Conversation
# Conflicts: # src/serena/cli.py
…from language servers that don't support the file type
…, tech stack, and updated suggested commands
# Conflicts: # src/solidlsp/language_servers/terraform_ls.py
1. Enhance SearchForPatternTool UX & Context Efficiency - Lowered default_max_tool_answer_chars from 150,000 to 50,000. 150k is too large, causing LLM context dilution (lost in the middle) and wasting tokens. 50k is a much safer sweet spot. - When a search result exceeds the max limit, instead of throwing a generic error, the tool now parses the results and returns a smart summary containing the total matches and the 'Top 10 files with the most matches'. This actively guides the AI Agent to iteratively refine its search query rather than hitting a dead end. 2. Clean up noisy logs and suppress misleading errors - asyncio / Windows: Added a custom exception handler to the main event loop to silently suppress ConnectionResetError [WinError 10054] and pipe transport errors. These happen naturally when clients/language servers close connections but were polluting the logs with scary tracebacks on Windows. - tools_base.py: Standard AI argument errors like ValueError and FileNotFoundError are now logged as simple WARNINGs instead of full ERROR stack traces. - TypeScript LS: Added a dummy handler for the $/typescriptVersion notification to eliminate the Unhandled method warning on startup. 3. Fix Terraform Language Server crash - Initialized self.completions_available = threading.Event() inside TerraformLS.__init__. Previously, the server would crash with an AttributeError when trying to call .set() on this missing attribute.
MischaPanch
left a comment
MischaPanch
left a comment
There was a problem hiding this comment.
Thanks for the contribution! For the future, it would be better to split up independent changes across different PRs, but since it's overall fairly small changes, this time we can let it slide
| # For a specific file, only use the language server that supports it | ||
| try: | ||
| lang_server = self._ls_manager.get_language_server(within_relative_path) | ||
| symbol_roots = lang_server.request_full_symbol_tree(within_relative_path=within_relative_path) |
There was a problem hiding this comment.
should be factored into a local function
| def window_log_message(msg: dict) -> None: | ||
| log.info(f"LSP: window/logMessage: {msg}") | ||
|
|
||
| def workspace_configuration(params: dict) -> list[dict]: |
There was a problem hiding this comment.
why is this needed? Pls add a block comment
| @@ -0,0 +1,20 @@ | |||
| # Serena Coding Conventions | |||
There was a problem hiding this comment.
pls rollback all changes to memories
| and rate limits may apply. | ||
| """ | ||
| default_max_tool_answer_chars: int = 150_000 | ||
| default_max_tool_answer_chars: int = 50_000 |
| loop.default_exception_handler(context) | ||
|
|
||
| try: | ||
| loop = asyncio.get_running_loop() |
There was a problem hiding this comment.
I have never seen excessive logs at startup and this is complex logic that dives into the MCP server's event loop. Is this really needed?
| file_to_matches[match.source_file_path].append(match.to_display_string()) | ||
| result = self._to_json(file_to_matches) | ||
|
|
||
| # Smart limit checking: if the result is too long, we provide a summary of the top 10 files instead of a generic failure message |
|
@phuongvm please submit one PR per logical change. In general, never submit PRs that contain several unrelated changes. The change you made to the search tool is something we would integrate, so do issue a PR on that. |
4 participants
🚀 What does this PR do?
This Pull Request introduces several stability fixes, cleans up noisy stack traces, and significantly improves the UX of the SearchForPatternTool when dealing with large
results.
📝 Detailed Changes & Benefits:
SearchForPatternToolUX & Context Efficiencyfor modern LLM context windows.
summary containing the total matches and the "Top 10 files with the most matches".
giving up.
errors.
WARNINGs instead of full ERROR stack traces.