[Yicheng-Mark]

Symptom

The Desktop Connectors per-server enable/disable switch does nothing: the state never changes, nothing persists to mcp_servers.<name>.enabled in config, and no error is ever shown. Meanwhile the CLI equivalent (hermes config set mcp_servers.<name>.enabled false) works against the same server, so the failure reads as "the toggle is dead". As reported in #135014 (first symptom).

Cause

The toggle went through the whole-map replace: setServerEnabled → writeEntry → persist() → saveMcpServers → PUT /api/mcp/servers with the entire mcp_servers map.

Two layers turn a backend rejection into silence:

1. The endpoint is the wrong shape for a toggle. replace_mcp_servers validates the whole map and 400s if any entry fails — one bad entry among eleven poisons every toggle, while the single-key CLI write path never touches that validation.
2. The failure is invisible. persist() awaited saveMcpServers but never read its { ok } outcome, then painted the React Query cache unconditionally. The optimistic paint flips the switch, the staleTime: 0 background refetch returns the untouched config and flips it back, and neither writeEntry's false return nor setServerEnabled notifies — the reporter sees "the state never changes (and does not persist to config)", with no error anywhere.

The per-server endpoint for exactly this toggle already exists (PUT /api/mcp/servers/{name}/enabled, exposed by setMcpServerEnabled), but only in ambient-scope form — the Capabilities surface needs the capabilityScoped(profile) routing every other desktop MCP write uses.

Fix

setServerEnabled stops round-tripping the whole map:

• New setMcpServerEnabledFor(name, enabled, profile?) in api/mcp.ts, scoped like saveMcpServers but hitting the per-server endpoint.
• setServerEnabled calls it, and:
  • on rejection (ok !== true or a thrown transport error) calls notifyError and returns — no cache paint, so the switch reflects reality;
  • after success, paints the config cache with withEnabled, mirrors into the draft, reloads, and probes when enabling — same post-write steps persist() performed, minus the map round-trip.
  • keeps writeEntry's guards: unknown server name and profilePending stay silent no-ops, and the profileEpoch check still fences cross-profile races.

The second symptom in the issue (clicking a row opening what looks like a different server's details) is deliberately not addressed here: it traces into how merged hosted+local cards pick their dialog and default way, which is a design question rather than a wiring bug, and belongs in its own discussion on the issue.

Regression tests (use-mcp-servers.test.tsx, new): the toggle PUTs the per-server endpoint and never the whole map; a rejected toggle notifies and leaves the cache unpainted; re-enabling drops the enabled key (the documented on-by-default) and probes; unknown server names short-circuit.

Fixes #135014 (toggle symptom)

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds a functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass — TS files; local verification was transpile-level (tsc, clean of syntax errors); suite runs in CI
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Windows 11

Documentation & Housekeeping

• I've updated relevant documentation — N/A, toggle persistence path only
• I've updated cli-config.yaml.example — N/A
• I've updated CONTRIBUTING.md or AGENTS.md — N/A
• I've considered cross-platform impact — request/reply path is platform-neutral
• I've updated tool descriptions/schemas — N/A

[Yicheng-Mark]

Superseded by #137863 — same fix and the same new test file, re-posted as a ready PR rather than a draft parked behind the open-PR limit.