Loading ToolBench
Coordinate index layer for LLM context — exposes cymatix knowledge store as MCP tools for Claude Code, Claude Desktop, and other MCP hosts. Provides retrieval, ingestion, session management, and introspection tools.
Server exposes 23 tools with significant definition gaps. Most tools lack descriptions adequate for LLM selection (many under 100 chars, some generic). Input schemas are present but sparse, many parameters lack type constraints and descriptions. No output schemas documented. Tool names are domain-specific but inconsistent in verb-noun convention (e.g., 'cymatix_context' vs 'cymatix_ingest'). 8 tools are aliases of others, creating redundancy. No error handling guidance visible. Security concerns: session_id and intent parameters suggest stateful session tracking, which conflicts with stateless MCP design. Overall pattern adherence is weak.
Register or update session presence in the registry
Federation/bridge inbox and signal state
Distill the session buffer into consolidated knowledge documents
Main retrieval tool — query compressed context for a topic from the knowledge store
Agent-safe packet with freshness labels and refresh plan per agent-context-index build spec
Alias for cymatix_fingerprint — document fingerprinting
Alias for cymatix_gene_get — fetch a single document by ID
8 tools are aliases (cymatix_document_*) duplicating core functionality. LLMs waste reasoning cycles choosing between cymatix_context and cymatix_document_query. Consolidate to one canonical name per operation.
Descriptions are too generic and short (avg ~50 chars). 'Query compressed context for a topic' does not explain WHEN to use cymatix_context vs cymatix_neighbors vs cymatix_splice_preview. LLMs cannot disambiguate. Expand to 100-200 chars with use-case guidance.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., does cymatix_context return gene_id for use in cymatix_gene_get?). Document return types for all 23 tools.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | <=2025-11-25 | v2 |
Alias for cymatix_neighbors — top-k SEMA neighbors for a query
Alias for cymatix_splice_preview — dry-run retrieval pipeline
Alias for cymatix_context — query compressed context for a topic
Fetch a single document by ID
Compressor / documents / upstream readiness probe
Record a Human-In-The-Loop pause event
Query recent HITL (Human-In-The-Loop) events
Add content to the knowledge store; content is compressed and stored for future context queries
Session and lifetime token counters
Top-k SEMA neighbors for a query (light)
Query the reread plan for an edit/ops task
Four-primitive introspection chart (ΣĒMA + cymatic + harmonic + neighbor set)
Documents authored by a handle, in chronological order
List active participants; filter by party, status, workspace
Dry-run retrieval pipeline (skip splice)
Knowledge store health metrics, size, compression ratio, and delta-epsilon history
session_id and intent parameters in cymatix_context suggest stateful session tracking. MCP 2026-07-28 spec requires stateless request handling, each request must be self-contained. Remove session_id; pass context via tool parameters or resources.
No error handling guidance. Tools return READ_ONLY or WRITE risk labels but no recovery hints. If cymatix_ingest fails, what should the LLM do? Add error responses with actionable next steps (e.g., 'Storage full. Try cymatix_consolidate() to free space.').