Loading ToolBench
Cognitive reasoning for AI agents — abstraction, contradiction detection, and insight surfacing. Plugin for cortex-engine.
This plugin provides 7 well-intentioned cognitive tools with clear domain focus (memory abstraction, contradiction detection, signal surfacing). Tool names are action-oriented (abstract, contradict, surface, resolve, query_explain). Descriptions are present and moderately detailed (80-180 chars), explaining WHAT and WHEN. However, input schemas lack critical details: parameter descriptions are minimal or missing for several fields, output schemas are not documented, and error handling is generic. The plugin is a cortex-engine adapter, not a standalone MCP server, transport mechanism is unknown/unclear. Code shows proper TypeScript typing and validation logic, but LLM-facing documentation (descriptions, param guidance) needs strengthening for agent reliability.
Propose a single higher-level concept (name and definition) that subsumes the given 2-10 memories. Does not write to the graph; returns the proposal only.
Record a contradiction between an observation and a belief or memory. Creates a CONTRADICTION signal. Provide observation_id and exactly one of belief_id or memory_id.
Prospective memory — set a reminder to surface when a trigger condition occurs. Example: trigger="when user mentions X", content="remind about Y". Use action=list to show pending, action=fire to mark as fired, action=cancel to delete.
Store an observation without embedding (fast path). Embedding happens during the next embed-pending job. Use for low-latency logging.
Semantic search over memory with one-sentence relevance explanations. Returns each result plus a "why" string explaining why that memory is relevant to the query.
Mark a signal as resolved with an optional note explaining how it was addressed.
Output schemas not documented. Tools return unstructured responses (e.g., abstract returns {proposed_name, proposed_definition} but no schema declaration). LLMs cannot plan downstream calls or validate response structure.
Parameter descriptions are sparse or missing. 'namespace' appears in all tools with identical minimal description. 'limit' in surface has no guidance on range or default behavior. 'action' enum in intention lacks per-value documentation.
Error handling is generic. Tools return {error: 'message'} without recovery guidance. E.g., 'Memory not found' does not suggest calling query_explain to search for the ID. Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
List unresolved signals — contradictions, tensions, and gaps in the knowledge graph.
Transport mechanism unknown. Package.json lists @fozikio/cortex-engine as peer dependency but does not declare MCP transport (HTTP, SSE, STDIO). Cannot assess protocol readiness without knowing how this plugin is exposed to MCP clients.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot distinguish safe reads (abstract, surface, query_explain) from writes (contradict, notice, resolve, intention). Increases risk of unintended side effects.