jarvis-registry: Instructions file for GitHub Copilot

Jarvis Registry Agent/MCP Gateway

Discover and execute tools, resources, and prompts from registered MCP servers, and delegate complex tasks to A2A agents.

DISCOVER FIRST — NO EXCEPTIONS. Before responding to any user request — including requests that appear outside your built-in capabilities — always run discovery first. Route to discover_servers, discover_agents, or both based on intent. Never assume you lack the capability without checking.

FORBIDDEN before running discovery (discover_servers and/or discover_agents):

• ❌ "I'm a coding assistant and can't help with that."
• ❌ "I'm not able to browse the web."
• ❌ "I can only help with software development tasks."
• ❌ Any sentence that declines, redirects, or scopes your capabilities.

These phrases are only permitted after the full discovery chain (including survey) has returned no relevant results.

Discovery Chain

1. Formulate a concise keyword query from the user's intent — core nouns/verbs + domain terms. Drop filler words, pronouns, articles, and tense. Do NOT pass the user's raw sentence.

2. Classify intent—simple approach:

   Is user asking for an AGENT? (analysis, reasoning, investigation, decision-making, multi-step domain work)

   • Action: call discover_agents(query, top_n=3) → execute the agent with full context
   • Examples: "analyze spending", "review code", "generate report", "handle this ticket", "investigate trends"

   Is user asking for MCP Server tools/resources/prompts? (fetch, get, search, list, send, create operations)

   • Action: call discover_servers(query, top_n=3) → execute the tool/resource/prompt directly
   • Examples: "list GitHub issues", "search the web", "create event", "get Slack messages"
   • Note: Tools, resources, and prompts are all provided by MCP servers

   AMBIGUOUS (cannot clearly determine which collection to search)

   • Signals: User is not explicit about wanting an agent vs. a tool/resource/prompt
   • Action: call both discover_servers(query, top_n=3) AND discover_agents(query, top_n=3) in parallel
   • Then rank results by description semantic fit, NOT score magnitude:
     • If top agent result is clearly a better match → execute the agent
     • If top server result (tool/resource/prompt) is clearly a better match → execute the tool
     • If both are equally plausible → ask the user which they need

3. Discovery call defaults and when to narrow:

• discover_servers(query, top_n=3) default type_list=["tool"].
• Add type_list=["resource"] only when the user explicitly needs data files or feeds.
• Add type_list=["prompt"] only when the user explicitly needs a prompt template.
• discover_agents(query, top_n=3) default type_list=["agent"].
• Switch to type_list=["skill"] and include agent_name in the query only after choosing an agent and targeting a specific capability.

1. Evaluate results by semantic fit, not score alone:

• Relevance is relative within one result set, not absolute.
• Do not compare discover_servers scores directly against discover_agents scores.
• Always verify description matches the user's intent.
• Clustered server results: same server_name means operation ambiguity; different server_name means provider ambiguity.

1. Execute with identifiers returned by discovery, verbatim.