* feat: store reasoning in conversation history for all providers
* refactor: address review feedback
- Move comments inside else block
- Combine reasoning checks into single if block
- Make comments more concise
* refactor: make comments more concise
* Fix preserveReasoning flag to control API reasoning inclusion
Changes:
1. Removed hardcoded <think> tag logic in streaming
- Previously hardcoded reasoning into assistant message text
- Now passes reasoning to addToApiConversationHistory as parameter
2. Updated buildCleanConversationHistory to respect preserveReasoning flag
- When preserveReasoning: true → reasoning block included in API requests
- When preserveReasoning: false/undefined → reasoning stripped from API
- Reasoning stored in history for all cases
3. Added temporary debug logs to base-openai-compatible-provider.ts
- Shows preserveReasoning flag value
- Logs reasoning blocks in incoming messages
- Logs <think> tags in converted messages sent to API
* Fix: Use api.getModel() directly instead of cachedStreamingModel
Addresses review comment: cachedStreamingModel is set during streaming but
buildCleanConversationHistory is called before streaming starts. Using the
cached value could cause stale model info when switching models between requests.
Now directly uses this.api.getModel().info.preserveReasoning to ensure we
always check the current model's flag, not a potentially stale cached value.
* Clean up comments in Task.ts
Removed outdated comment regarding model's preserveReasoning flag.
* fix: remove unnecessary reasoningBlock variable in task reasoning logic
- Filter out complete environment_details blocks before appending fresh ones
- Check for both opening and closing tags to ensure we're matching complete blocks
- Prevents stale environment data from being kept during task resume
- Add tests to verify deduplication logic and edge cases
* Improve read_file tool description with examples
- Add explicit JSON structure documentation
- Include three concrete examples (single file, with line ranges, multiple files)
- Clarify that 'path' is required and 'line_ranges' is optional
- Better explain line range format (1-based inclusive)
This addresses agent confusion by providing clear examples similar to the XML tool definition.
* Make read_file tool dynamic based on partialReadsEnabled setting
- Convert read_file from static export to createReadFileTool() factory function
- Add getNativeTools() function that accepts partialReadsEnabled parameter
- Create buildNativeToolsArray() helper to encapsulate tool building logic
- Update Task.ts to build native tools dynamically using maxReadFileLine setting
- When partialReadsEnabled is false, line_ranges parameter is excluded from schema
- Examples and descriptions adjust based on whether line ranges are supported
This matches the behavior of the XML tool definition which dynamically adjusts
its documentation based on settings, reducing confusion for agents.
* fix: format tool responses for native protocol
- Add toolResultFormatting utilities for protocol detection
- ReadFileTool now builds both XML and native formats
- Native format returns clean, readable text without XML tags
- Legacy conversation history conversion is protocol-aware
- All tests passing (55 total)
* refactor: use isNativeProtocol from @roo-code/types
Remove duplicate implementation and import from types package instead
* refactor: centralize toolProtocol configuration checks
- Created src/utils/toolProtocol.ts with getToolProtocolFromSettings() utility
- Replaced all direct vscode.workspace.getConfiguration() calls with centralized utility
- Updated 6 files to use the new utility function
- All tests pass and TypeScript compilation succeeds
* refactor: use isNativeProtocol function from types package
* fix: filter native tools by mode restrictions
Native tools are now filtered based on mode restrictions before being sent to the API, matching the behavior of XML tools. Previously, all native tools were sent to the API regardless of mode, causing the model to attempt using disallowed tools.
Changes:
- Created filterNativeToolsForMode() and filterMcpToolsForMode() utility functions
- Extracted filtering logic from Task.ts into dedicated module
- Applied same filtering approach used for XML tools in system prompt
- Added comprehensive test coverage (10 tests)
Impact:
- Model only sees tools allowed by current mode
- No more failed tool attempts due to mode restrictions
- Consistent behavior between XML and Native protocols
- Better UX with appropriate tool suggestions per mode
* refactor: eliminate repetitive tool checking using group-based approach
- Add getAvailableToolsInGroup() helper to check tools by group instead of individually
- Refactor filterNativeToolsForMode() to reuse getToolsForMode() instead of duplicating logic
- Simplify capabilities.ts by using group-based checks (60% reduction)
- Refactor rules.ts to use group helper (56% reduction)
- Remove debug console.log statements
- Update tests and snapshots
Benefits:
- Eliminates code duplication
- Leverages existing TOOL_GROUPS structure
- More maintainable - new tools in groups work automatically
- All tests passing (26/26)
* fix: add fallback to default mode when mode config not found
Ensures the agent always has functional tools even if:
- A custom mode is deleted while tasks still reference it
- Mode configuration becomes corrupted
- An invalid mode slug is provided
Without this fallback, the agent would have zero tools (not even
ask_followup_question or attempt_completion), completely breaking it.
* refactor(task): wrap initial user message in <feedback> instead of <task> to prevent focus drift after context-management
Rationale: After a successful context-management event, framing the next user block as feedback reduces model focus drift. Mentions parsing already supports <feedback>, and tool flows (attemptCompletion, responses) are aligned. No change to loop/persistence.
* refactor(mentions): drop <task> parsing; standardize on <feedback>; update tests
* Migrate conversation continuity to plugin-side encrypted reasoning items (Responses API)
Summary
We moved continuity off OpenAI servers and now maintain conversation state locally by persisting and replaying encrypted reasoning items. Requests are stateless (store=false) while retaining the performance/caching benefits of the Responses API.
Why
This aligns with how Roo manages context and simplifies our Responses API implementation while keeping all the benefits of continuity, caching, and latency improvements.
What changed
- All OpenAI models now use the Responses API; system instructions are passed via the top-level instructions field; requests include store=false and include=["reasoning.encrypted_content"].
- We persist encrypted reasoning items (type: "reasoning", encrypted_content, optional id) into API history and replay them on subsequent turns.
- Reasoning summaries default to summary: "auto" when supported; text.verbosity only when supported.
- Atomic persistence via safeWriteJson.
Removed
- previous_response_id flows, suppressPreviousResponseId/skipPrevResponseIdOnce, persistGpt5Metadata(), and GPT‑5 response ID metadata in UI messages.
Kept
- taskId and mode metadata for cross-provider features.
Result
- ZDR-friendly, stateless continuity with equal or better performance and a simpler codepath.
* fix(webview): remove unused metadata prop from ReasoningBlock render
* Responses API: retain response id for troubleshooting (not continuity)
Continuity is stateless via encrypted reasoning items that we persist and replay. We now capture the top-level response id in OpenAiNativeHandler and persist the assistant message id into api_conversation_history.json solely for debugging/correlation with provider logs; it is not used for continuity or control flow.
Also: silence request-body debug logging to avoid leaking prompts.
* remove DEPRECATED tests
* chore: remove unused Task types file to satisfy knip CI
* fix(task): properly type cleanConversationHistory and createMessage args in Task to address Dan's review
