Archive gating does not pair with workspace MCP server lifecycle (stdio servers survive archive; prompt discovery can re-wake Coder)

[ThomasK33]

Context

PR #3940's model-driven task_workspace_lifecycle archive path (refuseLiveUserActivity) pairs archive admission with streams, sends, queued turns, background bash, terminals, desktop sessions, workflows, executeBash, attachment staging, and file-completion refreshes. MCP server lifecycle is not paired.

Problem (Codex review round 19, P2)

Condition: the owned peer has a cached workspace stdio MCP server, or MCP prompt discovery (workspace.mcp.prompts.list) races a Snapshot/Coder-stop archive.

• The archive gate neither observes MCP server startup/leases nor stops cached servers; only workspace removal calls stopServers.
• A model-driven archive can therefore delete a managed checkout beneath a running stdio MCP process (losing post-snapshot writes), or a prompt-discovery request entering mid-archive can reconnect and re-wake a stopped Coder runtime.

Suggested direction

• Pair MCP server startup/prompt discovery with archive admission using the established synchronous-entry pattern (archivingWorkspaces guard + preflight admission counter checked by the refuseLiveUserActivity gate).
• Stop cached workspace MCP servers during archive before runtime cleanup (mirroring what removal does via stopServers), or refuse archive while servers hold leases.

References

• Archive gate: WorkspaceService.archiveUnlocked (refuseLiveUserActivity), pairing pattern in acquirePreflightAdmission
• MCP lifecycle: workspace MCP server cache / stopServers (removal path)
• Review thread: PR 🤖 feat: restore task_workspace_lifecycle with archive/unarchive for peer workspaces #3940, round 19 (src/node/services/taskService.ts)

---

Generated with mux • Model: anthropic:claude-fable-5 • Thinking: xhigh

[linear-code]

MUX-118