ACP: session/new blocks for 192s on a single unresponsive MCP server (no bounded MCP startup budget)

Describe the bug

In ACP mode (--acp), the session/new response is blocked until every
configured MCP server has finished connecting. A single unresponsive HTTP MCP
server therefore delays session creation by 192 seconds — three connection
attempts at roughly 59 s each — instead of failing fast and starting the session
without that server.

There is no bounded budget on this. An ACP client cannot distinguish "still
starting" from "wedged", so it either waits minutes or times out and kills a
process that was going to succeed.

Affected version

GitHub Copilot CLI 1.0.82-1

Windows 10/11 x64, WinGet install.

Steps to reproduce

Fully deterministic. Reproduces every time.

1. Create an MCP endpoint that accepts the connection and never replies —
   this stands in for any MCP server that is wedged or slow to start:

// stall-server.js
const http = require('http')
http.createServer((req) => {
  console.log(`received ${req.method} ${req.url} -- deliberately not responding`)
}).listen(39217, '127.0.0.1', () => console.log('listening on 127.0.0.1:39217'))

node stall-server.js

2. Point an MCP config at it:

{
  "mcpServers": {
    "stalled": { "type": "http", "url": "http://127.0.0.1:39217/mcp", "tools": ["*"] }
  }
}

3. Start Copilot in ACP mode with that config:

copilot --no-auto-update --log-dir ./logs --acp --additional-mcp-config @mcp.json

4. Over stdio, send initialize, then session/new, and time the response:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":1,"clientCapabilities":{"fs":{"readTextFile":false,"writeTextFile":false}}}}
{"jsonrpc":"2.0","id":2,"method":"session/new","params":{"cwd":"<abs path>","mcpServers":[]}}

Actual behavior

initialize returns in ~1 s. session/new takes 192 s:

RUN stall2: OK  initialize 1078ms  session/new 192355ms  sid=3a0719b1-...

The log shows the retry ladder — three attempts, each burning a ~59 s
connection timeout, each followed by a server/discover timeout and a legacy
initialize retry:

23:10:14.946  Workspace initialized: 3a0719b1-... (checkpoints: 0)
23:10:15.224  mcp discover_and_start_root: full replace_host path
23:10:15.752  mcp graph load: reload_with_config {"force":false}
23:10:18.747  mcp discover_and_start_root: incremental reconcile path
23:10:19.874  [WARNING] server/discover timed out; retrying with legacy initialize
23:11:18.876  Transient MCP connection failure {"server_name":"stalled",...}
23:11:18.876  Retrying connection to HTTP server stalled (attempt 2/3) after 500ms
23:11:20.392  [WARNING] server/discover timed out; retrying with legacy initialize
23:12:19.386  Transient MCP connection failure {"server_name":"stalled",...}
23:12:19.386  Retrying connection to HTTP server stalled (attempt 3/3) after 1000ms
23:12:21.395  [WARNING] server/discover timed out; retrying with legacy initialize

Two details that make this hard to diagnose from the client side:

• The wait is silent. Between connected to 127.0.0.1:39217 and the
  Transient MCP connection failure 59 s later, nothing is logged at all. The
  log simply appears to stop, which reads like a freeze rather than a retry.
• session/new has already succeeded internally. Workspace initialized
  is logged 3 minutes before the client gets its reply, so the session exists
  the whole time the client is blocked waiting for it.

Expected behavior

One of:

1. A bounded overall budget for MCP startup during session/new, after which
   the session is returned and unreachable servers are reported as degraded; or
2. MCP connection established asynchronously, so session/new returns promptly
   and servers become available as they connect.

Either way, an unreachable MCP server should degrade tool availability, not
delay session creation by minutes.

Baseline for comparison

Same client, same workspace, same machine:

MCP configuration	session/new
none	5.6 s
2 reachable public HTTP servers	~5 s (10/10 runs)
1 unresponsive HTTP server	192.4 s

Impact

Any ACP client with a startup timeout under ~3 minutes will kill a Copilot
process that was going to succeed. Because the connection wait is silent, the
resulting logs make it look like Copilot froze during MCP graph initialization,
which sends people looking in the wrong place. It also means one flaky or
slow-starting MCP server degrades session startup for the whole session, rather
than just its own tools.

Related observation (may be a separate issue)

Separately, and not reproduced deterministically, I have twice seen ACP
startup stop at mcp graph load: reload_with_config with no subsequent
connection attempt logged at all, on a workspace using locally-hosted MCP
servers. Unlike the case above, no connecting to line ever follows. Retrying
the same request minutes later succeeded, so it appears to be a race rather than
a configuration problem. I am mentioning it only as possible context — I cannot
currently reproduce it on demand and it may well be unrelated to the timeout
issue reported here.

Additional context

Possibly related, in the same startup-reconciliation area:

• #4636 — MCP servers from --additional-mcp-config removed during startup reconciliation
• #4392 — Post-authentication MCP client rebuild leaves orphaned stdio MCP server processes

Full logs for the reproduction available on request.

Start by running the provided stall-server.js reproduction with copilot --acp and timing initialize versus session/new; use the MCP timeout and retry log sequence as the baseline. Trace the ACP session/new entry point and MCP startup/reconciliation path, then verify that an unresponsive server no longer blocks session creation beyond a bounded budget and that the degraded server is reported.