ACMS provides 7 tools with explicit fastmcp registration, each with a description and input schema. However, quality is uneven: tool names follow verb_noun convention and are reasonably clear, but parameter descriptions are often generic and lack actionable constraints. Output schemas are completely undocumented, there is no visible specification of what each tool returns, only what it accepts. Error handling is absent from the visible code; no guidance on retryability, user-fixable vs fatal errors, or recovery paths. Parameter defaults exist but lack documentation of why they were chosen. The server targets Apple's container CLI, a niche domain, but the tool interface does not fully leverage Arcade patterns for clarity and LLM reasoning.

• For every tool, add a documented output schema with typed fields. E.g., acms_registry_login should return { token: string, expiry: string, registry_url: string }. This enables downstream tool chaining.
• Add enum constraints to 'format' (table|json|yaml|csv), 'scheme' (auto|basic|oauth|bearer), and 'platform' (linux/amd64|linux/arm64|etc) parameters. Use JSON Schema enum arrays.
• For numeric params, add min/max constraints: cpus [0.5-16], memory in MB [512-65536], port [1-65535]. Add format hints in descriptions.
• For acms_builder_delete and acms_container_build, add a 'dry_run' or 'confirm' parameter to enable safe preview and explicit user approval before destructive actions.
• Add a separate error-handling section to each tool definition documenting: (a) which errors are retryable (e.g., 'Registry unreachable'), (b) which are user-fixable (e.g., 'Invalid credentials'), and (c) what the agent should do in each case.
• Expand tool descriptions to 50-150 characters, answering: What does it do? When should the LLM call it? What does it return? E.g., 'Stop the running BuildKit builder container. Gracefully shuts down and pauses in-flight builds. Returns completion status and list of affected build IDs.'
• For acms_container_build, document whether repeated calls with same inputs are idempotent (e.g., does no_cache=true re-run from scratch, or does caching still apply?). This affects agent retry safety.
• Change 'memory' parameter type from string to number (bytes) and document the unit in the description. Or provide separate fields like 'memory_mb' with a numeric type and min/max constraint.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry