This wiki is the durable, deeply-linked knowledge base for the project. It is written for two audiences:
- Humans browsing on GitHub.
- Agents that load these markdown files into context.
Each page is self-contained, named after a single concept, and short enough to fit comfortably in a model's context window. Pages link out to source files at canonical paths (src/lib/...) so a reader can jump from concept → implementation in one click.
- Architecture — how the layers fit together.
- Commands — every command, what it does, what it maps to.
- Tools catalog — verbatim list of upstream MCP tools.
- Output contract — JSON envelope, exit codes, error codes.
- Auth — OAuth flow, token storage, refresh.
- MCP protocol notes — transport, headers, manual replay.
- Skills — agent-skills shipped from this repo and how to install them.
- Extending — adding new commands or supporting new servers.
- Releasing — tag-based publish flow.
- Troubleshooting — common failures and fixes.
- Official Swiggy MCP server manifest: https://github.com/Swiggy/swiggy-mcp-server-manifest
- Official Swiggy MCP builder docs: https://mcp.swiggy.com/builders/docs/
- Model Context Protocol spec: https://modelcontextprotocol.io
- skills.sh format: https://skills.sh/docs
- Code paths are absolute from the repo root, e.g.
src/lib/mcp.ts. - "Server" always refers to one of the three Swiggy MCP servers:
food,instamart,dineout. - "Tool" always refers to a single MCP tool (e.g.
search_restaurants). - "Layer A" = ergonomic CLI verbs. "Layer B" = generic
servers/tools/schema/call.