This adds support for models like Qwen3-Coder-Next that output XML-style
tool calls instead of native JSON tool calls when running via llama.cpp.
XML format supported:
<function=TOOL_NAME>
<parameter=PARAM_NAME>value</parameter>
</function>
Changes:
- Add XmlToolCallParser to detect and parse XML tool calls in text stream
- Integrate parser into Task.ts text chunk processing
- Convert XML tool calls to standard tool_call events for execution
- Add comprehensive tests for the parser
Fixes#11219
* fix(ai-sdk): preserve reasoning parts in message conversion
* fix(ai-sdk): convert message-level reasoning_content to reasoning part
* fix(task): remove invalid openai-compatible from reasoning allowlist
* feat(condense): add smart code folding with tree-sitter signatures
At context condensation time, use tree-sitter to generate folded code
signatures (function definitions, class declarations) for files read
during the conversation. Each file is included as its own <system-reminder>
block in the condensed summary, preserving structural awareness without
consuming excessive tokens.
- Add getFilesReadByRoo() method to FileContextTracker
- Create generateFoldedFileContext() using tree-sitter parsing
- Update summarizeConversation() to accept array of file sections
- Each file gets its own content block in the summary message
- Add comprehensive test coverage (12 tests)
* fix: skip tree-sitter error strings in folded file context
- Add isTreeSitterErrorString helper to detect error messages
- Skip files that return error strings instead of embedding them
- Add test for error string handling
* refactor: move generateFoldedFileContext() inside summarizeConversation()
- Update summarizeConversation() to accept filesReadByRoo, cwd, rooIgnoreController instead of pre-generated sections
- Move folded file context generation inside summarizeConversation() (lines 319-339)
- Update ContextManagementOptions type and manageContext() to pass new parameters
- Remove generateFoldedFileContext from Task.ts imports - folding now handled internally
- Update all tests to use new parameter signature
- Reduces Task.ts complexity by moving folding logic to summarization module
* fix: prioritize most recently read files in folded context
Files are now sorted by roo_read_date descending before folded context
generation, so if the character budget runs out, the most relevant
(recently read) files are included and older files are skipped.
* refactor: improve code quality in condense module
- Convert summarizeConversation to use options object instead of 11 positional params
- Extract duplicated getFilesReadByRoo error handling into helper method
- Remove unnecessary re-export of generateFoldedFileContext
- Update all test files to use new options object pattern
* fix: address roomote feedback - batch error logging and early budget exit
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: daniel-lxs <ricciodaniel98@gmail.com>
* fix: convert orphaned tool_results to text blocks after condensing
When condensing occurs after assistant sends tool_uses but before user responds,
the tool_use blocks get condensed away. User messages containing tool_results that
reference condensed tool_use_ids become orphaned and get filtered out by
getEffectiveApiHistory, causing user feedback to be lost.
This fix enhances the existing check in addToApiConversationHistory to detect when
the previous effective message is not an assistant and converts any tool_result
blocks to text blocks, preventing them from being filtered as orphans.
The conversion happens at the latest possible moment (message insertion) because:
- Tool results are created before we know if condensing will occur
- We need actual effective history state to make the decision
- This is the last checkpoint before orphan filtering happens
* Only include environment details in summary for automatic condensing
For automatic condensing (during attemptApiRequest), environment details
are included in the summary because the API request is already in progress
and the next user message won't have fresh environment details injected.
For manual condensing (via condenseContext button), environment details
are NOT included because fresh details will be injected on the very next
turn via getEnvironmentDetails() in recursivelyMakeClineRequests().
This uses the existing isAutomaticTrigger flag to differentiate behavior.
---------
Co-authored-by: Hannes Rudolph <hrudolph@gmail.com>
* chore: remove MULTI_FILE_APPLY_DIFF experiment
Remove the 'Enable concurrent file edits' experimental feature that
allowed editing multiple files in a single apply_diff call.
- Remove multiFileApplyDiff from experiment types and config
- Delete MultiFileSearchReplaceDiffStrategy class and tests
- Delete MultiApplyDiffTool wrapper and tests
- Remove experiment-specific code paths in Task.ts, generateSystemPrompt.ts, and presentAssistantMessage.ts
- Remove special handling in ExperimentalSettings.tsx
- Remove translations from all 18 locale files
The existing MultiSearchReplaceDiffStrategy continues to handle
multiple SEARCH/REPLACE blocks within a single file.
* fix: remove unused EXPERIMENT_IDS/experiments import from Task.ts
Addresses review feedback: removes the unused imports from
src/core/task/Task.ts that were left over after removing the
MULTI_FILE_APPLY_DIFF experiment routing code.
* fix(condense): remove custom condensing model option
Remove the ability to specify a different model/API configuration for
condensing conversations. Modern conversations include provider-specific
data (tool calls, reasoning blocks, thought signatures) that only the
originating model can properly understand and summarize.
Changes:
- Remove condensingApiHandler parameter from summarizeConversation()
- Remove condensingApiConfigId from context management and Task
- Remove API config dropdown for CONDENSE in settings UI
- Update telemetry to remove usedCustomApiHandler parameter
- Update related tests
Users can still customize the CONDENSE prompt text; only model selection
is removed.
* fix: remove condensingApiConfigId from types and test fixtures
---------
Co-authored-by: Roo Code <roomote@roocode.com>
* feat: warn users when too many MCP tools are enabled
- Add WarningRow component for displaying generic warnings with icon, title, message, and optional docs link
- Add TooManyToolsWarning component that shows when users have more than 40 MCP tools enabled
- Add MAX_MCP_TOOLS_THRESHOLD constant (40)
- Add i18n translations for the warning message
- Integrate warning into ChatView to display after task header
- Add comprehensive tests for both components
Closes ROO-542
* Moves constant to the right place
* Move it to the backend
* i18n
* Add actionlink that takes you to MCP settings in this case
* Add to MCP settings too
* Bump max tools up to 60 since github itself has 50+
* DRY
* Fix test
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: Bruno Bergher <bruno@roocode.com>
Co-authored-by: Matt Rubens <mrubens@users.noreply.github.com>
* feat: add optional mode field to slash command front matter
- Add mode field to Command interface
- Update command parsing to extract mode from frontmatter
- Modify RunSlashCommandTool to automatically switch mode when specified
- Add comprehensive tests for mode field parsing and switching
- Update existing tests to include mode field
* Make it work for manual slash commands too
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: Matt Rubens <mrubens@users.noreply.github.com>
* Add support for skills
* fix: use type-only import for ClineProvider and relative paths in skills section
---------
Co-authored-by: Roo Code <roomote@roocode.com>
