MCP server exposing OMC eval / introspection / error explainer to LLM clients
OMNIcode MCP server has critical definition gaps. Tool descriptions are present but generic (avg ~80 chars, below 194-char baseline). Input schemas exist but lack depth: most parameters have minimal type info and no validation constraints. No output schemas documented. Parameters lack descriptions in several tools (omc_list has empty input {}). Error handling is absent, no recovery guidance, no categorization. Tool naming is reasonable (verb_noun pattern) but compositions like omc_eval + omc_predict suggest potential multi-responsibility issues. No evidence of idempotency guarantees, permission gates, or audit trails. Schemas are inferred from Python source rather than explicitly registered in Rust MCP server code.
Compute the canonical hash without storing — dedup-check.
Apply substrate codec for OMC source code.
Execute OMC source code and return stdout as a string
Enumerate all stored entries.
Retrieve stored content by canonical hex hash. None on miss.
Predict the next n function completions after context using the OMC predictor
Search the OMC codebase for text windows similar to the query using the dual-scale Fibonacci-harmonic address index
omc_list has no input schema (empty {} input), cannot validate or guide LLM on parameters.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., hex_hash from omc_store for use in omc_lookup). Breaks tool chaining.
Parameter descriptions are minimal or missing. 'kind' in omc_store lists values (omc_fn, json, prose, blob) but no description of what each means or when to use. 'every_n' in omc_compress lacks explanation of sampling semantics.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Sidecar metadata: kind, attractor, distance, bytes, origin.
Store arbitrary content in the substrate-keyed kernel. Returns the canonical hex hash that addresses the stored entry.
No error handling guidance. Tools like omc_eval (WRITE risk) and omc_store have no documented error cases, recovery steps, or confirmation patterns. Agents cannot distinguish retryable from fatal failures.
Tool definitions appear to be inferred from Python source (experiments/transformerless_lm/omc_tools.py, tools/mcp_substrate/server.py) rather than explicitly registered in Rust MCP server code. Cannot verify actual Rust implementation or schema registration.