect_loop() — exact repeats AND oscillation patterns in last 6 tool calls
  • _assess_confidence() — success/failure ratio + findings count → 0-1 score
  • _detect_knowledge_gap() — "unknown", "not found", "no results" error patterns
  • _check_progress() — iterations since last meaningful finding or phase change
• MetacognitiveReport output model (loop_detected, confidence_level, knowledge_gap, progress_stalled, strategy_recommendation)
• Integration into think node: inject warnings into system prompt when issues detected
• Strategy recommendations: "CHANGE_STRATEGY", "ESCALATE" (ask user), "RESEARCH" (web_search)
• Confidence threshold for triggering deeper reasoning (ties into System 1/2 if implemented)

What already exists

• Basic 3-failure loop detection in prompts/base.py (lines 642-671)
• execution_trace in AgentState with full tool call history
• current_iteration counter
• Phase tracking (informational/exploitation/post_exploitation)

Implementation notes

• Pure Python, no infrastructure changes needed
• Runs as a synchronous check before each think node LLM call
• Warnings are injected as text into the system prompt (no architectural changes)
• Can be toggled via project settings if needed