arcade-mcp/libs/tests/core/test_errors.py
Francisco Or Something d9812621de
feat: add NetworkTransportError for no-response HTTP failures (#823)
## Summary

- Adds `NetworkTransportError` — a new sibling to `UpstreamError` under
`ToolExecutionError` — for failures where no complete HTTP response was
received from the upstream service (timeouts, connection errors, pool
exhaustion, DNS failures, decoding issues, redirect exhaustion)
- Routes client-construction bugs (`InvalidURL`, `UnsupportedProtocol`,
`MissingSchema`, `SSLError`, `InvalidHeader`, etc.) to existing
`FatalToolError` instead of `UpstreamError`
- Adds 3 new `ErrorKind` values: `NETWORK_TRANSPORT_RUNTIME_TIMEOUT`,
`_UNREACHABLE`, `_UNMAPPED` — operationally distinct telemetry slices
matching the UpstreamError pattern
- `UpstreamError` is unchanged and reserved for real HTTP responses with
status codes

Addresses Eric's feedback on #820: the `include_status_code=False`
post-init null-out workaround is replaced by a clean class hierarchy
where `NetworkTransportError.status_code` is natively `None`.

### Changes

| File | What |
|---|---|
| `arcade-core/errors.py` | 3 new `ErrorKind` values,
`NetworkTransportError` class, `is_network_transport_error` helper |
| `arcade-tdk/providers/http/error_adapter.py` | Full rewrite of httpx +
requests exception routing with 3-way split |
| `arcade-tdk/providers/graphql/error_adapter.py` |
`TransportConnectionFailed`/`TransportProtocolError` →
`NetworkTransportError` |
| `arcade-tdk/errors.py`, `arcade-mcp-server/exceptions.py` | Re-exports
|
| `pyproject.toml` × 3 | Version bumps: core 4.7.0, tdk 3.7.0,
mcp-server 1.20.0 |
| Tests × 3 | 33 new tests, 3 updated (2659 passed, 0 failures) |

### Exception routing table

| Exception | Target | Kind | can_retry |
|---|---|---|---|
| `httpx.HTTPStatusError`, `requests.HTTPError` (with response) |
`UpstreamError` | status-derived | status-derived |
| `httpx.TimeoutException`, `requests.Timeout` | `NetworkTransportError`
| `TIMEOUT` |  |
| `httpx.TransportError`, `requests.ConnectionError` |
`NetworkTransportError` | `UNREACHABLE` |  |
| `httpx.DecodingError`, `TooManyRedirects`, fallback |
`NetworkTransportError` | `UNMAPPED` | varies |
| `httpx.InvalidURL`/`UnsupportedProtocol`/`LocalProtocolError`,
`requests.MissingSchema`/`SSLError`/etc. | `FatalToolError` |
`TOOL_RUNTIME_FATAL` |  |

### Engine companion PR

ArcadeAI/monorepo — `feat/network-transport-error-kinds` adds the 3
`ErrorKind` constants to Go schemas + OpenAPI docs. No engine logic
changes needed (ErrorKind is a string alias, retry uses `can_retry` flag
only, telemetry auto-slices).

## Test plan

- [x] 2659 existing tests pass (0 failures)
- [x] 33 new routing + class tests added
- [x] mypy clean on arcade-core, arcade-tdk
- [ ] Verify engine telemetry dashboard auto-surfaces new
`NETWORK_TRANSPORT_*` kinds after deploy

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **Medium Risk**
> Changes the error taxonomy and classification helpers used for
retries/telemetry, so misclassification could affect operational
behavior, but the change is additive and covered by new tests.
> 
> **Overview**
> Adds a new error category for outbound request failures that never
yield a complete upstream response: `NetworkTransportError` (sibling to
`UpstreamError`) plus
`ErrorKind.NETWORK_TRANSPORT_RUNTIME_{TIMEOUT,UNREACHABLE,UNMAPPED}` and
matching `is_network_transport_error` classification helpers on both
`ToolkitError` and the wire-model `ToolCallError`.
> 
> Re-exports `NetworkTransportError` from `arcade-tdk` and
`arcade-mcp-server`, bumps package versions (`arcade-core` 4.7.0,
`arcade-tdk` 3.7.0, `arcade-mcp-server` 1.20.0) and dependency minimums,
and expands `core/test_errors.py` to cover the new kind
invariants/defaults and classification behavior.
> 
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
d2b89078729c6a67ba42684dc98445352238bc1d. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-16 18:29:13 -03:00

