fix(fetch_url): surface network errors as 400 with detail

Previously, when httpx.get raised an HTTPError (ConnectError for host
unreachable, ReadTimeout for slow servers, RemoteProtocolError, etc.)
the exception bubbled up through the route handler as a bare 500
"Internal Server Error". The LLM got no information about what
actually went wrong — could not tell whether the host was down, the
port was closed, DNS failed, TLS handshake broke, or the request
timed out. The only thing the agent could do was guess.

Wrap the redirect loop in try/except for httpx.HTTPError and
translate to ValueError. The existing exception handler in
server.py turns ValueError into HTTP 400 with the message in the
response detail, so the LLM now sees e.g.:

  fetch_url could not reach 'http://nm01.prod.internal:8042/':
  ConnectError: Connection refused. Check that the URL is reachable
  from the MCP service, the host is in Connection.url_allowlist, and
  the connection's auth/SSL settings are correct.

The original exception is chained via `raise ... from exc` so loguru
still records the full traceback with the original type, and the
`__cause__` attribute is set on the ValueError for programmatic
inspection.

Note: upstream HTTP 4xx/5xx responses (server replied, even with an
error status) are NOT translated — the FetchUrlResult carries the
status code and body so the LLM can read what the server actually
said. This is the intentional contrast with the no-response-at-all
case (which now has clear 400 detail).

Tests (3 new in tests/unit/test_fetch_url.py):
  - test_fetch_url_raises_400_with_detail_on_connect_error
    ConnectError("Connection refused") -> ValueError with
    "ConnectError", "Connection refused", the URL, and __cause__
    chained.
  - test_fetch_url_raises_400_with_detail_on_timeout
    ReadTimeout("Timed out reading") -> ValueError with
    "ReadTimeout", "Timed out reading", __cause__ chained.
  - test_fetch_url_returns_body_for_4xx_5xx_upstream
    Upstream 503 with body "Service Unavailable - try again later"
    -> FetchUrlResult(status_code=503, body=...). Proves the
    intentional contrast.

Route description in server.py updated with a new **Errors** section
explaining the two error paths (no response = 400 with detail, got
a response = body returned).

Tests: 401 passed (was 398, +3 net).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

spark_executor/server.py

@@ -557,7 +557,15 @@ def _update_job_file(req: UpdateJobFileRequest):
         "Connection's saved auth is reused, so the agent does not need cluster "
         "credentials.\n\n"
         "**Limits:** 30s timeout, redirects followed, response body capped at "
-        "1 MB (the response includes a `truncated` boolean when this kicks in)."
+        "1 MB (the response includes a `truncated` boolean when this kicks in).\n\n"
+        "**Errors:** when the host is unreachable (connect refused, DNS "
+        "failure, TLS handshake error, timeout, etc.) this tool returns "
+        "HTTP 400 with the exception class and message in the response "
+        "detail — e.g. `fetch_url could not reach ...: ConnectError: "
+        "Connection refused`. Upstream HTTP 4xx/5xx responses that DID "
+        "come back are returned in the result body with their status code "
+        "preserved (not translated to an error) so you can see what the "
+        "server actually said."
     ),
 )
 def _fetch_url(req: FetchUrlRequest):

spark_executor/tools/fetch_url.py

@@ -83,13 +83,27 @@ def fetch_url(url: str, connection_name: str) -> FetchUrlResult:
     verify = config.verify_for_httpx()
 
     for hop in range(_MAX_REDIRECTS + 1):
-        resp = httpx.get(
-            url,
-            auth=auth,
-            verify=verify,
-            timeout=_REQUEST_TIMEOUT_SECONDS,
-            follow_redirects=False,
-        )
+        try:
+            resp = httpx.get(
+                url,
+                auth=auth,
+                verify=verify,
+                timeout=_REQUEST_TIMEOUT_SECONDS,
+                follow_redirects=False,
+            )
+        except httpx.HTTPError as exc:
+            # Network-level failure (no response received). Surface the
+            # exception class + message so the LLM can act on it.
+            # ValueError -> 400 via the existing handler in server.py.
+            # The cause chain (`from exc`) preserves the original
+            # exception for loguru.
+            raise ValueError(
+                f"fetch_url could not reach {url!r}: "
+                f"{type(exc).__name__}: {exc}. "
+                f"Check that the URL is reachable from the MCP service, "
+                f"the host is in Connection.url_allowlist, and the "
+                f"connection's auth/SSL settings are correct."
+            ) from exc
         if resp.status_code not in _REDIRECT_STATUSES:
             logger.info(
                 f"fetch_url ok url={url} connection_name={connection_name} "

