bug-reports/mt5-mcp-build-6316.md
168 lines
8.7 KiB
Markdown
Vladislav Boyko
2026-11-13 14:18:39 -08:00
# MetaTrader 5 MCP servers, build 6316: faults in the file tools
Found on 2026-11-13 while comparing the tools that the two built-in MCP servers have in common.
## Environment
- MetaTrader 5 and MetaEditor, build 6316, 64-bit, running under Wine on Linux.
- Last tried on build 6327: faults 1 to 4 are reproduced as described, and the definitions
that faults 5 and 6 are about are unchanged.
Vladislav Boyko
2026-11-13 14:18:39 -08:00
- The servers as they name themselves in the handshake: `MetaTrader 5 MCP 1.0` (terminal) and
`MetaEditor 5 MCP 1.0` (MetaEditor), protocol `2025-06-18`.
- Every call below was sent to both servers with the same arguments (`tools/call` over
Streamable HTTP) and the results were compared.
- The test file is the standard `MQL5\Include\Object.mqh` as installed (33 lines, CRLF line
ends). Its line 8 is `//| Class CObject. |`
and its line 11 is `class CObject`. `<MQL5>` below stands for the absolute path of the
`MQL5` folder, as `get_workspace_info` returns it in `workspace.mql5_folder`.
## Summary
| # | Server | Tool | Fault |
| --- | --- | --- | --- |
| 1 | MetaEditor | `search_regex` | The pattern is lowercased when `case_sensitive` is false, escapes included: `\S` becomes `\s` |
| 2 | both | `search_regex` | `$` does not match at the end of a line in a file with CRLF line ends |
| 3 | MetaEditor | `list_directory`, `read_file_by_lines`, `search_regex` | An empty result where the terminal's server returns an error |
| 4 | MetaEditor | `get_time_information` | The description names two values that the result does not have |
| 5 | MetaEditor | `delete_file` | The description of `path` is the one of a tool that creates a file |
| 6 | MetaEditor | `read_binary_file`, `replace_text_in_file` | Descriptions less exact than the terminal's (not tested) |
Faults 1 and 2 matter most: the search returns a wrong result with no error.
## 1. MetaEditor, `search_regex`: the pattern is lowercased, escapes included
With `case_sensitive` false, which is the default, MetaEditor's server lowercases the whole
pattern before it is used. That changes the meaning of the escaped classes: `\S` becomes `\s`,
`\W` becomes `\w`, `\D` becomes `\d`, `\B` becomes `\b`.
To reproduce, call `search_regex` on MetaEditor's server:
```json
{
"path": "<MQL5>\\Include",
"pattern": "\\S\\sCObject",
"file_glob": "Object.mqh"
}
```
- Expected (and what the terminal's server returns): `total_matches` 3, in lines 8, 11 and 31,
and `pattern` in the result is `\S\sCObject`.
- MetaEditor's server returns: `total_matches` 5, in lines 14, 15, 18, 21 and 23 (none of the
right ones: these are the lines with two white-space characters before `CObject`), and
`pattern` in the result is `\s\scobject`.
With `"case_sensitive": true` added, both servers return the 3 right matches and the pattern
unchanged.
The terminal's server does not have this fault: it reports the pattern as it was given and
still matches without regard to case. (`search_text` lowercases its `query` on both servers,
which does no harm there: the text is literal.)
## 2. Both servers, `search_regex`: `$` does not match at the end of a line with CRLF
In a file with CRLF line ends, the end-of-line anchor matches only after the carriage return,
so the line seems to be matched with its `\r` still in it. CRLF is MetaEditor's own format for
source files (`get_workspace_info` returns `"newline": "CRLF"`).
To reproduce, call `search_regex` on either server with these patterns, each with
`"path": "<MQL5>\\Include"`, `"file_glob": "Object.mqh"` and `"case_sensitive": true`:
| `pattern` | Expected `total_matches` | Returned |
| --- | --- | --- |
| `CObject$` | 1 (line 11, `class CObject`) | 0 |
| `CObject\r$` | | 1 (line 11) |
| `^$` | 1 (line 16, the file's one empty line) | 0 |
| `^\r$` | | 1 (line 16) |
| `^class CObject` | 1 | 1 (line 11) |
`^` works as expected; only `$` is affected.
