Repository navigation
[BUG] A headless host has no supported way to add a monitored server: no CLI verb, and the web surface excludes the write tools #2256
Description
Activity
- added a commit that references this issue
on Aug 14, 2026 Fixed on
devin #2259.--add-server(also--add-servers) reads a JSON array from stdin — the same shapeadd_serverstakes — and delegates to that exact code path, so validation, dedupe, the in-process probe, encryption and theserver_idcomputation are shared rather than reimplemented. That last part is why a hand-written INSERT was never the workaround: the identity is a hash of the storage name, and getting it wrong splits history (#2158).Stdin rather than argv so a password never reaches the process list or shell history; empty stdin prints the JSON shape and two copy-paste pipelines to STDOUT rather than hanging. Exit 0 requires that something landed and nothing failed, so it works as a deployment gate — pure duplicates exit 0 since re-running a file is idempotent, and a batch where nothing landed exits 1. Windows is needed only for a managed (DPAPI) store credential, checked in that branch, so a Linux host with bring-your-own Postgres can use it.
Review caught three things worth naming: my block stole
BackfillRollupsAsync's[SupportedOSPlatform]attribute AND then its doc comment (two variants of the same insertion-boundary trap), and a real store failure was being reported as a parse failure becauseAddServersAsync's catch-all returns plain text rather than JSON. All fixed.@000al000 — this is what you needed for the Windows Server 2012 host. Ships with the next release; until then the Viewer's Add Server dialog from another machine or the MCP tool remain the paths.
- added a commit that references this issue
on Aug 14, 2026 - added a commit that references this issue
on Aug 19, 2026
Field report #2252. Operator runs the Darling service on Windows Server 2012, which cannot run the Viewer. They have the web viewer and a Viewer on a remote PC. There is no supported way for them to add a second server.
The three doors are all shut
darling.jsonconfig_monitored_serversonly when that table is empty; after the first start the store is authoritative and file edits are a silent no-op (#2254)add_servers/remove_serverare inDarlingWebEndpoints.ExcludedToolNames, deliberately kept off/api/read/*because they write the registry--backfill-rollups --collapse-legacy-slices --config --configure-firewall --configure-network --disable-mcp --disable-web --dry-run --enable-mcp --enable-web --encrypt-password --export-viewer-config --help --no-vacuum-full --print-viewer-connection --recompress-plan-dim --test-connection --vacuum-full --validate-config --versionWhat remains is a GUI on another machine (which then hits the DPAPI credential problem in #2255) or standing up an MCP client. Neither is a reasonable answer for a headless Windows host, and the operator reached the end of the documented surface before asking.
Why "just INSERT the row" is not the workaround
server_idisServerIdHelper.GetDeterministicHashCode(StorageName)where the storage name is built from host + database + read-only intent. A hand-written INSERT with any other value leaves the collectors writing history under the identity they compute while the registry points elsewhere — the split-history class in #2158. So the identity computation is exactly the part an operator cannot safely do by hand, and exactly whatadd_serversalready does correctly.Fix shape
An
--add-serverCLI verb that delegates to the same code path as theadd_serversMCP tool — identity computation, per-entry validation, and the in-process connection probe included, so it cannot diverge from the tool. It should also bumpconfig_service.config_version, since the service hot-swaps on that beacon each sweep and therefore needs no restart.Worth deciding alongside:
--remove-server/--list-serversfor symmetry. A verb that can only add leaves the same operator stuck on a rename or a decommission.--test-connectionshould gain the store-vs-file comparison from [BUG] Servers added to darling.json after the first seed are silently ignored, and --test-connection still validates them #2254 while it is being touched — the two issues are the same operator journey.Credential handling is the one wrinkle:
--encrypt-passwordalready exists, so a verb taking an already-encrypted blob (or integrated auth, which stores nothing) needs no new secret handling.