Agent Tooling

Dispatcher skill for three recurring "which surface" decisions when building out an agent-driven workflow. Each heading below is a short teaser — read the linked references/*.md file for the full criteria, tables, and gotchas.

MCP Server Selection

Whether a new capability belongs behind an MCP server versus a CLI call, a WebFetch, or no new tooling at all — plus the concrete anti-patterns (duplicating a CLI already in reach, unauthenticated public APIs, always-on connections, unmaintained servers) that should stop you before you wire one up. Read references/mcp-selection.md.

Mermaid Diagram Rendering

mmdc (mermaid-cli) as the sanctioned local renderer, the syntax gotchas that silently break a diagram (<br/> vs \n, comments inside subgraphs, special characters in edge labels), and why generating a Mermaid Live URL as a shortcut is a trap, not a convenience. Read references/mermaid-diagrams.md.

Skill vs Agent Mapping

How to decide whether a piece of reusable behavior should ship as a skill (on-demand knowledge loaded into an existing agent's context) or a dedicated agent (a separately-dispatched actor with its own tool scope and context window), plus the discovery patterns an agent uses to find skills at all. Read references/skill-vs-agent-mapping.md.

Thinking Patterns (frame before acting)

1. Config presence != runtime liveness. An MCP server listed as "connected," or a skill cited in a routing table, both describe intent to run — not confirmation it ran. Before trusting a mechanism, verify it actually fired (a log line, a matcher test, a liveness check) rather than trusting that it's configured.
2. A benchmark number is a claim about a specific comparison, not a universal ranking. A token-cost or latency benchmark for tool A vs tool B is only valid while that comparison holds — if the compared-against tool changes or disappears, the number needs re-measuring, not re-quoting.
3. An interactive CLI's silence about non-interactive flags is not permission to loop. Piping confirmation answers into a prompt that reads from the terminal directly (not stdin) doesn't fail loudly — it hangs or silently no-ops. Run --help once and look for --yes/--no-interactive/--force before piping anything at an unfamiliar CLI.