Files
turnstone/docs/diagrams/20-mcp-architecture.puml
T
Patrick Buckley eb2a119da9 refactor(mcp): remove periodic refresh, add manual refresh/reconnect controls
Deletes the _periodic_refresh task and its supporting state
(_refresh_task, _refresh_failures, _refresh_backoff_until,
_REFRESH_BACKOFF_BASE/MAX, _DEFAULT_REFRESH_INTERVAL, refresh_interval
kwarg) from MCPClientManager. Push notifications and operator-driven
manual refresh now cover all catalog-update needs; the long-running
4-hour timer was dead complexity that obscured the per-user pool
work to come.

Catalog freshness on auto-reconnect is preserved by scheduling an
unblocking _refresh_server task on the mcp-loop after _connect_one
succeeds; the calling thread returns immediately so half-open
recovery latency does not double. Adds MCPClientManager.reconnect_sync
(clears the circuit, closes any existing session, calls _connect_one,
clears stale catalog on failure).

Wires a new pair of operator endpoints —
POST /v1/api/admin/mcp-servers/{name}/refresh and
/v1/api/admin/mcp-servers/{name}/reconnect — that fan out to all
nodes through the existing _internal route family, with per-row
"Refresh" and "Reconnect" buttons in the MCP Servers admin tab.
The new node-internal paths /api/_internal/mcp-{refresh,reconnect}/
are gated to the approve scope to prevent direct unprivileged
reconnects bypassing the console's admin.mcp gate. Internal
endpoints return generic error messages and a filtered status
payload (no command/url) to keep transport details admin-gated.

Drops the [mcp] refresh_interval setting, the
--mcp-refresh-interval CLI flag, and the matching config-mapping
entry; updates docs/architecture.md, docs/tools.md,
docs/settings.md, and the three PlantUML diagrams that referenced
the periodic loop.

Tradeoffs (intentional):
- Idle nodes will not auto-rejoin a recovered MCP server until
  traffic arrives or an operator clicks Reconnect. The previous
  background reconnection loop is gone by design — push
  notifications + operator controls replace it.
- Console fan-out blocks on the slowest node (existing pattern);
  not changed here.

This is Phase 1 of the OAuth-MCP series — feature subtraction
ahead of per-user state.
2026-05-04 22:00:23 -07:00

