Persist last-used workspace selection across MCP session reconnects

[akornenvironmental]

The problem

The Render MCP server holds workspace selection in session-only memory. Every time the MCP client reconnects (a common occurrence in Claude Code: /login, /compact, fresh session resume, MCP server restart), the selection is lost and tools that need a workspace fail with no workspace set.

This is related to but distinct from #16 (which is about discoverability of the select_workspace tool). This issue is about persistence after the tool has been used: even when the LLM correctly calls select_workspace, the selection vanishes the next time the MCP client connects, and the LLM has to be prompted again to find and re-call select_workspace.

Reproduction (Claude Code)

1. Open a session in Claude Code, call any Render tool — it fails with no workspace set
2. Call list_workspaces, then select_workspace(ownerID=...) — works
3. /login, /compact, or quit and resume the session (claude --continue)
4. Call any Render tool — fails again with no workspace set

For users with one workspace, the auto-select-on-list behavior masks the issue partially, but tools that need a workspace before list_workspaces is called still fail. For users with multiple workspaces, every reconnect requires re-prompting.

Proposed fix

Persist the last-used workspace to a small per-user state file (e.g. ~/.render-mcp/state.json) keyed by the API key (or its hash). On connect, hydrate the in-memory workspace from that file. On select_workspace, write it back.

Alternative, lower-risk: persist on the server side keyed by API key. The MCP server already authenticates with the bearer token; treat the API key as the principal and store last-selected workspace per-key in Render's backend. This avoids client-side state files entirely and makes the selection follow the API key wherever it's used.

Why this matters

In long-running Claude Code sessions (which is the dominant use case), this causes constant friction. Users either:

• Get prompted to "pick a workspace" multiple times per session
• Have the LLM silently fail Render tools and burn turns figuring out why
• End up writing brittle SessionStart hooks to pre-call list_workspaces

Persisting the selection is the right fix and would close most of the workspace-related papercuts at once.

Environment

• Claude Code (Sonnet 4.6 / Opus 4.7)
• Render MCP server (hosted at mcp.render.com/mcp)
• HTTP transport with Bearer auth
• Single workspace (tea-d3tp...)

Happy to test a fix on the hosted endpoint if you want a guinea pig.