subagent-mcp

Use this checklist for repository work. The normative policy is

Build, package, install, configure, debug, and optimize Model Context Protocol (MCP) servers for Claude (Desktop, Code CLI) and Codex CLI on Windows and macOS. Use when building MCP tools, creating AI tool integrations, give Claude access to database or API or filesystem, connect AI agent to external tools or services, packaging stdio or HTTP servers, configuring mcpServers JSON or TOML, troubleshooting spawn ENOENT or connection errors, designing tool/resource/prompt schemas, implementing JSON-RPC handlers, publishing npm or Python MCP packages, setting up uvx or npx installable servers, debugging tool not appearing in Claude or Codex, writing secure MCP server code, server architecture review, MCP protocol compliance, FastMCP, @modelcontextprotocol/sdk, mcp Python library, MCP Inspector usage, Claude Desktop config, Claude Code CLI mcp add command, Codex config.toml, MCP server security, rate limiting MCP tools, MCP over SSE or HTTP transport, stdio transport pitfalls, environment variable injection for MCP, cross-platform MCP deployment, MCP server instructions field, tool search optimization, alwaysLoad, server descriptions for AI discovery, mcpb Desktop Extensions.

Impartially PROFILE the cross-provider sub-agent fleet against the FIXED 14 work-categories (directly benchmarked parents + 4 composite-inferred) whenever a new model ships (or on demand). Discover every model published in the recent window by the in-scope provider families, gather ALL public benchmark scores + statistics, map them onto the directly benchmarked parent categories (composites composed from parents, never directly benchmarked), then JUDGE each model+effort pairing into per-category tier rankings (best→worst) SOLELY from the discovered research, with a recorded rationale per tier placement. Emits exactly 3 artifacts: routing-table.json (lean), routing-table-audit.json (full provenance), research-seed-sites.json (accumulating learned source list). Single-family and multi-family are both fully-supported, first-class paths; provider mix is optional. The 14 categories are immutable inputs — this skill never derives, chooses, renames, reorders, or reshuffles them. Use when a new model is released, when asked to profile new model, re-profile models, re-profile the fleet, rebalance routing, update routing table, refresh model profiles, re-run model research, regenerate routing-table.json, refresh tier rankings, or to answer "which model for X now" after a model launch. Orchestrator-only pipeline: model discovery + maximalist benchmark research, pivotal-question interview, flagship judging + merge, 3-artifact emission, 3-pass adversarial validation, and provider/seed validators + scenario routing tests. Sub-agents dispatched via `mcp__subagent-mcp__launch_agent`; cross-family critics are available when ≥2 families are reachable.

Install, deploy, register, or set up the subagent-mcp addon (MCP server + orchestration-mode hooks) standards-compliant across every supported vendor (currently Claude Code CLI and Codex CLI; expandable). Ships ALL parts — MCP stdio server, per-turn hooks, and directive assets — to a PERMANENT, repo-decoupled location and wires each vendor with its official mechanism. Use when asked to "install subagent-mcp", "install the mcp", "install this", "deploy subagent-mcp", "set up subagent-mcp", "add subagent-mcp to Claude/Codex", "register the addon", "globally install subagent-mcp", "reinstall/repoint subagent-mcp", or "update the subagent-mcp install". HARD RULE: never install or run the addon from a worktree, %TEMP%/TMP/tmp, Downloads, or any temporary/ephemeral path — install only to a permanent global location. Stay compliant to the official vendor specs listed in references/compliance.md.

2 hooks across 2 events