The previous implementation buffered stderr from the native mcp-server process and only logged the full buffer content when the process exited. This prevented real-time viewing of logs from the server.
This change modifies the native process startup logic (`startProcess_locked`) to use `cmd.StderrPipe()` instead of assigning `cmd.Stderr` to a buffer. A separate goroutine is now launched within the process monitoring goroutine. This new goroutine uses a `bufio.Scanner` to continuously read lines from the stderr pipe and logs them using the Navidrome logger (`log.Info`) with an `[MCP-SERVER]` prefix.
This ensures logs from the native mcp-server appear in Navidrome's logs immediately as they are written.
(Note: Also includes update to McpServerPath constant to point to the native binary.)
Signed-off-by: Deluan <deluan@navidrome.org>
Added debug logging throughout the mcp-server components using the standard Go `log` package. All log messages are prefixed with `[MCP]` for easy identification.
This includes logging in: main function (startup, CLI execution, registration, serve loop), Tool handlers, Native and WASM fetcher implementations, Wikipedia, Wikidata, and DBpedia data fetching functions
Replaced previous `println` statements with `log.Println` or `log.Printf`.
Modified the MCP agent constructor to pre-compile the WASM module when detected. This shifts the costly compilation step out of the first API request path.
The `MCPWasm` implementation now stores the `wazero.CompiledModule` provided by the constructor and uses it directly for instantiation via `runtime.InstantiateModule()` when the agent is first used or restarted. This significantly speeds up the initialization during the first request.
Updated tests and cleanup logic to accommodate the shared nature of the pre-compiled module.
Centralized the argument preparation for GetArtistBiography and GetArtistURL within the main MCPAgent struct. Added callMCPTool to the mcpImplementation interface and removed the redundant GetArtist* methods from the native and WASM implementations.
Removed the embedded agents.Artist*Retriever interfaces from mcpImplementation as MCPAgent now directly fulfills these.
Also removed various redundant comments and leftover commented-out code from the agent, implementation, and test files.
Refactored the MCPAgent to delegate core functionality to separate implementations for native process and WASM execution.
Introduced an `mcpImplementation` interface (`MCPNative`, `MCPWasm`) to abstract the underlying execution method. The main `MCPAgent` now holds an instance of this interface, selected by the `mcpConstructor` based on the `McpServerPath` (native executable or `.wasm` file).
Shared Wazero resources (runtime, cache, WASI, host functions) are now initialized once in the constructor and passed to the `MCPWasm` implementation, improving resource management and potentially startup performance for WASM modules.
Updated tests (`mcp_agent_test.go`) to use a new testing constructor (`NewAgentForTesting`) which injects a mock client into the appropriate implementation. Assertions were adjusted to reflect the refactored error handling and structure. Also updated the `McpServerPath` to use a relative path.
Change the MCP server path in MCPAgent to point to the correct relative directory for the WASM file. This adjustment ensures proper initialization and access to the server resources, aligning with recent enhancements in the MCPAgent's handling of server types.
- Moved the native process handling logic from mcp_agent.go to a new file mcp_process_native.go for better organization.
- Introduced a new file mcp_host_functions.go to define and register host functions for WASM modules.
- Updated MCPAgent to ensure proper initialization and cleanup of both native and WASM processes, enhancing code clarity and maintainability.
- Added comments to clarify the purpose of changes and ensure future developers understand the structure.
Change the MCP server path in MCPAgent from a WASM file to a directory. This adjustment aligns with recent enhancements to the MCPAgent's handling of server types, ensuring proper initialization and access to the server resources.
Enhance the MCPAgent's WASM module initialization by granting access to the host filesystem. This is necessary for DNS lookups and other operations that may depend on filesystem access. Added comments to highlight the security implications of this change and the need for potential restrictions in the future.
Enhance MCPAgent to support both native executables and WASM modules using Wazero. This includes:
- Adding Wazero dependencies in go.mod and go.sum.
- Modifying MCPAgent to initialize a shared Wazero runtime and compile/load WASM modules.
- Implementing cleanup logic for WASM resources.
- Updating the process initialization to handle both native and WASM paths.
This change improves the flexibility of the MCPAgent in handling different server types.
Modify callMCPTool helper to check if the text content returned by\nthe MCP server starts with 'handler returned an error:'.\n\nIf it does, log a warning and return agents.ErrNotFound, treating\nthe error signaled by the external tool as if the data was not found.\nAdded test cases to verify this behavior.
Refactor the MCPAgent to reduce code duplication.\n\n- Consolidate GetArtistBiographyArgs and GetArtistURLArgs into a single\n ArtistArgs struct.\n- Extract common logic (initialization check, locking, tool calling,\n error handling, response validation) into a private callMCPTool helper method.\n- Simplify GetArtistBiography and GetArtistURL to delegate to callMCPTool.\n- Update tests to use the consolidated ArtistArgs struct.\n- Correct mutex locking in ensureClientInitialized to prevent race conditions.
Modify the createAgents function to call each agent constructor only once.\n\nPreviously, the constructor (init(ds)) was called first to check if the\nagent could be initialized, and then called again when appending the agent\nto the result slice. This caused unnecessary work and duplicate log messages.\n\nThe fix stores the result of the first call in the 'agent' variable and\nreuses this instance when appending, ensuring each constructor runs only once.
Implement unit tests for MCPAgent using a mocked MCP client.\n\n- Define an mcpClient interface and a mock implementation in the test file.\n- Modify MCPAgent to use the interface and add an exported ClientOverride field\n to allow injecting the mock during tests.\n- Export necessary constants and argument structs from the agent package.\n- Add test cases covering success, tool errors, empty responses, and pipe errors\n for both GetArtistBiography and GetArtistURL.\n- Fix agent logic to handle empty TextContent in responses correctly.\n- Remove previous placeholder tests and unreliable initialization test.
