Remote/HTTP MCP servers (e.g. atlassian) are stranded 'failed' after every /clear or session relaunch

Summary

Remote (HTTP) MCP server connections are reliably stranded in a failed state whenever the CLI performs a foreground-session handoff — which happens on every /clear, and apparently on session resume/relaunch too. The MCP connection graph is torn down and rebuilt during the handoff, and slower remote/OAuth-backed servers lose the reconnect race and never recover automatically.

Environment

• CLI version: 1.0.83 (macOS, confirmed already latest via in-app update check)
• MCP server affected: atlassian (https://mcp.atlassian.com/v2/mcp, HTTP transport with Authorization header)
• Other MCP servers configured (github-mcp-server, codegraph, local-rag — local/stdio) are not visibly affected, likely because their reconnect is fast enough to win the race.

Steps to reproduce

1. Start a Copilot CLI session with the atlassian MCP server configured (~/.copilot/mcp-config.json, HTTP transport).
2. Confirm it connects fine (/mcp show atlassian → connected, tools listed).
3. Run /clear (or relaunch/resume a session).
4. Run /mcp show atlassian again.

Expected

atlassian reconnects cleanly like the other MCP servers.

Actual

atlassian is left in:

{
  "name": "atlassian",
  "status": "failed",
  "error": "MCP server \"atlassian\" connection was cancelled"
}

It does not self-heal — it requires a manual reconnect action or a full CLI restart, and even a full restart reproduces the same failure deterministically (confirmed across two separate relaunches, ~20 minutes apart).

Root cause (from ~/.copilot/logs/process-*.log)

Every /clear/relaunch briefly registers a throwaway foreground session, then immediately unregisters it in favor of the real one, within milliseconds:

17:27:22.425Z [INFO] Registering foreground session: 6266dbfb-39da-4797-9265-c37d9655a3fd
17:27:28.717Z [INFO] Unregistering foreground session: 6266dbfb-39da-4797-9265-c37d9655a3fd
17:27:28.720Z [INFO] Registering foreground session: d89d1d9a-99c4-4247-86fe-a75b4da5daad
17:27:28.783Z [INFO] Closing session 6266dbfb-39da-4797-9265-c37d9655a3fd

This handoff triggers a full MCP graph reload while a reload lock is still held from the previous teardown:

17:05:05.167Z [WARNING] [rust:copilot_runtime::session::mcp::session_host] MCP reload lock still held at disposal; draining the graph without it

Every MCP server — including atlassian — gets re-initialized and then almost immediately cancelled mid-handshake:

17:27:26.104Z [INFO] [rust:rmcp::service] Service initialized as client {... "name": "atlassian-mcp-server" ...}
17:27:28.809Z [INFO] [rust:rmcp::service] task cancelled
17:27:28.809Z [INFO] [rust:rmcp::service] serve finished {"quit_reason":"Cancelled"}

Local/stdio servers appear to reconnect fast enough afterward to recover; atlassian's remote HTTP+OAuth handshake is slower and consistently loses the race, leaving it permanently failed with no automatic retry.

Confirmed not a network/auth issue independently:

• DNS resolves fine for mcp.atlassian.com.
• curl https://mcp.atlassian.com/v2/mcp returns 401 (server reachable, no valid header supplied in the manual test — expected).
• ATLASSIAN_MCP_AUTH_HEADER env var is set.

Impact

Every /clear (a very common action) breaks the Atlassian MCP integration and requires a manual reconnect or CLI restart to restore it — significant daily friction for anyone using an HTTP-based MCP server.

Suggested fix

• Don't tear down/reinitialize already-connected MCP servers on a foreground-session handoff that isn't actually changing MCP config.
• If a reload is unavoidable, retry a server that was cancelled mid-handshake instead of leaving it permanently in failed state.
• Respect the "MCP reload lock still held at disposal" case by waiting for the lock instead of draining the graph without it.

Reproduziere den Fehler mit einem in ~/.copilot/mcp-config.json konfigurierten HTTP server und untersuche anschließend die Logs zur Übergabe der Vordergrundsitzung und von copilot_runtime::session::mcp::session_host rund um die MCP reload lock. Vergleiche das Verhalten von /clear und dem erneuten Start mit der dokumentierten Abbruchsequenz. Erledigt ist die Aufgabe, wenn der Remote-Server nach einer der beiden Übergaben verbunden bleibt oder automatisch wiederhergestellt wird, ohne dass eine manuelle erneute Verbindung erforderlich ist.