235 lines
6.9 KiB
Plaintext

@startuml
!theme plain
title Turnstone — MCP Architecture (Resources, Prompts, Tools)
skinparam participant {
BackgroundColor<<mcp>> #E1BEE7
BackgroundColor<<session>> #C8E6C9
BackgroundColor<<storage>> #B3E5FC
BackgroundColor<<server>> #FFE0B2
BackgroundColor<<ui>> #E8EAF6
BackgroundColor<<registry>> #F8BBD0
}
participant "MCP Server\n(external)" as MCPSrv <<mcp>>
participant "MCPClientManager\n(mcp_client.py)" as MCPMgr <<mcp>>
participant "ChatSession\n(session.py)" as Session <<session>>
participant "StorageBackend\n(governance)" as Storage <<storage>>
participant "Server / Console\n(health + UI)" as UI <<server>>
participant "Console Admin UI\n(admin panel)" as Admin <<ui>>
participant "Database\n(mcp_servers table)" as DB <<storage>>
participant "MCPRegistryClient\n(mcp_registry.py)" as RegClient <<mcp>>
participant "MCP Registry\n(registry.modelcontextprotocol.io)" as Registry <<registry>>
== Admin-Driven Configuration ==
Admin -> DB : CRUD MCP server definitions\n(POST/PUT/DELETE /v1/api/admin/mcp-servers)
Admin -> UI : POST /v1/api/admin/mcp-servers/reload
UI -> MCPMgr : POST /_internal/mcp-reload\n(forwarded to each node)
MCPMgr -> MCPMgr : reconcile_sync()
note right
Diffs running servers against DB:
- New entries → connect
- Removed entries → disconnect
- Changed entries → reconnect
end note
== Registry Discovery & Install ==
Admin -> UI : GET /v1/api/admin/mcp-registry/search?search=...
UI -> RegClient : search(q, limit, cursor)
RegClient -> Registry : GET /v0.1/servers?search=...&latest=true
Registry --> RegClient : Server entries\n(remotes, packages, meta)
RegClient --> UI : RegistrySearchResult\n(annotated with installed status)
UI --> Admin : Search results\n(Install / Installed badges)
Admin -> UI : POST /v1/api/admin/mcp-registry/install
UI -> DB : create_mcp_server()\n(registry_name, version, meta)
UI -> MCPMgr : POST /_internal/mcp-reload\n(fan-out to nodes)
MCPMgr -> MCPMgr : reconcile_sync()
MCPMgr -> MCPSrv : connect to new server
note over RegClient, Registry
MCPRegistryClient is an async httpx client
targeting registry.modelcontextprotocol.io/v0.1.
resolve_install_config() translates registry
remotes/packages into mcp_servers rows.
end note
== Startup: Connection & Discovery ==
MCPMgr -> DB : load_mcp_config(storage=)\n(merge config file + DB)
MCPMgr -> MCPSrv : initialize (stdio or HTTP)
MCPSrv --> MCPMgr : capabilities\n(tools, resources, prompts)
MCPMgr -> MCPSrv : tools/list
MCPSrv --> MCPMgr : Tool[]
opt resources capability
MCPMgr -> MCPSrv : resources/list
MCPSrv --> MCPMgr : Resource[]
MCPMgr -> MCPSrv : resources/templates/list
MCPSrv --> MCPMgr : ResourceTemplate[]
end
opt prompts capability
MCPMgr -> MCPSrv : prompts/list
MCPSrv --> MCPMgr : Prompt[]
end
note over MCPMgr
Per-server storage:
_per_server_tools, _per_server_resources, _per_server_prompts
Copy-on-write rebuild into _tools, _resources, _prompts
Prefix: mcp__{server}__{name}
end note
MCPMgr -> Session : notify tool listeners
MCPMgr -> Session : notify resource listeners
== Governance Sync (on connect & refresh) ==
MCPMgr -> Storage : sync_prompts_to_storage()
note right
For each MCP prompt:
- Manual template exists? → skip
- MCP template exists? → update
(reset is_default=False)
- New? → create (origin="mcp",
readonly=True, is_default=False)
Removed prompts → delete
Protected by _sync_lock
end note
== set_storage() from entry point ==
UI -> MCPMgr : set_storage(backend)
note right
If servers already connected,
triggers immediate sync
end note
== Runtime: Tool Execution ==
Session -> Session : _prepare_mcp_tool(func_name, args)
note right
approval_label = func_name
(e.g. mcp__github__search)
needs_approval = True
end note
Session -> MCPMgr : call_tool_sync(name, args)
MCPMgr -> MCPSrv : tools/call
MCPSrv --> MCPMgr : ToolResult
MCPMgr --> Session : output (text)
== Runtime: Resource Read ==
Session -> Session : _prepare_read_resource(uri)
note right
approval_label = mcp_resource__{normalized_uri}
URI normalized (.. resolved)
needs_approval = True
end note
Session -> MCPMgr : read_resource_sync(uri)
MCPMgr -> MCPSrv : resources/read
MCPSrv --> MCPMgr : ReadResourceResult
MCPMgr --> Session : content (text/blob)
== Runtime: Prompt Invocation ==
Session -> Session : _prepare_use_prompt(name, arguments)
note right
approval_label = mcp__srv__prompt
Validated via is_mcp_prompt()
needs_approval = True
end note
Session -> MCPMgr : get_prompt_sync(name, args)
MCPMgr -> MCPSrv : prompts/get
MCPSrv --> MCPMgr : GetPromptResult
MCPMgr --> Session : messages [{role, content}]
== Resilience: Circuit Breaker & Stream Safety ==
note over MCPMgr
**Per-server circuit breaker**
CLOSED --(3 failures)--> OPEN
OPEN --(cooldown expires)--> half-open probe
Probe success --> CLOSED (trip_count decays by 1)
Probe failure --> OPEN (cooldown doubles, max 5 min)
McpError (protocol) does NOT trip breaker.
BrokenPipeError / EOFError evicts dead session.
All sync methods cancel orphaned futures on timeout.
Transport streams pre-closed before stack teardown
to avoid anyio cancel-scope CPU busy-loop (SDK #2147).
end note
Session -> MCPMgr : call_tool_sync()
MCPMgr -> MCPMgr : _cb_gate(server)\n[reject if circuit open]
MCPMgr -> MCPMgr : _cb_auto_reconnect()\n[if session gone + cooldown expired]
MCPMgr -> MCPSrv : tools/call
MCPSrv --> MCPMgr : result or error
MCPMgr -> MCPMgr : _cb_record_success()\nor _cb_record_failure()
== Three-Tier Refresh ==
group Push Notifications (debounced 5s per server)
MCPSrv -> MCPMgr : ToolListChangedNotification
MCPMgr -> MCPMgr : debounce check\n(skip if < 5s since last)
MCPMgr -> MCPMgr : _refresh_server_tools()
MCPSrv -> MCPMgr : ResourceListChangedNotification
MCPMgr -> MCPMgr : _refresh_server_resources()
MCPSrv -> MCPMgr : PromptListChangedNotification
MCPMgr -> MCPMgr : _refresh_server_prompts()
MCPMgr -> Storage : sync_prompts_to_storage()
end
group Manual Refresh
Session -> MCPMgr : refresh_sync()
note right
/mcp refresh [server] —
re-fetches catalog and
attempts reconnect for
disconnected servers.
end note
end
group Manual Reconnect
Session -> MCPMgr : reconnect_sync(name)
note right
Operator-driven via the
console admin panel —
tears down session, clears
circuit breaker, runs a
fresh handshake.
end note
end
== Policy Evaluation ==
note over Session
Tool policies use fnmatch on approval_label:
- mcp__github__* → allow (all GitHub tools/prompts)
- mcp_resource__file:///docs/* → allow
- mcp_resource__* → deny (block all resource reads)
- mcp__untrusted__* → ask
end note
== UI Visibility ==
UI -> MCPMgr : server_count, get_resources(), get_prompts()
note over UI
/health → mcp.servers, mcp.resources, mcp.prompts
Server UI: magenta status badge
Console: cluster status bar + node detail
System message: <mcp-resources> + <mcp-prompts> catalogs
end note
@enduml