Add support for retrieving artist URLs via the MCP agent.\n\n- Implement the agents.ArtistURLRetriever interface.\n- Add the GetArtistURL method, which calls the 'get_artist_url'\n tool on the external MCP server.\n- Define the necessary constants and argument struct.\n\nThis mirrors the existing implementation for GetArtistBiography, reusing\nthe persistent process and client logic.
Modify the MCPAgent to automatically attempt restarting the external\nserver process if it detects the process has exited.\n\n- Replaced sync.Once with mutex-protected checks (a.client == nil) to allow\n re-initialization.\n- The monitoring goroutine now cleans up agent state (nils client, cmd, stdin)\n upon process exit, signaling the need for restart.\n- ensureClientInitialized now attempts to start/initialize if the client is nil.\n- GetArtistBiography checks client validity again after locking to handle race\n conditions where the process might die just after initialization check.
Refactor MCPAgent to maintain a persistent external server process.\n\nInstead of starting and stopping the MCP server for each request, the agent\nnow uses sync.Once for lazy initialization on the first call.\nIt stores the running exec.Cmd, stdio pipes, and the mcp.Client\nin the agent struct.\n\nA sync.Mutex protects concurrent access to the client/pipes.\nA goroutine monitors the process using cmd.Wait() and logs if it exits\nunexpectedly.\n\nThis avoids the overhead of process creation/destruction on every\nmetadata retrieval request.
Add a new agent implementation MCPAgent in core/agents/mcp.\n\nThis agent interacts with an external Model Context Protocol (MCP) server\nusing the github.com/metoro-io/mcp-golang library via stdio.\n\nIt currently implements the ArtistBiographyRetriever interface to fetch\nbiographies by calling the 'get_artist_biography' tool on the MCP server.\nThe path to the MCP server executable is hardcoded for this PoC.\n\nIncludes basic Ginkgo test setup for the new agent.
Added Ginkgo/Gomega tests for userName and AbsolutePath functions in core/common.go. Tests cover normal and error cases, using existing mocks and helpers. This improves coverage and ensures correct behavior for user context extraction and library path resolution.
* fix(ui): ensure album tracks are always ordered by disc and track number (fixes#3720)
* refactor(ui): remove obsolete release date grouping logic from SongDatagrid and AlbumSongs
* fix(ui): ensure correct album track ordering in context menu and play button
* fix: Update album sort to use album_id instead of release_date
* refactor: Adjust filters in PlayButton and AlbumContextMenu
* fix: correct typo in comment regarding participants in GetMissingAndMatching function
* fix: prevent visual separation of tracks on same disc
Removes the leftover `releaseDate` check from the `firstTracksOfDiscs` calculation in `SongDatagridBody`. This check caused unnecessary `DiscSubtitleRow` insertions when tracks on the same disc had different release dates, leading to an incorrect visual grouping that resembled a multi-disc layout.
This change ensures disc subtitles are only shown when the actual `discNumber` changes, correcting the UI presentation issue reported in issue #3720 after PR #3975.
* fix: remove remaining releaseDate references in SongDatagrid
Cleaned up leftover `releaseDate` references in `SongDatagrid.jsx`:
- Removed `releaseDate` parameter and usage from `handlePlaySubset` in `DiscSubtitleRow`.
- Removed `releaseDate` prop passed to `AlbumContextMenu` in `DiscSubtitleRow`.
- Removed `releaseDate` from the drag item data in `SongDatagridRow`.
- Removed `releaseDate` parameter and the corresponding `else` block from the `playSubset` function in `SongDatagridBody`.
This ensures the component consistently uses `discNumber` for grouping and playback actions initiated from the disc subtitle, fully resolving the inconsistencies related to issue #3720.
* fix(server): bring back legacy date mappings
Signed-off-by: Deluan <deluan@navidrome.org>
* reuse the mapDates logic in the legacyReleaseDate function
Signed-off-by: Deluan <deluan@navidrome.org>
* fix mappings
Signed-off-by: Deluan <deluan@navidrome.org>
* show original and release dates in album grid
Signed-off-by: Deluan <deluan@navidrome.org>
* fix tests based on new year mapping
Signed-off-by: Deluan <deluan@navidrome.org>
* fix(subsonic): prefer returning original_year over (recording) year
when sorting albums
Signed-off-by: Deluan <deluan@navidrome.org>
* fix case when we don't have originalYear
Signed-off-by: Deluan <deluan@navidrome.org>
* show all dates in album's info, and remove the recording date from the album page
Signed-off-by: Deluan <deluan@navidrome.org>
* better?
Signed-off-by: Deluan <deluan@navidrome.org>
* add snapshot tests for Album Details
Signed-off-by: Deluan <deluan@navidrome.org>
* fix(subsonic): sort order for getAlbumList?type=byYear
Signed-off-by: Deluan <deluan@navidrome.org>
---------
Signed-off-by: Deluan <deluan@navidrome.org>
* chore(server): add more info to scrobble errors
Signed-off-by: Deluan <deluan@navidrome.org>
* chore(server): add more info to scrobble errors
Signed-off-by: Deluan <deluan@navidrome.org>
* chore(server): add more info to scrobble errors
Signed-off-by: Deluan <deluan@navidrome.org>
---------
Signed-off-by: Deluan <deluan@navidrome.org>
* chore(scanner): add trace logs
Signed-off-by: Deluan <deluan@navidrome.org>
* fix(bfr): force upgrade to read all folders. It was skipping folders for certain timezones
Signed-off-by: Deluan <deluan@navidrome.org>
---------
Signed-off-by: Deluan <deluan@navidrome.org>
