MCP server down at session start is unrecoverable for the session — even after account-level reconnect succeeds; no mid-session attach/reload

Summary

An MCP server that is unreachable at session start is permanently missing for the whole session — even after the underlying problem is fixed mid-session and claude mcp list confirms the server is ✔ Connected at the account/config level. There is no way to attach or reload it without abandoning the session.

Prior requests for this were closed as not planned and locked (#40059, #57207), so filing fresh with a concrete real-world failure.

What happened (real scenario, 2026-07-24)

1. The claude.ai Linear connector broke (OAuth needed re-auth). A Claude Code session (desktop app, v2.1.206) was started while it was broken → the session's tool roster has no mcp__claude_ai_Linear__* tools.
2. Mid-session, the user re-authenticated Linear on claude.ai. claude mcp list (run from inside the session via Bash) now shows claude.ai Linear: https://mcp.linear.app/mcp - ✔ Connected.
3. The running session still has no Linear tools: they're absent from the deferred-tool roster, ToolSearch finds nothing, and nothing re-evaluates the connector set after startup.
4. The user's reasonable ask — "reload the MCP server" — has no answer. The /mcp terminal dialog isn't available in the desktop app session, and even where it is, it can only reconnect servers that were attached at startup; it cannot attach one that failed/was added after startup.

The session had to work around it by spawning a nested headless claude -p run (which connects MCP fresh and did have the Linear tools) just to file one Linear issue. That works but is absurd as a UX answer, burns tokens, and loses all session context.

Why this is worse than it used to be

With deferred tools + ToolSearch, tool schemas are already loaded lazily — the architecture no longer assumes a fixed tool roster frozen at turn 0. The one thing still frozen at session start is the MCP server connection set. That makes any transient connector outage at boot (OAuth expiry, network blip, server deploy) a session-lifetime defect, invisible until the user needs the tool an hour into a long-running session.

Ask (any of these would resolve it)

• claude mcp reconnect <name> / a tool-callable equivalent that (re)attaches a configured or account-level connector into the running session and injects its tools into the deferred-tool roster; or
• periodic / on-demand re-evaluation of the claude.ai connector set for live sessions (e.g. on ToolSearch miss for an mcp__* pattern); or
• at minimum: retry failed-at-startup MCP servers on a backoff instead of marking them dead for the session.

Environment

• Claude Code v2.1.206, macOS (darwin 25.5.0), desktop app session
• Affected server: claude.ai Linear connector (https://mcp.linear.app/mcp), but the failure mode is generic to any MCP server down at session start

References

• #40059 — Allow reloading MCP servers without restarting the session (closed: not planned)
• #57207 — Add claude mcp reconnect <name> (closed: not planned)