Problem:
Closing auxiliary windows (e.g., from "open in editor") could prematurely dispose the shared singleton McpHub instance, shutting down MCP servers unexpectedly while other panels (like the sidebar) might still be active.
Solution:
Implemented reference counting directly within the McpHub class to manage its own lifecycle based on active clients (ClineProvider instances).
- Added `refCount`, `registerClient()`, and `unregisterClient()` methods to McpHub.
- McpHub now disposes itself only when the last registered client unregisters (`refCount` reaches 0).
- ClineProvider instances now call `mcpHub.registerClient()` upon initialization and `mcpHub.unregisterClient()` upon disposal.
- Removed the direct `mcpHub.dispose()` call from ClineProvider.dispose.
Benefit:
This ensures the shared McpHub instance remains active as long as at least one ClineProvider instance is using it. Cleanup now correctly occurs only when the last provider is closed or during full extension deactivation. This centralizes the resource's lifecycle logic within the resource class itself.
Files Changed:
- src/services/mcp/McpHub.ts
- src/core/webview/ClineProvider.ts
* refactor: Change read_file result format to XML
Modifies how the result of the `read_file` tool is presented
in the conversation history sent to the AI model.
Previously, the format was:
[read_file for 'path/to/file'] Result:
{content_or_error}
This commit changes the format to use XML tags for better
structure and potentially easier parsing by the model:
<file>
<path>path/to/file</path>
<content>
{content_or_error}
</content>
</file>
This change only affects the `read_file` tool result formatting
within the user context message constructed in `src/core/Cline.ts`.
Other tool result formats remain unchanged.
* fix: Update error message for better clarity in multi-search-replace strategy
Refines the error message returned when no sufficiently similar match is found during the multi-search-replace operation. The message now includes a clearer instruction to use the read_file tool for obtaining the latest file content before attempting to apply the diff again.
* refactor: Update readFileTool to return results in XML format
Modifies the `readFileTool` function to format the output as XML, enhancing the structure of the returned file content. This change aligns with previous updates to ensure consistent result formatting across tools.
* refactor: Simplify result handling for tool responses in Cline
This update refines the handling of tool responses in the `Cline` class by removing the XML formatting for `read_file` results and consolidating the logic for pushing results to the user message content. The changes ensure that all tool results are processed uniformly, improving code clarity and maintainability.
* test: Update read_file tests to validate XML formatted results
This commit modifies the assertions in the `read_file` tool tests to check for the expected XML format in the results. The changes ensure that the output structure aligns with recent updates to the tool's response formatting, enhancing test accuracy and reliability.
* test: Refactor assertions in read_file tests to use expected XML variable
This commit updates the `read_file` tool tests to utilize a predefined variable for the expected XML output, improving readability and maintainability of the test code. The changes ensure consistency in the expected results across multiple test cases.
* fix: Improve handling of escaped markers in apply_diff
unescapeMarkers() is supposed to fix escaped merge conflict markers so
that they match the code. However the function has a bug which requires
SEARCH and REPLACE strings in the markers to be replaced. This is not
part of merge conflict markers, so many valid cases were previously
missed.
* Revert test changes
* Add new test
---------
Co-authored-by: Matt Rubens <mrubens@users.noreply.github.com>
Currently all errors in apply_diff diff structure are assumed to be
related to merge conflicts. This is unfortunate because it is way more
likely that LLM simply made an error structuring the diff. The LLM is
then directed to fix escaping in the diff from which it doesn't recover.
It is safe to assume that missing or extra "<<<<<<< SEARCH" or ">>>>>>>
REPLACE" strings are likely bad diff structure, because while "<<<<<<<"
and ">>>>>>>" are common merge conflict markers, no tools add "SEARCH"
and "REPLACE" strings after them. This is likely output of the LLM
itself.
LLM not recovering has been observed in Claude 3.5 Sonnet and Claude 3.7
Sonnet models.
To address the issue, error message now mentions merge conflict markers
and their escaping only if it's clear that errors may come from this
source. In the rest of cases the error message repeats the expected diff
structure back to the LLM.
* Bedrock prompt caching
* remove whitespace only change to src/api/providers/base-provider.ts
* remove integration test that calls bedrock
* PR prep cleanup
* test file and redundant code cleanup
* updated multi-cache-point
* optimized cache strategy
* cache logic updated. commit before trying to simplify newConversationPlacements
* simplified initial cachePoint logic
* fix the missing caching params for claude 3.7 on bedrock
* undo unintentional changes to README.md files
* fix merge bug that hid use prompt cache checkbox
* prep for PR
* remove unintended changes
* revert .gitignore change
* undo unintended changes
* Issue 1998 - Unable to use ARN. https://github.com/RooVetGit/Roo-Code/issues/1998 - Simplify region extraction from Arns. Change example of ARN use from the foundational model ARN to an inference profile ARN which is what is needed.
* small updates to error message
* more flexible support for ARNs and hardening of ARN handling
* update comment on regex use for platform indipendent cosideration
* small code comment update
* remove duplicative logic to add a region to the arn when cross-region is selected
* Update README files from main branch
* improve error handling when on-demand throughput is not avialable for a foundational model
* fix prompt router bug
* minor tweaks for simplicity
* error messgae formatting
* set new version of bedrock runtime package
* fixes: 2229 - complete prompt doesn't work and appears like a no-op
* remove commented out logger.debug blocks per PR review request
* localize new prompt cache UI inputs
* single source of truth for AWS Region information
---------
Co-authored-by: Matt Rubens <mrubens@users.noreply.github.com>
* feat: Add support for Azure AI Inference Service with DeepSeek-V3 model
* refactor: extract Azure AI inference path to constant to avoid duplication
* fix(tests): update RequestyHandler tests to properly handle Azure inference and streaming
* fix(api): remove duplicate constant and update requesty tests
* refactor: remove unused isAzure property from OpenAiHandler
* refactor(openai): remove unused isAzure and extract Azure check
* Enable model switching on API failure
Bug: Cannot change model selection after API error due to UI state #1657
UI Bug: OpenRouter ran out of credits prevents user from switching models #1206
* Remove irrelevant code
make DropdownMenu scrollable with 80vh height limit
The DropdownMenu component now automatically becomes scrollable when its content
exceeds 80% of the viewport height. This enhances usability by:
- Limiting maximum height to 80vh using Tailwind's max-h-[80vh]
- Enabling vertical scrolling with overflow-y-auto
- Preserving dropdown positioning and accessibility
The multi-search-replace diff strategy previously did not correctly strip line numbers (`number | `) from the SEARCH block when the REPLACE block was empty. This occurred because the condition for stripping required both blocks to consistently have line numbers.
This prevented successful deletion operations when the SEARCH block content was copied from `read_file` output (which includes line numbers) and the REPLACE block was empty.
This commit updates the line number stripping condition in `applyDiff` to also trigger if the SEARCH block has line numbers and the REPLACE block is empty or contains only whitespace, resolving the bug.
Additionally, a new test case has been added to `multi-search-replace.test.ts` to specifically verify this deletion scenario. All tests now pass with this updated logic.
