Breaking Changes
All servers now default to lifecycle: "lazy" — they connect on first tool call instead of eagerly at session start. Set lifecycle: "keep-alive" or lifecycle: "eager" per-server to restore old behavior.
What's New
Lazy startup with metadata cache. Servers only connect when their tools are actually called. Persistent metadata cache (~/.pi/agent/mcp-cache.json) enables search/list/describe without live connections. Per-server config hashing with 7-day staleness.
Idle timeout. Connected servers disconnect after 10 minutes of inactivity (configurable via settings.idleTimeout or per-server idleTimeout). In-flight tracking prevents disconnect mid-request.
npx binary resolution. Resolves npx packages to direct binary paths, eliminating the ~143 MB npm parent process per server. Persistent cache at ~/.pi/agent/mcp-npx-cache.json.
Lifecycle options. lazy (default), eager (connect at startup, no auto-reconnect), keep-alive (unchanged).
mcp({ connect: "server" }) mode. Explicitly trigger connection and metadata refresh.
Failure backoff. Servers that fail to connect are skipped for 60 seconds.
Prefix-match fallback. Tool calls with unrecognized names try to match a server prefix and lazy-connect the matching server.
Fuzzy tool name matching. Hyphens and underscores treated as equivalent — resolve_library_id finds resolve-library-id.
Better errors. When a server is identified but the tool isn't found, the error lists available tools for self-correction.
Impact
| Scenario | Before | After |
|---|---|---|
| 1 session, 2 of 6 servers used | ~1,200 MB | ~200 MB |
| 4 sessions, 2 servers each | ~4,900 MB | ~800 MB |
| npm parent processes | 1 per npx server | 0 |
See CHANGELOG.md for the full list.