Drift Sentinel has 7 tools with basic schemas and descriptions, but significant gaps prevent production use. Tool names lack action verbs (ingest_traces, set_baseline are acceptable; explain_drift, calibrate, configure_suppression are vague). Most critically: 4 of 7 tools have incomplete input schemas with missing property descriptions (apply_calibration and configure_suppression have empty properties {}). Parameter descriptions are minimal (1-3 words: 'Agent identifier', 'Baseline name'). No output schemas documented. Error handling exists (SentinelError) but provides no recovery guidance. Tool descriptions are 10-15 words, meeting minimum length but lacking WHEN/WHY context. No tool annotations (readOnlyHint/destructiveHint). Composition is reasonable (single responsibility per tool), but parameter naming lacks type suffixes (keyRef is opaque; should be key_id or key_name).

• Add complete property descriptions to apply_calibration and configure_suppression schemas. Each property must have type, description, and constraints (e.g., 'configurationVersion: number, required, must be >= 0').
• Expand all parameter descriptions to 50-150 chars. Include format, range, and allowed values. E.g., 'agentId: string, 1-256 chars, alphanumeric + underscore, uniquely identifies the agent being monitored'.
• Document output schemas for all 7 tools. Specify returned fields, types, and what IDs/references are included for chaining. E.g., ingest_traces should return {traceSetRef: string, traceCount: number, ingestionTime: string}.
• Add tool annotations: mark set_baseline, apply_calibration, configure_suppression with destructiveHint=true; mark ingest_traces, check_drift, explain_drift, calibrate with readOnlyHint=true.
• Rename vague tools: 'explain_drift' → 'get_drift_explanation', 'calibrate' → 'propose_calibration', 'configure_suppression' → 'set_suppression_config'. Action verbs improve LLM intent inference.