mcp_skill_provider: Instructions file for Codex

Agent Notes for mcp_skill_provider

Project Basics

• Python project managed via pyproject.toml (setuptools backend).
• Package name: mcp_skill_provider.
• No CLI entrypoint — installed as a package and loaded by mcp_daemon_engine.
• Dev dependencies: pytest, black, ruff.
• Uses pyhumps (not humps) for decamelize — see pyproject.toml comment.

Architecture

mcp_skill_provider is a flat composition of mixins that expose Harness Engineering skill tools through mcp_daemon_engine.

mcp_skill_provider/
  __init__.py                # Exports MCPSkillProvider + MCP_CONFIGURATION
  mcp_skill_provider.py      # Facade class composing mixins
  mcp_configuration.py       # MCP_CONFIGURATION dict (tools + inputSchema)
  graphql_client.py          # GraphQLClient → HTTP/2 to harness_engineering_engine
  graphql_backed_processor.py # Root base: logger, setting, graphql_client
  skill_mixin.py             # search_skills, get_skill
  command_mixin.py           # run_command
  error_handler.py           # Decorators + skill-specific ErrorCode constants

How it connects to mcp_daemon_engine

mcp_daemon_engine loads MCP modules via the loadMcpConfiguration GraphQL mutation. The MCP_CONFIGURATION dict declares:

• Module name: mcp_skill_provider
• Class: mcp_skill_provider.mcp_skill_provider.MCPSkillProvider
• Three functions: search_skills, get_skill, run_command

GraphQLClient.execute_query() posts to the module's configured graphql_modules.harness_engineering_engine.endpoint with either a silvaengine_gateway JWT Bearer token (preferred) or x-api-key header. The graphql_modules key must match the real importable backend package name (harness_engineering_engine) — Graphql.get_graphql_schema() does a real importlib lookup on it, and class_name must be a bare attribute exported at that package's top level (HarnessEngineeringEngine), not a dotted path.

Code Style

• Follow SilvaEngine conventions: __future__ import, __author__ = "bibow" header, type hints.
• All GraphQL calls go through GraphQLClient.execute_query() — never import harness_engineering_engine directly.
• Field names returned from GraphQL are camelCase; the mixins call humps.decamelize() to normalize to snake_case before returning to the MCP caller.
• Errors use the error_handler decorator pattern: @handle_errors(operation_name="...") on every public method, with propagate_error_if_present() used to forward backend errors.