Architecture

RAYS-CORE is a terminal-first orchestration layer around static indexing, embeddings, LLM-guided analysis, planning, permission negotiation, code generation, optional shell execution and session memory.

High-level pipeline (edit mode)

1. Config — config.yaml (bundled + user defaults).
2. Indexing — populates .rays/ with symbol/relationship registries.
3. Vector DB — Chroma-compatible store for semantic candidate lookup.
4. Task analysis — classifies intent, extracts keywords/tools.
5. Symbol workflow — retrieval, merging, explicit-mention parsing.
6. Planning — structured implementation plans within permission envelopes.
7. Permissions — negotiates scopes before edits.
8. Anchoring — resolves where new symbols/files belong.
9. Execution — emits and applies edits.
10. Terminal engine — optional command intents with safety modes.
11. Memory — summaries and embeddings tied to .rays.

Alternate paths

• Default prompt — agent orchestrator (skills + MCP).
• /code — full coding pipeline (indexing, symbols, planning, execution).
• /chat — read-only retrieval and answer synthesis.
• /mcp — list configured MCP servers and connection status.

Agent orchestrator flow

1. Discover skills from skills/ and ~/.rays/skills/.
2. Connect MCP servers from config / ~/.rays/mcp.json.
3. LLM selects required skills and MCP servers.
4. LLM builds a spawn plan with a per-step spawn_reason.