tests/unit/test_fetch_url.py

@@ -327,3 +327,76 @@ def test_list_applications_request_limit_bounds():
         ListApplicationsRequest(connection_name="prod", limit=0)
     with pytest.raises(ValidationError):
         ListApplicationsRequest(connection_name="prod", limit=10001)
+
+
+# --- Network error handling (httpx.HTTPError -> 400 with detail) ---
+
+
+def test_fetch_url_raises_400_with_detail_on_connect_error(fresh_stores):
+    """When httpx.get raises ConnectError (host unreachable / port closed),
+    fetch_url must raise ValueError (-> 400) with the exception class
+    and message in the detail. Previously this bubbled up as a bare
+    500 Internal Server Error with no info."""
+    fetch_url.conn_store.save(
+        Connection(
+            name="prod",
+            master="yarn",
+            yarn_rm_url="http://rm.prod.internal:8088",
+            url_allowlist=["*.prod.internal"],
+        )
+    )
+    with patch(
+        "spark_executor.tools.fetch_url.httpx.get",
+        side_effect=httpx.ConnectError("Connection refused"),
+    ):
+        with pytest.raises(ValueError) as ei:
+            fetch_url.fetch_url("http://nm01.prod.internal:8042/", "prod")
+    msg = str(ei.value)
+    assert "ConnectError" in msg
+    assert "Connection refused" in msg
+    assert "nm01.prod.internal" in msg
+    # Cause chain preserved for loguru
+    assert isinstance(ei.value.__cause__, httpx.ConnectError)
+
+
+def test_fetch_url_raises_400_with_detail_on_timeout(fresh_stores):
+    """When httpx.get raises TimeoutException (request exceeded 30s),
+    fetch_url must raise ValueError with the exception details surfaced."""
+    fetch_url.conn_store.save(
+        Connection(
+            name="prod",
+            master="yarn",
+            yarn_rm_url="http://rm.prod.internal:8088",
+            url_allowlist=["*.prod.internal"],
+        )
+    )
+    with patch(
+        "spark_executor.tools.fetch_url.httpx.get",
+        side_effect=httpx.ReadTimeout("Timed out reading"),
+    ):
+        with pytest.raises(ValueError) as ei:
+            fetch_url.fetch_url("http://nm01.prod.internal:8042/", "prod")
+    msg = str(ei.value)
+    assert "ReadTimeout" in msg
+    assert "Timed out reading" in msg
+    assert isinstance(ei.value.__cause__, httpx.ReadTimeout)
+
+
+def test_fetch_url_returns_body_for_4xx_5xx_upstream(fresh_stores):
+    """Upstream HTTP errors (4xx/5xx responses that DID come back) are
+    NOT translated to ValueError — the FetchUrlResult carries the status
+    code and body so the LLM can see what the server actually said. This
+    is the intentional contrast with the no-response-at-all case."""
+    fetch_url.conn_store.save(
+        Connection(
+            name="prod",
+            master="yarn",
+            yarn_rm_url="http://rm.prod.internal:8088",
+            url_allowlist=["*.prod.internal"],
+        )
+    )
+    resp = httpx.Response(503, text="Service Unavailable - try again later")
+    with patch("spark_executor.tools.fetch_url.httpx.get", return_value=resp):
+        out = fetch_url.fetch_url("http://nm01.prod.internal:8042/", "prod")
+    assert out.status_code == 503
+    assert "Service Unavailable" in out.body