## 3. MetaEditor: an empty result where the terminal's server returns an error
In three cases MetaEditor's server reports success with nothing in it, where the terminal's
server returns `isError: true` and a message. A client cannot tell "nothing there" from "the
request was wrong".
| Tool | Arguments | MetaEditor | Terminal |
| --- | --- | --- | --- |
| `list_directory` | `"path": "<MQL5>\\nosuchdir"` (a folder that does not exist) | `"ok": true` and an empty `entries` | error: `Tool 'list_directory' specified folder does not exist` |
| `read_file_by_lines` | `"path": "<MQL5>\\Include\\Object.mqh"`, `"start_line": 100000` | an empty text | error: `Tool 'read_file_by_lines' start_line 100000 is out of file, only 33 lines available` |
| `search_regex` | `"path": "<MQL5>\\Include"`, `"pattern": "class ("`, `"file_glob": "*.mqh"` (an invalid pattern) | `"total_matches": 0` and an empty `results` | error: `Tool 'search_regex' search failed: invalid regular expression or search engine error` |
Expected: MetaEditor's server answers as the terminal's does.
## 4. MetaEditor, `get_time_information`: the description names values that are not returned
MetaEditor's description of the tool:
> Get current time information from the connected MetaTrader 5 terminal and the local computer.
> Returns UTC/GMT time, local computer time, local time offset from UTC in minutes, daylight
> saving time status, calculated trade server time, and last known server time if available.
Its result has four fields only, with the terminal running and connected to a trade server:
```json
{
"utc_time": "2026-11-13T21:10:49Z",
"local_time": "2026-11-13T13:10:49",
"local_utc_offset_minutes": -480,
"daylight_saving_time": false
}
```
There is no calculated trade server time and no last known server time. The terminal's server,
called at the same moment, returned the same four fields and
`"trade_server_last_known_time": "2026-11-13T23:10:48"`, and its description matches its
result (it names the fields and says when the last one is left out).
Expected: either the two values in MetaEditor's result, or a description that does not name
them.
## 5. MetaEditor, `delete_file`: the description of `path` belongs to another tool
In MetaEditor's `tools/list`, the parameter `path` of `delete_file` is described as:
> Absolute file path to create. The path must be under an allowed write root returned by
> get_workspace_info. Do not pass relative paths. Parent directories will be created
> automatically.
The terminal's server describes the same parameter of the same tool as:
> Absolute file path to delete. The path must be under an allowed write root returned by
> get_workspace_info. Do not pass relative paths.
To see it: `tools/list` on MetaEditor's server, the tool `delete_file`,
`inputSchema.properties.path.description`. Seen in the definition only: the tool was not
called.
## 6. MetaEditor: descriptions less exact than the terminal's (not tested)
Two more differences between the two servers' definitions of the same tools. Whether
MetaEditor's tools behave as the terminal's descriptions say was tested for the first only.
- `read_binary_file`, parameter `length`. MetaEditor: "Maximum number of bytes to read." The
terminal adds: "Use 0 to read all remaining bytes from the specified offset to the end of the
file." MetaEditor's server does read to the end with `length` 0, so only its description is
short of that sentence.
- `replace_text_in_file`. MetaEditor: "Replace literal case-sensitive text in a file while
preserving the rest of the file unchanged." The terminal's description adds that the search
is within a single line, that multiline search is not supported and that `search` must not
contain carriage-return or newline characters, and says the same in the descriptions of
`search`, `start_line` and `end_line`. Not called on either server.
## For comparison: what else differs between the two servers
Not faults, but differences a client meets when it uses the same tool on both servers:
- `list_directory`, `find_files_by_glob` and `find_files_by_name_keyword`: MetaEditor's server
adds `"git_status"` to every file entry (`"not under git control"` in this test); the
terminal's server has no such field.
- `get_workspace_info`: the terminal's result schema has four fields that MetaEditor's has not
(`capabilities.can_run_optimization`, `capabilities.can_read_tester_journal`,
`community.community_balance`, `community.community_currency`); the folders, the read, write
and execute roots, the limits and the editor settings were the same in both results.
Everything else that was tried gave the same result on both servers: `read_file` (a whole file,
`offset` and `length`, an offset past the end, a file that does not exist, a path outside the
read roots, a path with `..`), `read_file_by_lines` (ranges, `max_lines`, `encoding`),
`read_binary_file`, `search_text`, and the three listing tools but for `git_status`.