Architecture

llmAssistant is an intermediary Model Context Protocol (MCP) agent between client requests and default LLM-agent execution from AchillesAgentLib.

Component Overview

Component	Location	Role
Model Context Protocol contract declaration	`llmAssistant/mcp-config.json`	Declares tool contracts, schemas, command wrapper, and `TOOL_NAME` mapping.
Wrapper entrypoint	`llmAssistant/tools/llm_tool.sh`	Starts a new Node invocation for each tool call.
Dispatcher runtime	`llmAssistant/tools/llm_tool.mjs`	Parses input envelope, validates args, and routes execution to helper operations.
Commit-message helper	`llmAssistant/lib/git-commit-message.js`	Builds prompt from diffs and returns commit message text.
Conflict-resolution helper	`llmAssistant/lib/git-resolve-conflict.js`	Attempts deterministic merge first, then LLM-assisted conflict resolution.
Agent bootstrap	`llmAssistant/scripts/startAgent.sh`	Loads provider credentials from env or `.ploinky/.secrets` and starts AgentServer.

Runtime Boundary

Clients consume MCP tools only. They do not call Achilles LLM runtime directly. The agent owns prompt shaping, context clipping, and output normalization for each tool contract.

{
  "name": "llm_autocomplete",
  "command": "tools/llm_tool.sh",
  "cwd": "workspace",
  "env": {
    "TOOL_NAME": "llm_autocomplete"
  }
}

Intermediary Guarantee

Explorer and other agents receive deterministic contract outputs (ok, payload fields, explicit errors) while internal LLM provider changes remain behind this boundary.

The agent is a contract-oriented orchestration layer, not a generic free-form chat endpoint.