139 lines
5.8 KiB
Python

"""Tests for arcade_core.errors.
Covers the empty-message guard in ``ToolkitError.with_context()`` — without it,
``raise FatalToolError("")`` produces prefixed text like ``"...tool 'foo': "``
that carries no diagnostic payload in logs/agent output.
"""
import pytest
from arcade_core.errors import (
ErrorKind,
FatalToolError,
NetworkTransportError,
RetryableToolError,
ToolkitLoadError,
UpstreamError,
)
from arcade_core.schema import ToolCallError
@pytest.mark.parametrize("empty_message", ["", " ", "\t", "\n \n"])
def test_with_context_empty_message_substitutes_placeholder(empty_message):
err = FatalToolError(empty_message).with_context("my_tool")
# Prefix is preserved — kind, error type and tool name are still in the message.
assert "[TOOL_RUNTIME_FATAL]" in err.message
assert "my_tool" in err.message
# And the empty body is replaced with a recognizable placeholder so the
# message ends with diagnostic content rather than ``": "``.
assert "(no details provided)" in err.message
assert not err.message.endswith(": ")
def test_with_context_nonempty_message_unchanged():
err = FatalToolError("Spreadsheet not found").with_context("get_sheet")
assert err.message.endswith(": Spreadsheet not found")
assert "(no details provided)" not in err.message
def test_with_context_developer_message_with_empty_message_still_works():
# A non-empty developer_message is preserved alongside the placeholder body.
err = FatalToolError("", developer_message="trace: foo.py:42").with_context("my_tool")
assert "(no details provided)" in err.message
assert err.developer_message is not None
assert err.developer_message.endswith(": trace: foo.py:42")
def test_with_context_retryable_error_empty_message():
err = RetryableToolError(" ").with_context("flaky_tool")
assert "[TOOL_RUNTIME_RETRY]" in err.message
assert "(no details provided)" in err.message
def test_with_context_toolkit_load_error_empty_message():
err = ToolkitLoadError("").with_context("broken_toolkit")
assert "broken_toolkit" in err.message
assert "(no details provided)" in err.message
# ---- NetworkTransportError -------------------------------------------------
def test_network_transport_error_is_sibling_to_upstream_error():
"""NetworkTransportError and UpstreamError serve different semantic roles.
The classification helpers must not mix them up — any consumer keying on
``is_upstream_error`` (telemetry dashboards, retry playbooks) relies on
that distinction being clean.
"""
nte = NetworkTransportError("boom")
ue = UpstreamError("boom", status_code=500)
assert nte.is_network_transport_error is True
assert nte.is_upstream_error is False
assert ue.is_upstream_error is True
assert ue.is_network_transport_error is False
def test_network_transport_error_defaults():
err = NetworkTransportError("boom")
assert err.kind is ErrorKind.NETWORK_TRANSPORT_RUNTIME_UNMAPPED
assert err.can_retry is True
# No complete response was received, so there is no status code.
assert err.status_code is None
def test_network_transport_error_rejects_non_network_kind():
"""The class invariant — kind must be in the NETWORK_TRANSPORT_ namespace —
protects telemetry and classification helpers from accidental pollution."""
with pytest.raises(ValueError, match="NETWORK_TRANSPORT_"):
NetworkTransportError("x", kind=ErrorKind.UPSTREAM_RUNTIME_SERVER_ERROR)
def test_network_transport_error_to_payload_omits_status_code():
err = NetworkTransportError(
"timed out",
kind=ErrorKind.NETWORK_TRANSPORT_RUNTIME_TIMEOUT,
can_retry=True,
extra={"error_type": "PoolTimeout"},
)
payload = err.to_payload()
assert payload["status_code"] is None
assert payload["kind"] is ErrorKind.NETWORK_TRANSPORT_RUNTIME_TIMEOUT
assert payload["can_retry"] is True
assert payload["error_type"] == "PoolTimeout"
# ---- ToolCallError classification properties (wire-format schema) ----------
@pytest.mark.parametrize(
"kind, expected_tool, expected_upstream, expected_network, expected_toolkit",
[
(ErrorKind.TOOL_RUNTIME_FATAL, True, False, False, False),
(ErrorKind.TOOL_RUNTIME_RETRY, True, False, False, False),
(ErrorKind.TOOL_RUNTIME_CONTEXT_REQUIRED, True, False, False, False),
(ErrorKind.TOOL_RUNTIME_BAD_INPUT_VALUE, True, False, False, False),
(ErrorKind.TOOL_RUNTIME_BAD_OUTPUT_VALUE, True, False, False, False),
(ErrorKind.UPSTREAM_RUNTIME_SERVER_ERROR, False, True, False, False),
(ErrorKind.UPSTREAM_RUNTIME_AUTH_ERROR, False, True, False, False),
(ErrorKind.UPSTREAM_RUNTIME_NOT_FOUND, False, True, False, False),
(ErrorKind.UPSTREAM_RUNTIME_RATE_LIMIT, False, True, False, False),
(ErrorKind.UPSTREAM_RUNTIME_BAD_REQUEST, False, True, False, False),
(ErrorKind.UPSTREAM_RUNTIME_UNMAPPED, False, True, False, False),
(ErrorKind.NETWORK_TRANSPORT_RUNTIME_TIMEOUT, False, False, True, False),
(ErrorKind.NETWORK_TRANSPORT_RUNTIME_UNREACHABLE, False, False, True, False),
(ErrorKind.NETWORK_TRANSPORT_RUNTIME_UNMAPPED, False, False, True, False),
(ErrorKind.TOOLKIT_LOAD_FAILED, False, False, False, True),
],
)
def test_tool_call_error_classification_properties(
kind, expected_tool, expected_upstream, expected_network, expected_toolkit
):
"""ToolCallError (Pydantic wire-format model) classification helpers must
be consistent with the ToolkitError class hierarchy helpers."""
err = ToolCallError(message="test", kind=kind, can_retry=False)
assert err.is_tool_error is expected_tool
assert err.is_upstream_error is expected_upstream
assert err.is_network_transport_error is expected_network
assert err.is_toolkit_error is expected_toolkit