ArcRift exposes 8 tools with complete input schemas and descriptions. Naming follows verb_noun convention (recall_, store_, prune_, search_, list_, get_, identify_, index_). Descriptions are substantive (100-200 chars), explaining WHAT and WHEN. However, output schemas are not documented in the visible code, LLMs cannot predict response structure. Parameter descriptions are present but lack constraints (enums, ranges, formats). Error handling and recovery guidance are absent. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Security concerns: prune_memory and index_codebase are destructive but lack confirmation patterns.

• Document output schemas for all 8 tools. Include field names, types (string, number, array, object), and example values. E.g., recall_context returns {chunks: [{id, text, score, project}], total: number}.
• Add tool annotations: mark recall_context, search_memory, list_projects, get_project_summary, identify_active_project with readOnlyHint=true; mark prune_memory, store_memory, index_codebase with destructiveHint=true.
• Implement confirmation pattern for prune_memory and index_codebase. Add a 'confirm' boolean parameter (default false) or return a confirmation token that must be passed in a follow-up call.
• Add parameter constraints: topN as {type: number, minimum: 1, maximum: 6}; directoryPath with pattern for absolute paths; project with description of auto-create behavior.
• Enhance error descriptions with recovery guidance. E.g., 'No chunks found for prompt. Try: (1) broaden the search with search_memory(), (2) call store_memory() to add relevant context, (3) check project ID with list_projects().'.

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

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