fix(mcp): editing any field resets an external MCP server's restart_policy and auto_start

[AlexLiu190625]

What happens

PUT /api/mcp/servers/{id} with a payload that touches one unrelated field silently
rewrites two other columns on the shared MCPServer row. On a row stored with
managed="external" and restart_policy="always":

BEFORE: restart_policy='always'  auto_start=True
PUT   : {"description": "a new description"}
AFTER : restart_policy='no'      auto_start=None

The request returns 200. Nothing in the response or the UI says the restart settings
changed.

Mechanism

The route does not write the payload's fields onto the row. It rebuilds the whole
configuration and writes every field of the rebuilt object back, so any column the
round-trip drops comes back as a config default.

1. _build_server_config (src/xagent/web/api/mcp.py) starts from
   server.to_config_dict() and merges the payload into it.
2. MCPServer.to_config_dict() (src/xagent/core/tools/core/mcp/model.py) emits
   restart_policy, auto_start and the docker fields inside a
   if self.managed == "internal": block. For an external row it emits none of them,
   so the stored values never reach the config object.
3. MCPServerConfig (src/xagent/core/tools/core/mcp/data_config.py) then supplies its
   own defaults for the absent keys: restart_policy: str = Field("no", ...) and
   auto_start left as None.
4. _update_server_from_config (src/xagent/web/api/mcp.py) writes back every name in
   _MCP_SERVER_CONFIGURABLE_FIELDS, and that tuple contains both restart_policy and
   auto_start. The loop is an unconditional setattr per field — there is no
   "only if the payload carried it" guard — so the defaults land on the row.

The same round-trip drops docker_url, volumes and bind_ports for an external row
by the same mechanism.

How to reproduce

1. Store an MCP server row with managed="external", restart_policy="always" and
   auto_start=True.
2. PUT /api/mcp/servers/{id} with {"description": "a new description"}.
3. Read the row back: restart_policy is 'no' and auto_start is NULL.

Expected

A payload that does not name restart_policy or auto_start must not change them. More
generally, the rebuild should not be able to write a column whose stored value it never
read — either the round-trip carries every stored column through, or the write-back is
restricted to the fields the payload actually carried.

Related

• PUT /api/mcp/servers/{id} rewrites managed to "external" on every update #1817 — the same round-trip hard-writes managed to "external". Same function, same
  round-trip, different columns; carrying the stored managed through does not by itself
  stop an external row's restart_policy from being reset, so the two need separate
  fixes.
• fix(connectors): lock the MCP server definition row on edit #2060 — changes which payloads run the rebuild at all: a request naming only
  is_active and/or user_env no longer rebuilds, so this loss no longer happens on
  that path. Every payload that writes the shared definition row still rebuilds, so this
  issue is untouched there. fix(connectors): lock the MCP server definition row on edit #2060 does not fix it and does not claim to; its description
  records it as a known limitation.