The server defines 8 tools with good naming conventions (verb_noun pattern: current_context, resolve_url, resolve_port, render_env, status, slot_new, slot_release, list_all_projects). Descriptions are present and detailed (ranging 100-250 chars), explaining what each tool does and when to use it. Tool annotations are properly applied (readOnly, mutating, destroy). However, input schemas are not visible in the provided source code excerpt, only parameter names and descriptions are shown in comments. The schema structure (JSON type definitions, enums, constraints) cannot be verified. Error handling guidance is present in descriptions ('Refuses while any service is listening') but recovery paths are minimal. Composition is well-designed: tools have single responsibilities and chain via cwd/project/slot parameters. Parameter naming follows conventions (service, project, slot, cwd, confirm, cascade, format, infra_from). Some parameters accept free-form strings where enums would improve clarity (e.g., 'format' in render_env could be constrained to {dotenv, export, json, mise, direnv, claude-env} rather than relying on description).

• Expose full JSON Schema input definitions in tool registration. Verify all parameters have 'type' (string, boolean, integer) and 'required' array. Ensure enums are declared for 'format' (dotenv|export|json|mise|direnv|claude-env), 'proto' (http|tcp|ws), and 'tier' (app|infra) to prevent LLM hallucination.
• Document output schemas for each tool. Publish the ContextOut, ServiceInfo, StatusResult structures as JSON Schema in tool descriptions or a separate spec. Example: 'Returns {project: string, slot: string, slot_exists: boolean, services: [{name, proto, tier, label}]}'.
• Add error recovery guidance. When a tool refuses an operation (e.g. slot_release with running services), include: 'Call status() to list running processes, stop them, then retry with confirm=true.'
• Declare minimum/maximum for numeric parameters if any exist (page numbers, timeouts). Current code shows mostly string parameters, verify no numeric params lack bounds.
• Validate enum values at runtime and return actionable errors. If an LLM passes format='yaml', respond: 'Invalid format: yaml. Must be one of: dotenv, export, json, mise, direnv, claude-env.' This lets the agent self-correct.
• Add per-tool permission hints in descriptions (e.g. 'Requires write access to ledger' for slot_new/slot_release) to clarify agent privilege requirements.
• Test parameter name consistency across all tools. Verify 'cwd' is accepted by all tools that need it, 'project' and 'slot' parameters have consistent naming, and cross-tool references work without mapping.