Container Manager returns raw filesystem paths (/volume1/docker/...),
but SYNO.FileStation.* APIs expect paths without the volume prefix
(/docker/...). Add _to_filestation_path() to strip /volumeN and apply
it in _find_compose_path before any FileStation call.
Also switch directory probe from getinfo (returns truthy files array
with embedded code:408 for missing paths) to list (empty files array
for non-existent directories), and apply the same prefix stripping to
the error message shown when no compose file is found.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
SYNO.FileStation.Info/get is a system-info API and returns success
regardless of whether a path exists, so the probe always returned the
first candidate (docker-compose.yml) even when only compose.yaml
was present. SYNO.FileStation.List/getinfo returns {"files": [...]}
with an empty list for non-existent paths, enabling correct detection.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_find_compose_path was constructing {compose_base_path}/{project_name}
which didn't match the actual NAS path (e.g. /volume1/docker/frostiq/jenkins).
Now calls _find_project() first and uses project["path"] as the base
directory, with the old constructed path as a fallback only.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
login() called client.request() which called _ensure_initialized() which
tried to re-acquire _init_lock — deadlocking forever. Fix: set
_initializing=True while inside the init critical section so request()
skips the _ensure_initialized() guard when called from within init
(safe because query_api_info() already populated the API cache).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Write progress markers to stderr at each lazy-init step so Claude
Desktop logs show exactly where a timeout occurs
- Replace flat timeout=30 integer with httpx.Timeout(connect=10,
read=30, write=10, pool=5) to fail fast on connection issues
instead of waiting up to 90 s across three sequential requests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the server blocked at startup waiting for query_api_info()
and login() before starting the MCP protocol. Claude Desktop has a short
initialization timeout and dropped the server before the handshake started.
Changes:
- DsmClient: add _ensure_initialized() with asyncio.Lock for thread-safe
lazy init; called automatically at the start of request(), upload_text(),
and download_text() on the first use.
- cli.py serve: remove upfront query_api_info() and auth.login() calls;
the server now starts immediately ("MCP server ready" on stderr) and
connects to the NAS on the first tool invocation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Replace asyncio.run() with anyio.run() in serve command: FastMCP uses
anyio.create_task_group() internally, and anyio.run() ensures the correct
backend context. asyncio.run() can misbehave on Windows (ProactorEventLoop).
- Add SERVER STARTING / MCP server ready messages to stderr for diagnostics.
- Replace sys.exit(1) with early return + stderr write inside anyio context;
sys.exit inside anyio.run() is less predictable than a clean return.
- Eagerly import all modules at serve startup so ImportErrors surface on
stderr immediately instead of silently killing the process.
- Add __main__.py to allow python -m mcp_synology_container invocation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
marcus