Skip to content

Latest commit

 

History

History
60 lines (46 loc) · 1.92 KB

AGENTS.md

File metadata and controls

60 lines (46 loc) · 1.92 KB

AGENTS.md

STACKS

STACKS: [GO]

PROJECT OVERVIEW

Persistent shared memory infrastructure for Claude Code workstations. Single server (Docker on Unraid/NAS) stores observations in PostgreSQL 17 + pgvector, exposes 48 MCP tools via Streamable HTTP / SSE on port 37777.

RULES

Rule Description
No stubs Complete, working implementations only
No guessing Verify with tools before using
Reasoning first Document WHY before implementing
No silent patching Report every discrepancy found
No time estimates Prioritize by value/risk/dependencies, not phantom duration

CONVENTIONS

  • Language: Go 1.25+
  • Build: make build
  • Test: go test ./...
  • Database: PostgreSQL 17 + pgvector (HNSW cosine index)
  • Embedding: ONNX BGE (384D) or OpenAI-compatible REST API
  • Reranking: ONNX cross-encoder reranker
  • Worker: cmd/worker/main.go — HTTP API + MCP SSE + MCP Streamable HTTP + dashboard
  • Hooks: plugin/hooks/ — JavaScript hooks for Claude Code lifecycle
  • Plugin: plugin/ — Claude Code plugin definition + marketplace

KEY DIRECTORIES

cmd/worker/          — server entry point
internal/mcp/        — MCP protocol, 48 tool handlers (tools_*.go)
internal/search/     — hybrid search (tsvector + vector + BM25, RRF fusion)
internal/scoring/    — importance + relevance scoring
internal/embedding/  — ONNX BGE + OpenAI REST providers
internal/worker/sdk/ — observation extraction (LLM API or Claude CLI)
internal/learning/   — self-learning, LLM client
plugin/hooks/        — JS hooks (session-start, user-prompt, post-tool-use, stop)

INSTRUCTION HIERARCHY

System prompts > Task/delegation > Global rules > Project rules > Defaults

SKILL LOADING

  1. Project skills (.agent/skills/) override global skills
  2. Same-name project skill completely replaces global
  3. Skills are loaded by semantic description matching