[Infra][Agent] Implement MCP Server & Codex Skill Definition

[copparihollmann]

Context

To enable autonomous agents (like Claude Desktop or Codex) to interact with our compiler tooling, we need a standardized interface. The Model Context Protocol (MCP) provides this bridge, converting our Python functions into agent-callable tools. Additionally, defining a "Skill" with documentation references allows the agent to understand how and when to use these tools effectively.

Objective

Implement the mlirAgent MCP server, a unified CLI, and the .codex skill definition to expose our deterministic tools (build, compile, trace) to AI agents.

Scope of Work

1. Unified CLI (src/mlirAgent/cli.py): * Implement a central argparse entry point wrapping run_build, run_compile, verify_output, and trace_provenance.
   • Ensure it supports both direct shell usage and library import.
2. Configuration (src/mlirAgent/config.py): * Implement Config class to load paths from .env.
   • Add a validate() method to warn if IREE_SRC_PATH or ninja are missing.
3. MCP Server (src/mlirAgent/mcp_server.py): * Implement FastMCP("mlirEvolve").
   • Expose tools: build, compile_mlir, verify_ir, provenance_trace.
   • Ensure error handling returns structured JSON, not Python stack traces.
4. Codex Skill Definition (.codex/skills/mlir-evolve/):
   • SKILL.md: Define the persona and capabilities (Build, Dialect, Provenance).
   • references/*.md: Add repo-map.md, infra.md, and tool-conventions.md to guide the agent's reasoning.

Acceptance Criteria

• Test 1 (CLI): python src/mlirAgent/cli.py compile --help returns valid help text.
• Test 2 (Config): Config.validate() correctly identifies if IREE_SRC_PATH is invalid.
• Test 3 (MCP): Starting the MCP server exposes the provenance_trace tool with the correct argument schema (artifacts_root, source_filename, line).
• Test 4 (Documentation): The .codex directory structure is populated, and SKILL.md correctly references the src/mlirAgent/tools/ layout.