[Ewertonslv]

Summary

Failures from the anyio/httpx transport family reached the caller as a bare
Tool invocation failed: — an empty detail — so a caller could not tell a closed
stream from a read timeout from a broken resource.

Closes #6592

Root cause

The outer except BaseException handler of invoke_tool builds the caller-facing
message as error_message = str(root_cause) (mcpgateway/services/tool_service.py:7029
on main@5273986b3). str() of that whole transport family is '':

anyio 4.14.0, httpx 0.28.1
ClosedResourceError   ''
EndOfStream           ''
BrokenResourceError   ''
httpx.ReadTimeout     ''

The exception class is logged server-side but never reaches the message. The empty
string flows to all three consumers of error_message: the error span, the
ToolResult handed to the post-invoke hooks, and the final ToolInvocationError.

A second, related gap: a read timeout raised inside the MCP SDK's TaskGroup arrives
wrapped in an ExceptionGroup, so it misses the per-transport
except (asyncio.TimeoutError, httpx.TimeoutException) lanes that emit the typed
Tool invocation timed out after {effective_timeout}s message, and lands in the same
outer handler — where it also stringifies to ''.

Fix

Two changes at the one construction site:

• str(root_cause) or "<module>.<qualname>" — byte-identical for every exception whose
  str() is already non-empty (e.g. httpx.HTTPStatusError for a 503), so nothing that
  works today changes.
• A group-wrapped timeout is restated in the vocabulary the per-transport lanes already
  emit, so callers key on one wording regardless of how the timeout arrived.

This is the shape proposed by the reporter in the issue.

Tests

Three tests added to tests/unit/mcpgateway/services/test_tool_service.py, mirroring the
existing test_invoke_tool_error_exception_group_unwrapping:

• test_invoke_tool_error_empty_str_falls_back_to_class_name — anyio.ClosedResourceError
  now names its class instead of yielding an empty detail.
• test_invoke_tool_group_wrapped_timeout_uses_timeout_vocabulary — an ExceptionGroup-wrapped
  httpx.ReadTimeout is reported as a timeout.
• test_invoke_tool_error_non_empty_str_is_unchanged — regression guard: an exception carrying
  a real message still produces exactly Tool invocation failed: upstream said no.

Before/after (stash only the fix, keep the tests):

without the fix:  F F .
    assert 'ClosedResourceError' in 'Tool invocation failed: '
    assert 'Tool invocation timed out after' in 'Tool invocation failed: '
with the fix:     . . .

The regression test passes in both states, as it should.

Verification

• pytest tests/unit/mcpgateway/services/test_tool_service.py — passes (exit 0)
• pytest tests/unit/mcpgateway/services/test_tool_service_coverage.py — passes (exit 0)
• ruff check — All checks passed
• ruff format --diff — 2 files already formatted

Note on overlap

PR #6181 also touches tool_service.py error handling, but at a different level — it
changes the envelope (returning CallToolResult(isError=True) instead of raising),
while this changes the content of error_message. A textual merge conflict in that
region is possible; happy to rebase on whichever lands first.

[Ewertonslv]

Friendly ping — could a maintainer approve the fork workflows on this PR? Only DCO has run so far, so the test suite hasn't executed yet. Happy to rebase or adjust anything. Thanks!