Harneloop exposes 26 tools with consistent naming (verb_noun pattern) and descriptions present for all tools. However, critical gaps undermine quality: (1) Input schemas lack type information beyond basic type hints (Path, string, list[str], boolean), no JSON Schema with constraints, enums, or ranges; (2) Parameter descriptions are minimal (10-30 chars) and lack actionable guidance on format, constraints, or dependencies; (3) No output schemas documented, LLMs cannot predict return structure; (4) Error handling is absent from tool definitions; (5) No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite 15+ WRITE operations. Tool names are clear (init_unit, create_attempt_plan, promote_candidate) but descriptions are terse. Example: 'create_attempt_plan' has description 'Create an agent-authored attempt plan for testing a harness' (67 chars) but parameters like 'goal', 'method', 'action' lack format guidance. No evidence of pagination, batch operations, or composition patterns.

• Add JSON Schema with type, enum, minLength, maxLength, pattern, and default for all parameters. Example: 'template' should be enum: ['blank', 'standard', 'advanced'] with description 'Template type for unit initialization.'
• Expand parameter descriptions to 50-100 chars with format guidance. Example: 'goal' in create_attempt_plan should be 'The goal of the attempt (1-500 chars, plain text, no special characters).'
• Document output schemas for all tools. Example: create_attempt_plan should return {id: string, goal: string, method: string, created_at: ISO8601, observations: array}.
• Add tool annotations: mark WRITE tools with destructiveHint=true (init_unit, promote_candidate, rollback_unit) and readOnlyHint=true for READ tools (list_candidates, validate_unit, run_doctor).
• Add error handling guidance to tool descriptions. Example: 'If unit_root does not exist, returns error 'Unit not found at {path}. Create with init_unit first.' Retryable: no.'