[kid7st]

test/cli-mcp.test.ts's "MCP login gives a browser ticket, saves a grant/binding and disconnects without deleting the grant" timed out (45 s) on a loaded machine. The cause is the test's cost, not FastAgent.

Root cause

The test spawned the CLI 7 times, besides its --check polls and the two ticket steps' own processes, and each spawn is a cold start of the whole CLI. Timed per spawn in a Linux container (Node 26.13.0, 4 CPUs):

	One spawn	The test
Alone	0.24 s (0.65 s for a mcp login that starts a ticket)	7.8 s
In the shard	0.27 s	8 s
In the shard with 8 CPU-bound processes	2.4 to 8 s	33 to 43 s

The rest is the refused disconnect, which waits out BUSY_TIMEOUT_MS (5 s) for the storage the test holds, by design.

The fix

A process of its own only where the test needs one, as AGENTS.md asks of CLI tests:

• the two mcp logins that start a ticket, whose step command is process.argv[1], which in process is not the CLI;
• the disconnect refused while the test holds the storage, since a POSIX lock does not exclude its own process.

mcp list, the mcp login refused for having no --scopes, every --check, and the last disconnect now run in process. 3 spawns instead of 7 plus the polls.

Verification

Under the same load (the shard with 8 CPU-bound processes), the file took 24 s and 19 s, against 43 s and 33 s before. Alone it takes 8.7 to 10.5 s, a little more than before, since the worker now loads the CLI once. The whole suite passed on macOS and in the Linux container.