JSON-Native Lambda Calculus Engine with MCP protocol implementation and pipeline orchestration
MCP-PIF provides 4 tools with mixed quality. All tools have descriptions and visible schemas in Server.hs, but several critical gaps emerge: (1) Tool names lack action verbs (evolve, run, list, help are vague or non-standard); (2) Parameter descriptions are present but minimal, most under 50 chars, failing to explain WHEN or WHY to use parameters, violating pattern:tool-description baseline of 72 chars avg; (3) Input schemas are present but incomplete, 'code' parameter type is ambiguous ('string or JSON'), and input normalization logic in handleRun suggests the schema doesn't match actual behavior; (4) No output schemas documented, run/evolve responses return JSON.String or error, but downstream tools cannot parse structure; (5) No error recovery guidance, error responses are strings, not structured errors with actionable next steps; (6) compose-ability is poor, tools create/run lambda calculus, but there's no clear way for agents to chain results (e.g., does evolve return a tool_id usable in run?). This is a domain-specific tool (lambda calculus execution) not a general-purpose API wrapper, which limits its usefulness to specialized agents.
Create a new tool from lambda calculus code
Get help information about the server and available operations
List all registered tools with their descriptions and capabilities
Execute lambda calculus code with an input value
Tool names lack action verbs; 'evolve', 'run', 'help' do not follow verb_noun convention and are ambiguous for LLM intent inference.
Parameter type ambiguity: 'code' parameter in evolve and run accepts 'string or JSON' without clear schema or format guidance, forcing LLMs to guess serialization format.
All parameter descriptions are under baseline (194 chars avg); most are 40-65 chars. Descriptions lack context for WHEN to use parameters, what values are valid, and what side effects occur.
No output schemas documented; tools return unstructured JSON strings or errors. Agents cannot parse responses or plan downstream calls. 'Created tool xyz' tells LLM nothing about tool_id or how to reference the newly created tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 39 | 2026-07-28+ | v2 |
Error responses are freeform strings with no recovery guidance. A parse error in evolve gives 'Parse error: InvalidLiteral' but no hint on valid lambda syntax or how to fix it.
No composition support documented. After evolve creates a tool, how does an agent reference it in run? Code suggests tool lookup by name, but this is not documented in run's description or output schema.
help tool has optional 'category' parameter without enum constraint. Allowed values (syntax, operations, examples) are listed in description as prose, not machine-parseable.
Input normalization in run (normalizeInput function) is not reflected in schema or documentation. LLMs cannot know that string '42' is auto-converted to number 42.