What does this PR do?

Narrow fix for #14113 where stdio MCP servers silently fail to register in the launchd-managed gateway with a bare CancelledError, while the same config works via hermes mcp test, hermes chat -q, and the gateway's HTTP transport.

What's happening

_run_on_mcp_loop() (tools/mcp_tool.py:1544) polls is_interrupted() in a 0.1s loop and calls future.cancel() on the submitted coroutine if the calling thread's tid is in _interrupted_threads. That cancellation channel exists so a new user message aborts a long-running tool call mid-flight — exactly the right semantics for _make_tool_handler.

The same plumbing was also being used for registration / discovery / probe — call paths that run during gateway startup and /reload-mcp, scheduled onto the default executor's worker pool. run_agent.py's own clear_interrupt() (line 3619–3623) already warns that stale per-tid flags can survive a turn boundary and fire on unrelated work scheduled onto the recycled worker:

"guarantees no stale interrupt can survive a turn boundary and fire on a subsequent, unrelated tool call that happens to get scheduled onto the same recycled worker tid."

When a prior agent turn's cleanup was missed (crash, abandoned task, whatever), the executor worker's tid stays in _interrupted_threads. The next thing to run on that worker — discover_mcp_tools() at gateway boot or /reload-mcp — gets its stdio handshake cancelled before it finishes. asyncio.gather(..., return_exceptions=True) collects the bare CancelledError and the warning in the report is emitted.

The fix

One kwarg on _run_on_mcp_loop:

def _run_on_mcp_loop(coro, timeout: float = 30, *, respect_interrupt: bool = True):
    ...
    if respect_interrupt and is_interrupted():
        future.cancel()
        raise InterruptedError("User sent a new message")

• Default True — tool-call callers (_make_tool_handler._call_once and its five variants) keep the existing interrupt behavior; user interrupts still abort in-flight tool calls.
• Explicit False from register_mcp_servers() and probe_mcp_server_tools() — registration / probe cannot be cancelled by a stale flag on a recycled worker.

Also: _format_connect_error() now recognises a naked asyncio.CancelledError (no __cause__, no __context__) and emits a short actionable message instead of the bare class name. A chained CancelledError with a real cause still falls through to the existing flattening path so the underlying reason is preserved.

Scope limits

• Does not touch gateway/run.py, run_agent.py, or model_tools.py. The fix lives where the bug lives.
• Does not change tool-call interrupt semantics.
• Does not empirically confirm the stale-flag hypothesis against a live launchd environment — I don't have one. The fix is defensible on its own terms (registration shouldn't be cancellable by a stale user-interrupt flag), and removes the most plausible trigger for the reported symptom. If the reporter sees residual cancellations after this lands, the improved _format_connect_error message will point them to the next datapoint (parent-task cancel vs. some other source).

Related Issue

Fixes #14113

Type of Change

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

Changes Made

• tools/mcp_tool.py:
  • _run_on_mcp_loop: new respect_interrupt kwarg (keyword-only, default True)
  • register_mcp_servers and probe_mcp_server_tools: pass respect_interrupt=False
  • _format_connect_error: recognize naked CancelledError and emit an actionable hint
• tests/tools/test_mcp_tool.py:
  • TestRunOnMCPLoopInterrupts.test_respect_interrupt_false_ignores_stale_flag — simulates stale flag, asserts registration proceeds
  • TestRunOnMCPLoopInterrupts.test_respect_interrupt_default_still_cancels — regression guard, tool-call path still raises InterruptedError
  • TestFormatConnectError.test_naked_cancelled_error_has_actionable_message — new CancelledError branch
  • TestFormatConnectError.test_cancelled_error_with_cause_falls_through — chained cause still rendered
  • TestFormatConnectError.test_missing_node_executable_unchanged — existing FileNotFoundError path unaffected

How to Test

source .venv/bin/activate
scripts/run_tests.sh tests/tools/test_mcp_tool.py -q

Expected: 171 passed.

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 — only tests/tools/test_mcp_tool.py was run (171 passed); the full suite was not run
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS (Darwin 25.4.0). No live launchd repro — I don't have the reporter's environment.

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — added explanatory docstrings on the new respect_interrupt kwarg
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) — pure Python, no platform-specific code
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

▶ running pytest with 4 workers, hermetic env, in /…/hermes-agent
  (TZ=UTC LANG=C.UTF-8 PYTHONHASHSEED=0; all credential env vars unset)
........................................................................ [ 42%]
........................................................................ [ 84%]
...........................                                              [100%]
171 passed, 5 warnings in 8.01s