Five tools with clear verb-noun naming (submit_job, get_job, list_jobs, cancel_job, get_stats). All have descriptions (194 - 392 chars, within baseline). Input schemas use Zod with type definitions and parameter descriptions. Error handling returns typed errors with recovery guidance (e.g., 'Unknown job type' lists valid types). Output structure documented via toJobView(). Key gaps: no tool annotations (readOnlyHint/destructiveHint/idempotentHint), output schemas not formally documented in tool registration, and parameter constraints (e.g., maxAttempts min=1) are in Zod but not always reflected in descriptions.

• Add tool annotations: mark get_job, list_jobs, get_stats with readOnlyHint=true; mark submit_job with destructiveHint=true (enqueues work); mark cancel_job with destructiveHint=true and idempotentHint=true (safe to retry if already cancelled).
• Document output schemas in tool descriptions. E.g., 'Returns a job object with fields: id (string), type (string), status (queued|running|succeeded|failed|cancelled), payload (object), result (object|null), createdAt (ISO 8601), updatedAt (ISO 8601).'
• Expand get_stats description: 'Returns queue health metrics: counts of jobs by status, total queued, age of oldest queued job (seconds), and active configuration (timeouts, retry limits). Use this to monitor queue load before submitting high-priority jobs.'
• Reflect all Zod constraints in parameter descriptions. E.g., 'maxAttempts: Total attempts including the first (min 1, max 100, default 3).' This prevents LLMs from passing invalid values.
• Standardize error recovery guidance. When a job is not found or not cancellable, suggest next steps: 'Try list_jobs(status="queued") to find available jobs' or 'This job is already running; call get_job() to monitor progress.'

Inferred effective spec: 2025-06-18+.

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