Metacognitive AI agent oversight: adaptive CPI interrupts for alignment, reflection and safety
Two tools with complete input schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. vibe_check has 8 parameters, only 2 required; vibe_learn has 5 parameters, only 2 required. Both tools lack documented output schemas, making it unclear what LLMs should expect. Parameter descriptions are present but generic (e.g., 'Optional session ID for state management' does not explain what state is tracked or how sessionId affects behavior). Tool names lack action verbs (vibe_check, vibe_learn are noun-based rather than verb_noun pattern). No error handling guidance visible in tool definitions. Security concern: sessionId parameter suggests stateful session tracking, which contradicts stateless MCP protocol design.
Metacognitive questioning tool that identifies assumptions and breaks tunnel vision to prevent cascading errors
Pattern recognition system that tracks common errors and solutions to prevent recurring issues
Tool names lack action verbs. 'vibe_check' and 'vibe_learn' are noun-based; should follow verb_noun pattern (e.g., 'check_vibe', 'record_learning'). LLMs infer intent from names, ambiguous names reduce tool selection accuracy.
Output schemas not documented. Neither tool specifies what fields the response contains, forcing LLMs to guess structure. vibe_check likely returns analysis/questions; vibe_learn likely returns confirmation. Undocumented outputs break downstream tool chaining.
sessionId parameter suggests stateful session tracking, contradicting MCP's stateless protocol design (2026-07-28 spec). Each request should be self-contained. If session state is needed, pass it explicitly in each request or use server-side storage keyed by a request-scoped identifier.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Parameter descriptions lack actionable detail. 'Optional session ID for state management' does not explain what state is tracked, how it affects behavior, or when to use it. 'The agent's uncertainties' does not specify format (comma-separated? JSON array? free text?).
No error handling guidance in tool definitions. If vibe_check fails to reach the LLM provider (Gemini/OpenAI/OpenRouter), what should the agent do? Retry? Fall back? Ask the user? Undocumented error paths force LLMs to guess recovery strategies.