The Powernode MCP server (platform.* tools) is the primary knowledge source; file scanning is the fallback. These queries are mandatory protocol steps for non-trivial work, not optional. The full tool catalog lives in ../../reference/auto/mcp-tools.md (regenerate via rails mcp:generate_tool_catalog).
The per-subsystem variants of this protocol live in the nested
server/CLAUDE.md,frontend/CLAUDE.md, andworker/CLAUDE.md, which load automatically when you edit those trees.
platform.knowledge_health— baseline; identify stale/conflicting knowledgeplatform.learning_metrics— active learnings, recent contributions- If stale_count > 0 or conflicts, note for resolution
platform.code_index_statusrepository_id: "powernode-platform"— index freshness
- Search existing knowledge for the area:
platform.query_learnings,platform.search_knowledge,platform.search_knowledge_graph,platform.discover_skills,platform.code_semantic_search,platform.code_identifier_search - Understand impact:
platform.code_blast_radius,platform.code_file_skeleton - Apply discovered knowledge; fall back to file scanning only when MCP returns nothing; feed file-scan discoveries back into MCP
- Relying on a learning →
platform.reinforce_learning(prevents decay) - Using shared knowledge →
platform.rate_knowledge(4-5 helpful, 1-2 outdated) - Conflicting learnings →
platform.resolve_contradiction - Cross-cutting change →
platform.search_knowledge_graph; structure →platform.code_context_tree; impact →platform.code_blast_radius
Contribute via the knowledge-lifecycle.md responsibilities table. Skip for trivial fixes, speculative analysis, or knowledge that already exists. Self-check: "Did I create learnings for the critical findings?"
Claude Code invokes platform.* tools directly via the streamable-http MCP server in .claude/settings.json (powernode → http://localhost:3000/api/v1/mcp/message). No external daemon.